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  11  *      The old ExecutorMain() has been replaced by ExecutorStart(),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) 1996-2007, 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 1.286 2007/02/02 00:07:02 tgl Exp $
  30  *
  31  *-------------------------------------------------------------------------
  32  */
  33 #include "postgres.h"
  34
  35 #include "access/heapam.h"
  36 #include "access/reloptions.h"
  37 #include "access/transam.h"
  38 #include "access/xact.h"
  39 #include "catalog/heap.h"
  40 #include "catalog/namespace.h"
  41 #include "catalog/toasting.h"
  42 #include "commands/tablespace.h"
  43 #include "commands/trigger.h"
  44 #include "executor/execdebug.h"
  45 #include "executor/instrument.h"
  46 #include "executor/nodeSubplan.h"
  47 #include "miscadmin.h"
  48 #include "optimizer/clauses.h"
  49 #include "parser/parse_clause.h"
  50 #include "parser/parsetree.h"
  51 #include "storage/smgr.h"
  52 #include "utils/acl.h"
  53 #include "utils/lsyscache.h"
  54 #include "utils/memutils.h"
  55
  56
  57 typedef struct evalPlanQual
  58 {
  59         Index           rti;
  60         EState     *estate;
  61         PlanState  *planstate;
  62         struct evalPlanQual *next;      /* stack of active PlanQual plans */
  63         struct evalPlanQual *free;      /* list of free PlanQual plans */
  64 } evalPlanQual;
  65
  66 /* decls for local routines only used within this module */
  67 static void InitPlan(QueryDesc *queryDesc, int eflags);
  68 static void initResultRelInfo(ResultRelInfo *resultRelInfo,
  69                                   Index resultRelationIndex,
  70                                   List *rangeTable,
  71                                   CmdType operation,
  72                                   bool doInstrument);
  73 static TupleTableSlot *ExecutePlan(EState *estate, PlanState *planstate,
  74                         CmdType operation,
  75                         long numberTuples,
  76                         ScanDirection direction,
  77                         DestReceiver *dest);
  78 static void ExecSelect(TupleTableSlot *slot,
  79                    DestReceiver *dest, EState *estate);
  80 static void ExecInsert(TupleTableSlot *slot, ItemPointer tupleid,
  81                    TupleTableSlot *planSlot,
  82                    DestReceiver *dest, EState *estate);
  83 static void ExecDelete(ItemPointer tupleid,
  84                    TupleTableSlot *planSlot,
  85                    DestReceiver *dest, EState *estate);
  86 static void ExecUpdate(TupleTableSlot *slot, ItemPointer tupleid,
  87                    TupleTableSlot *planSlot,
  88                    DestReceiver *dest, EState *estate);
  89 static void ExecProcessReturning(ProjectionInfo *projectReturning,
  90                                          TupleTableSlot *tupleSlot,
  91                                          TupleTableSlot *planSlot,
  92                                          DestReceiver *dest);
  93 static TupleTableSlot *EvalPlanQualNext(EState *estate);
  94 static void EndEvalPlanQual(EState *estate);
  95 static void ExecCheckRTEPerms(RangeTblEntry *rte);
  96 static void ExecCheckXactReadOnly(Query *parsetree);
  97 static void EvalPlanQualStart(evalPlanQual *epq, EState *estate,
  98                                   evalPlanQual *priorepq);
  99 static void EvalPlanQualStop(evalPlanQual *epq);
 100 static void OpenIntoRel(QueryDesc *queryDesc);
 101 static void CloseIntoRel(QueryDesc *queryDesc);
 102 static void intorel_startup(DestReceiver *self, int operation, TupleDesc typeinfo);
 103 static void intorel_receive(TupleTableSlot *slot, DestReceiver *self);
 104 static void intorel_shutdown(DestReceiver *self);
 105 static void intorel_destroy(DestReceiver *self);
 106
 107 /* end of local decls */
 108
 109
 110 /* ----------------------------------------------------------------
 111  *              ExecutorStart
 112  *
 113  *              This routine must be called at the beginning of any execution of any
 114  *              query plan
 115  *
 116  * Takes a QueryDesc previously created by CreateQueryDesc (it's not real
 117  * clear why we bother to separate the two functions, but...).  The tupDesc
 118  * field of the QueryDesc is filled in to describe the tuples that will be
 119  * returned, and the internal fields (estate and planstate) are set up.
 120  *
 121  * eflags contains flag bits as described in executor.h.
 122  *
 123  * NB: the CurrentMemoryContext when this is called will become the parent
 124  * of the per-query context used for this Executor invocation.
 125  * ----------------------------------------------------------------
 126  */
 127 void
 128 ExecutorStart(QueryDesc *queryDesc, int eflags)
 129 {
 130         EState     *estate;
 131         MemoryContext oldcontext;
 132
 133         /* sanity checks: queryDesc must not be started already */
 134         Assert(queryDesc != NULL);
 135         Assert(queryDesc->estate == NULL);
 136
 137         /*
 138          * If the transaction is read-only, we need to check if any writes are
 139          * planned to non-temporary tables.  EXPLAIN is considered read-only.
 140          */
 141         if (XactReadOnly && !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
 142                 ExecCheckXactReadOnly(queryDesc->parsetree);
 143
 144         /*
 145          * Build EState, switch into per-query memory context for startup.
 146          */
 147         estate = CreateExecutorState();
 148         queryDesc->estate = estate;
 149
 150         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 151
 152         /*
 153          * Fill in parameters, if any, from queryDesc
 154          */
 155         estate->es_param_list_info = queryDesc->params;
 156
 157         if (queryDesc->plantree->nParamExec > 0)
 158                 estate->es_param_exec_vals = (ParamExecData *)
 159                         palloc0(queryDesc->plantree->nParamExec * sizeof(ParamExecData));
 160
 161         /*
 162          * Copy other important information into the EState
 163          */
 164         estate->es_snapshot = queryDesc->snapshot;
 165         estate->es_crosscheck_snapshot = queryDesc->crosscheck_snapshot;
 166         estate->es_instrument = queryDesc->doInstrument;
 167
 168         /*
 169          * Initialize the plan state tree
 170          */
 171         InitPlan(queryDesc, eflags);
 172
 173         MemoryContextSwitchTo(oldcontext);
 174 }
 175
 176 /* ----------------------------------------------------------------
 177  *              ExecutorRun
 178  *
 179  *              This is the main routine of the executor module. It accepts
 180  *              the query descriptor from the traffic cop and executes the
 181  *              query plan.
 182  *
 183  *              ExecutorStart must have been called already.
 184  *
 185  *              If direction is NoMovementScanDirection then nothing is done
 186  *              except to start up/shut down the destination.  Otherwise,
 187  *              we retrieve up to 'count' tuples in the specified direction.
 188  *
 189  *              Note: count = 0 is interpreted as no portal limit, i.e., run to
 190  *              completion.
 191  *
 192  * ----------------------------------------------------------------
 193  */
 194 TupleTableSlot *
 195 ExecutorRun(QueryDesc *queryDesc,
 196                         ScanDirection direction, long count)
 197 {
 198         EState     *estate;
 199         CmdType         operation;
 200         DestReceiver *dest;
 201         bool            sendTuples;
 202         TupleTableSlot *result;
 203         MemoryContext oldcontext;
 204
 205         /* sanity checks */
 206         Assert(queryDesc != NULL);
 207
 208         estate = queryDesc->estate;
 209
 210         Assert(estate != NULL);
 211
 212         /*
 213          * Switch into per-query memory context
 214          */
 215         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 216
 217         /*
 218          * extract information from the query descriptor and the query feature.
 219          */
 220         operation = queryDesc->operation;
 221         dest = queryDesc->dest;
 222
 223         /*
 224          * startup tuple receiver, if we will be emitting tuples
 225          */
 226         estate->es_processed = 0;
 227         estate->es_lastoid = InvalidOid;
 228
 229         sendTuples = (operation == CMD_SELECT ||
 230                                   queryDesc->parsetree->returningList);
 231
 232         if (sendTuples)
 233                 (*dest->rStartup) (dest, operation, queryDesc->tupDesc);
 234
 235         /*
 236          * run plan
 237          */
 238         if (ScanDirectionIsNoMovement(direction))
 239                 result = NULL;
 240         else
 241                 result = ExecutePlan(estate,
 242                                                          queryDesc->planstate,
 243                                                          operation,
 244                                                          count,
 245                                                          direction,
 246                                                          dest);
 247
 248         /*
 249          * shutdown tuple receiver, if we started it
 250          */
 251         if (sendTuples)
 252                 (*dest->rShutdown) (dest);
 253
 254         MemoryContextSwitchTo(oldcontext);
 255
 256         return result;
 257 }
 258
 259 /* ----------------------------------------------------------------
 260  *              ExecutorEnd
 261  *
 262  *              This routine must be called at the end of execution of any
 263  *              query plan
 264  * ----------------------------------------------------------------
 265  */
 266 void
 267 ExecutorEnd(QueryDesc *queryDesc)
 268 {
 269         EState     *estate;
 270         MemoryContext oldcontext;
 271
 272         /* sanity checks */
 273         Assert(queryDesc != NULL);
 274
 275         estate = queryDesc->estate;
 276
 277         Assert(estate != NULL);
 278
 279         /*
 280          * Switch into per-query memory context to run ExecEndPlan
 281          */
 282         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 283
 284         ExecEndPlan(queryDesc->planstate, estate);
 285
 286         /*
 287          * Close the SELECT INTO relation if any
 288          */
 289         if (estate->es_select_into)
 290                 CloseIntoRel(queryDesc);
 291
 292         /*
 293          * Must switch out of context before destroying it
 294          */
 295         MemoryContextSwitchTo(oldcontext);
 296
 297         /*
 298          * Release EState and per-query memory context.  This should release
 299          * everything the executor has allocated.
 300          */
 301         FreeExecutorState(estate);
 302
 303         /* Reset queryDesc fields that no longer point to anything */
 304         queryDesc->tupDesc = NULL;
 305         queryDesc->estate = NULL;
 306         queryDesc->planstate = NULL;
 307 }
 308
 309 /* ----------------------------------------------------------------
 310  *              ExecutorRewind
 311  *
 312  *              This routine may be called on an open queryDesc to rewind it
 313  *              to the start.
 314  * ----------------------------------------------------------------
 315  */
 316 void
 317 ExecutorRewind(QueryDesc *queryDesc)
 318 {
 319         EState     *estate;
 320         MemoryContext oldcontext;
 321
 322         /* sanity checks */
 323         Assert(queryDesc != NULL);
 324
 325         estate = queryDesc->estate;
 326
 327         Assert(estate != NULL);
 328
 329         /* It's probably not sensible to rescan updating queries */
 330         Assert(queryDesc->operation == CMD_SELECT);
 331
 332         /*
 333          * Switch into per-query memory context
 334          */
 335         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 336
 337         /*
 338          * rescan plan
 339          */
 340         ExecReScan(queryDesc->planstate, NULL);
 341
 342         MemoryContextSwitchTo(oldcontext);
 343 }
 344
 345
 346 /*
 347  * ExecCheckRTPerms
 348  *              Check access permissions for all relations listed in a range table.
 349  */
 350 void
 351 ExecCheckRTPerms(List *rangeTable)
 352 {
 353         ListCell   *l;
 354
 355         foreach(l, rangeTable)
 356         {
 357                 RangeTblEntry *rte = lfirst(l);
 358
 359                 ExecCheckRTEPerms(rte);
 360         }
 361 }
 362
 363 /*
 364  * ExecCheckRTEPerms
 365  *              Check access permissions for a single RTE.
 366  */
 367 static void
 368 ExecCheckRTEPerms(RangeTblEntry *rte)
 369 {
 370         AclMode         requiredPerms;
 371         Oid                     relOid;
 372         Oid                     userid;
 373
 374         /*
 375          * Only plain-relation RTEs need to be checked here.  Subquery RTEs are
 376          * checked by ExecInitSubqueryScan if the subquery is still a separate
 377          * subquery --- if it's been pulled up into our query level then the RTEs
 378          * are in our rangetable and will be checked here. Function RTEs are
 379          * checked by init_fcache when the function is prepared for execution.
 380          * Join and special RTEs need no checks.
 381          */
 382         if (rte->rtekind != RTE_RELATION)
 383                 return;
 384
 385         /*
 386          * No work if requiredPerms is empty.
 387          */
 388         requiredPerms = rte->requiredPerms;
 389         if (requiredPerms == 0)
 390                 return;
 391
 392         relOid = rte->relid;
 393
 394         /*
 395          * userid to check as: current user unless we have a setuid indication.
 396          *
 397          * Note: GetUserId() is presently fast enough that there's no harm in
 398          * calling it separately for each RTE.  If that stops being true, we could
 399          * call it once in ExecCheckRTPerms and pass the userid down from there.
 400          * But for now, no need for the extra clutter.
 401          */
 402         userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
 403
 404         /*
 405          * We must have *all* the requiredPerms bits, so use aclmask not aclcheck.
 406          */
 407         if (pg_class_aclmask(relOid, userid, requiredPerms, ACLMASK_ALL)
 408                 != requiredPerms)
 409                 aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
 410                                            get_rel_name(relOid));
 411 }
 412
 413 /*
 414  * Check that the query does not imply any writes to non-temp tables.
 415  */
 416 static void
 417 ExecCheckXactReadOnly(Query *parsetree)
 418 {
 419         ListCell   *l;
 420
 421         /*
 422          * CREATE TABLE AS or SELECT INTO?
 423          *
 424          * XXX should we allow this if the destination is temp?
 425          */
 426         if (parsetree->into != NULL)
 427                 goto fail;
 428
 429         /* Fail if write permissions are requested on any non-temp table */
 430         foreach(l, parsetree->rtable)
 431         {
 432                 RangeTblEntry *rte = lfirst(l);
 433
 434                 if (rte->rtekind == RTE_SUBQUERY)
 435                 {
 436                         ExecCheckXactReadOnly(rte->subquery);
 437                         continue;
 438                 }
 439
 440                 if (rte->rtekind != RTE_RELATION)
 441                         continue;
 442
 443                 if ((rte->requiredPerms & (~ACL_SELECT)) == 0)
 444                         continue;
 445
 446                 if (isTempNamespace(get_rel_namespace(rte->relid)))
 447                         continue;
 448
 449                 goto fail;
 450         }
 451
 452         return;
 453
 454 fail:
 455         ereport(ERROR,
 456                         (errcode(ERRCODE_READ_ONLY_SQL_TRANSACTION),
 457                          errmsg("transaction is read-only")));
 458 }
 459
 460
 461 /* ----------------------------------------------------------------
 462  *              InitPlan
 463  *
 464  *              Initializes the query plan: open files, allocate storage
 465  *              and start up the rule manager
 466  * ----------------------------------------------------------------
 467  */
 468 static void
 469 InitPlan(QueryDesc *queryDesc, int eflags)
 470 {
 471         CmdType         operation = queryDesc->operation;
 472         Query      *parseTree = queryDesc->parsetree;
 473         Plan       *plan = queryDesc->plantree;
 474         EState     *estate = queryDesc->estate;
 475         PlanState  *planstate;
 476         List       *rangeTable;
 477         TupleDesc       tupType;
 478         ListCell   *l;
 479
 480         /*
 481          * Do permissions checks.  It's sufficient to examine the query's top
 482          * rangetable here --- subplan RTEs will be checked during
 483          * ExecInitSubPlan().
 484          */
 485         ExecCheckRTPerms(parseTree->rtable);
 486
 487         /*
 488          * get information from query descriptor
 489          */
 490         rangeTable = parseTree->rtable;
 491
 492         /*
 493          * initialize the node's execution state
 494          */
 495         estate->es_range_table = rangeTable;
 496
 497         /*
 498          * if there is a result relation, initialize result relation stuff
 499          */
 500         if (parseTree->resultRelation)
 501         {
 502                 List       *resultRelations = parseTree->resultRelations;
 503                 int                     numResultRelations;
 504                 ResultRelInfo *resultRelInfos;
 505
 506                 if (resultRelations != NIL)
 507                 {
 508                         /*
 509                          * Multiple result relations (due to inheritance)
 510                          * parseTree->resultRelations identifies them all
 511                          */
 512                         ResultRelInfo *resultRelInfo;
 513
 514                         numResultRelations = list_length(resultRelations);
 515                         resultRelInfos = (ResultRelInfo *)
 516                                 palloc(numResultRelations * sizeof(ResultRelInfo));
 517                         resultRelInfo = resultRelInfos;
 518                         foreach(l, resultRelations)
 519                         {
 520                                 initResultRelInfo(resultRelInfo,
 521                                                                   lfirst_int(l),
 522                                                                   rangeTable,
 523                                                                   operation,
 524                                                                   estate->es_instrument);
 525                                 resultRelInfo++;
 526                         }
 527                 }
 528                 else
 529                 {
 530                         /*
 531                          * Single result relation identified by parseTree->resultRelation
 532                          */
 533                         numResultRelations = 1;
 534                         resultRelInfos = (ResultRelInfo *) palloc(sizeof(ResultRelInfo));
 535                         initResultRelInfo(resultRelInfos,
 536                                                           parseTree->resultRelation,
 537                                                           rangeTable,
 538                                                           operation,
 539                                                           estate->es_instrument);
 540                 }
 541
 542                 estate->es_result_relations = resultRelInfos;
 543                 estate->es_num_result_relations = numResultRelations;
 544                 /* Initialize to first or only result rel */
 545                 estate->es_result_relation_info = resultRelInfos;
 546         }
 547         else
 548         {
 549                 /*
 550                  * if no result relation, then set state appropriately
 551                  */
 552                 estate->es_result_relations = NULL;
 553                 estate->es_num_result_relations = 0;
 554                 estate->es_result_relation_info = NULL;
 555         }
 556
 557         /*
 558          * Detect whether we're doing SELECT INTO.  If so, set the es_into_oids
 559          * flag appropriately so that the plan tree will be initialized with the
 560          * correct tuple descriptors.  (Other SELECT INTO stuff comes later.)
 561          */
 562         estate->es_select_into = false;
 563         if (operation == CMD_SELECT && parseTree->into != NULL)
 564         {
 565                 estate->es_select_into = true;
 566                 estate->es_into_oids = interpretOidsOption(parseTree->intoOptions);
 567         }
 568
 569         /*
 570          * Have to lock relations selected FOR UPDATE/FOR SHARE before we
 571          * initialize the plan tree, else we'd be doing a lock upgrade.
 572          * While we are at it, build the ExecRowMark list.
 573          */
 574         estate->es_rowMarks = NIL;
 575         foreach(l, parseTree->rowMarks)
 576         {
 577                 RowMarkClause *rc = (RowMarkClause *) lfirst(l);
 578                 Oid                     relid = getrelid(rc->rti, rangeTable);
 579                 Relation        relation;
 580                 ExecRowMark *erm;
 581
 582                 relation = heap_open(relid, RowShareLock);
 583                 erm = (ExecRowMark *) palloc(sizeof(ExecRowMark));
 584                 erm->relation = relation;
 585                 erm->rti = rc->rti;
 586                 erm->forUpdate = rc->forUpdate;
 587                 erm->noWait = rc->noWait;
 588                 /* We'll set up ctidAttno below */
 589                 erm->ctidAttNo = InvalidAttrNumber;
 590                 estate->es_rowMarks = lappend(estate->es_rowMarks, erm);
 591         }
 592
 593         /*
 594          * initialize the executor "tuple" table.  We need slots for all the plan
 595          * nodes, plus possibly output slots for the junkfilter(s). At this point
 596          * we aren't sure if we need junkfilters, so just add slots for them
 597          * unconditionally.  Also, if it's not a SELECT, set up a slot for use for
 598          * trigger output tuples.
 599          */
 600         {
 601                 int                     nSlots = ExecCountSlotsNode(plan);
 602
 603                 if (parseTree->resultRelations != NIL)
 604                         nSlots += list_length(parseTree->resultRelations);
 605                 else
 606                         nSlots += 1;
 607                 if (operation != CMD_SELECT)
 608                         nSlots++;                       /* for es_trig_tuple_slot */
 609                 if (parseTree->returningLists)
 610                         nSlots++;                       /* for RETURNING projection */
 611
 612                 estate->es_tupleTable = ExecCreateTupleTable(nSlots);
 613
 614                 if (operation != CMD_SELECT)
 615                         estate->es_trig_tuple_slot =
 616                                 ExecAllocTableSlot(estate->es_tupleTable);
 617         }
 618
 619         /* mark EvalPlanQual not active */
 620         estate->es_topPlan = plan;
 621         estate->es_evalPlanQual = NULL;
 622         estate->es_evTupleNull = NULL;
 623         estate->es_evTuple = NULL;
 624         estate->es_useEvalPlan = false;
 625
 626         /*
 627          * initialize the private state information for all the nodes in the query
 628          * tree.  This opens files, allocates storage and leaves us ready to start
 629          * processing tuples.
 630          */
 631         planstate = ExecInitNode(plan, estate, eflags);
 632
 633         /*
 634          * Get the tuple descriptor describing the type of tuples to return. (this
 635          * is especially important if we are creating a relation with "SELECT
 636          * INTO")
 637          */
 638         tupType = ExecGetResultType(planstate);
 639
 640         /*
 641          * Initialize the junk filter if needed.  SELECT and INSERT queries need a
 642          * filter if there are any junk attrs in the tlist.  INSERT and SELECT
 643          * INTO also need a filter if the plan may return raw disk tuples (else
 644          * heap_insert will be scribbling on the source relation!). UPDATE and
 645          * DELETE always need a filter, since there's always a junk 'ctid'
 646          * attribute present --- no need to look first.
 647          */
 648         {
 649                 bool            junk_filter_needed = false;
 650                 ListCell   *tlist;
 651
 652                 switch (operation)
 653                 {
 654                         case CMD_SELECT:
 655                         case CMD_INSERT:
 656                                 foreach(tlist, plan->targetlist)
 657                                 {
 658                                         TargetEntry *tle = (TargetEntry *) lfirst(tlist);
 659
 660                                         if (tle->resjunk)
 661                                         {
 662                                                 junk_filter_needed = true;
 663                                                 break;
 664                                         }
 665                                 }
 666                                 if (!junk_filter_needed &&
 667                                         (operation == CMD_INSERT || estate->es_select_into) &&
 668                                         ExecMayReturnRawTuples(planstate))
 669                                         junk_filter_needed = true;
 670                                 break;
 671                         case CMD_UPDATE:
 672                         case CMD_DELETE:
 673                                 junk_filter_needed = true;
 674                                 break;
 675                         default:
 676                                 break;
 677                 }
 678
 679                 if (junk_filter_needed)
 680                 {
 681                         /*
 682                          * If there are multiple result relations, each one needs its own
 683                          * junk filter.  Note this is only possible for UPDATE/DELETE, so
 684                          * we can't be fooled by some needing a filter and some not.
 685                          */
 686                         if (parseTree->resultRelations != NIL)
 687                         {
 688                                 PlanState **appendplans;
 689                                 int                     as_nplans;
 690                                 ResultRelInfo *resultRelInfo;
 691                                 int                     i;
 692
 693                                 /* Top plan had better be an Append here. */
 694                                 Assert(IsA(plan, Append));
 695                                 Assert(((Append *) plan)->isTarget);
 696                                 Assert(IsA(planstate, AppendState));
 697                                 appendplans = ((AppendState *) planstate)->appendplans;
 698                                 as_nplans = ((AppendState *) planstate)->as_nplans;
 699                                 Assert(as_nplans == estate->es_num_result_relations);
 700                                 resultRelInfo = estate->es_result_relations;
 701                                 for (i = 0; i < as_nplans; i++)
 702                                 {
 703                                         PlanState  *subplan = appendplans[i];
 704                                         JunkFilter *j;
 705
 706                                         j = ExecInitJunkFilter(subplan->plan->targetlist,
 707                                                         resultRelInfo->ri_RelationDesc->rd_att->tdhasoid,
 708                                                                   ExecAllocTableSlot(estate->es_tupleTable));
 709                                         /*
 710                                          * Since it must be UPDATE/DELETE, there had better be
 711                                          * a "ctid" junk attribute in the tlist ... but ctid could
 712                                          * be at a different resno for each result relation.
 713                                          * We look up the ctid resnos now and save them in the
 714                                          * junkfilters.
 715                                          */
 716                                         j->jf_junkAttNo = ExecFindJunkAttribute(j, "ctid");
 717                                         if (!AttributeNumberIsValid(j->jf_junkAttNo))
 718                                                 elog(ERROR, "could not find junk ctid column");
 719                                         resultRelInfo->ri_junkFilter = j;
 720                                         resultRelInfo++;
 721                                 }
 722
 723                                 /*
 724                                  * Set active junkfilter too; at this point ExecInitAppend has
 725                                  * already selected an active result relation...
 726                                  */
 727                                 estate->es_junkFilter =
 728                                         estate->es_result_relation_info->ri_junkFilter;
 729                         }
 730                         else
 731                         {
 732                                 /* Normal case with just one JunkFilter */
 733                                 JunkFilter *j;
 734
 735                                 j = ExecInitJunkFilter(planstate->plan->targetlist,
 736                                                                            tupType->tdhasoid,
 737                                                                   ExecAllocTableSlot(estate->es_tupleTable));
 738                                 estate->es_junkFilter = j;
 739                                 if (estate->es_result_relation_info)
 740                                         estate->es_result_relation_info->ri_junkFilter = j;
 741
 742                                 if (operation == CMD_SELECT)
 743                                 {
 744                                         /* For SELECT, want to return the cleaned tuple type */
 745                                         tupType = j->jf_cleanTupType;
 746                                         /* For SELECT FOR UPDATE/SHARE, find the ctid attrs now */
 747                                         foreach(l, estate->es_rowMarks)
 748                                         {
 749                                                 ExecRowMark *erm = (ExecRowMark *) lfirst(l);
 750                                                 char            resname[32];
 751
 752                                                 snprintf(resname, sizeof(resname), "ctid%u", erm->rti);
 753                                                 erm->ctidAttNo = ExecFindJunkAttribute(j, resname);
 754                                                 if (!AttributeNumberIsValid(erm->ctidAttNo))
 755                                                         elog(ERROR, "could not find junk \"%s\" column",
 756                                                                  resname);
 757                                         }
 758                                 }
 759                                 else if (operation == CMD_UPDATE || operation == CMD_DELETE)
 760                                 {
 761                                         /* For UPDATE/DELETE, find the ctid junk attr now */
 762                                         j->jf_junkAttNo = ExecFindJunkAttribute(j, "ctid");
 763                                         if (!AttributeNumberIsValid(j->jf_junkAttNo))
 764                                                 elog(ERROR, "could not find junk ctid column");
 765                                 }
 766                         }
 767                 }
 768                 else
 769                         estate->es_junkFilter = NULL;
 770         }
 771
 772         /*
 773          * Initialize RETURNING projections if needed.
 774          */
 775         if (parseTree->returningLists)
 776         {
 777                 TupleTableSlot *slot;
 778                 ExprContext *econtext;
 779                 ResultRelInfo *resultRelInfo;
 780
 781                 /*
 782                  * We set QueryDesc.tupDesc to be the RETURNING rowtype in this case.
 783                  * We assume all the sublists will generate the same output tupdesc.
 784                  */
 785                 tupType = ExecTypeFromTL((List *) linitial(parseTree->returningLists),
 786                                                                  false);
 787
 788                 /* Set up a slot for the output of the RETURNING projection(s) */
 789                 slot = ExecAllocTableSlot(estate->es_tupleTable);
 790                 ExecSetSlotDescriptor(slot, tupType);
 791                 /* Need an econtext too */
 792                 econtext = CreateExprContext(estate);
 793
 794                 /*
 795                  * Build a projection for each result rel.      Note that any SubPlans in
 796                  * the RETURNING lists get attached to the topmost plan node.
 797                  */
 798                 Assert(list_length(parseTree->returningLists) == estate->es_num_result_relations);
 799                 resultRelInfo = estate->es_result_relations;
 800                 foreach(l, parseTree->returningLists)
 801                 {
 802                         List       *rlist = (List *) lfirst(l);
 803                         List       *rliststate;
 804
 805                         rliststate = (List *) ExecInitExpr((Expr *) rlist, planstate);
 806                         resultRelInfo->ri_projectReturning =
 807                                 ExecBuildProjectionInfo(rliststate, econtext, slot,
 808                                                                            resultRelInfo->ri_RelationDesc->rd_att);
 809                         resultRelInfo++;
 810                 }
 811
 812                 /*
 813                  * Because we already ran ExecInitNode() for the top plan node, any
 814                  * subplans we just attached to it won't have been initialized; so we
 815                  * have to do it here.  (Ugly, but the alternatives seem worse.)
 816                  */
 817                 foreach(l, planstate->subPlan)
 818                 {
 819                         SubPlanState *sstate = (SubPlanState *) lfirst(l);
 820
 821                         Assert(IsA(sstate, SubPlanState));
 822                         if (sstate->planstate == NULL)          /* already inited? */
 823                                 ExecInitSubPlan(sstate, estate, eflags);
 824                 }
 825         }
 826
 827         queryDesc->tupDesc = tupType;
 828         queryDesc->planstate = planstate;
 829
 830         /*
 831          * If doing SELECT INTO, initialize the "into" relation.  We must wait
 832          * till now so we have the "clean" result tuple type to create the new
 833          * table from.
 834          *
 835          * If EXPLAIN, skip creating the "into" relation.
 836          */
 837         if (estate->es_select_into && !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
 838                 OpenIntoRel(queryDesc);
 839 }
 840
 841 /*
 842  * Initialize ResultRelInfo data for one result relation
 843  */
 844 static void
 845 initResultRelInfo(ResultRelInfo *resultRelInfo,
 846                                   Index resultRelationIndex,
 847                                   List *rangeTable,
 848                                   CmdType operation,
 849                                   bool doInstrument)
 850 {
 851         Oid                     resultRelationOid;
 852         Relation        resultRelationDesc;
 853
 854         resultRelationOid = getrelid(resultRelationIndex, rangeTable);
 855         resultRelationDesc = heap_open(resultRelationOid, RowExclusiveLock);
 856
 857         switch (resultRelationDesc->rd_rel->relkind)
 858         {
 859                 case RELKIND_SEQUENCE:
 860                         ereport(ERROR,
 861                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 862                                          errmsg("cannot change sequence \"%s\"",
 863                                                         RelationGetRelationName(resultRelationDesc))));
 864                         break;
 865                 case RELKIND_TOASTVALUE:
 866                         ereport(ERROR,
 867                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 868                                          errmsg("cannot change TOAST relation \"%s\"",
 869                                                         RelationGetRelationName(resultRelationDesc))));
 870                         break;
 871                 case RELKIND_VIEW:
 872                         ereport(ERROR,
 873                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 874                                          errmsg("cannot change view \"%s\"",
 875                                                         RelationGetRelationName(resultRelationDesc))));
 876                         break;
 877         }
 878
 879         MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
 880         resultRelInfo->type = T_ResultRelInfo;
 881         resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
 882         resultRelInfo->ri_RelationDesc = resultRelationDesc;
 883         resultRelInfo->ri_NumIndices = 0;
 884         resultRelInfo->ri_IndexRelationDescs = NULL;
 885         resultRelInfo->ri_IndexRelationInfo = NULL;
 886         /* make a copy so as not to depend on relcache info not changing... */
 887         resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
 888         if (resultRelInfo->ri_TrigDesc)
 889         {
 890                 int                     n = resultRelInfo->ri_TrigDesc->numtriggers;
 891
 892                 resultRelInfo->ri_TrigFunctions = (FmgrInfo *)
 893                         palloc0(n * sizeof(FmgrInfo));
 894                 if (doInstrument)
 895                         resultRelInfo->ri_TrigInstrument = InstrAlloc(n);
 896                 else
 897                         resultRelInfo->ri_TrigInstrument = NULL;
 898         }
 899         else
 900         {
 901                 resultRelInfo->ri_TrigFunctions = NULL;
 902                 resultRelInfo->ri_TrigInstrument = NULL;
 903         }
 904         resultRelInfo->ri_ConstraintExprs = NULL;
 905         resultRelInfo->ri_junkFilter = NULL;
 906         resultRelInfo->ri_projectReturning = NULL;
 907
 908         /*
 909          * If there are indices on the result relation, open them and save
 910          * descriptors in the result relation info, so that we can add new index
 911          * entries for the tuples we add/update.  We need not do this for a
 912          * DELETE, however, since deletion doesn't affect indexes.
 913          */
 914         if (resultRelationDesc->rd_rel->relhasindex &&
 915                 operation != CMD_DELETE)
 916                 ExecOpenIndices(resultRelInfo);
 917 }
 918
 919 /*
 920  *              ExecContextForcesOids
 921  *
 922  * This is pretty grotty: when doing INSERT, UPDATE, or SELECT INTO,
 923  * we need to ensure that result tuples have space for an OID iff they are
 924  * going to be stored into a relation that has OIDs.  In other contexts
 925  * we are free to choose whether to leave space for OIDs in result tuples
 926  * (we generally don't want to, but we do if a physical-tlist optimization
 927  * is possible).  This routine checks the plan context and returns TRUE if the
 928  * choice is forced, FALSE if the choice is not forced.  In the TRUE case,
 929  * *hasoids is set to the required value.
 930  *
 931  * One reason this is ugly is that all plan nodes in the plan tree will emit
 932  * tuples with space for an OID, though we really only need the topmost node
 933  * to do so.  However, node types like Sort don't project new tuples but just
 934  * return their inputs, and in those cases the requirement propagates down
 935  * to the input node.  Eventually we might make this code smart enough to
 936  * recognize how far down the requirement really goes, but for now we just
 937  * make all plan nodes do the same thing if the top level forces the choice.
 938  *
 939  * We assume that estate->es_result_relation_info is already set up to
 940  * describe the target relation.  Note that in an UPDATE that spans an
 941  * inheritance tree, some of the target relations may have OIDs and some not.
 942  * We have to make the decisions on a per-relation basis as we initialize
 943  * each of the child plans of the topmost Append plan.
 944  *
 945  * SELECT INTO is even uglier, because we don't have the INTO relation's
 946  * descriptor available when this code runs; we have to look aside at a
 947  * flag set by InitPlan().
 948  */
 949 bool
 950 ExecContextForcesOids(PlanState *planstate, bool *hasoids)
 951 {
 952         if (planstate->state->es_select_into)
 953         {
 954                 *hasoids = planstate->state->es_into_oids;
 955                 return true;
 956         }
 957         else
 958         {
 959                 ResultRelInfo *ri = planstate->state->es_result_relation_info;
 960
 961                 if (ri != NULL)
 962                 {
 963                         Relation        rel = ri->ri_RelationDesc;
 964
 965                         if (rel != NULL)
 966                         {
 967                                 *hasoids = rel->rd_rel->relhasoids;
 968                                 return true;
 969                         }
 970                 }
 971         }
 972
 973         return false;
 974 }
 975
 976 /* ----------------------------------------------------------------
 977  *              ExecEndPlan
 978  *
 979  *              Cleans up the query plan -- closes files and frees up storage
 980  *
 981  * NOTE: we are no longer very worried about freeing storage per se
 982  * in this code; FreeExecutorState should be guaranteed to release all
 983  * memory that needs to be released.  What we are worried about doing
 984  * is closing relations and dropping buffer pins.  Thus, for example,
 985  * tuple tables must be cleared or dropped to ensure pins are released.
 986  * ----------------------------------------------------------------
 987  */
 988 void
 989 ExecEndPlan(PlanState *planstate, EState *estate)
 990 {
 991         ResultRelInfo *resultRelInfo;
 992         int                     i;
 993         ListCell   *l;
 994
 995         /*
 996          * shut down any PlanQual processing we were doing
 997          */
 998         if (estate->es_evalPlanQual != NULL)
 999                 EndEvalPlanQual(estate);
1000
1001         /*
1002          * shut down the node-type-specific query processing
1003          */
1004         ExecEndNode(planstate);
1005
1006         /*
1007          * destroy the executor "tuple" table.
1008          */
1009         ExecDropTupleTable(estate->es_tupleTable, true);
1010         estate->es_tupleTable = NULL;
1011
1012         /*
1013          * close the result relation(s) if any, but hold locks until xact commit.
1014          */
1015         resultRelInfo = estate->es_result_relations;
1016         for (i = estate->es_num_result_relations; i > 0; i--)
1017         {
1018                 /* Close indices and then the relation itself */
1019                 ExecCloseIndices(resultRelInfo);
1020                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
1021                 resultRelInfo++;
1022         }
1023
1024         /*
1025          * close any relations selected FOR UPDATE/FOR SHARE, again keeping locks
1026          */
1027         foreach(l, estate->es_rowMarks)
1028         {
1029                 ExecRowMark *erm = lfirst(l);
1030
1031                 heap_close(erm->relation, NoLock);
1032         }
1033 }
1034
1035 /* ----------------------------------------------------------------
1036  *              ExecutePlan
1037  *
1038  *              processes the query plan to retrieve 'numberTuples' tuples in the
1039  *              direction specified.
1040  *
1041  *              Retrieves all tuples if numberTuples is 0
1042  *
1043  *              result is either a slot containing the last tuple in the case
1044  *              of a SELECT or NULL otherwise.
1045  *
1046  * Note: the ctid attribute is a 'junk' attribute that is removed before the
1047  * user can see it
1048  * ----------------------------------------------------------------
1049  */
1050 static TupleTableSlot *
1051 ExecutePlan(EState *estate,
1052                         PlanState *planstate,
1053                         CmdType operation,
1054                         long numberTuples,
1055                         ScanDirection direction,
1056                         DestReceiver *dest)
1057 {
1058         JunkFilter *junkfilter;
1059         TupleTableSlot *planSlot;
1060         TupleTableSlot *slot;
1061         ItemPointer tupleid = NULL;
1062         ItemPointerData tuple_ctid;
1063         long            current_tuple_count;
1064         TupleTableSlot *result;
1065
1066         /*
1067          * initialize local variables
1068          */
1069         current_tuple_count = 0;
1070         result = NULL;
1071
1072         /*
1073          * Set the direction.
1074          */
1075         estate->es_direction = direction;
1076
1077         /*
1078          * Process BEFORE EACH STATEMENT triggers
1079          */
1080         switch (operation)
1081         {
1082                 case CMD_UPDATE:
1083                         ExecBSUpdateTriggers(estate, estate->es_result_relation_info);
1084                         break;
1085                 case CMD_DELETE:
1086                         ExecBSDeleteTriggers(estate, estate->es_result_relation_info);
1087                         break;
1088                 case CMD_INSERT:
1089                         ExecBSInsertTriggers(estate, estate->es_result_relation_info);
1090                         break;
1091                 default:
1092                         /* do nothing */
1093                         break;
1094         }
1095
1096         /*
1097          * Loop until we've processed the proper number of tuples from the plan.
1098          */
1099
1100         for (;;)
1101         {
1102                 /* Reset the per-output-tuple exprcontext */
1103                 ResetPerTupleExprContext(estate);
1104
1105                 /*
1106                  * Execute the plan and obtain a tuple
1107                  */
1108 lnext:  ;
1109                 if (estate->es_useEvalPlan)
1110                 {
1111                         planSlot = EvalPlanQualNext(estate);
1112                         if (TupIsNull(planSlot))
1113                                 planSlot = ExecProcNode(planstate);
1114                 }
1115                 else
1116                         planSlot = ExecProcNode(planstate);
1117
1118                 /*
1119                  * if the tuple is null, then we assume there is nothing more to
1120                  * process so we just return null...
1121                  */
1122                 if (TupIsNull(planSlot))
1123                 {
1124                         result = NULL;
1125                         break;
1126                 }
1127                 slot = planSlot;
1128
1129                 /*
1130                  * if we have a junk filter, then project a new tuple with the junk
1131                  * removed.
1132                  *
1133                  * Store this new "clean" tuple in the junkfilter's resultSlot.
1134                  * (Formerly, we stored it back over the "dirty" tuple, which is WRONG
1135                  * because that tuple slot has the wrong descriptor.)
1136                  *
1137                  * Also, extract all the junk information we need.
1138                  */
1139                 if ((junkfilter = estate->es_junkFilter) != NULL)
1140                 {
1141                         Datum           datum;
1142                         bool            isNull;
1143
1144                         /*
1145                          * extract the 'ctid' junk attribute.
1146                          */
1147                         if (operation == CMD_UPDATE || operation == CMD_DELETE)
1148                         {
1149                                 datum = ExecGetJunkAttribute(slot, junkfilter->jf_junkAttNo,
1150                                                                                          &isNull);
1151                                 /* shouldn't ever get a null result... */
1152                                 if (isNull)
1153                                         elog(ERROR, "ctid is NULL");
1154
1155                                 tupleid = (ItemPointer) DatumGetPointer(datum);
1156                                 tuple_ctid = *tupleid;  /* make sure we don't free the ctid!! */
1157                                 tupleid = &tuple_ctid;
1158                         }
1159
1160                         /*
1161                          * Process any FOR UPDATE or FOR SHARE locking requested.
1162                          */
1163                         else if (estate->es_rowMarks != NIL)
1164                         {
1165                                 ListCell   *l;
1166
1167                 lmark:  ;
1168                                 foreach(l, estate->es_rowMarks)
1169                                 {
1170                                         ExecRowMark *erm = lfirst(l);
1171                                         HeapTupleData tuple;
1172                                         Buffer          buffer;
1173                                         ItemPointerData update_ctid;
1174                                         TransactionId update_xmax;
1175                                         TupleTableSlot *newSlot;
1176                                         LockTupleMode lockmode;
1177                                         HTSU_Result test;
1178
1179                                         datum = ExecGetJunkAttribute(slot,
1180                                                                                                  erm->ctidAttNo,
1181                                                                                                  &isNull);
1182                                         /* shouldn't ever get a null result... */
1183                                         if (isNull)
1184                                                 elog(ERROR, "ctid is NULL");
1185
1186                                         tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
1187
1188                                         if (erm->forUpdate)
1189                                                 lockmode = LockTupleExclusive;
1190                                         else
1191                                                 lockmode = LockTupleShared;
1192
1193                                         test = heap_lock_tuple(erm->relation, &tuple, &buffer,
1194                                                                                    &update_ctid, &update_xmax,
1195                                                                                    estate->es_snapshot->curcid,
1196                                                                                    lockmode, erm->noWait);
1197                                         ReleaseBuffer(buffer);
1198                                         switch (test)
1199                                         {
1200                                                 case HeapTupleSelfUpdated:
1201                                                         /* treat it as deleted; do not process */
1202                                                         goto lnext;
1203
1204                                                 case HeapTupleMayBeUpdated:
1205                                                         break;
1206
1207                                                 case HeapTupleUpdated:
1208                                                         if (IsXactIsoLevelSerializable)
1209                                                                 ereport(ERROR,
1210                                                                  (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1211                                                                   errmsg("could not serialize access due to concurrent update")));
1212                                                         if (!ItemPointerEquals(&update_ctid,
1213                                                                                                    &tuple.t_self))
1214                                                         {
1215                                                                 /* updated, so look at updated version */
1216                                                                 newSlot = EvalPlanQual(estate,
1217                                                                                                            erm->rti,
1218                                                                                                            &update_ctid,
1219                                                                                                            update_xmax,
1220                                                                                                 estate->es_snapshot->curcid);
1221                                                                 if (!TupIsNull(newSlot))
1222                                                                 {
1223                                                                         slot = planSlot = newSlot;
1224                                                                         estate->es_useEvalPlan = true;
1225                                                                         goto lmark;
1226                                                                 }
1227                                                         }
1228
1229                                                         /*
1230                                                          * if tuple was deleted or PlanQual failed for
1231                                                          * updated tuple - we must not return this tuple!
1232                                                          */
1233                                                         goto lnext;
1234
1235                                                 default:
1236                                                         elog(ERROR, "unrecognized heap_lock_tuple status: %u",
1237                                                                  test);
1238                                                         return NULL;
1239                                         }
1240                                 }
1241                         }
1242
1243                         /*
1244                          * Create a new "clean" tuple with all junk attributes removed. We
1245                          * don't need to do this for DELETE, however (there will in fact
1246                          * be no non-junk attributes in a DELETE!)
1247                          */
1248                         if (operation != CMD_DELETE)
1249                                 slot = ExecFilterJunk(junkfilter, slot);
1250                 }
1251
1252                 /*
1253                  * now that we have a tuple, do the appropriate thing with it.. either
1254                  * return it to the user, add it to a relation someplace, delete it
1255                  * from a relation, or modify some of its attributes.
1256                  */
1257                 switch (operation)
1258                 {
1259                         case CMD_SELECT:
1260                                 ExecSelect(slot, dest, estate);
1261                                 result = slot;
1262                                 break;
1263
1264                         case CMD_INSERT:
1265                                 ExecInsert(slot, tupleid, planSlot, dest, estate);
1266                                 result = NULL;
1267                                 break;
1268
1269                         case CMD_DELETE:
1270                                 ExecDelete(tupleid, planSlot, dest, estate);
1271                                 result = NULL;
1272                                 break;
1273
1274                         case CMD_UPDATE:
1275                                 ExecUpdate(slot, tupleid, planSlot, dest, estate);
1276                                 result = NULL;
1277                                 break;
1278
1279                         default:
1280                                 elog(ERROR, "unrecognized operation code: %d",
1281                                          (int) operation);
1282                                 result = NULL;
1283                                 break;
1284                 }
1285
1286                 /*
1287                  * check our tuple count.. if we've processed the proper number then
1288                  * quit, else loop again and process more tuples.  Zero numberTuples
1289                  * means no limit.
1290                  */
1291                 current_tuple_count++;
1292                 if (numberTuples && numberTuples == current_tuple_count)
1293                         break;
1294         }
1295
1296         /*
1297          * Process AFTER EACH STATEMENT triggers
1298          */
1299         switch (operation)
1300         {
1301                 case CMD_UPDATE:
1302                         ExecASUpdateTriggers(estate, estate->es_result_relation_info);
1303                         break;
1304                 case CMD_DELETE:
1305                         ExecASDeleteTriggers(estate, estate->es_result_relation_info);
1306                         break;
1307                 case CMD_INSERT:
1308                         ExecASInsertTriggers(estate, estate->es_result_relation_info);
1309                         break;
1310                 default:
1311                         /* do nothing */
1312                         break;
1313         }
1314
1315         /*
1316          * here, result is either a slot containing a tuple in the case of a
1317          * SELECT or NULL otherwise.
1318          */
1319         return result;
1320 }
1321
1322 /* ----------------------------------------------------------------
1323  *              ExecSelect
1324  *
1325  *              SELECTs are easy.. we just pass the tuple to the appropriate
1326  *              output function.
1327  * ----------------------------------------------------------------
1328  */
1329 static void
1330 ExecSelect(TupleTableSlot *slot,
1331                    DestReceiver *dest,
1332                    EState *estate)
1333 {
1334         (*dest->receiveSlot) (slot, dest);
1335         IncrRetrieved();
1336         (estate->es_processed)++;
1337 }
1338
1339 /* ----------------------------------------------------------------
1340  *              ExecInsert
1341  *
1342  *              INSERTs are trickier.. we have to insert the tuple into
1343  *              the base relation and insert appropriate tuples into the
1344  *              index relations.
1345  * ----------------------------------------------------------------
1346  */
1347 static void
1348 ExecInsert(TupleTableSlot *slot,
1349                    ItemPointer tupleid,
1350                    TupleTableSlot *planSlot,
1351                    DestReceiver *dest,
1352                    EState *estate)
1353 {
1354         HeapTuple       tuple;
1355         ResultRelInfo *resultRelInfo;
1356         Relation        resultRelationDesc;
1357         Oid                     newId;
1358
1359         /*
1360          * get the heap tuple out of the tuple table slot, making sure we have a
1361          * writable copy
1362          */
1363         tuple = ExecMaterializeSlot(slot);
1364
1365         /*
1366          * get information on the (current) result relation
1367          */
1368         resultRelInfo = estate->es_result_relation_info;
1369         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1370
1371         /* BEFORE ROW INSERT Triggers */
1372         if (resultRelInfo->ri_TrigDesc &&
1373                 resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_INSERT] > 0)
1374         {
1375                 HeapTuple       newtuple;
1376
1377                 newtuple = ExecBRInsertTriggers(estate, resultRelInfo, tuple);
1378
1379                 if (newtuple == NULL)   /* "do nothing" */
1380                         return;
1381
1382                 if (newtuple != tuple)  /* modified by Trigger(s) */
1383                 {
1384                         /*
1385                          * Put the modified tuple into a slot for convenience of routines
1386                          * below.  We assume the tuple was allocated in per-tuple memory
1387                          * context, and therefore will go away by itself. The tuple table
1388                          * slot should not try to clear it.
1389                          */
1390                         TupleTableSlot *newslot = estate->es_trig_tuple_slot;
1391
1392                         if (newslot->tts_tupleDescriptor != slot->tts_tupleDescriptor)
1393                                 ExecSetSlotDescriptor(newslot, slot->tts_tupleDescriptor);
1394                         ExecStoreTuple(newtuple, newslot, InvalidBuffer, false);
1395                         slot = newslot;
1396                         tuple = newtuple;
1397                 }
1398         }
1399
1400         /*
1401          * Check the constraints of the tuple
1402          */
1403         if (resultRelationDesc->rd_att->constr)
1404                 ExecConstraints(resultRelInfo, slot, estate);
1405
1406         /*
1407          * insert the tuple
1408          *
1409          * Note: heap_insert returns the tid (location) of the new tuple in the
1410          * t_self field.
1411          */
1412         newId = heap_insert(resultRelationDesc, tuple,
1413                                                 estate->es_snapshot->curcid,
1414                                                 true, true);
1415
1416         IncrAppended();
1417         (estate->es_processed)++;
1418         estate->es_lastoid = newId;
1419         setLastTid(&(tuple->t_self));
1420
1421         /*
1422          * insert index entries for tuple
1423          */
1424         if (resultRelInfo->ri_NumIndices > 0)
1425                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1426
1427         /* AFTER ROW INSERT Triggers */
1428         ExecARInsertTriggers(estate, resultRelInfo, tuple);
1429
1430         /* Process RETURNING if present */
1431         if (resultRelInfo->ri_projectReturning)
1432                 ExecProcessReturning(resultRelInfo->ri_projectReturning,
1433                                                          slot, planSlot, dest);
1434 }
1435
1436 /* ----------------------------------------------------------------
1437  *              ExecDelete
1438  *
1439  *              DELETE is like UPDATE, except that we delete the tuple and no
1440  *              index modifications are needed
1441  * ----------------------------------------------------------------
1442  */
1443 static void
1444 ExecDelete(ItemPointer tupleid,
1445                    TupleTableSlot *planSlot,
1446                    DestReceiver *dest,
1447                    EState *estate)
1448 {
1449         ResultRelInfo *resultRelInfo;
1450         Relation        resultRelationDesc;
1451         HTSU_Result result;
1452         ItemPointerData update_ctid;
1453         TransactionId update_xmax;
1454
1455         /*
1456          * get information on the (current) result relation
1457          */
1458         resultRelInfo = estate->es_result_relation_info;
1459         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1460
1461         /* BEFORE ROW DELETE Triggers */
1462         if (resultRelInfo->ri_TrigDesc &&
1463                 resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_DELETE] > 0)
1464         {
1465                 bool            dodelete;
1466
1467                 dodelete = ExecBRDeleteTriggers(estate, resultRelInfo, tupleid,
1468                                                                                 estate->es_snapshot->curcid);
1469
1470                 if (!dodelete)                  /* "do nothing" */
1471                         return;
1472         }
1473
1474         /*
1475          * delete the tuple
1476          *
1477          * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check that
1478          * the row to be deleted is visible to that snapshot, and throw a can't-
1479          * serialize error if not.      This is a special-case behavior needed for
1480          * referential integrity updates in serializable transactions.
1481          */
1482 ldelete:;
1483         result = heap_delete(resultRelationDesc, tupleid,
1484                                                  &update_ctid, &update_xmax,
1485                                                  estate->es_snapshot->curcid,
1486                                                  estate->es_crosscheck_snapshot,
1487                                                  true /* wait for commit */ );
1488         switch (result)
1489         {
1490                 case HeapTupleSelfUpdated:
1491                         /* already deleted by self; nothing to do */
1492                         return;
1493
1494                 case HeapTupleMayBeUpdated:
1495                         break;
1496
1497                 case HeapTupleUpdated:
1498                         if (IsXactIsoLevelSerializable)
1499                                 ereport(ERROR,
1500                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1501                                                  errmsg("could not serialize access due to concurrent update")));
1502                         else if (!ItemPointerEquals(tupleid, &update_ctid))
1503                         {
1504                                 TupleTableSlot *epqslot;
1505
1506                                 epqslot = EvalPlanQual(estate,
1507                                                                            resultRelInfo->ri_RangeTableIndex,
1508                                                                            &update_ctid,
1509                                                                            update_xmax,
1510                                                                            estate->es_snapshot->curcid);
1511                                 if (!TupIsNull(epqslot))
1512                                 {
1513                                         *tupleid = update_ctid;
1514                                         goto ldelete;
1515                                 }
1516                         }
1517                         /* tuple already deleted; nothing to do */
1518                         return;
1519
1520                 default:
1521                         elog(ERROR, "unrecognized heap_delete status: %u", result);
1522                         return;
1523         }
1524
1525         IncrDeleted();
1526         (estate->es_processed)++;
1527
1528         /*
1529          * Note: Normally one would think that we have to delete index tuples
1530          * associated with the heap tuple now...
1531          *
1532          * ... but in POSTGRES, we have no need to do this because VACUUM will
1533          * take care of it later.  We can't delete index tuples immediately
1534          * anyway, since the tuple is still visible to other transactions.
1535          */
1536
1537         /* AFTER ROW DELETE Triggers */
1538         ExecARDeleteTriggers(estate, resultRelInfo, tupleid);
1539
1540         /* Process RETURNING if present */
1541         if (resultRelInfo->ri_projectReturning)
1542         {
1543                 /*
1544                  * We have to put the target tuple into a slot, which means first we
1545                  * gotta fetch it.      We can use the trigger tuple slot.
1546                  */
1547                 TupleTableSlot *slot = estate->es_trig_tuple_slot;
1548                 HeapTupleData deltuple;
1549                 Buffer          delbuffer;
1550
1551                 deltuple.t_self = *tupleid;
1552                 if (!heap_fetch(resultRelationDesc, SnapshotAny,
1553                                                 &deltuple, &delbuffer, false, NULL))
1554                         elog(ERROR, "failed to fetch deleted tuple for DELETE RETURNING");
1555
1556                 if (slot->tts_tupleDescriptor != RelationGetDescr(resultRelationDesc))
1557                         ExecSetSlotDescriptor(slot, RelationGetDescr(resultRelationDesc));
1558                 ExecStoreTuple(&deltuple, slot, InvalidBuffer, false);
1559
1560                 ExecProcessReturning(resultRelInfo->ri_projectReturning,
1561                                                          slot, planSlot, dest);
1562
1563                 ExecClearTuple(slot);
1564                 ReleaseBuffer(delbuffer);
1565         }
1566 }
1567
1568 /* ----------------------------------------------------------------
1569  *              ExecUpdate
1570  *
1571  *              note: we can't run UPDATE queries with transactions
1572  *              off because UPDATEs are actually INSERTs and our
1573  *              scan will mistakenly loop forever, updating the tuple
1574  *              it just inserted..      This should be fixed but until it
1575  *              is, we don't want to get stuck in an infinite loop
1576  *              which corrupts your database..
1577  * ----------------------------------------------------------------
1578  */
1579 static void
1580 ExecUpdate(TupleTableSlot *slot,
1581                    ItemPointer tupleid,
1582                    TupleTableSlot *planSlot,
1583                    DestReceiver *dest,
1584                    EState *estate)
1585 {
1586         HeapTuple       tuple;
1587         ResultRelInfo *resultRelInfo;
1588         Relation        resultRelationDesc;
1589         HTSU_Result result;
1590         ItemPointerData update_ctid;
1591         TransactionId update_xmax;
1592
1593         /*
1594          * abort the operation if not running transactions
1595          */
1596         if (IsBootstrapProcessingMode())
1597                 elog(ERROR, "cannot UPDATE during bootstrap");
1598
1599         /*
1600          * get the heap tuple out of the tuple table slot, making sure we have a
1601          * writable copy
1602          */
1603         tuple = ExecMaterializeSlot(slot);
1604
1605         /*
1606          * get information on the (current) result relation
1607          */
1608         resultRelInfo = estate->es_result_relation_info;
1609         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1610
1611         /* BEFORE ROW UPDATE Triggers */
1612         if (resultRelInfo->ri_TrigDesc &&
1613                 resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_UPDATE] > 0)
1614         {
1615                 HeapTuple       newtuple;
1616
1617                 newtuple = ExecBRUpdateTriggers(estate, resultRelInfo,
1618                                                                                 tupleid, tuple,
1619                                                                                 estate->es_snapshot->curcid);
1620
1621                 if (newtuple == NULL)   /* "do nothing" */
1622                         return;
1623
1624                 if (newtuple != tuple)  /* modified by Trigger(s) */
1625                 {
1626                         /*
1627                          * Put the modified tuple into a slot for convenience of routines
1628                          * below.  We assume the tuple was allocated in per-tuple memory
1629                          * context, and therefore will go away by itself. The tuple table
1630                          * slot should not try to clear it.
1631                          */
1632                         TupleTableSlot *newslot = estate->es_trig_tuple_slot;
1633
1634                         if (newslot->tts_tupleDescriptor != slot->tts_tupleDescriptor)
1635                                 ExecSetSlotDescriptor(newslot, slot->tts_tupleDescriptor);
1636                         ExecStoreTuple(newtuple, newslot, InvalidBuffer, false);
1637                         slot = newslot;
1638                         tuple = newtuple;
1639                 }
1640         }
1641
1642         /*
1643          * Check the constraints of the tuple
1644          *
1645          * If we generate a new candidate tuple after EvalPlanQual testing, we
1646          * must loop back here and recheck constraints.  (We don't need to redo
1647          * triggers, however.  If there are any BEFORE triggers then trigger.c
1648          * will have done heap_lock_tuple to lock the correct tuple, so there's no
1649          * need to do them again.)
1650          */
1651 lreplace:;
1652         if (resultRelationDesc->rd_att->constr)
1653                 ExecConstraints(resultRelInfo, slot, estate);
1654
1655         /*
1656          * replace the heap tuple
1657          *
1658          * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check that
1659          * the row to be updated is visible to that snapshot, and throw a can't-
1660          * serialize error if not.      This is a special-case behavior needed for
1661          * referential integrity updates in serializable transactions.
1662          */
1663         result = heap_update(resultRelationDesc, tupleid, tuple,
1664                                                  &update_ctid, &update_xmax,
1665                                                  estate->es_snapshot->curcid,
1666                                                  estate->es_crosscheck_snapshot,
1667                                                  true /* wait for commit */ );
1668         switch (result)
1669         {
1670                 case HeapTupleSelfUpdated:
1671                         /* already deleted by self; nothing to do */
1672                         return;
1673
1674                 case HeapTupleMayBeUpdated:
1675                         break;
1676
1677                 case HeapTupleUpdated:
1678                         if (IsXactIsoLevelSerializable)
1679                                 ereport(ERROR,
1680                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1681                                                  errmsg("could not serialize access due to concurrent update")));
1682                         else if (!ItemPointerEquals(tupleid, &update_ctid))
1683                         {
1684                                 TupleTableSlot *epqslot;
1685
1686                                 epqslot = EvalPlanQual(estate,
1687                                                                            resultRelInfo->ri_RangeTableIndex,
1688                                                                            &update_ctid,
1689                                                                            update_xmax,
1690                                                                            estate->es_snapshot->curcid);
1691                                 if (!TupIsNull(epqslot))
1692                                 {
1693                                         *tupleid = update_ctid;
1694                                         slot = ExecFilterJunk(estate->es_junkFilter, epqslot);
1695                                         tuple = ExecMaterializeSlot(slot);
1696                                         goto lreplace;
1697                                 }
1698                         }
1699                         /* tuple already deleted; nothing to do */
1700                         return;
1701
1702                 default:
1703                         elog(ERROR, "unrecognized heap_update status: %u", result);
1704                         return;
1705         }
1706
1707         IncrReplaced();
1708         (estate->es_processed)++;
1709
1710         /*
1711          * Note: instead of having to update the old index tuples associated with
1712          * the heap tuple, all we do is form and insert new index tuples. This is
1713          * because UPDATEs are actually DELETEs and INSERTs, and index tuple
1714          * deletion is done later by VACUUM (see notes in ExecDelete).  All we do
1715          * here is insert new index tuples.  -cim 9/27/89
1716          */
1717
1718         /*
1719          * insert index entries for tuple
1720          *
1721          * Note: heap_update returns the tid (location) of the new tuple in the
1722          * t_self field.
1723          */
1724         if (resultRelInfo->ri_NumIndices > 0)
1725                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1726
1727         /* AFTER ROW UPDATE Triggers */
1728         ExecARUpdateTriggers(estate, resultRelInfo, tupleid, tuple);
1729
1730         /* Process RETURNING if present */
1731         if (resultRelInfo->ri_projectReturning)
1732                 ExecProcessReturning(resultRelInfo->ri_projectReturning,
1733                                                          slot, planSlot, dest);
1734 }
1735
1736 /*
1737  * ExecRelCheck --- check that tuple meets constraints for result relation
1738  */
1739 static const char *
1740 ExecRelCheck(ResultRelInfo *resultRelInfo,
1741                          TupleTableSlot *slot, EState *estate)
1742 {
1743         Relation        rel = resultRelInfo->ri_RelationDesc;
1744         int                     ncheck = rel->rd_att->constr->num_check;
1745         ConstrCheck *check = rel->rd_att->constr->check;
1746         ExprContext *econtext;
1747         MemoryContext oldContext;
1748         List       *qual;
1749         int                     i;
1750
1751         /*
1752          * If first time through for this result relation, build expression
1753          * nodetrees for rel's constraint expressions.  Keep them in the per-query
1754          * memory context so they'll survive throughout the query.
1755          */
1756         if (resultRelInfo->ri_ConstraintExprs == NULL)
1757         {
1758                 oldContext = MemoryContextSwitchTo(estate->es_query_cxt);
1759                 resultRelInfo->ri_ConstraintExprs =
1760                         (List **) palloc(ncheck * sizeof(List *));
1761                 for (i = 0; i < ncheck; i++)
1762                 {
1763                         /* ExecQual wants implicit-AND form */
1764                         qual = make_ands_implicit(stringToNode(check[i].ccbin));
1765                         resultRelInfo->ri_ConstraintExprs[i] = (List *)
1766                                 ExecPrepareExpr((Expr *) qual, estate);
1767                 }
1768                 MemoryContextSwitchTo(oldContext);
1769         }
1770
1771         /*
1772          * We will use the EState's per-tuple context for evaluating constraint
1773          * expressions (creating it if it's not already there).
1774          */
1775         econtext = GetPerTupleExprContext(estate);
1776
1777         /* Arrange for econtext's scan tuple to be the tuple under test */
1778         econtext->ecxt_scantuple = slot;
1779
1780         /* And evaluate the constraints */
1781         for (i = 0; i < ncheck; i++)
1782         {
1783                 qual = resultRelInfo->ri_ConstraintExprs[i];
1784
1785                 /*
1786                  * NOTE: SQL92 specifies that a NULL result from a constraint
1787                  * expression is not to be treated as a failure.  Therefore, tell
1788                  * ExecQual to return TRUE for NULL.
1789                  */
1790                 if (!ExecQual(qual, econtext, true))
1791                         return check[i].ccname;
1792         }
1793
1794         /* NULL result means no error */
1795         return NULL;
1796 }
1797
1798 void
1799 ExecConstraints(ResultRelInfo *resultRelInfo,
1800                                 TupleTableSlot *slot, EState *estate)
1801 {
1802         Relation        rel = resultRelInfo->ri_RelationDesc;
1803         TupleConstr *constr = rel->rd_att->constr;
1804
1805         Assert(constr);
1806
1807         if (constr->has_not_null)
1808         {
1809                 int                     natts = rel->rd_att->natts;
1810                 int                     attrChk;
1811
1812                 for (attrChk = 1; attrChk <= natts; attrChk++)
1813                 {
1814                         if (rel->rd_att->attrs[attrChk - 1]->attnotnull &&
1815                                 slot_attisnull(slot, attrChk))
1816                                 ereport(ERROR,
1817                                                 (errcode(ERRCODE_NOT_NULL_VIOLATION),
1818                                                  errmsg("null value in column \"%s\" violates not-null constraint",
1819                                                 NameStr(rel->rd_att->attrs[attrChk - 1]->attname))));
1820                 }
1821         }
1822
1823         if (constr->num_check > 0)
1824         {
1825                 const char *failed;
1826
1827                 if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
1828                         ereport(ERROR,
1829                                         (errcode(ERRCODE_CHECK_VIOLATION),
1830                                          errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
1831                                                         RelationGetRelationName(rel), failed)));
1832         }
1833 }
1834
1835 /*
1836  * ExecProcessReturning --- evaluate a RETURNING list and send to dest
1837  *
1838  * projectReturning: RETURNING projection info for current result rel
1839  * tupleSlot: slot holding tuple actually inserted/updated/deleted
1840  * planSlot: slot holding tuple returned by top plan node
1841  * dest: where to send the output
1842  */
1843 static void
1844 ExecProcessReturning(ProjectionInfo *projectReturning,
1845                                          TupleTableSlot *tupleSlot,
1846                                          TupleTableSlot *planSlot,
1847                                          DestReceiver *dest)
1848 {
1849         ExprContext *econtext = projectReturning->pi_exprContext;
1850         TupleTableSlot *retSlot;
1851
1852         /*
1853          * Reset per-tuple memory context to free any expression evaluation
1854          * storage allocated in the previous cycle.
1855          */
1856         ResetExprContext(econtext);
1857
1858         /* Make tuple and any needed join variables available to ExecProject */
1859         econtext->ecxt_scantuple = tupleSlot;
1860         econtext->ecxt_outertuple = planSlot;
1861
1862         /* Compute the RETURNING expressions */
1863         retSlot = ExecProject(projectReturning, NULL);
1864
1865         /* Send to dest */
1866         (*dest->receiveSlot) (retSlot, dest);
1867
1868         ExecClearTuple(retSlot);
1869 }
1870
1871 /*
1872  * Check a modified tuple to see if we want to process its updated version
1873  * under READ COMMITTED rules.
1874  *
1875  * See backend/executor/README for some info about how this works.
1876  *
1877  *      estate - executor state data
1878  *      rti - rangetable index of table containing tuple
1879  *      *tid - t_ctid from the outdated tuple (ie, next updated version)
1880  *      priorXmax - t_xmax from the outdated tuple
1881  *      curCid - command ID of current command of my transaction
1882  *
1883  * *tid is also an output parameter: it's modified to hold the TID of the
1884  * latest version of the tuple (note this may be changed even on failure)
1885  *
1886  * Returns a slot containing the new candidate update/delete tuple, or
1887  * NULL if we determine we shouldn't process the row.
1888  */
1889 TupleTableSlot *
1890 EvalPlanQual(EState *estate, Index rti,
1891                          ItemPointer tid, TransactionId priorXmax, CommandId curCid)
1892 {
1893         evalPlanQual *epq;
1894         EState     *epqstate;
1895         Relation        relation;
1896         HeapTupleData tuple;
1897         HeapTuple       copyTuple = NULL;
1898         bool            endNode;
1899
1900         Assert(rti != 0);
1901
1902         /*
1903          * find relation containing target tuple
1904          */
1905         if (estate->es_result_relation_info != NULL &&
1906                 estate->es_result_relation_info->ri_RangeTableIndex == rti)
1907                 relation = estate->es_result_relation_info->ri_RelationDesc;
1908         else
1909         {
1910                 ListCell   *l;
1911
1912                 relation = NULL;
1913                 foreach(l, estate->es_rowMarks)
1914                 {
1915                         if (((ExecRowMark *) lfirst(l))->rti == rti)
1916                         {
1917                                 relation = ((ExecRowMark *) lfirst(l))->relation;
1918                                 break;
1919                         }
1920                 }
1921                 if (relation == NULL)
1922                         elog(ERROR, "could not find RowMark for RT index %u", rti);
1923         }
1924
1925         /*
1926          * fetch tid tuple
1927          *
1928          * Loop here to deal with updated or busy tuples
1929          */
1930         tuple.t_self = *tid;
1931         for (;;)
1932         {
1933                 Buffer          buffer;
1934
1935                 if (heap_fetch(relation, SnapshotDirty, &tuple, &buffer, true, NULL))
1936                 {
1937                         /*
1938                          * If xmin isn't what we're expecting, the slot must have been
1939                          * recycled and reused for an unrelated tuple.  This implies that
1940                          * the latest version of the row was deleted, so we need do
1941                          * nothing.  (Should be safe to examine xmin without getting
1942                          * buffer's content lock, since xmin never changes in an existing
1943                          * tuple.)
1944                          */
1945                         if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
1946                                                                          priorXmax))
1947                         {
1948                                 ReleaseBuffer(buffer);
1949                                 return NULL;
1950                         }
1951
1952                         /* otherwise xmin should not be dirty... */
1953                         if (TransactionIdIsValid(SnapshotDirty->xmin))
1954                                 elog(ERROR, "t_xmin is uncommitted in tuple to be updated");
1955
1956                         /*
1957                          * If tuple is being updated by other transaction then we have to
1958                          * wait for its commit/abort.
1959                          */
1960                         if (TransactionIdIsValid(SnapshotDirty->xmax))
1961                         {
1962                                 ReleaseBuffer(buffer);
1963                                 XactLockTableWait(SnapshotDirty->xmax);
1964                                 continue;               /* loop back to repeat heap_fetch */
1965                         }
1966
1967                         /*
1968                          * If tuple was inserted by our own transaction, we have to check
1969                          * cmin against curCid: cmin >= curCid means our command cannot
1970                          * see the tuple, so we should ignore it.  Without this we are
1971                          * open to the "Halloween problem" of indefinitely re-updating the
1972                          * same tuple.  (We need not check cmax because
1973                          * HeapTupleSatisfiesDirty will consider a tuple deleted by our
1974                          * transaction dead, regardless of cmax.)  We just checked that
1975                          * priorXmax == xmin, so we can test that variable instead of
1976                          * doing HeapTupleHeaderGetXmin again.
1977                          */
1978                         if (TransactionIdIsCurrentTransactionId(priorXmax) &&
1979                                 HeapTupleHeaderGetCmin(tuple.t_data) >= curCid)
1980                         {
1981                                 ReleaseBuffer(buffer);
1982                                 return NULL;
1983                         }
1984
1985                         /*
1986                          * We got tuple - now copy it for use by recheck query.
1987                          */
1988                         copyTuple = heap_copytuple(&tuple);
1989                         ReleaseBuffer(buffer);
1990                         break;
1991                 }
1992
1993                 /*
1994                  * If the referenced slot was actually empty, the latest version of
1995                  * the row must have been deleted, so we need do nothing.
1996                  */
1997                 if (tuple.t_data == NULL)
1998                 {
1999                         ReleaseBuffer(buffer);
2000                         return NULL;
2001                 }
2002
2003                 /*
2004                  * As above, if xmin isn't what we're expecting, do nothing.
2005                  */
2006                 if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
2007                                                                  priorXmax))
2008                 {
2009                         ReleaseBuffer(buffer);
2010                         return NULL;
2011                 }
2012
2013                 /*
2014                  * If we get here, the tuple was found but failed SnapshotDirty.
2015                  * Assuming the xmin is either a committed xact or our own xact (as it
2016                  * certainly should be if we're trying to modify the tuple), this must
2017                  * mean that the row was updated or deleted by either a committed xact
2018                  * or our own xact.  If it was deleted, we can ignore it; if it was
2019                  * updated then chain up to the next version and repeat the whole
2020                  * test.
2021                  *
2022                  * As above, it should be safe to examine xmax and t_ctid without the
2023                  * buffer content lock, because they can't be changing.
2024                  */
2025                 if (ItemPointerEquals(&tuple.t_self, &tuple.t_data->t_ctid))
2026                 {
2027                         /* deleted, so forget about it */
2028                         ReleaseBuffer(buffer);
2029                         return NULL;
2030                 }
2031
2032                 /* updated, so look at the updated row */
2033                 tuple.t_self = tuple.t_data->t_ctid;
2034                 /* updated row should have xmin matching this xmax */
2035                 priorXmax = HeapTupleHeaderGetXmax(tuple.t_data);
2036                 ReleaseBuffer(buffer);
2037                 /* loop back to fetch next in chain */
2038         }
2039
2040         /*
2041          * For UPDATE/DELETE we have to return tid of actual row we're executing
2042          * PQ for.
2043          */
2044         *tid = tuple.t_self;
2045
2046         /*
2047          * Need to run a recheck subquery.      Find or create a PQ stack entry.
2048          */
2049         epq = estate->es_evalPlanQual;
2050         endNode = true;
2051
2052         if (epq != NULL && epq->rti == 0)
2053         {
2054                 /* Top PQ stack entry is idle, so re-use it */
2055                 Assert(!(estate->es_useEvalPlan) && epq->next == NULL);
2056                 epq->rti = rti;
2057                 endNode = false;
2058         }
2059
2060         /*
2061          * If this is request for another RTE - Ra, - then we have to check wasn't
2062          * PlanQual requested for Ra already and if so then Ra' row was updated
2063          * again and we have to re-start old execution for Ra and forget all what
2064          * we done after Ra was suspended. Cool? -:))
2065          */
2066         if (epq != NULL && epq->rti != rti &&
2067                 epq->estate->es_evTuple[rti - 1] != NULL)
2068         {
2069                 do
2070                 {
2071                         evalPlanQual *oldepq;
2072
2073                         /* stop execution */
2074                         EvalPlanQualStop(epq);
2075                         /* pop previous PlanQual from the stack */
2076                         oldepq = epq->next;
2077                         Assert(oldepq && oldepq->rti != 0);
2078                         /* push current PQ to freePQ stack */
2079                         oldepq->free = epq;
2080                         epq = oldepq;
2081                         estate->es_evalPlanQual = epq;
2082                 } while (epq->rti != rti);
2083         }
2084
2085         /*
2086          * If we are requested for another RTE then we have to suspend execution
2087          * of current PlanQual and start execution for new one.
2088          */
2089         if (epq == NULL || epq->rti != rti)
2090         {
2091                 /* try to reuse plan used previously */
2092                 evalPlanQual *newepq = (epq != NULL) ? epq->free : NULL;
2093
2094                 if (newepq == NULL)             /* first call or freePQ stack is empty */
2095                 {
2096                         newepq = (evalPlanQual *) palloc0(sizeof(evalPlanQual));
2097                         newepq->free = NULL;
2098                         newepq->estate = NULL;
2099                         newepq->planstate = NULL;
2100                 }
2101                 else
2102                 {
2103                         /* recycle previously used PlanQual */
2104                         Assert(newepq->estate == NULL);
2105                         epq->free = NULL;
2106                 }
2107                 /* push current PQ to the stack */
2108                 newepq->next = epq;
2109                 epq = newepq;
2110                 estate->es_evalPlanQual = epq;
2111                 epq->rti = rti;
2112                 endNode = false;
2113         }
2114
2115         Assert(epq->rti == rti);
2116
2117         /*
2118          * Ok - we're requested for the same RTE.  Unfortunately we still have to
2119          * end and restart execution of the plan, because ExecReScan wouldn't
2120          * ensure that upper plan nodes would reset themselves.  We could make
2121          * that work if insertion of the target tuple were integrated with the
2122          * Param mechanism somehow, so that the upper plan nodes know that their
2123          * children's outputs have changed.
2124          *
2125          * Note that the stack of free evalPlanQual nodes is quite useless at the
2126          * moment, since it only saves us from pallocing/releasing the
2127          * evalPlanQual nodes themselves.  But it will be useful once we implement
2128          * ReScan instead of end/restart for re-using PlanQual nodes.
2129          */
2130         if (endNode)
2131         {
2132                 /* stop execution */
2133                 EvalPlanQualStop(epq);
2134         }
2135
2136         /*
2137          * Initialize new recheck query.
2138          *
2139          * Note: if we were re-using PlanQual plans via ExecReScan, we'd need to
2140          * instead copy down changeable state from the top plan (including
2141          * es_result_relation_info, es_junkFilter) and reset locally changeable
2142          * state in the epq (including es_param_exec_vals, es_evTupleNull).
2143          */
2144         EvalPlanQualStart(epq, estate, epq->next);
2145
2146         /*
2147          * free old RTE' tuple, if any, and store target tuple where relation's
2148          * scan node will see it
2149          */
2150         epqstate = epq->estate;
2151         if (epqstate->es_evTuple[rti - 1] != NULL)
2152                 heap_freetuple(epqstate->es_evTuple[rti - 1]);
2153         epqstate->es_evTuple[rti - 1] = copyTuple;
2154
2155         return EvalPlanQualNext(estate);
2156 }
2157
2158 static TupleTableSlot *
2159 EvalPlanQualNext(EState *estate)
2160 {
2161         evalPlanQual *epq = estate->es_evalPlanQual;
2162         MemoryContext oldcontext;
2163         TupleTableSlot *slot;
2164
2165         Assert(epq->rti != 0);
2166
2167 lpqnext:;
2168         oldcontext = MemoryContextSwitchTo(epq->estate->es_query_cxt);
2169         slot = ExecProcNode(epq->planstate);
2170         MemoryContextSwitchTo(oldcontext);
2171
2172         /*
2173          * No more tuples for this PQ. Continue previous one.
2174          */
2175         if (TupIsNull(slot))
2176         {
2177                 evalPlanQual *oldepq;
2178
2179                 /* stop execution */
2180                 EvalPlanQualStop(epq);
2181                 /* pop old PQ from the stack */
2182                 oldepq = epq->next;
2183                 if (oldepq == NULL)
2184                 {
2185                         /* this is the first (oldest) PQ - mark as free */
2186                         epq->rti = 0;
2187                         estate->es_useEvalPlan = false;
2188                         /* and continue Query execution */
2189                         return NULL;
2190                 }
2191                 Assert(oldepq->rti != 0);
2192                 /* push current PQ to freePQ stack */
2193                 oldepq->free = epq;
2194                 epq = oldepq;
2195                 estate->es_evalPlanQual = epq;
2196                 goto lpqnext;
2197         }
2198
2199         return slot;
2200 }
2201
2202 static void
2203 EndEvalPlanQual(EState *estate)
2204 {
2205         evalPlanQual *epq = estate->es_evalPlanQual;
2206
2207         if (epq->rti == 0)                      /* plans already shutdowned */
2208         {
2209                 Assert(epq->next == NULL);
2210                 return;
2211         }
2212
2213         for (;;)
2214         {
2215                 evalPlanQual *oldepq;
2216
2217                 /* stop execution */
2218                 EvalPlanQualStop(epq);
2219                 /* pop old PQ from the stack */
2220                 oldepq = epq->next;
2221                 if (oldepq == NULL)
2222                 {
2223                         /* this is the first (oldest) PQ - mark as free */
2224                         epq->rti = 0;
2225                         estate->es_useEvalPlan = false;
2226                         break;
2227                 }
2228                 Assert(oldepq->rti != 0);
2229                 /* push current PQ to freePQ stack */
2230                 oldepq->free = epq;
2231                 epq = oldepq;
2232                 estate->es_evalPlanQual = epq;
2233         }
2234 }
2235
2236 /*
2237  * Start execution of one level of PlanQual.
2238  *
2239  * This is a cut-down version of ExecutorStart(): we copy some state from
2240  * the top-level estate rather than initializing it fresh.
2241  */
2242 static void
2243 EvalPlanQualStart(evalPlanQual *epq, EState *estate, evalPlanQual *priorepq)
2244 {
2245         EState     *epqstate;
2246         int                     rtsize;
2247         MemoryContext oldcontext;
2248
2249         rtsize = list_length(estate->es_range_table);
2250
2251         /*
2252          * It's tempting to think about using CreateSubExecutorState here, but
2253          * at present we can't because of memory leakage concerns ...
2254          */
2255         epq->estate = epqstate = CreateExecutorState();
2256
2257         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2258
2259         /*
2260          * The epqstates share the top query's copy of unchanging state such as
2261          * the snapshot, rangetable, result-rel info, and external Param info.
2262          * They need their own copies of local state, including a tuple table,
2263          * es_param_exec_vals, etc.
2264          */
2265         epqstate->es_direction = ForwardScanDirection;
2266         epqstate->es_snapshot = estate->es_snapshot;
2267         epqstate->es_crosscheck_snapshot = estate->es_crosscheck_snapshot;
2268         epqstate->es_range_table = estate->es_range_table;
2269         epqstate->es_result_relations = estate->es_result_relations;
2270         epqstate->es_num_result_relations = estate->es_num_result_relations;
2271         epqstate->es_result_relation_info = estate->es_result_relation_info;
2272         epqstate->es_junkFilter = estate->es_junkFilter;
2273         epqstate->es_into_relation_descriptor = estate->es_into_relation_descriptor;
2274         epqstate->es_into_relation_use_wal = estate->es_into_relation_use_wal;
2275         epqstate->es_param_list_info = estate->es_param_list_info;
2276         if (estate->es_topPlan->nParamExec > 0)
2277                 epqstate->es_param_exec_vals = (ParamExecData *)
2278                         palloc0(estate->es_topPlan->nParamExec * sizeof(ParamExecData));
2279         epqstate->es_rowMarks = estate->es_rowMarks;
2280         epqstate->es_instrument = estate->es_instrument;
2281         epqstate->es_select_into = estate->es_select_into;
2282         epqstate->es_into_oids = estate->es_into_oids;
2283         epqstate->es_topPlan = estate->es_topPlan;
2284
2285         /*
2286          * Each epqstate must have its own es_evTupleNull state, but all the stack
2287          * entries share es_evTuple state.      This allows sub-rechecks to inherit
2288          * the value being examined by an outer recheck.
2289          */
2290         epqstate->es_evTupleNull = (bool *) palloc0(rtsize * sizeof(bool));
2291         if (priorepq == NULL)
2292                 /* first PQ stack entry */
2293                 epqstate->es_evTuple = (HeapTuple *)
2294                         palloc0(rtsize * sizeof(HeapTuple));
2295         else
2296                 /* later stack entries share the same storage */
2297                 epqstate->es_evTuple = priorepq->estate->es_evTuple;
2298
2299         epqstate->es_tupleTable =
2300                 ExecCreateTupleTable(estate->es_tupleTable->size);
2301
2302         epq->planstate = ExecInitNode(estate->es_topPlan, epqstate, 0);
2303
2304         MemoryContextSwitchTo(oldcontext);
2305 }
2306
2307 /*
2308  * End execution of one level of PlanQual.
2309  *
2310  * This is a cut-down version of ExecutorEnd(); basically we want to do most
2311  * of the normal cleanup, but *not* close result relations (which we are
2312  * just sharing from the outer query).
2313  */
2314 static void
2315 EvalPlanQualStop(evalPlanQual *epq)
2316 {
2317         EState     *epqstate = epq->estate;
2318         MemoryContext oldcontext;
2319
2320         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2321
2322         ExecEndNode(epq->planstate);
2323
2324         ExecDropTupleTable(epqstate->es_tupleTable, true);
2325         epqstate->es_tupleTable = NULL;
2326
2327         if (epqstate->es_evTuple[epq->rti - 1] != NULL)
2328         {
2329                 heap_freetuple(epqstate->es_evTuple[epq->rti - 1]);
2330                 epqstate->es_evTuple[epq->rti - 1] = NULL;
2331         }
2332
2333         MemoryContextSwitchTo(oldcontext);
2334
2335         FreeExecutorState(epqstate);
2336
2337         epq->estate = NULL;
2338         epq->planstate = NULL;
2339 }
2340
2341
2342 /*
2343  * Support for SELECT INTO (a/k/a CREATE TABLE AS)
2344  *
2345  * We implement SELECT INTO by diverting SELECT's normal output with
2346  * a specialized DestReceiver type.
2347  *
2348  * TODO: remove some of the INTO-specific cruft from EState, and keep
2349  * it in the DestReceiver instead.
2350  */
2351
2352 typedef struct
2353 {
2354         DestReceiver pub;                       /* publicly-known function pointers */
2355         EState     *estate;                     /* EState we are working with */
2356 } DR_intorel;
2357
2358 /*
2359  * OpenIntoRel --- actually create the SELECT INTO target relation
2360  *
2361  * This also replaces QueryDesc->dest with the special DestReceiver for
2362  * SELECT INTO.  We assume that the correct result tuple type has already
2363  * been placed in queryDesc->tupDesc.
2364  */
2365 static void
2366 OpenIntoRel(QueryDesc *queryDesc)
2367 {
2368         Query      *parseTree = queryDesc->parsetree;
2369         EState     *estate = queryDesc->estate;
2370         Relation        intoRelationDesc;
2371         char       *intoName;
2372         Oid                     namespaceId;
2373         Oid                     tablespaceId;
2374         Datum           reloptions;
2375         AclResult       aclresult;
2376         Oid                     intoRelationId;
2377         TupleDesc       tupdesc;
2378         DR_intorel *myState;
2379
2380         /*
2381          * Check consistency of arguments
2382          */
2383         if (parseTree->intoOnCommit != ONCOMMIT_NOOP && !parseTree->into->istemp)
2384                 ereport(ERROR,
2385                                 (errcode(ERRCODE_INVALID_TABLE_DEFINITION),
2386                                  errmsg("ON COMMIT can only be used on temporary tables")));
2387
2388         /*
2389          * Find namespace to create in, check its permissions
2390          */
2391         intoName = parseTree->into->relname;
2392         namespaceId = RangeVarGetCreationNamespace(parseTree->into);
2393
2394         aclresult = pg_namespace_aclcheck(namespaceId, GetUserId(),
2395                                                                           ACL_CREATE);
2396         if (aclresult != ACLCHECK_OK)
2397                 aclcheck_error(aclresult, ACL_KIND_NAMESPACE,
2398                                            get_namespace_name(namespaceId));
2399
2400         /*
2401          * Select tablespace to use.  If not specified, use default_tablespace
2402          * (which may in turn default to database's default).
2403          */
2404         if (parseTree->intoTableSpaceName)
2405         {
2406                 tablespaceId = get_tablespace_oid(parseTree->intoTableSpaceName);
2407                 if (!OidIsValid(tablespaceId))
2408                         ereport(ERROR,
2409                                         (errcode(ERRCODE_UNDEFINED_OBJECT),
2410                                          errmsg("tablespace \"%s\" does not exist",
2411                                                         parseTree->intoTableSpaceName)));
2412         }
2413         else if (parseTree->into->istemp)
2414         {
2415                 tablespaceId = GetTempTablespace();
2416         }
2417         else
2418         {
2419                 tablespaceId = GetDefaultTablespace();
2420                 /* note InvalidOid is OK in this case */
2421         }
2422
2423         /* Check permissions except when using the database's default space */
2424         if (OidIsValid(tablespaceId))
2425         {
2426                 AclResult       aclresult;
2427
2428                 aclresult = pg_tablespace_aclcheck(tablespaceId, GetUserId(),
2429                                                                                    ACL_CREATE);
2430
2431                 if (aclresult != ACLCHECK_OK)
2432                         aclcheck_error(aclresult, ACL_KIND_TABLESPACE,
2433                                                    get_tablespace_name(tablespaceId));
2434         }
2435
2436         /* Parse and validate any reloptions */
2437         reloptions = transformRelOptions((Datum) 0,
2438                                                                          parseTree->intoOptions,
2439                                                                          true,
2440                                                                          false);
2441         (void) heap_reloptions(RELKIND_RELATION, reloptions, true);
2442
2443         /* have to copy the actual tupdesc to get rid of any constraints */
2444         tupdesc = CreateTupleDescCopy(queryDesc->tupDesc);
2445
2446         /* Now we can actually create the new relation */
2447         intoRelationId = heap_create_with_catalog(intoName,
2448                                                                                           namespaceId,
2449                                                                                           tablespaceId,
2450                                                                                           InvalidOid,
2451                                                                                           GetUserId(),
2452                                                                                           tupdesc,
2453                                                                                           RELKIND_RELATION,
2454                                                                                           false,
2455                                                                                           true,
2456                                                                                           0,
2457                                                                                           parseTree->intoOnCommit,
2458                                                                                           reloptions,
2459                                                                                           allowSystemTableMods);
2460
2461         FreeTupleDesc(tupdesc);
2462
2463         /*
2464          * Advance command counter so that the newly-created relation's catalog
2465          * tuples will be visible to heap_open.
2466          */
2467         CommandCounterIncrement();
2468
2469         /*
2470          * If necessary, create a TOAST table for the INTO relation. Note that
2471          * AlterTableCreateToastTable ends with CommandCounterIncrement(), so that
2472          * the TOAST table will be visible for insertion.
2473          */
2474         AlterTableCreateToastTable(intoRelationId);
2475
2476         /*
2477          * And open the constructed table for writing.
2478          */
2479         intoRelationDesc = heap_open(intoRelationId, AccessExclusiveLock);
2480
2481         /* use_wal off requires rd_targblock be initially invalid */
2482         Assert(intoRelationDesc->rd_targblock == InvalidBlockNumber);
2483
2484         /*
2485          * We can skip WAL-logging the insertions, unless PITR is in use.
2486          *
2487          * Note that for a non-temp INTO table, this is safe only because we know
2488          * that the catalog changes above will have been WAL-logged, and so
2489          * RecordTransactionCommit will think it needs to WAL-log the eventual
2490          * transaction commit.  Else the commit might be lost, even though all the
2491          * data is safely fsync'd ...
2492          */
2493         estate->es_into_relation_use_wal = XLogArchivingActive();
2494         estate->es_into_relation_descriptor = intoRelationDesc;
2495
2496         /*
2497          * Now replace the query's DestReceiver with one for SELECT INTO
2498          */
2499         queryDesc->dest = CreateDestReceiver(DestIntoRel, NULL);
2500         myState = (DR_intorel *) queryDesc->dest;
2501         Assert(myState->pub.mydest == DestIntoRel);
2502         myState->estate = estate;
2503 }
2504
2505 /*
2506  * CloseIntoRel --- clean up SELECT INTO at ExecutorEnd time
2507  */
2508 static void
2509 CloseIntoRel(QueryDesc *queryDesc)
2510 {
2511         EState     *estate = queryDesc->estate;
2512
2513         /* OpenIntoRel might never have gotten called */
2514         if (estate->es_into_relation_descriptor)
2515         {
2516                 /*
2517                  * If we skipped using WAL, and it's not a temp relation, we must
2518                  * force the relation down to disk before it's safe to commit the
2519                  * transaction.  This requires forcing out any dirty buffers and then
2520                  * doing a forced fsync.
2521                  */
2522                 if (!estate->es_into_relation_use_wal &&
2523                         !estate->es_into_relation_descriptor->rd_istemp)
2524                         heap_sync(estate->es_into_relation_descriptor);
2525
2526                 /* close rel, but keep lock until commit */
2527                 heap_close(estate->es_into_relation_descriptor, NoLock);
2528
2529                 estate->es_into_relation_descriptor = NULL;
2530         }
2531 }
2532
2533 /*
2534  * CreateIntoRelDestReceiver -- create a suitable DestReceiver object
2535  *
2536  * Since CreateDestReceiver doesn't accept the parameters we'd need,
2537  * we just leave the private fields empty here.  OpenIntoRel will
2538  * fill them in.
2539  */
2540 DestReceiver *
2541 CreateIntoRelDestReceiver(void)
2542 {
2543         DR_intorel *self = (DR_intorel *) palloc(sizeof(DR_intorel));
2544
2545         self->pub.receiveSlot = intorel_receive;
2546         self->pub.rStartup = intorel_startup;
2547         self->pub.rShutdown = intorel_shutdown;
2548         self->pub.rDestroy = intorel_destroy;
2549         self->pub.mydest = DestIntoRel;
2550
2551         self->estate = NULL;
2552
2553         return (DestReceiver *) self;
2554 }
2555
2556 /*
2557  * intorel_startup --- executor startup
2558  */
2559 static void
2560 intorel_startup(DestReceiver *self, int operation, TupleDesc typeinfo)
2561 {
2562         /* no-op */
2563 }
2564
2565 /*
2566  * intorel_receive --- receive one tuple
2567  */
2568 static void
2569 intorel_receive(TupleTableSlot *slot, DestReceiver *self)
2570 {
2571         DR_intorel *myState = (DR_intorel *) self;
2572         EState     *estate = myState->estate;
2573         HeapTuple       tuple;
2574
2575         tuple = ExecCopySlotTuple(slot);
2576
2577         heap_insert(estate->es_into_relation_descriptor,
2578                                 tuple,
2579                                 estate->es_snapshot->curcid,
2580                                 estate->es_into_relation_use_wal,
2581                                 false);                 /* never any point in using FSM */
2582
2583         /* We know this is a newly created relation, so there are no indexes */
2584
2585         heap_freetuple(tuple);
2586
2587         IncrAppended();
2588 }
2589
2590 /*
2591  * intorel_shutdown --- executor end
2592  */
2593 static void
2594 intorel_shutdown(DestReceiver *self)
2595 {
2596         /* no-op */
2597 }
2598
2599 /*
2600  * intorel_destroy --- release DestReceiver object
2601  */
2602 static void
2603 intorel_destroy(DestReceiver *self)
2604 {
2605         pfree(self);
2606 }