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

   1 /*-------------------------------------------------------------------------
   2  *
   3  * nodeModifyTable.c
   4  *        routines to handle ModifyTable nodes.
   5  *
   6  * Portions Copyright (c) 1996-2017, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        src/backend/executor/nodeModifyTable.c
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15 /* INTERFACE ROUTINES
  16  *              ExecInitModifyTable - initialize the ModifyTable node
  17  *              ExecModifyTable         - retrieve the next tuple from the node
  18  *              ExecEndModifyTable      - shut down the ModifyTable node
  19  *              ExecReScanModifyTable - rescan the ModifyTable node
  20  *
  21  *       NOTES
  22  *              Each ModifyTable node contains a list of one or more subplans,
  23  *              much like an Append node.  There is one subplan per result relation.
  24  *              The key reason for this is that in an inherited UPDATE command, each
  25  *              result relation could have a different schema (more or different
  26  *              columns) requiring a different plan tree to produce it.  In an
  27  *              inherited DELETE, all the subplans should produce the same output
  28  *              rowtype, but we might still find that different plans are appropriate
  29  *              for different child relations.
  30  *
  31  *              If the query specifies RETURNING, then the ModifyTable returns a
  32  *              RETURNING tuple after completing each row insert, update, or delete.
  33  *              It must be called again to continue the operation.  Without RETURNING,
  34  *              we just loop within the node until all the work is done, then
  35  *              return NULL.  This avoids useless call/return overhead.
  36  */
  37
  38 #include "postgres.h"
  39
  40 #include "access/htup_details.h"
  41 #include "access/xact.h"
  42 #include "commands/trigger.h"
  43 #include "executor/executor.h"
  44 #include "executor/nodeModifyTable.h"
  45 #include "foreign/fdwapi.h"
  46 #include "miscadmin.h"
  47 #include "nodes/nodeFuncs.h"
  48 #include "parser/parsetree.h"
  49 #include "storage/bufmgr.h"
  50 #include "storage/lmgr.h"
  51 #include "utils/builtins.h"
  52 #include "utils/memutils.h"
  53 #include "utils/rel.h"
  54 #include "utils/tqual.h"
  55
  56
  57 static bool ExecOnConflictUpdate(ModifyTableState *mtstate,
  58                                          ResultRelInfo *resultRelInfo,
  59                                          ItemPointer conflictTid,
  60                                          TupleTableSlot *planSlot,
  61                                          TupleTableSlot *excludedSlot,
  62                                          EState *estate,
  63                                          bool canSetTag,
  64                                          TupleTableSlot **returning);
  65
  66 /*
  67  * Verify that the tuples to be produced by INSERT or UPDATE match the
  68  * target relation's rowtype
  69  *
  70  * We do this to guard against stale plans.  If plan invalidation is
  71  * functioning properly then we should never get a failure here, but better
  72  * safe than sorry.  Note that this is called after we have obtained lock
  73  * on the target rel, so the rowtype can't change underneath us.
  74  *
  75  * The plan output is represented by its targetlist, because that makes
  76  * handling the dropped-column case easier.
  77  */
  78 static void
  79 ExecCheckPlanOutput(Relation resultRel, List *targetList)
  80 {
  81         TupleDesc       resultDesc = RelationGetDescr(resultRel);
  82         int                     attno = 0;
  83         ListCell   *lc;
  84
  85         foreach(lc, targetList)
  86         {
  87                 TargetEntry *tle = (TargetEntry *) lfirst(lc);
  88                 Form_pg_attribute attr;
  89
  90                 if (tle->resjunk)
  91                         continue;                       /* ignore junk tlist items */
  92
  93                 if (attno >= resultDesc->natts)
  94                         ereport(ERROR,
  95                                         (errcode(ERRCODE_DATATYPE_MISMATCH),
  96                                          errmsg("table row type and query-specified row type do not match"),
  97                                          errdetail("Query has too many columns.")));
  98                 attr = resultDesc->attrs[attno++];
  99
 100                 if (!attr->attisdropped)
 101                 {
 102                         /* Normal case: demand type match */
 103                         if (exprType((Node *) tle->expr) != attr->atttypid)
 104                                 ereport(ERROR,
 105                                                 (errcode(ERRCODE_DATATYPE_MISMATCH),
 106                                                  errmsg("table row type and query-specified row type do not match"),
 107                                                  errdetail("Table has type %s at ordinal position %d, but query expects %s.",
 108                                                                    format_type_be(attr->atttypid),
 109                                                                    attno,
 110                                                          format_type_be(exprType((Node *) tle->expr)))));
 111                 }
 112                 else
 113                 {
 114                         /*
 115                          * For a dropped column, we can't check atttypid (it's likely 0).
 116                          * In any case the planner has most likely inserted an INT4 null.
 117                          * What we insist on is just *some* NULL constant.
 118                          */
 119                         if (!IsA(tle->expr, Const) ||
 120                                 !((Const *) tle->expr)->constisnull)
 121                                 ereport(ERROR,
 122                                                 (errcode(ERRCODE_DATATYPE_MISMATCH),
 123                                                  errmsg("table row type and query-specified row type do not match"),
 124                                                  errdetail("Query provides a value for a dropped column at ordinal position %d.",
 125                                                                    attno)));
 126                 }
 127         }
 128         if (attno != resultDesc->natts)
 129                 ereport(ERROR,
 130                                 (errcode(ERRCODE_DATATYPE_MISMATCH),
 131                   errmsg("table row type and query-specified row type do not match"),
 132                                  errdetail("Query has too few columns.")));
 133 }
 134
 135 /*
 136  * ExecProcessReturning --- evaluate a RETURNING list
 137  *
 138  * projectReturning: RETURNING projection info for current result rel
 139  * tupleSlot: slot holding tuple actually inserted/updated/deleted
 140  * planSlot: slot holding tuple returned by top subplan node
 141  *
 142  * Note: If tupleSlot is NULL, the FDW should have already provided econtext's
 143  * scan tuple.
 144  *
 145  * Returns a slot holding the result tuple
 146  */
 147 static TupleTableSlot *
 148 ExecProcessReturning(ResultRelInfo *resultRelInfo,
 149                                          TupleTableSlot *tupleSlot,
 150                                          TupleTableSlot *planSlot)
 151 {
 152         ProjectionInfo *projectReturning = resultRelInfo->ri_projectReturning;
 153         ExprContext *econtext = projectReturning->pi_exprContext;
 154
 155         /*
 156          * Reset per-tuple memory context to free any expression evaluation
 157          * storage allocated in the previous cycle.
 158          */
 159         ResetExprContext(econtext);
 160
 161         /* Make tuple and any needed join variables available to ExecProject */
 162         if (tupleSlot)
 163                 econtext->ecxt_scantuple = tupleSlot;
 164         else
 165         {
 166                 HeapTuple       tuple;
 167
 168                 /*
 169                  * RETURNING expressions might reference the tableoid column, so
 170                  * initialize t_tableOid before evaluating them.
 171                  */
 172                 Assert(!TupIsNull(econtext->ecxt_scantuple));
 173                 tuple = ExecMaterializeSlot(econtext->ecxt_scantuple);
 174                 tuple->t_tableOid = RelationGetRelid(resultRelInfo->ri_RelationDesc);
 175         }
 176         econtext->ecxt_outertuple = planSlot;
 177
 178         /* Compute the RETURNING expressions */
 179         return ExecProject(projectReturning);
 180 }
 181
 182 /*
 183  * ExecCheckHeapTupleVisible -- verify heap tuple is visible
 184  *
 185  * It would not be consistent with guarantees of the higher isolation levels to
 186  * proceed with avoiding insertion (taking speculative insertion's alternative
 187  * path) on the basis of another tuple that is not visible to MVCC snapshot.
 188  * Check for the need to raise a serialization failure, and do so as necessary.
 189  */
 190 static void
 191 ExecCheckHeapTupleVisible(EState *estate,
 192                                                   HeapTuple tuple,
 193                                                   Buffer buffer)
 194 {
 195         if (!IsolationUsesXactSnapshot())
 196                 return;
 197
 198         /*
 199          * We need buffer pin and lock to call HeapTupleSatisfiesVisibility.
 200          * Caller should be holding pin, but not lock.
 201          */
 202         LockBuffer(buffer, BUFFER_LOCK_SHARE);
 203         if (!HeapTupleSatisfiesVisibility(tuple, estate->es_snapshot, buffer))
 204         {
 205                 /*
 206                  * We should not raise a serialization failure if the conflict is
 207                  * against a tuple inserted by our own transaction, even if it's not
 208                  * visible to our snapshot.  (This would happen, for example, if
 209                  * conflicting keys are proposed for insertion in a single command.)
 210                  */
 211                 if (!TransactionIdIsCurrentTransactionId(HeapTupleHeaderGetXmin(tuple->t_data)))
 212                         ereport(ERROR,
 213                                         (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
 214                          errmsg("could not serialize access due to concurrent update")));
 215         }
 216         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
 217 }
 218
 219 /*
 220  * ExecCheckTIDVisible -- convenience variant of ExecCheckHeapTupleVisible()
 221  */
 222 static void
 223 ExecCheckTIDVisible(EState *estate,
 224                                         ResultRelInfo *relinfo,
 225                                         ItemPointer tid)
 226 {
 227         Relation        rel = relinfo->ri_RelationDesc;
 228         Buffer          buffer;
 229         HeapTupleData tuple;
 230
 231         /* Redundantly check isolation level */
 232         if (!IsolationUsesXactSnapshot())
 233                 return;
 234
 235         tuple.t_self = *tid;
 236         if (!heap_fetch(rel, SnapshotAny, &tuple, &buffer, false, NULL))
 237                 elog(ERROR, "failed to fetch conflicting tuple for ON CONFLICT");
 238         ExecCheckHeapTupleVisible(estate, &tuple, buffer);
 239         ReleaseBuffer(buffer);
 240 }
 241
 242 /* ----------------------------------------------------------------
 243  *              ExecInsert
 244  *
 245  *              For INSERT, we have to insert the tuple into the target relation
 246  *              and insert appropriate tuples into the index relations.
 247  *
 248  *              Returns RETURNING result if any, otherwise NULL.
 249  * ----------------------------------------------------------------
 250  */
 251 static TupleTableSlot *
 252 ExecInsert(ModifyTableState *mtstate,
 253                    TupleTableSlot *slot,
 254                    TupleTableSlot *planSlot,
 255                    List *arbiterIndexes,
 256                    OnConflictAction onconflict,
 257                    EState *estate,
 258                    bool canSetTag)
 259 {
 260         HeapTuple       tuple;
 261         ResultRelInfo *resultRelInfo;
 262         ResultRelInfo *saved_resultRelInfo = NULL;
 263         Relation        resultRelationDesc;
 264         Oid                     newId;
 265         List       *recheckIndexes = NIL;
 266         TupleTableSlot *oldslot = slot,
 267                            *result = NULL;
 268
 269         /*
 270          * get the heap tuple out of the tuple table slot, making sure we have a
 271          * writable copy
 272          */
 273         tuple = ExecMaterializeSlot(slot);
 274
 275         /*
 276          * get information on the (current) result relation
 277          */
 278         resultRelInfo = estate->es_result_relation_info;
 279
 280         /* Determine the partition to heap_insert the tuple into */
 281         if (mtstate->mt_partition_dispatch_info)
 282         {
 283                 int                     leaf_part_index;
 284                 TupleConversionMap *map;
 285
 286                 /*
 287                  * Away we go ... If we end up not finding a partition after all,
 288                  * ExecFindPartition() does not return and errors out instead.
 289                  * Otherwise, the returned value is to be used as an index into arrays
 290                  * mt_partitions[] and mt_partition_tupconv_maps[] that will get us
 291                  * the ResultRelInfo and TupleConversionMap for the partition,
 292                  * respectively.
 293                  */
 294                 leaf_part_index = ExecFindPartition(resultRelInfo,
 295                                                                                  mtstate->mt_partition_dispatch_info,
 296                                                                                         slot,
 297                                                                                         estate);
 298                 Assert(leaf_part_index >= 0 &&
 299                            leaf_part_index < mtstate->mt_num_partitions);
 300
 301                 /*
 302                  * Save the old ResultRelInfo and switch to the one corresponding to
 303                  * the selected partition.
 304                  */
 305                 saved_resultRelInfo = resultRelInfo;
 306                 resultRelInfo = mtstate->mt_partitions + leaf_part_index;
 307
 308                 /* We do not yet have a way to insert into a foreign partition */
 309                 if (resultRelInfo->ri_FdwRoutine)
 310                         ereport(ERROR,
 311                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 312                                  errmsg("cannot route inserted tuples to a foreign table")));
 313
 314                 /* For ExecInsertIndexTuples() to work on the partition's indexes */
 315                 estate->es_result_relation_info = resultRelInfo;
 316
 317                 /*
 318                  * We might need to convert from the parent rowtype to the partition
 319                  * rowtype.
 320                  */
 321                 map = mtstate->mt_partition_tupconv_maps[leaf_part_index];
 322                 if (map)
 323                 {
 324                         Relation        partrel = resultRelInfo->ri_RelationDesc;
 325
 326                         tuple = do_convert_tuple(tuple, map);
 327
 328                         /*
 329                          * We must use the partition's tuple descriptor from this point
 330                          * on, until we're finished dealing with the partition. Use the
 331                          * dedicated slot for that.
 332                          */
 333                         slot = mtstate->mt_partition_tuple_slot;
 334                         Assert(slot != NULL);
 335                         ExecSetSlotDescriptor(slot, RelationGetDescr(partrel));
 336                         ExecStoreTuple(tuple, slot, InvalidBuffer, true);
 337                 }
 338         }
 339
 340         resultRelationDesc = resultRelInfo->ri_RelationDesc;
 341
 342         /*
 343          * If the result relation has OIDs, force the tuple's OID to zero so that
 344          * heap_insert will assign a fresh OID.  Usually the OID already will be
 345          * zero at this point, but there are corner cases where the plan tree can
 346          * return a tuple extracted literally from some table with the same
 347          * rowtype.
 348          *
 349          * XXX if we ever wanted to allow users to assign their own OIDs to new
 350          * rows, this'd be the place to do it.  For the moment, we make a point of
 351          * doing this before calling triggers, so that a user-supplied trigger
 352          * could hack the OID if desired.
 353          */
 354         if (resultRelationDesc->rd_rel->relhasoids)
 355                 HeapTupleSetOid(tuple, InvalidOid);
 356
 357         /*
 358          * BEFORE ROW INSERT Triggers.
 359          *
 360          * Note: We fire BEFORE ROW TRIGGERS for every attempted insertion in an
 361          * INSERT ... ON CONFLICT statement.  We cannot check for constraint
 362          * violations before firing these triggers, because they can change the
 363          * values to insert.  Also, they can run arbitrary user-defined code with
 364          * side-effects that we can't cancel by just not inserting the tuple.
 365          */
 366         if (resultRelInfo->ri_TrigDesc &&
 367                 resultRelInfo->ri_TrigDesc->trig_insert_before_row)
 368         {
 369                 slot = ExecBRInsertTriggers(estate, resultRelInfo, slot);
 370
 371                 if (slot == NULL)               /* "do nothing" */
 372                         return NULL;
 373
 374                 /* trigger might have changed tuple */
 375                 tuple = ExecMaterializeSlot(slot);
 376         }
 377
 378         /* INSTEAD OF ROW INSERT Triggers */
 379         if (resultRelInfo->ri_TrigDesc &&
 380                 resultRelInfo->ri_TrigDesc->trig_insert_instead_row)
 381         {
 382                 slot = ExecIRInsertTriggers(estate, resultRelInfo, slot);
 383
 384                 if (slot == NULL)               /* "do nothing" */
 385                         return NULL;
 386
 387                 /* trigger might have changed tuple */
 388                 tuple = ExecMaterializeSlot(slot);
 389
 390                 newId = InvalidOid;
 391         }
 392         else if (resultRelInfo->ri_FdwRoutine)
 393         {
 394                 /*
 395                  * insert into foreign table: let the FDW do it
 396                  */
 397                 slot = resultRelInfo->ri_FdwRoutine->ExecForeignInsert(estate,
 398                                                                                                                            resultRelInfo,
 399                                                                                                                            slot,
 400                                                                                                                            planSlot);
 401
 402                 if (slot == NULL)               /* "do nothing" */
 403                         return NULL;
 404
 405                 /* FDW might have changed tuple */
 406                 tuple = ExecMaterializeSlot(slot);
 407
 408                 /*
 409                  * AFTER ROW Triggers or RETURNING expressions might reference the
 410                  * tableoid column, so initialize t_tableOid before evaluating them.
 411                  */
 412                 tuple->t_tableOid = RelationGetRelid(resultRelationDesc);
 413
 414                 newId = InvalidOid;
 415         }
 416         else
 417         {
 418                 /*
 419                  * Constraints might reference the tableoid column, so initialize
 420                  * t_tableOid before evaluating them.
 421                  */
 422                 tuple->t_tableOid = RelationGetRelid(resultRelationDesc);
 423
 424                 /*
 425                  * Check any RLS INSERT WITH CHECK policies
 426                  *
 427                  * ExecWithCheckOptions() will skip any WCOs which are not of the kind
 428                  * we are looking for at this point.
 429                  */
 430                 if (resultRelInfo->ri_WithCheckOptions != NIL)
 431                         ExecWithCheckOptions(WCO_RLS_INSERT_CHECK,
 432                                                                  resultRelInfo, slot, estate);
 433
 434                 /*
 435                  * Check the constraints of the tuple
 436                  */
 437                 if (resultRelationDesc->rd_att->constr || resultRelInfo->ri_PartitionCheck)
 438                         ExecConstraints(resultRelInfo, slot, oldslot, estate);
 439
 440                 if (onconflict != ONCONFLICT_NONE && resultRelInfo->ri_NumIndices > 0)
 441                 {
 442                         /* Perform a speculative insertion. */
 443                         uint32          specToken;
 444                         ItemPointerData conflictTid;
 445                         bool            specConflict;
 446
 447                         /*
 448                          * Do a non-conclusive check for conflicts first.
 449                          *
 450                          * We're not holding any locks yet, so this doesn't guarantee that
 451                          * the later insert won't conflict.  But it avoids leaving behind
 452                          * a lot of canceled speculative insertions, if you run a lot of
 453                          * INSERT ON CONFLICT statements that do conflict.
 454                          *
 455                          * We loop back here if we find a conflict below, either during
 456                          * the pre-check, or when we re-check after inserting the tuple
 457                          * speculatively.
 458                          */
 459         vlock:
 460                         specConflict = false;
 461                         if (!ExecCheckIndexConstraints(slot, estate, &conflictTid,
 462                                                                                    arbiterIndexes))
 463                         {
 464                                 /* committed conflict tuple found */
 465                                 if (onconflict == ONCONFLICT_UPDATE)
 466                                 {
 467                                         /*
 468                                          * In case of ON CONFLICT DO UPDATE, execute the UPDATE
 469                                          * part.  Be prepared to retry if the UPDATE fails because
 470                                          * of another concurrent UPDATE/DELETE to the conflict
 471                                          * tuple.
 472                                          */
 473                                         TupleTableSlot *returning = NULL;
 474
 475                                         if (ExecOnConflictUpdate(mtstate, resultRelInfo,
 476                                                                                          &conflictTid, planSlot, slot,
 477                                                                                          estate, canSetTag, &returning))
 478                                         {
 479                                                 InstrCountFiltered2(&mtstate->ps, 1);
 480                                                 return returning;
 481                                         }
 482                                         else
 483                                                 goto vlock;
 484                                 }
 485                                 else
 486                                 {
 487                                         /*
 488                                          * In case of ON CONFLICT DO NOTHING, do nothing. However,
 489                                          * verify that the tuple is visible to the executor's MVCC
 490                                          * snapshot at higher isolation levels.
 491                                          */
 492                                         Assert(onconflict == ONCONFLICT_NOTHING);
 493                                         ExecCheckTIDVisible(estate, resultRelInfo, &conflictTid);
 494                                         InstrCountFiltered2(&mtstate->ps, 1);
 495                                         return NULL;
 496                                 }
 497                         }
 498
 499                         /*
 500                          * Before we start insertion proper, acquire our "speculative
 501                          * insertion lock".  Others can use that to wait for us to decide
 502                          * if we're going to go ahead with the insertion, instead of
 503                          * waiting for the whole transaction to complete.
 504                          */
 505                         specToken = SpeculativeInsertionLockAcquire(GetCurrentTransactionId());
 506                         HeapTupleHeaderSetSpeculativeToken(tuple->t_data, specToken);
 507
 508                         /* insert the tuple, with the speculative token */
 509                         newId = heap_insert(resultRelationDesc, tuple,
 510                                                                 estate->es_output_cid,
 511                                                                 HEAP_INSERT_SPECULATIVE,
 512                                                                 NULL);
 513
 514                         /* insert index entries for tuple */
 515                         recheckIndexes = ExecInsertIndexTuples(slot, &(tuple->t_self),
 516                                                                                                  estate, true, &specConflict,
 517                                                                                                    arbiterIndexes);
 518
 519                         /* adjust the tuple's state accordingly */
 520                         if (!specConflict)
 521                                 heap_finish_speculative(resultRelationDesc, tuple);
 522                         else
 523                                 heap_abort_speculative(resultRelationDesc, tuple);
 524
 525                         /*
 526                          * Wake up anyone waiting for our decision.  They will re-check
 527                          * the tuple, see that it's no longer speculative, and wait on our
 528                          * XID as if this was a regularly inserted tuple all along.  Or if
 529                          * we killed the tuple, they will see it's dead, and proceed as if
 530                          * the tuple never existed.
 531                          */
 532                         SpeculativeInsertionLockRelease(GetCurrentTransactionId());
 533
 534                         /*
 535                          * If there was a conflict, start from the beginning.  We'll do
 536                          * the pre-check again, which will now find the conflicting tuple
 537                          * (unless it aborts before we get there).
 538                          */
 539                         if (specConflict)
 540                         {
 541                                 list_free(recheckIndexes);
 542                                 goto vlock;
 543                         }
 544
 545                         /* Since there was no insertion conflict, we're done */
 546                 }
 547                 else
 548                 {
 549                         /*
 550                          * insert the tuple normally.
 551                          *
 552                          * Note: heap_insert returns the tid (location) of the new tuple
 553                          * in the t_self field.
 554                          */
 555                         newId = heap_insert(resultRelationDesc, tuple,
 556                                                                 estate->es_output_cid,
 557                                                                 0, NULL);
 558
 559                         /* insert index entries for tuple */
 560                         if (resultRelInfo->ri_NumIndices > 0)
 561                                 recheckIndexes = ExecInsertIndexTuples(slot, &(tuple->t_self),
 562                                                                                                            estate, false, NULL,
 563                                                                                                            arbiterIndexes);
 564                 }
 565         }
 566
 567         if (canSetTag)
 568         {
 569                 (estate->es_processed)++;
 570                 estate->es_lastoid = newId;
 571                 setLastTid(&(tuple->t_self));
 572         }
 573
 574         /* AFTER ROW INSERT Triggers */
 575         ExecARInsertTriggers(estate, resultRelInfo, tuple, recheckIndexes);
 576
 577         list_free(recheckIndexes);
 578
 579         /*
 580          * Check any WITH CHECK OPTION constraints from parent views.  We are
 581          * required to do this after testing all constraints and uniqueness
 582          * violations per the SQL spec, so we do it after actually inserting the
 583          * record into the heap and all indexes.
 584          *
 585          * ExecWithCheckOptions will elog(ERROR) if a violation is found, so the
 586          * tuple will never be seen, if it violates the WITH CHECK OPTION.
 587          *
 588          * ExecWithCheckOptions() will skip any WCOs which are not of the kind we
 589          * are looking for at this point.
 590          */
 591         if (resultRelInfo->ri_WithCheckOptions != NIL)
 592                 ExecWithCheckOptions(WCO_VIEW_CHECK, resultRelInfo, slot, estate);
 593
 594         /* Process RETURNING if present */
 595         if (resultRelInfo->ri_projectReturning)
 596                 result = ExecProcessReturning(resultRelInfo, slot, planSlot);
 597
 598         if (saved_resultRelInfo)
 599                 estate->es_result_relation_info = saved_resultRelInfo;
 600
 601         return result;
 602 }
 603
 604 /* ----------------------------------------------------------------
 605  *              ExecDelete
 606  *
 607  *              DELETE is like UPDATE, except that we delete the tuple and no
 608  *              index modifications are needed.
 609  *
 610  *              When deleting from a table, tupleid identifies the tuple to
 611  *              delete and oldtuple is NULL.  When deleting from a view,
 612  *              oldtuple is passed to the INSTEAD OF triggers and identifies
 613  *              what to delete, and tupleid is invalid.  When deleting from a
 614  *              foreign table, tupleid is invalid; the FDW has to figure out
 615  *              which row to delete using data from the planSlot.  oldtuple is
 616  *              passed to foreign table triggers; it is NULL when the foreign
 617  *              table has no relevant triggers.
 618  *
 619  *              Returns RETURNING result if any, otherwise NULL.
 620  * ----------------------------------------------------------------
 621  */
 622 static TupleTableSlot *
 623 ExecDelete(ItemPointer tupleid,
 624                    HeapTuple oldtuple,
 625                    TupleTableSlot *planSlot,
 626                    EPQState *epqstate,
 627                    EState *estate,
 628                    bool canSetTag)
 629 {
 630         ResultRelInfo *resultRelInfo;
 631         Relation        resultRelationDesc;
 632         HTSU_Result result;
 633         HeapUpdateFailureData hufd;
 634         TupleTableSlot *slot = NULL;
 635
 636         /*
 637          * get information on the (current) result relation
 638          */
 639         resultRelInfo = estate->es_result_relation_info;
 640         resultRelationDesc = resultRelInfo->ri_RelationDesc;
 641
 642         /* BEFORE ROW DELETE Triggers */
 643         if (resultRelInfo->ri_TrigDesc &&
 644                 resultRelInfo->ri_TrigDesc->trig_delete_before_row)
 645         {
 646                 bool            dodelete;
 647
 648                 dodelete = ExecBRDeleteTriggers(estate, epqstate, resultRelInfo,
 649                                                                                 tupleid, oldtuple);
 650
 651                 if (!dodelete)                  /* "do nothing" */
 652                         return NULL;
 653         }
 654
 655         /* INSTEAD OF ROW DELETE Triggers */
 656         if (resultRelInfo->ri_TrigDesc &&
 657                 resultRelInfo->ri_TrigDesc->trig_delete_instead_row)
 658         {
 659                 bool            dodelete;
 660
 661                 Assert(oldtuple != NULL);
 662                 dodelete = ExecIRDeleteTriggers(estate, resultRelInfo, oldtuple);
 663
 664                 if (!dodelete)                  /* "do nothing" */
 665                         return NULL;
 666         }
 667         else if (resultRelInfo->ri_FdwRoutine)
 668         {
 669                 HeapTuple       tuple;
 670
 671                 /*
 672                  * delete from foreign table: let the FDW do it
 673                  *
 674                  * We offer the trigger tuple slot as a place to store RETURNING data,
 675                  * although the FDW can return some other slot if it wants.  Set up
 676                  * the slot's tupdesc so the FDW doesn't need to do that for itself.
 677                  */
 678                 slot = estate->es_trig_tuple_slot;
 679                 if (slot->tts_tupleDescriptor != RelationGetDescr(resultRelationDesc))
 680                         ExecSetSlotDescriptor(slot, RelationGetDescr(resultRelationDesc));
 681
 682                 slot = resultRelInfo->ri_FdwRoutine->ExecForeignDelete(estate,
 683                                                                                                                            resultRelInfo,
 684                                                                                                                            slot,
 685                                                                                                                            planSlot);
 686
 687                 if (slot == NULL)               /* "do nothing" */
 688                         return NULL;
 689
 690                 /*
 691                  * RETURNING expressions might reference the tableoid column, so
 692                  * initialize t_tableOid before evaluating them.
 693                  */
 694                 if (slot->tts_isempty)
 695                         ExecStoreAllNullTuple(slot);
 696                 tuple = ExecMaterializeSlot(slot);
 697                 tuple->t_tableOid = RelationGetRelid(resultRelationDesc);
 698         }
 699         else
 700         {
 701                 /*
 702                  * delete the tuple
 703                  *
 704                  * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check
 705                  * that the row to be deleted is visible to that snapshot, and throw a
 706                  * can't-serialize error if not. This is a special-case behavior
 707                  * needed for referential integrity updates in transaction-snapshot
 708                  * mode transactions.
 709                  */
 710 ldelete:;
 711                 result = heap_delete(resultRelationDesc, tupleid,
 712                                                          estate->es_output_cid,
 713                                                          estate->es_crosscheck_snapshot,
 714                                                          true /* wait for commit */ ,
 715                                                          &hufd);
 716                 switch (result)
 717                 {
 718                         case HeapTupleSelfUpdated:
 719
 720                                 /*
 721                                  * The target tuple was already updated or deleted by the
 722                                  * current command, or by a later command in the current
 723                                  * transaction.  The former case is possible in a join DELETE
 724                                  * where multiple tuples join to the same target tuple. This
 725                                  * is somewhat questionable, but Postgres has always allowed
 726                                  * it: we just ignore additional deletion attempts.
 727                                  *
 728                                  * The latter case arises if the tuple is modified by a
 729                                  * command in a BEFORE trigger, or perhaps by a command in a
 730                                  * volatile function used in the query.  In such situations we
 731                                  * should not ignore the deletion, but it is equally unsafe to
 732                                  * proceed.  We don't want to discard the original DELETE
 733                                  * while keeping the triggered actions based on its deletion;
 734                                  * and it would be no better to allow the original DELETE
 735                                  * while discarding updates that it triggered.  The row update
 736                                  * carries some information that might be important according
 737                                  * to business rules; so throwing an error is the only safe
 738                                  * course.
 739                                  *
 740                                  * If a trigger actually intends this type of interaction, it
 741                                  * can re-execute the DELETE and then return NULL to cancel
 742                                  * the outer delete.
 743                                  */
 744                                 if (hufd.cmax != estate->es_output_cid)
 745                                         ereport(ERROR,
 746                                                         (errcode(ERRCODE_TRIGGERED_DATA_CHANGE_VIOLATION),
 747                                                          errmsg("tuple to be updated was already modified by an operation triggered by the current command"),
 748                                                          errhint("Consider using an AFTER trigger instead of a BEFORE trigger to propagate changes to other rows.")));
 749
 750                                 /* Else, already deleted by self; nothing to do */
 751                                 return NULL;
 752
 753                         case HeapTupleMayBeUpdated:
 754                                 break;
 755
 756                         case HeapTupleUpdated:
 757                                 if (IsolationUsesXactSnapshot())
 758                                         ereport(ERROR,
 759                                                         (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
 760                                                          errmsg("could not serialize access due to concurrent update")));
 761                                 if (!ItemPointerEquals(tupleid, &hufd.ctid))
 762                                 {
 763                                         TupleTableSlot *epqslot;
 764
 765                                         epqslot = EvalPlanQual(estate,
 766                                                                                    epqstate,
 767                                                                                    resultRelationDesc,
 768                                                                                    resultRelInfo->ri_RangeTableIndex,
 769                                                                                    LockTupleExclusive,
 770                                                                                    &hufd.ctid,
 771                                                                                    hufd.xmax);
 772                                         if (!TupIsNull(epqslot))
 773                                         {
 774                                                 *tupleid = hufd.ctid;
 775                                                 goto ldelete;
 776                                         }
 777                                 }
 778                                 /* tuple already deleted; nothing to do */
 779                                 return NULL;
 780
 781                         default:
 782                                 elog(ERROR, "unrecognized heap_delete status: %u", result);
 783                                 return NULL;
 784                 }
 785
 786                 /*
 787                  * Note: Normally one would think that we have to delete index tuples
 788                  * associated with the heap tuple now...
 789                  *
 790                  * ... but in POSTGRES, we have no need to do this because VACUUM will
 791                  * take care of it later.  We can't delete index tuples immediately
 792                  * anyway, since the tuple is still visible to other transactions.
 793                  */
 794         }
 795
 796         if (canSetTag)
 797                 (estate->es_processed)++;
 798
 799         /* AFTER ROW DELETE Triggers */
 800         ExecARDeleteTriggers(estate, resultRelInfo, tupleid, oldtuple);
 801
 802         /* Process RETURNING if present */
 803         if (resultRelInfo->ri_projectReturning)
 804         {
 805                 /*
 806                  * We have to put the target tuple into a slot, which means first we
 807                  * gotta fetch it.  We can use the trigger tuple slot.
 808                  */
 809                 TupleTableSlot *rslot;
 810                 HeapTupleData deltuple;
 811                 Buffer          delbuffer;
 812
 813                 if (resultRelInfo->ri_FdwRoutine)
 814                 {
 815                         /* FDW must have provided a slot containing the deleted row */
 816                         Assert(!TupIsNull(slot));
 817                         delbuffer = InvalidBuffer;
 818                 }
 819                 else
 820                 {
 821                         slot = estate->es_trig_tuple_slot;
 822                         if (oldtuple != NULL)
 823                         {
 824                                 deltuple = *oldtuple;
 825                                 delbuffer = InvalidBuffer;
 826                         }
 827                         else
 828                         {
 829                                 deltuple.t_self = *tupleid;
 830                                 if (!heap_fetch(resultRelationDesc, SnapshotAny,
 831                                                                 &deltuple, &delbuffer, false, NULL))
 832                                         elog(ERROR, "failed to fetch deleted tuple for DELETE RETURNING");
 833                         }
 834
 835                         if (slot->tts_tupleDescriptor != RelationGetDescr(resultRelationDesc))
 836                                 ExecSetSlotDescriptor(slot, RelationGetDescr(resultRelationDesc));
 837                         ExecStoreTuple(&deltuple, slot, InvalidBuffer, false);
 838                 }
 839
 840                 rslot = ExecProcessReturning(resultRelInfo, slot, planSlot);
 841
 842                 /*
 843                  * Before releasing the target tuple again, make sure rslot has a
 844                  * local copy of any pass-by-reference values.
 845                  */
 846                 ExecMaterializeSlot(rslot);
 847
 848                 ExecClearTuple(slot);
 849                 if (BufferIsValid(delbuffer))
 850                         ReleaseBuffer(delbuffer);
 851
 852                 return rslot;
 853         }
 854
 855         return NULL;
 856 }
 857
 858 /* ----------------------------------------------------------------
 859  *              ExecUpdate
 860  *
 861  *              note: we can't run UPDATE queries with transactions
 862  *              off because UPDATEs are actually INSERTs and our
 863  *              scan will mistakenly loop forever, updating the tuple
 864  *              it just inserted..  This should be fixed but until it
 865  *              is, we don't want to get stuck in an infinite loop
 866  *              which corrupts your database..
 867  *
 868  *              When updating a table, tupleid identifies the tuple to
 869  *              update and oldtuple is NULL.  When updating a view, oldtuple
 870  *              is passed to the INSTEAD OF triggers and identifies what to
 871  *              update, and tupleid is invalid.  When updating a foreign table,
 872  *              tupleid is invalid; the FDW has to figure out which row to
 873  *              update using data from the planSlot.  oldtuple is passed to
 874  *              foreign table triggers; it is NULL when the foreign table has
 875  *              no relevant triggers.
 876  *
 877  *              Returns RETURNING result if any, otherwise NULL.
 878  * ----------------------------------------------------------------
 879  */
 880 static TupleTableSlot *
 881 ExecUpdate(ItemPointer tupleid,
 882                    HeapTuple oldtuple,
 883                    TupleTableSlot *slot,
 884                    TupleTableSlot *planSlot,
 885                    EPQState *epqstate,
 886                    EState *estate,
 887                    bool canSetTag)
 888 {
 889         HeapTuple       tuple;
 890         ResultRelInfo *resultRelInfo;
 891         Relation        resultRelationDesc;
 892         HTSU_Result result;
 893         HeapUpdateFailureData hufd;
 894         List       *recheckIndexes = NIL;
 895
 896         /*
 897          * abort the operation if not running transactions
 898          */
 899         if (IsBootstrapProcessingMode())
 900                 elog(ERROR, "cannot UPDATE during bootstrap");
 901
 902         /*
 903          * get the heap tuple out of the tuple table slot, making sure we have a
 904          * writable copy
 905          */
 906         tuple = ExecMaterializeSlot(slot);
 907
 908         /*
 909          * get information on the (current) result relation
 910          */
 911         resultRelInfo = estate->es_result_relation_info;
 912         resultRelationDesc = resultRelInfo->ri_RelationDesc;
 913
 914         /* BEFORE ROW UPDATE Triggers */
 915         if (resultRelInfo->ri_TrigDesc &&
 916                 resultRelInfo->ri_TrigDesc->trig_update_before_row)
 917         {
 918                 slot = ExecBRUpdateTriggers(estate, epqstate, resultRelInfo,
 919                                                                         tupleid, oldtuple, slot);
 920
 921                 if (slot == NULL)               /* "do nothing" */
 922                         return NULL;
 923
 924                 /* trigger might have changed tuple */
 925                 tuple = ExecMaterializeSlot(slot);
 926         }
 927
 928         /* INSTEAD OF ROW UPDATE Triggers */
 929         if (resultRelInfo->ri_TrigDesc &&
 930                 resultRelInfo->ri_TrigDesc->trig_update_instead_row)
 931         {
 932                 slot = ExecIRUpdateTriggers(estate, resultRelInfo,
 933                                                                         oldtuple, slot);
 934
 935                 if (slot == NULL)               /* "do nothing" */
 936                         return NULL;
 937
 938                 /* trigger might have changed tuple */
 939                 tuple = ExecMaterializeSlot(slot);
 940         }
 941         else if (resultRelInfo->ri_FdwRoutine)
 942         {
 943                 /*
 944                  * update in foreign table: let the FDW do it
 945                  */
 946                 slot = resultRelInfo->ri_FdwRoutine->ExecForeignUpdate(estate,
 947                                                                                                                            resultRelInfo,
 948                                                                                                                            slot,
 949                                                                                                                            planSlot);
 950
 951                 if (slot == NULL)               /* "do nothing" */
 952                         return NULL;
 953
 954                 /* FDW might have changed tuple */
 955                 tuple = ExecMaterializeSlot(slot);
 956
 957                 /*
 958                  * AFTER ROW Triggers or RETURNING expressions might reference the
 959                  * tableoid column, so initialize t_tableOid before evaluating them.
 960                  */
 961                 tuple->t_tableOid = RelationGetRelid(resultRelationDesc);
 962         }
 963         else
 964         {
 965                 LockTupleMode lockmode;
 966
 967                 /*
 968                  * Constraints might reference the tableoid column, so initialize
 969                  * t_tableOid before evaluating them.
 970                  */
 971                 tuple->t_tableOid = RelationGetRelid(resultRelationDesc);
 972
 973                 /*
 974                  * Check any RLS UPDATE WITH CHECK policies
 975                  *
 976                  * If we generate a new candidate tuple after EvalPlanQual testing, we
 977                  * must loop back here and recheck any RLS policies and constraints.
 978                  * (We don't need to redo triggers, however.  If there are any BEFORE
 979                  * triggers then trigger.c will have done heap_lock_tuple to lock the
 980                  * correct tuple, so there's no need to do them again.)
 981                  *
 982                  * ExecWithCheckOptions() will skip any WCOs which are not of the kind
 983                  * we are looking for at this point.
 984                  */
 985 lreplace:;
 986                 if (resultRelInfo->ri_WithCheckOptions != NIL)
 987                         ExecWithCheckOptions(WCO_RLS_UPDATE_CHECK,
 988                                                                  resultRelInfo, slot, estate);
 989
 990                 /*
 991                  * Check the constraints of the tuple.  Note that we pass the same
 992                  * slot for the orig_slot argument, because unlike ExecInsert(), no
 993                  * tuple-routing is performed here, hence the slot remains unchanged.
 994                  */
 995                 if (resultRelationDesc->rd_att->constr || resultRelInfo->ri_PartitionCheck)
 996                         ExecConstraints(resultRelInfo, slot, slot, estate);
 997
 998                 /*
 999                  * replace the heap tuple
1000                  *
1001                  * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check
1002                  * that the row to be updated is visible to that snapshot, and throw a
1003                  * can't-serialize error if not. This is a special-case behavior
1004                  * needed for referential integrity updates in transaction-snapshot
1005                  * mode transactions.
1006                  */
1007                 result = heap_update(resultRelationDesc, tupleid, tuple,
1008                                                          estate->es_output_cid,
1009                                                          estate->es_crosscheck_snapshot,
1010                                                          true /* wait for commit */ ,
1011                                                          &hufd, &lockmode);
1012                 switch (result)
1013                 {
1014                         case HeapTupleSelfUpdated:
1015
1016                                 /*
1017                                  * The target tuple was already updated or deleted by the
1018                                  * current command, or by a later command in the current
1019                                  * transaction.  The former case is possible in a join UPDATE
1020                                  * where multiple tuples join to the same target tuple. This
1021                                  * is pretty questionable, but Postgres has always allowed it:
1022                                  * we just execute the first update action and ignore
1023                                  * additional update attempts.
1024                                  *
1025                                  * The latter case arises if the tuple is modified by a
1026                                  * command in a BEFORE trigger, or perhaps by a command in a
1027                                  * volatile function used in the query.  In such situations we
1028                                  * should not ignore the update, but it is equally unsafe to
1029                                  * proceed.  We don't want to discard the original UPDATE
1030                                  * while keeping the triggered actions based on it; and we
1031                                  * have no principled way to merge this update with the
1032                                  * previous ones.  So throwing an error is the only safe
1033                                  * course.
1034                                  *
1035                                  * If a trigger actually intends this type of interaction, it
1036                                  * can re-execute the UPDATE (assuming it can figure out how)
1037                                  * and then return NULL to cancel the outer update.
1038                                  */
1039                                 if (hufd.cmax != estate->es_output_cid)
1040                                         ereport(ERROR,
1041                                                         (errcode(ERRCODE_TRIGGERED_DATA_CHANGE_VIOLATION),
1042                                                          errmsg("tuple to be updated was already modified by an operation triggered by the current command"),
1043                                                          errhint("Consider using an AFTER trigger instead of a BEFORE trigger to propagate changes to other rows.")));
1044
1045                                 /* Else, already updated by self; nothing to do */
1046                                 return NULL;
1047
1048                         case HeapTupleMayBeUpdated:
1049                                 break;
1050
1051                         case HeapTupleUpdated:
1052                                 if (IsolationUsesXactSnapshot())
1053                                         ereport(ERROR,
1054                                                         (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1055                                                          errmsg("could not serialize access due to concurrent update")));
1056                                 if (!ItemPointerEquals(tupleid, &hufd.ctid))
1057                                 {
1058                                         TupleTableSlot *epqslot;
1059
1060                                         epqslot = EvalPlanQual(estate,
1061                                                                                    epqstate,
1062                                                                                    resultRelationDesc,
1063                                                                                    resultRelInfo->ri_RangeTableIndex,
1064                                                                                    lockmode,
1065                                                                                    &hufd.ctid,
1066                                                                                    hufd.xmax);
1067                                         if (!TupIsNull(epqslot))
1068                                         {
1069                                                 *tupleid = hufd.ctid;
1070                                                 slot = ExecFilterJunk(resultRelInfo->ri_junkFilter, epqslot);
1071                                                 tuple = ExecMaterializeSlot(slot);
1072                                                 goto lreplace;
1073                                         }
1074                                 }
1075                                 /* tuple already deleted; nothing to do */
1076                                 return NULL;
1077
1078                         default:
1079                                 elog(ERROR, "unrecognized heap_update status: %u", result);
1080                                 return NULL;
1081                 }
1082
1083                 /*
1084                  * Note: instead of having to update the old index tuples associated
1085                  * with the heap tuple, all we do is form and insert new index tuples.
1086                  * This is because UPDATEs are actually DELETEs and INSERTs, and index
1087                  * tuple deletion is done later by VACUUM (see notes in ExecDelete).
1088                  * All we do here is insert new index tuples.  -cim 9/27/89
1089                  */
1090
1091                 /*
1092                  * insert index entries for tuple
1093                  *
1094                  * Note: heap_update returns the tid (location) of the new tuple in
1095                  * the t_self field.
1096                  *
1097                  * If it's a HOT update, we mustn't insert new index entries.
1098                  */
1099                 if (resultRelInfo->ri_NumIndices > 0 && !HeapTupleIsHeapOnly(tuple))
1100                         recheckIndexes = ExecInsertIndexTuples(slot, &(tuple->t_self),
1101                                                                                                    estate, false, NULL, NIL);
1102         }
1103
1104         if (canSetTag)
1105                 (estate->es_processed)++;
1106
1107         /* AFTER ROW UPDATE Triggers */
1108         ExecARUpdateTriggers(estate, resultRelInfo, tupleid, oldtuple, tuple,
1109                                                  recheckIndexes);
1110
1111         list_free(recheckIndexes);
1112
1113         /*
1114          * Check any WITH CHECK OPTION constraints from parent views.  We are
1115          * required to do this after testing all constraints and uniqueness
1116          * violations per the SQL spec, so we do it after actually updating the
1117          * record in the heap and all indexes.
1118          *
1119          * ExecWithCheckOptions() will skip any WCOs which are not of the kind we
1120          * are looking for at this point.
1121          */
1122         if (resultRelInfo->ri_WithCheckOptions != NIL)
1123                 ExecWithCheckOptions(WCO_VIEW_CHECK, resultRelInfo, slot, estate);
1124
1125         /* Process RETURNING if present */
1126         if (resultRelInfo->ri_projectReturning)
1127                 return ExecProcessReturning(resultRelInfo, slot, planSlot);
1128
1129         return NULL;
1130 }
1131
1132 /*
1133  * ExecOnConflictUpdate --- execute UPDATE of INSERT ON CONFLICT DO UPDATE
1134  *
1135  * Try to lock tuple for update as part of speculative insertion.  If
1136  * a qual originating from ON CONFLICT DO UPDATE is satisfied, update
1137  * (but still lock row, even though it may not satisfy estate's
1138  * snapshot).
1139  *
1140  * Returns true if if we're done (with or without an update), or false if
1141  * the caller must retry the INSERT from scratch.
1142  */
1143 static bool
1144 ExecOnConflictUpdate(ModifyTableState *mtstate,
1145                                          ResultRelInfo *resultRelInfo,
1146                                          ItemPointer conflictTid,
1147                                          TupleTableSlot *planSlot,
1148                                          TupleTableSlot *excludedSlot,
1149                                          EState *estate,
1150                                          bool canSetTag,
1151                                          TupleTableSlot **returning)
1152 {
1153         ExprContext *econtext = mtstate->ps.ps_ExprContext;
1154         Relation        relation = resultRelInfo->ri_RelationDesc;
1155         List       *onConflictSetWhere = resultRelInfo->ri_onConflictSetWhere;
1156         HeapTupleData tuple;
1157         HeapUpdateFailureData hufd;
1158         LockTupleMode lockmode;
1159         HTSU_Result test;
1160         Buffer          buffer;
1161
1162         /* Determine lock mode to use */
1163         lockmode = ExecUpdateLockMode(estate, resultRelInfo);
1164
1165         /*
1166          * Lock tuple for update.  Don't follow updates when tuple cannot be
1167          * locked without doing so.  A row locking conflict here means our
1168          * previous conclusion that the tuple is conclusively committed is not
1169          * true anymore.
1170          */
1171         tuple.t_self = *conflictTid;
1172         test = heap_lock_tuple(relation, &tuple, estate->es_output_cid,
1173                                                    lockmode, LockWaitBlock, false, &buffer,
1174                                                    &hufd);
1175         switch (test)
1176         {
1177                 case HeapTupleMayBeUpdated:
1178                         /* success! */
1179                         break;
1180
1181                 case HeapTupleInvisible:
1182
1183                         /*
1184                          * This can occur when a just inserted tuple is updated again in
1185                          * the same command. E.g. because multiple rows with the same
1186                          * conflicting key values are inserted.
1187                          *
1188                          * This is somewhat similar to the ExecUpdate()
1189                          * HeapTupleSelfUpdated case.  We do not want to proceed because
1190                          * it would lead to the same row being updated a second time in
1191                          * some unspecified order, and in contrast to plain UPDATEs
1192                          * there's no historical behavior to break.
1193                          *
1194                          * It is the user's responsibility to prevent this situation from
1195                          * occurring.  These problems are why SQL-2003 similarly specifies
1196                          * that for SQL MERGE, an exception must be raised in the event of
1197                          * an attempt to update the same row twice.
1198                          */
1199                         if (TransactionIdIsCurrentTransactionId(HeapTupleHeaderGetXmin(tuple.t_data)))
1200                                 ereport(ERROR,
1201                                                 (errcode(ERRCODE_CARDINALITY_VIOLATION),
1202                                                  errmsg("ON CONFLICT DO UPDATE command cannot affect row a second time"),
1203                                                  errhint("Ensure that no rows proposed for insertion within the same command have duplicate constrained values.")));
1204
1205                         /* This shouldn't happen */
1206                         elog(ERROR, "attempted to lock invisible tuple");
1207
1208                 case HeapTupleSelfUpdated:
1209
1210                         /*
1211                          * This state should never be reached. As a dirty snapshot is used
1212                          * to find conflicting tuples, speculative insertion wouldn't have
1213                          * seen this row to conflict with.
1214                          */
1215                         elog(ERROR, "unexpected self-updated tuple");
1216
1217                 case HeapTupleUpdated:
1218                         if (IsolationUsesXactSnapshot())
1219                                 ereport(ERROR,
1220                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1221                                                  errmsg("could not serialize access due to concurrent update")));
1222
1223                         /*
1224                          * Tell caller to try again from the very start.
1225                          *
1226                          * It does not make sense to use the usual EvalPlanQual() style
1227                          * loop here, as the new version of the row might not conflict
1228                          * anymore, or the conflicting tuple has actually been deleted.
1229                          */
1230                         ReleaseBuffer(buffer);
1231                         return false;
1232
1233                 default:
1234                         elog(ERROR, "unrecognized heap_lock_tuple status: %u", test);
1235         }
1236
1237         /*
1238          * Success, the tuple is locked.
1239          *
1240          * Reset per-tuple memory context to free any expression evaluation
1241          * storage allocated in the previous cycle.
1242          */
1243         ResetExprContext(econtext);
1244
1245         /*
1246          * Verify that the tuple is visible to our MVCC snapshot if the current
1247          * isolation level mandates that.
1248          *
1249          * It's not sufficient to rely on the check within ExecUpdate() as e.g.
1250          * CONFLICT ... WHERE clause may prevent us from reaching that.
1251          *
1252          * This means we only ever continue when a new command in the current
1253          * transaction could see the row, even though in READ COMMITTED mode the
1254          * tuple will not be visible according to the current statement's
1255          * snapshot.  This is in line with the way UPDATE deals with newer tuple
1256          * versions.
1257          */
1258         ExecCheckHeapTupleVisible(estate, &tuple, buffer);
1259
1260         /* Store target's existing tuple in the state's dedicated slot */
1261         ExecStoreTuple(&tuple, mtstate->mt_existing, buffer, false);
1262
1263         /*
1264          * Make tuple and any needed join variables available to ExecQual and
1265          * ExecProject.  The EXCLUDED tuple is installed in ecxt_innertuple, while
1266          * the target's existing tuple is installed in the scantuple.  EXCLUDED
1267          * has been made to reference INNER_VAR in setrefs.c, but there is no
1268          * other redirection.
1269          */
1270         econtext->ecxt_scantuple = mtstate->mt_existing;
1271         econtext->ecxt_innertuple = excludedSlot;
1272         econtext->ecxt_outertuple = NULL;
1273
1274         if (!ExecQual(onConflictSetWhere, econtext, false))
1275         {
1276                 ReleaseBuffer(buffer);
1277                 InstrCountFiltered1(&mtstate->ps, 1);
1278                 return true;                    /* done with the tuple */
1279         }
1280
1281         if (resultRelInfo->ri_WithCheckOptions != NIL)
1282         {
1283                 /*
1284                  * Check target's existing tuple against UPDATE-applicable USING
1285                  * security barrier quals (if any), enforced here as RLS checks/WCOs.
1286                  *
1287                  * The rewriter creates UPDATE RLS checks/WCOs for UPDATE security
1288                  * quals, and stores them as WCOs of "kind" WCO_RLS_CONFLICT_CHECK,
1289                  * but that's almost the extent of its special handling for ON
1290                  * CONFLICT DO UPDATE.
1291                  *
1292                  * The rewriter will also have associated UPDATE applicable straight
1293                  * RLS checks/WCOs for the benefit of the ExecUpdate() call that
1294                  * follows.  INSERTs and UPDATEs naturally have mutually exclusive WCO
1295                  * kinds, so there is no danger of spurious over-enforcement in the
1296                  * INSERT or UPDATE path.
1297                  */
1298                 ExecWithCheckOptions(WCO_RLS_CONFLICT_CHECK, resultRelInfo,
1299                                                          mtstate->mt_existing,
1300                                                          mtstate->ps.state);
1301         }
1302
1303         /* Project the new tuple version */
1304         ExecProject(resultRelInfo->ri_onConflictSetProj);
1305
1306         /*
1307          * Note that it is possible that the target tuple has been modified in
1308          * this session, after the above heap_lock_tuple. We choose to not error
1309          * out in that case, in line with ExecUpdate's treatment of similar cases.
1310          * This can happen if an UPDATE is triggered from within ExecQual(),
1311          * ExecWithCheckOptions() or ExecProject() above, e.g. by selecting from a
1312          * wCTE in the ON CONFLICT's SET.
1313          */
1314
1315         /* Execute UPDATE with projection */
1316         *returning = ExecUpdate(&tuple.t_self, NULL,
1317                                                         mtstate->mt_conflproj, planSlot,
1318                                                         &mtstate->mt_epqstate, mtstate->ps.state,
1319                                                         canSetTag);
1320
1321         ReleaseBuffer(buffer);
1322         return true;
1323 }
1324
1325
1326 /*
1327  * Process BEFORE EACH STATEMENT triggers
1328  */
1329 static void
1330 fireBSTriggers(ModifyTableState *node)
1331 {
1332         switch (node->operation)
1333         {
1334                 case CMD_INSERT:
1335                         ExecBSInsertTriggers(node->ps.state, node->resultRelInfo);
1336                         if (node->mt_onconflict == ONCONFLICT_UPDATE)
1337                                 ExecBSUpdateTriggers(node->ps.state,
1338                                                                          node->resultRelInfo);
1339                         break;
1340                 case CMD_UPDATE:
1341                         ExecBSUpdateTriggers(node->ps.state, node->resultRelInfo);
1342                         break;
1343                 case CMD_DELETE:
1344                         ExecBSDeleteTriggers(node->ps.state, node->resultRelInfo);
1345                         break;
1346                 default:
1347                         elog(ERROR, "unknown operation");
1348                         break;
1349         }
1350 }
1351
1352 /*
1353  * Process AFTER EACH STATEMENT triggers
1354  */
1355 static void
1356 fireASTriggers(ModifyTableState *node)
1357 {
1358         switch (node->operation)
1359         {
1360                 case CMD_INSERT:
1361                         if (node->mt_onconflict == ONCONFLICT_UPDATE)
1362                                 ExecASUpdateTriggers(node->ps.state,
1363                                                                          node->resultRelInfo);
1364                         ExecASInsertTriggers(node->ps.state, node->resultRelInfo);
1365                         break;
1366                 case CMD_UPDATE:
1367                         ExecASUpdateTriggers(node->ps.state, node->resultRelInfo);
1368                         break;
1369                 case CMD_DELETE:
1370                         ExecASDeleteTriggers(node->ps.state, node->resultRelInfo);
1371                         break;
1372                 default:
1373                         elog(ERROR, "unknown operation");
1374                         break;
1375         }
1376 }
1377
1378
1379 /* ----------------------------------------------------------------
1380  *         ExecModifyTable
1381  *
1382  *              Perform table modifications as required, and return RETURNING results
1383  *              if needed.
1384  * ----------------------------------------------------------------
1385  */
1386 TupleTableSlot *
1387 ExecModifyTable(ModifyTableState *node)
1388 {
1389         EState     *estate = node->ps.state;
1390         CmdType         operation = node->operation;
1391         ResultRelInfo *saved_resultRelInfo;
1392         ResultRelInfo *resultRelInfo;
1393         PlanState  *subplanstate;
1394         JunkFilter *junkfilter;
1395         TupleTableSlot *slot;
1396         TupleTableSlot *planSlot;
1397         ItemPointer tupleid = NULL;
1398         ItemPointerData tuple_ctid;
1399         HeapTupleData oldtupdata;
1400         HeapTuple       oldtuple;
1401
1402         /*
1403          * This should NOT get called during EvalPlanQual; we should have passed a
1404          * subplan tree to EvalPlanQual, instead.  Use a runtime test not just
1405          * Assert because this condition is easy to miss in testing.  (Note:
1406          * although ModifyTable should not get executed within an EvalPlanQual
1407          * operation, we do have to allow it to be initialized and shut down in
1408          * case it is within a CTE subplan.  Hence this test must be here, not in
1409          * ExecInitModifyTable.)
1410          */
1411         if (estate->es_epqTuple != NULL)
1412                 elog(ERROR, "ModifyTable should not be called during EvalPlanQual");
1413
1414         /*
1415          * If we've already completed processing, don't try to do more.  We need
1416          * this test because ExecPostprocessPlan might call us an extra time, and
1417          * our subplan's nodes aren't necessarily robust against being called
1418          * extra times.
1419          */
1420         if (node->mt_done)
1421                 return NULL;
1422
1423         /*
1424          * On first call, fire BEFORE STATEMENT triggers before proceeding.
1425          */
1426         if (node->fireBSTriggers)
1427         {
1428                 fireBSTriggers(node);
1429                 node->fireBSTriggers = false;
1430         }
1431
1432         /* Preload local variables */
1433         resultRelInfo = node->resultRelInfo + node->mt_whichplan;
1434         subplanstate = node->mt_plans[node->mt_whichplan];
1435         junkfilter = resultRelInfo->ri_junkFilter;
1436
1437         /*
1438          * es_result_relation_info must point to the currently active result
1439          * relation while we are within this ModifyTable node.  Even though
1440          * ModifyTable nodes can't be nested statically, they can be nested
1441          * dynamically (since our subplan could include a reference to a modifying
1442          * CTE).  So we have to save and restore the caller's value.
1443          */
1444         saved_resultRelInfo = estate->es_result_relation_info;
1445
1446         estate->es_result_relation_info = resultRelInfo;
1447
1448         /*
1449          * Fetch rows from subplan(s), and execute the required table modification
1450          * for each row.
1451          */
1452         for (;;)
1453         {
1454                 /*
1455                  * Reset the per-output-tuple exprcontext.  This is needed because
1456                  * triggers expect to use that context as workspace.  It's a bit ugly
1457                  * to do this below the top level of the plan, however.  We might need
1458                  * to rethink this later.
1459                  */
1460                 ResetPerTupleExprContext(estate);
1461
1462                 planSlot = ExecProcNode(subplanstate);
1463
1464                 if (TupIsNull(planSlot))
1465                 {
1466                         /* advance to next subplan if any */
1467                         node->mt_whichplan++;
1468                         if (node->mt_whichplan < node->mt_nplans)
1469                         {
1470                                 resultRelInfo++;
1471                                 subplanstate = node->mt_plans[node->mt_whichplan];
1472                                 junkfilter = resultRelInfo->ri_junkFilter;
1473                                 estate->es_result_relation_info = resultRelInfo;
1474                                 EvalPlanQualSetPlan(&node->mt_epqstate, subplanstate->plan,
1475                                                                         node->mt_arowmarks[node->mt_whichplan]);
1476                                 continue;
1477                         }
1478                         else
1479                                 break;
1480                 }
1481
1482                 /*
1483                  * If resultRelInfo->ri_usesFdwDirectModify is true, all we need to do
1484                  * here is compute the RETURNING expressions.
1485                  */
1486                 if (resultRelInfo->ri_usesFdwDirectModify)
1487                 {
1488                         Assert(resultRelInfo->ri_projectReturning);
1489
1490                         /*
1491                          * A scan slot containing the data that was actually inserted,
1492                          * updated or deleted has already been made available to
1493                          * ExecProcessReturning by IterateDirectModify, so no need to
1494                          * provide it here.
1495                          */
1496                         slot = ExecProcessReturning(resultRelInfo, NULL, planSlot);
1497
1498                         estate->es_result_relation_info = saved_resultRelInfo;
1499                         return slot;
1500                 }
1501
1502                 EvalPlanQualSetSlot(&node->mt_epqstate, planSlot);
1503                 slot = planSlot;
1504
1505                 oldtuple = NULL;
1506                 if (junkfilter != NULL)
1507                 {
1508                         /*
1509                          * extract the 'ctid' or 'wholerow' junk attribute.
1510                          */
1511                         if (operation == CMD_UPDATE || operation == CMD_DELETE)
1512                         {
1513                                 char            relkind;
1514                                 Datum           datum;
1515                                 bool            isNull;
1516
1517                                 relkind = resultRelInfo->ri_RelationDesc->rd_rel->relkind;
1518                                 if (relkind == RELKIND_RELATION || relkind == RELKIND_MATVIEW)
1519                                 {
1520                                         datum = ExecGetJunkAttribute(slot,
1521                                                                                                  junkfilter->jf_junkAttNo,
1522                                                                                                  &isNull);
1523                                         /* shouldn't ever get a null result... */
1524                                         if (isNull)
1525                                                 elog(ERROR, "ctid is NULL");
1526
1527                                         tupleid = (ItemPointer) DatumGetPointer(datum);
1528                                         tuple_ctid = *tupleid;          /* be sure we don't free
1529                                                                                                  * ctid!! */
1530                                         tupleid = &tuple_ctid;
1531                                 }
1532
1533                                 /*
1534                                  * Use the wholerow attribute, when available, to reconstruct
1535                                  * the old relation tuple.
1536                                  *
1537                                  * Foreign table updates have a wholerow attribute when the
1538                                  * relation has an AFTER ROW trigger.  Note that the wholerow
1539                                  * attribute does not carry system columns.  Foreign table
1540                                  * triggers miss seeing those, except that we know enough here
1541                                  * to set t_tableOid.  Quite separately from this, the FDW may
1542                                  * fetch its own junk attrs to identify the row.
1543                                  *
1544                                  * Other relevant relkinds, currently limited to views, always
1545                                  * have a wholerow attribute.
1546                                  */
1547                                 else if (AttributeNumberIsValid(junkfilter->jf_junkAttNo))
1548                                 {
1549                                         datum = ExecGetJunkAttribute(slot,
1550                                                                                                  junkfilter->jf_junkAttNo,
1551                                                                                                  &isNull);
1552                                         /* shouldn't ever get a null result... */
1553                                         if (isNull)
1554                                                 elog(ERROR, "wholerow is NULL");
1555
1556                                         oldtupdata.t_data = DatumGetHeapTupleHeader(datum);
1557                                         oldtupdata.t_len =
1558                                                 HeapTupleHeaderGetDatumLength(oldtupdata.t_data);
1559                                         ItemPointerSetInvalid(&(oldtupdata.t_self));
1560                                         /* Historically, view triggers see invalid t_tableOid. */
1561                                         oldtupdata.t_tableOid =
1562                                                 (relkind == RELKIND_VIEW) ? InvalidOid :
1563                                                 RelationGetRelid(resultRelInfo->ri_RelationDesc);
1564
1565                                         oldtuple = &oldtupdata;
1566                                 }
1567                                 else
1568                                         Assert(relkind == RELKIND_FOREIGN_TABLE);
1569                         }
1570
1571                         /*
1572                          * apply the junkfilter if needed.
1573                          */
1574                         if (operation != CMD_DELETE)
1575                                 slot = ExecFilterJunk(junkfilter, slot);
1576                 }
1577
1578                 switch (operation)
1579                 {
1580                         case CMD_INSERT:
1581                                 slot = ExecInsert(node, slot, planSlot,
1582                                                                 node->mt_arbiterindexes, node->mt_onconflict,
1583                                                                   estate, node->canSetTag);
1584                                 break;
1585                         case CMD_UPDATE:
1586                                 slot = ExecUpdate(tupleid, oldtuple, slot, planSlot,
1587                                                                 &node->mt_epqstate, estate, node->canSetTag);
1588                                 break;
1589                         case CMD_DELETE:
1590                                 slot = ExecDelete(tupleid, oldtuple, planSlot,
1591                                                                 &node->mt_epqstate, estate, node->canSetTag);
1592                                 break;
1593                         default:
1594                                 elog(ERROR, "unknown operation");
1595                                 break;
1596                 }
1597
1598                 /*
1599                  * If we got a RETURNING result, return it to caller.  We'll continue
1600                  * the work on next call.
1601                  */
1602                 if (slot)
1603                 {
1604                         estate->es_result_relation_info = saved_resultRelInfo;
1605                         return slot;
1606                 }
1607         }
1608
1609         /* Restore es_result_relation_info before exiting */
1610         estate->es_result_relation_info = saved_resultRelInfo;
1611
1612         /*
1613          * We're done, but fire AFTER STATEMENT triggers before exiting.
1614          */
1615         fireASTriggers(node);
1616
1617         node->mt_done = true;
1618
1619         return NULL;
1620 }
1621
1622 /* ----------------------------------------------------------------
1623  *              ExecInitModifyTable
1624  * ----------------------------------------------------------------
1625  */
1626 ModifyTableState *
1627 ExecInitModifyTable(ModifyTable *node, EState *estate, int eflags)
1628 {
1629         ModifyTableState *mtstate;
1630         CmdType         operation = node->operation;
1631         int                     nplans = list_length(node->plans);
1632         ResultRelInfo *saved_resultRelInfo;
1633         ResultRelInfo *resultRelInfo;
1634         TupleDesc       tupDesc;
1635         Plan       *subplan;
1636         ListCell   *l;
1637         int                     i;
1638         Relation        rel;
1639
1640         /* check for unsupported flags */
1641         Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
1642
1643         /*
1644          * create state structure
1645          */
1646         mtstate = makeNode(ModifyTableState);
1647         mtstate->ps.plan = (Plan *) node;
1648         mtstate->ps.state = estate;
1649         mtstate->ps.targetlist = NIL;           /* not actually used */
1650
1651         mtstate->operation = operation;
1652         mtstate->canSetTag = node->canSetTag;
1653         mtstate->mt_done = false;
1654
1655         mtstate->mt_plans = (PlanState **) palloc0(sizeof(PlanState *) * nplans);
1656         mtstate->resultRelInfo = estate->es_result_relations + node->resultRelIndex;
1657         mtstate->mt_arowmarks = (List **) palloc0(sizeof(List *) * nplans);
1658         mtstate->mt_nplans = nplans;
1659         mtstate->mt_onconflict = node->onConflictAction;
1660         mtstate->mt_arbiterindexes = node->arbiterIndexes;
1661
1662         /* set up epqstate with dummy subplan data for the moment */
1663         EvalPlanQualInit(&mtstate->mt_epqstate, estate, NULL, NIL, node->epqParam);
1664         mtstate->fireBSTriggers = true;
1665
1666         /*
1667          * call ExecInitNode on each of the plans to be executed and save the
1668          * results into the array "mt_plans".  This is also a convenient place to
1669          * verify that the proposed target relations are valid and open their
1670          * indexes for insertion of new index entries.  Note we *must* set
1671          * estate->es_result_relation_info correctly while we initialize each
1672          * sub-plan; ExecContextForcesOids depends on that!
1673          */
1674         saved_resultRelInfo = estate->es_result_relation_info;
1675
1676         resultRelInfo = mtstate->resultRelInfo;
1677         i = 0;
1678         foreach(l, node->plans)
1679         {
1680                 subplan = (Plan *) lfirst(l);
1681
1682                 /* Initialize the usesFdwDirectModify flag */
1683                 resultRelInfo->ri_usesFdwDirectModify = bms_is_member(i,
1684                                                                                                  node->fdwDirectModifyPlans);
1685
1686                 /*
1687                  * Verify result relation is a valid target for the current operation
1688                  */
1689                 CheckValidResultRel(resultRelInfo->ri_RelationDesc, operation);
1690
1691                 /*
1692                  * If there are indices on the result relation, open them and save
1693                  * descriptors in the result relation info, so that we can add new
1694                  * index entries for the tuples we add/update.  We need not do this
1695                  * for a DELETE, however, since deletion doesn't affect indexes. Also,
1696                  * inside an EvalPlanQual operation, the indexes might be open
1697                  * already, since we share the resultrel state with the original
1698                  * query.
1699                  */
1700                 if (resultRelInfo->ri_RelationDesc->rd_rel->relhasindex &&
1701                         operation != CMD_DELETE &&
1702                         resultRelInfo->ri_IndexRelationDescs == NULL)
1703                         ExecOpenIndices(resultRelInfo, mtstate->mt_onconflict != ONCONFLICT_NONE);
1704
1705                 /* Now init the plan for this result rel */
1706                 estate->es_result_relation_info = resultRelInfo;
1707                 mtstate->mt_plans[i] = ExecInitNode(subplan, estate, eflags);
1708
1709                 /* Also let FDWs init themselves for foreign-table result rels */
1710                 if (!resultRelInfo->ri_usesFdwDirectModify &&
1711                         resultRelInfo->ri_FdwRoutine != NULL &&
1712                         resultRelInfo->ri_FdwRoutine->BeginForeignModify != NULL)
1713                 {
1714                         List       *fdw_private = (List *) list_nth(node->fdwPrivLists, i);
1715
1716                         resultRelInfo->ri_FdwRoutine->BeginForeignModify(mtstate,
1717                                                                                                                          resultRelInfo,
1718                                                                                                                          fdw_private,
1719                                                                                                                          i,
1720                                                                                                                          eflags);
1721                 }
1722
1723                 resultRelInfo++;
1724                 i++;
1725         }
1726
1727         estate->es_result_relation_info = saved_resultRelInfo;
1728
1729         /* The root table RT index is at the head of the partitioned_rels list */
1730         if (node->partitioned_rels)
1731         {
1732                 Index   root_rti;
1733                 Oid             root_oid;
1734
1735                 root_rti = linitial_int(node->partitioned_rels);
1736                 root_oid = getrelid(root_rti, estate->es_range_table);
1737                 rel = heap_open(root_oid, NoLock);      /* locked by InitPlan */
1738         }
1739         else
1740                 rel = mtstate->resultRelInfo->ri_RelationDesc;
1741
1742         /* Build state for INSERT tuple routing */
1743         if (operation == CMD_INSERT &&
1744                 rel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
1745         {
1746                 PartitionDispatch *partition_dispatch_info;
1747                 ResultRelInfo *partitions;
1748                 TupleConversionMap **partition_tupconv_maps;
1749                 TupleTableSlot *partition_tuple_slot;
1750                 int                     num_parted,
1751                                         num_partitions;
1752
1753                 ExecSetupPartitionTupleRouting(rel,
1754                                                                            &partition_dispatch_info,
1755                                                                            &partitions,
1756                                                                            &partition_tupconv_maps,
1757                                                                            &partition_tuple_slot,
1758                                                                            &num_parted, &num_partitions);
1759                 mtstate->mt_partition_dispatch_info = partition_dispatch_info;
1760                 mtstate->mt_num_dispatch = num_parted;
1761                 mtstate->mt_partitions = partitions;
1762                 mtstate->mt_num_partitions = num_partitions;
1763                 mtstate->mt_partition_tupconv_maps = partition_tupconv_maps;
1764                 mtstate->mt_partition_tuple_slot = partition_tuple_slot;
1765         }
1766
1767         /*
1768          * Initialize any WITH CHECK OPTION constraints if needed.
1769          */
1770         resultRelInfo = mtstate->resultRelInfo;
1771         i = 0;
1772         foreach(l, node->withCheckOptionLists)
1773         {
1774                 List       *wcoList = (List *) lfirst(l);
1775                 List       *wcoExprs = NIL;
1776                 ListCell   *ll;
1777
1778                 foreach(ll, wcoList)
1779                 {
1780                         WithCheckOption *wco = (WithCheckOption *) lfirst(ll);
1781                         ExprState  *wcoExpr = ExecInitExpr((Expr *) wco->qual,
1782                                                                                            mtstate->mt_plans[i]);
1783
1784                         wcoExprs = lappend(wcoExprs, wcoExpr);
1785                 }
1786
1787                 resultRelInfo->ri_WithCheckOptions = wcoList;
1788                 resultRelInfo->ri_WithCheckOptionExprs = wcoExprs;
1789                 resultRelInfo++;
1790                 i++;
1791         }
1792
1793         /*
1794          * Build WITH CHECK OPTION constraints for each leaf partition rel.
1795          * Note that we didn't build the withCheckOptionList for each partition
1796          * within the planner, but simple translation of the varattnos for each
1797          * partition will suffice.  This only occurs for the INSERT case;
1798          * UPDATE/DELETE cases are handled above.
1799          */
1800         if (node->withCheckOptionLists != NIL && mtstate->mt_num_partitions > 0)
1801         {
1802                 List            *wcoList;
1803
1804                 Assert(operation == CMD_INSERT);
1805                 resultRelInfo = mtstate->mt_partitions;
1806                 wcoList = linitial(node->withCheckOptionLists);
1807                 for (i = 0; i < mtstate->mt_num_partitions; i++)
1808                 {
1809                         Relation        partrel = resultRelInfo->ri_RelationDesc;
1810                         List       *mapped_wcoList;
1811                         List       *wcoExprs = NIL;
1812                         ListCell   *ll;
1813
1814                         /* varno = node->nominalRelation */
1815                         mapped_wcoList = map_partition_varattnos(wcoList,
1816                                                                                                          node->nominalRelation,
1817                                                                                                          partrel, rel);
1818                         foreach(ll, mapped_wcoList)
1819                         {
1820                                 WithCheckOption *wco = (WithCheckOption *) lfirst(ll);
1821                                 ExprState  *wcoExpr = ExecInitExpr((Expr *) wco->qual,
1822                                                                                            mtstate->mt_plans[i]);
1823
1824                                 wcoExprs = lappend(wcoExprs, wcoExpr);
1825                         }
1826
1827                         resultRelInfo->ri_WithCheckOptions = mapped_wcoList;
1828                         resultRelInfo->ri_WithCheckOptionExprs = wcoExprs;
1829                         resultRelInfo++;
1830                 }
1831         }
1832
1833         /*
1834          * Initialize RETURNING projections if needed.
1835          */
1836         if (node->returningLists)
1837         {
1838                 TupleTableSlot *slot;
1839                 ExprContext *econtext;
1840                 List       *returningList;
1841
1842                 /*
1843                  * Initialize result tuple slot and assign its rowtype using the first
1844                  * RETURNING list.  We assume the rest will look the same.
1845                  */
1846                 tupDesc = ExecTypeFromTL((List *) linitial(node->returningLists),
1847                                                                  false);
1848
1849                 /* Set up a slot for the output of the RETURNING projection(s) */
1850                 ExecInitResultTupleSlot(estate, &mtstate->ps);
1851                 ExecAssignResultType(&mtstate->ps, tupDesc);
1852                 slot = mtstate->ps.ps_ResultTupleSlot;
1853
1854                 /* Need an econtext too */
1855                 econtext = CreateExprContext(estate);
1856                 mtstate->ps.ps_ExprContext = econtext;
1857
1858                 /*
1859                  * Build a projection for each result rel.
1860                  */
1861                 resultRelInfo = mtstate->resultRelInfo;
1862                 foreach(l, node->returningLists)
1863                 {
1864                         List       *rlist = (List *) lfirst(l);
1865                         List       *rliststate;
1866
1867                         rliststate = (List *) ExecInitExpr((Expr *) rlist, &mtstate->ps);
1868                         resultRelInfo->ri_projectReturning =
1869                                 ExecBuildProjectionInfo(rliststate, econtext, slot,
1870                                                                          resultRelInfo->ri_RelationDesc->rd_att);
1871                         resultRelInfo++;
1872                 }
1873
1874                 /*
1875                  * Build a projection for each leaf partition rel.  Note that we
1876                  * didn't build the returningList for each partition within the
1877                  * planner, but simple translation of the varattnos for each partition
1878                  * will suffice.  This only occurs for the INSERT case; UPDATE/DELETE
1879                  * are handled above.
1880                  */
1881                 resultRelInfo = mtstate->mt_partitions;
1882                 returningList = linitial(node->returningLists);
1883                 for (i = 0; i < mtstate->mt_num_partitions; i++)
1884                 {
1885                         Relation        partrel = resultRelInfo->ri_RelationDesc;
1886                         List       *rlist,
1887                                            *rliststate;
1888
1889                         /* varno = node->nominalRelation */
1890                         rlist = map_partition_varattnos(returningList,
1891                                                                                         node->nominalRelation,
1892                                                                                         partrel, rel);
1893                         rliststate = (List *) ExecInitExpr((Expr *) rlist, &mtstate->ps);
1894                         resultRelInfo->ri_projectReturning =
1895                                 ExecBuildProjectionInfo(rliststate, econtext, slot,
1896                                                                          resultRelInfo->ri_RelationDesc->rd_att);
1897                         resultRelInfo++;
1898                 }
1899         }
1900         else
1901         {
1902                 /*
1903                  * We still must construct a dummy result tuple type, because InitPlan
1904                  * expects one (maybe should change that?).
1905                  */
1906                 tupDesc = ExecTypeFromTL(NIL, false);
1907                 ExecInitResultTupleSlot(estate, &mtstate->ps);
1908                 ExecAssignResultType(&mtstate->ps, tupDesc);
1909
1910                 mtstate->ps.ps_ExprContext = NULL;
1911         }
1912
1913         /* Close the root partitioned rel if we opened it above. */
1914         if (rel != mtstate->resultRelInfo->ri_RelationDesc)
1915                 heap_close(rel, NoLock);
1916
1917         /*
1918          * If needed, Initialize target list, projection and qual for ON CONFLICT
1919          * DO UPDATE.
1920          */
1921         resultRelInfo = mtstate->resultRelInfo;
1922         if (node->onConflictAction == ONCONFLICT_UPDATE)
1923         {
1924                 ExprContext *econtext;
1925                 ExprState  *setexpr;
1926                 TupleDesc       tupDesc;
1927
1928                 /* insert may only have one plan, inheritance is not expanded */
1929                 Assert(nplans == 1);
1930
1931                 /* already exists if created by RETURNING processing above */
1932                 if (mtstate->ps.ps_ExprContext == NULL)
1933                         ExecAssignExprContext(estate, &mtstate->ps);
1934
1935                 econtext = mtstate->ps.ps_ExprContext;
1936
1937                 /* initialize slot for the existing tuple */
1938                 mtstate->mt_existing = ExecInitExtraTupleSlot(mtstate->ps.state);
1939                 ExecSetSlotDescriptor(mtstate->mt_existing,
1940                                                           resultRelInfo->ri_RelationDesc->rd_att);
1941
1942                 /* carried forward solely for the benefit of explain */
1943                 mtstate->mt_excludedtlist = node->exclRelTlist;
1944
1945                 /* create target slot for UPDATE SET projection */
1946                 tupDesc = ExecTypeFromTL((List *) node->onConflictSet,
1947                                                  resultRelInfo->ri_RelationDesc->rd_rel->relhasoids);
1948                 mtstate->mt_conflproj = ExecInitExtraTupleSlot(mtstate->ps.state);
1949                 ExecSetSlotDescriptor(mtstate->mt_conflproj, tupDesc);
1950
1951                 /* build UPDATE SET expression and projection state */
1952                 setexpr = ExecInitExpr((Expr *) node->onConflictSet, &mtstate->ps);
1953                 resultRelInfo->ri_onConflictSetProj =
1954                         ExecBuildProjectionInfo((List *) setexpr, econtext,
1955                                                                         mtstate->mt_conflproj,
1956                                                                         resultRelInfo->ri_RelationDesc->rd_att);
1957
1958                 /* build DO UPDATE WHERE clause expression */
1959                 if (node->onConflictWhere)
1960                 {
1961                         ExprState  *qualexpr;
1962
1963                         qualexpr = ExecInitExpr((Expr *) node->onConflictWhere,
1964                                                                         &mtstate->ps);
1965
1966                         resultRelInfo->ri_onConflictSetWhere = (List *) qualexpr;
1967                 }
1968         }
1969
1970         /*
1971          * If we have any secondary relations in an UPDATE or DELETE, they need to
1972          * be treated like non-locked relations in SELECT FOR UPDATE, ie, the
1973          * EvalPlanQual mechanism needs to be told about them.  Locate the
1974          * relevant ExecRowMarks.
1975          */
1976         foreach(l, node->rowMarks)
1977         {
1978                 PlanRowMark *rc = castNode(PlanRowMark, lfirst(l));
1979                 ExecRowMark *erm;
1980
1981                 /* ignore "parent" rowmarks; they are irrelevant at runtime */
1982                 if (rc->isParent)
1983                         continue;
1984
1985                 /* find ExecRowMark (same for all subplans) */
1986                 erm = ExecFindRowMark(estate, rc->rti, false);
1987
1988                 /* build ExecAuxRowMark for each subplan */
1989                 for (i = 0; i < nplans; i++)
1990                 {
1991                         ExecAuxRowMark *aerm;
1992
1993                         subplan = mtstate->mt_plans[i]->plan;
1994                         aerm = ExecBuildAuxRowMark(erm, subplan->targetlist);
1995                         mtstate->mt_arowmarks[i] = lappend(mtstate->mt_arowmarks[i], aerm);
1996                 }
1997         }
1998
1999         /* select first subplan */
2000         mtstate->mt_whichplan = 0;
2001         subplan = (Plan *) linitial(node->plans);
2002         EvalPlanQualSetPlan(&mtstate->mt_epqstate, subplan,
2003                                                 mtstate->mt_arowmarks[0]);
2004
2005         /*
2006          * Initialize the junk filter(s) if needed.  INSERT queries need a filter
2007          * if there are any junk attrs in the tlist.  UPDATE and DELETE always
2008          * need a filter, since there's always a junk 'ctid' or 'wholerow'
2009          * attribute present --- no need to look first.
2010          *
2011          * If there are multiple result relations, each one needs its own junk
2012          * filter.  Note multiple rels are only possible for UPDATE/DELETE, so we
2013          * can't be fooled by some needing a filter and some not.
2014          *
2015          * This section of code is also a convenient place to verify that the
2016          * output of an INSERT or UPDATE matches the target table(s).
2017          */
2018         {
2019                 bool            junk_filter_needed = false;
2020
2021                 switch (operation)
2022                 {
2023                         case CMD_INSERT:
2024                                 foreach(l, subplan->targetlist)
2025                                 {
2026                                         TargetEntry *tle = (TargetEntry *) lfirst(l);
2027
2028                                         if (tle->resjunk)
2029                                         {
2030                                                 junk_filter_needed = true;
2031                                                 break;
2032                                         }
2033                                 }
2034                                 break;
2035                         case CMD_UPDATE:
2036                         case CMD_DELETE:
2037                                 junk_filter_needed = true;
2038                                 break;
2039                         default:
2040                                 elog(ERROR, "unknown operation");
2041                                 break;
2042                 }
2043
2044                 if (junk_filter_needed)
2045                 {
2046                         resultRelInfo = mtstate->resultRelInfo;
2047                         for (i = 0; i < nplans; i++)
2048                         {
2049                                 JunkFilter *j;
2050
2051                                 subplan = mtstate->mt_plans[i]->plan;
2052                                 if (operation == CMD_INSERT || operation == CMD_UPDATE)
2053                                         ExecCheckPlanOutput(resultRelInfo->ri_RelationDesc,
2054                                                                                 subplan->targetlist);
2055
2056                                 j = ExecInitJunkFilter(subplan->targetlist,
2057                                                         resultRelInfo->ri_RelationDesc->rd_att->tdhasoid,
2058                                                                            ExecInitExtraTupleSlot(estate));
2059
2060                                 if (operation == CMD_UPDATE || operation == CMD_DELETE)
2061                                 {
2062                                         /* For UPDATE/DELETE, find the appropriate junk attr now */
2063                                         char            relkind;
2064
2065                                         relkind = resultRelInfo->ri_RelationDesc->rd_rel->relkind;
2066                                         if (relkind == RELKIND_RELATION ||
2067                                                 relkind == RELKIND_MATVIEW ||
2068                                                 relkind == RELKIND_PARTITIONED_TABLE)
2069                                         {
2070                                                 j->jf_junkAttNo = ExecFindJunkAttribute(j, "ctid");
2071                                                 if (!AttributeNumberIsValid(j->jf_junkAttNo))
2072                                                         elog(ERROR, "could not find junk ctid column");
2073                                         }
2074                                         else if (relkind == RELKIND_FOREIGN_TABLE)
2075                                         {
2076                                                 /*
2077                                                  * When there is an AFTER trigger, there should be a
2078                                                  * wholerow attribute.
2079                                                  */
2080                                                 j->jf_junkAttNo = ExecFindJunkAttribute(j, "wholerow");
2081                                         }
2082                                         else
2083                                         {
2084                                                 j->jf_junkAttNo = ExecFindJunkAttribute(j, "wholerow");
2085                                                 if (!AttributeNumberIsValid(j->jf_junkAttNo))
2086                                                         elog(ERROR, "could not find junk wholerow column");
2087                                         }
2088                                 }
2089
2090                                 resultRelInfo->ri_junkFilter = j;
2091                                 resultRelInfo++;
2092                         }
2093                 }
2094                 else
2095                 {
2096                         if (operation == CMD_INSERT)
2097                                 ExecCheckPlanOutput(mtstate->resultRelInfo->ri_RelationDesc,
2098                                                                         subplan->targetlist);
2099                 }
2100         }
2101
2102         /*
2103          * Set up a tuple table slot for use for trigger output tuples. In a plan
2104          * containing multiple ModifyTable nodes, all can share one such slot, so
2105          * we keep it in the estate.
2106          */
2107         if (estate->es_trig_tuple_slot == NULL)
2108                 estate->es_trig_tuple_slot = ExecInitExtraTupleSlot(estate);
2109
2110         /*
2111          * Lastly, if this is not the primary (canSetTag) ModifyTable node, add it
2112          * to estate->es_auxmodifytables so that it will be run to completion by
2113          * ExecPostprocessPlan.  (It'd actually work fine to add the primary
2114          * ModifyTable node too, but there's no need.)  Note the use of lcons not
2115          * lappend: we need later-initialized ModifyTable nodes to be shut down
2116          * before earlier ones.  This ensures that we don't throw away RETURNING
2117          * rows that need to be seen by a later CTE subplan.
2118          */
2119         if (!mtstate->canSetTag)
2120                 estate->es_auxmodifytables = lcons(mtstate,
2121                                                                                    estate->es_auxmodifytables);
2122
2123         return mtstate;
2124 }
2125
2126 /* ----------------------------------------------------------------
2127  *              ExecEndModifyTable
2128  *
2129  *              Shuts down the plan.
2130  *
2131  *              Returns nothing of interest.
2132  * ----------------------------------------------------------------
2133  */
2134 void
2135 ExecEndModifyTable(ModifyTableState *node)
2136 {
2137         int                     i;
2138
2139         /*
2140          * Allow any FDWs to shut down
2141          */
2142         for (i = 0; i < node->mt_nplans; i++)
2143         {
2144                 ResultRelInfo *resultRelInfo = node->resultRelInfo + i;
2145
2146                 if (!resultRelInfo->ri_usesFdwDirectModify &&
2147                         resultRelInfo->ri_FdwRoutine != NULL &&
2148                         resultRelInfo->ri_FdwRoutine->EndForeignModify != NULL)
2149                         resultRelInfo->ri_FdwRoutine->EndForeignModify(node->ps.state,
2150                                                                                                                    resultRelInfo);
2151         }
2152
2153         /*
2154          * Close all the partitioned tables, leaf partitions, and their indices
2155          *
2156          * Remember node->mt_partition_dispatch_info[0] corresponds to the root
2157          * partitioned table, which we must not try to close, because it is the
2158          * main target table of the query that will be closed by ExecEndPlan().
2159          * Also, tupslot is NULL for the root partitioned table.
2160          */
2161         for (i = 1; i < node->mt_num_dispatch; i++)
2162         {
2163                 PartitionDispatch pd = node->mt_partition_dispatch_info[i];
2164
2165                 heap_close(pd->reldesc, NoLock);
2166                 ExecDropSingleTupleTableSlot(pd->tupslot);
2167         }
2168         for (i = 0; i < node->mt_num_partitions; i++)
2169         {
2170                 ResultRelInfo *resultRelInfo = node->mt_partitions + i;
2171
2172                 ExecCloseIndices(resultRelInfo);
2173                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
2174         }
2175
2176         /* Release the standalone partition tuple descriptor, if any */
2177         if (node->mt_partition_tuple_slot)
2178                 ExecDropSingleTupleTableSlot(node->mt_partition_tuple_slot);
2179
2180         /*
2181          * Free the exprcontext
2182          */
2183         ExecFreeExprContext(&node->ps);
2184
2185         /*
2186          * clean out the tuple table
2187          */
2188         ExecClearTuple(node->ps.ps_ResultTupleSlot);
2189
2190         /*
2191          * Terminate EPQ execution if active
2192          */
2193         EvalPlanQualEnd(&node->mt_epqstate);
2194
2195         /*
2196          * shut down subplans
2197          */
2198         for (i = 0; i < node->mt_nplans; i++)
2199                 ExecEndNode(node->mt_plans[i]);
2200 }
2201
2202 void
2203 ExecReScanModifyTable(ModifyTableState *node)
2204 {
2205         /*
2206          * Currently, we don't need to support rescan on ModifyTable nodes. The
2207          * semantics of that would be a bit debatable anyway.
2208          */
2209         elog(ERROR, "ExecReScanModifyTable is not implemented");
2210 }