granicus.if.org Git - postgresql/blob - src/backend/executor/execMain.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * execMain.c
   4  *        top level executor interface routines
   5  *
   6  * INTERFACE ROUTINES
   7  *      ExecutorStart()
   8  *      ExecutorRun()
   9  *      ExecutorFinish()
  10  *      ExecutorEnd()
  11  *
  12  *      These four procedures are the external interface to the executor.
  13  *      In each case, the query descriptor is required as an argument.
  14  *
  15  *      ExecutorStart must be called at the beginning of execution of any
  16  *      query plan and ExecutorEnd must always be called at the end of
  17  *      execution of a plan (unless it is aborted due to error).
  18  *
  19  *      ExecutorRun accepts direction and count arguments that specify whether
  20  *      the plan is to be executed forwards, backwards, and for how many tuples.
  21  *      In some cases ExecutorRun may be called multiple times to process all
  22  *      the tuples for a plan.  It is also acceptable to stop short of executing
  23  *      the whole plan (but only if it is a SELECT).
  24  *
  25  *      ExecutorFinish must be called after the final ExecutorRun call and
  26  *      before ExecutorEnd.  This can be omitted only in case of EXPLAIN,
  27  *      which should also omit ExecutorRun.
  28  *
  29  * Portions Copyright (c) 1996-2018, PostgreSQL Global Development Group
  30  * Portions Copyright (c) 1994, Regents of the University of California
  31  *
  32  *
  33  * IDENTIFICATION
  34  *        src/backend/executor/execMain.c
  35  *
  36  *-------------------------------------------------------------------------
  37  */
  38 #include "postgres.h"
  39
  40 #include "access/htup_details.h"
  41 #include "access/sysattr.h"
  42 #include "access/transam.h"
  43 #include "access/xact.h"
  44 #include "catalog/namespace.h"
  45 #include "catalog/partition.h"
  46 #include "catalog/pg_publication.h"
  47 #include "commands/matview.h"
  48 #include "commands/trigger.h"
  49 #include "executor/execdebug.h"
  50 #include "foreign/fdwapi.h"
  51 #include "jit/jit.h"
  52 #include "mb/pg_wchar.h"
  53 #include "miscadmin.h"
  54 #include "optimizer/clauses.h"
  55 #include "parser/parsetree.h"
  56 #include "rewrite/rewriteManip.h"
  57 #include "storage/bufmgr.h"
  58 #include "storage/lmgr.h"
  59 #include "tcop/utility.h"
  60 #include "utils/acl.h"
  61 #include "utils/lsyscache.h"
  62 #include "utils/memutils.h"
  63 #include "utils/rls.h"
  64 #include "utils/ruleutils.h"
  65 #include "utils/snapmgr.h"
  66 #include "utils/tqual.h"
  67
  68
  69 /* Hooks for plugins to get control in ExecutorStart/Run/Finish/End */
  70 ExecutorStart_hook_type ExecutorStart_hook = NULL;
  71 ExecutorRun_hook_type ExecutorRun_hook = NULL;
  72 ExecutorFinish_hook_type ExecutorFinish_hook = NULL;
  73 ExecutorEnd_hook_type ExecutorEnd_hook = NULL;
  74
  75 /* Hook for plugin to get control in ExecCheckRTPerms() */
  76 ExecutorCheckPerms_hook_type ExecutorCheckPerms_hook = NULL;
  77
  78 /* decls for local routines only used within this module */
  79 static void InitPlan(QueryDesc *queryDesc, int eflags);
  80 static void CheckValidRowMarkRel(Relation rel, RowMarkType markType);
  81 static void ExecPostprocessPlan(EState *estate);
  82 static void ExecEndPlan(PlanState *planstate, EState *estate);
  83 static void ExecutePlan(EState *estate, PlanState *planstate,
  84                         bool use_parallel_mode,
  85                         CmdType operation,
  86                         bool sendTuples,
  87                         uint64 numberTuples,
  88                         ScanDirection direction,
  89                         DestReceiver *dest,
  90                         bool execute_once);
  91 static bool ExecCheckRTEPerms(RangeTblEntry *rte);
  92 static bool ExecCheckRTEPermsModified(Oid relOid, Oid userid,
  93                                                   Bitmapset *modifiedCols,
  94                                                   AclMode requiredPerms);
  95 static void ExecCheckXactReadOnly(PlannedStmt *plannedstmt);
  96 static char *ExecBuildSlotValueDescription(Oid reloid,
  97                                                           TupleTableSlot *slot,
  98                                                           TupleDesc tupdesc,
  99                                                           Bitmapset *modifiedCols,
 100                                                           int maxfieldlen);
 101 static void EvalPlanQualStart(EPQState *epqstate, EState *parentestate,
 102                                   Plan *planTree);
 103
 104 /*
 105  * Note that GetUpdatedColumns() also exists in commands/trigger.c.  There does
 106  * not appear to be any good header to put it into, given the structures that
 107  * it uses, so we let them be duplicated.  Be sure to update both if one needs
 108  * to be changed, however.
 109  */
 110 #define GetInsertedColumns(relinfo, estate) \
 111         (rt_fetch((relinfo)->ri_RangeTableIndex, (estate)->es_range_table)->insertedCols)
 112 #define GetUpdatedColumns(relinfo, estate) \
 113         (rt_fetch((relinfo)->ri_RangeTableIndex, (estate)->es_range_table)->updatedCols)
 114
 115 /* end of local decls */
 116
 117
 118 /* ----------------------------------------------------------------
 119  *              ExecutorStart
 120  *
 121  *              This routine must be called at the beginning of any execution of any
 122  *              query plan
 123  *
 124  * Takes a QueryDesc previously created by CreateQueryDesc (which is separate
 125  * only because some places use QueryDescs for utility commands).  The tupDesc
 126  * field of the QueryDesc is filled in to describe the tuples that will be
 127  * returned, and the internal fields (estate and planstate) are set up.
 128  *
 129  * eflags contains flag bits as described in executor.h.
 130  *
 131  * NB: the CurrentMemoryContext when this is called will become the parent
 132  * of the per-query context used for this Executor invocation.
 133  *
 134  * We provide a function hook variable that lets loadable plugins
 135  * get control when ExecutorStart is called.  Such a plugin would
 136  * normally call standard_ExecutorStart().
 137  *
 138  * ----------------------------------------------------------------
 139  */
 140 void
 141 ExecutorStart(QueryDesc *queryDesc, int eflags)
 142 {
 143         if (ExecutorStart_hook)
 144                 (*ExecutorStart_hook) (queryDesc, eflags);
 145         else
 146                 standard_ExecutorStart(queryDesc, eflags);
 147 }
 148
 149 void
 150 standard_ExecutorStart(QueryDesc *queryDesc, int eflags)
 151 {
 152         EState     *estate;
 153         MemoryContext oldcontext;
 154
 155         /* sanity checks: queryDesc must not be started already */
 156         Assert(queryDesc != NULL);
 157         Assert(queryDesc->estate == NULL);
 158
 159         /*
 160          * If the transaction is read-only, we need to check if any writes are
 161          * planned to non-temporary tables.  EXPLAIN is considered read-only.
 162          *
 163          * Don't allow writes in parallel mode.  Supporting UPDATE and DELETE
 164          * would require (a) storing the combocid hash in shared memory, rather
 165          * than synchronizing it just once at the start of parallelism, and (b) an
 166          * alternative to heap_update()'s reliance on xmax for mutual exclusion.
 167          * INSERT may have no such troubles, but we forbid it to simplify the
 168          * checks.
 169          *
 170          * We have lower-level defenses in CommandCounterIncrement and elsewhere
 171          * against performing unsafe operations in parallel mode, but this gives a
 172          * more user-friendly error message.
 173          */
 174         if ((XactReadOnly || IsInParallelMode()) &&
 175                 !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
 176                 ExecCheckXactReadOnly(queryDesc->plannedstmt);
 177
 178         /*
 179          * Build EState, switch into per-query memory context for startup.
 180          */
 181         estate = CreateExecutorState();
 182         queryDesc->estate = estate;
 183
 184         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 185
 186         /*
 187          * Fill in external parameters, if any, from queryDesc; and allocate
 188          * workspace for internal parameters
 189          */
 190         estate->es_param_list_info = queryDesc->params;
 191
 192         if (queryDesc->plannedstmt->paramExecTypes != NIL)
 193         {
 194                 int                     nParamExec;
 195
 196                 nParamExec = list_length(queryDesc->plannedstmt->paramExecTypes);
 197                 estate->es_param_exec_vals = (ParamExecData *)
 198                         palloc0(nParamExec * sizeof(ParamExecData));
 199         }
 200
 201         estate->es_sourceText = queryDesc->sourceText;
 202
 203         /*
 204          * Fill in the query environment, if any, from queryDesc.
 205          */
 206         estate->es_queryEnv = queryDesc->queryEnv;
 207
 208         /*
 209          * If non-read-only query, set the command ID to mark output tuples with
 210          */
 211         switch (queryDesc->operation)
 212         {
 213                 case CMD_SELECT:
 214
 215                         /*
 216                          * SELECT FOR [KEY] UPDATE/SHARE and modifying CTEs need to mark
 217                          * tuples
 218                          */
 219                         if (queryDesc->plannedstmt->rowMarks != NIL ||
 220                                 queryDesc->plannedstmt->hasModifyingCTE)
 221                                 estate->es_output_cid = GetCurrentCommandId(true);
 222
 223                         /*
 224                          * A SELECT without modifying CTEs can't possibly queue triggers,
 225                          * so force skip-triggers mode. This is just a marginal efficiency
 226                          * hack, since AfterTriggerBeginQuery/AfterTriggerEndQuery aren't
 227                          * all that expensive, but we might as well do it.
 228                          */
 229                         if (!queryDesc->plannedstmt->hasModifyingCTE)
 230                                 eflags |= EXEC_FLAG_SKIP_TRIGGERS;
 231                         break;
 232
 233                 case CMD_INSERT:
 234                 case CMD_DELETE:
 235                 case CMD_UPDATE:
 236                         estate->es_output_cid = GetCurrentCommandId(true);
 237                         break;
 238
 239                 default:
 240                         elog(ERROR, "unrecognized operation code: %d",
 241                                  (int) queryDesc->operation);
 242                         break;
 243         }
 244
 245         /*
 246          * Copy other important information into the EState
 247          */
 248         estate->es_snapshot = RegisterSnapshot(queryDesc->snapshot);
 249         estate->es_crosscheck_snapshot = RegisterSnapshot(queryDesc->crosscheck_snapshot);
 250         estate->es_top_eflags = eflags;
 251         estate->es_instrument = queryDesc->instrument_options;
 252         estate->es_jit_flags = queryDesc->plannedstmt->jitFlags;
 253
 254         /*
 255          * Set up an AFTER-trigger statement context, unless told not to, or
 256          * unless it's EXPLAIN-only mode (when ExecutorFinish won't be called).
 257          */
 258         if (!(eflags & (EXEC_FLAG_SKIP_TRIGGERS | EXEC_FLAG_EXPLAIN_ONLY)))
 259                 AfterTriggerBeginQuery();
 260
 261         /*
 262          * Initialize the plan state tree
 263          */
 264         InitPlan(queryDesc, eflags);
 265
 266         MemoryContextSwitchTo(oldcontext);
 267 }
 268
 269 /* ----------------------------------------------------------------
 270  *              ExecutorRun
 271  *
 272  *              This is the main routine of the executor module. It accepts
 273  *              the query descriptor from the traffic cop and executes the
 274  *              query plan.
 275  *
 276  *              ExecutorStart must have been called already.
 277  *
 278  *              If direction is NoMovementScanDirection then nothing is done
 279  *              except to start up/shut down the destination.  Otherwise,
 280  *              we retrieve up to 'count' tuples in the specified direction.
 281  *
 282  *              Note: count = 0 is interpreted as no portal limit, i.e., run to
 283  *              completion.  Also note that the count limit is only applied to
 284  *              retrieved tuples, not for instance to those inserted/updated/deleted
 285  *              by a ModifyTable plan node.
 286  *
 287  *              There is no return value, but output tuples (if any) are sent to
 288  *              the destination receiver specified in the QueryDesc; and the number
 289  *              of tuples processed at the top level can be found in
 290  *              estate->es_processed.
 291  *
 292  *              We provide a function hook variable that lets loadable plugins
 293  *              get control when ExecutorRun is called.  Such a plugin would
 294  *              normally call standard_ExecutorRun().
 295  *
 296  * ----------------------------------------------------------------
 297  */
 298 void
 299 ExecutorRun(QueryDesc *queryDesc,
 300                         ScanDirection direction, uint64 count,
 301                         bool execute_once)
 302 {
 303         if (ExecutorRun_hook)
 304                 (*ExecutorRun_hook) (queryDesc, direction, count, execute_once);
 305         else
 306                 standard_ExecutorRun(queryDesc, direction, count, execute_once);
 307 }
 308
 309 void
 310 standard_ExecutorRun(QueryDesc *queryDesc,
 311                                          ScanDirection direction, uint64 count, bool execute_once)
 312 {
 313         EState     *estate;
 314         CmdType         operation;
 315         DestReceiver *dest;
 316         bool            sendTuples;
 317         MemoryContext oldcontext;
 318
 319         /* sanity checks */
 320         Assert(queryDesc != NULL);
 321
 322         estate = queryDesc->estate;
 323
 324         Assert(estate != NULL);
 325         Assert(!(estate->es_top_eflags & EXEC_FLAG_EXPLAIN_ONLY));
 326
 327         /*
 328          * Switch into per-query memory context
 329          */
 330         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 331
 332         /* Allow instrumentation of Executor overall runtime */
 333         if (queryDesc->totaltime)
 334                 InstrStartNode(queryDesc->totaltime);
 335
 336         /*
 337          * extract information from the query descriptor and the query feature.
 338          */
 339         operation = queryDesc->operation;
 340         dest = queryDesc->dest;
 341
 342         /*
 343          * startup tuple receiver, if we will be emitting tuples
 344          */
 345         estate->es_processed = 0;
 346         estate->es_lastoid = InvalidOid;
 347
 348         sendTuples = (operation == CMD_SELECT ||
 349                                   queryDesc->plannedstmt->hasReturning);
 350
 351         if (sendTuples)
 352                 dest->rStartup(dest, operation, queryDesc->tupDesc);
 353
 354         /*
 355          * run plan
 356          */
 357         if (!ScanDirectionIsNoMovement(direction))
 358         {
 359                 if (execute_once && queryDesc->already_executed)
 360                         elog(ERROR, "can't re-execute query flagged for single execution");
 361                 queryDesc->already_executed = true;
 362
 363                 ExecutePlan(estate,
 364                                         queryDesc->planstate,
 365                                         queryDesc->plannedstmt->parallelModeNeeded,
 366                                         operation,
 367                                         sendTuples,
 368                                         count,
 369                                         direction,
 370                                         dest,
 371                                         execute_once);
 372         }
 373
 374         /*
 375          * shutdown tuple receiver, if we started it
 376          */
 377         if (sendTuples)
 378                 dest->rShutdown(dest);
 379
 380         if (queryDesc->totaltime)
 381                 InstrStopNode(queryDesc->totaltime, estate->es_processed);
 382
 383         MemoryContextSwitchTo(oldcontext);
 384 }
 385
 386 /* ----------------------------------------------------------------
 387  *              ExecutorFinish
 388  *
 389  *              This routine must be called after the last ExecutorRun call.
 390  *              It performs cleanup such as firing AFTER triggers.  It is
 391  *              separate from ExecutorEnd because EXPLAIN ANALYZE needs to
 392  *              include these actions in the total runtime.
 393  *
 394  *              We provide a function hook variable that lets loadable plugins
 395  *              get control when ExecutorFinish is called.  Such a plugin would
 396  *              normally call standard_ExecutorFinish().
 397  *
 398  * ----------------------------------------------------------------
 399  */
 400 void
 401 ExecutorFinish(QueryDesc *queryDesc)
 402 {
 403         if (ExecutorFinish_hook)
 404                 (*ExecutorFinish_hook) (queryDesc);
 405         else
 406                 standard_ExecutorFinish(queryDesc);
 407 }
 408
 409 void
 410 standard_ExecutorFinish(QueryDesc *queryDesc)
 411 {
 412         EState     *estate;
 413         MemoryContext oldcontext;
 414
 415         /* sanity checks */
 416         Assert(queryDesc != NULL);
 417
 418         estate = queryDesc->estate;
 419
 420         Assert(estate != NULL);
 421         Assert(!(estate->es_top_eflags & EXEC_FLAG_EXPLAIN_ONLY));
 422
 423         /* This should be run once and only once per Executor instance */
 424         Assert(!estate->es_finished);
 425
 426         /* Switch into per-query memory context */
 427         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 428
 429         /* Allow instrumentation of Executor overall runtime */
 430         if (queryDesc->totaltime)
 431                 InstrStartNode(queryDesc->totaltime);
 432
 433         /* Run ModifyTable nodes to completion */
 434         ExecPostprocessPlan(estate);
 435
 436         /* Execute queued AFTER triggers, unless told not to */
 437         if (!(estate->es_top_eflags & EXEC_FLAG_SKIP_TRIGGERS))
 438                 AfterTriggerEndQuery(estate);
 439
 440         if (queryDesc->totaltime)
 441                 InstrStopNode(queryDesc->totaltime, 0);
 442
 443         MemoryContextSwitchTo(oldcontext);
 444
 445         estate->es_finished = true;
 446 }
 447
 448 /* ----------------------------------------------------------------
 449  *              ExecutorEnd
 450  *
 451  *              This routine must be called at the end of execution of any
 452  *              query plan
 453  *
 454  *              We provide a function hook variable that lets loadable plugins
 455  *              get control when ExecutorEnd is called.  Such a plugin would
 456  *              normally call standard_ExecutorEnd().
 457  *
 458  * ----------------------------------------------------------------
 459  */
 460 void
 461 ExecutorEnd(QueryDesc *queryDesc)
 462 {
 463         if (ExecutorEnd_hook)
 464                 (*ExecutorEnd_hook) (queryDesc);
 465         else
 466                 standard_ExecutorEnd(queryDesc);
 467 }
 468
 469 void
 470 standard_ExecutorEnd(QueryDesc *queryDesc)
 471 {
 472         EState     *estate;
 473         MemoryContext oldcontext;
 474
 475         /* sanity checks */
 476         Assert(queryDesc != NULL);
 477
 478         estate = queryDesc->estate;
 479
 480         Assert(estate != NULL);
 481
 482         /*
 483          * Check that ExecutorFinish was called, unless in EXPLAIN-only mode. This
 484          * Assert is needed because ExecutorFinish is new as of 9.1, and callers
 485          * might forget to call it.
 486          */
 487         Assert(estate->es_finished ||
 488                    (estate->es_top_eflags & EXEC_FLAG_EXPLAIN_ONLY));
 489
 490         /*
 491          * Switch into per-query memory context to run ExecEndPlan
 492          */
 493         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 494
 495         ExecEndPlan(queryDesc->planstate, estate);
 496
 497         /* do away with our snapshots */
 498         UnregisterSnapshot(estate->es_snapshot);
 499         UnregisterSnapshot(estate->es_crosscheck_snapshot);
 500
 501         /* release JIT context, if allocated */
 502         if (estate->es_jit)
 503                 jit_release_context(estate->es_jit);
 504
 505         /*
 506          * Must switch out of context before destroying it
 507          */
 508         MemoryContextSwitchTo(oldcontext);
 509
 510         /*
 511          * Release EState and per-query memory context.  This should release
 512          * everything the executor has allocated.
 513          */
 514         FreeExecutorState(estate);
 515
 516         /* Reset queryDesc fields that no longer point to anything */
 517         queryDesc->tupDesc = NULL;
 518         queryDesc->estate = NULL;
 519         queryDesc->planstate = NULL;
 520         queryDesc->totaltime = NULL;
 521 }
 522
 523 /* ----------------------------------------------------------------
 524  *              ExecutorRewind
 525  *
 526  *              This routine may be called on an open queryDesc to rewind it
 527  *              to the start.
 528  * ----------------------------------------------------------------
 529  */
 530 void
 531 ExecutorRewind(QueryDesc *queryDesc)
 532 {
 533         EState     *estate;
 534         MemoryContext oldcontext;
 535
 536         /* sanity checks */
 537         Assert(queryDesc != NULL);
 538
 539         estate = queryDesc->estate;
 540
 541         Assert(estate != NULL);
 542
 543         /* It's probably not sensible to rescan updating queries */
 544         Assert(queryDesc->operation == CMD_SELECT);
 545
 546         /*
 547          * Switch into per-query memory context
 548          */
 549         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 550
 551         /*
 552          * rescan plan
 553          */
 554         ExecReScan(queryDesc->planstate);
 555
 556         MemoryContextSwitchTo(oldcontext);
 557 }
 558
 559
 560 /*
 561  * ExecCheckRTPerms
 562  *              Check access permissions for all relations listed in a range table.
 563  *
 564  * Returns true if permissions are adequate.  Otherwise, throws an appropriate
 565  * error if ereport_on_violation is true, or simply returns false otherwise.
 566  *
 567  * Note that this does NOT address row level security policies (aka: RLS).  If
 568  * rows will be returned to the user as a result of this permission check
 569  * passing, then RLS also needs to be consulted (and check_enable_rls()).
 570  *
 571  * See rewrite/rowsecurity.c.
 572  */
 573 bool
 574 ExecCheckRTPerms(List *rangeTable, bool ereport_on_violation)
 575 {
 576         ListCell   *l;
 577         bool            result = true;
 578
 579         foreach(l, rangeTable)
 580         {
 581                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 582
 583                 result = ExecCheckRTEPerms(rte);
 584                 if (!result)
 585                 {
 586                         Assert(rte->rtekind == RTE_RELATION);
 587                         if (ereport_on_violation)
 588                                 aclcheck_error(ACLCHECK_NO_PRIV, get_relkind_objtype(get_rel_relkind(rte->relid)),
 589                                                            get_rel_name(rte->relid));
 590                         return false;
 591                 }
 592         }
 593
 594         if (ExecutorCheckPerms_hook)
 595                 result = (*ExecutorCheckPerms_hook) (rangeTable,
 596                                                                                          ereport_on_violation);
 597         return result;
 598 }
 599
 600 /*
 601  * ExecCheckRTEPerms
 602  *              Check access permissions for a single RTE.
 603  */
 604 static bool
 605 ExecCheckRTEPerms(RangeTblEntry *rte)
 606 {
 607         AclMode         requiredPerms;
 608         AclMode         relPerms;
 609         AclMode         remainingPerms;
 610         Oid                     relOid;
 611         Oid                     userid;
 612
 613         /*
 614          * Only plain-relation RTEs need to be checked here.  Function RTEs are
 615          * checked when the function is prepared for execution.  Join, subquery,
 616          * and special RTEs need no checks.
 617          */
 618         if (rte->rtekind != RTE_RELATION)
 619                 return true;
 620
 621         /*
 622          * No work if requiredPerms is empty.
 623          */
 624         requiredPerms = rte->requiredPerms;
 625         if (requiredPerms == 0)
 626                 return true;
 627
 628         relOid = rte->relid;
 629
 630         /*
 631          * userid to check as: current user unless we have a setuid indication.
 632          *
 633          * Note: GetUserId() is presently fast enough that there's no harm in
 634          * calling it separately for each RTE.  If that stops being true, we could
 635          * call it once in ExecCheckRTPerms and pass the userid down from there.
 636          * But for now, no need for the extra clutter.
 637          */
 638         userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
 639
 640         /*
 641          * We must have *all* the requiredPerms bits, but some of the bits can be
 642          * satisfied from column-level rather than relation-level permissions.
 643          * First, remove any bits that are satisfied by relation permissions.
 644          */
 645         relPerms = pg_class_aclmask(relOid, userid, requiredPerms, ACLMASK_ALL);
 646         remainingPerms = requiredPerms & ~relPerms;
 647         if (remainingPerms != 0)
 648         {
 649                 int                     col = -1;
 650
 651                 /*
 652                  * If we lack any permissions that exist only as relation permissions,
 653                  * we can fail straight away.
 654                  */
 655                 if (remainingPerms & ~(ACL_SELECT | ACL_INSERT | ACL_UPDATE))
 656                         return false;
 657
 658                 /*
 659                  * Check to see if we have the needed privileges at column level.
 660                  *
 661                  * Note: failures just report a table-level error; it would be nicer
 662                  * to report a column-level error if we have some but not all of the
 663                  * column privileges.
 664                  */
 665                 if (remainingPerms & ACL_SELECT)
 666                 {
 667                         /*
 668                          * When the query doesn't explicitly reference any columns (for
 669                          * example, SELECT COUNT(*) FROM table), allow the query if we
 670                          * have SELECT on any column of the rel, as per SQL spec.
 671                          */
 672                         if (bms_is_empty(rte->selectedCols))
 673                         {
 674                                 if (pg_attribute_aclcheck_all(relOid, userid, ACL_SELECT,
 675                                                                                           ACLMASK_ANY) != ACLCHECK_OK)
 676                                         return false;
 677                         }
 678
 679                         while ((col = bms_next_member(rte->selectedCols, col)) >= 0)
 680                         {
 681                                 /* bit #s are offset by FirstLowInvalidHeapAttributeNumber */
 682                                 AttrNumber      attno = col + FirstLowInvalidHeapAttributeNumber;
 683
 684                                 if (attno == InvalidAttrNumber)
 685                                 {
 686                                         /* Whole-row reference, must have priv on all cols */
 687                                         if (pg_attribute_aclcheck_all(relOid, userid, ACL_SELECT,
 688                                                                                                   ACLMASK_ALL) != ACLCHECK_OK)
 689                                                 return false;
 690                                 }
 691                                 else
 692                                 {
 693                                         if (pg_attribute_aclcheck(relOid, attno, userid,
 694                                                                                           ACL_SELECT) != ACLCHECK_OK)
 695                                                 return false;
 696                                 }
 697                         }
 698                 }
 699
 700                 /*
 701                  * Basically the same for the mod columns, for both INSERT and UPDATE
 702                  * privilege as specified by remainingPerms.
 703                  */
 704                 if (remainingPerms & ACL_INSERT && !ExecCheckRTEPermsModified(relOid,
 705                                                                                                                                           userid,
 706                                                                                                                                           rte->insertedCols,
 707                                                                                                                                           ACL_INSERT))
 708                         return false;
 709
 710                 if (remainingPerms & ACL_UPDATE && !ExecCheckRTEPermsModified(relOid,
 711                                                                                                                                           userid,
 712                                                                                                                                           rte->updatedCols,
 713                                                                                                                                           ACL_UPDATE))
 714                         return false;
 715         }
 716         return true;
 717 }
 718
 719 /*
 720  * ExecCheckRTEPermsModified
 721  *              Check INSERT or UPDATE access permissions for a single RTE (these
 722  *              are processed uniformly).
 723  */
 724 static bool
 725 ExecCheckRTEPermsModified(Oid relOid, Oid userid, Bitmapset *modifiedCols,
 726                                                   AclMode requiredPerms)
 727 {
 728         int                     col = -1;
 729
 730         /*
 731          * When the query doesn't explicitly update any columns, allow the query
 732          * if we have permission on any column of the rel.  This is to handle
 733          * SELECT FOR UPDATE as well as possible corner cases in UPDATE.
 734          */
 735         if (bms_is_empty(modifiedCols))
 736         {
 737                 if (pg_attribute_aclcheck_all(relOid, userid, requiredPerms,
 738                                                                           ACLMASK_ANY) != ACLCHECK_OK)
 739                         return false;
 740         }
 741
 742         while ((col = bms_next_member(modifiedCols, col)) >= 0)
 743         {
 744                 /* bit #s are offset by FirstLowInvalidHeapAttributeNumber */
 745                 AttrNumber      attno = col + FirstLowInvalidHeapAttributeNumber;
 746
 747                 if (attno == InvalidAttrNumber)
 748                 {
 749                         /* whole-row reference can't happen here */
 750                         elog(ERROR, "whole-row update is not implemented");
 751                 }
 752                 else
 753                 {
 754                         if (pg_attribute_aclcheck(relOid, attno, userid,
 755                                                                           requiredPerms) != ACLCHECK_OK)
 756                                 return false;
 757                 }
 758         }
 759         return true;
 760 }
 761
 762 /*
 763  * Check that the query does not imply any writes to non-temp tables;
 764  * unless we're in parallel mode, in which case don't even allow writes
 765  * to temp tables.
 766  *
 767  * Note: in a Hot Standby this would need to reject writes to temp
 768  * tables just as we do in parallel mode; but an HS standby can't have created
 769  * any temp tables in the first place, so no need to check that.
 770  */
 771 static void
 772 ExecCheckXactReadOnly(PlannedStmt *plannedstmt)
 773 {
 774         ListCell   *l;
 775
 776         /*
 777          * Fail if write permissions are requested in parallel mode for table
 778          * (temp or non-temp), otherwise fail for any non-temp table.
 779          */
 780         foreach(l, plannedstmt->rtable)
 781         {
 782                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 783
 784                 if (rte->rtekind != RTE_RELATION)
 785                         continue;
 786
 787                 if ((rte->requiredPerms & (~ACL_SELECT)) == 0)
 788                         continue;
 789
 790                 if (isTempNamespace(get_rel_namespace(rte->relid)))
 791                         continue;
 792
 793                 PreventCommandIfReadOnly(CreateCommandTag((Node *) plannedstmt));
 794         }
 795
 796         if (plannedstmt->commandType != CMD_SELECT || plannedstmt->hasModifyingCTE)
 797                 PreventCommandIfParallelMode(CreateCommandTag((Node *) plannedstmt));
 798 }
 799
 800
 801 /* ----------------------------------------------------------------
 802  *              InitPlan
 803  *
 804  *              Initializes the query plan: open files, allocate storage
 805  *              and start up the rule manager
 806  * ----------------------------------------------------------------
 807  */
 808 static void
 809 InitPlan(QueryDesc *queryDesc, int eflags)
 810 {
 811         CmdType         operation = queryDesc->operation;
 812         PlannedStmt *plannedstmt = queryDesc->plannedstmt;
 813         Plan       *plan = plannedstmt->planTree;
 814         List       *rangeTable = plannedstmt->rtable;
 815         EState     *estate = queryDesc->estate;
 816         PlanState  *planstate;
 817         TupleDesc       tupType;
 818         ListCell   *l;
 819         int                     i;
 820
 821         /*
 822          * Do permissions checks
 823          */
 824         ExecCheckRTPerms(rangeTable, true);
 825
 826         /*
 827          * initialize the node's execution state
 828          */
 829         estate->es_range_table = rangeTable;
 830         estate->es_plannedstmt = plannedstmt;
 831
 832         /*
 833          * initialize result relation stuff, and open/lock the result rels.
 834          *
 835          * We must do this before initializing the plan tree, else we might try to
 836          * do a lock upgrade if a result rel is also a source rel.
 837          */
 838         if (plannedstmt->resultRelations)
 839         {
 840                 List       *resultRelations = plannedstmt->resultRelations;
 841                 int                     numResultRelations = list_length(resultRelations);
 842                 ResultRelInfo *resultRelInfos;
 843                 ResultRelInfo *resultRelInfo;
 844
 845                 resultRelInfos = (ResultRelInfo *)
 846                         palloc(numResultRelations * sizeof(ResultRelInfo));
 847                 resultRelInfo = resultRelInfos;
 848                 foreach(l, resultRelations)
 849                 {
 850                         Index           resultRelationIndex = lfirst_int(l);
 851                         Oid                     resultRelationOid;
 852                         Relation        resultRelation;
 853
 854                         resultRelationOid = getrelid(resultRelationIndex, rangeTable);
 855                         resultRelation = heap_open(resultRelationOid, RowExclusiveLock);
 856
 857                         InitResultRelInfo(resultRelInfo,
 858                                                           resultRelation,
 859                                                           resultRelationIndex,
 860                                                           NULL,
 861                                                           estate->es_instrument);
 862                         resultRelInfo++;
 863                 }
 864                 estate->es_result_relations = resultRelInfos;
 865                 estate->es_num_result_relations = numResultRelations;
 866                 /* es_result_relation_info is NULL except when within ModifyTable */
 867                 estate->es_result_relation_info = NULL;
 868
 869                 /*
 870                  * In the partitioned result relation case, lock the non-leaf result
 871                  * relations too.  A subset of these are the roots of respective
 872                  * partitioned tables, for which we also allocate ResulRelInfos.
 873                  */
 874                 estate->es_root_result_relations = NULL;
 875                 estate->es_num_root_result_relations = 0;
 876                 if (plannedstmt->nonleafResultRelations)
 877                 {
 878                         int                     num_roots = list_length(plannedstmt->rootResultRelations);
 879
 880                         /*
 881                          * Firstly, build ResultRelInfos for all the partitioned table
 882                          * roots, because we will need them to fire the statement-level
 883                          * triggers, if any.
 884                          */
 885                         resultRelInfos = (ResultRelInfo *)
 886                                 palloc(num_roots * sizeof(ResultRelInfo));
 887                         resultRelInfo = resultRelInfos;
 888                         foreach(l, plannedstmt->rootResultRelations)
 889                         {
 890                                 Index           resultRelIndex = lfirst_int(l);
 891                                 Oid                     resultRelOid;
 892                                 Relation        resultRelDesc;
 893
 894                                 resultRelOid = getrelid(resultRelIndex, rangeTable);
 895                                 resultRelDesc = heap_open(resultRelOid, RowExclusiveLock);
 896                                 InitResultRelInfo(resultRelInfo,
 897                                                                   resultRelDesc,
 898                                                                   lfirst_int(l),
 899                                                                   NULL,
 900                                                                   estate->es_instrument);
 901                                 resultRelInfo++;
 902                         }
 903
 904                         estate->es_root_result_relations = resultRelInfos;
 905                         estate->es_num_root_result_relations = num_roots;
 906
 907                         /* Simply lock the rest of them. */
 908                         foreach(l, plannedstmt->nonleafResultRelations)
 909                         {
 910                                 Index           resultRelIndex = lfirst_int(l);
 911
 912                                 /* We locked the roots above. */
 913                                 if (!list_member_int(plannedstmt->rootResultRelations,
 914                                                                          resultRelIndex))
 915                                         LockRelationOid(getrelid(resultRelIndex, rangeTable),
 916                                                                         RowExclusiveLock);
 917                         }
 918                 }
 919         }
 920         else
 921         {
 922                 /*
 923                  * if no result relation, then set state appropriately
 924                  */
 925                 estate->es_result_relations = NULL;
 926                 estate->es_num_result_relations = 0;
 927                 estate->es_result_relation_info = NULL;
 928                 estate->es_root_result_relations = NULL;
 929                 estate->es_num_root_result_relations = 0;
 930         }
 931
 932         /*
 933          * Similarly, we have to lock relations selected FOR [KEY] UPDATE/SHARE
 934          * before we initialize the plan tree, else we'd be risking lock upgrades.
 935          * While we are at it, build the ExecRowMark list.  Any partitioned child
 936          * tables are ignored here (because isParent=true) and will be locked by
 937          * the first Append or MergeAppend node that references them.  (Note that
 938          * the RowMarks corresponding to partitioned child tables are present in
 939          * the same list as the rest, i.e., plannedstmt->rowMarks.)
 940          */
 941         estate->es_rowMarks = NIL;
 942         foreach(l, plannedstmt->rowMarks)
 943         {
 944                 PlanRowMark *rc = (PlanRowMark *) lfirst(l);
 945                 Oid                     relid;
 946                 Relation        relation;
 947                 ExecRowMark *erm;
 948
 949                 /* ignore "parent" rowmarks; they are irrelevant at runtime */
 950                 if (rc->isParent)
 951                         continue;
 952
 953                 /* get relation's OID (will produce InvalidOid if subquery) */
 954                 relid = getrelid(rc->rti, rangeTable);
 955
 956                 /*
 957                  * If you change the conditions under which rel locks are acquired
 958                  * here, be sure to adjust ExecOpenScanRelation to match.
 959                  */
 960                 switch (rc->markType)
 961                 {
 962                         case ROW_MARK_EXCLUSIVE:
 963                         case ROW_MARK_NOKEYEXCLUSIVE:
 964                         case ROW_MARK_SHARE:
 965                         case ROW_MARK_KEYSHARE:
 966                                 relation = heap_open(relid, RowShareLock);
 967                                 break;
 968                         case ROW_MARK_REFERENCE:
 969                                 relation = heap_open(relid, AccessShareLock);
 970                                 break;
 971                         case ROW_MARK_COPY:
 972                                 /* no physical table access is required */
 973                                 relation = NULL;
 974                                 break;
 975                         default:
 976                                 elog(ERROR, "unrecognized markType: %d", rc->markType);
 977                                 relation = NULL;        /* keep compiler quiet */
 978                                 break;
 979                 }
 980
 981                 /* Check that relation is a legal target for marking */
 982                 if (relation)
 983                         CheckValidRowMarkRel(relation, rc->markType);
 984
 985                 erm = (ExecRowMark *) palloc(sizeof(ExecRowMark));
 986                 erm->relation = relation;
 987                 erm->relid = relid;
 988                 erm->rti = rc->rti;
 989                 erm->prti = rc->prti;
 990                 erm->rowmarkId = rc->rowmarkId;
 991                 erm->markType = rc->markType;
 992                 erm->strength = rc->strength;
 993                 erm->waitPolicy = rc->waitPolicy;
 994                 erm->ermActive = false;
 995                 ItemPointerSetInvalid(&(erm->curCtid));
 996                 erm->ermExtra = NULL;
 997                 estate->es_rowMarks = lappend(estate->es_rowMarks, erm);
 998         }
 999
1000         /*
1001          * Initialize the executor's tuple table to empty.
1002          */
1003         estate->es_tupleTable = NIL;
1004         estate->es_trig_tuple_slot = NULL;
1005         estate->es_trig_oldtup_slot = NULL;
1006         estate->es_trig_newtup_slot = NULL;
1007
1008         /* mark EvalPlanQual not active */
1009         estate->es_epqTuple = NULL;
1010         estate->es_epqTupleSet = NULL;
1011         estate->es_epqScanDone = NULL;
1012
1013         /*
1014          * Initialize private state information for each SubPlan.  We must do this
1015          * before running ExecInitNode on the main query tree, since
1016          * ExecInitSubPlan expects to be able to find these entries.
1017          */
1018         Assert(estate->es_subplanstates == NIL);
1019         i = 1;                                          /* subplan indices count from 1 */
1020         foreach(l, plannedstmt->subplans)
1021         {
1022                 Plan       *subplan = (Plan *) lfirst(l);
1023                 PlanState  *subplanstate;
1024                 int                     sp_eflags;
1025
1026                 /*
1027                  * A subplan will never need to do BACKWARD scan nor MARK/RESTORE. If
1028                  * it is a parameterless subplan (not initplan), we suggest that it be
1029                  * prepared to handle REWIND efficiently; otherwise there is no need.
1030                  */
1031                 sp_eflags = eflags
1032                         & (EXEC_FLAG_EXPLAIN_ONLY | EXEC_FLAG_WITH_NO_DATA);
1033                 if (bms_is_member(i, plannedstmt->rewindPlanIDs))
1034                         sp_eflags |= EXEC_FLAG_REWIND;
1035
1036                 subplanstate = ExecInitNode(subplan, estate, sp_eflags);
1037
1038                 estate->es_subplanstates = lappend(estate->es_subplanstates,
1039                                                                                    subplanstate);
1040
1041                 i++;
1042         }
1043
1044         /*
1045          * Initialize the private state information for all the nodes in the query
1046          * tree.  This opens files, allocates storage and leaves us ready to start
1047          * processing tuples.
1048          */
1049         planstate = ExecInitNode(plan, estate, eflags);
1050
1051         /*
1052          * Get the tuple descriptor describing the type of tuples to return.
1053          */
1054         tupType = ExecGetResultType(planstate);
1055
1056         /*
1057          * Initialize the junk filter if needed.  SELECT queries need a filter if
1058          * there are any junk attrs in the top-level tlist.
1059          */
1060         if (operation == CMD_SELECT)
1061         {
1062                 bool            junk_filter_needed = false;
1063                 ListCell   *tlist;
1064
1065                 foreach(tlist, plan->targetlist)
1066                 {
1067                         TargetEntry *tle = (TargetEntry *) lfirst(tlist);
1068
1069                         if (tle->resjunk)
1070                         {
1071                                 junk_filter_needed = true;
1072                                 break;
1073                         }
1074                 }
1075
1076                 if (junk_filter_needed)
1077                 {
1078                         JunkFilter *j;
1079
1080                         j = ExecInitJunkFilter(planstate->plan->targetlist,
1081                                                                    tupType->tdhasoid,
1082                                                                    ExecInitExtraTupleSlot(estate, NULL));
1083                         estate->es_junkFilter = j;
1084
1085                         /* Want to return the cleaned tuple type */
1086                         tupType = j->jf_cleanTupType;
1087                 }
1088         }
1089
1090         queryDesc->tupDesc = tupType;
1091         queryDesc->planstate = planstate;
1092 }
1093
1094 /*
1095  * Check that a proposed result relation is a legal target for the operation
1096  *
1097  * Generally the parser and/or planner should have noticed any such mistake
1098  * already, but let's make sure.
1099  *
1100  * Note: when changing this function, you probably also need to look at
1101  * CheckValidRowMarkRel.
1102  */
1103 void
1104 CheckValidResultRel(ResultRelInfo *resultRelInfo, CmdType operation)
1105 {
1106         Relation        resultRel = resultRelInfo->ri_RelationDesc;
1107         TriggerDesc *trigDesc = resultRel->trigdesc;
1108         FdwRoutine *fdwroutine;
1109
1110         switch (resultRel->rd_rel->relkind)
1111         {
1112                 case RELKIND_RELATION:
1113                 case RELKIND_PARTITIONED_TABLE:
1114                         CheckCmdReplicaIdentity(resultRel, operation);
1115                         break;
1116                 case RELKIND_SEQUENCE:
1117                         ereport(ERROR,
1118                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1119                                          errmsg("cannot change sequence \"%s\"",
1120                                                         RelationGetRelationName(resultRel))));
1121                         break;
1122                 case RELKIND_TOASTVALUE:
1123                         ereport(ERROR,
1124                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1125                                          errmsg("cannot change TOAST relation \"%s\"",
1126                                                         RelationGetRelationName(resultRel))));
1127                         break;
1128                 case RELKIND_VIEW:
1129
1130                         /*
1131                          * Okay only if there's a suitable INSTEAD OF trigger.  Messages
1132                          * here should match rewriteHandler.c's rewriteTargetView, except
1133                          * that we omit errdetail because we haven't got the information
1134                          * handy (and given that we really shouldn't get here anyway, it's
1135                          * not worth great exertion to get).
1136                          */
1137                         switch (operation)
1138                         {
1139                                 case CMD_INSERT:
1140                                         if (!trigDesc || !trigDesc->trig_insert_instead_row)
1141                                                 ereport(ERROR,
1142                                                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1143                                                                  errmsg("cannot insert into view \"%s\"",
1144                                                                                 RelationGetRelationName(resultRel)),
1145                                                                  errhint("To enable inserting into the view, provide an INSTEAD OF INSERT trigger or an unconditional ON INSERT DO INSTEAD rule.")));
1146                                         break;
1147                                 case CMD_UPDATE:
1148                                         if (!trigDesc || !trigDesc->trig_update_instead_row)
1149                                                 ereport(ERROR,
1150                                                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1151                                                                  errmsg("cannot update view \"%s\"",
1152                                                                                 RelationGetRelationName(resultRel)),
1153                                                                  errhint("To enable updating the view, provide an INSTEAD OF UPDATE trigger or an unconditional ON UPDATE DO INSTEAD rule.")));
1154                                         break;
1155                                 case CMD_DELETE:
1156                                         if (!trigDesc || !trigDesc->trig_delete_instead_row)
1157                                                 ereport(ERROR,
1158                                                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1159                                                                  errmsg("cannot delete from view \"%s\"",
1160                                                                                 RelationGetRelationName(resultRel)),
1161                                                                  errhint("To enable deleting from the view, provide an INSTEAD OF DELETE trigger or an unconditional ON DELETE DO INSTEAD rule.")));
1162                                         break;
1163                                 default:
1164                                         elog(ERROR, "unrecognized CmdType: %d", (int) operation);
1165                                         break;
1166                         }
1167                         break;
1168                 case RELKIND_MATVIEW:
1169                         if (!MatViewIncrementalMaintenanceIsEnabled())
1170                                 ereport(ERROR,
1171                                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1172                                                  errmsg("cannot change materialized view \"%s\"",
1173                                                                 RelationGetRelationName(resultRel))));
1174                         break;
1175                 case RELKIND_FOREIGN_TABLE:
1176                         /* Okay only if the FDW supports it */
1177                         fdwroutine = resultRelInfo->ri_FdwRoutine;
1178                         switch (operation)
1179                         {
1180                                 case CMD_INSERT:
1181
1182                                         /*
1183                                          * If foreign partition to do tuple-routing for, skip the
1184                                          * check; it's disallowed elsewhere.
1185                                          */
1186                                         if (resultRelInfo->ri_PartitionRoot)
1187                                                 break;
1188                                         if (fdwroutine->ExecForeignInsert == NULL)
1189                                                 ereport(ERROR,
1190                                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1191                                                                  errmsg("cannot insert into foreign table \"%s\"",
1192                                                                                 RelationGetRelationName(resultRel))));
1193                                         if (fdwroutine->IsForeignRelUpdatable != NULL &&
1194                                                 (fdwroutine->IsForeignRelUpdatable(resultRel) & (1 << CMD_INSERT)) == 0)
1195                                                 ereport(ERROR,
1196                                                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1197                                                                  errmsg("foreign table \"%s\" does not allow inserts",
1198                                                                                 RelationGetRelationName(resultRel))));
1199                                         break;
1200                                 case CMD_UPDATE:
1201                                         if (fdwroutine->ExecForeignUpdate == NULL)
1202                                                 ereport(ERROR,
1203                                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1204                                                                  errmsg("cannot update foreign table \"%s\"",
1205                                                                                 RelationGetRelationName(resultRel))));
1206                                         if (fdwroutine->IsForeignRelUpdatable != NULL &&
1207                                                 (fdwroutine->IsForeignRelUpdatable(resultRel) & (1 << CMD_UPDATE)) == 0)
1208                                                 ereport(ERROR,
1209                                                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1210                                                                  errmsg("foreign table \"%s\" does not allow updates",
1211                                                                                 RelationGetRelationName(resultRel))));
1212                                         break;
1213                                 case CMD_DELETE:
1214                                         if (fdwroutine->ExecForeignDelete == NULL)
1215                                                 ereport(ERROR,
1216                                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1217                                                                  errmsg("cannot delete from foreign table \"%s\"",
1218                                                                                 RelationGetRelationName(resultRel))));
1219                                         if (fdwroutine->IsForeignRelUpdatable != NULL &&
1220                                                 (fdwroutine->IsForeignRelUpdatable(resultRel) & (1 << CMD_DELETE)) == 0)
1221                                                 ereport(ERROR,
1222                                                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1223                                                                  errmsg("foreign table \"%s\" does not allow deletes",
1224                                                                                 RelationGetRelationName(resultRel))));
1225                                         break;
1226                                 default:
1227                                         elog(ERROR, "unrecognized CmdType: %d", (int) operation);
1228                                         break;
1229                         }
1230                         break;
1231                 default:
1232                         ereport(ERROR,
1233                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1234                                          errmsg("cannot change relation \"%s\"",
1235                                                         RelationGetRelationName(resultRel))));
1236                         break;
1237         }
1238 }
1239
1240 /*
1241  * Check that a proposed rowmark target relation is a legal target
1242  *
1243  * In most cases parser and/or planner should have noticed this already, but
1244  * they don't cover all cases.
1245  */
1246 static void
1247 CheckValidRowMarkRel(Relation rel, RowMarkType markType)
1248 {
1249         FdwRoutine *fdwroutine;
1250
1251         switch (rel->rd_rel->relkind)
1252         {
1253                 case RELKIND_RELATION:
1254                 case RELKIND_PARTITIONED_TABLE:
1255                         /* OK */
1256                         break;
1257                 case RELKIND_SEQUENCE:
1258                         /* Must disallow this because we don't vacuum sequences */
1259                         ereport(ERROR,
1260                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1261                                          errmsg("cannot lock rows in sequence \"%s\"",
1262                                                         RelationGetRelationName(rel))));
1263                         break;
1264                 case RELKIND_TOASTVALUE:
1265                         /* We could allow this, but there seems no good reason to */
1266                         ereport(ERROR,
1267                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1268                                          errmsg("cannot lock rows in TOAST relation \"%s\"",
1269                                                         RelationGetRelationName(rel))));
1270                         break;
1271                 case RELKIND_VIEW:
1272                         /* Should not get here; planner should have expanded the view */
1273                         ereport(ERROR,
1274                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1275                                          errmsg("cannot lock rows in view \"%s\"",
1276                                                         RelationGetRelationName(rel))));
1277                         break;
1278                 case RELKIND_MATVIEW:
1279                         /* Allow referencing a matview, but not actual locking clauses */
1280                         if (markType != ROW_MARK_REFERENCE)
1281                                 ereport(ERROR,
1282                                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1283                                                  errmsg("cannot lock rows in materialized view \"%s\"",
1284                                                                 RelationGetRelationName(rel))));
1285                         break;
1286                 case RELKIND_FOREIGN_TABLE:
1287                         /* Okay only if the FDW supports it */
1288                         fdwroutine = GetFdwRoutineForRelation(rel, false);
1289                         if (fdwroutine->RefetchForeignRow == NULL)
1290                                 ereport(ERROR,
1291                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1292                                                  errmsg("cannot lock rows in foreign table \"%s\"",
1293                                                                 RelationGetRelationName(rel))));
1294                         break;
1295                 default:
1296                         ereport(ERROR,
1297                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1298                                          errmsg("cannot lock rows in relation \"%s\"",
1299                                                         RelationGetRelationName(rel))));
1300                         break;
1301         }
1302 }
1303
1304 /*
1305  * Initialize ResultRelInfo data for one result relation
1306  *
1307  * Caution: before Postgres 9.1, this function included the relkind checking
1308  * that's now in CheckValidResultRel, and it also did ExecOpenIndices if
1309  * appropriate.  Be sure callers cover those needs.
1310  */
1311 void
1312 InitResultRelInfo(ResultRelInfo *resultRelInfo,
1313                                   Relation resultRelationDesc,
1314                                   Index resultRelationIndex,
1315                                   Relation partition_root,
1316                                   int instrument_options)
1317 {
1318         List       *partition_check = NIL;
1319
1320         MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
1321         resultRelInfo->type = T_ResultRelInfo;
1322         resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
1323         resultRelInfo->ri_RelationDesc = resultRelationDesc;
1324         resultRelInfo->ri_NumIndices = 0;
1325         resultRelInfo->ri_IndexRelationDescs = NULL;
1326         resultRelInfo->ri_IndexRelationInfo = NULL;
1327         /* make a copy so as not to depend on relcache info not changing... */
1328         resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
1329         if (resultRelInfo->ri_TrigDesc)
1330         {
1331                 int                     n = resultRelInfo->ri_TrigDesc->numtriggers;
1332
1333                 resultRelInfo->ri_TrigFunctions = (FmgrInfo *)
1334                         palloc0(n * sizeof(FmgrInfo));
1335                 resultRelInfo->ri_TrigWhenExprs = (ExprState **)
1336                         palloc0(n * sizeof(ExprState *));
1337                 if (instrument_options)
1338                         resultRelInfo->ri_TrigInstrument = InstrAlloc(n, instrument_options);
1339         }
1340         else
1341         {
1342                 resultRelInfo->ri_TrigFunctions = NULL;
1343                 resultRelInfo->ri_TrigWhenExprs = NULL;
1344                 resultRelInfo->ri_TrigInstrument = NULL;
1345         }
1346         if (resultRelationDesc->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1347                 resultRelInfo->ri_FdwRoutine = GetFdwRoutineForRelation(resultRelationDesc, true);
1348         else
1349                 resultRelInfo->ri_FdwRoutine = NULL;
1350
1351         /* The following fields are set later if needed */
1352         resultRelInfo->ri_FdwState = NULL;
1353         resultRelInfo->ri_usesFdwDirectModify = false;
1354         resultRelInfo->ri_ConstraintExprs = NULL;
1355         resultRelInfo->ri_junkFilter = NULL;
1356         resultRelInfo->ri_projectReturning = NULL;
1357         resultRelInfo->ri_onConflictArbiterIndexes = NIL;
1358         resultRelInfo->ri_onConflict = NULL;
1359
1360         /*
1361          * Partition constraint, which also includes the partition constraint of
1362          * all the ancestors that are partitions.  Note that it will be checked
1363          * even in the case of tuple-routing where this table is the target leaf
1364          * partition, if there any BR triggers defined on the table.  Although
1365          * tuple-routing implicitly preserves the partition constraint of the
1366          * target partition for a given row, the BR triggers may change the row
1367          * such that the constraint is no longer satisfied, which we must fail for
1368          * by checking it explicitly.
1369          *
1370          * If this is a partitioned table, the partition constraint (if any) of a
1371          * given row will be checked just before performing tuple-routing.
1372          */
1373         partition_check = RelationGetPartitionQual(resultRelationDesc);
1374
1375         resultRelInfo->ri_PartitionCheck = partition_check;
1376         resultRelInfo->ri_PartitionRoot = partition_root;
1377 }
1378
1379 /*
1380  *              ExecGetTriggerResultRel
1381  *
1382  * Get a ResultRelInfo for a trigger target relation.  Most of the time,
1383  * triggers are fired on one of the result relations of the query, and so
1384  * we can just return a member of the es_result_relations array, the
1385  * es_root_result_relations array (if any), or the es_leaf_result_relations
1386  * list (if any).  (Note: in self-join situations there might be multiple
1387  * members with the same OID; if so it doesn't matter which one we pick.)
1388  * However, it is sometimes necessary to fire triggers on other relations;
1389  * this happens mainly when an RI update trigger queues additional triggers
1390  * on other relations, which will be processed in the context of the outer
1391  * query.  For efficiency's sake, we want to have a ResultRelInfo for those
1392  * triggers too; that can avoid repeated re-opening of the relation.  (It
1393  * also provides a way for EXPLAIN ANALYZE to report the runtimes of such
1394  * triggers.)  So we make additional ResultRelInfo's as needed, and save them
1395  * in es_trig_target_relations.
1396  */
1397 ResultRelInfo *
1398 ExecGetTriggerResultRel(EState *estate, Oid relid)
1399 {
1400         ResultRelInfo *rInfo;
1401         int                     nr;
1402         ListCell   *l;
1403         Relation        rel;
1404         MemoryContext oldcontext;
1405
1406         /* First, search through the query result relations */
1407         rInfo = estate->es_result_relations;
1408         nr = estate->es_num_result_relations;
1409         while (nr > 0)
1410         {
1411                 if (RelationGetRelid(rInfo->ri_RelationDesc) == relid)
1412                         return rInfo;
1413                 rInfo++;
1414                 nr--;
1415         }
1416         /* Second, search through the root result relations, if any */
1417         rInfo = estate->es_root_result_relations;
1418         nr = estate->es_num_root_result_relations;
1419         while (nr > 0)
1420         {
1421                 if (RelationGetRelid(rInfo->ri_RelationDesc) == relid)
1422                         return rInfo;
1423                 rInfo++;
1424                 nr--;
1425         }
1426         /*
1427          * Third, search through the result relations that were created during
1428          * tuple routing, if any.
1429          */
1430         foreach(l, estate->es_tuple_routing_result_relations)
1431         {
1432                 rInfo = (ResultRelInfo *) lfirst(l);
1433                 if (RelationGetRelid(rInfo->ri_RelationDesc) == relid)
1434                         return rInfo;
1435         }
1436         /* Nope, but maybe we already made an extra ResultRelInfo for it */
1437         foreach(l, estate->es_trig_target_relations)
1438         {
1439                 rInfo = (ResultRelInfo *) lfirst(l);
1440                 if (RelationGetRelid(rInfo->ri_RelationDesc) == relid)
1441                         return rInfo;
1442         }
1443         /* Nope, so we need a new one */
1444
1445         /*
1446          * Open the target relation's relcache entry.  We assume that an
1447          * appropriate lock is still held by the backend from whenever the trigger
1448          * event got queued, so we need take no new lock here.  Also, we need not
1449          * recheck the relkind, so no need for CheckValidResultRel.
1450          */
1451         rel = heap_open(relid, NoLock);
1452
1453         /*
1454          * Make the new entry in the right context.
1455          */
1456         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
1457         rInfo = makeNode(ResultRelInfo);
1458         InitResultRelInfo(rInfo,
1459                                           rel,
1460                                           0,            /* dummy rangetable index */
1461                                           NULL,
1462                                           estate->es_instrument);
1463         estate->es_trig_target_relations =
1464                 lappend(estate->es_trig_target_relations, rInfo);
1465         MemoryContextSwitchTo(oldcontext);
1466
1467         /*
1468          * Currently, we don't need any index information in ResultRelInfos used
1469          * only for triggers, so no need to call ExecOpenIndices.
1470          */
1471
1472         return rInfo;
1473 }
1474
1475 /*
1476  * Close any relations that have been opened by ExecGetTriggerResultRel().
1477  */
1478 void
1479 ExecCleanUpTriggerState(EState *estate)
1480 {
1481         ListCell   *l;
1482
1483         foreach(l, estate->es_trig_target_relations)
1484         {
1485                 ResultRelInfo *resultRelInfo = (ResultRelInfo *) lfirst(l);
1486
1487                 /* Close indices and then the relation itself */
1488                 ExecCloseIndices(resultRelInfo);
1489                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
1490         }
1491 }
1492
1493 /*
1494  *              ExecContextForcesOids
1495  *
1496  * This is pretty grotty: when doing INSERT, UPDATE, or CREATE TABLE AS,
1497  * we need to ensure that result tuples have space for an OID iff they are
1498  * going to be stored into a relation that has OIDs.  In other contexts
1499  * we are free to choose whether to leave space for OIDs in result tuples
1500  * (we generally don't want to, but we do if a physical-tlist optimization
1501  * is possible).  This routine checks the plan context and returns true if the
1502  * choice is forced, false if the choice is not forced.  In the true case,
1503  * *hasoids is set to the required value.
1504  *
1505  * One reason this is ugly is that all plan nodes in the plan tree will emit
1506  * tuples with space for an OID, though we really only need the topmost node
1507  * to do so.  However, node types like Sort don't project new tuples but just
1508  * return their inputs, and in those cases the requirement propagates down
1509  * to the input node.  Eventually we might make this code smart enough to
1510  * recognize how far down the requirement really goes, but for now we just
1511  * make all plan nodes do the same thing if the top level forces the choice.
1512  *
1513  * We assume that if we are generating tuples for INSERT or UPDATE,
1514  * estate->es_result_relation_info is already set up to describe the target
1515  * relation.  Note that in an UPDATE that spans an inheritance tree, some of
1516  * the target relations may have OIDs and some not.  We have to make the
1517  * decisions on a per-relation basis as we initialize each of the subplans of
1518  * the ModifyTable node, so ModifyTable has to set es_result_relation_info
1519  * while initializing each subplan.
1520  *
1521  * CREATE TABLE AS is even uglier, because we don't have the target relation's
1522  * descriptor available when this code runs; we have to look aside at the
1523  * flags passed to ExecutorStart().
1524  */
1525 bool
1526 ExecContextForcesOids(PlanState *planstate, bool *hasoids)
1527 {
1528         ResultRelInfo *ri = planstate->state->es_result_relation_info;
1529
1530         if (ri != NULL)
1531         {
1532                 Relation        rel = ri->ri_RelationDesc;
1533
1534                 if (rel != NULL)
1535                 {
1536                         *hasoids = rel->rd_rel->relhasoids;
1537                         return true;
1538                 }
1539         }
1540
1541         if (planstate->state->es_top_eflags & EXEC_FLAG_WITH_OIDS)
1542         {
1543                 *hasoids = true;
1544                 return true;
1545         }
1546         if (planstate->state->es_top_eflags & EXEC_FLAG_WITHOUT_OIDS)
1547         {
1548                 *hasoids = false;
1549                 return true;
1550         }
1551
1552         return false;
1553 }
1554
1555 /* ----------------------------------------------------------------
1556  *              ExecPostprocessPlan
1557  *
1558  *              Give plan nodes a final chance to execute before shutdown
1559  * ----------------------------------------------------------------
1560  */
1561 static void
1562 ExecPostprocessPlan(EState *estate)
1563 {
1564         ListCell   *lc;
1565
1566         /*
1567          * Make sure nodes run forward.
1568          */
1569         estate->es_direction = ForwardScanDirection;
1570
1571         /*
1572          * Run any secondary ModifyTable nodes to completion, in case the main
1573          * query did not fetch all rows from them.  (We do this to ensure that
1574          * such nodes have predictable results.)
1575          */
1576         foreach(lc, estate->es_auxmodifytables)
1577         {
1578                 PlanState  *ps = (PlanState *) lfirst(lc);
1579
1580                 for (;;)
1581                 {
1582                         TupleTableSlot *slot;
1583
1584                         /* Reset the per-output-tuple exprcontext each time */
1585                         ResetPerTupleExprContext(estate);
1586
1587                         slot = ExecProcNode(ps);
1588
1589                         if (TupIsNull(slot))
1590                                 break;
1591                 }
1592         }
1593 }
1594
1595 /* ----------------------------------------------------------------
1596  *              ExecEndPlan
1597  *
1598  *              Cleans up the query plan -- closes files and frees up storage
1599  *
1600  * NOTE: we are no longer very worried about freeing storage per se
1601  * in this code; FreeExecutorState should be guaranteed to release all
1602  * memory that needs to be released.  What we are worried about doing
1603  * is closing relations and dropping buffer pins.  Thus, for example,
1604  * tuple tables must be cleared or dropped to ensure pins are released.
1605  * ----------------------------------------------------------------
1606  */
1607 static void
1608 ExecEndPlan(PlanState *planstate, EState *estate)
1609 {
1610         ResultRelInfo *resultRelInfo;
1611         int                     i;
1612         ListCell   *l;
1613
1614         /*
1615          * shut down the node-type-specific query processing
1616          */
1617         ExecEndNode(planstate);
1618
1619         /*
1620          * for subplans too
1621          */
1622         foreach(l, estate->es_subplanstates)
1623         {
1624                 PlanState  *subplanstate = (PlanState *) lfirst(l);
1625
1626                 ExecEndNode(subplanstate);
1627         }
1628
1629         /*
1630          * destroy the executor's tuple table.  Actually we only care about
1631          * releasing buffer pins and tupdesc refcounts; there's no need to pfree
1632          * the TupleTableSlots, since the containing memory context is about to go
1633          * away anyway.
1634          */
1635         ExecResetTupleTable(estate->es_tupleTable, false);
1636
1637         /*
1638          * close the result relation(s) if any, but hold locks until xact commit.
1639          */
1640         resultRelInfo = estate->es_result_relations;
1641         for (i = estate->es_num_result_relations; i > 0; i--)
1642         {
1643                 /* Close indices and then the relation itself */
1644                 ExecCloseIndices(resultRelInfo);
1645                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
1646                 resultRelInfo++;
1647         }
1648
1649         /* Close the root target relation(s). */
1650         resultRelInfo = estate->es_root_result_relations;
1651         for (i = estate->es_num_root_result_relations; i > 0; i--)
1652         {
1653                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
1654                 resultRelInfo++;
1655         }
1656
1657         /* likewise close any trigger target relations */
1658         ExecCleanUpTriggerState(estate);
1659
1660         /*
1661          * close any relations selected FOR [KEY] UPDATE/SHARE, again keeping
1662          * locks
1663          */
1664         foreach(l, estate->es_rowMarks)
1665         {
1666                 ExecRowMark *erm = (ExecRowMark *) lfirst(l);
1667
1668                 if (erm->relation)
1669                         heap_close(erm->relation, NoLock);
1670         }
1671 }
1672
1673 /* ----------------------------------------------------------------
1674  *              ExecutePlan
1675  *
1676  *              Processes the query plan until we have retrieved 'numberTuples' tuples,
1677  *              moving in the specified direction.
1678  *
1679  *              Runs to completion if numberTuples is 0
1680  *
1681  * Note: the ctid attribute is a 'junk' attribute that is removed before the
1682  * user can see it
1683  * ----------------------------------------------------------------
1684  */
1685 static void
1686 ExecutePlan(EState *estate,
1687                         PlanState *planstate,
1688                         bool use_parallel_mode,
1689                         CmdType operation,
1690                         bool sendTuples,
1691                         uint64 numberTuples,
1692                         ScanDirection direction,
1693                         DestReceiver *dest,
1694                         bool execute_once)
1695 {
1696         TupleTableSlot *slot;
1697         uint64          current_tuple_count;
1698
1699         /*
1700          * initialize local variables
1701          */
1702         current_tuple_count = 0;
1703
1704         /*
1705          * Set the direction.
1706          */
1707         estate->es_direction = direction;
1708
1709         /*
1710          * If the plan might potentially be executed multiple times, we must force
1711          * it to run without parallelism, because we might exit early.
1712          */
1713         if (!execute_once)
1714                 use_parallel_mode = false;
1715
1716         estate->es_use_parallel_mode = use_parallel_mode;
1717         if (use_parallel_mode)
1718                 EnterParallelMode();
1719
1720         /*
1721          * Loop until we've processed the proper number of tuples from the plan.
1722          */
1723         for (;;)
1724         {
1725                 /* Reset the per-output-tuple exprcontext */
1726                 ResetPerTupleExprContext(estate);
1727
1728                 /*
1729                  * Execute the plan and obtain a tuple
1730                  */
1731                 slot = ExecProcNode(planstate);
1732
1733                 /*
1734                  * if the tuple is null, then we assume there is nothing more to
1735                  * process so we just end the loop...
1736                  */
1737                 if (TupIsNull(slot))
1738                 {
1739                         /* Allow nodes to release or shut down resources. */
1740                         (void) ExecShutdownNode(planstate);
1741                         break;
1742                 }
1743
1744                 /*
1745                  * If we have a junk filter, then project a new tuple with the junk
1746                  * removed.
1747                  *
1748                  * Store this new "clean" tuple in the junkfilter's resultSlot.
1749                  * (Formerly, we stored it back over the "dirty" tuple, which is WRONG
1750                  * because that tuple slot has the wrong descriptor.)
1751                  */
1752                 if (estate->es_junkFilter != NULL)
1753                         slot = ExecFilterJunk(estate->es_junkFilter, slot);
1754
1755                 /*
1756                  * If we are supposed to send the tuple somewhere, do so. (In
1757                  * practice, this is probably always the case at this point.)
1758                  */
1759                 if (sendTuples)
1760                 {
1761                         /*
1762                          * If we are not able to send the tuple, we assume the destination
1763                          * has closed and no more tuples can be sent. If that's the case,
1764                          * end the loop.
1765                          */
1766                         if (!dest->receiveSlot(slot, dest))
1767                                 break;
1768                 }
1769
1770                 /*
1771                  * Count tuples processed, if this is a SELECT.  (For other operation
1772                  * types, the ModifyTable plan node must count the appropriate
1773                  * events.)
1774                  */
1775                 if (operation == CMD_SELECT)
1776                         (estate->es_processed)++;
1777
1778                 /*
1779                  * check our tuple count.. if we've processed the proper number then
1780                  * quit, else loop again and process more tuples.  Zero numberTuples
1781                  * means no limit.
1782                  */
1783                 current_tuple_count++;
1784                 if (numberTuples && numberTuples == current_tuple_count)
1785                 {
1786                         /* Allow nodes to release or shut down resources. */
1787                         (void) ExecShutdownNode(planstate);
1788                         break;
1789                 }
1790         }
1791
1792         if (use_parallel_mode)
1793                 ExitParallelMode();
1794 }
1795
1796
1797 /*
1798  * ExecRelCheck --- check that tuple meets constraints for result relation
1799  *
1800  * Returns NULL if OK, else name of failed check constraint
1801  */
1802 static const char *
1803 ExecRelCheck(ResultRelInfo *resultRelInfo,
1804                          TupleTableSlot *slot, EState *estate)
1805 {
1806         Relation        rel = resultRelInfo->ri_RelationDesc;
1807         int                     ncheck = rel->rd_att->constr->num_check;
1808         ConstrCheck *check = rel->rd_att->constr->check;
1809         ExprContext *econtext;
1810         MemoryContext oldContext;
1811         int                     i;
1812
1813         /*
1814          * If first time through for this result relation, build expression
1815          * nodetrees for rel's constraint expressions.  Keep them in the per-query
1816          * memory context so they'll survive throughout the query.
1817          */
1818         if (resultRelInfo->ri_ConstraintExprs == NULL)
1819         {
1820                 oldContext = MemoryContextSwitchTo(estate->es_query_cxt);
1821                 resultRelInfo->ri_ConstraintExprs =
1822                         (ExprState **) palloc(ncheck * sizeof(ExprState *));
1823                 for (i = 0; i < ncheck; i++)
1824                 {
1825                         Expr       *checkconstr;
1826
1827                         checkconstr = stringToNode(check[i].ccbin);
1828                         resultRelInfo->ri_ConstraintExprs[i] =
1829                                 ExecPrepareExpr(checkconstr, estate);
1830                 }
1831                 MemoryContextSwitchTo(oldContext);
1832         }
1833
1834         /*
1835          * We will use the EState's per-tuple context for evaluating constraint
1836          * expressions (creating it if it's not already there).
1837          */
1838         econtext = GetPerTupleExprContext(estate);
1839
1840         /* Arrange for econtext's scan tuple to be the tuple under test */
1841         econtext->ecxt_scantuple = slot;
1842
1843         /* And evaluate the constraints */
1844         for (i = 0; i < ncheck; i++)
1845         {
1846                 ExprState  *checkconstr = resultRelInfo->ri_ConstraintExprs[i];
1847
1848                 /*
1849                  * NOTE: SQL specifies that a NULL result from a constraint expression
1850                  * is not to be treated as a failure.  Therefore, use ExecCheck not
1851                  * ExecQual.
1852                  */
1853                 if (!ExecCheck(checkconstr, econtext))
1854                         return check[i].ccname;
1855         }
1856
1857         /* NULL result means no error */
1858         return NULL;
1859 }
1860
1861 /*
1862  * ExecPartitionCheck --- check that tuple meets the partition constraint.
1863  *
1864  * Exported in executor.h for outside use.
1865  * Returns true if it meets the partition constraint, else returns false.
1866  */
1867 bool
1868 ExecPartitionCheck(ResultRelInfo *resultRelInfo, TupleTableSlot *slot,
1869                                    EState *estate)
1870 {
1871         ExprContext *econtext;
1872
1873         /*
1874          * If first time through, build expression state tree for the partition
1875          * check expression.  Keep it in the per-query memory context so they'll
1876          * survive throughout the query.
1877          */
1878         if (resultRelInfo->ri_PartitionCheckExpr == NULL)
1879         {
1880                 List       *qual = resultRelInfo->ri_PartitionCheck;
1881
1882                 resultRelInfo->ri_PartitionCheckExpr = ExecPrepareCheck(qual, estate);
1883         }
1884
1885         /*
1886          * We will use the EState's per-tuple context for evaluating constraint
1887          * expressions (creating it if it's not already there).
1888          */
1889         econtext = GetPerTupleExprContext(estate);
1890
1891         /* Arrange for econtext's scan tuple to be the tuple under test */
1892         econtext->ecxt_scantuple = slot;
1893
1894         /*
1895          * As in case of the catalogued constraints, we treat a NULL result as
1896          * success here, not a failure.
1897          */
1898         return ExecCheck(resultRelInfo->ri_PartitionCheckExpr, econtext);
1899 }
1900
1901 /*
1902  * ExecPartitionCheckEmitError - Form and emit an error message after a failed
1903  * partition constraint check.
1904  */
1905 void
1906 ExecPartitionCheckEmitError(ResultRelInfo *resultRelInfo,
1907                                                         TupleTableSlot *slot,
1908                                                         EState *estate)
1909 {
1910         Relation        rel = resultRelInfo->ri_RelationDesc;
1911         Relation        orig_rel = rel;
1912         TupleDesc       tupdesc = RelationGetDescr(rel);
1913         char       *val_desc;
1914         Bitmapset  *modifiedCols;
1915         Bitmapset  *insertedCols;
1916         Bitmapset  *updatedCols;
1917
1918         /*
1919          * Need to first convert the tuple to the root partitioned table's row
1920          * type. For details, check similar comments in ExecConstraints().
1921          */
1922         if (resultRelInfo->ri_PartitionRoot)
1923         {
1924                 HeapTuple       tuple = ExecFetchSlotTuple(slot);
1925                 TupleDesc       old_tupdesc = RelationGetDescr(rel);
1926                 TupleConversionMap *map;
1927
1928                 rel = resultRelInfo->ri_PartitionRoot;
1929                 tupdesc = RelationGetDescr(rel);
1930                 /* a reverse map */
1931                 map = convert_tuples_by_name(old_tupdesc, tupdesc,
1932                                                                          gettext_noop("could not convert row type"));
1933                 if (map != NULL)
1934                 {
1935                         tuple = do_convert_tuple(tuple, map);
1936                         ExecSetSlotDescriptor(slot, tupdesc);
1937                         ExecStoreTuple(tuple, slot, InvalidBuffer, false);
1938                 }
1939         }
1940
1941         insertedCols = GetInsertedColumns(resultRelInfo, estate);
1942         updatedCols = GetUpdatedColumns(resultRelInfo, estate);
1943         modifiedCols = bms_union(insertedCols, updatedCols);
1944         val_desc = ExecBuildSlotValueDescription(RelationGetRelid(rel),
1945                                                                                          slot,
1946                                                                                          tupdesc,
1947                                                                                          modifiedCols,
1948                                                                                          64);
1949         ereport(ERROR,
1950                         (errcode(ERRCODE_CHECK_VIOLATION),
1951                          errmsg("new row for relation \"%s\" violates partition constraint",
1952                                         RelationGetRelationName(orig_rel)),
1953                          val_desc ? errdetail("Failing row contains %s.", val_desc) : 0));
1954 }
1955
1956 /*
1957  * ExecConstraints - check constraints of the tuple in 'slot'
1958  *
1959  * This checks the traditional NOT NULL and check constraints, and if
1960  * requested, checks the partition constraint.
1961  *
1962  * Note: 'slot' contains the tuple to check the constraints of, which may
1963  * have been converted from the original input tuple after tuple routing.
1964  * 'resultRelInfo' is the original result relation, before tuple routing.
1965  */
1966 void
1967 ExecConstraints(ResultRelInfo *resultRelInfo,
1968                                 TupleTableSlot *slot, EState *estate,
1969                                 bool check_partition_constraint)
1970 {
1971         Relation        rel = resultRelInfo->ri_RelationDesc;
1972         TupleDesc       tupdesc = RelationGetDescr(rel);
1973         TupleConstr *constr = tupdesc->constr;
1974         Bitmapset  *modifiedCols;
1975         Bitmapset  *insertedCols;
1976         Bitmapset  *updatedCols;
1977
1978         Assert(constr || resultRelInfo->ri_PartitionCheck);
1979
1980         if (constr && constr->has_not_null)
1981         {
1982                 int                     natts = tupdesc->natts;
1983                 int                     attrChk;
1984
1985                 for (attrChk = 1; attrChk <= natts; attrChk++)
1986                 {
1987                         Form_pg_attribute att = TupleDescAttr(tupdesc, attrChk - 1);
1988
1989                         if (att->attnotnull && slot_attisnull(slot, attrChk))
1990                         {
1991                                 char       *val_desc;
1992                                 Relation        orig_rel = rel;
1993                                 TupleDesc       orig_tupdesc = RelationGetDescr(rel);
1994
1995                                 /*
1996                                  * If the tuple has been routed, it's been converted to the
1997                                  * partition's rowtype, which might differ from the root
1998                                  * table's.  We must convert it back to the root table's
1999                                  * rowtype so that val_desc shown error message matches the
2000                                  * input tuple.
2001                                  */
2002                                 if (resultRelInfo->ri_PartitionRoot)
2003                                 {
2004                                         HeapTuple       tuple = ExecFetchSlotTuple(slot);
2005                                         TupleConversionMap *map;
2006
2007                                         rel = resultRelInfo->ri_PartitionRoot;
2008                                         tupdesc = RelationGetDescr(rel);
2009                                         /* a reverse map */
2010                                         map = convert_tuples_by_name(orig_tupdesc, tupdesc,
2011                                                                                                  gettext_noop("could not convert row type"));
2012                                         if (map != NULL)
2013                                         {
2014                                                 tuple = do_convert_tuple(tuple, map);
2015                                                 ExecSetSlotDescriptor(slot, tupdesc);
2016                                                 ExecStoreTuple(tuple, slot, InvalidBuffer, false);
2017                                         }
2018                                 }
2019
2020                                 insertedCols = GetInsertedColumns(resultRelInfo, estate);
2021                                 updatedCols = GetUpdatedColumns(resultRelInfo, estate);
2022                                 modifiedCols = bms_union(insertedCols, updatedCols);
2023                                 val_desc = ExecBuildSlotValueDescription(RelationGetRelid(rel),
2024                                                                                                                  slot,
2025                                                                                                                  tupdesc,
2026                                                                                                                  modifiedCols,
2027                                                                                                                  64);
2028
2029                                 ereport(ERROR,
2030                                                 (errcode(ERRCODE_NOT_NULL_VIOLATION),
2031                                                  errmsg("null value in column \"%s\" violates not-null constraint",
2032                                                                 NameStr(att->attname)),
2033                                                  val_desc ? errdetail("Failing row contains %s.", val_desc) : 0,
2034                                                  errtablecol(orig_rel, attrChk)));
2035                         }
2036                 }
2037         }
2038
2039         if (constr && constr->num_check > 0)
2040         {
2041                 const char *failed;
2042
2043                 if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
2044                 {
2045                         char       *val_desc;
2046                         Relation        orig_rel = rel;
2047
2048                         /* See the comment above. */
2049                         if (resultRelInfo->ri_PartitionRoot)
2050                         {
2051                                 HeapTuple       tuple = ExecFetchSlotTuple(slot);
2052                                 TupleDesc       old_tupdesc = RelationGetDescr(rel);
2053                                 TupleConversionMap *map;
2054
2055                                 rel = resultRelInfo->ri_PartitionRoot;
2056                                 tupdesc = RelationGetDescr(rel);
2057                                 /* a reverse map */
2058                                 map = convert_tuples_by_name(old_tupdesc, tupdesc,
2059                                                                                          gettext_noop("could not convert row type"));
2060                                 if (map != NULL)
2061                                 {
2062                                         tuple = do_convert_tuple(tuple, map);
2063                                         ExecSetSlotDescriptor(slot, tupdesc);
2064                                         ExecStoreTuple(tuple, slot, InvalidBuffer, false);
2065                                 }
2066                         }
2067
2068                         insertedCols = GetInsertedColumns(resultRelInfo, estate);
2069                         updatedCols = GetUpdatedColumns(resultRelInfo, estate);
2070                         modifiedCols = bms_union(insertedCols, updatedCols);
2071                         val_desc = ExecBuildSlotValueDescription(RelationGetRelid(rel),
2072                                                                                                          slot,
2073                                                                                                          tupdesc,
2074                                                                                                          modifiedCols,
2075                                                                                                          64);
2076                         ereport(ERROR,
2077                                         (errcode(ERRCODE_CHECK_VIOLATION),
2078                                          errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
2079                                                         RelationGetRelationName(orig_rel), failed),
2080                                          val_desc ? errdetail("Failing row contains %s.", val_desc) : 0,
2081                                          errtableconstraint(orig_rel, failed)));
2082                 }
2083         }
2084
2085         if (check_partition_constraint && resultRelInfo->ri_PartitionCheck &&
2086                 !ExecPartitionCheck(resultRelInfo, slot, estate))
2087                 ExecPartitionCheckEmitError(resultRelInfo, slot, estate);
2088 }
2089
2090
2091 /*
2092  * ExecWithCheckOptions -- check that tuple satisfies any WITH CHECK OPTIONs
2093  * of the specified kind.
2094  *
2095  * Note that this needs to be called multiple times to ensure that all kinds of
2096  * WITH CHECK OPTIONs are handled (both those from views which have the WITH
2097  * CHECK OPTION set and from row level security policies).  See ExecInsert()
2098  * and ExecUpdate().
2099  */
2100 void
2101 ExecWithCheckOptions(WCOKind kind, ResultRelInfo *resultRelInfo,
2102                                          TupleTableSlot *slot, EState *estate)
2103 {
2104         Relation        rel = resultRelInfo->ri_RelationDesc;
2105         TupleDesc       tupdesc = RelationGetDescr(rel);
2106         ExprContext *econtext;
2107         ListCell   *l1,
2108                            *l2;
2109
2110         /*
2111          * We will use the EState's per-tuple context for evaluating constraint
2112          * expressions (creating it if it's not already there).
2113          */
2114         econtext = GetPerTupleExprContext(estate);
2115
2116         /* Arrange for econtext's scan tuple to be the tuple under test */
2117         econtext->ecxt_scantuple = slot;
2118
2119         /* Check each of the constraints */
2120         forboth(l1, resultRelInfo->ri_WithCheckOptions,
2121                         l2, resultRelInfo->ri_WithCheckOptionExprs)
2122         {
2123                 WithCheckOption *wco = (WithCheckOption *) lfirst(l1);
2124                 ExprState  *wcoExpr = (ExprState *) lfirst(l2);
2125
2126                 /*
2127                  * Skip any WCOs which are not the kind we are looking for at this
2128                  * time.
2129                  */
2130                 if (wco->kind != kind)
2131                         continue;
2132
2133                 /*
2134                  * WITH CHECK OPTION checks are intended to ensure that the new tuple
2135                  * is visible (in the case of a view) or that it passes the
2136                  * 'with-check' policy (in the case of row security). If the qual
2137                  * evaluates to NULL or FALSE, then the new tuple won't be included in
2138                  * the view or doesn't pass the 'with-check' policy for the table.
2139                  */
2140                 if (!ExecQual(wcoExpr, econtext))
2141                 {
2142                         char       *val_desc;
2143                         Bitmapset  *modifiedCols;
2144                         Bitmapset  *insertedCols;
2145                         Bitmapset  *updatedCols;
2146
2147                         switch (wco->kind)
2148                         {
2149                                         /*
2150                                          * For WITH CHECK OPTIONs coming from views, we might be
2151                                          * able to provide the details on the row, depending on
2152                                          * the permissions on the relation (that is, if the user
2153                                          * could view it directly anyway).  For RLS violations, we
2154                                          * don't include the data since we don't know if the user
2155                                          * should be able to view the tuple as that depends on the
2156                                          * USING policy.
2157                                          */
2158                                 case WCO_VIEW_CHECK:
2159                                         /* See the comment in ExecConstraints(). */
2160                                         if (resultRelInfo->ri_PartitionRoot)
2161                                         {
2162                                                 HeapTuple       tuple = ExecFetchSlotTuple(slot);
2163                                                 TupleDesc       old_tupdesc = RelationGetDescr(rel);
2164                                                 TupleConversionMap *map;
2165
2166                                                 rel = resultRelInfo->ri_PartitionRoot;
2167                                                 tupdesc = RelationGetDescr(rel);
2168                                                 /* a reverse map */
2169                                                 map = convert_tuples_by_name(old_tupdesc, tupdesc,
2170                                                                                                          gettext_noop("could not convert row type"));
2171                                                 if (map != NULL)
2172                                                 {
2173                                                         tuple = do_convert_tuple(tuple, map);
2174                                                         ExecSetSlotDescriptor(slot, tupdesc);
2175                                                         ExecStoreTuple(tuple, slot, InvalidBuffer, false);
2176                                                 }
2177                                         }
2178
2179                                         insertedCols = GetInsertedColumns(resultRelInfo, estate);
2180                                         updatedCols = GetUpdatedColumns(resultRelInfo, estate);
2181                                         modifiedCols = bms_union(insertedCols, updatedCols);
2182                                         val_desc = ExecBuildSlotValueDescription(RelationGetRelid(rel),
2183                                                                                                                          slot,
2184                                                                                                                          tupdesc,
2185                                                                                                                          modifiedCols,
2186                                                                                                                          64);
2187
2188                                         ereport(ERROR,
2189                                                         (errcode(ERRCODE_WITH_CHECK_OPTION_VIOLATION),
2190                                                          errmsg("new row violates check option for view \"%s\"",
2191                                                                         wco->relname),
2192                                                          val_desc ? errdetail("Failing row contains %s.",
2193                                                                                                   val_desc) : 0));
2194                                         break;
2195                                 case WCO_RLS_INSERT_CHECK:
2196                                 case WCO_RLS_UPDATE_CHECK:
2197                                         if (wco->polname != NULL)
2198                                                 ereport(ERROR,
2199                                                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
2200                                                                  errmsg("new row violates row-level security policy \"%s\" for table \"%s\"",
2201                                                                                 wco->polname, wco->relname)));
2202                                         else
2203                                                 ereport(ERROR,
2204                                                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
2205                                                                  errmsg("new row violates row-level security policy for table \"%s\"",
2206                                                                                 wco->relname)));
2207                                         break;
2208                                 case WCO_RLS_CONFLICT_CHECK:
2209                                         if (wco->polname != NULL)
2210                                                 ereport(ERROR,
2211                                                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
2212                                                                  errmsg("new row violates row-level security policy \"%s\" (USING expression) for table \"%s\"",
2213                                                                                 wco->polname, wco->relname)));
2214                                         else
2215                                                 ereport(ERROR,
2216                                                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
2217                                                                  errmsg("new row violates row-level security policy (USING expression) for table \"%s\"",
2218                                                                                 wco->relname)));
2219                                         break;
2220                                 default:
2221                                         elog(ERROR, "unrecognized WCO kind: %u", wco->kind);
2222                                         break;
2223                         }
2224                 }
2225         }
2226 }
2227
2228 /*
2229  * ExecBuildSlotValueDescription -- construct a string representing a tuple
2230  *
2231  * This is intentionally very similar to BuildIndexValueDescription, but
2232  * unlike that function, we truncate long field values (to at most maxfieldlen
2233  * bytes).  That seems necessary here since heap field values could be very
2234  * long, whereas index entries typically aren't so wide.
2235  *
2236  * Also, unlike the case with index entries, we need to be prepared to ignore
2237  * dropped columns.  We used to use the slot's tuple descriptor to decode the
2238  * data, but the slot's descriptor doesn't identify dropped columns, so we
2239  * now need to be passed the relation's descriptor.
2240  *
2241  * Note that, like BuildIndexValueDescription, if the user does not have
2242  * permission to view any of the columns involved, a NULL is returned.  Unlike
2243  * BuildIndexValueDescription, if the user has access to view a subset of the
2244  * column involved, that subset will be returned with a key identifying which
2245  * columns they are.
2246  */
2247 static char *
2248 ExecBuildSlotValueDescription(Oid reloid,
2249                                                           TupleTableSlot *slot,
2250                                                           TupleDesc tupdesc,
2251                                                           Bitmapset *modifiedCols,
2252                                                           int maxfieldlen)
2253 {
2254         StringInfoData buf;
2255         StringInfoData collist;
2256         bool            write_comma = false;
2257         bool            write_comma_collist = false;
2258         int                     i;
2259         AclResult       aclresult;
2260         bool            table_perm = false;
2261         bool            any_perm = false;
2262
2263         /*
2264          * Check if RLS is enabled and should be active for the relation; if so,
2265          * then don't return anything.  Otherwise, go through normal permission
2266          * checks.
2267          */
2268         if (check_enable_rls(reloid, InvalidOid, true) == RLS_ENABLED)
2269                 return NULL;
2270
2271         initStringInfo(&buf);
2272
2273         appendStringInfoChar(&buf, '(');
2274
2275         /*
2276          * Check if the user has permissions to see the row.  Table-level SELECT
2277          * allows access to all columns.  If the user does not have table-level
2278          * SELECT then we check each column and include those the user has SELECT
2279          * rights on.  Additionally, we always include columns the user provided
2280          * data for.
2281          */
2282         aclresult = pg_class_aclcheck(reloid, GetUserId(), ACL_SELECT);
2283         if (aclresult != ACLCHECK_OK)
2284         {
2285                 /* Set up the buffer for the column list */
2286                 initStringInfo(&collist);
2287                 appendStringInfoChar(&collist, '(');
2288         }
2289         else
2290                 table_perm = any_perm = true;
2291
2292         /* Make sure the tuple is fully deconstructed */
2293         slot_getallattrs(slot);
2294
2295         for (i = 0; i < tupdesc->natts; i++)
2296         {
2297                 bool            column_perm = false;
2298                 char       *val;
2299                 int                     vallen;
2300                 Form_pg_attribute att = TupleDescAttr(tupdesc, i);
2301
2302                 /* ignore dropped columns */
2303                 if (att->attisdropped)
2304                         continue;
2305
2306                 if (!table_perm)
2307                 {
2308                         /*
2309                          * No table-level SELECT, so need to make sure they either have
2310                          * SELECT rights on the column or that they have provided the data
2311                          * for the column.  If not, omit this column from the error
2312                          * message.
2313                          */
2314                         aclresult = pg_attribute_aclcheck(reloid, att->attnum,
2315                                                                                           GetUserId(), ACL_SELECT);
2316                         if (bms_is_member(att->attnum - FirstLowInvalidHeapAttributeNumber,
2317                                                           modifiedCols) || aclresult == ACLCHECK_OK)
2318                         {
2319                                 column_perm = any_perm = true;
2320
2321                                 if (write_comma_collist)
2322                                         appendStringInfoString(&collist, ", ");
2323                                 else
2324                                         write_comma_collist = true;
2325
2326                                 appendStringInfoString(&collist, NameStr(att->attname));
2327                         }
2328                 }
2329
2330                 if (table_perm || column_perm)
2331                 {
2332                         if (slot->tts_isnull[i])
2333                                 val = "null";
2334                         else
2335                         {
2336                                 Oid                     foutoid;
2337                                 bool            typisvarlena;
2338
2339                                 getTypeOutputInfo(att->atttypid,
2340                                                                   &foutoid, &typisvarlena);
2341                                 val = OidOutputFunctionCall(foutoid, slot->tts_values[i]);
2342                         }
2343
2344                         if (write_comma)
2345                                 appendStringInfoString(&buf, ", ");
2346                         else
2347                                 write_comma = true;
2348
2349                         /* truncate if needed */
2350                         vallen = strlen(val);
2351                         if (vallen <= maxfieldlen)
2352                                 appendStringInfoString(&buf, val);
2353                         else
2354                         {
2355                                 vallen = pg_mbcliplen(val, vallen, maxfieldlen);
2356                                 appendBinaryStringInfo(&buf, val, vallen);
2357                                 appendStringInfoString(&buf, "...");
2358                         }
2359                 }
2360         }
2361
2362         /* If we end up with zero columns being returned, then return NULL. */
2363         if (!any_perm)
2364                 return NULL;
2365
2366         appendStringInfoChar(&buf, ')');
2367
2368         if (!table_perm)
2369         {
2370                 appendStringInfoString(&collist, ") = ");
2371                 appendStringInfoString(&collist, buf.data);
2372
2373                 return collist.data;
2374         }
2375
2376         return buf.data;
2377 }
2378
2379
2380 /*
2381  * ExecUpdateLockMode -- find the appropriate UPDATE tuple lock mode for a
2382  * given ResultRelInfo
2383  */
2384 LockTupleMode
2385 ExecUpdateLockMode(EState *estate, ResultRelInfo *relinfo)
2386 {
2387         Bitmapset  *keyCols;
2388         Bitmapset  *updatedCols;
2389
2390         /*
2391          * Compute lock mode to use.  If columns that are part of the key have not
2392          * been modified, then we can use a weaker lock, allowing for better
2393          * concurrency.
2394          */
2395         updatedCols = GetUpdatedColumns(relinfo, estate);
2396         keyCols = RelationGetIndexAttrBitmap(relinfo->ri_RelationDesc,
2397                                                                                  INDEX_ATTR_BITMAP_KEY);
2398
2399         if (bms_overlap(keyCols, updatedCols))
2400                 return LockTupleExclusive;
2401
2402         return LockTupleNoKeyExclusive;
2403 }
2404
2405 /*
2406  * ExecFindRowMark -- find the ExecRowMark struct for given rangetable index
2407  *
2408  * If no such struct, either return NULL or throw error depending on missing_ok
2409  */
2410 ExecRowMark *
2411 ExecFindRowMark(EState *estate, Index rti, bool missing_ok)
2412 {
2413         ListCell   *lc;
2414
2415         foreach(lc, estate->es_rowMarks)
2416         {
2417                 ExecRowMark *erm = (ExecRowMark *) lfirst(lc);
2418
2419                 if (erm->rti == rti)
2420                         return erm;
2421         }
2422         if (!missing_ok)
2423                 elog(ERROR, "failed to find ExecRowMark for rangetable index %u", rti);
2424         return NULL;
2425 }
2426
2427 /*
2428  * ExecBuildAuxRowMark -- create an ExecAuxRowMark struct
2429  *
2430  * Inputs are the underlying ExecRowMark struct and the targetlist of the
2431  * input plan node (not planstate node!).  We need the latter to find out
2432  * the column numbers of the resjunk columns.
2433  */
2434 ExecAuxRowMark *
2435 ExecBuildAuxRowMark(ExecRowMark *erm, List *targetlist)
2436 {
2437         ExecAuxRowMark *aerm = (ExecAuxRowMark *) palloc0(sizeof(ExecAuxRowMark));
2438         char            resname[32];
2439
2440         aerm->rowmark = erm;
2441
2442         /* Look up the resjunk columns associated with this rowmark */
2443         if (erm->markType != ROW_MARK_COPY)
2444         {
2445                 /* need ctid for all methods other than COPY */
2446                 snprintf(resname, sizeof(resname), "ctid%u", erm->rowmarkId);
2447                 aerm->ctidAttNo = ExecFindJunkAttributeInTlist(targetlist,
2448                                                                                                            resname);
2449                 if (!AttributeNumberIsValid(aerm->ctidAttNo))
2450                         elog(ERROR, "could not find junk %s column", resname);
2451         }
2452         else
2453         {
2454                 /* need wholerow if COPY */
2455                 snprintf(resname, sizeof(resname), "wholerow%u", erm->rowmarkId);
2456                 aerm->wholeAttNo = ExecFindJunkAttributeInTlist(targetlist,
2457                                                                                                                 resname);
2458                 if (!AttributeNumberIsValid(aerm->wholeAttNo))
2459                         elog(ERROR, "could not find junk %s column", resname);
2460         }
2461
2462         /* if child rel, need tableoid */
2463         if (erm->rti != erm->prti)
2464         {
2465                 snprintf(resname, sizeof(resname), "tableoid%u", erm->rowmarkId);
2466                 aerm->toidAttNo = ExecFindJunkAttributeInTlist(targetlist,
2467                                                                                                            resname);
2468                 if (!AttributeNumberIsValid(aerm->toidAttNo))
2469                         elog(ERROR, "could not find junk %s column", resname);
2470         }
2471
2472         return aerm;
2473 }
2474
2475
2476 /*
2477  * EvalPlanQual logic --- recheck modified tuple(s) to see if we want to
2478  * process the updated version under READ COMMITTED rules.
2479  *
2480  * See backend/executor/README for some info about how this works.
2481  */
2482
2483
2484 /*
2485  * Check a modified tuple to see if we want to process its updated version
2486  * under READ COMMITTED rules.
2487  *
2488  *      estate - outer executor state data
2489  *      epqstate - state for EvalPlanQual rechecking
2490  *      relation - table containing tuple
2491  *      rti - rangetable index of table containing tuple
2492  *      lockmode - requested tuple lock mode
2493  *      *tid - t_ctid from the outdated tuple (ie, next updated version)
2494  *      priorXmax - t_xmax from the outdated tuple
2495  *
2496  * *tid is also an output parameter: it's modified to hold the TID of the
2497  * latest version of the tuple (note this may be changed even on failure)
2498  *
2499  * Returns a slot containing the new candidate update/delete tuple, or
2500  * NULL if we determine we shouldn't process the row.
2501  *
2502  * Note: properly, lockmode should be declared as enum LockTupleMode,
2503  * but we use "int" to avoid having to include heapam.h in executor.h.
2504  */
2505 TupleTableSlot *
2506 EvalPlanQual(EState *estate, EPQState *epqstate,
2507                          Relation relation, Index rti, int lockmode,
2508                          ItemPointer tid, TransactionId priorXmax)
2509 {
2510         TupleTableSlot *slot;
2511         HeapTuple       copyTuple;
2512
2513         Assert(rti > 0);
2514
2515         /*
2516          * Get and lock the updated version of the row; if fail, return NULL.
2517          */
2518         copyTuple = EvalPlanQualFetch(estate, relation, lockmode, LockWaitBlock,
2519                                                                   tid, priorXmax);
2520
2521         if (copyTuple == NULL)
2522                 return NULL;
2523
2524         /*
2525          * For UPDATE/DELETE we have to return tid of actual row we're executing
2526          * PQ for.
2527          */
2528         *tid = copyTuple->t_self;
2529
2530         /*
2531          * Need to run a recheck subquery.  Initialize or reinitialize EPQ state.
2532          */
2533         EvalPlanQualBegin(epqstate, estate);
2534
2535         /*
2536          * Free old test tuple, if any, and store new tuple where relation's scan
2537          * node will see it
2538          */
2539         EvalPlanQualSetTuple(epqstate, rti, copyTuple);
2540
2541         /*
2542          * Fetch any non-locked source rows
2543          */
2544         EvalPlanQualFetchRowMarks(epqstate);
2545
2546         /*
2547          * Run the EPQ query.  We assume it will return at most one tuple.
2548          */
2549         slot = EvalPlanQualNext(epqstate);
2550
2551         /*
2552          * If we got a tuple, force the slot to materialize the tuple so that it
2553          * is not dependent on any local state in the EPQ query (in particular,
2554          * it's highly likely that the slot contains references to any pass-by-ref
2555          * datums that may be present in copyTuple).  As with the next step, this
2556          * is to guard against early re-use of the EPQ query.
2557          */
2558         if (!TupIsNull(slot))
2559                 (void) ExecMaterializeSlot(slot);
2560
2561         /*
2562          * Clear out the test tuple.  This is needed in case the EPQ query is
2563          * re-used to test a tuple for a different relation.  (Not clear that can
2564          * really happen, but let's be safe.)
2565          */
2566         EvalPlanQualSetTuple(epqstate, rti, NULL);
2567
2568         return slot;
2569 }
2570
2571 /*
2572  * Fetch a copy of the newest version of an outdated tuple
2573  *
2574  *      estate - executor state data
2575  *      relation - table containing tuple
2576  *      lockmode - requested tuple lock mode
2577  *      wait_policy - requested lock wait policy
2578  *      *tid - t_ctid from the outdated tuple (ie, next updated version)
2579  *      priorXmax - t_xmax from the outdated tuple
2580  *
2581  * Returns a palloc'd copy of the newest tuple version, or NULL if we find
2582  * that there is no newest version (ie, the row was deleted not updated).
2583  * We also return NULL if the tuple is locked and the wait policy is to skip
2584  * such tuples.
2585  *
2586  * If successful, we have locked the newest tuple version, so caller does not
2587  * need to worry about it changing anymore.
2588  *
2589  * Note: properly, lockmode should be declared as enum LockTupleMode,
2590  * but we use "int" to avoid having to include heapam.h in executor.h.
2591  */
2592 HeapTuple
2593 EvalPlanQualFetch(EState *estate, Relation relation, int lockmode,
2594                                   LockWaitPolicy wait_policy,
2595                                   ItemPointer tid, TransactionId priorXmax)
2596 {
2597         HeapTuple       copyTuple = NULL;
2598         HeapTupleData tuple;
2599         SnapshotData SnapshotDirty;
2600
2601         /*
2602          * fetch target tuple
2603          *
2604          * Loop here to deal with updated or busy tuples
2605          */
2606         InitDirtySnapshot(SnapshotDirty);
2607         tuple.t_self = *tid;
2608         for (;;)
2609         {
2610                 Buffer          buffer;
2611
2612                 if (heap_fetch(relation, &SnapshotDirty, &tuple, &buffer, true, NULL))
2613                 {
2614                         HTSU_Result test;
2615                         HeapUpdateFailureData hufd;
2616
2617                         /*
2618                          * If xmin isn't what we're expecting, the slot must have been
2619                          * recycled and reused for an unrelated tuple.  This implies that
2620                          * the latest version of the row was deleted, so we need do
2621                          * nothing.  (Should be safe to examine xmin without getting
2622                          * buffer's content lock.  We assume reading a TransactionId to be
2623                          * atomic, and Xmin never changes in an existing tuple, except to
2624                          * invalid or frozen, and neither of those can match priorXmax.)
2625                          */
2626                         if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
2627                                                                          priorXmax))
2628                         {
2629                                 ReleaseBuffer(buffer);
2630                                 return NULL;
2631                         }
2632
2633                         /* otherwise xmin should not be dirty... */
2634                         if (TransactionIdIsValid(SnapshotDirty.xmin))
2635                                 elog(ERROR, "t_xmin is uncommitted in tuple to be updated");
2636
2637                         /*
2638                          * If tuple is being updated by other transaction then we have to
2639                          * wait for its commit/abort, or die trying.
2640                          */
2641                         if (TransactionIdIsValid(SnapshotDirty.xmax))
2642                         {
2643                                 ReleaseBuffer(buffer);
2644                                 switch (wait_policy)
2645                                 {
2646                                         case LockWaitBlock:
2647                                                 XactLockTableWait(SnapshotDirty.xmax,
2648                                                                                   relation, &tuple.t_self,
2649                                                                                   XLTW_FetchUpdated);
2650                                                 break;
2651                                         case LockWaitSkip:
2652                                                 if (!ConditionalXactLockTableWait(SnapshotDirty.xmax))
2653                                                         return NULL;    /* skip instead of waiting */
2654                                                 break;
2655                                         case LockWaitError:
2656                                                 if (!ConditionalXactLockTableWait(SnapshotDirty.xmax))
2657                                                         ereport(ERROR,
2658                                                                         (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
2659                                                                          errmsg("could not obtain lock on row in relation \"%s\"",
2660                                                                                         RelationGetRelationName(relation))));
2661                                                 break;
2662                                 }
2663                                 continue;               /* loop back to repeat heap_fetch */
2664                         }
2665
2666                         /*
2667                          * If tuple was inserted by our own transaction, we have to check
2668                          * cmin against es_output_cid: cmin >= current CID means our
2669                          * command cannot see the tuple, so we should ignore it. Otherwise
2670                          * heap_lock_tuple() will throw an error, and so would any later
2671                          * attempt to update or delete the tuple.  (We need not check cmax
2672                          * because HeapTupleSatisfiesDirty will consider a tuple deleted
2673                          * by our transaction dead, regardless of cmax.) We just checked
2674                          * that priorXmax == xmin, so we can test that variable instead of
2675                          * doing HeapTupleHeaderGetXmin again.
2676                          */
2677                         if (TransactionIdIsCurrentTransactionId(priorXmax) &&
2678                                 HeapTupleHeaderGetCmin(tuple.t_data) >= estate->es_output_cid)
2679                         {
2680                                 ReleaseBuffer(buffer);
2681                                 return NULL;
2682                         }
2683
2684                         /*
2685                          * This is a live tuple, so now try to lock it.
2686                          */
2687                         test = heap_lock_tuple(relation, &tuple,
2688                                                                    estate->es_output_cid,
2689                                                                    lockmode, wait_policy,
2690                                                                    false, &buffer, &hufd);
2691                         /* We now have two pins on the buffer, get rid of one */
2692                         ReleaseBuffer(buffer);
2693
2694                         switch (test)
2695                         {
2696                                 case HeapTupleSelfUpdated:
2697
2698                                         /*
2699                                          * The target tuple was already updated or deleted by the
2700                                          * current command, or by a later command in the current
2701                                          * transaction.  We *must* ignore the tuple in the former
2702                                          * case, so as to avoid the "Halloween problem" of
2703                                          * repeated update attempts.  In the latter case it might
2704                                          * be sensible to fetch the updated tuple instead, but
2705                                          * doing so would require changing heap_update and
2706                                          * heap_delete to not complain about updating "invisible"
2707                                          * tuples, which seems pretty scary (heap_lock_tuple will
2708                                          * not complain, but few callers expect
2709                                          * HeapTupleInvisible, and we're not one of them).  So for
2710                                          * now, treat the tuple as deleted and do not process.
2711                                          */
2712                                         ReleaseBuffer(buffer);
2713                                         return NULL;
2714
2715                                 case HeapTupleMayBeUpdated:
2716                                         /* successfully locked */
2717                                         break;
2718
2719                                 case HeapTupleUpdated:
2720                                         ReleaseBuffer(buffer);
2721                                         if (IsolationUsesXactSnapshot())
2722                                                 ereport(ERROR,
2723                                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
2724                                                                  errmsg("could not serialize access due to concurrent update")));
2725
2726                                         /* Should not encounter speculative tuple on recheck */
2727                                         Assert(!HeapTupleHeaderIsSpeculative(tuple.t_data));
2728                                         if (!ItemPointerEquals(&hufd.ctid, &tuple.t_self))
2729                                         {
2730                                                 /* it was updated, so look at the updated version */
2731                                                 tuple.t_self = hufd.ctid;
2732                                                 /* updated row should have xmin matching this xmax */
2733                                                 priorXmax = hufd.xmax;
2734                                                 continue;
2735                                         }
2736                                         /* tuple was deleted, so give up */
2737                                         return NULL;
2738
2739                                 case HeapTupleWouldBlock:
2740                                         ReleaseBuffer(buffer);
2741                                         return NULL;
2742
2743                                 case HeapTupleInvisible:
2744                                         elog(ERROR, "attempted to lock invisible tuple");
2745
2746                                 default:
2747                                         ReleaseBuffer(buffer);
2748                                         elog(ERROR, "unrecognized heap_lock_tuple status: %u",
2749                                                  test);
2750                                         return NULL;    /* keep compiler quiet */
2751                         }
2752
2753                         /*
2754                          * We got tuple - now copy it for use by recheck query.
2755                          */
2756                         copyTuple = heap_copytuple(&tuple);
2757                         ReleaseBuffer(buffer);
2758                         break;
2759                 }
2760
2761                 /*
2762                  * If the referenced slot was actually empty, the latest version of
2763                  * the row must have been deleted, so we need do nothing.
2764                  */
2765                 if (tuple.t_data == NULL)
2766                 {
2767                         ReleaseBuffer(buffer);
2768                         return NULL;
2769                 }
2770
2771                 /*
2772                  * As above, if xmin isn't what we're expecting, do nothing.
2773                  */
2774                 if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
2775                                                                  priorXmax))
2776                 {
2777                         ReleaseBuffer(buffer);
2778                         return NULL;
2779                 }
2780
2781                 /*
2782                  * If we get here, the tuple was found but failed SnapshotDirty.
2783                  * Assuming the xmin is either a committed xact or our own xact (as it
2784                  * certainly should be if we're trying to modify the tuple), this must
2785                  * mean that the row was updated or deleted by either a committed xact
2786                  * or our own xact.  If it was deleted, we can ignore it; if it was
2787                  * updated then chain up to the next version and repeat the whole
2788                  * process.
2789                  *
2790                  * As above, it should be safe to examine xmax and t_ctid without the
2791                  * buffer content lock, because they can't be changing.
2792                  */
2793                 if (ItemPointerEquals(&tuple.t_self, &tuple.t_data->t_ctid))
2794                 {
2795                         /* deleted, so forget about it */
2796                         ReleaseBuffer(buffer);
2797                         return NULL;
2798                 }
2799
2800                 /* updated, so look at the updated row */
2801                 tuple.t_self = tuple.t_data->t_ctid;
2802                 /* updated row should have xmin matching this xmax */
2803                 priorXmax = HeapTupleHeaderGetUpdateXid(tuple.t_data);
2804                 ReleaseBuffer(buffer);
2805                 /* loop back to fetch next in chain */
2806         }
2807
2808         /*
2809          * Return the copied tuple
2810          */
2811         return copyTuple;
2812 }
2813
2814 /*
2815  * EvalPlanQualInit -- initialize during creation of a plan state node
2816  * that might need to invoke EPQ processing.
2817  *
2818  * Note: subplan/auxrowmarks can be NULL/NIL if they will be set later
2819  * with EvalPlanQualSetPlan.
2820  */
2821 void
2822 EvalPlanQualInit(EPQState *epqstate, EState *estate,
2823                                  Plan *subplan, List *auxrowmarks, int epqParam)
2824 {
2825         /* Mark the EPQ state inactive */
2826         epqstate->estate = NULL;
2827         epqstate->planstate = NULL;
2828         epqstate->origslot = NULL;
2829         /* ... and remember data that EvalPlanQualBegin will need */
2830         epqstate->plan = subplan;
2831         epqstate->arowMarks = auxrowmarks;
2832         epqstate->epqParam = epqParam;
2833 }
2834
2835 /*
2836  * EvalPlanQualSetPlan -- set or change subplan of an EPQState.
2837  *
2838  * We need this so that ModifyTable can deal with multiple subplans.
2839  */
2840 void
2841 EvalPlanQualSetPlan(EPQState *epqstate, Plan *subplan, List *auxrowmarks)
2842 {
2843         /* If we have a live EPQ query, shut it down */
2844         EvalPlanQualEnd(epqstate);
2845         /* And set/change the plan pointer */
2846         epqstate->plan = subplan;
2847         /* The rowmarks depend on the plan, too */
2848         epqstate->arowMarks = auxrowmarks;
2849 }
2850
2851 /*
2852  * Install one test tuple into EPQ state, or clear test tuple if tuple == NULL
2853  *
2854  * NB: passed tuple must be palloc'd; it may get freed later
2855  */
2856 void
2857 EvalPlanQualSetTuple(EPQState *epqstate, Index rti, HeapTuple tuple)
2858 {
2859         EState     *estate = epqstate->estate;
2860
2861         Assert(rti > 0);
2862
2863         /*
2864          * free old test tuple, if any, and store new tuple where relation's scan
2865          * node will see it
2866          */
2867         if (estate->es_epqTuple[rti - 1] != NULL)
2868                 heap_freetuple(estate->es_epqTuple[rti - 1]);
2869         estate->es_epqTuple[rti - 1] = tuple;
2870         estate->es_epqTupleSet[rti - 1] = true;
2871 }
2872
2873 /*
2874  * Fetch back the current test tuple (if any) for the specified RTI
2875  */
2876 HeapTuple
2877 EvalPlanQualGetTuple(EPQState *epqstate, Index rti)
2878 {
2879         EState     *estate = epqstate->estate;
2880
2881         Assert(rti > 0);
2882
2883         return estate->es_epqTuple[rti - 1];
2884 }
2885
2886 /*
2887  * Fetch the current row values for any non-locked relations that need
2888  * to be scanned by an EvalPlanQual operation.  origslot must have been set
2889  * to contain the current result row (top-level row) that we need to recheck.
2890  */
2891 void
2892 EvalPlanQualFetchRowMarks(EPQState *epqstate)
2893 {
2894         ListCell   *l;
2895
2896         Assert(epqstate->origslot != NULL);
2897
2898         foreach(l, epqstate->arowMarks)
2899         {
2900                 ExecAuxRowMark *aerm = (ExecAuxRowMark *) lfirst(l);
2901                 ExecRowMark *erm = aerm->rowmark;
2902                 Datum           datum;
2903                 bool            isNull;
2904                 HeapTupleData tuple;
2905
2906                 if (RowMarkRequiresRowShareLock(erm->markType))
2907                         elog(ERROR, "EvalPlanQual doesn't support locking rowmarks");
2908
2909                 /* clear any leftover test tuple for this rel */
2910                 EvalPlanQualSetTuple(epqstate, erm->rti, NULL);
2911
2912                 /* if child rel, must check whether it produced this row */
2913                 if (erm->rti != erm->prti)
2914                 {
2915                         Oid                     tableoid;
2916
2917                         datum = ExecGetJunkAttribute(epqstate->origslot,
2918                                                                                  aerm->toidAttNo,
2919                                                                                  &isNull);
2920                         /* non-locked rels could be on the inside of outer joins */
2921                         if (isNull)
2922                                 continue;
2923                         tableoid = DatumGetObjectId(datum);
2924
2925                         Assert(OidIsValid(erm->relid));
2926                         if (tableoid != erm->relid)
2927                         {
2928                                 /* this child is inactive right now */
2929                                 continue;
2930                         }
2931                 }
2932
2933                 if (erm->markType == ROW_MARK_REFERENCE)
2934                 {
2935                         HeapTuple       copyTuple;
2936
2937                         Assert(erm->relation != NULL);
2938
2939                         /* fetch the tuple's ctid */
2940                         datum = ExecGetJunkAttribute(epqstate->origslot,
2941                                                                                  aerm->ctidAttNo,
2942                                                                                  &isNull);
2943                         /* non-locked rels could be on the inside of outer joins */
2944                         if (isNull)
2945                                 continue;
2946
2947                         /* fetch requests on foreign tables must be passed to their FDW */
2948                         if (erm->relation->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
2949                         {
2950                                 FdwRoutine *fdwroutine;
2951                                 bool            updated = false;
2952
2953                                 fdwroutine = GetFdwRoutineForRelation(erm->relation, false);
2954                                 /* this should have been checked already, but let's be safe */
2955                                 if (fdwroutine->RefetchForeignRow == NULL)
2956                                         ereport(ERROR,
2957                                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2958                                                          errmsg("cannot lock rows in foreign table \"%s\"",
2959                                                                         RelationGetRelationName(erm->relation))));
2960                                 copyTuple = fdwroutine->RefetchForeignRow(epqstate->estate,
2961                                                                                                                   erm,
2962                                                                                                                   datum,
2963                                                                                                                   &updated);
2964                                 if (copyTuple == NULL)
2965                                         elog(ERROR, "failed to fetch tuple for EvalPlanQual recheck");
2966
2967                                 /*
2968                                  * Ideally we'd insist on updated == false here, but that
2969                                  * assumes that FDWs can track that exactly, which they might
2970                                  * not be able to.  So just ignore the flag.
2971                                  */
2972                         }
2973                         else
2974                         {
2975                                 /* ordinary table, fetch the tuple */
2976                                 Buffer          buffer;
2977
2978                                 tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
2979                                 if (!heap_fetch(erm->relation, SnapshotAny, &tuple, &buffer,
2980                                                                 false, NULL))
2981                                         elog(ERROR, "failed to fetch tuple for EvalPlanQual recheck");
2982
2983                                 /* successful, copy tuple */
2984                                 copyTuple = heap_copytuple(&tuple);
2985                                 ReleaseBuffer(buffer);
2986                         }
2987
2988                         /* store tuple */
2989                         EvalPlanQualSetTuple(epqstate, erm->rti, copyTuple);
2990                 }
2991                 else
2992                 {
2993                         HeapTupleHeader td;
2994
2995                         Assert(erm->markType == ROW_MARK_COPY);
2996
2997                         /* fetch the whole-row Var for the relation */
2998                         datum = ExecGetJunkAttribute(epqstate->origslot,
2999                                                                                  aerm->wholeAttNo,
3000                                                                                  &isNull);
3001                         /* non-locked rels could be on the inside of outer joins */
3002                         if (isNull)
3003                                 continue;
3004                         td = DatumGetHeapTupleHeader(datum);
3005
3006                         /* build a temporary HeapTuple control structure */
3007                         tuple.t_len = HeapTupleHeaderGetDatumLength(td);
3008                         tuple.t_data = td;
3009                         /* relation might be a foreign table, if so provide tableoid */
3010                         tuple.t_tableOid = erm->relid;
3011                         /* also copy t_ctid in case there's valid data there */
3012                         tuple.t_self = td->t_ctid;
3013
3014                         /* copy and store tuple */
3015                         EvalPlanQualSetTuple(epqstate, erm->rti,
3016                                                                  heap_copytuple(&tuple));
3017                 }
3018         }
3019 }
3020
3021 /*
3022  * Fetch the next row (if any) from EvalPlanQual testing
3023  *
3024  * (In practice, there should never be more than one row...)
3025  */
3026 TupleTableSlot *
3027 EvalPlanQualNext(EPQState *epqstate)
3028 {
3029         MemoryContext oldcontext;
3030         TupleTableSlot *slot;
3031
3032         oldcontext = MemoryContextSwitchTo(epqstate->estate->es_query_cxt);
3033         slot = ExecProcNode(epqstate->planstate);
3034         MemoryContextSwitchTo(oldcontext);
3035
3036         return slot;
3037 }
3038
3039 /*
3040  * Initialize or reset an EvalPlanQual state tree
3041  */
3042 void
3043 EvalPlanQualBegin(EPQState *epqstate, EState *parentestate)
3044 {
3045         EState     *estate = epqstate->estate;
3046
3047         if (estate == NULL)
3048         {
3049                 /* First time through, so create a child EState */
3050                 EvalPlanQualStart(epqstate, parentestate, epqstate->plan);
3051         }
3052         else
3053         {
3054                 /*
3055                  * We already have a suitable child EPQ tree, so just reset it.
3056                  */
3057                 int                     rtsize = list_length(parentestate->es_range_table);
3058                 PlanState  *planstate = epqstate->planstate;
3059
3060                 MemSet(estate->es_epqScanDone, 0, rtsize * sizeof(bool));
3061
3062                 /* Recopy current values of parent parameters */
3063                 if (parentestate->es_plannedstmt->paramExecTypes != NIL)
3064                 {
3065                         int                     i;
3066
3067                         i = list_length(parentestate->es_plannedstmt->paramExecTypes);
3068
3069                         while (--i >= 0)
3070                         {
3071                                 /* copy value if any, but not execPlan link */
3072                                 estate->es_param_exec_vals[i].value =
3073                                         parentestate->es_param_exec_vals[i].value;
3074                                 estate->es_param_exec_vals[i].isnull =
3075                                         parentestate->es_param_exec_vals[i].isnull;
3076                         }
3077                 }
3078
3079                 /*
3080                  * Mark child plan tree as needing rescan at all scan nodes.  The
3081                  * first ExecProcNode will take care of actually doing the rescan.
3082                  */
3083                 planstate->chgParam = bms_add_member(planstate->chgParam,
3084                                                                                          epqstate->epqParam);
3085         }
3086 }
3087
3088 /*
3089  * Start execution of an EvalPlanQual plan tree.
3090  *
3091  * This is a cut-down version of ExecutorStart(): we copy some state from
3092  * the top-level estate rather than initializing it fresh.
3093  */
3094 static void
3095 EvalPlanQualStart(EPQState *epqstate, EState *parentestate, Plan *planTree)
3096 {
3097         EState     *estate;
3098         int                     rtsize;
3099         MemoryContext oldcontext;
3100         ListCell   *l;
3101
3102         rtsize = list_length(parentestate->es_range_table);
3103
3104         epqstate->estate = estate = CreateExecutorState();
3105
3106         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
3107
3108         /*
3109          * Child EPQ EStates share the parent's copy of unchanging state such as
3110          * the snapshot, rangetable, result-rel info, and external Param info.
3111          * They need their own copies of local state, including a tuple table,
3112          * es_param_exec_vals, etc.
3113          *
3114          * The ResultRelInfo array management is trickier than it looks.  We
3115          * create a fresh array for the child but copy all the content from the
3116          * parent.  This is because it's okay for the child to share any
3117          * per-relation state the parent has already created --- but if the child
3118          * sets up any ResultRelInfo fields, such as its own junkfilter, that
3119          * state must *not* propagate back to the parent.  (For one thing, the
3120          * pointed-to data is in a memory context that won't last long enough.)
3121          */
3122         estate->es_direction = ForwardScanDirection;
3123         estate->es_snapshot = parentestate->es_snapshot;
3124         estate->es_crosscheck_snapshot = parentestate->es_crosscheck_snapshot;
3125         estate->es_range_table = parentestate->es_range_table;
3126         estate->es_plannedstmt = parentestate->es_plannedstmt;
3127         estate->es_junkFilter = parentestate->es_junkFilter;
3128         estate->es_output_cid = parentestate->es_output_cid;
3129         if (parentestate->es_num_result_relations > 0)
3130         {
3131                 int                     numResultRelations = parentestate->es_num_result_relations;
3132                 ResultRelInfo *resultRelInfos;
3133
3134                 resultRelInfos = (ResultRelInfo *)
3135                         palloc(numResultRelations * sizeof(ResultRelInfo));
3136                 memcpy(resultRelInfos, parentestate->es_result_relations,
3137                            numResultRelations * sizeof(ResultRelInfo));
3138                 estate->es_result_relations = resultRelInfos;
3139                 estate->es_num_result_relations = numResultRelations;
3140         }
3141         /* es_result_relation_info must NOT be copied */
3142         /* es_trig_target_relations must NOT be copied */
3143         estate->es_rowMarks = parentestate->es_rowMarks;
3144         estate->es_top_eflags = parentestate->es_top_eflags;
3145         estate->es_instrument = parentestate->es_instrument;
3146         /* es_auxmodifytables must NOT be copied */
3147
3148         /*
3149          * The external param list is simply shared from parent.  The internal
3150          * param workspace has to be local state, but we copy the initial values
3151          * from the parent, so as to have access to any param values that were
3152          * already set from other parts of the parent's plan tree.
3153          */
3154         estate->es_param_list_info = parentestate->es_param_list_info;
3155         if (parentestate->es_plannedstmt->paramExecTypes != NIL)
3156         {
3157                 int                     i;
3158
3159                 i = list_length(parentestate->es_plannedstmt->paramExecTypes);
3160                 estate->es_param_exec_vals = (ParamExecData *)
3161                         palloc0(i * sizeof(ParamExecData));
3162                 while (--i >= 0)
3163                 {
3164                         /* copy value if any, but not execPlan link */
3165                         estate->es_param_exec_vals[i].value =
3166                                 parentestate->es_param_exec_vals[i].value;
3167                         estate->es_param_exec_vals[i].isnull =
3168                                 parentestate->es_param_exec_vals[i].isnull;
3169                 }
3170         }
3171
3172         /*
3173          * Each EState must have its own es_epqScanDone state, but if we have
3174          * nested EPQ checks they should share es_epqTuple arrays.  This allows
3175          * sub-rechecks to inherit the values being examined by an outer recheck.
3176          */
3177         estate->es_epqScanDone = (bool *) palloc0(rtsize * sizeof(bool));
3178         if (parentestate->es_epqTuple != NULL)
3179         {
3180                 estate->es_epqTuple = parentestate->es_epqTuple;
3181                 estate->es_epqTupleSet = parentestate->es_epqTupleSet;
3182         }
3183         else
3184         {
3185                 estate->es_epqTuple = (HeapTuple *)
3186                         palloc0(rtsize * sizeof(HeapTuple));
3187                 estate->es_epqTupleSet = (bool *)
3188                         palloc0(rtsize * sizeof(bool));
3189         }
3190
3191         /*
3192          * Each estate also has its own tuple table.
3193          */
3194         estate->es_tupleTable = NIL;
3195
3196         /*
3197          * Initialize private state information for each SubPlan.  We must do this
3198          * before running ExecInitNode on the main query tree, since
3199          * ExecInitSubPlan expects to be able to find these entries. Some of the
3200          * SubPlans might not be used in the part of the plan tree we intend to
3201          * run, but since it's not easy to tell which, we just initialize them
3202          * all.
3203          */
3204         Assert(estate->es_subplanstates == NIL);
3205         foreach(l, parentestate->es_plannedstmt->subplans)
3206         {
3207                 Plan       *subplan = (Plan *) lfirst(l);
3208                 PlanState  *subplanstate;
3209
3210                 subplanstate = ExecInitNode(subplan, estate, 0);
3211                 estate->es_subplanstates = lappend(estate->es_subplanstates,
3212                                                                                    subplanstate);
3213         }
3214
3215         /*
3216          * Initialize the private state information for all the nodes in the part
3217          * of the plan tree we need to run.  This opens files, allocates storage
3218          * and leaves us ready to start processing tuples.
3219          */
3220         epqstate->planstate = ExecInitNode(planTree, estate, 0);
3221
3222         MemoryContextSwitchTo(oldcontext);
3223 }
3224
3225 /*
3226  * EvalPlanQualEnd -- shut down at termination of parent plan state node,
3227  * or if we are done with the current EPQ child.
3228  *
3229  * This is a cut-down version of ExecutorEnd(); basically we want to do most
3230  * of the normal cleanup, but *not* close result relations (which we are
3231  * just sharing from the outer query).  We do, however, have to close any
3232  * trigger target relations that got opened, since those are not shared.
3233  * (There probably shouldn't be any of the latter, but just in case...)
3234  */
3235 void
3236 EvalPlanQualEnd(EPQState *epqstate)
3237 {
3238         EState     *estate = epqstate->estate;
3239         MemoryContext oldcontext;
3240         ListCell   *l;
3241
3242         if (estate == NULL)
3243                 return;                                 /* idle, so nothing to do */
3244
3245         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
3246
3247         ExecEndNode(epqstate->planstate);
3248
3249         foreach(l, estate->es_subplanstates)
3250         {
3251                 PlanState  *subplanstate = (PlanState *) lfirst(l);
3252
3253                 ExecEndNode(subplanstate);
3254         }
3255
3256         /* throw away the per-estate tuple table */
3257         ExecResetTupleTable(estate->es_tupleTable, false);
3258
3259         /* close any trigger target relations attached to this EState */
3260         ExecCleanUpTriggerState(estate);
3261
3262         MemoryContextSwitchTo(oldcontext);
3263
3264         FreeExecutorState(estate);
3265
3266         /* Mark EPQState idle */
3267         epqstate->estate = NULL;
3268         epqstate->planstate = NULL;
3269         epqstate->origslot = NULL;
3270 }