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  11  *      The old ExecutorMain() has been replaced by ExecutorStart(),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) 1996-2010, 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 1.354 2010/08/05 14:45:02 rhaas Exp $
  30  *
  31  *-------------------------------------------------------------------------
  32  */
  33 #include "postgres.h"
  34
  35 #include "access/reloptions.h"
  36 #include "access/sysattr.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 "miscadmin.h"
  47 #include "optimizer/clauses.h"
  48 #include "parser/parse_clause.h"
  49 #include "parser/parsetree.h"
  50 #include "storage/bufmgr.h"
  51 #include "storage/lmgr.h"
  52 #include "storage/smgr.h"
  53 #include "tcop/utility.h"
  54 #include "utils/acl.h"
  55 #include "utils/lsyscache.h"
  56 #include "utils/memutils.h"
  57 #include "utils/snapmgr.h"
  58 #include "utils/tqual.h"
  59
  60
  61 /* Hooks for plugins to get control in ExecutorStart/Run/End() */
  62 ExecutorStart_hook_type ExecutorStart_hook = NULL;
  63 ExecutorRun_hook_type ExecutorRun_hook = NULL;
  64 ExecutorEnd_hook_type ExecutorEnd_hook = NULL;
  65
  66 /* Hook for plugin to get control in ExecCheckRTPerms() */
  67 ExecutorCheckPerms_hook_type ExecutorCheckPerms_hook = NULL;
  68
  69 /* decls for local routines only used within this module */
  70 static void InitPlan(QueryDesc *queryDesc, int eflags);
  71 static void ExecEndPlan(PlanState *planstate, EState *estate);
  72 static void ExecutePlan(EState *estate, PlanState *planstate,
  73                         CmdType operation,
  74                         bool sendTuples,
  75                         long numberTuples,
  76                         ScanDirection direction,
  77                         DestReceiver *dest);
  78 static bool ExecCheckRTEPerms(RangeTblEntry *rte);
  79 static void ExecCheckXactReadOnly(PlannedStmt *plannedstmt);
  80 static void EvalPlanQualStart(EPQState *epqstate, EState *parentestate,
  81                                   Plan *planTree);
  82 static void OpenIntoRel(QueryDesc *queryDesc);
  83 static void CloseIntoRel(QueryDesc *queryDesc);
  84 static void intorel_startup(DestReceiver *self, int operation, TupleDesc typeinfo);
  85 static void intorel_receive(TupleTableSlot *slot, DestReceiver *self);
  86 static void intorel_shutdown(DestReceiver *self);
  87 static void intorel_destroy(DestReceiver *self);
  88
  89 /* end of local decls */
  90
  91
  92 /* ----------------------------------------------------------------
  93  *              ExecutorStart
  94  *
  95  *              This routine must be called at the beginning of any execution of any
  96  *              query plan
  97  *
  98  * Takes a QueryDesc previously created by CreateQueryDesc (it's not real
  99  * clear why we bother to separate the two functions, but...).  The tupDesc
 100  * field of the QueryDesc is filled in to describe the tuples that will be
 101  * returned, and the internal fields (estate and planstate) are set up.
 102  *
 103  * eflags contains flag bits as described in executor.h.
 104  *
 105  * NB: the CurrentMemoryContext when this is called will become the parent
 106  * of the per-query context used for this Executor invocation.
 107  *
 108  * We provide a function hook variable that lets loadable plugins
 109  * get control when ExecutorStart is called.  Such a plugin would
 110  * normally call standard_ExecutorStart().
 111  *
 112  * ----------------------------------------------------------------
 113  */
 114 void
 115 ExecutorStart(QueryDesc *queryDesc, int eflags)
 116 {
 117         if (ExecutorStart_hook)
 118                 (*ExecutorStart_hook) (queryDesc, eflags);
 119         else
 120                 standard_ExecutorStart(queryDesc, eflags);
 121 }
 122
 123 void
 124 standard_ExecutorStart(QueryDesc *queryDesc, int eflags)
 125 {
 126         EState     *estate;
 127         MemoryContext oldcontext;
 128
 129         /* sanity checks: queryDesc must not be started already */
 130         Assert(queryDesc != NULL);
 131         Assert(queryDesc->estate == NULL);
 132
 133         /*
 134          * If the transaction is read-only, we need to check if any writes are
 135          * planned to non-temporary tables.  EXPLAIN is considered read-only.
 136          */
 137         if (XactReadOnly && !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
 138                 ExecCheckXactReadOnly(queryDesc->plannedstmt);
 139
 140         /*
 141          * Build EState, switch into per-query memory context for startup.
 142          */
 143         estate = CreateExecutorState();
 144         queryDesc->estate = estate;
 145
 146         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 147
 148         /*
 149          * Fill in external parameters, if any, from queryDesc; and allocate
 150          * workspace for internal parameters
 151          */
 152         estate->es_param_list_info = queryDesc->params;
 153
 154         if (queryDesc->plannedstmt->nParamExec > 0)
 155                 estate->es_param_exec_vals = (ParamExecData *)
 156                         palloc0(queryDesc->plannedstmt->nParamExec * sizeof(ParamExecData));
 157
 158         /*
 159          * If non-read-only query, set the command ID to mark output tuples with
 160          */
 161         switch (queryDesc->operation)
 162         {
 163                 case CMD_SELECT:
 164                         /* SELECT INTO and SELECT FOR UPDATE/SHARE need to mark tuples */
 165                         if (queryDesc->plannedstmt->intoClause != NULL ||
 166                                 queryDesc->plannedstmt->rowMarks != NIL)
 167                                 estate->es_output_cid = GetCurrentCommandId(true);
 168                         break;
 169
 170                 case CMD_INSERT:
 171                 case CMD_DELETE:
 172                 case CMD_UPDATE:
 173                         estate->es_output_cid = GetCurrentCommandId(true);
 174                         break;
 175
 176                 default:
 177                         elog(ERROR, "unrecognized operation code: %d",
 178                                  (int) queryDesc->operation);
 179                         break;
 180         }
 181
 182         /*
 183          * Copy other important information into the EState
 184          */
 185         estate->es_snapshot = RegisterSnapshot(queryDesc->snapshot);
 186         estate->es_crosscheck_snapshot = RegisterSnapshot(queryDesc->crosscheck_snapshot);
 187         estate->es_instrument = queryDesc->instrument_options;
 188
 189         /*
 190          * Initialize the plan state tree
 191          */
 192         InitPlan(queryDesc, eflags);
 193
 194         MemoryContextSwitchTo(oldcontext);
 195 }
 196
 197 /* ----------------------------------------------------------------
 198  *              ExecutorRun
 199  *
 200  *              This is the main routine of the executor module. It accepts
 201  *              the query descriptor from the traffic cop and executes the
 202  *              query plan.
 203  *
 204  *              ExecutorStart must have been called already.
 205  *
 206  *              If direction is NoMovementScanDirection then nothing is done
 207  *              except to start up/shut down the destination.  Otherwise,
 208  *              we retrieve up to 'count' tuples in the specified direction.
 209  *
 210  *              Note: count = 0 is interpreted as no portal limit, i.e., run to
 211  *              completion.
 212  *
 213  *              There is no return value, but output tuples (if any) are sent to
 214  *              the destination receiver specified in the QueryDesc; and the number
 215  *              of tuples processed at the top level can be found in
 216  *              estate->es_processed.
 217  *
 218  *              We provide a function hook variable that lets loadable plugins
 219  *              get control when ExecutorRun is called.  Such a plugin would
 220  *              normally call standard_ExecutorRun().
 221  *
 222  * ----------------------------------------------------------------
 223  */
 224 void
 225 ExecutorRun(QueryDesc *queryDesc,
 226                         ScanDirection direction, long count)
 227 {
 228         if (ExecutorRun_hook)
 229                 (*ExecutorRun_hook) (queryDesc, direction, count);
 230         else
 231                 standard_ExecutorRun(queryDesc, direction, count);
 232 }
 233
 234 void
 235 standard_ExecutorRun(QueryDesc *queryDesc,
 236                                          ScanDirection direction, long count)
 237 {
 238         EState     *estate;
 239         CmdType         operation;
 240         DestReceiver *dest;
 241         bool            sendTuples;
 242         MemoryContext oldcontext;
 243
 244         /* sanity checks */
 245         Assert(queryDesc != NULL);
 246
 247         estate = queryDesc->estate;
 248
 249         Assert(estate != NULL);
 250
 251         /*
 252          * Switch into per-query memory context
 253          */
 254         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 255
 256         /* Allow instrumentation of ExecutorRun overall runtime */
 257         if (queryDesc->totaltime)
 258                 InstrStartNode(queryDesc->totaltime);
 259
 260         /*
 261          * extract information from the query descriptor and the query feature.
 262          */
 263         operation = queryDesc->operation;
 264         dest = queryDesc->dest;
 265
 266         /*
 267          * startup tuple receiver, if we will be emitting tuples
 268          */
 269         estate->es_processed = 0;
 270         estate->es_lastoid = InvalidOid;
 271
 272         sendTuples = (operation == CMD_SELECT ||
 273                                   queryDesc->plannedstmt->hasReturning);
 274
 275         if (sendTuples)
 276                 (*dest->rStartup) (dest, operation, queryDesc->tupDesc);
 277
 278         /*
 279          * run plan
 280          */
 281         if (!ScanDirectionIsNoMovement(direction))
 282                 ExecutePlan(estate,
 283                                         queryDesc->planstate,
 284                                         operation,
 285                                         sendTuples,
 286                                         count,
 287                                         direction,
 288                                         dest);
 289
 290         /*
 291          * shutdown tuple receiver, if we started it
 292          */
 293         if (sendTuples)
 294                 (*dest->rShutdown) (dest);
 295
 296         if (queryDesc->totaltime)
 297                 InstrStopNode(queryDesc->totaltime, estate->es_processed);
 298
 299         MemoryContextSwitchTo(oldcontext);
 300 }
 301
 302 /* ----------------------------------------------------------------
 303  *              ExecutorEnd
 304  *
 305  *              This routine must be called at the end of execution of any
 306  *              query plan
 307  *
 308  *              We provide a function hook variable that lets loadable plugins
 309  *              get control when ExecutorEnd is called.  Such a plugin would
 310  *              normally call standard_ExecutorEnd().
 311  *
 312  * ----------------------------------------------------------------
 313  */
 314 void
 315 ExecutorEnd(QueryDesc *queryDesc)
 316 {
 317         if (ExecutorEnd_hook)
 318                 (*ExecutorEnd_hook) (queryDesc);
 319         else
 320                 standard_ExecutorEnd(queryDesc);
 321 }
 322
 323 void
 324 standard_ExecutorEnd(QueryDesc *queryDesc)
 325 {
 326         EState     *estate;
 327         MemoryContext oldcontext;
 328
 329         /* sanity checks */
 330         Assert(queryDesc != NULL);
 331
 332         estate = queryDesc->estate;
 333
 334         Assert(estate != NULL);
 335
 336         /*
 337          * Switch into per-query memory context to run ExecEndPlan
 338          */
 339         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 340
 341         ExecEndPlan(queryDesc->planstate, estate);
 342
 343         /*
 344          * Close the SELECT INTO relation if any
 345          */
 346         if (estate->es_select_into)
 347                 CloseIntoRel(queryDesc);
 348
 349         /* do away with our snapshots */
 350         UnregisterSnapshot(estate->es_snapshot);
 351         UnregisterSnapshot(estate->es_crosscheck_snapshot);
 352
 353         /*
 354          * Must switch out of context before destroying it
 355          */
 356         MemoryContextSwitchTo(oldcontext);
 357
 358         /*
 359          * Release EState and per-query memory context.  This should release
 360          * everything the executor has allocated.
 361          */
 362         FreeExecutorState(estate);
 363
 364         /* Reset queryDesc fields that no longer point to anything */
 365         queryDesc->tupDesc = NULL;
 366         queryDesc->estate = NULL;
 367         queryDesc->planstate = NULL;
 368         queryDesc->totaltime = NULL;
 369 }
 370
 371 /* ----------------------------------------------------------------
 372  *              ExecutorRewind
 373  *
 374  *              This routine may be called on an open queryDesc to rewind it
 375  *              to the start.
 376  * ----------------------------------------------------------------
 377  */
 378 void
 379 ExecutorRewind(QueryDesc *queryDesc)
 380 {
 381         EState     *estate;
 382         MemoryContext oldcontext;
 383
 384         /* sanity checks */
 385         Assert(queryDesc != NULL);
 386
 387         estate = queryDesc->estate;
 388
 389         Assert(estate != NULL);
 390
 391         /* It's probably not sensible to rescan updating queries */
 392         Assert(queryDesc->operation == CMD_SELECT);
 393
 394         /*
 395          * Switch into per-query memory context
 396          */
 397         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 398
 399         /*
 400          * rescan plan
 401          */
 402         ExecReScan(queryDesc->planstate);
 403
 404         MemoryContextSwitchTo(oldcontext);
 405 }
 406
 407
 408 /*
 409  * ExecCheckRTPerms
 410  *              Check access permissions for all relations listed in a range table.
 411  *
 412  * Returns true if permissions are adequate.  Otherwise, throws an appropriate
 413  * error if ereport_on_violation is true, or simply returns false otherwise.
 414  */
 415 bool
 416 ExecCheckRTPerms(List *rangeTable, bool ereport_on_violation)
 417 {
 418         ListCell   *l;
 419         bool            result = true;
 420
 421         foreach(l, rangeTable)
 422         {
 423                 RangeTblEntry  *rte = (RangeTblEntry *) lfirst(l);
 424
 425                 result = ExecCheckRTEPerms(rte);
 426                 if (!result)
 427                 {
 428                         Assert(rte->rtekind == RTE_RELATION);
 429                         if (ereport_on_violation)
 430                                 aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
 431                                                            get_rel_name(rte->relid));
 432                         return false;
 433                 }
 434         }
 435
 436         if (ExecutorCheckPerms_hook)
 437                 result = (*ExecutorCheckPerms_hook)(rangeTable,
 438                                                                                         ereport_on_violation);
 439         return result;
 440 }
 441
 442 /*
 443  * ExecCheckRTEPerms
 444  *              Check access permissions for a single RTE.
 445  */
 446 static bool
 447 ExecCheckRTEPerms(RangeTblEntry *rte)
 448 {
 449         AclMode         requiredPerms;
 450         AclMode         relPerms;
 451         AclMode         remainingPerms;
 452         Oid                     relOid;
 453         Oid                     userid;
 454         Bitmapset  *tmpset;
 455         int                     col;
 456
 457         /*
 458          * Only plain-relation RTEs need to be checked here.  Function RTEs are
 459          * checked by init_fcache when the function is prepared for execution.
 460          * Join, subquery, and special RTEs need no checks.
 461          */
 462         if (rte->rtekind != RTE_RELATION)
 463                 return true;
 464
 465         /*
 466          * No work if requiredPerms is empty.
 467          */
 468         requiredPerms = rte->requiredPerms;
 469         if (requiredPerms == 0)
 470                 return true;
 471
 472         relOid = rte->relid;
 473
 474         /*
 475          * userid to check as: current user unless we have a setuid indication.
 476          *
 477          * Note: GetUserId() is presently fast enough that there's no harm in
 478          * calling it separately for each RTE.  If that stops being true, we could
 479          * call it once in ExecCheckRTPerms and pass the userid down from there.
 480          * But for now, no need for the extra clutter.
 481          */
 482         userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
 483
 484         /*
 485          * We must have *all* the requiredPerms bits, but some of the bits can be
 486          * satisfied from column-level rather than relation-level permissions.
 487          * First, remove any bits that are satisfied by relation permissions.
 488          */
 489         relPerms = pg_class_aclmask(relOid, userid, requiredPerms, ACLMASK_ALL);
 490         remainingPerms = requiredPerms & ~relPerms;
 491         if (remainingPerms != 0)
 492         {
 493                 /*
 494                  * If we lack any permissions that exist only as relation permissions,
 495                  * we can fail straight away.
 496                  */
 497                 if (remainingPerms & ~(ACL_SELECT | ACL_INSERT | ACL_UPDATE))
 498                         return false;
 499
 500                 /*
 501                  * Check to see if we have the needed privileges at column level.
 502                  *
 503                  * Note: failures just report a table-level error; it would be nicer
 504                  * to report a column-level error if we have some but not all of the
 505                  * column privileges.
 506                  */
 507                 if (remainingPerms & ACL_SELECT)
 508                 {
 509                         /*
 510                          * When the query doesn't explicitly reference any columns (for
 511                          * example, SELECT COUNT(*) FROM table), allow the query if we
 512                          * have SELECT on any column of the rel, as per SQL spec.
 513                          */
 514                         if (bms_is_empty(rte->selectedCols))
 515                         {
 516                                 if (pg_attribute_aclcheck_all(relOid, userid, ACL_SELECT,
 517                                                                                           ACLMASK_ANY) != ACLCHECK_OK)
 518                                         return false;
 519                         }
 520
 521                         tmpset = bms_copy(rte->selectedCols);
 522                         while ((col = bms_first_member(tmpset)) >= 0)
 523                         {
 524                                 /* remove the column number offset */
 525                                 col += FirstLowInvalidHeapAttributeNumber;
 526                                 if (col == InvalidAttrNumber)
 527                                 {
 528                                         /* Whole-row reference, must have priv on all cols */
 529                                         if (pg_attribute_aclcheck_all(relOid, userid, ACL_SELECT,
 530                                                                                                   ACLMASK_ALL) != ACLCHECK_OK)
 531                                                 return false;
 532                                 }
 533                                 else
 534                                 {
 535                                         if (pg_attribute_aclcheck(relOid, col, userid,
 536                                                                                           ACL_SELECT) != ACLCHECK_OK)
 537                                                 return false;
 538                                 }
 539                         }
 540                         bms_free(tmpset);
 541                 }
 542
 543                 /*
 544                  * Basically the same for the mod columns, with either INSERT or
 545                  * UPDATE privilege as specified by remainingPerms.
 546                  */
 547                 remainingPerms &= ~ACL_SELECT;
 548                 if (remainingPerms != 0)
 549                 {
 550                         /*
 551                          * When the query doesn't explicitly change any columns, allow the
 552                          * query if we have permission on any column of the rel.  This is
 553                          * to handle SELECT FOR UPDATE as well as possible corner cases in
 554                          * INSERT and UPDATE.
 555                          */
 556                         if (bms_is_empty(rte->modifiedCols))
 557                         {
 558                                 if (pg_attribute_aclcheck_all(relOid, userid, remainingPerms,
 559                                                                                           ACLMASK_ANY) != ACLCHECK_OK)
 560                                         return false;
 561                         }
 562
 563                         tmpset = bms_copy(rte->modifiedCols);
 564                         while ((col = bms_first_member(tmpset)) >= 0)
 565                         {
 566                                 /* remove the column number offset */
 567                                 col += FirstLowInvalidHeapAttributeNumber;
 568                                 if (col == InvalidAttrNumber)
 569                                 {
 570                                         /* whole-row reference can't happen here */
 571                                         elog(ERROR, "whole-row update is not implemented");
 572                                 }
 573                                 else
 574                                 {
 575                                         if (pg_attribute_aclcheck(relOid, col, userid,
 576                                                                                           remainingPerms) != ACLCHECK_OK)
 577                                                 return false;
 578                                 }
 579                         }
 580                         bms_free(tmpset);
 581                 }
 582         }
 583         return true;
 584 }
 585
 586 /*
 587  * Check that the query does not imply any writes to non-temp tables.
 588  *
 589  * Note: in a Hot Standby slave this would need to reject writes to temp
 590  * tables as well; but an HS slave can't have created any temp tables
 591  * in the first place, so no need to check that.
 592  */
 593 static void
 594 ExecCheckXactReadOnly(PlannedStmt *plannedstmt)
 595 {
 596         ListCell   *l;
 597
 598         /*
 599          * CREATE TABLE AS or SELECT INTO?
 600          *
 601          * XXX should we allow this if the destination is temp?  Considering that
 602          * it would still require catalog changes, probably not.
 603          */
 604         if (plannedstmt->intoClause != NULL)
 605                 PreventCommandIfReadOnly(CreateCommandTag((Node *) plannedstmt));
 606
 607         /* Fail if write permissions are requested on any non-temp table */
 608         foreach(l, plannedstmt->rtable)
 609         {
 610                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 611
 612                 if (rte->rtekind != RTE_RELATION)
 613                         continue;
 614
 615                 if ((rte->requiredPerms & (~ACL_SELECT)) == 0)
 616                         continue;
 617
 618                 if (isTempNamespace(get_rel_namespace(rte->relid)))
 619                         continue;
 620
 621                 PreventCommandIfReadOnly(CreateCommandTag((Node *) plannedstmt));
 622         }
 623 }
 624
 625
 626 /* ----------------------------------------------------------------
 627  *              InitPlan
 628  *
 629  *              Initializes the query plan: open files, allocate storage
 630  *              and start up the rule manager
 631  * ----------------------------------------------------------------
 632  */
 633 static void
 634 InitPlan(QueryDesc *queryDesc, int eflags)
 635 {
 636         CmdType         operation = queryDesc->operation;
 637         PlannedStmt *plannedstmt = queryDesc->plannedstmt;
 638         Plan       *plan = plannedstmt->planTree;
 639         List       *rangeTable = plannedstmt->rtable;
 640         EState     *estate = queryDesc->estate;
 641         PlanState  *planstate;
 642         TupleDesc       tupType;
 643         ListCell   *l;
 644         int                     i;
 645
 646         /*
 647          * Do permissions checks
 648          */
 649         ExecCheckRTPerms(rangeTable, true);
 650
 651         /*
 652          * initialize the node's execution state
 653          */
 654         estate->es_range_table = rangeTable;
 655         estate->es_plannedstmt = plannedstmt;
 656
 657         /*
 658          * initialize result relation stuff, and open/lock the result rels.
 659          *
 660          * We must do this before initializing the plan tree, else we might try to
 661          * do a lock upgrade if a result rel is also a source rel.
 662          */
 663         if (plannedstmt->resultRelations)
 664         {
 665                 List       *resultRelations = plannedstmt->resultRelations;
 666                 int                     numResultRelations = list_length(resultRelations);
 667                 ResultRelInfo *resultRelInfos;
 668                 ResultRelInfo *resultRelInfo;
 669
 670                 resultRelInfos = (ResultRelInfo *)
 671                         palloc(numResultRelations * sizeof(ResultRelInfo));
 672                 resultRelInfo = resultRelInfos;
 673                 foreach(l, resultRelations)
 674                 {
 675                         Index           resultRelationIndex = lfirst_int(l);
 676                         Oid                     resultRelationOid;
 677                         Relation        resultRelation;
 678
 679                         resultRelationOid = getrelid(resultRelationIndex, rangeTable);
 680                         resultRelation = heap_open(resultRelationOid, RowExclusiveLock);
 681                         InitResultRelInfo(resultRelInfo,
 682                                                           resultRelation,
 683                                                           resultRelationIndex,
 684                                                           operation,
 685                                                           estate->es_instrument);
 686                         resultRelInfo++;
 687                 }
 688                 estate->es_result_relations = resultRelInfos;
 689                 estate->es_num_result_relations = numResultRelations;
 690                 /* es_result_relation_info is NULL except when within ModifyTable */
 691                 estate->es_result_relation_info = NULL;
 692         }
 693         else
 694         {
 695                 /*
 696                  * if no result relation, then set state appropriately
 697                  */
 698                 estate->es_result_relations = NULL;
 699                 estate->es_num_result_relations = 0;
 700                 estate->es_result_relation_info = NULL;
 701         }
 702
 703         /*
 704          * Similarly, we have to lock relations selected FOR UPDATE/FOR SHARE
 705          * before we initialize the plan tree, else we'd be risking lock upgrades.
 706          * While we are at it, build the ExecRowMark list.
 707          */
 708         estate->es_rowMarks = NIL;
 709         foreach(l, plannedstmt->rowMarks)
 710         {
 711                 PlanRowMark *rc = (PlanRowMark *) lfirst(l);
 712                 Oid                     relid;
 713                 Relation        relation;
 714                 ExecRowMark *erm;
 715
 716                 /* ignore "parent" rowmarks; they are irrelevant at runtime */
 717                 if (rc->isParent)
 718                         continue;
 719
 720                 switch (rc->markType)
 721                 {
 722                         case ROW_MARK_EXCLUSIVE:
 723                         case ROW_MARK_SHARE:
 724                                 relid = getrelid(rc->rti, rangeTable);
 725                                 relation = heap_open(relid, RowShareLock);
 726                                 break;
 727                         case ROW_MARK_REFERENCE:
 728                                 relid = getrelid(rc->rti, rangeTable);
 729                                 relation = heap_open(relid, AccessShareLock);
 730                                 break;
 731                         case ROW_MARK_COPY:
 732                                 /* there's no real table here ... */
 733                                 relation = NULL;
 734                                 break;
 735                         default:
 736                                 elog(ERROR, "unrecognized markType: %d", rc->markType);
 737                                 relation = NULL;        /* keep compiler quiet */
 738                                 break;
 739                 }
 740
 741                 erm = (ExecRowMark *) palloc(sizeof(ExecRowMark));
 742                 erm->relation = relation;
 743                 erm->rti = rc->rti;
 744                 erm->prti = rc->prti;
 745                 erm->markType = rc->markType;
 746                 erm->noWait = rc->noWait;
 747                 erm->ctidAttNo = rc->ctidAttNo;
 748                 erm->toidAttNo = rc->toidAttNo;
 749                 erm->wholeAttNo = rc->wholeAttNo;
 750                 ItemPointerSetInvalid(&(erm->curCtid));
 751                 estate->es_rowMarks = lappend(estate->es_rowMarks, erm);
 752         }
 753
 754         /*
 755          * Detect whether we're doing SELECT INTO.  If so, set the es_into_oids
 756          * flag appropriately so that the plan tree will be initialized with the
 757          * correct tuple descriptors.  (Other SELECT INTO stuff comes later.)
 758          */
 759         estate->es_select_into = false;
 760         if (operation == CMD_SELECT && plannedstmt->intoClause != NULL)
 761         {
 762                 estate->es_select_into = true;
 763                 estate->es_into_oids = interpretOidsOption(plannedstmt->intoClause->options);
 764         }
 765
 766         /*
 767          * Initialize the executor's tuple table to empty.
 768          */
 769         estate->es_tupleTable = NIL;
 770         estate->es_trig_tuple_slot = NULL;
 771         estate->es_trig_oldtup_slot = NULL;
 772
 773         /* mark EvalPlanQual not active */
 774         estate->es_epqTuple = NULL;
 775         estate->es_epqTupleSet = NULL;
 776         estate->es_epqScanDone = NULL;
 777
 778         /*
 779          * Initialize private state information for each SubPlan.  We must do this
 780          * before running ExecInitNode on the main query tree, since
 781          * ExecInitSubPlan expects to be able to find these entries.
 782          */
 783         Assert(estate->es_subplanstates == NIL);
 784         i = 1;                                          /* subplan indices count from 1 */
 785         foreach(l, plannedstmt->subplans)
 786         {
 787                 Plan       *subplan = (Plan *) lfirst(l);
 788                 PlanState  *subplanstate;
 789                 int                     sp_eflags;
 790
 791                 /*
 792                  * A subplan will never need to do BACKWARD scan nor MARK/RESTORE. If
 793                  * it is a parameterless subplan (not initplan), we suggest that it be
 794                  * prepared to handle REWIND efficiently; otherwise there is no need.
 795                  */
 796                 sp_eflags = eflags & EXEC_FLAG_EXPLAIN_ONLY;
 797                 if (bms_is_member(i, plannedstmt->rewindPlanIDs))
 798                         sp_eflags |= EXEC_FLAG_REWIND;
 799
 800                 subplanstate = ExecInitNode(subplan, estate, sp_eflags);
 801
 802                 estate->es_subplanstates = lappend(estate->es_subplanstates,
 803                                                                                    subplanstate);
 804
 805                 i++;
 806         }
 807
 808         /*
 809          * Initialize the private state information for all the nodes in the query
 810          * tree.  This opens files, allocates storage and leaves us ready to start
 811          * processing tuples.
 812          */
 813         planstate = ExecInitNode(plan, estate, eflags);
 814
 815         /*
 816          * Get the tuple descriptor describing the type of tuples to return. (this
 817          * is especially important if we are creating a relation with "SELECT
 818          * INTO")
 819          */
 820         tupType = ExecGetResultType(planstate);
 821
 822         /*
 823          * Initialize the junk filter if needed.  SELECT queries need a filter if
 824          * there are any junk attrs in the top-level tlist.
 825          */
 826         if (operation == CMD_SELECT)
 827         {
 828                 bool            junk_filter_needed = false;
 829                 ListCell   *tlist;
 830
 831                 foreach(tlist, plan->targetlist)
 832                 {
 833                         TargetEntry *tle = (TargetEntry *) lfirst(tlist);
 834
 835                         if (tle->resjunk)
 836                         {
 837                                 junk_filter_needed = true;
 838                                 break;
 839                         }
 840                 }
 841
 842                 if (junk_filter_needed)
 843                 {
 844                         JunkFilter *j;
 845
 846                         j = ExecInitJunkFilter(planstate->plan->targetlist,
 847                                                                    tupType->tdhasoid,
 848                                                                    ExecInitExtraTupleSlot(estate));
 849                         estate->es_junkFilter = j;
 850
 851                         /* Want to return the cleaned tuple type */
 852                         tupType = j->jf_cleanTupType;
 853                 }
 854         }
 855
 856         queryDesc->tupDesc = tupType;
 857         queryDesc->planstate = planstate;
 858
 859         /*
 860          * If doing SELECT INTO, initialize the "into" relation.  We must wait
 861          * till now so we have the "clean" result tuple type to create the new
 862          * table from.
 863          *
 864          * If EXPLAIN, skip creating the "into" relation.
 865          */
 866         if (estate->es_select_into && !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
 867                 OpenIntoRel(queryDesc);
 868 }
 869
 870 /*
 871  * Initialize ResultRelInfo data for one result relation
 872  */
 873 void
 874 InitResultRelInfo(ResultRelInfo *resultRelInfo,
 875                                   Relation resultRelationDesc,
 876                                   Index resultRelationIndex,
 877                                   CmdType operation,
 878                                   int instrument_options)
 879 {
 880         /*
 881          * Check valid relkind ... parser and/or planner should have noticed this
 882          * already, but let's make sure.
 883          */
 884         switch (resultRelationDesc->rd_rel->relkind)
 885         {
 886                 case RELKIND_RELATION:
 887                         /* OK */
 888                         break;
 889                 case RELKIND_SEQUENCE:
 890                         ereport(ERROR,
 891                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 892                                          errmsg("cannot change sequence \"%s\"",
 893                                                         RelationGetRelationName(resultRelationDesc))));
 894                         break;
 895                 case RELKIND_TOASTVALUE:
 896                         ereport(ERROR,
 897                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 898                                          errmsg("cannot change TOAST relation \"%s\"",
 899                                                         RelationGetRelationName(resultRelationDesc))));
 900                         break;
 901                 case RELKIND_VIEW:
 902                         ereport(ERROR,
 903                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 904                                          errmsg("cannot change view \"%s\"",
 905                                                         RelationGetRelationName(resultRelationDesc))));
 906                         break;
 907                 default:
 908                         ereport(ERROR,
 909                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 910                                          errmsg("cannot change relation \"%s\"",
 911                                                         RelationGetRelationName(resultRelationDesc))));
 912                         break;
 913         }
 914
 915         /* OK, fill in the node */
 916         MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
 917         resultRelInfo->type = T_ResultRelInfo;
 918         resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
 919         resultRelInfo->ri_RelationDesc = resultRelationDesc;
 920         resultRelInfo->ri_NumIndices = 0;
 921         resultRelInfo->ri_IndexRelationDescs = NULL;
 922         resultRelInfo->ri_IndexRelationInfo = NULL;
 923         /* make a copy so as not to depend on relcache info not changing... */
 924         resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
 925         if (resultRelInfo->ri_TrigDesc)
 926         {
 927                 int                     n = resultRelInfo->ri_TrigDesc->numtriggers;
 928
 929                 resultRelInfo->ri_TrigFunctions = (FmgrInfo *)
 930                         palloc0(n * sizeof(FmgrInfo));
 931                 resultRelInfo->ri_TrigWhenExprs = (List **)
 932                         palloc0(n * sizeof(List *));
 933                 if (instrument_options)
 934                         resultRelInfo->ri_TrigInstrument = InstrAlloc(n, instrument_options);
 935         }
 936         else
 937         {
 938                 resultRelInfo->ri_TrigFunctions = NULL;
 939                 resultRelInfo->ri_TrigWhenExprs = NULL;
 940                 resultRelInfo->ri_TrigInstrument = NULL;
 941         }
 942         resultRelInfo->ri_ConstraintExprs = NULL;
 943         resultRelInfo->ri_junkFilter = NULL;
 944         resultRelInfo->ri_projectReturning = NULL;
 945
 946         /*
 947          * If there are indices on the result relation, open them and save
 948          * descriptors in the result relation info, so that we can add new index
 949          * entries for the tuples we add/update.  We need not do this for a
 950          * DELETE, however, since deletion doesn't affect indexes.
 951          */
 952         if (resultRelationDesc->rd_rel->relhasindex &&
 953                 operation != CMD_DELETE)
 954                 ExecOpenIndices(resultRelInfo);
 955 }
 956
 957 /*
 958  *              ExecGetTriggerResultRel
 959  *
 960  * Get a ResultRelInfo for a trigger target relation.  Most of the time,
 961  * triggers are fired on one of the result relations of the query, and so
 962  * we can just return a member of the es_result_relations array.  (Note: in
 963  * self-join situations there might be multiple members with the same OID;
 964  * if so it doesn't matter which one we pick.)  However, it is sometimes
 965  * necessary to fire triggers on other relations; this happens mainly when an
 966  * RI update trigger queues additional triggers on other relations, which will
 967  * be processed in the context of the outer query.      For efficiency's sake,
 968  * we want to have a ResultRelInfo for those triggers too; that can avoid
 969  * repeated re-opening of the relation.  (It also provides a way for EXPLAIN
 970  * ANALYZE to report the runtimes of such triggers.)  So we make additional
 971  * ResultRelInfo's as needed, and save them in es_trig_target_relations.
 972  */
 973 ResultRelInfo *
 974 ExecGetTriggerResultRel(EState *estate, Oid relid)
 975 {
 976         ResultRelInfo *rInfo;
 977         int                     nr;
 978         ListCell   *l;
 979         Relation        rel;
 980         MemoryContext oldcontext;
 981
 982         /* First, search through the query result relations */
 983         rInfo = estate->es_result_relations;
 984         nr = estate->es_num_result_relations;
 985         while (nr > 0)
 986         {
 987                 if (RelationGetRelid(rInfo->ri_RelationDesc) == relid)
 988                         return rInfo;
 989                 rInfo++;
 990                 nr--;
 991         }
 992         /* Nope, but maybe we already made an extra ResultRelInfo for it */
 993         foreach(l, estate->es_trig_target_relations)
 994         {
 995                 rInfo = (ResultRelInfo *) lfirst(l);
 996                 if (RelationGetRelid(rInfo->ri_RelationDesc) == relid)
 997                         return rInfo;
 998         }
 999         /* Nope, so we need a new one */
1000
1001         /*
1002          * Open the target relation's relcache entry.  We assume that an
1003          * appropriate lock is still held by the backend from whenever the trigger
1004          * event got queued, so we need take no new lock here.
1005          */
1006         rel = heap_open(relid, NoLock);
1007
1008         /*
1009          * Make the new entry in the right context.  Currently, we don't need any
1010          * index information in ResultRelInfos used only for triggers, so tell
1011          * InitResultRelInfo it's a DELETE.
1012          */
1013         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
1014         rInfo = makeNode(ResultRelInfo);
1015         InitResultRelInfo(rInfo,
1016                                           rel,
1017                                           0,            /* dummy rangetable index */
1018                                           CMD_DELETE,
1019                                           estate->es_instrument);
1020         estate->es_trig_target_relations =
1021                 lappend(estate->es_trig_target_relations, rInfo);
1022         MemoryContextSwitchTo(oldcontext);
1023
1024         return rInfo;
1025 }
1026
1027 /*
1028  *              ExecContextForcesOids
1029  *
1030  * This is pretty grotty: when doing INSERT, UPDATE, or SELECT INTO,
1031  * we need to ensure that result tuples have space for an OID iff they are
1032  * going to be stored into a relation that has OIDs.  In other contexts
1033  * we are free to choose whether to leave space for OIDs in result tuples
1034  * (we generally don't want to, but we do if a physical-tlist optimization
1035  * is possible).  This routine checks the plan context and returns TRUE if the
1036  * choice is forced, FALSE if the choice is not forced.  In the TRUE case,
1037  * *hasoids is set to the required value.
1038  *
1039  * One reason this is ugly is that all plan nodes in the plan tree will emit
1040  * tuples with space for an OID, though we really only need the topmost node
1041  * to do so.  However, node types like Sort don't project new tuples but just
1042  * return their inputs, and in those cases the requirement propagates down
1043  * to the input node.  Eventually we might make this code smart enough to
1044  * recognize how far down the requirement really goes, but for now we just
1045  * make all plan nodes do the same thing if the top level forces the choice.
1046  *
1047  * We assume that if we are generating tuples for INSERT or UPDATE,
1048  * estate->es_result_relation_info is already set up to describe the target
1049  * relation.  Note that in an UPDATE that spans an inheritance tree, some of
1050  * the target relations may have OIDs and some not.  We have to make the
1051  * decisions on a per-relation basis as we initialize each of the subplans of
1052  * the ModifyTable node, so ModifyTable has to set es_result_relation_info
1053  * while initializing each subplan.
1054  *
1055  * SELECT INTO is even uglier, because we don't have the INTO relation's
1056  * descriptor available when this code runs; we have to look aside at a
1057  * flag set by InitPlan().
1058  */
1059 bool
1060 ExecContextForcesOids(PlanState *planstate, bool *hasoids)
1061 {
1062         ResultRelInfo *ri = planstate->state->es_result_relation_info;
1063
1064         if (ri != NULL)
1065         {
1066                 Relation        rel = ri->ri_RelationDesc;
1067
1068                 if (rel != NULL)
1069                 {
1070                         *hasoids = rel->rd_rel->relhasoids;
1071                         return true;
1072                 }
1073         }
1074
1075         if (planstate->state->es_select_into)
1076         {
1077                 *hasoids = planstate->state->es_into_oids;
1078                 return true;
1079         }
1080
1081         return false;
1082 }
1083
1084 /* ----------------------------------------------------------------
1085  *              ExecEndPlan
1086  *
1087  *              Cleans up the query plan -- closes files and frees up storage
1088  *
1089  * NOTE: we are no longer very worried about freeing storage per se
1090  * in this code; FreeExecutorState should be guaranteed to release all
1091  * memory that needs to be released.  What we are worried about doing
1092  * is closing relations and dropping buffer pins.  Thus, for example,
1093  * tuple tables must be cleared or dropped to ensure pins are released.
1094  * ----------------------------------------------------------------
1095  */
1096 static void
1097 ExecEndPlan(PlanState *planstate, EState *estate)
1098 {
1099         ResultRelInfo *resultRelInfo;
1100         int                     i;
1101         ListCell   *l;
1102
1103         /*
1104          * shut down the node-type-specific query processing
1105          */
1106         ExecEndNode(planstate);
1107
1108         /*
1109          * for subplans too
1110          */
1111         foreach(l, estate->es_subplanstates)
1112         {
1113                 PlanState  *subplanstate = (PlanState *) lfirst(l);
1114
1115                 ExecEndNode(subplanstate);
1116         }
1117
1118         /*
1119          * destroy the executor's tuple table.  Actually we only care about
1120          * releasing buffer pins and tupdesc refcounts; there's no need to pfree
1121          * the TupleTableSlots, since the containing memory context is about to go
1122          * away anyway.
1123          */
1124         ExecResetTupleTable(estate->es_tupleTable, false);
1125
1126         /*
1127          * close the result relation(s) if any, but hold locks until xact commit.
1128          */
1129         resultRelInfo = estate->es_result_relations;
1130         for (i = estate->es_num_result_relations; i > 0; i--)
1131         {
1132                 /* Close indices and then the relation itself */
1133                 ExecCloseIndices(resultRelInfo);
1134                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
1135                 resultRelInfo++;
1136         }
1137
1138         /*
1139          * likewise close any trigger target relations
1140          */
1141         foreach(l, estate->es_trig_target_relations)
1142         {
1143                 resultRelInfo = (ResultRelInfo *) lfirst(l);
1144                 /* Close indices and then the relation itself */
1145                 ExecCloseIndices(resultRelInfo);
1146                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
1147         }
1148
1149         /*
1150          * close any relations selected FOR UPDATE/FOR SHARE, again keeping locks
1151          */
1152         foreach(l, estate->es_rowMarks)
1153         {
1154                 ExecRowMark *erm = (ExecRowMark *) lfirst(l);
1155
1156                 if (erm->relation)
1157                         heap_close(erm->relation, NoLock);
1158         }
1159 }
1160
1161 /* ----------------------------------------------------------------
1162  *              ExecutePlan
1163  *
1164  *              Processes the query plan until we have processed 'numberTuples' tuples,
1165  *              moving in the specified direction.
1166  *
1167  *              Runs to completion if numberTuples is 0
1168  *
1169  * Note: the ctid attribute is a 'junk' attribute that is removed before the
1170  * user can see it
1171  * ----------------------------------------------------------------
1172  */
1173 static void
1174 ExecutePlan(EState *estate,
1175                         PlanState *planstate,
1176                         CmdType operation,
1177                         bool sendTuples,
1178                         long numberTuples,
1179                         ScanDirection direction,
1180                         DestReceiver *dest)
1181 {
1182         TupleTableSlot *slot;
1183         long            current_tuple_count;
1184
1185         /*
1186          * initialize local variables
1187          */
1188         current_tuple_count = 0;
1189
1190         /*
1191          * Set the direction.
1192          */
1193         estate->es_direction = direction;
1194
1195         /*
1196          * Loop until we've processed the proper number of tuples from the plan.
1197          */
1198         for (;;)
1199         {
1200                 /* Reset the per-output-tuple exprcontext */
1201                 ResetPerTupleExprContext(estate);
1202
1203                 /*
1204                  * Execute the plan and obtain a tuple
1205                  */
1206                 slot = ExecProcNode(planstate);
1207
1208                 /*
1209                  * if the tuple is null, then we assume there is nothing more to
1210                  * process so we just end the loop...
1211                  */
1212                 if (TupIsNull(slot))
1213                         break;
1214
1215                 /*
1216                  * If we have a junk filter, then project a new tuple with the junk
1217                  * removed.
1218                  *
1219                  * Store this new "clean" tuple in the junkfilter's resultSlot.
1220                  * (Formerly, we stored it back over the "dirty" tuple, which is WRONG
1221                  * because that tuple slot has the wrong descriptor.)
1222                  */
1223                 if (estate->es_junkFilter != NULL)
1224                         slot = ExecFilterJunk(estate->es_junkFilter, slot);
1225
1226                 /*
1227                  * If we are supposed to send the tuple somewhere, do so. (In
1228                  * practice, this is probably always the case at this point.)
1229                  */
1230                 if (sendTuples)
1231                         (*dest->receiveSlot) (slot, dest);
1232
1233                 /*
1234                  * Count tuples processed, if this is a SELECT.  (For other operation
1235                  * types, the ModifyTable plan node must count the appropriate
1236                  * events.)
1237                  */
1238                 if (operation == CMD_SELECT)
1239                         (estate->es_processed)++;
1240
1241                 /*
1242                  * check our tuple count.. if we've processed the proper number then
1243                  * quit, else loop again and process more tuples.  Zero numberTuples
1244                  * means no limit.
1245                  */
1246                 current_tuple_count++;
1247                 if (numberTuples && numberTuples == current_tuple_count)
1248                         break;
1249         }
1250 }
1251
1252
1253 /*
1254  * ExecRelCheck --- check that tuple meets constraints for result relation
1255  */
1256 static const char *
1257 ExecRelCheck(ResultRelInfo *resultRelInfo,
1258                          TupleTableSlot *slot, EState *estate)
1259 {
1260         Relation        rel = resultRelInfo->ri_RelationDesc;
1261         int                     ncheck = rel->rd_att->constr->num_check;
1262         ConstrCheck *check = rel->rd_att->constr->check;
1263         ExprContext *econtext;
1264         MemoryContext oldContext;
1265         List       *qual;
1266         int                     i;
1267
1268         /*
1269          * If first time through for this result relation, build expression
1270          * nodetrees for rel's constraint expressions.  Keep them in the per-query
1271          * memory context so they'll survive throughout the query.
1272          */
1273         if (resultRelInfo->ri_ConstraintExprs == NULL)
1274         {
1275                 oldContext = MemoryContextSwitchTo(estate->es_query_cxt);
1276                 resultRelInfo->ri_ConstraintExprs =
1277                         (List **) palloc(ncheck * sizeof(List *));
1278                 for (i = 0; i < ncheck; i++)
1279                 {
1280                         /* ExecQual wants implicit-AND form */
1281                         qual = make_ands_implicit(stringToNode(check[i].ccbin));
1282                         resultRelInfo->ri_ConstraintExprs[i] = (List *)
1283                                 ExecPrepareExpr((Expr *) qual, estate);
1284                 }
1285                 MemoryContextSwitchTo(oldContext);
1286         }
1287
1288         /*
1289          * We will use the EState's per-tuple context for evaluating constraint
1290          * expressions (creating it if it's not already there).
1291          */
1292         econtext = GetPerTupleExprContext(estate);
1293
1294         /* Arrange for econtext's scan tuple to be the tuple under test */
1295         econtext->ecxt_scantuple = slot;
1296
1297         /* And evaluate the constraints */
1298         for (i = 0; i < ncheck; i++)
1299         {
1300                 qual = resultRelInfo->ri_ConstraintExprs[i];
1301
1302                 /*
1303                  * NOTE: SQL92 specifies that a NULL result from a constraint
1304                  * expression is not to be treated as a failure.  Therefore, tell
1305                  * ExecQual to return TRUE for NULL.
1306                  */
1307                 if (!ExecQual(qual, econtext, true))
1308                         return check[i].ccname;
1309         }
1310
1311         /* NULL result means no error */
1312         return NULL;
1313 }
1314
1315 void
1316 ExecConstraints(ResultRelInfo *resultRelInfo,
1317                                 TupleTableSlot *slot, EState *estate)
1318 {
1319         Relation        rel = resultRelInfo->ri_RelationDesc;
1320         TupleConstr *constr = rel->rd_att->constr;
1321
1322         Assert(constr);
1323
1324         if (constr->has_not_null)
1325         {
1326                 int                     natts = rel->rd_att->natts;
1327                 int                     attrChk;
1328
1329                 for (attrChk = 1; attrChk <= natts; attrChk++)
1330                 {
1331                         if (rel->rd_att->attrs[attrChk - 1]->attnotnull &&
1332                                 slot_attisnull(slot, attrChk))
1333                                 ereport(ERROR,
1334                                                 (errcode(ERRCODE_NOT_NULL_VIOLATION),
1335                                                  errmsg("null value in column \"%s\" violates not-null constraint",
1336                                                 NameStr(rel->rd_att->attrs[attrChk - 1]->attname))));
1337                 }
1338         }
1339
1340         if (constr->num_check > 0)
1341         {
1342                 const char *failed;
1343
1344                 if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
1345                         ereport(ERROR,
1346                                         (errcode(ERRCODE_CHECK_VIOLATION),
1347                                          errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
1348                                                         RelationGetRelationName(rel), failed)));
1349         }
1350 }
1351
1352
1353 /*
1354  * EvalPlanQual logic --- recheck modified tuple(s) to see if we want to
1355  * process the updated version under READ COMMITTED rules.
1356  *
1357  * See backend/executor/README for some info about how this works.
1358  */
1359
1360
1361 /*
1362  * Check a modified tuple to see if we want to process its updated version
1363  * under READ COMMITTED rules.
1364  *
1365  *      estate - outer executor state data
1366  *      epqstate - state for EvalPlanQual rechecking
1367  *      relation - table containing tuple
1368  *      rti - rangetable index of table containing tuple
1369  *      *tid - t_ctid from the outdated tuple (ie, next updated version)
1370  *      priorXmax - t_xmax from the outdated tuple
1371  *
1372  * *tid is also an output parameter: it's modified to hold the TID of the
1373  * latest version of the tuple (note this may be changed even on failure)
1374  *
1375  * Returns a slot containing the new candidate update/delete tuple, or
1376  * NULL if we determine we shouldn't process the row.
1377  */
1378 TupleTableSlot *
1379 EvalPlanQual(EState *estate, EPQState *epqstate,
1380                          Relation relation, Index rti,
1381                          ItemPointer tid, TransactionId priorXmax)
1382 {
1383         TupleTableSlot *slot;
1384         HeapTuple       copyTuple;
1385
1386         Assert(rti > 0);
1387
1388         /*
1389          * Get and lock the updated version of the row; if fail, return NULL.
1390          */
1391         copyTuple = EvalPlanQualFetch(estate, relation, LockTupleExclusive,
1392                                                                   tid, priorXmax);
1393
1394         if (copyTuple == NULL)
1395                 return NULL;
1396
1397         /*
1398          * For UPDATE/DELETE we have to return tid of actual row we're executing
1399          * PQ for.
1400          */
1401         *tid = copyTuple->t_self;
1402
1403         /*
1404          * Need to run a recheck subquery.      Initialize or reinitialize EPQ state.
1405          */
1406         EvalPlanQualBegin(epqstate, estate);
1407
1408         /*
1409          * Free old test tuple, if any, and store new tuple where relation's scan
1410          * node will see it
1411          */
1412         EvalPlanQualSetTuple(epqstate, rti, copyTuple);
1413
1414         /*
1415          * Fetch any non-locked source rows
1416          */
1417         EvalPlanQualFetchRowMarks(epqstate);
1418
1419         /*
1420          * Run the EPQ query.  We assume it will return at most one tuple.
1421          */
1422         slot = EvalPlanQualNext(epqstate);
1423
1424         /*
1425          * If we got a tuple, force the slot to materialize the tuple so that it
1426          * is not dependent on any local state in the EPQ query (in particular,
1427          * it's highly likely that the slot contains references to any pass-by-ref
1428          * datums that may be present in copyTuple).  As with the next step, this
1429          * is to guard against early re-use of the EPQ query.
1430          */
1431         if (!TupIsNull(slot))
1432                 (void) ExecMaterializeSlot(slot);
1433
1434         /*
1435          * Clear out the test tuple.  This is needed in case the EPQ query is
1436          * re-used to test a tuple for a different relation.  (Not clear that can
1437          * really happen, but let's be safe.)
1438          */
1439         EvalPlanQualSetTuple(epqstate, rti, NULL);
1440
1441         return slot;
1442 }
1443
1444 /*
1445  * Fetch a copy of the newest version of an outdated tuple
1446  *
1447  *      estate - executor state data
1448  *      relation - table containing tuple
1449  *      lockmode - requested tuple lock mode
1450  *      *tid - t_ctid from the outdated tuple (ie, next updated version)
1451  *      priorXmax - t_xmax from the outdated tuple
1452  *
1453  * Returns a palloc'd copy of the newest tuple version, or NULL if we find
1454  * that there is no newest version (ie, the row was deleted not updated).
1455  * If successful, we have locked the newest tuple version, so caller does not
1456  * need to worry about it changing anymore.
1457  *
1458  * Note: properly, lockmode should be declared as enum LockTupleMode,
1459  * but we use "int" to avoid having to include heapam.h in executor.h.
1460  */
1461 HeapTuple
1462 EvalPlanQualFetch(EState *estate, Relation relation, int lockmode,
1463                                   ItemPointer tid, TransactionId priorXmax)
1464 {
1465         HeapTuple       copyTuple = NULL;
1466         HeapTupleData tuple;
1467         SnapshotData SnapshotDirty;
1468
1469         /*
1470          * fetch target tuple
1471          *
1472          * Loop here to deal with updated or busy tuples
1473          */
1474         InitDirtySnapshot(SnapshotDirty);
1475         tuple.t_self = *tid;
1476         for (;;)
1477         {
1478                 Buffer          buffer;
1479
1480                 if (heap_fetch(relation, &SnapshotDirty, &tuple, &buffer, true, NULL))
1481                 {
1482                         HTSU_Result test;
1483                         ItemPointerData update_ctid;
1484                         TransactionId update_xmax;
1485
1486                         /*
1487                          * If xmin isn't what we're expecting, the slot must have been
1488                          * recycled and reused for an unrelated tuple.  This implies that
1489                          * the latest version of the row was deleted, so we need do
1490                          * nothing.  (Should be safe to examine xmin without getting
1491                          * buffer's content lock, since xmin never changes in an existing
1492                          * tuple.)
1493                          */
1494                         if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
1495                                                                          priorXmax))
1496                         {
1497                                 ReleaseBuffer(buffer);
1498                                 return NULL;
1499                         }
1500
1501                         /* otherwise xmin should not be dirty... */
1502                         if (TransactionIdIsValid(SnapshotDirty.xmin))
1503                                 elog(ERROR, "t_xmin is uncommitted in tuple to be updated");
1504
1505                         /*
1506                          * If tuple is being updated by other transaction then we have to
1507                          * wait for its commit/abort.
1508                          */
1509                         if (TransactionIdIsValid(SnapshotDirty.xmax))
1510                         {
1511                                 ReleaseBuffer(buffer);
1512                                 XactLockTableWait(SnapshotDirty.xmax);
1513                                 continue;               /* loop back to repeat heap_fetch */
1514                         }
1515
1516                         /*
1517                          * If tuple was inserted by our own transaction, we have to check
1518                          * cmin against es_output_cid: cmin >= current CID means our
1519                          * command cannot see the tuple, so we should ignore it.  Without
1520                          * this we are open to the "Halloween problem" of indefinitely
1521                          * re-updating the same tuple. (We need not check cmax because
1522                          * HeapTupleSatisfiesDirty will consider a tuple deleted by our
1523                          * transaction dead, regardless of cmax.)  We just checked that
1524                          * priorXmax == xmin, so we can test that variable instead of
1525                          * doing HeapTupleHeaderGetXmin again.
1526                          */
1527                         if (TransactionIdIsCurrentTransactionId(priorXmax) &&
1528                                 HeapTupleHeaderGetCmin(tuple.t_data) >= estate->es_output_cid)
1529                         {
1530                                 ReleaseBuffer(buffer);
1531                                 return NULL;
1532                         }
1533
1534                         /*
1535                          * This is a live tuple, so now try to lock it.
1536                          */
1537                         test = heap_lock_tuple(relation, &tuple, &buffer,
1538                                                                    &update_ctid, &update_xmax,
1539                                                                    estate->es_output_cid,
1540                                                                    lockmode, false);
1541                         /* We now have two pins on the buffer, get rid of one */
1542                         ReleaseBuffer(buffer);
1543
1544                         switch (test)
1545                         {
1546                                 case HeapTupleSelfUpdated:
1547                                         /* treat it as deleted; do not process */
1548                                         ReleaseBuffer(buffer);
1549                                         return NULL;
1550
1551                                 case HeapTupleMayBeUpdated:
1552                                         /* successfully locked */
1553                                         break;
1554
1555                                 case HeapTupleUpdated:
1556                                         ReleaseBuffer(buffer);
1557                                         if (IsXactIsoLevelSerializable)
1558                                                 ereport(ERROR,
1559                                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1560                                                                  errmsg("could not serialize access due to concurrent update")));
1561                                         if (!ItemPointerEquals(&update_ctid, &tuple.t_self))
1562                                         {
1563                                                 /* it was updated, so look at the updated version */
1564                                                 tuple.t_self = update_ctid;
1565                                                 /* updated row should have xmin matching this xmax */
1566                                                 priorXmax = update_xmax;
1567                                                 continue;
1568                                         }
1569                                         /* tuple was deleted, so give up */
1570                                         return NULL;
1571
1572                                 default:
1573                                         ReleaseBuffer(buffer);
1574                                         elog(ERROR, "unrecognized heap_lock_tuple status: %u",
1575                                                  test);
1576                                         return NULL;    /* keep compiler quiet */
1577                         }
1578
1579                         /*
1580                          * We got tuple - now copy it for use by recheck query.
1581                          */
1582                         copyTuple = heap_copytuple(&tuple);
1583                         ReleaseBuffer(buffer);
1584                         break;
1585                 }
1586
1587                 /*
1588                  * If the referenced slot was actually empty, the latest version of
1589                  * the row must have been deleted, so we need do nothing.
1590                  */
1591                 if (tuple.t_data == NULL)
1592                 {
1593                         ReleaseBuffer(buffer);
1594                         return NULL;
1595                 }
1596
1597                 /*
1598                  * As above, if xmin isn't what we're expecting, do nothing.
1599                  */
1600                 if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
1601                                                                  priorXmax))
1602                 {
1603                         ReleaseBuffer(buffer);
1604                         return NULL;
1605                 }
1606
1607                 /*
1608                  * If we get here, the tuple was found but failed SnapshotDirty.
1609                  * Assuming the xmin is either a committed xact or our own xact (as it
1610                  * certainly should be if we're trying to modify the tuple), this must
1611                  * mean that the row was updated or deleted by either a committed xact
1612                  * or our own xact.  If it was deleted, we can ignore it; if it was
1613                  * updated then chain up to the next version and repeat the whole
1614                  * process.
1615                  *
1616                  * As above, it should be safe to examine xmax and t_ctid without the
1617                  * buffer content lock, because they can't be changing.
1618                  */
1619                 if (ItemPointerEquals(&tuple.t_self, &tuple.t_data->t_ctid))
1620                 {
1621                         /* deleted, so forget about it */
1622                         ReleaseBuffer(buffer);
1623                         return NULL;
1624                 }
1625
1626                 /* updated, so look at the updated row */
1627                 tuple.t_self = tuple.t_data->t_ctid;
1628                 /* updated row should have xmin matching this xmax */
1629                 priorXmax = HeapTupleHeaderGetXmax(tuple.t_data);
1630                 ReleaseBuffer(buffer);
1631                 /* loop back to fetch next in chain */
1632         }
1633
1634         /*
1635          * Return the copied tuple
1636          */
1637         return copyTuple;
1638 }
1639
1640 /*
1641  * EvalPlanQualInit -- initialize during creation of a plan state node
1642  * that might need to invoke EPQ processing.
1643  * Note: subplan can be NULL if it will be set later with EvalPlanQualSetPlan.
1644  */
1645 void
1646 EvalPlanQualInit(EPQState *epqstate, EState *estate,
1647                                  Plan *subplan, int epqParam)
1648 {
1649         /* Mark the EPQ state inactive */
1650         epqstate->estate = NULL;
1651         epqstate->planstate = NULL;
1652         epqstate->origslot = NULL;
1653         /* ... and remember data that EvalPlanQualBegin will need */
1654         epqstate->plan = subplan;
1655         epqstate->rowMarks = NIL;
1656         epqstate->epqParam = epqParam;
1657 }
1658
1659 /*
1660  * EvalPlanQualSetPlan -- set or change subplan of an EPQState.
1661  *
1662  * We need this so that ModifyTuple can deal with multiple subplans.
1663  */
1664 void
1665 EvalPlanQualSetPlan(EPQState *epqstate, Plan *subplan)
1666 {
1667         /* If we have a live EPQ query, shut it down */
1668         EvalPlanQualEnd(epqstate);
1669         /* And set/change the plan pointer */
1670         epqstate->plan = subplan;
1671 }
1672
1673 /*
1674  * EvalPlanQualAddRowMark -- add an ExecRowMark that EPQ needs to handle.
1675  *
1676  * Currently, only non-locking RowMarks are supported.
1677  */
1678 void
1679 EvalPlanQualAddRowMark(EPQState *epqstate, ExecRowMark *erm)
1680 {
1681         if (RowMarkRequiresRowShareLock(erm->markType))
1682                 elog(ERROR, "EvalPlanQual doesn't support locking rowmarks");
1683         epqstate->rowMarks = lappend(epqstate->rowMarks, erm);
1684 }
1685
1686 /*
1687  * Install one test tuple into EPQ state, or clear test tuple if tuple == NULL
1688  *
1689  * NB: passed tuple must be palloc'd; it may get freed later
1690  */
1691 void
1692 EvalPlanQualSetTuple(EPQState *epqstate, Index rti, HeapTuple tuple)
1693 {
1694         EState     *estate = epqstate->estate;
1695
1696         Assert(rti > 0);
1697
1698         /*
1699          * free old test tuple, if any, and store new tuple where relation's scan
1700          * node will see it
1701          */
1702         if (estate->es_epqTuple[rti - 1] != NULL)
1703                 heap_freetuple(estate->es_epqTuple[rti - 1]);
1704         estate->es_epqTuple[rti - 1] = tuple;
1705         estate->es_epqTupleSet[rti - 1] = true;
1706 }
1707
1708 /*
1709  * Fetch back the current test tuple (if any) for the specified RTI
1710  */
1711 HeapTuple
1712 EvalPlanQualGetTuple(EPQState *epqstate, Index rti)
1713 {
1714         EState     *estate = epqstate->estate;
1715
1716         Assert(rti > 0);
1717
1718         return estate->es_epqTuple[rti - 1];
1719 }
1720
1721 /*
1722  * Fetch the current row values for any non-locked relations that need
1723  * to be scanned by an EvalPlanQual operation.  origslot must have been set
1724  * to contain the current result row (top-level row) that we need to recheck.
1725  */
1726 void
1727 EvalPlanQualFetchRowMarks(EPQState *epqstate)
1728 {
1729         ListCell   *l;
1730
1731         Assert(epqstate->origslot != NULL);
1732
1733         foreach(l, epqstate->rowMarks)
1734         {
1735                 ExecRowMark *erm = (ExecRowMark *) lfirst(l);
1736                 Datum           datum;
1737                 bool            isNull;
1738                 HeapTupleData tuple;
1739
1740                 /* clear any leftover test tuple for this rel */
1741                 EvalPlanQualSetTuple(epqstate, erm->rti, NULL);
1742
1743                 if (erm->relation)
1744                 {
1745                         Buffer          buffer;
1746
1747                         Assert(erm->markType == ROW_MARK_REFERENCE);
1748
1749                         /* if child rel, must check whether it produced this row */
1750                         if (erm->rti != erm->prti)
1751                         {
1752                                 Oid                     tableoid;
1753
1754                                 datum = ExecGetJunkAttribute(epqstate->origslot,
1755                                                                                          erm->toidAttNo,
1756                                                                                          &isNull);
1757                                 /* non-locked rels could be on the inside of outer joins */
1758                                 if (isNull)
1759                                         continue;
1760                                 tableoid = DatumGetObjectId(datum);
1761
1762                                 if (tableoid != RelationGetRelid(erm->relation))
1763                                 {
1764                                         /* this child is inactive right now */
1765                                         continue;
1766                                 }
1767                         }
1768
1769                         /* fetch the tuple's ctid */
1770                         datum = ExecGetJunkAttribute(epqstate->origslot,
1771                                                                                  erm->ctidAttNo,
1772                                                                                  &isNull);
1773                         /* non-locked rels could be on the inside of outer joins */
1774                         if (isNull)
1775                                 continue;
1776                         tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
1777
1778                         /* okay, fetch the tuple */
1779                         if (!heap_fetch(erm->relation, SnapshotAny, &tuple, &buffer,
1780                                                         false, NULL))
1781                                 elog(ERROR, "failed to fetch tuple for EvalPlanQual recheck");
1782
1783                         /* successful, copy and store tuple */
1784                         EvalPlanQualSetTuple(epqstate, erm->rti,
1785                                                                  heap_copytuple(&tuple));
1786                         ReleaseBuffer(buffer);
1787                 }
1788                 else
1789                 {
1790                         HeapTupleHeader td;
1791
1792                         Assert(erm->markType == ROW_MARK_COPY);
1793
1794                         /* fetch the whole-row Var for the relation */
1795                         datum = ExecGetJunkAttribute(epqstate->origslot,
1796                                                                                  erm->wholeAttNo,
1797                                                                                  &isNull);
1798                         /* non-locked rels could be on the inside of outer joins */
1799                         if (isNull)
1800                                 continue;
1801                         td = DatumGetHeapTupleHeader(datum);
1802
1803                         /* build a temporary HeapTuple control structure */
1804                         tuple.t_len = HeapTupleHeaderGetDatumLength(td);
1805                         ItemPointerSetInvalid(&(tuple.t_self));
1806                         tuple.t_tableOid = InvalidOid;
1807                         tuple.t_data = td;
1808
1809                         /* copy and store tuple */
1810                         EvalPlanQualSetTuple(epqstate, erm->rti,
1811                                                                  heap_copytuple(&tuple));
1812                 }
1813         }
1814 }
1815
1816 /*
1817  * Fetch the next row (if any) from EvalPlanQual testing
1818  *
1819  * (In practice, there should never be more than one row...)
1820  */
1821 TupleTableSlot *
1822 EvalPlanQualNext(EPQState *epqstate)
1823 {
1824         MemoryContext oldcontext;
1825         TupleTableSlot *slot;
1826
1827         oldcontext = MemoryContextSwitchTo(epqstate->estate->es_query_cxt);
1828         slot = ExecProcNode(epqstate->planstate);
1829         MemoryContextSwitchTo(oldcontext);
1830
1831         return slot;
1832 }
1833
1834 /*
1835  * Initialize or reset an EvalPlanQual state tree
1836  */
1837 void
1838 EvalPlanQualBegin(EPQState *epqstate, EState *parentestate)
1839 {
1840         EState     *estate = epqstate->estate;
1841
1842         if (estate == NULL)
1843         {
1844                 /* First time through, so create a child EState */
1845                 EvalPlanQualStart(epqstate, parentestate, epqstate->plan);
1846         }
1847         else
1848         {
1849                 /*
1850                  * We already have a suitable child EPQ tree, so just reset it.
1851                  */
1852                 int                     rtsize = list_length(parentestate->es_range_table);
1853                 PlanState  *planstate = epqstate->planstate;
1854
1855                 MemSet(estate->es_epqScanDone, 0, rtsize * sizeof(bool));
1856
1857                 /* Recopy current values of parent parameters */
1858                 if (parentestate->es_plannedstmt->nParamExec > 0)
1859                 {
1860                         int                     i = parentestate->es_plannedstmt->nParamExec;
1861
1862                         while (--i >= 0)
1863                         {
1864                                 /* copy value if any, but not execPlan link */
1865                                 estate->es_param_exec_vals[i].value =
1866                                         parentestate->es_param_exec_vals[i].value;
1867                                 estate->es_param_exec_vals[i].isnull =
1868                                         parentestate->es_param_exec_vals[i].isnull;
1869                         }
1870                 }
1871
1872                 /*
1873                  * Mark child plan tree as needing rescan at all scan nodes.  The
1874                  * first ExecProcNode will take care of actually doing the rescan.
1875                  */
1876                 planstate->chgParam = bms_add_member(planstate->chgParam,
1877                                                                                          epqstate->epqParam);
1878         }
1879 }
1880
1881 /*
1882  * Start execution of an EvalPlanQual plan tree.
1883  *
1884  * This is a cut-down version of ExecutorStart(): we copy some state from
1885  * the top-level estate rather than initializing it fresh.
1886  */
1887 static void
1888 EvalPlanQualStart(EPQState *epqstate, EState *parentestate, Plan *planTree)
1889 {
1890         EState     *estate;
1891         int                     rtsize;
1892         MemoryContext oldcontext;
1893         ListCell   *l;
1894
1895         rtsize = list_length(parentestate->es_range_table);
1896
1897         epqstate->estate = estate = CreateExecutorState();
1898
1899         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
1900
1901         /*
1902          * Child EPQ EStates share the parent's copy of unchanging state such as
1903          * the snapshot, rangetable, result-rel info, and external Param info.
1904          * They need their own copies of local state, including a tuple table,
1905          * es_param_exec_vals, etc.
1906          */
1907         estate->es_direction = ForwardScanDirection;
1908         estate->es_snapshot = parentestate->es_snapshot;
1909         estate->es_crosscheck_snapshot = parentestate->es_crosscheck_snapshot;
1910         estate->es_range_table = parentestate->es_range_table;
1911         estate->es_plannedstmt = parentestate->es_plannedstmt;
1912         estate->es_junkFilter = parentestate->es_junkFilter;
1913         estate->es_output_cid = parentestate->es_output_cid;
1914         estate->es_result_relations = parentestate->es_result_relations;
1915         estate->es_num_result_relations = parentestate->es_num_result_relations;
1916         estate->es_result_relation_info = parentestate->es_result_relation_info;
1917         /* es_trig_target_relations must NOT be copied */
1918         estate->es_rowMarks = parentestate->es_rowMarks;
1919         estate->es_instrument = parentestate->es_instrument;
1920         estate->es_select_into = parentestate->es_select_into;
1921         estate->es_into_oids = parentestate->es_into_oids;
1922
1923         /*
1924          * The external param list is simply shared from parent.  The internal
1925          * param workspace has to be local state, but we copy the initial values
1926          * from the parent, so as to have access to any param values that were
1927          * already set from other parts of the parent's plan tree.
1928          */
1929         estate->es_param_list_info = parentestate->es_param_list_info;
1930         if (parentestate->es_plannedstmt->nParamExec > 0)
1931         {
1932                 int                     i = parentestate->es_plannedstmt->nParamExec;
1933
1934                 estate->es_param_exec_vals = (ParamExecData *)
1935                         palloc0(i * sizeof(ParamExecData));
1936                 while (--i >= 0)
1937                 {
1938                         /* copy value if any, but not execPlan link */
1939                         estate->es_param_exec_vals[i].value =
1940                                 parentestate->es_param_exec_vals[i].value;
1941                         estate->es_param_exec_vals[i].isnull =
1942                                 parentestate->es_param_exec_vals[i].isnull;
1943                 }
1944         }
1945
1946         /*
1947          * Each EState must have its own es_epqScanDone state, but if we have
1948          * nested EPQ checks they should share es_epqTuple arrays.      This allows
1949          * sub-rechecks to inherit the values being examined by an outer recheck.
1950          */
1951         estate->es_epqScanDone = (bool *) palloc0(rtsize * sizeof(bool));
1952         if (parentestate->es_epqTuple != NULL)
1953         {
1954                 estate->es_epqTuple = parentestate->es_epqTuple;
1955                 estate->es_epqTupleSet = parentestate->es_epqTupleSet;
1956         }
1957         else
1958         {
1959                 estate->es_epqTuple = (HeapTuple *)
1960                         palloc0(rtsize * sizeof(HeapTuple));
1961                 estate->es_epqTupleSet = (bool *)
1962                         palloc0(rtsize * sizeof(bool));
1963         }
1964
1965         /*
1966          * Each estate also has its own tuple table.
1967          */
1968         estate->es_tupleTable = NIL;
1969
1970         /*
1971          * Initialize private state information for each SubPlan.  We must do this
1972          * before running ExecInitNode on the main query tree, since
1973          * ExecInitSubPlan expects to be able to find these entries. Some of the
1974          * SubPlans might not be used in the part of the plan tree we intend to
1975          * run, but since it's not easy to tell which, we just initialize them
1976          * all.
1977          */
1978         Assert(estate->es_subplanstates == NIL);
1979         foreach(l, parentestate->es_plannedstmt->subplans)
1980         {
1981                 Plan       *subplan = (Plan *) lfirst(l);
1982                 PlanState  *subplanstate;
1983
1984                 subplanstate = ExecInitNode(subplan, estate, 0);
1985
1986                 estate->es_subplanstates = lappend(estate->es_subplanstates,
1987                                                                                    subplanstate);
1988         }
1989
1990         /*
1991          * Initialize the private state information for all the nodes in the part
1992          * of the plan tree we need to run.  This opens files, allocates storage
1993          * and leaves us ready to start processing tuples.
1994          */
1995         epqstate->planstate = ExecInitNode(planTree, estate, 0);
1996
1997         MemoryContextSwitchTo(oldcontext);
1998 }
1999
2000 /*
2001  * EvalPlanQualEnd -- shut down at termination of parent plan state node,
2002  * or if we are done with the current EPQ child.
2003  *
2004  * This is a cut-down version of ExecutorEnd(); basically we want to do most
2005  * of the normal cleanup, but *not* close result relations (which we are
2006  * just sharing from the outer query).  We do, however, have to close any
2007  * trigger target relations that got opened, since those are not shared.
2008  * (There probably shouldn't be any of the latter, but just in case...)
2009  */
2010 void
2011 EvalPlanQualEnd(EPQState *epqstate)
2012 {
2013         EState     *estate = epqstate->estate;
2014         MemoryContext oldcontext;
2015         ListCell   *l;
2016
2017         if (estate == NULL)
2018                 return;                                 /* idle, so nothing to do */
2019
2020         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
2021
2022         ExecEndNode(epqstate->planstate);
2023
2024         foreach(l, estate->es_subplanstates)
2025         {
2026                 PlanState  *subplanstate = (PlanState *) lfirst(l);
2027
2028                 ExecEndNode(subplanstate);
2029         }
2030
2031         /* throw away the per-estate tuple table */
2032         ExecResetTupleTable(estate->es_tupleTable, false);
2033
2034         /* close any trigger target relations attached to this EState */
2035         foreach(l, estate->es_trig_target_relations)
2036         {
2037                 ResultRelInfo *resultRelInfo = (ResultRelInfo *) lfirst(l);
2038
2039                 /* Close indices and then the relation itself */
2040                 ExecCloseIndices(resultRelInfo);
2041                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
2042         }
2043
2044         MemoryContextSwitchTo(oldcontext);
2045
2046         FreeExecutorState(estate);
2047
2048         /* Mark EPQState idle */
2049         epqstate->estate = NULL;
2050         epqstate->planstate = NULL;
2051         epqstate->origslot = NULL;
2052 }
2053
2054
2055 /*
2056  * Support for SELECT INTO (a/k/a CREATE TABLE AS)
2057  *
2058  * We implement SELECT INTO by diverting SELECT's normal output with
2059  * a specialized DestReceiver type.
2060  */
2061
2062 typedef struct
2063 {
2064         DestReceiver pub;                       /* publicly-known function pointers */
2065         EState     *estate;                     /* EState we are working with */
2066         Relation        rel;                    /* Relation to write to */
2067         int                     hi_options;             /* heap_insert performance options */
2068         BulkInsertState bistate;        /* bulk insert state */
2069 } DR_intorel;
2070
2071 /*
2072  * OpenIntoRel --- actually create the SELECT INTO target relation
2073  *
2074  * This also replaces QueryDesc->dest with the special DestReceiver for
2075  * SELECT INTO.  We assume that the correct result tuple type has already
2076  * been placed in queryDesc->tupDesc.
2077  */
2078 static void
2079 OpenIntoRel(QueryDesc *queryDesc)
2080 {
2081         IntoClause *into = queryDesc->plannedstmt->intoClause;
2082         EState     *estate = queryDesc->estate;
2083         Relation        intoRelationDesc;
2084         char       *intoName;
2085         Oid                     namespaceId;
2086         Oid                     tablespaceId;
2087         Datum           reloptions;
2088         AclResult       aclresult;
2089         Oid                     intoRelationId;
2090         TupleDesc       tupdesc;
2091         DR_intorel *myState;
2092         static char *validnsps[] = HEAP_RELOPT_NAMESPACES;
2093
2094         Assert(into);
2095
2096         /*
2097          * XXX This code needs to be kept in sync with DefineRelation(). Maybe we
2098          * should try to use that function instead.
2099          */
2100
2101         /*
2102          * Check consistency of arguments
2103          */
2104         if (into->onCommit != ONCOMMIT_NOOP && !into->rel->istemp)
2105                 ereport(ERROR,
2106                                 (errcode(ERRCODE_INVALID_TABLE_DEFINITION),
2107                                  errmsg("ON COMMIT can only be used on temporary tables")));
2108
2109         /*
2110          * Security check: disallow creating temp tables from security-restricted
2111          * code.  This is needed because calling code might not expect untrusted
2112          * tables to appear in pg_temp at the front of its search path.
2113          */
2114         if (into->rel->istemp && InSecurityRestrictedOperation())
2115                 ereport(ERROR,
2116                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
2117                                  errmsg("cannot create temporary table within security-restricted operation")));
2118
2119         /*
2120          * Find namespace to create in, check its permissions
2121          */
2122         intoName = into->rel->relname;
2123         namespaceId = RangeVarGetCreationNamespace(into->rel);
2124
2125         aclresult = pg_namespace_aclcheck(namespaceId, GetUserId(),
2126                                                                           ACL_CREATE);
2127         if (aclresult != ACLCHECK_OK)
2128                 aclcheck_error(aclresult, ACL_KIND_NAMESPACE,
2129                                            get_namespace_name(namespaceId));
2130
2131         /*
2132          * Select tablespace to use.  If not specified, use default tablespace
2133          * (which may in turn default to database's default).
2134          */
2135         if (into->tableSpaceName)
2136         {
2137                 tablespaceId = get_tablespace_oid(into->tableSpaceName, false);
2138         }
2139         else
2140         {
2141                 tablespaceId = GetDefaultTablespace(into->rel->istemp);
2142                 /* note InvalidOid is OK in this case */
2143         }
2144
2145         /* Check permissions except when using the database's default space */
2146         if (OidIsValid(tablespaceId) && tablespaceId != MyDatabaseTableSpace)
2147         {
2148                 AclResult       aclresult;
2149
2150                 aclresult = pg_tablespace_aclcheck(tablespaceId, GetUserId(),
2151                                                                                    ACL_CREATE);
2152
2153                 if (aclresult != ACLCHECK_OK)
2154                         aclcheck_error(aclresult, ACL_KIND_TABLESPACE,
2155                                                    get_tablespace_name(tablespaceId));
2156         }
2157
2158         /* Parse and validate any reloptions */
2159         reloptions = transformRelOptions((Datum) 0,
2160                                                                          into->options,
2161                                                                          NULL,
2162                                                                          validnsps,
2163                                                                          true,
2164                                                                          false);
2165         (void) heap_reloptions(RELKIND_RELATION, reloptions, true);
2166
2167         /* Copy the tupdesc because heap_create_with_catalog modifies it */
2168         tupdesc = CreateTupleDescCopy(queryDesc->tupDesc);
2169
2170         /* Now we can actually create the new relation */
2171         intoRelationId = heap_create_with_catalog(intoName,
2172                                                                                           namespaceId,
2173                                                                                           tablespaceId,
2174                                                                                           InvalidOid,
2175                                                                                           InvalidOid,
2176                                                                                           InvalidOid,
2177                                                                                           GetUserId(),
2178                                                                                           tupdesc,
2179                                                                                           NIL,
2180                                                                                           RELKIND_RELATION,
2181                                                                                           false,
2182                                                                                           false,
2183                                                                                           true,
2184                                                                                           0,
2185                                                                                           into->onCommit,
2186                                                                                           reloptions,
2187                                                                                           true,
2188                                                                                           allowSystemTableMods,
2189                                                                                           false);
2190         Assert(intoRelationId != InvalidOid);
2191
2192         FreeTupleDesc(tupdesc);
2193
2194         /*
2195          * Advance command counter so that the newly-created relation's catalog
2196          * tuples will be visible to heap_open.
2197          */
2198         CommandCounterIncrement();
2199
2200         /*
2201          * If necessary, create a TOAST table for the INTO relation. Note that
2202          * AlterTableCreateToastTable ends with CommandCounterIncrement(), so that
2203          * the TOAST table will be visible for insertion.
2204          */
2205         reloptions = transformRelOptions((Datum) 0,
2206                                                                          into->options,
2207                                                                          "toast",
2208                                                                          validnsps,
2209                                                                          true,
2210                                                                          false);
2211
2212         (void) heap_reloptions(RELKIND_TOASTVALUE, reloptions, true);
2213
2214         AlterTableCreateToastTable(intoRelationId, reloptions);
2215
2216         /*
2217          * And open the constructed table for writing.
2218          */
2219         intoRelationDesc = heap_open(intoRelationId, AccessExclusiveLock);
2220
2221         /*
2222          * Now replace the query's DestReceiver with one for SELECT INTO
2223          */
2224         queryDesc->dest = CreateDestReceiver(DestIntoRel);
2225         myState = (DR_intorel *) queryDesc->dest;
2226         Assert(myState->pub.mydest == DestIntoRel);
2227         myState->estate = estate;
2228         myState->rel = intoRelationDesc;
2229
2230         /*
2231          * We can skip WAL-logging the insertions, unless PITR or streaming
2232          * replication is in use. We can skip the FSM in any case.
2233          */
2234         myState->hi_options = HEAP_INSERT_SKIP_FSM |
2235                 (XLogIsNeeded() ? 0 : HEAP_INSERT_SKIP_WAL);
2236         myState->bistate = GetBulkInsertState();
2237
2238         /* Not using WAL requires smgr_targblock be initially invalid */
2239         Assert(RelationGetTargetBlock(intoRelationDesc) == InvalidBlockNumber);
2240 }
2241
2242 /*
2243  * CloseIntoRel --- clean up SELECT INTO at ExecutorEnd time
2244  */
2245 static void
2246 CloseIntoRel(QueryDesc *queryDesc)
2247 {
2248         DR_intorel *myState = (DR_intorel *) queryDesc->dest;
2249
2250         /* OpenIntoRel might never have gotten called */
2251         if (myState && myState->pub.mydest == DestIntoRel && myState->rel)
2252         {
2253                 FreeBulkInsertState(myState->bistate);
2254
2255                 /* If we skipped using WAL, must heap_sync before commit */
2256                 if (myState->hi_options & HEAP_INSERT_SKIP_WAL)
2257                         heap_sync(myState->rel);
2258
2259                 /* close rel, but keep lock until commit */
2260                 heap_close(myState->rel, NoLock);
2261
2262                 myState->rel = NULL;
2263         }
2264 }
2265
2266 /*
2267  * CreateIntoRelDestReceiver -- create a suitable DestReceiver object
2268  */
2269 DestReceiver *
2270 CreateIntoRelDestReceiver(void)
2271 {
2272         DR_intorel *self = (DR_intorel *) palloc0(sizeof(DR_intorel));
2273
2274         self->pub.receiveSlot = intorel_receive;
2275         self->pub.rStartup = intorel_startup;
2276         self->pub.rShutdown = intorel_shutdown;
2277         self->pub.rDestroy = intorel_destroy;
2278         self->pub.mydest = DestIntoRel;
2279
2280         /* private fields will be set by OpenIntoRel */
2281
2282         return (DestReceiver *) self;
2283 }
2284
2285 /*
2286  * intorel_startup --- executor startup
2287  */
2288 static void
2289 intorel_startup(DestReceiver *self, int operation, TupleDesc typeinfo)
2290 {
2291         /* no-op */
2292 }
2293
2294 /*
2295  * intorel_receive --- receive one tuple
2296  */
2297 static void
2298 intorel_receive(TupleTableSlot *slot, DestReceiver *self)
2299 {
2300         DR_intorel *myState = (DR_intorel *) self;
2301         HeapTuple       tuple;
2302
2303         /*
2304          * get the heap tuple out of the tuple table slot, making sure we have a
2305          * writable copy
2306          */
2307         tuple = ExecMaterializeSlot(slot);
2308
2309         /*
2310          * force assignment of new OID (see comments in ExecInsert)
2311          */
2312         if (myState->rel->rd_rel->relhasoids)
2313                 HeapTupleSetOid(tuple, InvalidOid);
2314
2315         heap_insert(myState->rel,
2316                                 tuple,
2317                                 myState->estate->es_output_cid,
2318                                 myState->hi_options,
2319                                 myState->bistate);
2320
2321         /* We know this is a newly created relation, so there are no indexes */
2322 }
2323
2324 /*
2325  * intorel_shutdown --- executor end
2326  */
2327 static void
2328 intorel_shutdown(DestReceiver *self)
2329 {
2330         /* no-op */
2331 }
2332
2333 /*
2334  * intorel_destroy --- release DestReceiver object
2335  */
2336 static void
2337 intorel_destroy(DestReceiver *self)
2338 {
2339         pfree(self);
2340 }