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