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) 1996-2017, PostgreSQL Global Development Group
  30  * Portions Copyright (c) 1994, Regents of the University of California
  31  *
  32  *
  33  * IDENTIFICATION
  34  *        src/backend/executor/execMain.c
  35  *
  36  *-------------------------------------------------------------------------
  37  */
  38 #include "postgres.h"
  39
  40 #include "access/htup_details.h"
  41 #include "access/sysattr.h"
  42 #include "access/transam.h"
  43 #include "access/xact.h"
  44 #include "catalog/namespace.h"
  45 #include "catalog/partition.h"
  46 #include "catalog/pg_publication.h"
  47 #include "commands/matview.h"
  48 #include "commands/trigger.h"
  49 #include "executor/execdebug.h"
  50 #include "foreign/fdwapi.h"
  51 #include "mb/pg_wchar.h"
  52 #include "miscadmin.h"
  53 #include "optimizer/clauses.h"
  54 #include "parser/parsetree.h"
  55 #include "rewrite/rewriteManip.h"
  56 #include "storage/bufmgr.h"
  57 #include "storage/lmgr.h"
  58 #include "tcop/utility.h"
  59 #include "utils/acl.h"
  60 #include "utils/lsyscache.h"
  61 #include "utils/memutils.h"
  62 #include "utils/rls.h"
  63 #include "utils/ruleutils.h"
  64 #include "utils/snapmgr.h"
  65 #include "utils/tqual.h"
  66
  67
  68 /* Hooks for plugins to get control in ExecutorStart/Run/Finish/End */
  69 ExecutorStart_hook_type ExecutorStart_hook = NULL;
  70 ExecutorRun_hook_type ExecutorRun_hook = NULL;
  71 ExecutorFinish_hook_type ExecutorFinish_hook = NULL;
  72 ExecutorEnd_hook_type ExecutorEnd_hook = NULL;
  73
  74 /* Hook for plugin to get control in ExecCheckRTPerms() */
  75 ExecutorCheckPerms_hook_type ExecutorCheckPerms_hook = NULL;
  76
  77 /* decls for local routines only used within this module */
  78 static void InitPlan(QueryDesc *queryDesc, int eflags);
  79 static void CheckValidRowMarkRel(Relation rel, RowMarkType markType);
  80 static void ExecPostprocessPlan(EState *estate);
  81 static void ExecEndPlan(PlanState *planstate, EState *estate);
  82 static void ExecutePlan(EState *estate, PlanState *planstate,
  83                         bool use_parallel_mode,
  84                         CmdType operation,
  85                         bool sendTuples,
  86                         uint64 numberTuples,
  87                         ScanDirection direction,
  88                         DestReceiver *dest,
  89                         bool execute_once);
  90 static bool ExecCheckRTEPerms(RangeTblEntry *rte);
  91 static bool ExecCheckRTEPermsModified(Oid relOid, Oid userid,
  92                                                   Bitmapset *modifiedCols,
  93                                                   AclMode requiredPerms);
  94 static void ExecCheckXactReadOnly(PlannedStmt *plannedstmt);
  95 static char *ExecBuildSlotValueDescription(Oid reloid,
  96                                                           TupleTableSlot *slot,
  97                                                           TupleDesc tupdesc,
  98                                                           Bitmapset *modifiedCols,
  99                                                           int maxfieldlen);
 100 static char *ExecBuildSlotPartitionKeyDescription(Relation rel,
 101                                                                          Datum *values,
 102                                                                          bool *isnull,
 103                                                                          int maxfieldlen);
 104 static void EvalPlanQualStart(EPQState *epqstate, EState *parentestate,
 105                                   Plan *planTree);
 106 static void ExecPartitionCheck(ResultRelInfo *resultRelInfo,
 107                                    TupleTableSlot *slot, EState *estate);
 108
 109 /*
 110  * Note that GetUpdatedColumns() also exists in commands/trigger.c.  There does
 111  * not appear to be any good header to put it into, given the structures that
 112  * it uses, so we let them be duplicated.  Be sure to update both if one needs
 113  * to be changed, however.
 114  */
 115 #define GetInsertedColumns(relinfo, estate) \
 116         (rt_fetch((relinfo)->ri_RangeTableIndex, (estate)->es_range_table)->insertedCols)
 117 #define GetUpdatedColumns(relinfo, estate) \
 118         (rt_fetch((relinfo)->ri_RangeTableIndex, (estate)->es_range_table)->updatedCols)
 119
 120 /* end of local decls */
 121
 122
 123 /* ----------------------------------------------------------------
 124  *              ExecutorStart
 125  *
 126  *              This routine must be called at the beginning of any execution of any
 127  *              query plan
 128  *
 129  * Takes a QueryDesc previously created by CreateQueryDesc (which is separate
 130  * only because some places use QueryDescs for utility commands).  The tupDesc
 131  * field of the QueryDesc is filled in to describe the tuples that will be
 132  * returned, and the internal fields (estate and planstate) are set up.
 133  *
 134  * eflags contains flag bits as described in executor.h.
 135  *
 136  * NB: the CurrentMemoryContext when this is called will become the parent
 137  * of the per-query context used for this Executor invocation.
 138  *
 139  * We provide a function hook variable that lets loadable plugins
 140  * get control when ExecutorStart is called.  Such a plugin would
 141  * normally call standard_ExecutorStart().
 142  *
 143  * ----------------------------------------------------------------
 144  */
 145 void
 146 ExecutorStart(QueryDesc *queryDesc, int eflags)
 147 {
 148         if (ExecutorStart_hook)
 149                 (*ExecutorStart_hook) (queryDesc, eflags);
 150         else
 151                 standard_ExecutorStart(queryDesc, eflags);
 152 }
 153
 154 void
 155 standard_ExecutorStart(QueryDesc *queryDesc, int eflags)
 156 {
 157         EState     *estate;
 158         MemoryContext oldcontext;
 159
 160         /* sanity checks: queryDesc must not be started already */
 161         Assert(queryDesc != NULL);
 162         Assert(queryDesc->estate == NULL);
 163
 164         /*
 165          * If the transaction is read-only, we need to check if any writes are
 166          * planned to non-temporary tables.  EXPLAIN is considered read-only.
 167          *
 168          * Don't allow writes in parallel mode.  Supporting UPDATE and DELETE
 169          * would require (a) storing the combocid hash in shared memory, rather
 170          * than synchronizing it just once at the start of parallelism, and (b) an
 171          * alternative to heap_update()'s reliance on xmax for mutual exclusion.
 172          * INSERT may have no such troubles, but we forbid it to simplify the
 173          * checks.
 174          *
 175          * We have lower-level defenses in CommandCounterIncrement and elsewhere
 176          * against performing unsafe operations in parallel mode, but this gives a
 177          * more user-friendly error message.
 178          */
 179         if ((XactReadOnly || IsInParallelMode()) &&
 180                 !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
 181                 ExecCheckXactReadOnly(queryDesc->plannedstmt);
 182
 183         /*
 184          * Build EState, switch into per-query memory context for startup.
 185          */
 186         estate = CreateExecutorState();
 187         queryDesc->estate = estate;
 188
 189         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 190
 191         /*
 192          * Fill in external parameters, if any, from queryDesc; and allocate
 193          * workspace for internal parameters
 194          */
 195         estate->es_param_list_info = queryDesc->params;
 196
 197         if (queryDesc->plannedstmt->nParamExec > 0)
 198                 estate->es_param_exec_vals = (ParamExecData *)
 199                         palloc0(queryDesc->plannedstmt->nParamExec * sizeof(ParamExecData));
 200
 201         estate->es_sourceText = queryDesc->sourceText;
 202
 203         /*
 204          * Fill in the query environment, if any, from queryDesc.
 205          */
 206         estate->es_queryEnv = queryDesc->queryEnv;
 207
 208         /*
 209          * If non-read-only query, set the command ID to mark output tuples with
 210          */
 211         switch (queryDesc->operation)
 212         {
 213                 case CMD_SELECT:
 214
 215                         /*
 216                          * SELECT FOR [KEY] UPDATE/SHARE and modifying CTEs need to mark
 217                          * tuples
 218                          */
 219                         if (queryDesc->plannedstmt->rowMarks != NIL ||
 220                                 queryDesc->plannedstmt->hasModifyingCTE)
 221                                 estate->es_output_cid = GetCurrentCommandId(true);
 222
 223                         /*
 224                          * A SELECT without modifying CTEs can't possibly queue triggers,
 225                          * so force skip-triggers mode. This is just a marginal efficiency
 226                          * hack, since AfterTriggerBeginQuery/AfterTriggerEndQuery aren't
 227                          * all that expensive, but we might as well do it.
 228                          */
 229                         if (!queryDesc->plannedstmt->hasModifyingCTE)
 230                                 eflags |= EXEC_FLAG_SKIP_TRIGGERS;
 231                         break;
 232
 233                 case CMD_INSERT:
 234                 case CMD_DELETE:
 235                 case CMD_UPDATE:
 236                         estate->es_output_cid = GetCurrentCommandId(true);
 237                         break;
 238
 239                 default:
 240                         elog(ERROR, "unrecognized operation code: %d",
 241                                  (int) queryDesc->operation);
 242                         break;
 243         }
 244
 245         /*
 246          * Copy other important information into the EState
 247          */
 248         estate->es_snapshot = RegisterSnapshot(queryDesc->snapshot);
 249         estate->es_crosscheck_snapshot = RegisterSnapshot(queryDesc->crosscheck_snapshot);
 250         estate->es_top_eflags = eflags;
 251         estate->es_instrument = queryDesc->instrument_options;
 252
 253         /*
 254          * Initialize the plan state tree
 255          */
 256         InitPlan(queryDesc, eflags);
 257
 258         /*
 259          * Set up an AFTER-trigger statement context, unless told not to, or
 260          * unless it's EXPLAIN-only mode (when ExecutorFinish won't be called).
 261          */
 262         if (!(eflags & (EXEC_FLAG_SKIP_TRIGGERS | EXEC_FLAG_EXPLAIN_ONLY)))
 263                 AfterTriggerBeginQuery();
 264
 265         MemoryContextSwitchTo(oldcontext);
 266 }
 267
 268 /* ----------------------------------------------------------------
 269  *              ExecutorRun
 270  *
 271  *              This is the main routine of the executor module. It accepts
 272  *              the query descriptor from the traffic cop and executes the
 273  *              query plan.
 274  *
 275  *              ExecutorStart must have been called already.
 276  *
 277  *              If direction is NoMovementScanDirection then nothing is done
 278  *              except to start up/shut down the destination.  Otherwise,
 279  *              we retrieve up to 'count' tuples in the specified direction.
 280  *
 281  *              Note: count = 0 is interpreted as no portal limit, i.e., run to
 282  *              completion.  Also note that the count limit is only applied to
 283  *              retrieved tuples, not for instance to those inserted/updated/deleted
 284  *              by a ModifyTable plan node.
 285  *
 286  *              There is no return value, but output tuples (if any) are sent to
 287  *              the destination receiver specified in the QueryDesc; and the number
 288  *              of tuples processed at the top level can be found in
 289  *              estate->es_processed.
 290  *
 291  *              We provide a function hook variable that lets loadable plugins
 292  *              get control when ExecutorRun is called.  Such a plugin would
 293  *              normally call standard_ExecutorRun().
 294  *
 295  * ----------------------------------------------------------------
 296  */
 297 void
 298 ExecutorRun(QueryDesc *queryDesc,
 299                         ScanDirection direction, uint64 count,
 300                         bool execute_once)
 301 {
 302         if (ExecutorRun_hook)
 303                 (*ExecutorRun_hook) (queryDesc, direction, count, execute_once);
 304         else
 305                 standard_ExecutorRun(queryDesc, direction, count, execute_once);
 306 }
 307
 308 void
 309 standard_ExecutorRun(QueryDesc *queryDesc,
 310                                          ScanDirection direction, uint64 count, bool execute_once)
 311 {
 312         EState     *estate;
 313         CmdType         operation;
 314         DestReceiver *dest;
 315         bool            sendTuples;
 316         MemoryContext oldcontext;
 317
 318         /* sanity checks */
 319         Assert(queryDesc != NULL);
 320
 321         estate = queryDesc->estate;
 322
 323         Assert(estate != NULL);
 324         Assert(!(estate->es_top_eflags & EXEC_FLAG_EXPLAIN_ONLY));
 325
 326         /*
 327          * Switch into per-query memory context
 328          */
 329         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 330
 331         /* Allow instrumentation of Executor overall runtime */
 332         if (queryDesc->totaltime)
 333                 InstrStartNode(queryDesc->totaltime);
 334
 335         /*
 336          * extract information from the query descriptor and the query feature.
 337          */
 338         operation = queryDesc->operation;
 339         dest = queryDesc->dest;
 340
 341         /*
 342          * startup tuple receiver, if we will be emitting tuples
 343          */
 344         estate->es_processed = 0;
 345         estate->es_lastoid = InvalidOid;
 346
 347         sendTuples = (operation == CMD_SELECT ||
 348                                   queryDesc->plannedstmt->hasReturning);
 349
 350         if (sendTuples)
 351                 (*dest->rStartup) (dest, operation, queryDesc->tupDesc);
 352
 353         /*
 354          * run plan
 355          */
 356         if (!ScanDirectionIsNoMovement(direction))
 357         {
 358                 if (execute_once && queryDesc->already_executed)
 359                         elog(ERROR, "can't re-execute query flagged for single execution");
 360                 queryDesc->already_executed = true;
 361
 362                 ExecutePlan(estate,
 363                                         queryDesc->planstate,
 364                                         queryDesc->plannedstmt->parallelModeNeeded,
 365                                         operation,
 366                                         sendTuples,
 367                                         count,
 368                                         direction,
 369                                         dest,
 370                                         execute_once);
 371         }
 372
 373         /*
 374          * shutdown tuple receiver, if we started it
 375          */
 376         if (sendTuples)
 377                 (*dest->rShutdown) (dest);
 378
 379         if (queryDesc->totaltime)
 380                 InstrStopNode(queryDesc->totaltime, estate->es_processed);
 381
 382         MemoryContextSwitchTo(oldcontext);
 383 }
 384
 385 /* ----------------------------------------------------------------
 386  *              ExecutorFinish
 387  *
 388  *              This routine must be called after the last ExecutorRun call.
 389  *              It performs cleanup such as firing AFTER triggers.  It is
 390  *              separate from ExecutorEnd because EXPLAIN ANALYZE needs to
 391  *              include these actions in the total runtime.
 392  *
 393  *              We provide a function hook variable that lets loadable plugins
 394  *              get control when ExecutorFinish is called.  Such a plugin would
 395  *              normally call standard_ExecutorFinish().
 396  *
 397  * ----------------------------------------------------------------
 398  */
 399 void
 400 ExecutorFinish(QueryDesc *queryDesc)
 401 {
 402         if (ExecutorFinish_hook)
 403                 (*ExecutorFinish_hook) (queryDesc);
 404         else
 405                 standard_ExecutorFinish(queryDesc);
 406 }
 407
 408 void
 409 standard_ExecutorFinish(QueryDesc *queryDesc)
 410 {
 411         EState     *estate;
 412         MemoryContext oldcontext;
 413
 414         /* sanity checks */
 415         Assert(queryDesc != NULL);
 416
 417         estate = queryDesc->estate;
 418
 419         Assert(estate != NULL);
 420         Assert(!(estate->es_top_eflags & EXEC_FLAG_EXPLAIN_ONLY));
 421
 422         /* This should be run once and only once per Executor instance */
 423         Assert(!estate->es_finished);
 424
 425         /* Switch into per-query memory context */
 426         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 427
 428         /* Allow instrumentation of Executor overall runtime */
 429         if (queryDesc->totaltime)
 430                 InstrStartNode(queryDesc->totaltime);
 431
 432         /* Run ModifyTable nodes to completion */
 433         ExecPostprocessPlan(estate);
 434
 435         /* Execute queued AFTER triggers, unless told not to */
 436         if (!(estate->es_top_eflags & EXEC_FLAG_SKIP_TRIGGERS))
 437                 AfterTriggerEndQuery(estate);
 438
 439         if (queryDesc->totaltime)
 440                 InstrStopNode(queryDesc->totaltime, 0);
 441
 442         MemoryContextSwitchTo(oldcontext);
 443
 444         estate->es_finished = true;
 445 }
 446
 447 /* ----------------------------------------------------------------
 448  *              ExecutorEnd
 449  *
 450  *              This routine must be called at the end of execution of any
 451  *              query plan
 452  *
 453  *              We provide a function hook variable that lets loadable plugins
 454  *              get control when ExecutorEnd is called.  Such a plugin would
 455  *              normally call standard_ExecutorEnd().
 456  *
 457  * ----------------------------------------------------------------
 458  */
 459 void
 460 ExecutorEnd(QueryDesc *queryDesc)
 461 {
 462         if (ExecutorEnd_hook)
 463                 (*ExecutorEnd_hook) (queryDesc);
 464         else
 465                 standard_ExecutorEnd(queryDesc);
 466 }
 467
 468 void
 469 standard_ExecutorEnd(QueryDesc *queryDesc)
 470 {
 471         EState     *estate;
 472         MemoryContext oldcontext;
 473
 474         /* sanity checks */
 475         Assert(queryDesc != NULL);
 476
 477         estate = queryDesc->estate;
 478
 479         Assert(estate != NULL);
 480
 481         /*
 482          * Check that ExecutorFinish was called, unless in EXPLAIN-only mode. This
 483          * Assert is needed because ExecutorFinish is new as of 9.1, and callers
 484          * might forget to call it.
 485          */
 486         Assert(estate->es_finished ||
 487                    (estate->es_top_eflags & EXEC_FLAG_EXPLAIN_ONLY));
 488
 489         /*
 490          * Switch into per-query memory context to run ExecEndPlan
 491          */
 492         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 493
 494         ExecEndPlan(queryDesc->planstate, estate);
 495
 496         /* do away with our snapshots */
 497         UnregisterSnapshot(estate->es_snapshot);
 498         UnregisterSnapshot(estate->es_crosscheck_snapshot);
 499
 500         /*
 501          * Must switch out of context before destroying it
 502          */
 503         MemoryContextSwitchTo(oldcontext);
 504
 505         /*
 506          * Release EState and per-query memory context.  This should release
 507          * everything the executor has allocated.
 508          */
 509         FreeExecutorState(estate);
 510
 511         /* Reset queryDesc fields that no longer point to anything */
 512         queryDesc->tupDesc = NULL;
 513         queryDesc->estate = NULL;
 514         queryDesc->planstate = NULL;
 515         queryDesc->totaltime = NULL;
 516 }
 517
 518 /* ----------------------------------------------------------------
 519  *              ExecutorRewind
 520  *
 521  *              This routine may be called on an open queryDesc to rewind it
 522  *              to the start.
 523  * ----------------------------------------------------------------
 524  */
 525 void
 526 ExecutorRewind(QueryDesc *queryDesc)
 527 {
 528         EState     *estate;
 529         MemoryContext oldcontext;
 530
 531         /* sanity checks */
 532         Assert(queryDesc != NULL);
 533
 534         estate = queryDesc->estate;
 535
 536         Assert(estate != NULL);
 537
 538         /* It's probably not sensible to rescan updating queries */
 539         Assert(queryDesc->operation == CMD_SELECT);
 540
 541         /*
 542          * Switch into per-query memory context
 543          */
 544         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 545
 546         /*
 547          * rescan plan
 548          */
 549         ExecReScan(queryDesc->planstate);
 550
 551         MemoryContextSwitchTo(oldcontext);
 552 }
 553
 554
 555 /*
 556  * ExecCheckRTPerms
 557  *              Check access permissions for all relations listed in a range table.
 558  *
 559  * Returns true if permissions are adequate.  Otherwise, throws an appropriate
 560  * error if ereport_on_violation is true, or simply returns false otherwise.
 561  *
 562  * Note that this does NOT address row level security policies (aka: RLS).  If
 563  * rows will be returned to the user as a result of this permission check
 564  * passing, then RLS also needs to be consulted (and check_enable_rls()).
 565  *
 566  * See rewrite/rowsecurity.c.
 567  */
 568 bool
 569 ExecCheckRTPerms(List *rangeTable, bool ereport_on_violation)
 570 {
 571         ListCell   *l;
 572         bool            result = true;
 573
 574         foreach(l, rangeTable)
 575         {
 576                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 577
 578                 result = ExecCheckRTEPerms(rte);
 579                 if (!result)
 580                 {
 581                         Assert(rte->rtekind == RTE_RELATION);
 582                         if (ereport_on_violation)
 583                                 aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
 584                                                            get_rel_name(rte->relid));
 585                         return false;
 586                 }
 587         }
 588
 589         if (ExecutorCheckPerms_hook)
 590                 result = (*ExecutorCheckPerms_hook) (rangeTable,
 591                                                                                          ereport_on_violation);
 592         return result;
 593 }
 594
 595 /*
 596  * ExecCheckRTEPerms
 597  *              Check access permissions for a single RTE.
 598  */
 599 static bool
 600 ExecCheckRTEPerms(RangeTblEntry *rte)
 601 {
 602         AclMode         requiredPerms;
 603         AclMode         relPerms;
 604         AclMode         remainingPerms;
 605         Oid                     relOid;
 606         Oid                     userid;
 607
 608         /*
 609          * Only plain-relation RTEs need to be checked here.  Function RTEs are
 610          * checked when the function is prepared for execution.  Join, subquery,
 611          * and special RTEs need no checks.
 612          */
 613         if (rte->rtekind != RTE_RELATION)
 614                 return true;
 615
 616         /*
 617          * No work if requiredPerms is empty.
 618          */
 619         requiredPerms = rte->requiredPerms;
 620         if (requiredPerms == 0)
 621                 return true;
 622
 623         relOid = rte->relid;
 624
 625         /*
 626          * userid to check as: current user unless we have a setuid indication.
 627          *
 628          * Note: GetUserId() is presently fast enough that there's no harm in
 629          * calling it separately for each RTE.  If that stops being true, we could
 630          * call it once in ExecCheckRTPerms and pass the userid down from there.
 631          * But for now, no need for the extra clutter.
 632          */
 633         userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
 634
 635         /*
 636          * We must have *all* the requiredPerms bits, but some of the bits can be
 637          * satisfied from column-level rather than relation-level permissions.
 638          * First, remove any bits that are satisfied by relation permissions.
 639          */
 640         relPerms = pg_class_aclmask(relOid, userid, requiredPerms, ACLMASK_ALL);
 641         remainingPerms = requiredPerms & ~relPerms;
 642         if (remainingPerms != 0)
 643         {
 644                 int                     col = -1;
 645
 646                 /*
 647                  * If we lack any permissions that exist only as relation permissions,
 648                  * we can fail straight away.
 649                  */
 650                 if (remainingPerms & ~(ACL_SELECT | ACL_INSERT | ACL_UPDATE))
 651                         return false;
 652
 653                 /*
 654                  * Check to see if we have the needed privileges at column level.
 655                  *
 656                  * Note: failures just report a table-level error; it would be nicer
 657                  * to report a column-level error if we have some but not all of the
 658                  * column privileges.
 659                  */
 660                 if (remainingPerms & ACL_SELECT)
 661                 {
 662                         /*
 663                          * When the query doesn't explicitly reference any columns (for
 664                          * example, SELECT COUNT(*) FROM table), allow the query if we
 665                          * have SELECT on any column of the rel, as per SQL spec.
 666                          */
 667                         if (bms_is_empty(rte->selectedCols))
 668                         {
 669                                 if (pg_attribute_aclcheck_all(relOid, userid, ACL_SELECT,
 670                                                                                           ACLMASK_ANY) != ACLCHECK_OK)
 671                                         return false;
 672                         }
 673
 674                         while ((col = bms_next_member(rte->selectedCols, col)) >= 0)
 675                         {
 676                                 /* bit #s are offset by FirstLowInvalidHeapAttributeNumber */
 677                                 AttrNumber      attno = col + FirstLowInvalidHeapAttributeNumber;
 678
 679                                 if (attno == InvalidAttrNumber)
 680                                 {
 681                                         /* Whole-row reference, must have priv on all cols */
 682                                         if (pg_attribute_aclcheck_all(relOid, userid, ACL_SELECT,
 683                                                                                                   ACLMASK_ALL) != ACLCHECK_OK)
 684                                                 return false;
 685                                 }
 686                                 else
 687                                 {
 688                                         if (pg_attribute_aclcheck(relOid, attno, userid,
 689                                                                                           ACL_SELECT) != ACLCHECK_OK)
 690                                                 return false;
 691                                 }
 692                         }
 693                 }
 694
 695                 /*
 696                  * Basically the same for the mod columns, for both INSERT and UPDATE
 697                  * privilege as specified by remainingPerms.
 698                  */
 699                 if (remainingPerms & ACL_INSERT && !ExecCheckRTEPermsModified(relOid,
 700                                                                                                                                           userid,
 701                                                                                                                    rte->insertedCols,
 702                                                                                                                                  ACL_INSERT))
 703                         return false;
 704
 705                 if (remainingPerms & ACL_UPDATE && !ExecCheckRTEPermsModified(relOid,
 706                                                                                                                                           userid,
 707                                                                                                                         rte->updatedCols,
 708                                                                                                                                  ACL_UPDATE))
 709                         return false;
 710         }
 711         return true;
 712 }
 713
 714 /*
 715  * ExecCheckRTEPermsModified
 716  *              Check INSERT or UPDATE access permissions for a single RTE (these
 717  *              are processed uniformly).
 718  */
 719 static bool
 720 ExecCheckRTEPermsModified(Oid relOid, Oid userid, Bitmapset *modifiedCols,
 721                                                   AclMode requiredPerms)
 722 {
 723         int                     col = -1;
 724
 725         /*
 726          * When the query doesn't explicitly update any columns, allow the query
 727          * if we have permission on any column of the rel.  This is to handle
 728          * SELECT FOR UPDATE as well as possible corner cases in UPDATE.
 729          */
 730         if (bms_is_empty(modifiedCols))
 731         {
 732                 if (pg_attribute_aclcheck_all(relOid, userid, requiredPerms,
 733                                                                           ACLMASK_ANY) != ACLCHECK_OK)
 734                         return false;
 735         }
 736
 737         while ((col = bms_next_member(modifiedCols, col)) >= 0)
 738         {
 739                 /* bit #s are offset by FirstLowInvalidHeapAttributeNumber */
 740                 AttrNumber      attno = col + FirstLowInvalidHeapAttributeNumber;
 741
 742                 if (attno == InvalidAttrNumber)
 743                 {
 744                         /* whole-row reference can't happen here */
 745                         elog(ERROR, "whole-row update is not implemented");
 746                 }
 747                 else
 748                 {
 749                         if (pg_attribute_aclcheck(relOid, attno, userid,
 750                                                                           requiredPerms) != ACLCHECK_OK)
 751                                 return false;
 752                 }
 753         }
 754         return true;
 755 }
 756
 757 /*
 758  * Check that the query does not imply any writes to non-temp tables;
 759  * unless we're in parallel mode, in which case don't even allow writes
 760  * to temp tables.
 761  *
 762  * Note: in a Hot Standby slave this would need to reject writes to temp
 763  * tables just as we do in parallel mode; but an HS slave can't have created
 764  * any temp tables in the first place, so no need to check that.
 765  */
 766 static void
 767 ExecCheckXactReadOnly(PlannedStmt *plannedstmt)
 768 {
 769         ListCell   *l;
 770
 771         /*
 772          * Fail if write permissions are requested in parallel mode for table
 773          * (temp or non-temp), otherwise fail for any non-temp table.
 774          */
 775         foreach(l, plannedstmt->rtable)
 776         {
 777                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 778
 779                 if (rte->rtekind != RTE_RELATION)
 780                         continue;
 781
 782                 if ((rte->requiredPerms & (~ACL_SELECT)) == 0)
 783                         continue;
 784
 785                 if (isTempNamespace(get_rel_namespace(rte->relid)))
 786                         continue;
 787
 788                 PreventCommandIfReadOnly(CreateCommandTag((Node *) plannedstmt));
 789         }
 790
 791         if (plannedstmt->commandType != CMD_SELECT || plannedstmt->hasModifyingCTE)
 792                 PreventCommandIfParallelMode(CreateCommandTag((Node *) plannedstmt));
 793 }
 794
 795
 796 /* ----------------------------------------------------------------
 797  *              InitPlan
 798  *
 799  *              Initializes the query plan: open files, allocate storage
 800  *              and start up the rule manager
 801  * ----------------------------------------------------------------
 802  */
 803 static void
 804 InitPlan(QueryDesc *queryDesc, int eflags)
 805 {
 806         CmdType         operation = queryDesc->operation;
 807         PlannedStmt *plannedstmt = queryDesc->plannedstmt;
 808         Plan       *plan = plannedstmt->planTree;
 809         List       *rangeTable = plannedstmt->rtable;
 810         EState     *estate = queryDesc->estate;
 811         PlanState  *planstate;
 812         TupleDesc       tupType;
 813         ListCell   *l;
 814         int                     i;
 815
 816         /*
 817          * Do permissions checks
 818          */
 819         ExecCheckRTPerms(rangeTable, true);
 820
 821         /*
 822          * initialize the node's execution state
 823          */
 824         estate->es_range_table = rangeTable;
 825         estate->es_plannedstmt = plannedstmt;
 826
 827         /*
 828          * initialize result relation stuff, and open/lock the result rels.
 829          *
 830          * We must do this before initializing the plan tree, else we might try to
 831          * do a lock upgrade if a result rel is also a source rel.
 832          */
 833         if (plannedstmt->resultRelations)
 834         {
 835                 List       *resultRelations = plannedstmt->resultRelations;
 836                 int                     numResultRelations = list_length(resultRelations);
 837                 ResultRelInfo *resultRelInfos;
 838                 ResultRelInfo *resultRelInfo;
 839
 840                 resultRelInfos = (ResultRelInfo *)
 841                         palloc(numResultRelations * sizeof(ResultRelInfo));
 842                 resultRelInfo = resultRelInfos;
 843                 foreach(l, resultRelations)
 844                 {
 845                         Index           resultRelationIndex = lfirst_int(l);
 846                         Oid                     resultRelationOid;
 847                         Relation        resultRelation;
 848
 849                         resultRelationOid = getrelid(resultRelationIndex, rangeTable);
 850                         resultRelation = heap_open(resultRelationOid, RowExclusiveLock);
 851
 852                         InitResultRelInfo(resultRelInfo,
 853                                                           resultRelation,
 854                                                           resultRelationIndex,
 855                                                           NULL,
 856                                                           estate->es_instrument);
 857                         resultRelInfo++;
 858                 }
 859                 estate->es_result_relations = resultRelInfos;
 860                 estate->es_num_result_relations = numResultRelations;
 861                 /* es_result_relation_info is NULL except when within ModifyTable */
 862                 estate->es_result_relation_info = NULL;
 863
 864                 /*
 865                  * In the partitioned result relation case, lock the non-leaf result
 866                  * relations too.  A subset of these are the roots of respective
 867                  * partitioned tables, for which we also allocate ResulRelInfos.
 868                  */
 869                 estate->es_root_result_relations = NULL;
 870                 estate->es_num_root_result_relations = 0;
 871                 if (plannedstmt->nonleafResultRelations)
 872                 {
 873                         int                     num_roots = list_length(plannedstmt->rootResultRelations);
 874
 875                         /*
 876                          * Firstly, build ResultRelInfos for all the partitioned table
 877                          * roots, because we will need them to fire the statement-level
 878                          * triggers, if any.
 879                          */
 880                         resultRelInfos = (ResultRelInfo *)
 881                                 palloc(num_roots * sizeof(ResultRelInfo));
 882                         resultRelInfo = resultRelInfos;
 883                         foreach(l, plannedstmt->rootResultRelations)
 884                         {
 885                                 Index           resultRelIndex = lfirst_int(l);
 886                                 Oid                     resultRelOid;
 887                                 Relation        resultRelDesc;
 888
 889                                 resultRelOid = getrelid(resultRelIndex, rangeTable);
 890                                 resultRelDesc = heap_open(resultRelOid, RowExclusiveLock);
 891                                 InitResultRelInfo(resultRelInfo,
 892                                                                   resultRelDesc,
 893                                                                   lfirst_int(l),
 894                                                                   NULL,
 895                                                                   estate->es_instrument);
 896                                 resultRelInfo++;
 897                         }
 898
 899                         estate->es_root_result_relations = resultRelInfos;
 900                         estate->es_num_root_result_relations = num_roots;
 901
 902                         /* Simply lock the rest of them. */
 903                         foreach(l, plannedstmt->nonleafResultRelations)
 904                         {
 905                                 Index           resultRelIndex = lfirst_int(l);
 906
 907                                 /* We locked the roots above. */
 908                                 if (!list_member_int(plannedstmt->rootResultRelations,
 909                                                                          resultRelIndex))
 910                                         LockRelationOid(getrelid(resultRelIndex, rangeTable),
 911                                                                         RowExclusiveLock);
 912                         }
 913                 }
 914         }
 915         else
 916         {
 917                 /*
 918                  * if no result relation, then set state appropriately
 919                  */
 920                 estate->es_result_relations = NULL;
 921                 estate->es_num_result_relations = 0;
 922                 estate->es_result_relation_info = NULL;
 923                 estate->es_root_result_relations = NULL;
 924                 estate->es_num_root_result_relations = 0;
 925         }
 926
 927         /*
 928          * Similarly, we have to lock relations selected FOR [KEY] UPDATE/SHARE
 929          * before we initialize the plan tree, else we'd be risking lock upgrades.
 930          * While we are at it, build the ExecRowMark list.  Any partitioned child
 931          * tables are ignored here (because isParent=true) and will be locked by
 932          * the first Append or MergeAppend node that references them.  (Note that
 933          * the RowMarks corresponding to partitioned child tables are present in
 934          * the same list as the rest, i.e., plannedstmt->rowMarks.)
 935          */
 936         estate->es_rowMarks = NIL;
 937         foreach(l, plannedstmt->rowMarks)
 938         {
 939                 PlanRowMark *rc = (PlanRowMark *) lfirst(l);
 940                 Oid                     relid;
 941                 Relation        relation;
 942                 ExecRowMark *erm;
 943
 944                 /* ignore "parent" rowmarks; they are irrelevant at runtime */
 945                 if (rc->isParent)
 946                         continue;
 947
 948                 /* get relation's OID (will produce InvalidOid if subquery) */
 949                 relid = getrelid(rc->rti, rangeTable);
 950
 951                 /*
 952                  * If you change the conditions under which rel locks are acquired
 953                  * here, be sure to adjust ExecOpenScanRelation to match.
 954                  */
 955                 switch (rc->markType)
 956                 {
 957                         case ROW_MARK_EXCLUSIVE:
 958                         case ROW_MARK_NOKEYEXCLUSIVE:
 959                         case ROW_MARK_SHARE:
 960                         case ROW_MARK_KEYSHARE:
 961                                 relation = heap_open(relid, RowShareLock);
 962                                 break;
 963                         case ROW_MARK_REFERENCE:
 964                                 relation = heap_open(relid, AccessShareLock);
 965                                 break;
 966                         case ROW_MARK_COPY:
 967                                 /* no physical table access is required */
 968                                 relation = NULL;
 969                                 break;
 970                         default:
 971                                 elog(ERROR, "unrecognized markType: %d", rc->markType);
 972                                 relation = NULL;        /* keep compiler quiet */
 973                                 break;
 974                 }
 975
 976                 /* Check that relation is a legal target for marking */
 977                 if (relation)
 978                         CheckValidRowMarkRel(relation, rc->markType);
 979
 980                 erm = (ExecRowMark *) palloc(sizeof(ExecRowMark));
 981                 erm->relation = relation;
 982                 erm->relid = relid;
 983                 erm->rti = rc->rti;
 984                 erm->prti = rc->prti;
 985                 erm->rowmarkId = rc->rowmarkId;
 986                 erm->markType = rc->markType;
 987                 erm->strength = rc->strength;
 988                 erm->waitPolicy = rc->waitPolicy;
 989                 erm->ermActive = false;
 990                 ItemPointerSetInvalid(&(erm->curCtid));
 991                 erm->ermExtra = NULL;
 992                 estate->es_rowMarks = lappend(estate->es_rowMarks, erm);
 993         }
 994
 995         /*
 996          * Initialize the executor's tuple table to empty.
 997          */
 998         estate->es_tupleTable = NIL;
 999         estate->es_trig_tuple_slot = NULL;
1000         estate->es_trig_oldtup_slot = NULL;
1001         estate->es_trig_newtup_slot = NULL;
1002
1003         /* mark EvalPlanQual not active */
1004         estate->es_epqTuple = NULL;
1005         estate->es_epqTupleSet = NULL;
1006         estate->es_epqScanDone = NULL;
1007
1008         /*
1009          * Initialize private state information for each SubPlan.  We must do this
1010          * before running ExecInitNode on the main query tree, since
1011          * ExecInitSubPlan expects to be able to find these entries.
1012          */
1013         Assert(estate->es_subplanstates == NIL);
1014         i = 1;                                          /* subplan indices count from 1 */
1015         foreach(l, plannedstmt->subplans)
1016         {
1017                 Plan       *subplan = (Plan *) lfirst(l);
1018                 PlanState  *subplanstate;
1019                 int                     sp_eflags;
1020
1021                 /*
1022                  * A subplan will never need to do BACKWARD scan nor MARK/RESTORE. If
1023                  * it is a parameterless subplan (not initplan), we suggest that it be
1024                  * prepared to handle REWIND efficiently; otherwise there is no need.
1025                  */
1026                 sp_eflags = eflags
1027                         & (EXEC_FLAG_EXPLAIN_ONLY | EXEC_FLAG_WITH_NO_DATA);
1028                 if (bms_is_member(i, plannedstmt->rewindPlanIDs))
1029                         sp_eflags |= EXEC_FLAG_REWIND;
1030
1031                 subplanstate = ExecInitNode(subplan, estate, sp_eflags);
1032
1033                 estate->es_subplanstates = lappend(estate->es_subplanstates,
1034                                                                                    subplanstate);
1035
1036                 i++;
1037         }
1038
1039         /*
1040          * Initialize the private state information for all the nodes in the query
1041          * tree.  This opens files, allocates storage and leaves us ready to start
1042          * processing tuples.
1043          */
1044         planstate = ExecInitNode(plan, estate, eflags);
1045
1046         /*
1047          * Get the tuple descriptor describing the type of tuples to return.
1048          */
1049         tupType = ExecGetResultType(planstate);
1050
1051         /*
1052          * Initialize the junk filter if needed.  SELECT queries need a filter if
1053          * there are any junk attrs in the top-level tlist.
1054          */
1055         if (operation == CMD_SELECT)
1056         {
1057                 bool            junk_filter_needed = false;
1058                 ListCell   *tlist;
1059
1060                 foreach(tlist, plan->targetlist)
1061                 {
1062                         TargetEntry *tle = (TargetEntry *) lfirst(tlist);
1063
1064                         if (tle->resjunk)
1065                         {
1066                                 junk_filter_needed = true;
1067                                 break;
1068                         }
1069                 }
1070
1071                 if (junk_filter_needed)
1072                 {
1073                         JunkFilter *j;
1074
1075                         j = ExecInitJunkFilter(planstate->plan->targetlist,
1076                                                                    tupType->tdhasoid,
1077                                                                    ExecInitExtraTupleSlot(estate));
1078                         estate->es_junkFilter = j;
1079
1080                         /* Want to return the cleaned tuple type */
1081                         tupType = j->jf_cleanTupType;
1082                 }
1083         }
1084
1085         queryDesc->tupDesc = tupType;
1086         queryDesc->planstate = planstate;
1087 }
1088
1089 /*
1090  * Check that a proposed result relation is a legal target for the operation
1091  *
1092  * Generally the parser and/or planner should have noticed any such mistake
1093  * already, but let's make sure.
1094  *
1095  * Note: when changing this function, you probably also need to look at
1096  * CheckValidRowMarkRel.
1097  */
1098 void
1099 CheckValidResultRel(Relation resultRel, CmdType operation)
1100 {
1101         TriggerDesc *trigDesc = resultRel->trigdesc;
1102         FdwRoutine *fdwroutine;
1103
1104         switch (resultRel->rd_rel->relkind)
1105         {
1106                 case RELKIND_RELATION:
1107                 case RELKIND_PARTITIONED_TABLE:
1108                         CheckCmdReplicaIdentity(resultRel, operation);
1109                         break;
1110                 case RELKIND_SEQUENCE:
1111                         ereport(ERROR,
1112                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1113                                          errmsg("cannot change sequence \"%s\"",
1114                                                         RelationGetRelationName(resultRel))));
1115                         break;
1116                 case RELKIND_TOASTVALUE:
1117                         ereport(ERROR,
1118                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1119                                          errmsg("cannot change TOAST relation \"%s\"",
1120                                                         RelationGetRelationName(resultRel))));
1121                         break;
1122                 case RELKIND_VIEW:
1123
1124                         /*
1125                          * Okay only if there's a suitable INSTEAD OF trigger.  Messages
1126                          * here should match rewriteHandler.c's rewriteTargetView, except
1127                          * that we omit errdetail because we haven't got the information
1128                          * handy (and given that we really shouldn't get here anyway, it's
1129                          * not worth great exertion to get).
1130                          */
1131                         switch (operation)
1132                         {
1133                                 case CMD_INSERT:
1134                                         if (!trigDesc || !trigDesc->trig_insert_instead_row)
1135                                                 ereport(ERROR,
1136                                                   (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1137                                                    errmsg("cannot insert into view \"%s\"",
1138                                                                   RelationGetRelationName(resultRel)),
1139                                                    errhint("To enable inserting into the view, provide an INSTEAD OF INSERT trigger or an unconditional ON INSERT DO INSTEAD rule.")));
1140                                         break;
1141                                 case CMD_UPDATE:
1142                                         if (!trigDesc || !trigDesc->trig_update_instead_row)
1143                                                 ereport(ERROR,
1144                                                   (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1145                                                    errmsg("cannot update view \"%s\"",
1146                                                                   RelationGetRelationName(resultRel)),
1147                                                    errhint("To enable updating the view, provide an INSTEAD OF UPDATE trigger or an unconditional ON UPDATE DO INSTEAD rule.")));
1148                                         break;
1149                                 case CMD_DELETE:
1150                                         if (!trigDesc || !trigDesc->trig_delete_instead_row)
1151                                                 ereport(ERROR,
1152                                                   (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1153                                                    errmsg("cannot delete from view \"%s\"",
1154                                                                   RelationGetRelationName(resultRel)),
1155                                                    errhint("To enable deleting from the view, provide an INSTEAD OF DELETE trigger or an unconditional ON DELETE DO INSTEAD rule.")));
1156                                         break;
1157                                 default:
1158                                         elog(ERROR, "unrecognized CmdType: %d", (int) operation);
1159                                         break;
1160                         }
1161                         break;
1162                 case RELKIND_MATVIEW:
1163                         if (!MatViewIncrementalMaintenanceIsEnabled())
1164                                 ereport(ERROR,
1165                                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1166                                                  errmsg("cannot change materialized view \"%s\"",
1167                                                                 RelationGetRelationName(resultRel))));
1168                         break;
1169                 case RELKIND_FOREIGN_TABLE:
1170                         /* Okay only if the FDW supports it */
1171                         fdwroutine = GetFdwRoutineForRelation(resultRel, false);
1172                         switch (operation)
1173                         {
1174                                 case CMD_INSERT:
1175                                         if (fdwroutine->ExecForeignInsert == NULL)
1176                                                 ereport(ERROR,
1177                                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1178                                                         errmsg("cannot insert into foreign table \"%s\"",
1179                                                                    RelationGetRelationName(resultRel))));
1180                                         if (fdwroutine->IsForeignRelUpdatable != NULL &&
1181                                                 (fdwroutine->IsForeignRelUpdatable(resultRel) & (1 << CMD_INSERT)) == 0)
1182                                                 ereport(ERROR,
1183                                                   (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1184                                                 errmsg("foreign table \"%s\" does not allow inserts",
1185                                                            RelationGetRelationName(resultRel))));
1186                                         break;
1187                                 case CMD_UPDATE:
1188                                         if (fdwroutine->ExecForeignUpdate == NULL)
1189                                                 ereport(ERROR,
1190                                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1191                                                                  errmsg("cannot update foreign table \"%s\"",
1192                                                                                 RelationGetRelationName(resultRel))));
1193                                         if (fdwroutine->IsForeignRelUpdatable != NULL &&
1194                                                 (fdwroutine->IsForeignRelUpdatable(resultRel) & (1 << CMD_UPDATE)) == 0)
1195                                                 ereport(ERROR,
1196                                                   (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1197                                                 errmsg("foreign table \"%s\" does not allow updates",
1198                                                            RelationGetRelationName(resultRel))));
1199                                         break;
1200                                 case CMD_DELETE:
1201                                         if (fdwroutine->ExecForeignDelete == NULL)
1202                                                 ereport(ERROR,
1203                                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1204                                                         errmsg("cannot delete from foreign table \"%s\"",
1205                                                                    RelationGetRelationName(resultRel))));
1206                                         if (fdwroutine->IsForeignRelUpdatable != NULL &&
1207                                                 (fdwroutine->IsForeignRelUpdatable(resultRel) & (1 << CMD_DELETE)) == 0)
1208                                                 ereport(ERROR,
1209                                                   (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1210                                                 errmsg("foreign table \"%s\" does not allow deletes",
1211                                                            RelationGetRelationName(resultRel))));
1212                                         break;
1213                                 default:
1214                                         elog(ERROR, "unrecognized CmdType: %d", (int) operation);
1215                                         break;
1216                         }
1217                         break;
1218                 default:
1219                         ereport(ERROR,
1220                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1221                                          errmsg("cannot change relation \"%s\"",
1222                                                         RelationGetRelationName(resultRel))));
1223                         break;
1224         }
1225 }
1226
1227 /*
1228  * Check that a proposed rowmark target relation is a legal target
1229  *
1230  * In most cases parser and/or planner should have noticed this already, but
1231  * they don't cover all cases.
1232  */
1233 static void
1234 CheckValidRowMarkRel(Relation rel, RowMarkType markType)
1235 {
1236         FdwRoutine *fdwroutine;
1237
1238         switch (rel->rd_rel->relkind)
1239         {
1240                 case RELKIND_RELATION:
1241                 case RELKIND_PARTITIONED_TABLE:
1242                         /* OK */
1243                         break;
1244                 case RELKIND_SEQUENCE:
1245                         /* Must disallow this because we don't vacuum sequences */
1246                         ereport(ERROR,
1247                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1248                                          errmsg("cannot lock rows in sequence \"%s\"",
1249                                                         RelationGetRelationName(rel))));
1250                         break;
1251                 case RELKIND_TOASTVALUE:
1252                         /* We could allow this, but there seems no good reason to */
1253                         ereport(ERROR,
1254                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1255                                          errmsg("cannot lock rows in TOAST relation \"%s\"",
1256                                                         RelationGetRelationName(rel))));
1257                         break;
1258                 case RELKIND_VIEW:
1259                         /* Should not get here; planner should have expanded the view */
1260                         ereport(ERROR,
1261                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1262                                          errmsg("cannot lock rows in view \"%s\"",
1263                                                         RelationGetRelationName(rel))));
1264                         break;
1265                 case RELKIND_MATVIEW:
1266                         /* Allow referencing a matview, but not actual locking clauses */
1267                         if (markType != ROW_MARK_REFERENCE)
1268                                 ereport(ERROR,
1269                                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1270                                            errmsg("cannot lock rows in materialized view \"%s\"",
1271                                                           RelationGetRelationName(rel))));
1272                         break;
1273                 case RELKIND_FOREIGN_TABLE:
1274                         /* Okay only if the FDW supports it */
1275                         fdwroutine = GetFdwRoutineForRelation(rel, false);
1276                         if (fdwroutine->RefetchForeignRow == NULL)
1277                                 ereport(ERROR,
1278                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1279                                                  errmsg("cannot lock rows in foreign table \"%s\"",
1280                                                                 RelationGetRelationName(rel))));
1281                         break;
1282                 default:
1283                         ereport(ERROR,
1284                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1285                                          errmsg("cannot lock rows in relation \"%s\"",
1286                                                         RelationGetRelationName(rel))));
1287                         break;
1288         }
1289 }
1290
1291 /*
1292  * Initialize ResultRelInfo data for one result relation
1293  *
1294  * Caution: before Postgres 9.1, this function included the relkind checking
1295  * that's now in CheckValidResultRel, and it also did ExecOpenIndices if
1296  * appropriate.  Be sure callers cover those needs.
1297  */
1298 void
1299 InitResultRelInfo(ResultRelInfo *resultRelInfo,
1300                                   Relation resultRelationDesc,
1301                                   Index resultRelationIndex,
1302                                   Relation partition_root,
1303                                   int instrument_options)
1304 {
1305         List       *partition_check = NIL;
1306
1307         MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
1308         resultRelInfo->type = T_ResultRelInfo;
1309         resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
1310         resultRelInfo->ri_RelationDesc = resultRelationDesc;
1311         resultRelInfo->ri_NumIndices = 0;
1312         resultRelInfo->ri_IndexRelationDescs = NULL;
1313         resultRelInfo->ri_IndexRelationInfo = NULL;
1314         /* make a copy so as not to depend on relcache info not changing... */
1315         resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
1316         if (resultRelInfo->ri_TrigDesc)
1317         {
1318                 int                     n = resultRelInfo->ri_TrigDesc->numtriggers;
1319
1320                 resultRelInfo->ri_TrigFunctions = (FmgrInfo *)
1321                         palloc0(n * sizeof(FmgrInfo));
1322                 resultRelInfo->ri_TrigWhenExprs = (ExprState **)
1323                         palloc0(n * sizeof(ExprState *));
1324                 if (instrument_options)
1325                         resultRelInfo->ri_TrigInstrument = InstrAlloc(n, instrument_options);
1326         }
1327         else
1328         {
1329                 resultRelInfo->ri_TrigFunctions = NULL;
1330                 resultRelInfo->ri_TrigWhenExprs = NULL;
1331                 resultRelInfo->ri_TrigInstrument = NULL;
1332         }
1333         if (resultRelationDesc->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1334                 resultRelInfo->ri_FdwRoutine = GetFdwRoutineForRelation(resultRelationDesc, true);
1335         else
1336                 resultRelInfo->ri_FdwRoutine = NULL;
1337         resultRelInfo->ri_FdwState = NULL;
1338         resultRelInfo->ri_usesFdwDirectModify = false;
1339         resultRelInfo->ri_ConstraintExprs = NULL;
1340         resultRelInfo->ri_junkFilter = NULL;
1341         resultRelInfo->ri_projectReturning = NULL;
1342
1343         /*
1344          * Partition constraint, which also includes the partition constraint of
1345          * all the ancestors that are partitions.  Note that it will be checked
1346          * even in the case of tuple-routing where this table is the target leaf
1347          * partition, if there any BR triggers defined on the table.  Although
1348          * tuple-routing implicitly preserves the partition constraint of the
1349          * target partition for a given row, the BR triggers may change the row
1350          * such that the constraint is no longer satisfied, which we must fail for
1351          * by checking it explicitly.
1352          *
1353          * If this is a partitioned table, the partition constraint (if any) of a
1354          * given row will be checked just before performing tuple-routing.
1355          */
1356         partition_check = RelationGetPartitionQual(resultRelationDesc);
1357
1358         resultRelInfo->ri_PartitionCheck = partition_check;
1359         resultRelInfo->ri_PartitionRoot = partition_root;
1360 }
1361
1362 /*
1363  *              ExecGetTriggerResultRel
1364  *
1365  * Get a ResultRelInfo for a trigger target relation.  Most of the time,
1366  * triggers are fired on one of the result relations of the query, and so
1367  * we can just return a member of the es_result_relations array.  (Note: in
1368  * self-join situations there might be multiple members with the same OID;
1369  * if so it doesn't matter which one we pick.)  However, it is sometimes
1370  * necessary to fire triggers on other relations; this happens mainly when an
1371  * RI update trigger queues additional triggers on other relations, which will
1372  * be processed in the context of the outer query.  For efficiency's sake,
1373  * we want to have a ResultRelInfo for those triggers too; that can avoid
1374  * repeated re-opening of the relation.  (It also provides a way for EXPLAIN
1375  * ANALYZE to report the runtimes of such triggers.)  So we make additional
1376  * ResultRelInfo's as needed, and save them in es_trig_target_relations.
1377  */
1378 ResultRelInfo *
1379 ExecGetTriggerResultRel(EState *estate, Oid relid)
1380 {
1381         ResultRelInfo *rInfo;
1382         int                     nr;
1383         ListCell   *l;
1384         Relation        rel;
1385         MemoryContext oldcontext;
1386
1387         /* First, search through the query result relations */
1388         rInfo = estate->es_result_relations;
1389         nr = estate->es_num_result_relations;
1390         while (nr > 0)
1391         {
1392                 if (RelationGetRelid(rInfo->ri_RelationDesc) == relid)
1393                         return rInfo;
1394                 rInfo++;
1395                 nr--;
1396         }
1397         /* Nope, but maybe we already made an extra ResultRelInfo for it */
1398         foreach(l, estate->es_trig_target_relations)
1399         {
1400                 rInfo = (ResultRelInfo *) lfirst(l);
1401                 if (RelationGetRelid(rInfo->ri_RelationDesc) == relid)
1402                         return rInfo;
1403         }
1404         /* Nope, so we need a new one */
1405
1406         /*
1407          * Open the target relation's relcache entry.  We assume that an
1408          * appropriate lock is still held by the backend from whenever the trigger
1409          * event got queued, so we need take no new lock here.  Also, we need not
1410          * recheck the relkind, so no need for CheckValidResultRel.
1411          */
1412         rel = heap_open(relid, NoLock);
1413
1414         /*
1415          * Make the new entry in the right context.
1416          */
1417         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
1418         rInfo = makeNode(ResultRelInfo);
1419         InitResultRelInfo(rInfo,
1420                                           rel,
1421                                           0,            /* dummy rangetable index */
1422                                           NULL,
1423                                           estate->es_instrument);
1424         estate->es_trig_target_relations =
1425                 lappend(estate->es_trig_target_relations, rInfo);
1426         MemoryContextSwitchTo(oldcontext);
1427
1428         /*
1429          * Currently, we don't need any index information in ResultRelInfos used
1430          * only for triggers, so no need to call ExecOpenIndices.
1431          */
1432
1433         return rInfo;
1434 }
1435
1436 /*
1437  * Close any relations that have been opened by ExecGetTriggerResultRel().
1438  */
1439 void
1440 ExecCleanUpTriggerState(EState *estate)
1441 {
1442         ListCell   *l;
1443
1444         foreach(l, estate->es_trig_target_relations)
1445         {
1446                 ResultRelInfo *resultRelInfo = (ResultRelInfo *) lfirst(l);
1447
1448                 /* Close indices and then the relation itself */
1449                 ExecCloseIndices(resultRelInfo);
1450                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
1451         }
1452 }
1453
1454 /*
1455  *              ExecContextForcesOids
1456  *
1457  * This is pretty grotty: when doing INSERT, UPDATE, or CREATE TABLE AS,
1458  * we need to ensure that result tuples have space for an OID iff they are
1459  * going to be stored into a relation that has OIDs.  In other contexts
1460  * we are free to choose whether to leave space for OIDs in result tuples
1461  * (we generally don't want to, but we do if a physical-tlist optimization
1462  * is possible).  This routine checks the plan context and returns TRUE if the
1463  * choice is forced, FALSE if the choice is not forced.  In the TRUE case,
1464  * *hasoids is set to the required value.
1465  *
1466  * One reason this is ugly is that all plan nodes in the plan tree will emit
1467  * tuples with space for an OID, though we really only need the topmost node
1468  * to do so.  However, node types like Sort don't project new tuples but just
1469  * return their inputs, and in those cases the requirement propagates down
1470  * to the input node.  Eventually we might make this code smart enough to
1471  * recognize how far down the requirement really goes, but for now we just
1472  * make all plan nodes do the same thing if the top level forces the choice.
1473  *
1474  * We assume that if we are generating tuples for INSERT or UPDATE,
1475  * estate->es_result_relation_info is already set up to describe the target
1476  * relation.  Note that in an UPDATE that spans an inheritance tree, some of
1477  * the target relations may have OIDs and some not.  We have to make the
1478  * decisions on a per-relation basis as we initialize each of the subplans of
1479  * the ModifyTable node, so ModifyTable has to set es_result_relation_info
1480  * while initializing each subplan.
1481  *
1482  * CREATE TABLE AS is even uglier, because we don't have the target relation's
1483  * descriptor available when this code runs; we have to look aside at the
1484  * flags passed to ExecutorStart().
1485  */
1486 bool
1487 ExecContextForcesOids(PlanState *planstate, bool *hasoids)
1488 {
1489         ResultRelInfo *ri = planstate->state->es_result_relation_info;
1490
1491         if (ri != NULL)
1492         {
1493                 Relation        rel = ri->ri_RelationDesc;
1494
1495                 if (rel != NULL)
1496                 {
1497                         *hasoids = rel->rd_rel->relhasoids;
1498                         return true;
1499                 }
1500         }
1501
1502         if (planstate->state->es_top_eflags & EXEC_FLAG_WITH_OIDS)
1503         {
1504                 *hasoids = true;
1505                 return true;
1506         }
1507         if (planstate->state->es_top_eflags & EXEC_FLAG_WITHOUT_OIDS)
1508         {
1509                 *hasoids = false;
1510                 return true;
1511         }
1512
1513         return false;
1514 }
1515
1516 /* ----------------------------------------------------------------
1517  *              ExecPostprocessPlan
1518  *
1519  *              Give plan nodes a final chance to execute before shutdown
1520  * ----------------------------------------------------------------
1521  */
1522 static void
1523 ExecPostprocessPlan(EState *estate)
1524 {
1525         ListCell   *lc;
1526
1527         /*
1528          * Make sure nodes run forward.
1529          */
1530         estate->es_direction = ForwardScanDirection;
1531
1532         /*
1533          * Run any secondary ModifyTable nodes to completion, in case the main
1534          * query did not fetch all rows from them.  (We do this to ensure that
1535          * such nodes have predictable results.)
1536          */
1537         foreach(lc, estate->es_auxmodifytables)
1538         {
1539                 PlanState  *ps = (PlanState *) lfirst(lc);
1540
1541                 for (;;)
1542                 {
1543                         TupleTableSlot *slot;
1544
1545                         /* Reset the per-output-tuple exprcontext each time */
1546                         ResetPerTupleExprContext(estate);
1547
1548                         slot = ExecProcNode(ps);
1549
1550                         if (TupIsNull(slot))
1551                                 break;
1552                 }
1553         }
1554 }
1555
1556 /* ----------------------------------------------------------------
1557  *              ExecEndPlan
1558  *
1559  *              Cleans up the query plan -- closes files and frees up storage
1560  *
1561  * NOTE: we are no longer very worried about freeing storage per se
1562  * in this code; FreeExecutorState should be guaranteed to release all
1563  * memory that needs to be released.  What we are worried about doing
1564  * is closing relations and dropping buffer pins.  Thus, for example,
1565  * tuple tables must be cleared or dropped to ensure pins are released.
1566  * ----------------------------------------------------------------
1567  */
1568 static void
1569 ExecEndPlan(PlanState *planstate, EState *estate)
1570 {
1571         ResultRelInfo *resultRelInfo;
1572         int                     i;
1573         ListCell   *l;
1574
1575         /*
1576          * shut down the node-type-specific query processing
1577          */
1578         ExecEndNode(planstate);
1579
1580         /*
1581          * for subplans too
1582          */
1583         foreach(l, estate->es_subplanstates)
1584         {
1585                 PlanState  *subplanstate = (PlanState *) lfirst(l);
1586
1587                 ExecEndNode(subplanstate);
1588         }
1589
1590         /*
1591          * destroy the executor's tuple table.  Actually we only care about
1592          * releasing buffer pins and tupdesc refcounts; there's no need to pfree
1593          * the TupleTableSlots, since the containing memory context is about to go
1594          * away anyway.
1595          */
1596         ExecResetTupleTable(estate->es_tupleTable, false);
1597
1598         /*
1599          * close the result relation(s) if any, but hold locks until xact commit.
1600          */
1601         resultRelInfo = estate->es_result_relations;
1602         for (i = estate->es_num_result_relations; i > 0; i--)
1603         {
1604                 /* Close indices and then the relation itself */
1605                 ExecCloseIndices(resultRelInfo);
1606                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
1607                 resultRelInfo++;
1608         }
1609
1610         /* Close the root target relation(s). */
1611         resultRelInfo = estate->es_root_result_relations;
1612         for (i = estate->es_num_root_result_relations; i > 0; i--)
1613         {
1614                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
1615                 resultRelInfo++;
1616         }
1617
1618         /* likewise close any trigger target relations */
1619         ExecCleanUpTriggerState(estate);
1620
1621         /*
1622          * close any relations selected FOR [KEY] UPDATE/SHARE, again keeping
1623          * locks
1624          */
1625         foreach(l, estate->es_rowMarks)
1626         {
1627                 ExecRowMark *erm = (ExecRowMark *) lfirst(l);
1628
1629                 if (erm->relation)
1630                         heap_close(erm->relation, NoLock);
1631         }
1632 }
1633
1634 /* ----------------------------------------------------------------
1635  *              ExecutePlan
1636  *
1637  *              Processes the query plan until we have retrieved 'numberTuples' tuples,
1638  *              moving in the specified direction.
1639  *
1640  *              Runs to completion if numberTuples is 0
1641  *
1642  * Note: the ctid attribute is a 'junk' attribute that is removed before the
1643  * user can see it
1644  * ----------------------------------------------------------------
1645  */
1646 static void
1647 ExecutePlan(EState *estate,
1648                         PlanState *planstate,
1649                         bool use_parallel_mode,
1650                         CmdType operation,
1651                         bool sendTuples,
1652                         uint64 numberTuples,
1653                         ScanDirection direction,
1654                         DestReceiver *dest,
1655                         bool execute_once)
1656 {
1657         TupleTableSlot *slot;
1658         uint64          current_tuple_count;
1659
1660         /*
1661          * initialize local variables
1662          */
1663         current_tuple_count = 0;
1664
1665         /*
1666          * Set the direction.
1667          */
1668         estate->es_direction = direction;
1669
1670         /*
1671          * If the plan might potentially be executed multiple times, we must force
1672          * it to run without parallelism, because we might exit early.  Also
1673          * disable parallelism when writing into a relation, because no database
1674          * changes are allowed in parallel mode.
1675          */
1676         if (!execute_once || dest->mydest == DestIntoRel)
1677                 use_parallel_mode = false;
1678
1679         if (use_parallel_mode)
1680                 EnterParallelMode();
1681
1682         /*
1683          * Loop until we've processed the proper number of tuples from the plan.
1684          */
1685         for (;;)
1686         {
1687                 /* Reset the per-output-tuple exprcontext */
1688                 ResetPerTupleExprContext(estate);
1689
1690                 /*
1691                  * Execute the plan and obtain a tuple
1692                  */
1693                 slot = ExecProcNode(planstate);
1694
1695                 /*
1696                  * if the tuple is null, then we assume there is nothing more to
1697                  * process so we just end the loop...
1698                  */
1699                 if (TupIsNull(slot))
1700                 {
1701                         /* Allow nodes to release or shut down resources. */
1702                         (void) ExecShutdownNode(planstate);
1703                         break;
1704                 }
1705
1706                 /*
1707                  * If we have a junk filter, then project a new tuple with the junk
1708                  * removed.
1709                  *
1710                  * Store this new "clean" tuple in the junkfilter's resultSlot.
1711                  * (Formerly, we stored it back over the "dirty" tuple, which is WRONG
1712                  * because that tuple slot has the wrong descriptor.)
1713                  */
1714                 if (estate->es_junkFilter != NULL)
1715                         slot = ExecFilterJunk(estate->es_junkFilter, slot);
1716
1717                 /*
1718                  * If we are supposed to send the tuple somewhere, do so. (In
1719                  * practice, this is probably always the case at this point.)
1720                  */
1721                 if (sendTuples)
1722                 {
1723                         /*
1724                          * If we are not able to send the tuple, we assume the destination
1725                          * has closed and no more tuples can be sent. If that's the case,
1726                          * end the loop.
1727                          */
1728                         if (!((*dest->receiveSlot) (slot, dest)))
1729                                 break;
1730                 }
1731
1732                 /*
1733                  * Count tuples processed, if this is a SELECT.  (For other operation
1734                  * types, the ModifyTable plan node must count the appropriate
1735                  * events.)
1736                  */
1737                 if (operation == CMD_SELECT)
1738                         (estate->es_processed)++;
1739
1740                 /*
1741                  * check our tuple count.. if we've processed the proper number then
1742                  * quit, else loop again and process more tuples.  Zero numberTuples
1743                  * means no limit.
1744                  */
1745                 current_tuple_count++;
1746                 if (numberTuples && numberTuples == current_tuple_count)
1747                 {
1748                         /* Allow nodes to release or shut down resources. */
1749                         (void) ExecShutdownNode(planstate);
1750                         break;
1751                 }
1752         }
1753
1754         if (use_parallel_mode)
1755                 ExitParallelMode();
1756 }
1757
1758
1759 /*
1760  * ExecRelCheck --- check that tuple meets constraints for result relation
1761  *
1762  * Returns NULL if OK, else name of failed check constraint
1763  */
1764 static const char *
1765 ExecRelCheck(ResultRelInfo *resultRelInfo,
1766                          TupleTableSlot *slot, EState *estate)
1767 {
1768         Relation        rel = resultRelInfo->ri_RelationDesc;
1769         int                     ncheck = rel->rd_att->constr->num_check;
1770         ConstrCheck *check = rel->rd_att->constr->check;
1771         ExprContext *econtext;
1772         MemoryContext oldContext;
1773         int                     i;
1774
1775         /*
1776          * If first time through for this result relation, build expression
1777          * nodetrees for rel's constraint expressions.  Keep them in the per-query
1778          * memory context so they'll survive throughout the query.
1779          */
1780         if (resultRelInfo->ri_ConstraintExprs == NULL)
1781         {
1782                 oldContext = MemoryContextSwitchTo(estate->es_query_cxt);
1783                 resultRelInfo->ri_ConstraintExprs =
1784                         (ExprState **) palloc(ncheck * sizeof(ExprState *));
1785                 for (i = 0; i < ncheck; i++)
1786                 {
1787                         Expr       *checkconstr;
1788
1789                         checkconstr = stringToNode(check[i].ccbin);
1790                         resultRelInfo->ri_ConstraintExprs[i] =
1791                                 ExecPrepareExpr(checkconstr, estate);
1792                 }
1793                 MemoryContextSwitchTo(oldContext);
1794         }
1795
1796         /*
1797          * We will use the EState's per-tuple context for evaluating constraint
1798          * expressions (creating it if it's not already there).
1799          */
1800         econtext = GetPerTupleExprContext(estate);
1801
1802         /* Arrange for econtext's scan tuple to be the tuple under test */
1803         econtext->ecxt_scantuple = slot;
1804
1805         /* And evaluate the constraints */
1806         for (i = 0; i < ncheck; i++)
1807         {
1808                 ExprState  *checkconstr = resultRelInfo->ri_ConstraintExprs[i];
1809
1810                 /*
1811                  * NOTE: SQL specifies that a NULL result from a constraint expression
1812                  * is not to be treated as a failure.  Therefore, use ExecCheck not
1813                  * ExecQual.
1814                  */
1815                 if (!ExecCheck(checkconstr, econtext))
1816                         return check[i].ccname;
1817         }
1818
1819         /* NULL result means no error */
1820         return NULL;
1821 }
1822
1823 /*
1824  * ExecPartitionCheck --- check that tuple meets the partition constraint.
1825  */
1826 static void
1827 ExecPartitionCheck(ResultRelInfo *resultRelInfo, TupleTableSlot *slot,
1828                                    EState *estate)
1829 {
1830         Relation        rel = resultRelInfo->ri_RelationDesc;
1831         TupleDesc       tupdesc = RelationGetDescr(rel);
1832         Bitmapset  *modifiedCols;
1833         Bitmapset  *insertedCols;
1834         Bitmapset  *updatedCols;
1835         ExprContext *econtext;
1836
1837         /*
1838          * If first time through, build expression state tree for the partition
1839          * check expression.  Keep it in the per-query memory context so they'll
1840          * survive throughout the query.
1841          */
1842         if (resultRelInfo->ri_PartitionCheckExpr == NULL)
1843         {
1844                 List       *qual = resultRelInfo->ri_PartitionCheck;
1845
1846                 resultRelInfo->ri_PartitionCheckExpr = ExecPrepareCheck(qual, estate);
1847         }
1848
1849         /*
1850          * We will use the EState's per-tuple context for evaluating constraint
1851          * expressions (creating it if it's not already there).
1852          */
1853         econtext = GetPerTupleExprContext(estate);
1854
1855         /* Arrange for econtext's scan tuple to be the tuple under test */
1856         econtext->ecxt_scantuple = slot;
1857
1858         /*
1859          * As in case of the catalogued constraints, we treat a NULL result as
1860          * success here, not a failure.
1861          */
1862         if (!ExecCheck(resultRelInfo->ri_PartitionCheckExpr, econtext))
1863         {
1864                 char       *val_desc;
1865                 Relation        orig_rel = rel;
1866
1867                 /* See the comment above. */
1868                 if (resultRelInfo->ri_PartitionRoot)
1869                 {
1870                         HeapTuple       tuple = ExecFetchSlotTuple(slot);
1871                         TupleDesc       old_tupdesc = RelationGetDescr(rel);
1872                         TupleConversionMap *map;
1873
1874                         rel = resultRelInfo->ri_PartitionRoot;
1875                         tupdesc = RelationGetDescr(rel);
1876                         /* a reverse map */
1877                         map = convert_tuples_by_name(old_tupdesc, tupdesc,
1878                                                                  gettext_noop("could not convert row type"));
1879                         if (map != NULL)
1880                         {
1881                                 tuple = do_convert_tuple(tuple, map);
1882                                 ExecStoreTuple(tuple, slot, InvalidBuffer, false);
1883                         }
1884                 }
1885
1886                 insertedCols = GetInsertedColumns(resultRelInfo, estate);
1887                 updatedCols = GetUpdatedColumns(resultRelInfo, estate);
1888                 modifiedCols = bms_union(insertedCols, updatedCols);
1889                 val_desc = ExecBuildSlotValueDescription(RelationGetRelid(rel),
1890                                                                                                  slot,
1891                                                                                                  tupdesc,
1892                                                                                                  modifiedCols,
1893                                                                                                  64);
1894                 ereport(ERROR,
1895                                 (errcode(ERRCODE_CHECK_VIOLATION),
1896                   errmsg("new row for relation \"%s\" violates partition constraint",
1897                                  RelationGetRelationName(orig_rel)),
1898                         val_desc ? errdetail("Failing row contains %s.", val_desc) : 0));
1899         }
1900 }
1901
1902 /*
1903  * ExecConstraints - check constraints of the tuple in 'slot'
1904  *
1905  * This checks the traditional NOT NULL and check constraints, as well as
1906  * the partition constraint, if any.
1907  *
1908  * Note: 'slot' contains the tuple to check the constraints of, which may
1909  * have been converted from the original input tuple after tuple routing.
1910  * 'resultRelInfo' is the original result relation, before tuple routing.
1911  */
1912 void
1913 ExecConstraints(ResultRelInfo *resultRelInfo,
1914                                 TupleTableSlot *slot, EState *estate)
1915 {
1916         Relation        rel = resultRelInfo->ri_RelationDesc;
1917         TupleDesc       tupdesc = RelationGetDescr(rel);
1918         TupleConstr *constr = tupdesc->constr;
1919         Bitmapset  *modifiedCols;
1920         Bitmapset  *insertedCols;
1921         Bitmapset  *updatedCols;
1922
1923         Assert(constr || resultRelInfo->ri_PartitionCheck);
1924
1925         if (constr && constr->has_not_null)
1926         {
1927                 int                     natts = tupdesc->natts;
1928                 int                     attrChk;
1929
1930                 for (attrChk = 1; attrChk <= natts; attrChk++)
1931                 {
1932                         if (tupdesc->attrs[attrChk - 1]->attnotnull &&
1933                                 slot_attisnull(slot, attrChk))
1934                         {
1935                                 char       *val_desc;
1936                                 Relation        orig_rel = rel;
1937                                 TupleDesc       orig_tupdesc = RelationGetDescr(rel);
1938
1939                                 /*
1940                                  * If the tuple has been routed, it's been converted to the
1941                                  * partition's rowtype, which might differ from the root
1942                                  * table's.  We must convert it back to the root table's
1943                                  * rowtype so that val_desc shown error message matches the
1944                                  * input tuple.
1945                                  */
1946                                 if (resultRelInfo->ri_PartitionRoot)
1947                                 {
1948                                         HeapTuple       tuple = ExecFetchSlotTuple(slot);
1949                                         TupleConversionMap *map;
1950
1951                                         rel = resultRelInfo->ri_PartitionRoot;
1952                                         tupdesc = RelationGetDescr(rel);
1953                                         /* a reverse map */
1954                                         map = convert_tuples_by_name(orig_tupdesc, tupdesc,
1955                                                                  gettext_noop("could not convert row type"));
1956                                         if (map != NULL)
1957                                         {
1958                                                 tuple = do_convert_tuple(tuple, map);
1959                                                 ExecStoreTuple(tuple, slot, InvalidBuffer, false);
1960                                         }
1961                                 }
1962
1963                                 insertedCols = GetInsertedColumns(resultRelInfo, estate);
1964                                 updatedCols = GetUpdatedColumns(resultRelInfo, estate);
1965                                 modifiedCols = bms_union(insertedCols, updatedCols);
1966                                 val_desc = ExecBuildSlotValueDescription(RelationGetRelid(rel),
1967                                                                                                                  slot,
1968                                                                                                                  tupdesc,
1969                                                                                                                  modifiedCols,
1970                                                                                                                  64);
1971
1972                                 ereport(ERROR,
1973                                                 (errcode(ERRCODE_NOT_NULL_VIOLATION),
1974                                                  errmsg("null value in column \"%s\" violates not-null constraint",
1975                                                  NameStr(orig_tupdesc->attrs[attrChk - 1]->attname)),
1976                                                  val_desc ? errdetail("Failing row contains %s.", val_desc) : 0,
1977                                                  errtablecol(orig_rel, attrChk)));
1978                         }
1979                 }
1980         }
1981
1982         if (constr && constr->num_check > 0)
1983         {
1984                 const char *failed;
1985
1986                 if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
1987                 {
1988                         char       *val_desc;
1989                         Relation        orig_rel = rel;
1990
1991                         /* See the comment above. */
1992                         if (resultRelInfo->ri_PartitionRoot)
1993                         {
1994                                 HeapTuple       tuple = ExecFetchSlotTuple(slot);
1995                                 TupleDesc       old_tupdesc = RelationGetDescr(rel);
1996                                 TupleConversionMap *map;
1997
1998                                 rel = resultRelInfo->ri_PartitionRoot;
1999                                 tupdesc = RelationGetDescr(rel);
2000                                 /* a reverse map */
2001                                 map = convert_tuples_by_name(old_tupdesc, tupdesc,
2002                                                                  gettext_noop("could not convert row type"));
2003                                 if (map != NULL)
2004                                 {
2005                                         tuple = do_convert_tuple(tuple, map);
2006                                         ExecStoreTuple(tuple, slot, InvalidBuffer, false);
2007                                 }
2008                         }
2009
2010                         insertedCols = GetInsertedColumns(resultRelInfo, estate);
2011                         updatedCols = GetUpdatedColumns(resultRelInfo, estate);
2012                         modifiedCols = bms_union(insertedCols, updatedCols);
2013                         val_desc = ExecBuildSlotValueDescription(RelationGetRelid(rel),
2014                                                                                                          slot,
2015                                                                                                          tupdesc,
2016                                                                                                          modifiedCols,
2017                                                                                                          64);
2018                         ereport(ERROR,
2019                                         (errcode(ERRCODE_CHECK_VIOLATION),
2020                                          errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
2021                                                         RelationGetRelationName(orig_rel), failed),
2022                           val_desc ? errdetail("Failing row contains %s.", val_desc) : 0,
2023                                          errtableconstraint(orig_rel, failed)));
2024                 }
2025         }
2026
2027         if (resultRelInfo->ri_PartitionCheck)
2028                 ExecPartitionCheck(resultRelInfo, slot, estate);
2029 }
2030
2031
2032 /*
2033  * ExecWithCheckOptions -- check that tuple satisfies any WITH CHECK OPTIONs
2034  * of the specified kind.
2035  *
2036  * Note that this needs to be called multiple times to ensure that all kinds of
2037  * WITH CHECK OPTIONs are handled (both those from views which have the WITH
2038  * CHECK OPTION set and from row level security policies).  See ExecInsert()
2039  * and ExecUpdate().
2040  */
2041 void
2042 ExecWithCheckOptions(WCOKind kind, ResultRelInfo *resultRelInfo,
2043                                          TupleTableSlot *slot, EState *estate)
2044 {
2045         Relation        rel = resultRelInfo->ri_RelationDesc;
2046         TupleDesc       tupdesc = RelationGetDescr(rel);
2047         ExprContext *econtext;
2048         ListCell   *l1,
2049                            *l2;
2050
2051         /*
2052          * We will use the EState's per-tuple context for evaluating constraint
2053          * expressions (creating it if it's not already there).
2054          */
2055         econtext = GetPerTupleExprContext(estate);
2056
2057         /* Arrange for econtext's scan tuple to be the tuple under test */
2058         econtext->ecxt_scantuple = slot;
2059
2060         /* Check each of the constraints */
2061         forboth(l1, resultRelInfo->ri_WithCheckOptions,
2062                         l2, resultRelInfo->ri_WithCheckOptionExprs)
2063         {
2064                 WithCheckOption *wco = (WithCheckOption *) lfirst(l1);
2065                 ExprState  *wcoExpr = (ExprState *) lfirst(l2);
2066
2067                 /*
2068                  * Skip any WCOs which are not the kind we are looking for at this
2069                  * time.
2070                  */
2071                 if (wco->kind != kind)
2072                         continue;
2073
2074                 /*
2075                  * WITH CHECK OPTION checks are intended to ensure that the new tuple
2076                  * is visible (in the case of a view) or that it passes the
2077                  * 'with-check' policy (in the case of row security). If the qual
2078                  * evaluates to NULL or FALSE, then the new tuple won't be included in
2079                  * the view or doesn't pass the 'with-check' policy for the table.
2080                  */
2081                 if (!ExecQual(wcoExpr, econtext))
2082                 {
2083                         char       *val_desc;
2084                         Bitmapset  *modifiedCols;
2085                         Bitmapset  *insertedCols;
2086                         Bitmapset  *updatedCols;
2087
2088                         switch (wco->kind)
2089                         {
2090                                         /*
2091                                          * For WITH CHECK OPTIONs coming from views, we might be
2092                                          * able to provide the details on the row, depending on
2093                                          * the permissions on the relation (that is, if the user
2094                                          * could view it directly anyway).  For RLS violations, we
2095                                          * don't include the data since we don't know if the user
2096                                          * should be able to view the tuple as that depends on the
2097                                          * USING policy.
2098                                          */
2099                                 case WCO_VIEW_CHECK:
2100                                         insertedCols = GetInsertedColumns(resultRelInfo, estate);
2101                                         updatedCols = GetUpdatedColumns(resultRelInfo, estate);
2102                                         modifiedCols = bms_union(insertedCols, updatedCols);
2103                                         val_desc = ExecBuildSlotValueDescription(RelationGetRelid(rel),
2104                                                                                                                          slot,
2105                                                                                                                          tupdesc,
2106                                                                                                                          modifiedCols,
2107                                                                                                                          64);
2108
2109                                         ereport(ERROR,
2110                                                         (errcode(ERRCODE_WITH_CHECK_OPTION_VIOLATION),
2111                                           errmsg("new row violates check option for view \"%s\"",
2112                                                          wco->relname),
2113                                                          val_desc ? errdetail("Failing row contains %s.",
2114                                                                                                   val_desc) : 0));
2115                                         break;
2116                                 case WCO_RLS_INSERT_CHECK:
2117                                 case WCO_RLS_UPDATE_CHECK:
2118                                         if (wco->polname != NULL)
2119                                                 ereport(ERROR,
2120                                                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
2121                                                                  errmsg("new row violates row-level security policy \"%s\" for table \"%s\"",
2122                                                                                 wco->polname, wco->relname)));
2123                                         else
2124                                                 ereport(ERROR,
2125                                                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
2126                                                                  errmsg("new row violates row-level security policy for table \"%s\"",
2127                                                                                 wco->relname)));
2128                                         break;
2129                                 case WCO_RLS_CONFLICT_CHECK:
2130                                         if (wco->polname != NULL)
2131                                                 ereport(ERROR,
2132                                                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
2133                                                                  errmsg("new row violates row-level security policy \"%s\" (USING expression) for table \"%s\"",
2134                                                                                 wco->polname, wco->relname)));
2135                                         else
2136                                                 ereport(ERROR,
2137                                                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
2138                                                                  errmsg("new row violates row-level security policy (USING expression) for table \"%s\"",
2139                                                                                 wco->relname)));
2140                                         break;
2141                                 default:
2142                                         elog(ERROR, "unrecognized WCO kind: %u", wco->kind);
2143                                         break;
2144                         }
2145                 }
2146         }
2147 }
2148
2149 /*
2150  * ExecBuildSlotValueDescription -- construct a string representing a tuple
2151  *
2152  * This is intentionally very similar to BuildIndexValueDescription, but
2153  * unlike that function, we truncate long field values (to at most maxfieldlen
2154  * bytes).  That seems necessary here since heap field values could be very
2155  * long, whereas index entries typically aren't so wide.
2156  *
2157  * Also, unlike the case with index entries, we need to be prepared to ignore
2158  * dropped columns.  We used to use the slot's tuple descriptor to decode the
2159  * data, but the slot's descriptor doesn't identify dropped columns, so we
2160  * now need to be passed the relation's descriptor.
2161  *
2162  * Note that, like BuildIndexValueDescription, if the user does not have
2163  * permission to view any of the columns involved, a NULL is returned.  Unlike
2164  * BuildIndexValueDescription, if the user has access to view a subset of the
2165  * column involved, that subset will be returned with a key identifying which
2166  * columns they are.
2167  */
2168 static char *
2169 ExecBuildSlotValueDescription(Oid reloid,
2170                                                           TupleTableSlot *slot,
2171                                                           TupleDesc tupdesc,
2172                                                           Bitmapset *modifiedCols,
2173                                                           int maxfieldlen)
2174 {
2175         StringInfoData buf;
2176         StringInfoData collist;
2177         bool            write_comma = false;
2178         bool            write_comma_collist = false;
2179         int                     i;
2180         AclResult       aclresult;
2181         bool            table_perm = false;
2182         bool            any_perm = false;
2183
2184         /*
2185          * Check if RLS is enabled and should be active for the relation; if so,
2186          * then don't return anything.  Otherwise, go through normal permission
2187          * checks.
2188          */
2189         if (check_enable_rls(reloid, InvalidOid, true) == RLS_ENABLED)
2190                 return NULL;
2191
2192         initStringInfo(&buf);
2193
2194         appendStringInfoChar(&buf, '(');
2195
2196         /*
2197          * Check if the user has permissions to see the row.  Table-level SELECT
2198          * allows access to all columns.  If the user does not have table-level
2199          * SELECT then we check each column and include those the user has SELECT
2200          * rights on.  Additionally, we always include columns the user provided
2201          * data for.
2202          */
2203         aclresult = pg_class_aclcheck(reloid, GetUserId(), ACL_SELECT);
2204         if (aclresult != ACLCHECK_OK)
2205         {
2206                 /* Set up the buffer for the column list */
2207                 initStringInfo(&collist);
2208                 appendStringInfoChar(&collist, '(');
2209         }
2210         else
2211                 table_perm = any_perm = true;
2212
2213         /* Make sure the tuple is fully deconstructed */
2214         slot_getallattrs(slot);
2215
2216         for (i = 0; i < tupdesc->natts; i++)
2217         {
2218                 bool            column_perm = false;
2219                 char       *val;
2220                 int                     vallen;
2221
2222                 /* ignore dropped columns */
2223                 if (tupdesc->attrs[i]->attisdropped)
2224                         continue;
2225
2226                 if (!table_perm)
2227                 {
2228                         /*
2229                          * No table-level SELECT, so need to make sure they either have
2230                          * SELECT rights on the column or that they have provided the data
2231                          * for the column.  If not, omit this column from the error
2232                          * message.
2233                          */
2234                         aclresult = pg_attribute_aclcheck(reloid, tupdesc->attrs[i]->attnum,
2235                                                                                           GetUserId(), ACL_SELECT);
2236                         if (bms_is_member(tupdesc->attrs[i]->attnum - FirstLowInvalidHeapAttributeNumber,
2237                                                           modifiedCols) || aclresult == ACLCHECK_OK)
2238                         {
2239                                 column_perm = any_perm = true;
2240
2241                                 if (write_comma_collist)
2242                                         appendStringInfoString(&collist, ", ");
2243                                 else
2244                                         write_comma_collist = true;
2245
2246                                 appendStringInfoString(&collist, NameStr(tupdesc->attrs[i]->attname));
2247                         }
2248                 }
2249
2250                 if (table_perm || column_perm)
2251                 {
2252                         if (slot->tts_isnull[i])
2253                                 val = "null";
2254                         else
2255                         {
2256                                 Oid                     foutoid;
2257                                 bool            typisvarlena;
2258
2259                                 getTypeOutputInfo(tupdesc->attrs[i]->atttypid,
2260                                                                   &foutoid, &typisvarlena);
2261                                 val = OidOutputFunctionCall(foutoid, slot->tts_values[i]);
2262                         }
2263
2264                         if (write_comma)
2265                                 appendStringInfoString(&buf, ", ");
2266                         else
2267                                 write_comma = true;
2268
2269                         /* truncate if needed */
2270                         vallen = strlen(val);
2271                         if (vallen <= maxfieldlen)
2272                                 appendStringInfoString(&buf, val);
2273                         else
2274                         {
2275                                 vallen = pg_mbcliplen(val, vallen, maxfieldlen);
2276                                 appendBinaryStringInfo(&buf, val, vallen);
2277                                 appendStringInfoString(&buf, "...");
2278                         }
2279                 }
2280         }
2281
2282         /* If we end up with zero columns being returned, then return NULL. */
2283         if (!any_perm)
2284                 return NULL;
2285
2286         appendStringInfoChar(&buf, ')');
2287
2288         if (!table_perm)
2289         {
2290                 appendStringInfoString(&collist, ") = ");
2291                 appendStringInfoString(&collist, buf.data);
2292
2293                 return collist.data;
2294         }
2295
2296         return buf.data;
2297 }
2298
2299
2300 /*
2301  * ExecUpdateLockMode -- find the appropriate UPDATE tuple lock mode for a
2302  * given ResultRelInfo
2303  */
2304 LockTupleMode
2305 ExecUpdateLockMode(EState *estate, ResultRelInfo *relinfo)
2306 {
2307         Bitmapset  *keyCols;
2308         Bitmapset  *updatedCols;
2309
2310         /*
2311          * Compute lock mode to use.  If columns that are part of the key have not
2312          * been modified, then we can use a weaker lock, allowing for better
2313          * concurrency.
2314          */
2315         updatedCols = GetUpdatedColumns(relinfo, estate);
2316         keyCols = RelationGetIndexAttrBitmap(relinfo->ri_RelationDesc,
2317                                                                                  INDEX_ATTR_BITMAP_KEY);
2318
2319         if (bms_overlap(keyCols, updatedCols))
2320                 return LockTupleExclusive;
2321
2322         return LockTupleNoKeyExclusive;
2323 }
2324
2325 /*
2326  * ExecFindRowMark -- find the ExecRowMark struct for given rangetable index
2327  *
2328  * If no such struct, either return NULL or throw error depending on missing_ok
2329  */
2330 ExecRowMark *
2331 ExecFindRowMark(EState *estate, Index rti, bool missing_ok)
2332 {
2333         ListCell   *lc;
2334
2335         foreach(lc, estate->es_rowMarks)
2336         {
2337                 ExecRowMark *erm = (ExecRowMark *) lfirst(lc);
2338
2339                 if (erm->rti == rti)
2340                         return erm;
2341         }
2342         if (!missing_ok)
2343                 elog(ERROR, "failed to find ExecRowMark for rangetable index %u", rti);
2344         return NULL;
2345 }
2346
2347 /*
2348  * ExecBuildAuxRowMark -- create an ExecAuxRowMark struct
2349  *
2350  * Inputs are the underlying ExecRowMark struct and the targetlist of the
2351  * input plan node (not planstate node!).  We need the latter to find out
2352  * the column numbers of the resjunk columns.
2353  */
2354 ExecAuxRowMark *
2355 ExecBuildAuxRowMark(ExecRowMark *erm, List *targetlist)
2356 {
2357         ExecAuxRowMark *aerm = (ExecAuxRowMark *) palloc0(sizeof(ExecAuxRowMark));
2358         char            resname[32];
2359
2360         aerm->rowmark = erm;
2361
2362         /* Look up the resjunk columns associated with this rowmark */
2363         if (erm->markType != ROW_MARK_COPY)
2364         {
2365                 /* need ctid for all methods other than COPY */
2366                 snprintf(resname, sizeof(resname), "ctid%u", erm->rowmarkId);
2367                 aerm->ctidAttNo = ExecFindJunkAttributeInTlist(targetlist,
2368                                                                                                            resname);
2369                 if (!AttributeNumberIsValid(aerm->ctidAttNo))
2370                         elog(ERROR, "could not find junk %s column", resname);
2371         }
2372         else
2373         {
2374                 /* need wholerow if COPY */
2375                 snprintf(resname, sizeof(resname), "wholerow%u", erm->rowmarkId);
2376                 aerm->wholeAttNo = ExecFindJunkAttributeInTlist(targetlist,
2377                                                                                                                 resname);
2378                 if (!AttributeNumberIsValid(aerm->wholeAttNo))
2379                         elog(ERROR, "could not find junk %s column", resname);
2380         }
2381
2382         /* if child rel, need tableoid */
2383         if (erm->rti != erm->prti)
2384         {
2385                 snprintf(resname, sizeof(resname), "tableoid%u", erm->rowmarkId);
2386                 aerm->toidAttNo = ExecFindJunkAttributeInTlist(targetlist,
2387                                                                                                            resname);
2388                 if (!AttributeNumberIsValid(aerm->toidAttNo))
2389                         elog(ERROR, "could not find junk %s column", resname);
2390         }
2391
2392         return aerm;
2393 }
2394
2395
2396 /*
2397  * EvalPlanQual logic --- recheck modified tuple(s) to see if we want to
2398  * process the updated version under READ COMMITTED rules.
2399  *
2400  * See backend/executor/README for some info about how this works.
2401  */
2402
2403
2404 /*
2405  * Check a modified tuple to see if we want to process its updated version
2406  * under READ COMMITTED rules.
2407  *
2408  *      estate - outer executor state data
2409  *      epqstate - state for EvalPlanQual rechecking
2410  *      relation - table containing tuple
2411  *      rti - rangetable index of table containing tuple
2412  *      lockmode - requested tuple lock mode
2413  *      *tid - t_ctid from the outdated tuple (ie, next updated version)
2414  *      priorXmax - t_xmax from the outdated tuple
2415  *
2416  * *tid is also an output parameter: it's modified to hold the TID of the
2417  * latest version of the tuple (note this may be changed even on failure)
2418  *
2419  * Returns a slot containing the new candidate update/delete tuple, or
2420  * NULL if we determine we shouldn't process the row.
2421  *
2422  * Note: properly, lockmode should be declared as enum LockTupleMode,
2423  * but we use "int" to avoid having to include heapam.h in executor.h.
2424  */
2425 TupleTableSlot *
2426 EvalPlanQual(EState *estate, EPQState *epqstate,
2427                          Relation relation, Index rti, int lockmode,
2428                          ItemPointer tid, TransactionId priorXmax)
2429 {
2430         TupleTableSlot *slot;
2431         HeapTuple       copyTuple;
2432
2433         Assert(rti > 0);
2434
2435         /*
2436          * Get and lock the updated version of the row; if fail, return NULL.
2437          */
2438         copyTuple = EvalPlanQualFetch(estate, relation, lockmode, LockWaitBlock,
2439                                                                   tid, priorXmax);
2440
2441         if (copyTuple == NULL)
2442                 return NULL;
2443
2444         /*
2445          * For UPDATE/DELETE we have to return tid of actual row we're executing
2446          * PQ for.
2447          */
2448         *tid = copyTuple->t_self;
2449
2450         /*
2451          * Need to run a recheck subquery.  Initialize or reinitialize EPQ state.
2452          */
2453         EvalPlanQualBegin(epqstate, estate);
2454
2455         /*
2456          * Free old test tuple, if any, and store new tuple where relation's scan
2457          * node will see it
2458          */
2459         EvalPlanQualSetTuple(epqstate, rti, copyTuple);
2460
2461         /*
2462          * Fetch any non-locked source rows
2463          */
2464         EvalPlanQualFetchRowMarks(epqstate);
2465
2466         /*
2467          * Run the EPQ query.  We assume it will return at most one tuple.
2468          */
2469         slot = EvalPlanQualNext(epqstate);
2470
2471         /*
2472          * If we got a tuple, force the slot to materialize the tuple so that it
2473          * is not dependent on any local state in the EPQ query (in particular,
2474          * it's highly likely that the slot contains references to any pass-by-ref
2475          * datums that may be present in copyTuple).  As with the next step, this
2476          * is to guard against early re-use of the EPQ query.
2477          */
2478         if (!TupIsNull(slot))
2479                 (void) ExecMaterializeSlot(slot);
2480
2481         /*
2482          * Clear out the test tuple.  This is needed in case the EPQ query is
2483          * re-used to test a tuple for a different relation.  (Not clear that can
2484          * really happen, but let's be safe.)
2485          */
2486         EvalPlanQualSetTuple(epqstate, rti, NULL);
2487
2488         return slot;
2489 }
2490
2491 /*
2492  * Fetch a copy of the newest version of an outdated tuple
2493  *
2494  *      estate - executor state data
2495  *      relation - table containing tuple
2496  *      lockmode - requested tuple lock mode
2497  *      wait_policy - requested lock wait policy
2498  *      *tid - t_ctid from the outdated tuple (ie, next updated version)
2499  *      priorXmax - t_xmax from the outdated tuple
2500  *
2501  * Returns a palloc'd copy of the newest tuple version, or NULL if we find
2502  * that there is no newest version (ie, the row was deleted not updated).
2503  * We also return NULL if the tuple is locked and the wait policy is to skip
2504  * such tuples.
2505  *
2506  * If successful, we have locked the newest tuple version, so caller does not
2507  * need to worry about it changing anymore.
2508  *
2509  * Note: properly, lockmode should be declared as enum LockTupleMode,
2510  * but we use "int" to avoid having to include heapam.h in executor.h.
2511  */
2512 HeapTuple
2513 EvalPlanQualFetch(EState *estate, Relation relation, int lockmode,
2514                                   LockWaitPolicy wait_policy,
2515                                   ItemPointer tid, TransactionId priorXmax)
2516 {
2517         HeapTuple       copyTuple = NULL;
2518         HeapTupleData tuple;
2519         SnapshotData SnapshotDirty;
2520
2521         /*
2522          * fetch target tuple
2523          *
2524          * Loop here to deal with updated or busy tuples
2525          */
2526         InitDirtySnapshot(SnapshotDirty);
2527         tuple.t_self = *tid;
2528         for (;;)
2529         {
2530                 Buffer          buffer;
2531
2532                 if (heap_fetch(relation, &SnapshotDirty, &tuple, &buffer, true, NULL))
2533                 {
2534                         HTSU_Result test;
2535                         HeapUpdateFailureData hufd;
2536
2537                         /*
2538                          * If xmin isn't what we're expecting, the slot must have been
2539                          * recycled and reused for an unrelated tuple.  This implies that
2540                          * the latest version of the row was deleted, so we need do
2541                          * nothing.  (Should be safe to examine xmin without getting
2542                          * buffer's content lock.  We assume reading a TransactionId to be
2543                          * atomic, and Xmin never changes in an existing tuple, except to
2544                          * invalid or frozen, and neither of those can match priorXmax.)
2545                          */
2546                         if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
2547                                                                          priorXmax))
2548                         {
2549                                 ReleaseBuffer(buffer);
2550                                 return NULL;
2551                         }
2552
2553                         /* otherwise xmin should not be dirty... */
2554                         if (TransactionIdIsValid(SnapshotDirty.xmin))
2555                                 elog(ERROR, "t_xmin is uncommitted in tuple to be updated");
2556
2557                         /*
2558                          * If tuple is being updated by other transaction then we have to
2559                          * wait for its commit/abort, or die trying.
2560                          */
2561                         if (TransactionIdIsValid(SnapshotDirty.xmax))
2562                         {
2563                                 ReleaseBuffer(buffer);
2564                                 switch (wait_policy)
2565                                 {
2566                                         case LockWaitBlock:
2567                                                 XactLockTableWait(SnapshotDirty.xmax,
2568                                                                                   relation, &tuple.t_self,
2569                                                                                   XLTW_FetchUpdated);
2570                                                 break;
2571                                         case LockWaitSkip:
2572                                                 if (!ConditionalXactLockTableWait(SnapshotDirty.xmax))
2573                                                         return NULL;            /* skip instead of waiting */
2574                                                 break;
2575                                         case LockWaitError:
2576                                                 if (!ConditionalXactLockTableWait(SnapshotDirty.xmax))
2577                                                         ereport(ERROR,
2578                                                                         (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
2579                                                                          errmsg("could not obtain lock on row in relation \"%s\"",
2580                                                                                 RelationGetRelationName(relation))));
2581                                                 break;
2582                                 }
2583                                 continue;               /* loop back to repeat heap_fetch */
2584                         }
2585
2586                         /*
2587                          * If tuple was inserted by our own transaction, we have to check
2588                          * cmin against es_output_cid: cmin >= current CID means our
2589                          * command cannot see the tuple, so we should ignore it. Otherwise
2590                          * heap_lock_tuple() will throw an error, and so would any later
2591                          * attempt to update or delete the tuple.  (We need not check cmax
2592                          * because HeapTupleSatisfiesDirty will consider a tuple deleted
2593                          * by our transaction dead, regardless of cmax.) We just checked
2594                          * that priorXmax == xmin, so we can test that variable instead of
2595                          * doing HeapTupleHeaderGetXmin again.
2596                          */
2597                         if (TransactionIdIsCurrentTransactionId(priorXmax) &&
2598                                 HeapTupleHeaderGetCmin(tuple.t_data) >= estate->es_output_cid)
2599                         {
2600                                 ReleaseBuffer(buffer);
2601                                 return NULL;
2602                         }
2603
2604                         /*
2605                          * This is a live tuple, so now try to lock it.
2606                          */
2607                         test = heap_lock_tuple(relation, &tuple,
2608                                                                    estate->es_output_cid,
2609                                                                    lockmode, wait_policy,
2610                                                                    false, &buffer, &hufd);
2611                         /* We now have two pins on the buffer, get rid of one */
2612                         ReleaseBuffer(buffer);
2613
2614                         switch (test)
2615                         {
2616                                 case HeapTupleSelfUpdated:
2617
2618                                         /*
2619                                          * The target tuple was already updated or deleted by the
2620                                          * current command, or by a later command in the current
2621                                          * transaction.  We *must* ignore the tuple in the former
2622                                          * case, so as to avoid the "Halloween problem" of
2623                                          * repeated update attempts.  In the latter case it might
2624                                          * be sensible to fetch the updated tuple instead, but
2625                                          * doing so would require changing heap_update and
2626                                          * heap_delete to not complain about updating "invisible"
2627                                          * tuples, which seems pretty scary (heap_lock_tuple will
2628                                          * not complain, but few callers expect
2629                                          * HeapTupleInvisible, and we're not one of them).  So for
2630                                          * now, treat the tuple as deleted and do not process.
2631                                          */
2632                                         ReleaseBuffer(buffer);
2633                                         return NULL;
2634
2635                                 case HeapTupleMayBeUpdated:
2636                                         /* successfully locked */
2637                                         break;
2638
2639                                 case HeapTupleUpdated:
2640                                         ReleaseBuffer(buffer);
2641                                         if (IsolationUsesXactSnapshot())
2642                                                 ereport(ERROR,
2643                                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
2644                                                                  errmsg("could not serialize access due to concurrent update")));
2645
2646                                         /* Should not encounter speculative tuple on recheck */
2647                                         Assert(!HeapTupleHeaderIsSpeculative(tuple.t_data));
2648                                         if (!ItemPointerEquals(&hufd.ctid, &tuple.t_self))
2649                                         {
2650                                                 /* it was updated, so look at the updated version */
2651                                                 tuple.t_self = hufd.ctid;
2652                                                 /* updated row should have xmin matching this xmax */
2653                                                 priorXmax = hufd.xmax;
2654                                                 continue;
2655                                         }
2656                                         /* tuple was deleted, so give up */
2657                                         return NULL;
2658
2659                                 case HeapTupleWouldBlock:
2660                                         ReleaseBuffer(buffer);
2661                                         return NULL;
2662
2663                                 case HeapTupleInvisible:
2664                                         elog(ERROR, "attempted to lock invisible tuple");
2665
2666                                 default:
2667                                         ReleaseBuffer(buffer);
2668                                         elog(ERROR, "unrecognized heap_lock_tuple status: %u",
2669                                                  test);
2670                                         return NULL;    /* keep compiler quiet */
2671                         }
2672
2673                         /*
2674                          * We got tuple - now copy it for use by recheck query.
2675                          */
2676                         copyTuple = heap_copytuple(&tuple);
2677                         ReleaseBuffer(buffer);
2678                         break;
2679                 }
2680
2681                 /*
2682                  * If the referenced slot was actually empty, the latest version of
2683                  * the row must have been deleted, so we need do nothing.
2684                  */
2685                 if (tuple.t_data == NULL)
2686                 {
2687                         ReleaseBuffer(buffer);
2688                         return NULL;
2689                 }
2690
2691                 /*
2692                  * As above, if xmin isn't what we're expecting, do nothing.
2693                  */
2694                 if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
2695                                                                  priorXmax))
2696                 {
2697                         ReleaseBuffer(buffer);
2698                         return NULL;
2699                 }
2700
2701                 /*
2702                  * If we get here, the tuple was found but failed SnapshotDirty.
2703                  * Assuming the xmin is either a committed xact or our own xact (as it
2704                  * certainly should be if we're trying to modify the tuple), this must
2705                  * mean that the row was updated or deleted by either a committed xact
2706                  * or our own xact.  If it was deleted, we can ignore it; if it was
2707                  * updated then chain up to the next version and repeat the whole
2708                  * process.
2709                  *
2710                  * As above, it should be safe to examine xmax and t_ctid without the
2711                  * buffer content lock, because they can't be changing.
2712                  */
2713                 if (ItemPointerEquals(&tuple.t_self, &tuple.t_data->t_ctid))
2714                 {
2715                         /* deleted, so forget about it */
2716                         ReleaseBuffer(buffer);
2717                         return NULL;
2718                 }
2719
2720                 /* updated, so look at the updated row */
2721                 tuple.t_self = tuple.t_data->t_ctid;
2722                 /* updated row should have xmin matching this xmax */
2723                 priorXmax = HeapTupleHeaderGetUpdateXid(tuple.t_data);
2724                 ReleaseBuffer(buffer);
2725                 /* loop back to fetch next in chain */
2726         }
2727
2728         /*
2729          * Return the copied tuple
2730          */
2731         return copyTuple;
2732 }
2733
2734 /*
2735  * EvalPlanQualInit -- initialize during creation of a plan state node
2736  * that might need to invoke EPQ processing.
2737  *
2738  * Note: subplan/auxrowmarks can be NULL/NIL if they will be set later
2739  * with EvalPlanQualSetPlan.
2740  */
2741 void
2742 EvalPlanQualInit(EPQState *epqstate, EState *estate,
2743                                  Plan *subplan, List *auxrowmarks, int epqParam)
2744 {
2745         /* Mark the EPQ state inactive */
2746         epqstate->estate = NULL;
2747         epqstate->planstate = NULL;
2748         epqstate->origslot = NULL;
2749         /* ... and remember data that EvalPlanQualBegin will need */
2750         epqstate->plan = subplan;
2751         epqstate->arowMarks = auxrowmarks;
2752         epqstate->epqParam = epqParam;
2753 }
2754
2755 /*
2756  * EvalPlanQualSetPlan -- set or change subplan of an EPQState.
2757  *
2758  * We need this so that ModifyTable can deal with multiple subplans.
2759  */
2760 void
2761 EvalPlanQualSetPlan(EPQState *epqstate, Plan *subplan, List *auxrowmarks)
2762 {
2763         /* If we have a live EPQ query, shut it down */
2764         EvalPlanQualEnd(epqstate);
2765         /* And set/change the plan pointer */
2766         epqstate->plan = subplan;
2767         /* The rowmarks depend on the plan, too */
2768         epqstate->arowMarks = auxrowmarks;
2769 }
2770
2771 /*
2772  * Install one test tuple into EPQ state, or clear test tuple if tuple == NULL
2773  *
2774  * NB: passed tuple must be palloc'd; it may get freed later
2775  */
2776 void
2777 EvalPlanQualSetTuple(EPQState *epqstate, Index rti, HeapTuple tuple)
2778 {
2779         EState     *estate = epqstate->estate;
2780
2781         Assert(rti > 0);
2782
2783         /*
2784          * free old test tuple, if any, and store new tuple where relation's scan
2785          * node will see it
2786          */
2787         if (estate->es_epqTuple[rti - 1] != NULL)
2788                 heap_freetuple(estate->es_epqTuple[rti - 1]);
2789         estate->es_epqTuple[rti - 1] = tuple;
2790         estate->es_epqTupleSet[rti - 1] = true;
2791 }
2792
2793 /*
2794  * Fetch back the current test tuple (if any) for the specified RTI
2795  */
2796 HeapTuple
2797 EvalPlanQualGetTuple(EPQState *epqstate, Index rti)
2798 {
2799         EState     *estate = epqstate->estate;
2800
2801         Assert(rti > 0);
2802
2803         return estate->es_epqTuple[rti - 1];
2804 }
2805
2806 /*
2807  * Fetch the current row values for any non-locked relations that need
2808  * to be scanned by an EvalPlanQual operation.  origslot must have been set
2809  * to contain the current result row (top-level row) that we need to recheck.
2810  */
2811 void
2812 EvalPlanQualFetchRowMarks(EPQState *epqstate)
2813 {
2814         ListCell   *l;
2815
2816         Assert(epqstate->origslot != NULL);
2817
2818         foreach(l, epqstate->arowMarks)
2819         {
2820                 ExecAuxRowMark *aerm = (ExecAuxRowMark *) lfirst(l);
2821                 ExecRowMark *erm = aerm->rowmark;
2822                 Datum           datum;
2823                 bool            isNull;
2824                 HeapTupleData tuple;
2825
2826                 if (RowMarkRequiresRowShareLock(erm->markType))
2827                         elog(ERROR, "EvalPlanQual doesn't support locking rowmarks");
2828
2829                 /* clear any leftover test tuple for this rel */
2830                 EvalPlanQualSetTuple(epqstate, erm->rti, NULL);
2831
2832                 /* if child rel, must check whether it produced this row */
2833                 if (erm->rti != erm->prti)
2834                 {
2835                         Oid                     tableoid;
2836
2837                         datum = ExecGetJunkAttribute(epqstate->origslot,
2838                                                                                  aerm->toidAttNo,
2839                                                                                  &isNull);
2840                         /* non-locked rels could be on the inside of outer joins */
2841                         if (isNull)
2842                                 continue;
2843                         tableoid = DatumGetObjectId(datum);
2844
2845                         Assert(OidIsValid(erm->relid));
2846                         if (tableoid != erm->relid)
2847                         {
2848                                 /* this child is inactive right now */
2849                                 continue;
2850                         }
2851                 }
2852
2853                 if (erm->markType == ROW_MARK_REFERENCE)
2854                 {
2855                         HeapTuple       copyTuple;
2856
2857                         Assert(erm->relation != NULL);
2858
2859                         /* fetch the tuple's ctid */
2860                         datum = ExecGetJunkAttribute(epqstate->origslot,
2861                                                                                  aerm->ctidAttNo,
2862                                                                                  &isNull);
2863                         /* non-locked rels could be on the inside of outer joins */
2864                         if (isNull)
2865                                 continue;
2866
2867                         /* fetch requests on foreign tables must be passed to their FDW */
2868                         if (erm->relation->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
2869                         {
2870                                 FdwRoutine *fdwroutine;
2871                                 bool            updated = false;
2872
2873                                 fdwroutine = GetFdwRoutineForRelation(erm->relation, false);
2874                                 /* this should have been checked already, but let's be safe */
2875                                 if (fdwroutine->RefetchForeignRow == NULL)
2876                                         ereport(ERROR,
2877                                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2878                                                    errmsg("cannot lock rows in foreign table \"%s\"",
2879                                                                   RelationGetRelationName(erm->relation))));
2880                                 copyTuple = fdwroutine->RefetchForeignRow(epqstate->estate,
2881                                                                                                                   erm,
2882                                                                                                                   datum,
2883                                                                                                                   &updated);
2884                                 if (copyTuple == NULL)
2885                                         elog(ERROR, "failed to fetch tuple for EvalPlanQual recheck");
2886
2887                                 /*
2888                                  * Ideally we'd insist on updated == false here, but that
2889                                  * assumes that FDWs can track that exactly, which they might
2890                                  * not be able to.  So just ignore the flag.
2891                                  */
2892                         }
2893                         else
2894                         {
2895                                 /* ordinary table, fetch the tuple */
2896                                 Buffer          buffer;
2897
2898                                 tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
2899                                 if (!heap_fetch(erm->relation, SnapshotAny, &tuple, &buffer,
2900                                                                 false, NULL))
2901                                         elog(ERROR, "failed to fetch tuple for EvalPlanQual recheck");
2902
2903                                 /* successful, copy tuple */
2904                                 copyTuple = heap_copytuple(&tuple);
2905                                 ReleaseBuffer(buffer);
2906                         }
2907
2908                         /* store tuple */
2909                         EvalPlanQualSetTuple(epqstate, erm->rti, copyTuple);
2910                 }
2911                 else
2912                 {
2913                         HeapTupleHeader td;
2914
2915                         Assert(erm->markType == ROW_MARK_COPY);
2916
2917                         /* fetch the whole-row Var for the relation */
2918                         datum = ExecGetJunkAttribute(epqstate->origslot,
2919                                                                                  aerm->wholeAttNo,
2920                                                                                  &isNull);
2921                         /* non-locked rels could be on the inside of outer joins */
2922                         if (isNull)
2923                                 continue;
2924                         td = DatumGetHeapTupleHeader(datum);
2925
2926                         /* build a temporary HeapTuple control structure */
2927                         tuple.t_len = HeapTupleHeaderGetDatumLength(td);
2928                         tuple.t_data = td;
2929                         /* relation might be a foreign table, if so provide tableoid */
2930                         tuple.t_tableOid = erm->relid;
2931                         /* also copy t_ctid in case there's valid data there */
2932                         tuple.t_self = td->t_ctid;
2933
2934                         /* copy and store tuple */
2935                         EvalPlanQualSetTuple(epqstate, erm->rti,
2936                                                                  heap_copytuple(&tuple));
2937                 }
2938         }
2939 }
2940
2941 /*
2942  * Fetch the next row (if any) from EvalPlanQual testing
2943  *
2944  * (In practice, there should never be more than one row...)
2945  */
2946 TupleTableSlot *
2947 EvalPlanQualNext(EPQState *epqstate)
2948 {
2949         MemoryContext oldcontext;
2950         TupleTableSlot *slot;
2951
2952         oldcontext = MemoryContextSwitchTo(epqstate->estate->es_query_cxt);
2953         slot = ExecProcNode(epqstate->planstate);
2954         MemoryContextSwitchTo(oldcontext);
2955
2956         return slot;
2957 }
2958
2959 /*
2960  * Initialize or reset an EvalPlanQual state tree
2961  */
2962 void
2963 EvalPlanQualBegin(EPQState *epqstate, EState *parentestate)
2964 {
2965         EState     *estate = epqstate->estate;
2966
2967         if (estate == NULL)
2968         {
2969                 /* First time through, so create a child EState */
2970                 EvalPlanQualStart(epqstate, parentestate, epqstate->plan);
2971         }
2972         else
2973         {
2974                 /*
2975                  * We already have a suitable child EPQ tree, so just reset it.
2976                  */
2977                 int                     rtsize = list_length(parentestate->es_range_table);
2978                 PlanState  *planstate = epqstate->planstate;
2979
2980                 MemSet(estate->es_epqScanDone, 0, rtsize * sizeof(bool));
2981
2982                 /* Recopy current values of parent parameters */
2983                 if (parentestate->es_plannedstmt->nParamExec > 0)
2984                 {
2985                         int                     i = parentestate->es_plannedstmt->nParamExec;
2986
2987                         while (--i >= 0)
2988                         {
2989                                 /* copy value if any, but not execPlan link */
2990                                 estate->es_param_exec_vals[i].value =
2991                                         parentestate->es_param_exec_vals[i].value;
2992                                 estate->es_param_exec_vals[i].isnull =
2993                                         parentestate->es_param_exec_vals[i].isnull;
2994                         }
2995                 }
2996
2997                 /*
2998                  * Mark child plan tree as needing rescan at all scan nodes.  The
2999                  * first ExecProcNode will take care of actually doing the rescan.
3000                  */
3001                 planstate->chgParam = bms_add_member(planstate->chgParam,
3002                                                                                          epqstate->epqParam);
3003         }
3004 }
3005
3006 /*
3007  * Start execution of an EvalPlanQual plan tree.
3008  *
3009  * This is a cut-down version of ExecutorStart(): we copy some state from
3010  * the top-level estate rather than initializing it fresh.
3011  */
3012 static void
3013 EvalPlanQualStart(EPQState *epqstate, EState *parentestate, Plan *planTree)
3014 {
3015         EState     *estate;
3016         int                     rtsize;
3017         MemoryContext oldcontext;
3018         ListCell   *l;
3019
3020         rtsize = list_length(parentestate->es_range_table);
3021
3022         epqstate->estate = estate = CreateExecutorState();
3023
3024         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
3025
3026         /*
3027          * Child EPQ EStates share the parent's copy of unchanging state such as
3028          * the snapshot, rangetable, result-rel info, and external Param info.
3029          * They need their own copies of local state, including a tuple table,
3030          * es_param_exec_vals, etc.
3031          *
3032          * The ResultRelInfo array management is trickier than it looks.  We
3033          * create a fresh array for the child but copy all the content from the
3034          * parent.  This is because it's okay for the child to share any
3035          * per-relation state the parent has already created --- but if the child
3036          * sets up any ResultRelInfo fields, such as its own junkfilter, that
3037          * state must *not* propagate back to the parent.  (For one thing, the
3038          * pointed-to data is in a memory context that won't last long enough.)
3039          */
3040         estate->es_direction = ForwardScanDirection;
3041         estate->es_snapshot = parentestate->es_snapshot;
3042         estate->es_crosscheck_snapshot = parentestate->es_crosscheck_snapshot;
3043         estate->es_range_table = parentestate->es_range_table;
3044         estate->es_plannedstmt = parentestate->es_plannedstmt;
3045         estate->es_junkFilter = parentestate->es_junkFilter;
3046         estate->es_output_cid = parentestate->es_output_cid;
3047         if (parentestate->es_num_result_relations > 0)
3048         {
3049                 int                     numResultRelations = parentestate->es_num_result_relations;
3050                 ResultRelInfo *resultRelInfos;
3051
3052                 resultRelInfos = (ResultRelInfo *)
3053                         palloc(numResultRelations * sizeof(ResultRelInfo));
3054                 memcpy(resultRelInfos, parentestate->es_result_relations,
3055                            numResultRelations * sizeof(ResultRelInfo));
3056                 estate->es_result_relations = resultRelInfos;
3057                 estate->es_num_result_relations = numResultRelations;
3058         }
3059         /* es_result_relation_info must NOT be copied */
3060         /* es_trig_target_relations must NOT be copied */
3061         estate->es_rowMarks = parentestate->es_rowMarks;
3062         estate->es_top_eflags = parentestate->es_top_eflags;
3063         estate->es_instrument = parentestate->es_instrument;
3064         /* es_auxmodifytables must NOT be copied */
3065
3066         /*
3067          * The external param list is simply shared from parent.  The internal
3068          * param workspace has to be local state, but we copy the initial values
3069          * from the parent, so as to have access to any param values that were
3070          * already set from other parts of the parent's plan tree.
3071          */
3072         estate->es_param_list_info = parentestate->es_param_list_info;
3073         if (parentestate->es_plannedstmt->nParamExec > 0)
3074         {
3075                 int                     i = parentestate->es_plannedstmt->nParamExec;
3076
3077                 estate->es_param_exec_vals = (ParamExecData *)
3078                         palloc0(i * sizeof(ParamExecData));
3079                 while (--i >= 0)
3080                 {
3081                         /* copy value if any, but not execPlan link */
3082                         estate->es_param_exec_vals[i].value =
3083                                 parentestate->es_param_exec_vals[i].value;
3084                         estate->es_param_exec_vals[i].isnull =
3085                                 parentestate->es_param_exec_vals[i].isnull;
3086                 }
3087         }
3088
3089         /*
3090          * Each EState must have its own es_epqScanDone state, but if we have
3091          * nested EPQ checks they should share es_epqTuple arrays.  This allows
3092          * sub-rechecks to inherit the values being examined by an outer recheck.
3093          */
3094         estate->es_epqScanDone = (bool *) palloc0(rtsize * sizeof(bool));
3095         if (parentestate->es_epqTuple != NULL)
3096         {
3097                 estate->es_epqTuple = parentestate->es_epqTuple;
3098                 estate->es_epqTupleSet = parentestate->es_epqTupleSet;
3099         }
3100         else
3101         {
3102                 estate->es_epqTuple = (HeapTuple *)
3103                         palloc0(rtsize * sizeof(HeapTuple));
3104                 estate->es_epqTupleSet = (bool *)
3105                         palloc0(rtsize * sizeof(bool));
3106         }
3107
3108         /*
3109          * Each estate also has its own tuple table.
3110          */
3111         estate->es_tupleTable = NIL;
3112
3113         /*
3114          * Initialize private state information for each SubPlan.  We must do this
3115          * before running ExecInitNode on the main query tree, since
3116          * ExecInitSubPlan expects to be able to find these entries. Some of the
3117          * SubPlans might not be used in the part of the plan tree we intend to
3118          * run, but since it's not easy to tell which, we just initialize them
3119          * all.
3120          */
3121         Assert(estate->es_subplanstates == NIL);
3122         foreach(l, parentestate->es_plannedstmt->subplans)
3123         {
3124                 Plan       *subplan = (Plan *) lfirst(l);
3125                 PlanState  *subplanstate;
3126
3127                 subplanstate = ExecInitNode(subplan, estate, 0);
3128                 estate->es_subplanstates = lappend(estate->es_subplanstates,
3129                                                                                    subplanstate);
3130         }
3131
3132         /*
3133          * Initialize the private state information for all the nodes in the part
3134          * of the plan tree we need to run.  This opens files, allocates storage
3135          * and leaves us ready to start processing tuples.
3136          */
3137         epqstate->planstate = ExecInitNode(planTree, estate, 0);
3138
3139         MemoryContextSwitchTo(oldcontext);
3140 }
3141
3142 /*
3143  * EvalPlanQualEnd -- shut down at termination of parent plan state node,
3144  * or if we are done with the current EPQ child.
3145  *
3146  * This is a cut-down version of ExecutorEnd(); basically we want to do most
3147  * of the normal cleanup, but *not* close result relations (which we are
3148  * just sharing from the outer query).  We do, however, have to close any
3149  * trigger target relations that got opened, since those are not shared.
3150  * (There probably shouldn't be any of the latter, but just in case...)
3151  */
3152 void
3153 EvalPlanQualEnd(EPQState *epqstate)
3154 {
3155         EState     *estate = epqstate->estate;
3156         MemoryContext oldcontext;
3157         ListCell   *l;
3158
3159         if (estate == NULL)
3160                 return;                                 /* idle, so nothing to do */
3161
3162         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
3163
3164         ExecEndNode(epqstate->planstate);
3165
3166         foreach(l, estate->es_subplanstates)
3167         {
3168                 PlanState  *subplanstate = (PlanState *) lfirst(l);
3169
3170                 ExecEndNode(subplanstate);
3171         }
3172
3173         /* throw away the per-estate tuple table */
3174         ExecResetTupleTable(estate->es_tupleTable, false);
3175
3176         /* close any trigger target relations attached to this EState */
3177         ExecCleanUpTriggerState(estate);
3178
3179         MemoryContextSwitchTo(oldcontext);
3180
3181         FreeExecutorState(estate);
3182
3183         /* Mark EPQState idle */
3184         epqstate->estate = NULL;
3185         epqstate->planstate = NULL;
3186         epqstate->origslot = NULL;
3187 }
3188
3189 /*
3190  * ExecSetupPartitionTupleRouting - set up information needed during
3191  * tuple routing for partitioned tables
3192  *
3193  * Output arguments:
3194  * 'pd' receives an array of PartitionDispatch objects with one entry for
3195  *              every partitioned table in the partition tree
3196  * 'partitions' receives an array of ResultRelInfo objects with one entry for
3197  *              every leaf partition in the partition tree
3198  * 'tup_conv_maps' receives an array of TupleConversionMap objects with one
3199  *              entry for every leaf partition (required to convert input tuple based
3200  *              on the root table's rowtype to a leaf partition's rowtype after tuple
3201  *              routing is done
3202  * 'partition_tuple_slot' receives a standalone TupleTableSlot to be used
3203  *              to manipulate any given leaf partition's rowtype after that partition
3204  *              is chosen by tuple-routing.
3205  * 'num_parted' receives the number of partitioned tables in the partition
3206  *              tree (= the number of entries in the 'pd' output array)
3207  * 'num_partitions' receives the number of leaf partitions in the partition
3208  *              tree (= the number of entries in the 'partitions' and 'tup_conv_maps'
3209  *              output arrays
3210  *
3211  * Note that all the relations in the partition tree are locked using the
3212  * RowExclusiveLock mode upon return from this function.
3213  */
3214 void
3215 ExecSetupPartitionTupleRouting(Relation rel,
3216                                                            PartitionDispatch **pd,
3217                                                            ResultRelInfo **partitions,
3218                                                            TupleConversionMap ***tup_conv_maps,
3219                                                            TupleTableSlot **partition_tuple_slot,
3220                                                            int *num_parted, int *num_partitions)
3221 {
3222         TupleDesc       tupDesc = RelationGetDescr(rel);
3223         List       *leaf_parts;
3224         ListCell   *cell;
3225         int                     i;
3226         ResultRelInfo *leaf_part_rri;
3227
3228         /* Get the tuple-routing information and lock partitions */
3229         *pd = RelationGetPartitionDispatchInfo(rel, RowExclusiveLock, num_parted,
3230                                                                                    &leaf_parts);
3231         *num_partitions = list_length(leaf_parts);
3232         *partitions = (ResultRelInfo *) palloc(*num_partitions *
3233                                                                                    sizeof(ResultRelInfo));
3234         *tup_conv_maps = (TupleConversionMap **) palloc0(*num_partitions *
3235                                                                                            sizeof(TupleConversionMap *));
3236
3237         /*
3238          * Initialize an empty slot that will be used to manipulate tuples of any
3239          * given partition's rowtype.  It is attached to the caller-specified node
3240          * (such as ModifyTableState) and released when the node finishes
3241          * processing.
3242          */
3243         *partition_tuple_slot = MakeTupleTableSlot();
3244
3245         leaf_part_rri = *partitions;
3246         i = 0;
3247         foreach(cell, leaf_parts)
3248         {
3249                 Relation        partrel;
3250                 TupleDesc       part_tupdesc;
3251
3252                 /*
3253                  * We locked all the partitions above including the leaf partitions.
3254                  * Note that each of the relations in *partitions are eventually
3255                  * closed by the caller.
3256                  */
3257                 partrel = heap_open(lfirst_oid(cell), NoLock);
3258                 part_tupdesc = RelationGetDescr(partrel);
3259
3260                 /*
3261                  * Verify result relation is a valid target for the current operation.
3262                  */
3263                 CheckValidResultRel(partrel, CMD_INSERT);
3264
3265                 /*
3266                  * Save a tuple conversion map to convert a tuple routed to this
3267                  * partition from the parent's type to the partition's.
3268                  */
3269                 (*tup_conv_maps)[i] = convert_tuples_by_name(tupDesc, part_tupdesc,
3270                                                                  gettext_noop("could not convert row type"));
3271
3272                 InitResultRelInfo(leaf_part_rri,
3273                                                   partrel,
3274                                                   1,    /* dummy */
3275                                                   rel,
3276                                                   0);
3277
3278                 /*
3279                  * Open partition indices (remember we do not support ON CONFLICT in
3280                  * case of partitioned tables, so we do not need support information
3281                  * for speculative insertion)
3282                  */
3283                 if (leaf_part_rri->ri_RelationDesc->rd_rel->relhasindex &&
3284                         leaf_part_rri->ri_IndexRelationDescs == NULL)
3285                         ExecOpenIndices(leaf_part_rri, false);
3286
3287                 leaf_part_rri++;
3288                 i++;
3289         }
3290 }
3291
3292 /*
3293  * ExecFindPartition -- Find a leaf partition in the partition tree rooted
3294  * at parent, for the heap tuple contained in *slot
3295  *
3296  * estate must be non-NULL; we'll need it to compute any expressions in the
3297  * partition key(s)
3298  *
3299  * If no leaf partition is found, this routine errors out with the appropriate
3300  * error message, else it returns the leaf partition sequence number returned
3301  * by get_partition_for_tuple() unchanged.
3302  */
3303 int
3304 ExecFindPartition(ResultRelInfo *resultRelInfo, PartitionDispatch *pd,
3305                                   TupleTableSlot *slot, EState *estate)
3306 {
3307         int                     result;
3308         PartitionDispatchData *failed_at;
3309         TupleTableSlot *failed_slot;
3310
3311         /*
3312          * First check the root table's partition constraint, if any.  No point in
3313          * routing the tuple it if it doesn't belong in the root table itself.
3314          */
3315         if (resultRelInfo->ri_PartitionCheck)
3316                 ExecPartitionCheck(resultRelInfo, slot, estate);
3317
3318         result = get_partition_for_tuple(pd, slot, estate,
3319                                                                          &failed_at, &failed_slot);
3320         if (result < 0)
3321         {
3322                 Relation        failed_rel;
3323                 Datum           key_values[PARTITION_MAX_KEYS];
3324                 bool            key_isnull[PARTITION_MAX_KEYS];
3325                 char       *val_desc;
3326                 ExprContext *ecxt = GetPerTupleExprContext(estate);
3327
3328                 failed_rel = failed_at->reldesc;
3329                 ecxt->ecxt_scantuple = failed_slot;
3330                 FormPartitionKeyDatum(failed_at, failed_slot, estate,
3331                                                           key_values, key_isnull);
3332                 val_desc = ExecBuildSlotPartitionKeyDescription(failed_rel,
3333                                                                                                                 key_values,
3334                                                                                                                 key_isnull,
3335                                                                                                                 64);
3336                 Assert(OidIsValid(RelationGetRelid(failed_rel)));
3337                 ereport(ERROR,
3338                                 (errcode(ERRCODE_CHECK_VIOLATION),
3339                                  errmsg("no partition of relation \"%s\" found for row",
3340                                                 RelationGetRelationName(failed_rel)),
3341                                  val_desc ? errdetail("Partition key of the failing row contains %s.", val_desc) : 0));
3342         }
3343
3344         return result;
3345 }
3346
3347 /*
3348  * BuildSlotPartitionKeyDescription
3349  *
3350  * This works very much like BuildIndexValueDescription() and is currently
3351  * used for building error messages when ExecFindPartition() fails to find
3352  * partition for a row.
3353  */
3354 static char *
3355 ExecBuildSlotPartitionKeyDescription(Relation rel,
3356                                                                          Datum *values,
3357                                                                          bool *isnull,
3358                                                                          int maxfieldlen)
3359 {
3360         StringInfoData buf;
3361         PartitionKey key = RelationGetPartitionKey(rel);
3362         int                     partnatts = get_partition_natts(key);
3363         int                     i;
3364         Oid                     relid = RelationGetRelid(rel);
3365         AclResult       aclresult;
3366
3367         if (check_enable_rls(relid, InvalidOid, true) == RLS_ENABLED)
3368                 return NULL;
3369
3370         /* If the user has table-level access, just go build the description. */
3371         aclresult = pg_class_aclcheck(relid, GetUserId(), ACL_SELECT);
3372         if (aclresult != ACLCHECK_OK)
3373         {
3374                 /*
3375                  * Step through the columns of the partition key and make sure the
3376                  * user has SELECT rights on all of them.
3377                  */
3378                 for (i = 0; i < partnatts; i++)
3379                 {
3380                         AttrNumber      attnum = get_partition_col_attnum(key, i);
3381
3382                         /*
3383                          * If this partition key column is an expression, we return no
3384                          * detail rather than try to figure out what column(s) the
3385                          * expression includes and if the user has SELECT rights on them.
3386                          */
3387                         if (attnum == InvalidAttrNumber ||
3388                                 pg_attribute_aclcheck(relid, attnum, GetUserId(),
3389                                                                           ACL_SELECT) != ACLCHECK_OK)
3390                                 return NULL;
3391                 }
3392         }
3393
3394         initStringInfo(&buf);
3395         appendStringInfo(&buf, "(%s) = (",
3396                                          pg_get_partkeydef_columns(relid, true));
3397
3398         for (i = 0; i < partnatts; i++)
3399         {
3400                 char       *val;
3401                 int                     vallen;
3402
3403                 if (isnull[i])
3404                         val = "null";
3405                 else
3406                 {
3407                         Oid                     foutoid;
3408                         bool            typisvarlena;
3409
3410                         getTypeOutputInfo(get_partition_col_typid(key, i),
3411                                                           &foutoid, &typisvarlena);
3412                         val = OidOutputFunctionCall(foutoid, values[i]);
3413                 }
3414
3415                 if (i > 0)
3416                         appendStringInfoString(&buf, ", ");
3417
3418                 /* truncate if needed */
3419                 vallen = strlen(val);
3420                 if (vallen <= maxfieldlen)
3421                         appendStringInfoString(&buf, val);
3422                 else
3423                 {
3424                         vallen = pg_mbcliplen(val, vallen, maxfieldlen);
3425                         appendBinaryStringInfo(&buf, val, vallen);
3426                         appendStringInfoString(&buf, "...");
3427                 }
3428         }
3429
3430         appendStringInfoChar(&buf, ')');
3431
3432         return buf.data;
3433 }