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

   1 /*-------------------------------------------------------------------------
   2  *
   3  * execIndexing.c
   4  *        routines for inserting index tuples and enforcing unique and
   5  *        exclusion constraints.
   6  *
   7  * ExecInsertIndexTuples() is the main entry point.  It's called after
   8  * inserting a tuple to the heap, and it inserts corresponding index tuples
   9  * into all indexes.  At the same time, it enforces any unique and
  10  * exclusion constraints:
  11  *
  12  * Unique Indexes
  13  * --------------
  14  *
  15  * Enforcing a unique constraint is straightforward.  When the index AM
  16  * inserts the tuple to the index, it also checks that there are no
  17  * conflicting tuples in the index already.  It does so atomically, so that
  18  * even if two backends try to insert the same key concurrently, only one
  19  * of them will succeed.  All the logic to ensure atomicity, and to wait
  20  * for in-progress transactions to finish, is handled by the index AM.
  21  *
  22  * If a unique constraint is deferred, we request the index AM to not
  23  * throw an error if a conflict is found.  Instead, we make note that there
  24  * was a conflict and return the list of indexes with conflicts to the
  25  * caller.  The caller must re-check them later, by calling index_insert()
  26  * with the UNIQUE_CHECK_EXISTING option.
  27  *
  28  * Exclusion Constraints
  29  * ---------------------
  30  *
  31  * Exclusion constraints are different from unique indexes in that when the
  32  * tuple is inserted to the index, the index AM does not check for
  33  * duplicate keys at the same time.  After the insertion, we perform a
  34  * separate scan on the index to check for conflicting tuples, and if one
  35  * is found, we throw an error and the transaction is aborted.  If the
  36  * conflicting tuple's inserter or deleter is in-progress, we wait for it
  37  * to finish first.
  38  *
  39  * There is a chance of deadlock, if two backends insert a tuple at the
  40  * same time, and then perform the scan to check for conflicts.  They will
  41  * find each other's tuple, and both try to wait for each other.  The
  42  * deadlock detector will detect that, and abort one of the transactions.
  43  * That's fairly harmless, as one of them was bound to abort with a
  44  * "duplicate key error" anyway, although you get a different error
  45  * message.
  46  *
  47  * If an exclusion constraint is deferred, we still perform the conflict
  48  * checking scan immediately after inserting the index tuple.  But instead
  49  * of throwing an error if a conflict is found, we return that information
  50  * to the caller.  The caller must re-check them later by calling
  51  * check_exclusion_constraint().
  52  *
  53  * Speculative insertion
  54  * ---------------------
  55  *
  56  * Speculative insertion is a two-phase mechanism used to implement
  57  * INSERT ... ON CONFLICT DO UPDATE/NOTHING.  The tuple is first inserted
  58  * to the heap and update the indexes as usual, but if a constraint is
  59  * violated, we can still back out the insertion without aborting the whole
  60  * transaction.  In an INSERT ... ON CONFLICT statement, if a conflict is
  61  * detected, the inserted tuple is backed out and the ON CONFLICT action is
  62  * executed instead.
  63  *
  64  * Insertion to a unique index works as usual: the index AM checks for
  65  * duplicate keys atomically with the insertion.  But instead of throwing
  66  * an error on a conflict, the speculatively inserted heap tuple is backed
  67  * out.
  68  *
  69  * Exclusion constraints are slightly more complicated.  As mentioned
  70  * earlier, there is a risk of deadlock when two backends insert the same
  71  * key concurrently.  That was not a problem for regular insertions, when
  72  * one of the transactions has to be aborted anyway, but with a speculative
  73  * insertion we cannot let a deadlock happen, because we only want to back
  74  * out the speculatively inserted tuple on conflict, not abort the whole
  75  * transaction.
  76  *
  77  * When a backend detects that the speculative insertion conflicts with
  78  * another in-progress tuple, it has two options:
  79  *
  80  * 1. back out the speculatively inserted tuple, then wait for the other
  81  *        transaction, and retry. Or,
  82  * 2. wait for the other transaction, with the speculatively inserted tuple
  83  *        still in place.
  84  *
  85  * If two backends insert at the same time, and both try to wait for each
  86  * other, they will deadlock.  So option 2 is not acceptable.  Option 1
  87  * avoids the deadlock, but it is prone to a livelock instead.  Both
  88  * transactions will wake up immediately as the other transaction backs
  89  * out.  Then they both retry, and conflict with each other again, lather,
  90  * rinse, repeat.
  91  *
  92  * To avoid the livelock, one of the backends must back out first, and then
  93  * wait, while the other one waits without backing out.  It doesn't matter
  94  * which one backs out, so we employ an arbitrary rule that the transaction
  95  * with the higher XID backs out.
  96  *
  97  *
  98  * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
  99  * Portions Copyright (c) 1994, Regents of the University of California
 100  *
 101  *
 102  * IDENTIFICATION
 103  *        src/backend/executor/execIndexing.c
 104  *
 105  *-------------------------------------------------------------------------
 106  */
 107 #include "postgres.h"
 108
 109 #include "access/relscan.h"
 110 #include "access/xact.h"
 111 #include "catalog/index.h"
 112 #include "executor/executor.h"
 113 #include "nodes/nodeFuncs.h"
 114 #include "storage/lmgr.h"
 115 #include "utils/tqual.h"
 116
 117 /* waitMode argument to check_exclusion_or_unique_constraint() */
 118 typedef enum
 119 {
 120         CEOUC_WAIT,
 121         CEOUC_NOWAIT,
 122         CEOUC_LIVELOCK_PREVENTING_WAIT
 123 } CEOUC_WAIT_MODE;
 124
 125 static bool check_exclusion_or_unique_constraint(Relation heap, Relation index,
 126                                                                          IndexInfo *indexInfo,
 127                                                                          ItemPointer tupleid,
 128                                                                          Datum *values, bool *isnull,
 129                                                                          EState *estate, bool newIndex,
 130                                                                          CEOUC_WAIT_MODE waitMode,
 131                                                                          bool errorOK,
 132                                                                          ItemPointer conflictTid);
 133
 134 static bool index_recheck_constraint(Relation index, Oid *constr_procs,
 135                                                  Datum *existing_values, bool *existing_isnull,
 136                                                  Datum *new_values);
 137
 138 /* ----------------------------------------------------------------
 139  *              ExecOpenIndices
 140  *
 141  *              Find the indices associated with a result relation, open them,
 142  *              and save information about them in the result ResultRelInfo.
 143  *
 144  *              At entry, caller has already opened and locked
 145  *              resultRelInfo->ri_RelationDesc.
 146  * ----------------------------------------------------------------
 147  */
 148 void
 149 ExecOpenIndices(ResultRelInfo *resultRelInfo, bool speculative)
 150 {
 151         Relation        resultRelation = resultRelInfo->ri_RelationDesc;
 152         List       *indexoidlist;
 153         ListCell   *l;
 154         int                     len,
 155                                 i;
 156         RelationPtr relationDescs;
 157         IndexInfo **indexInfoArray;
 158
 159         resultRelInfo->ri_NumIndices = 0;
 160
 161         /* fast path if no indexes */
 162         if (!RelationGetForm(resultRelation)->relhasindex)
 163                 return;
 164
 165         /*
 166          * Get cached list of index OIDs
 167          */
 168         indexoidlist = RelationGetIndexList(resultRelation);
 169         len = list_length(indexoidlist);
 170         if (len == 0)
 171                 return;
 172
 173         /*
 174          * allocate space for result arrays
 175          */
 176         relationDescs = (RelationPtr) palloc(len * sizeof(Relation));
 177         indexInfoArray = (IndexInfo **) palloc(len * sizeof(IndexInfo *));
 178
 179         resultRelInfo->ri_NumIndices = len;
 180         resultRelInfo->ri_IndexRelationDescs = relationDescs;
 181         resultRelInfo->ri_IndexRelationInfo = indexInfoArray;
 182
 183         /*
 184          * For each index, open the index relation and save pg_index info. We
 185          * acquire RowExclusiveLock, signifying we will update the index.
 186          *
 187          * Note: we do this even if the index is not indisready; it's not worth
 188          * the trouble to optimize for the case where it isn't.
 189          */
 190         i = 0;
 191         foreach(l, indexoidlist)
 192         {
 193                 Oid                     indexOid = lfirst_oid(l);
 194                 Relation        indexDesc;
 195                 IndexInfo  *ii;
 196
 197                 indexDesc = index_open(indexOid, RowExclusiveLock);
 198
 199                 /* extract index key information from the index's pg_index info */
 200                 ii = BuildIndexInfo(indexDesc);
 201
 202                 /*
 203                  * If the indexes are to be used for speculative insertion, add extra
 204                  * information required by unique index entries.
 205                  */
 206                 if (speculative && ii->ii_Unique)
 207                         BuildSpeculativeIndexInfo(indexDesc, ii);
 208
 209                 relationDescs[i] = indexDesc;
 210                 indexInfoArray[i] = ii;
 211                 i++;
 212         }
 213
 214         list_free(indexoidlist);
 215 }
 216
 217 /* ----------------------------------------------------------------
 218  *              ExecCloseIndices
 219  *
 220  *              Close the index relations stored in resultRelInfo
 221  * ----------------------------------------------------------------
 222  */
 223 void
 224 ExecCloseIndices(ResultRelInfo *resultRelInfo)
 225 {
 226         int                     i;
 227         int                     numIndices;
 228         RelationPtr indexDescs;
 229
 230         numIndices = resultRelInfo->ri_NumIndices;
 231         indexDescs = resultRelInfo->ri_IndexRelationDescs;
 232
 233         for (i = 0; i < numIndices; i++)
 234         {
 235                 if (indexDescs[i] == NULL)
 236                         continue;                       /* shouldn't happen? */
 237
 238                 /* Drop lock acquired by ExecOpenIndices */
 239                 index_close(indexDescs[i], RowExclusiveLock);
 240         }
 241
 242         /*
 243          * XXX should free indexInfo array here too?  Currently we assume that
 244          * such stuff will be cleaned up automatically in FreeExecutorState.
 245          */
 246 }
 247
 248 /* ----------------------------------------------------------------
 249  *              ExecInsertIndexTuples
 250  *
 251  *              This routine takes care of inserting index tuples
 252  *              into all the relations indexing the result relation
 253  *              when a heap tuple is inserted into the result relation.
 254  *
 255  *              Unique and exclusion constraints are enforced at the same
 256  *              time.  This returns a list of index OIDs for any unique or
 257  *              exclusion constraints that are deferred and that had
 258  *              potential (unconfirmed) conflicts.  (if noDupErr == true,
 259  *              the same is done for non-deferred constraints, but report
 260  *              if conflict was speculative or deferred conflict to caller)
 261  *
 262  *              If 'arbiterIndexes' is nonempty, noDupErr applies only to
 263  *              those indexes.  NIL means noDupErr applies to all indexes.
 264  *
 265  *              CAUTION: this must not be called for a HOT update.
 266  *              We can't defend against that here for lack of info.
 267  *              Should we change the API to make it safer?
 268  * ----------------------------------------------------------------
 269  */
 270 List *
 271 ExecInsertIndexTuples(TupleTableSlot *slot,
 272                                           ItemPointer tupleid,
 273                                           EState *estate,
 274                                           bool noDupErr,
 275                                           bool *specConflict,
 276                                           List *arbiterIndexes)
 277 {
 278         List       *result = NIL;
 279         ResultRelInfo *resultRelInfo;
 280         int                     i;
 281         int                     numIndices;
 282         RelationPtr relationDescs;
 283         Relation        heapRelation;
 284         IndexInfo **indexInfoArray;
 285         ExprContext *econtext;
 286         Datum           values[INDEX_MAX_KEYS];
 287         bool            isnull[INDEX_MAX_KEYS];
 288
 289         /*
 290          * Get information from the result relation info structure.
 291          */
 292         resultRelInfo = estate->es_result_relation_info;
 293         numIndices = resultRelInfo->ri_NumIndices;
 294         relationDescs = resultRelInfo->ri_IndexRelationDescs;
 295         indexInfoArray = resultRelInfo->ri_IndexRelationInfo;
 296         heapRelation = resultRelInfo->ri_RelationDesc;
 297
 298         /*
 299          * We will use the EState's per-tuple context for evaluating predicates
 300          * and index expressions (creating it if it's not already there).
 301          */
 302         econtext = GetPerTupleExprContext(estate);
 303
 304         /* Arrange for econtext's scan tuple to be the tuple under test */
 305         econtext->ecxt_scantuple = slot;
 306
 307         /*
 308          * for each index, form and insert the index tuple
 309          */
 310         for (i = 0; i < numIndices; i++)
 311         {
 312                 Relation        indexRelation = relationDescs[i];
 313                 IndexInfo  *indexInfo;
 314                 bool            applyNoDupErr;
 315                 IndexUniqueCheck checkUnique;
 316                 bool            satisfiesConstraint;
 317
 318                 if (indexRelation == NULL)
 319                         continue;
 320
 321                 indexInfo = indexInfoArray[i];
 322
 323                 /* If the index is marked as read-only, ignore it */
 324                 if (!indexInfo->ii_ReadyForInserts)
 325                         continue;
 326
 327                 /* Check for partial index */
 328                 if (indexInfo->ii_Predicate != NIL)
 329                 {
 330                         ExprState  *predicate;
 331
 332                         /*
 333                          * If predicate state not set up yet, create it (in the estate's
 334                          * per-query context)
 335                          */
 336                         predicate = indexInfo->ii_PredicateState;
 337                         if (predicate == NULL)
 338                         {
 339                                 predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
 340                                 indexInfo->ii_PredicateState = predicate;
 341                         }
 342
 343                         /* Skip this index-update if the predicate isn't satisfied */
 344                         if (!ExecQual(predicate, econtext))
 345                                 continue;
 346                 }
 347
 348                 /*
 349                  * FormIndexDatum fills in its values and isnull parameters with the
 350                  * appropriate values for the column(s) of the index.
 351                  */
 352                 FormIndexDatum(indexInfo,
 353                                            slot,
 354                                            estate,
 355                                            values,
 356                                            isnull);
 357
 358                 /* Check whether to apply noDupErr to this index */
 359                 applyNoDupErr = noDupErr &&
 360                         (arbiterIndexes == NIL ||
 361                          list_member_oid(arbiterIndexes,
 362                                                          indexRelation->rd_index->indexrelid));
 363
 364                 /*
 365                  * The index AM does the actual insertion, plus uniqueness checking.
 366                  *
 367                  * For an immediate-mode unique index, we just tell the index AM to
 368                  * throw error if not unique.
 369                  *
 370                  * For a deferrable unique index, we tell the index AM to just detect
 371                  * possible non-uniqueness, and we add the index OID to the result
 372                  * list if further checking is needed.
 373                  *
 374                  * For a speculative insertion (used by INSERT ... ON CONFLICT), do
 375                  * the same as for a deferrable unique index.
 376                  */
 377                 if (!indexRelation->rd_index->indisunique)
 378                         checkUnique = UNIQUE_CHECK_NO;
 379                 else if (applyNoDupErr)
 380                         checkUnique = UNIQUE_CHECK_PARTIAL;
 381                 else if (indexRelation->rd_index->indimmediate)
 382                         checkUnique = UNIQUE_CHECK_YES;
 383                 else
 384                         checkUnique = UNIQUE_CHECK_PARTIAL;
 385
 386                 satisfiesConstraint =
 387                         index_insert(indexRelation, /* index relation */
 388                                                  values,        /* array of index Datums */
 389                                                  isnull,        /* null flags */
 390                                                  tupleid,       /* tid of heap tuple */
 391                                                  heapRelation,  /* heap relation */
 392                                                  checkUnique,   /* type of uniqueness check to do */
 393                                                  indexInfo);    /* index AM may need this */
 394
 395                 /*
 396                  * If the index has an associated exclusion constraint, check that.
 397                  * This is simpler than the process for uniqueness checks since we
 398                  * always insert first and then check.  If the constraint is deferred,
 399                  * we check now anyway, but don't throw error on violation or wait for
 400                  * a conclusive outcome from a concurrent insertion; instead we'll
 401                  * queue a recheck event.  Similarly, noDupErr callers (speculative
 402                  * inserters) will recheck later, and wait for a conclusive outcome
 403                  * then.
 404                  *
 405                  * An index for an exclusion constraint can't also be UNIQUE (not an
 406                  * essential property, we just don't allow it in the grammar), so no
 407                  * need to preserve the prior state of satisfiesConstraint.
 408                  */
 409                 if (indexInfo->ii_ExclusionOps != NULL)
 410                 {
 411                         bool            violationOK;
 412                         CEOUC_WAIT_MODE waitMode;
 413
 414                         if (applyNoDupErr)
 415                         {
 416                                 violationOK = true;
 417                                 waitMode = CEOUC_LIVELOCK_PREVENTING_WAIT;
 418                         }
 419                         else if (!indexRelation->rd_index->indimmediate)
 420                         {
 421                                 violationOK = true;
 422                                 waitMode = CEOUC_NOWAIT;
 423                         }
 424                         else
 425                         {
 426                                 violationOK = false;
 427                                 waitMode = CEOUC_WAIT;
 428                         }
 429
 430                         satisfiesConstraint =
 431                                 check_exclusion_or_unique_constraint(heapRelation,
 432                                                                                                          indexRelation, indexInfo,
 433                                                                                                          tupleid, values, isnull,
 434                                                                                                          estate, false,
 435                                                                                                          waitMode, violationOK, NULL);
 436                 }
 437
 438                 if ((checkUnique == UNIQUE_CHECK_PARTIAL ||
 439                          indexInfo->ii_ExclusionOps != NULL) &&
 440                         !satisfiesConstraint)
 441                 {
 442                         /*
 443                          * The tuple potentially violates the uniqueness or exclusion
 444                          * constraint, so make a note of the index so that we can re-check
 445                          * it later.  Speculative inserters are told if there was a
 446                          * speculative conflict, since that always requires a restart.
 447                          */
 448                         result = lappend_oid(result, RelationGetRelid(indexRelation));
 449                         if (indexRelation->rd_index->indimmediate && specConflict)
 450                                 *specConflict = true;
 451                 }
 452         }
 453
 454         return result;
 455 }
 456
 457 /* ----------------------------------------------------------------
 458  *              ExecCheckIndexConstraints
 459  *
 460  *              This routine checks if a tuple violates any unique or
 461  *              exclusion constraints.  Returns true if there is no conflict.
 462  *              Otherwise returns false, and the TID of the conflicting
 463  *              tuple is returned in *conflictTid.
 464  *
 465  *              If 'arbiterIndexes' is given, only those indexes are checked.
 466  *              NIL means all indexes.
 467  *
 468  *              Note that this doesn't lock the values in any way, so it's
 469  *              possible that a conflicting tuple is inserted immediately
 470  *              after this returns.  But this can be used for a pre-check
 471  *              before insertion.
 472  * ----------------------------------------------------------------
 473  */
 474 bool
 475 ExecCheckIndexConstraints(TupleTableSlot *slot,
 476                                                   EState *estate, ItemPointer conflictTid,
 477                                                   List *arbiterIndexes)
 478 {
 479         ResultRelInfo *resultRelInfo;
 480         int                     i;
 481         int                     numIndices;
 482         RelationPtr relationDescs;
 483         Relation        heapRelation;
 484         IndexInfo **indexInfoArray;
 485         ExprContext *econtext;
 486         Datum           values[INDEX_MAX_KEYS];
 487         bool            isnull[INDEX_MAX_KEYS];
 488         ItemPointerData invalidItemPtr;
 489         bool            checkedIndex = false;
 490
 491         ItemPointerSetInvalid(conflictTid);
 492         ItemPointerSetInvalid(&invalidItemPtr);
 493
 494         /*
 495          * Get information from the result relation info structure.
 496          */
 497         resultRelInfo = estate->es_result_relation_info;
 498         numIndices = resultRelInfo->ri_NumIndices;
 499         relationDescs = resultRelInfo->ri_IndexRelationDescs;
 500         indexInfoArray = resultRelInfo->ri_IndexRelationInfo;
 501         heapRelation = resultRelInfo->ri_RelationDesc;
 502
 503         /*
 504          * We will use the EState's per-tuple context for evaluating predicates
 505          * and index expressions (creating it if it's not already there).
 506          */
 507         econtext = GetPerTupleExprContext(estate);
 508
 509         /* Arrange for econtext's scan tuple to be the tuple under test */
 510         econtext->ecxt_scantuple = slot;
 511
 512         /*
 513          * For each index, form index tuple and check if it satisfies the
 514          * constraint.
 515          */
 516         for (i = 0; i < numIndices; i++)
 517         {
 518                 Relation        indexRelation = relationDescs[i];
 519                 IndexInfo  *indexInfo;
 520                 bool            satisfiesConstraint;
 521
 522                 if (indexRelation == NULL)
 523                         continue;
 524
 525                 indexInfo = indexInfoArray[i];
 526
 527                 if (!indexInfo->ii_Unique && !indexInfo->ii_ExclusionOps)
 528                         continue;
 529
 530                 /* If the index is marked as read-only, ignore it */
 531                 if (!indexInfo->ii_ReadyForInserts)
 532                         continue;
 533
 534                 /* When specific arbiter indexes requested, only examine them */
 535                 if (arbiterIndexes != NIL &&
 536                         !list_member_oid(arbiterIndexes,
 537                                                          indexRelation->rd_index->indexrelid))
 538                         continue;
 539
 540                 if (!indexRelation->rd_index->indimmediate)
 541                         ereport(ERROR,
 542                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 543                                          errmsg("ON CONFLICT does not support deferrable unique constraints/exclusion constraints as arbiters"),
 544                                          errtableconstraint(heapRelation,
 545                                                                                 RelationGetRelationName(indexRelation))));
 546
 547                 checkedIndex = true;
 548
 549                 /* Check for partial index */
 550                 if (indexInfo->ii_Predicate != NIL)
 551                 {
 552                         ExprState  *predicate;
 553
 554                         /*
 555                          * If predicate state not set up yet, create it (in the estate's
 556                          * per-query context)
 557                          */
 558                         predicate = indexInfo->ii_PredicateState;
 559                         if (predicate == NULL)
 560                         {
 561                                 predicate = ExecPrepareQual(indexInfo->ii_Predicate, estate);
 562                                 indexInfo->ii_PredicateState = predicate;
 563                         }
 564
 565                         /* Skip this index-update if the predicate isn't satisfied */
 566                         if (!ExecQual(predicate, econtext))
 567                                 continue;
 568                 }
 569
 570                 /*
 571                  * FormIndexDatum fills in its values and isnull parameters with the
 572                  * appropriate values for the column(s) of the index.
 573                  */
 574                 FormIndexDatum(indexInfo,
 575                                            slot,
 576                                            estate,
 577                                            values,
 578                                            isnull);
 579
 580                 satisfiesConstraint =
 581                         check_exclusion_or_unique_constraint(heapRelation, indexRelation,
 582                                                                                                  indexInfo, &invalidItemPtr,
 583                                                                                                  values, isnull, estate, false,
 584                                                                                                  CEOUC_WAIT, true,
 585                                                                                                  conflictTid);
 586                 if (!satisfiesConstraint)
 587                         return false;
 588         }
 589
 590         if (arbiterIndexes != NIL && !checkedIndex)
 591                 elog(ERROR, "unexpected failure to find arbiter index");
 592
 593         return true;
 594 }
 595
 596 /*
 597  * Check for violation of an exclusion or unique constraint
 598  *
 599  * heap: the table containing the new tuple
 600  * index: the index supporting the constraint
 601  * indexInfo: info about the index, including the exclusion properties
 602  * tupleid: heap TID of the new tuple we have just inserted (invalid if we
 603  *              haven't inserted a new tuple yet)
 604  * values, isnull: the *index* column values computed for the new tuple
 605  * estate: an EState we can do evaluation in
 606  * newIndex: if true, we are trying to build a new index (this affects
 607  *              only the wording of error messages)
 608  * waitMode: whether to wait for concurrent inserters/deleters
 609  * violationOK: if true, don't throw error for violation
 610  * conflictTid: if not-NULL, the TID of the conflicting tuple is returned here
 611  *
 612  * Returns true if OK, false if actual or potential violation
 613  *
 614  * 'waitMode' determines what happens if a conflict is detected with a tuple
 615  * that was inserted or deleted by a transaction that's still running.
 616  * CEOUC_WAIT means that we wait for the transaction to commit, before
 617  * throwing an error or returning.  CEOUC_NOWAIT means that we report the
 618  * violation immediately; so the violation is only potential, and the caller
 619  * must recheck sometime later.  This behavior is convenient for deferred
 620  * exclusion checks; we need not bother queuing a deferred event if there is
 621  * definitely no conflict at insertion time.
 622  *
 623  * CEOUC_LIVELOCK_PREVENTING_WAIT is like CEOUC_NOWAIT, but we will sometimes
 624  * wait anyway, to prevent livelocking if two transactions try inserting at
 625  * the same time.  This is used with speculative insertions, for INSERT ON
 626  * CONFLICT statements. (See notes in file header)
 627  *
 628  * If violationOK is true, we just report the potential or actual violation to
 629  * the caller by returning 'false'.  Otherwise we throw a descriptive error
 630  * message here.  When violationOK is false, a false result is impossible.
 631  *
 632  * Note: The indexam is normally responsible for checking unique constraints,
 633  * so this normally only needs to be used for exclusion constraints.  But this
 634  * function is also called when doing a "pre-check" for conflicts on a unique
 635  * constraint, when doing speculative insertion.  Caller may use the returned
 636  * conflict TID to take further steps.
 637  */
 638 static bool
 639 check_exclusion_or_unique_constraint(Relation heap, Relation index,
 640                                                                          IndexInfo *indexInfo,
 641                                                                          ItemPointer tupleid,
 642                                                                          Datum *values, bool *isnull,
 643                                                                          EState *estate, bool newIndex,
 644                                                                          CEOUC_WAIT_MODE waitMode,
 645                                                                          bool violationOK,
 646                                                                          ItemPointer conflictTid)
 647 {
 648         Oid                *constr_procs;
 649         uint16     *constr_strats;
 650         Oid                *index_collations = index->rd_indcollation;
 651         int                     indnkeyatts = IndexRelationGetNumberOfKeyAttributes(index);
 652         IndexScanDesc index_scan;
 653         HeapTuple       tup;
 654         ScanKeyData scankeys[INDEX_MAX_KEYS];
 655         SnapshotData DirtySnapshot;
 656         int                     i;
 657         bool            conflict;
 658         bool            found_self;
 659         ExprContext *econtext;
 660         TupleTableSlot *existing_slot;
 661         TupleTableSlot *save_scantuple;
 662
 663         if (indexInfo->ii_ExclusionOps)
 664         {
 665                 constr_procs = indexInfo->ii_ExclusionProcs;
 666                 constr_strats = indexInfo->ii_ExclusionStrats;
 667         }
 668         else
 669         {
 670                 constr_procs = indexInfo->ii_UniqueProcs;
 671                 constr_strats = indexInfo->ii_UniqueStrats;
 672         }
 673
 674         /*
 675          * If any of the input values are NULL, the constraint check is assumed to
 676          * pass (i.e., we assume the operators are strict).
 677          */
 678         for (i = 0; i < indnkeyatts; i++)
 679         {
 680                 if (isnull[i])
 681                         return true;
 682         }
 683
 684         /*
 685          * Search the tuples that are in the index for any violations, including
 686          * tuples that aren't visible yet.
 687          */
 688         InitDirtySnapshot(DirtySnapshot);
 689
 690         for (i = 0; i < indnkeyatts; i++)
 691         {
 692                 ScanKeyEntryInitialize(&scankeys[i],
 693                                                            0,
 694                                                            i + 1,
 695                                                            constr_strats[i],
 696                                                            InvalidOid,
 697                                                            index_collations[i],
 698                                                            constr_procs[i],
 699                                                            values[i]);
 700         }
 701
 702         /*
 703          * Need a TupleTableSlot to put existing tuples in.
 704          *
 705          * To use FormIndexDatum, we have to make the econtext's scantuple point
 706          * to this slot.  Be sure to save and restore caller's value for
 707          * scantuple.
 708          */
 709         existing_slot = MakeSingleTupleTableSlot(RelationGetDescr(heap),
 710                                                                                          &TTSOpsHeapTuple);
 711
 712         econtext = GetPerTupleExprContext(estate);
 713         save_scantuple = econtext->ecxt_scantuple;
 714         econtext->ecxt_scantuple = existing_slot;
 715
 716         /*
 717          * May have to restart scan from this point if a potential conflict is
 718          * found.
 719          */
 720 retry:
 721         conflict = false;
 722         found_self = false;
 723         index_scan = index_beginscan(heap, index, &DirtySnapshot, indnkeyatts, 0);
 724         index_rescan(index_scan, scankeys, indnkeyatts, NULL, 0);
 725
 726         while ((tup = index_getnext(index_scan,
 727                                                                 ForwardScanDirection)) != NULL)
 728         {
 729                 TransactionId xwait;
 730                 ItemPointerData ctid_wait;
 731                 XLTW_Oper       reason_wait;
 732                 Datum           existing_values[INDEX_MAX_KEYS];
 733                 bool            existing_isnull[INDEX_MAX_KEYS];
 734                 char       *error_new;
 735                 char       *error_existing;
 736
 737                 /*
 738                  * Ignore the entry for the tuple we're trying to check.
 739                  */
 740                 if (ItemPointerIsValid(tupleid) &&
 741                         ItemPointerEquals(tupleid, &tup->t_self))
 742                 {
 743                         if (found_self)         /* should not happen */
 744                                 elog(ERROR, "found self tuple multiple times in index \"%s\"",
 745                                          RelationGetRelationName(index));
 746                         found_self = true;
 747                         continue;
 748                 }
 749
 750                 /*
 751                  * Extract the index column values and isnull flags from the existing
 752                  * tuple.
 753                  */
 754                 ExecStoreHeapTuple(tup, existing_slot, false);
 755                 FormIndexDatum(indexInfo, existing_slot, estate,
 756                                            existing_values, existing_isnull);
 757
 758                 /* If lossy indexscan, must recheck the condition */
 759                 if (index_scan->xs_recheck)
 760                 {
 761                         if (!index_recheck_constraint(index,
 762                                                                                   constr_procs,
 763                                                                                   existing_values,
 764                                                                                   existing_isnull,
 765                                                                                   values))
 766                                 continue;               /* tuple doesn't actually match, so no
 767                                                                  * conflict */
 768                 }
 769
 770                 /*
 771                  * At this point we have either a conflict or a potential conflict.
 772                  *
 773                  * If an in-progress transaction is affecting the visibility of this
 774                  * tuple, we need to wait for it to complete and then recheck (unless
 775                  * the caller requested not to).  For simplicity we do rechecking by
 776                  * just restarting the whole scan --- this case probably doesn't
 777                  * happen often enough to be worth trying harder, and anyway we don't
 778                  * want to hold any index internal locks while waiting.
 779                  */
 780                 xwait = TransactionIdIsValid(DirtySnapshot.xmin) ?
 781                         DirtySnapshot.xmin : DirtySnapshot.xmax;
 782
 783                 if (TransactionIdIsValid(xwait) &&
 784                         (waitMode == CEOUC_WAIT ||
 785                          (waitMode == CEOUC_LIVELOCK_PREVENTING_WAIT &&
 786                           DirtySnapshot.speculativeToken &&
 787                           TransactionIdPrecedes(GetCurrentTransactionId(), xwait))))
 788                 {
 789                         ctid_wait = tup->t_data->t_ctid;
 790                         reason_wait = indexInfo->ii_ExclusionOps ?
 791                                 XLTW_RecheckExclusionConstr : XLTW_InsertIndex;
 792                         index_endscan(index_scan);
 793                         if (DirtySnapshot.speculativeToken)
 794                                 SpeculativeInsertionWait(DirtySnapshot.xmin,
 795                                                                                  DirtySnapshot.speculativeToken);
 796                         else
 797                                 XactLockTableWait(xwait, heap, &ctid_wait, reason_wait);
 798                         goto retry;
 799                 }
 800
 801                 /*
 802                  * We have a definite conflict (or a potential one, but the caller
 803                  * didn't want to wait).  Return it to caller, or report it.
 804                  */
 805                 if (violationOK)
 806                 {
 807                         conflict = true;
 808                         if (conflictTid)
 809                                 *conflictTid = tup->t_self;
 810                         break;
 811                 }
 812
 813                 error_new = BuildIndexValueDescription(index, values, isnull);
 814                 error_existing = BuildIndexValueDescription(index, existing_values,
 815                                                                                                         existing_isnull);
 816                 if (newIndex)
 817                         ereport(ERROR,
 818                                         (errcode(ERRCODE_EXCLUSION_VIOLATION),
 819                                          errmsg("could not create exclusion constraint \"%s\"",
 820                                                         RelationGetRelationName(index)),
 821                                          error_new && error_existing ?
 822                                          errdetail("Key %s conflicts with key %s.",
 823                                                            error_new, error_existing) :
 824                                          errdetail("Key conflicts exist."),
 825                                          errtableconstraint(heap,
 826                                                                                 RelationGetRelationName(index))));
 827                 else
 828                         ereport(ERROR,
 829                                         (errcode(ERRCODE_EXCLUSION_VIOLATION),
 830                                          errmsg("conflicting key value violates exclusion constraint \"%s\"",
 831                                                         RelationGetRelationName(index)),
 832                                          error_new && error_existing ?
 833                                          errdetail("Key %s conflicts with existing key %s.",
 834                                                            error_new, error_existing) :
 835                                          errdetail("Key conflicts with existing key."),
 836                                          errtableconstraint(heap,
 837                                                                                 RelationGetRelationName(index))));
 838         }
 839
 840         index_endscan(index_scan);
 841
 842         /*
 843          * Ordinarily, at this point the search should have found the originally
 844          * inserted tuple (if any), unless we exited the loop early because of
 845          * conflict.  However, it is possible to define exclusion constraints for
 846          * which that wouldn't be true --- for instance, if the operator is <>. So
 847          * we no longer complain if found_self is still false.
 848          */
 849
 850         econtext->ecxt_scantuple = save_scantuple;
 851
 852         ExecDropSingleTupleTableSlot(existing_slot);
 853
 854         return !conflict;
 855 }
 856
 857 /*
 858  * Check for violation of an exclusion constraint
 859  *
 860  * This is a dumbed down version of check_exclusion_or_unique_constraint
 861  * for external callers. They don't need all the special modes.
 862  */
 863 void
 864 check_exclusion_constraint(Relation heap, Relation index,
 865                                                    IndexInfo *indexInfo,
 866                                                    ItemPointer tupleid,
 867                                                    Datum *values, bool *isnull,
 868                                                    EState *estate, bool newIndex)
 869 {
 870         (void) check_exclusion_or_unique_constraint(heap, index, indexInfo, tupleid,
 871                                                                                                 values, isnull,
 872                                                                                                 estate, newIndex,
 873                                                                                                 CEOUC_WAIT, false, NULL);
 874 }
 875
 876 /*
 877  * Check existing tuple's index values to see if it really matches the
 878  * exclusion condition against the new_values.  Returns true if conflict.
 879  */
 880 static bool
 881 index_recheck_constraint(Relation index, Oid *constr_procs,
 882                                                  Datum *existing_values, bool *existing_isnull,
 883                                                  Datum *new_values)
 884 {
 885         int                     indnkeyatts = IndexRelationGetNumberOfKeyAttributes(index);
 886         int                     i;
 887
 888         for (i = 0; i < indnkeyatts; i++)
 889         {
 890                 /* Assume the exclusion operators are strict */
 891                 if (existing_isnull[i])
 892                         return false;
 893
 894                 if (!DatumGetBool(OidFunctionCall2Coll(constr_procs[i],
 895                                                                                            index->rd_indcollation[i],
 896                                                                                            existing_values[i],
 897                                                                                            new_values[i])))
 898                         return false;
 899         }
 900
 901         return true;
 902 }