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) 1996-2017, 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 IndexIsReady; 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                         List       *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 == NIL)
 338                         {
 339                                 predicate = (List *)
 340                                         ExecPrepareExpr((Expr *) indexInfo->ii_Predicate,
 341                                                                         estate);
 342                                 indexInfo->ii_PredicateState = predicate;
 343                         }
 344
 345                         /* Skip this index-update if the predicate isn't satisfied */
 346                         if (!ExecQual(predicate, econtext, false))
 347                                 continue;
 348                 }
 349
 350                 /*
 351                  * FormIndexDatum fills in its values and isnull parameters with the
 352                  * appropriate values for the column(s) of the index.
 353                  */
 354                 FormIndexDatum(indexInfo,
 355                                            slot,
 356                                            estate,
 357                                            values,
 358                                            isnull);
 359
 360                 /* Check whether to apply noDupErr to this index */
 361                 applyNoDupErr = noDupErr &&
 362                         (arbiterIndexes == NIL ||
 363                          list_member_oid(arbiterIndexes,
 364                                                          indexRelation->rd_index->indexrelid));
 365
 366                 /*
 367                  * The index AM does the actual insertion, plus uniqueness checking.
 368                  *
 369                  * For an immediate-mode unique index, we just tell the index AM to
 370                  * throw error if not unique.
 371                  *
 372                  * For a deferrable unique index, we tell the index AM to just detect
 373                  * possible non-uniqueness, and we add the index OID to the result
 374                  * list if further checking is needed.
 375                  *
 376                  * For a speculative insertion (used by INSERT ... ON CONFLICT), do
 377                  * the same as for a deferrable unique index.
 378                  */
 379                 if (!indexRelation->rd_index->indisunique)
 380                         checkUnique = UNIQUE_CHECK_NO;
 381                 else if (applyNoDupErr)
 382                         checkUnique = UNIQUE_CHECK_PARTIAL;
 383                 else if (indexRelation->rd_index->indimmediate)
 384                         checkUnique = UNIQUE_CHECK_YES;
 385                 else
 386                         checkUnique = UNIQUE_CHECK_PARTIAL;
 387
 388                 satisfiesConstraint =
 389                         index_insert(indexRelation, /* index relation */
 390                                                  values,        /* array of index Datums */
 391                                                  isnull,        /* null flags */
 392                                                  tupleid,               /* tid of heap tuple */
 393                                                  heapRelation,  /* heap relation */
 394                                                  checkUnique);  /* type of uniqueness check to do */
 395
 396                 /*
 397                  * If the index has an associated exclusion constraint, check that.
 398                  * This is simpler than the process for uniqueness checks since we
 399                  * always insert first and then check.  If the constraint is deferred,
 400                  * we check now anyway, but don't throw error on violation or wait for
 401                  * a conclusive outcome from a concurrent insertion; instead we'll
 402                  * queue a recheck event.  Similarly, noDupErr callers (speculative
 403                  * inserters) will recheck later, and wait for a conclusive outcome
 404                  * then.
 405                  *
 406                  * An index for an exclusion constraint can't also be UNIQUE (not an
 407                  * essential property, we just don't allow it in the grammar), so no
 408                  * need to preserve the prior state of satisfiesConstraint.
 409                  */
 410                 if (indexInfo->ii_ExclusionOps != NULL)
 411                 {
 412                         bool            violationOK;
 413                         CEOUC_WAIT_MODE waitMode;
 414
 415                         if (applyNoDupErr)
 416                         {
 417                                 violationOK = true;
 418                                 waitMode = CEOUC_LIVELOCK_PREVENTING_WAIT;
 419                         }
 420                         else if (!indexRelation->rd_index->indimmediate)
 421                         {
 422                                 violationOK = true;
 423                                 waitMode = CEOUC_NOWAIT;
 424                         }
 425                         else
 426                         {
 427                                 violationOK = false;
 428                                 waitMode = CEOUC_WAIT;
 429                         }
 430
 431                         satisfiesConstraint =
 432                                 check_exclusion_or_unique_constraint(heapRelation,
 433                                                                                                          indexRelation, indexInfo,
 434                                                                                                          tupleid, values, isnull,
 435                                                                                                          estate, false,
 436                                                                                                 waitMode, violationOK, NULL);
 437                 }
 438
 439                 if ((checkUnique == UNIQUE_CHECK_PARTIAL ||
 440                          indexInfo->ii_ExclusionOps != NULL) &&
 441                         !satisfiesConstraint)
 442                 {
 443                         /*
 444                          * The tuple potentially violates the uniqueness or exclusion
 445                          * constraint, so make a note of the index so that we can re-check
 446                          * it later.  Speculative inserters are told if there was a
 447                          * speculative conflict, since that always requires a restart.
 448                          */
 449                         result = lappend_oid(result, RelationGetRelid(indexRelation));
 450                         if (indexRelation->rd_index->indimmediate && specConflict)
 451                                 *specConflict = true;
 452                 }
 453         }
 454
 455         return result;
 456 }
 457
 458 /* ----------------------------------------------------------------
 459  *              ExecCheckIndexConstraints
 460  *
 461  *              This routine checks if a tuple violates any unique or
 462  *              exclusion constraints.  Returns true if there is no conflict.
 463  *              Otherwise returns false, and the TID of the conflicting
 464  *              tuple is returned in *conflictTid.
 465  *
 466  *              If 'arbiterIndexes' is given, only those indexes are checked.
 467  *              NIL means all indexes.
 468  *
 469  *              Note that this doesn't lock the values in any way, so it's
 470  *              possible that a conflicting tuple is inserted immediately
 471  *              after this returns.  But this can be used for a pre-check
 472  *              before insertion.
 473  * ----------------------------------------------------------------
 474  */
 475 bool
 476 ExecCheckIndexConstraints(TupleTableSlot *slot,
 477                                                   EState *estate, ItemPointer conflictTid,
 478                                                   List *arbiterIndexes)
 479 {
 480         ResultRelInfo *resultRelInfo;
 481         int                     i;
 482         int                     numIndices;
 483         RelationPtr relationDescs;
 484         Relation        heapRelation;
 485         IndexInfo **indexInfoArray;
 486         ExprContext *econtext;
 487         Datum           values[INDEX_MAX_KEYS];
 488         bool            isnull[INDEX_MAX_KEYS];
 489         ItemPointerData invalidItemPtr;
 490         bool            checkedIndex = false;
 491
 492         ItemPointerSetInvalid(conflictTid);
 493         ItemPointerSetInvalid(&invalidItemPtr);
 494
 495         /*
 496          * Get information from the result relation info structure.
 497          */
 498         resultRelInfo = estate->es_result_relation_info;
 499         numIndices = resultRelInfo->ri_NumIndices;
 500         relationDescs = resultRelInfo->ri_IndexRelationDescs;
 501         indexInfoArray = resultRelInfo->ri_IndexRelationInfo;
 502         heapRelation = resultRelInfo->ri_RelationDesc;
 503
 504         /*
 505          * We will use the EState's per-tuple context for evaluating predicates
 506          * and index expressions (creating it if it's not already there).
 507          */
 508         econtext = GetPerTupleExprContext(estate);
 509
 510         /* Arrange for econtext's scan tuple to be the tuple under test */
 511         econtext->ecxt_scantuple = slot;
 512
 513         /*
 514          * For each index, form index tuple and check if it satisfies the
 515          * constraint.
 516          */
 517         for (i = 0; i < numIndices; i++)
 518         {
 519                 Relation        indexRelation = relationDescs[i];
 520                 IndexInfo  *indexInfo;
 521                 bool            satisfiesConstraint;
 522
 523                 if (indexRelation == NULL)
 524                         continue;
 525
 526                 indexInfo = indexInfoArray[i];
 527
 528                 if (!indexInfo->ii_Unique && !indexInfo->ii_ExclusionOps)
 529                         continue;
 530
 531                 /* If the index is marked as read-only, ignore it */
 532                 if (!indexInfo->ii_ReadyForInserts)
 533                         continue;
 534
 535                 /* When specific arbiter indexes requested, only examine them */
 536                 if (arbiterIndexes != NIL &&
 537                         !list_member_oid(arbiterIndexes,
 538                                                          indexRelation->rd_index->indexrelid))
 539                         continue;
 540
 541                 if (!indexRelation->rd_index->indimmediate)
 542                         ereport(ERROR,
 543                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 544                                          errmsg("ON CONFLICT does not support deferrable unique constraints/exclusion constraints as arbiters"),
 545                                          errtableconstraint(heapRelation,
 546                                                                    RelationGetRelationName(indexRelation))));
 547
 548                 checkedIndex = true;
 549
 550                 /* Check for partial index */
 551                 if (indexInfo->ii_Predicate != NIL)
 552                 {
 553                         List       *predicate;
 554
 555                         /*
 556                          * If predicate state not set up yet, create it (in the estate's
 557                          * per-query context)
 558                          */
 559                         predicate = indexInfo->ii_PredicateState;
 560                         if (predicate == NIL)
 561                         {
 562                                 predicate = (List *)
 563                                         ExecPrepareExpr((Expr *) indexInfo->ii_Predicate,
 564                                                                         estate);
 565                                 indexInfo->ii_PredicateState = predicate;
 566                         }
 567
 568                         /* Skip this index-update if the predicate isn't satisfied */
 569                         if (!ExecQual(predicate, econtext, false))
 570                                 continue;
 571                 }
 572
 573                 /*
 574                  * FormIndexDatum fills in its values and isnull parameters with the
 575                  * appropriate values for the column(s) of the index.
 576                  */
 577                 FormIndexDatum(indexInfo,
 578                                            slot,
 579                                            estate,
 580                                            values,
 581                                            isnull);
 582
 583                 satisfiesConstraint =
 584                         check_exclusion_or_unique_constraint(heapRelation, indexRelation,
 585                                                                                                  indexInfo, &invalidItemPtr,
 586                                                                                            values, isnull, estate, false,
 587                                                                                                  CEOUC_WAIT, true,
 588                                                                                                  conflictTid);
 589                 if (!satisfiesConstraint)
 590                         return false;
 591         }
 592
 593         if (arbiterIndexes != NIL && !checkedIndex)
 594                 elog(ERROR, "unexpected failure to find arbiter index");
 595
 596         return true;
 597 }
 598
 599 /*
 600  * Check for violation of an exclusion or unique constraint
 601  *
 602  * heap: the table containing the new tuple
 603  * index: the index supporting the constraint
 604  * indexInfo: info about the index, including the exclusion properties
 605  * tupleid: heap TID of the new tuple we have just inserted (invalid if we
 606  *              haven't inserted a new tuple yet)
 607  * values, isnull: the *index* column values computed for the new tuple
 608  * estate: an EState we can do evaluation in
 609  * newIndex: if true, we are trying to build a new index (this affects
 610  *              only the wording of error messages)
 611  * waitMode: whether to wait for concurrent inserters/deleters
 612  * violationOK: if true, don't throw error for violation
 613  * conflictTid: if not-NULL, the TID of the conflicting tuple is returned here
 614  *
 615  * Returns true if OK, false if actual or potential violation
 616  *
 617  * 'waitMode' determines what happens if a conflict is detected with a tuple
 618  * that was inserted or deleted by a transaction that's still running.
 619  * CEOUC_WAIT means that we wait for the transaction to commit, before
 620  * throwing an error or returning.  CEOUC_NOWAIT means that we report the
 621  * violation immediately; so the violation is only potential, and the caller
 622  * must recheck sometime later.  This behavior is convenient for deferred
 623  * exclusion checks; we need not bother queuing a deferred event if there is
 624  * definitely no conflict at insertion time.
 625  *
 626  * CEOUC_LIVELOCK_PREVENTING_WAIT is like CEOUC_NOWAIT, but we will sometimes
 627  * wait anyway, to prevent livelocking if two transactions try inserting at
 628  * the same time.  This is used with speculative insertions, for INSERT ON
 629  * CONFLICT statements. (See notes in file header)
 630  *
 631  * If violationOK is true, we just report the potential or actual violation to
 632  * the caller by returning 'false'.  Otherwise we throw a descriptive error
 633  * message here.  When violationOK is false, a false result is impossible.
 634  *
 635  * Note: The indexam is normally responsible for checking unique constraints,
 636  * so this normally only needs to be used for exclusion constraints.  But this
 637  * function is also called when doing a "pre-check" for conflicts on a unique
 638  * constraint, when doing speculative insertion.  Caller may use the returned
 639  * conflict TID to take further steps.
 640  */
 641 static bool
 642 check_exclusion_or_unique_constraint(Relation heap, Relation index,
 643                                                                          IndexInfo *indexInfo,
 644                                                                          ItemPointer tupleid,
 645                                                                          Datum *values, bool *isnull,
 646                                                                          EState *estate, bool newIndex,
 647                                                                          CEOUC_WAIT_MODE waitMode,
 648                                                                          bool violationOK,
 649                                                                          ItemPointer conflictTid)
 650 {
 651         Oid                *constr_procs;
 652         uint16     *constr_strats;
 653         Oid                *index_collations = index->rd_indcollation;
 654         int                     index_natts = index->rd_index->indnatts;
 655         IndexScanDesc index_scan;
 656         HeapTuple       tup;
 657         ScanKeyData scankeys[INDEX_MAX_KEYS];
 658         SnapshotData DirtySnapshot;
 659         int                     i;
 660         bool            conflict;
 661         bool            found_self;
 662         ExprContext *econtext;
 663         TupleTableSlot *existing_slot;
 664         TupleTableSlot *save_scantuple;
 665
 666         if (indexInfo->ii_ExclusionOps)
 667         {
 668                 constr_procs = indexInfo->ii_ExclusionProcs;
 669                 constr_strats = indexInfo->ii_ExclusionStrats;
 670         }
 671         else
 672         {
 673                 constr_procs = indexInfo->ii_UniqueProcs;
 674                 constr_strats = indexInfo->ii_UniqueStrats;
 675         }
 676
 677         /*
 678          * If any of the input values are NULL, the constraint check is assumed to
 679          * pass (i.e., we assume the operators are strict).
 680          */
 681         for (i = 0; i < index_natts; i++)
 682         {
 683                 if (isnull[i])
 684                         return true;
 685         }
 686
 687         /*
 688          * Search the tuples that are in the index for any violations, including
 689          * tuples that aren't visible yet.
 690          */
 691         InitDirtySnapshot(DirtySnapshot);
 692
 693         for (i = 0; i < index_natts; i++)
 694         {
 695                 ScanKeyEntryInitialize(&scankeys[i],
 696                                                            0,
 697                                                            i + 1,
 698                                                            constr_strats[i],
 699                                                            InvalidOid,
 700                                                            index_collations[i],
 701                                                            constr_procs[i],
 702                                                            values[i]);
 703         }
 704
 705         /*
 706          * Need a TupleTableSlot to put existing tuples in.
 707          *
 708          * To use FormIndexDatum, we have to make the econtext's scantuple point
 709          * to this slot.  Be sure to save and restore caller's value for
 710          * scantuple.
 711          */
 712         existing_slot = MakeSingleTupleTableSlot(RelationGetDescr(heap));
 713
 714         econtext = GetPerTupleExprContext(estate);
 715         save_scantuple = econtext->ecxt_scantuple;
 716         econtext->ecxt_scantuple = existing_slot;
 717
 718         /*
 719          * May have to restart scan from this point if a potential conflict is
 720          * found.
 721          */
 722 retry:
 723         conflict = false;
 724         found_self = false;
 725         index_scan = index_beginscan(heap, index, &DirtySnapshot, index_natts, 0);
 726         index_rescan(index_scan, scankeys, index_natts, NULL, 0);
 727
 728         while ((tup = index_getnext(index_scan,
 729                                                                 ForwardScanDirection)) != NULL)
 730         {
 731                 TransactionId xwait;
 732                 ItemPointerData ctid_wait;
 733                 XLTW_Oper       reason_wait;
 734                 Datum           existing_values[INDEX_MAX_KEYS];
 735                 bool            existing_isnull[INDEX_MAX_KEYS];
 736                 char       *error_new;
 737                 char       *error_existing;
 738
 739                 /*
 740                  * Ignore the entry for the tuple we're trying to check.
 741                  */
 742                 if (ItemPointerIsValid(tupleid) &&
 743                         ItemPointerEquals(tupleid, &tup->t_self))
 744                 {
 745                         if (found_self)         /* should not happen */
 746                                 elog(ERROR, "found self tuple multiple times in index \"%s\"",
 747                                          RelationGetRelationName(index));
 748                         found_self = true;
 749                         continue;
 750                 }
 751
 752                 /*
 753                  * Extract the index column values and isnull flags from the existing
 754                  * tuple.
 755                  */
 756                 ExecStoreTuple(tup, existing_slot, InvalidBuffer, false);
 757                 FormIndexDatum(indexInfo, existing_slot, estate,
 758                                            existing_values, existing_isnull);
 759
 760                 /* If lossy indexscan, must recheck the condition */
 761                 if (index_scan->xs_recheck)
 762                 {
 763                         if (!index_recheck_constraint(index,
 764                                                                                   constr_procs,
 765                                                                                   existing_values,
 766                                                                                   existing_isnull,
 767                                                                                   values))
 768                                 continue;               /* tuple doesn't actually match, so no
 769                                                                  * conflict */
 770                 }
 771
 772                 /*
 773                  * At this point we have either a conflict or a potential conflict.
 774                  *
 775                  * If an in-progress transaction is affecting the visibility of this
 776                  * tuple, we need to wait for it to complete and then recheck (unless
 777                  * the caller requested not to).  For simplicity we do rechecking by
 778                  * just restarting the whole scan --- this case probably doesn't
 779                  * happen often enough to be worth trying harder, and anyway we don't
 780                  * want to hold any index internal locks while waiting.
 781                  */
 782                 xwait = TransactionIdIsValid(DirtySnapshot.xmin) ?
 783                         DirtySnapshot.xmin : DirtySnapshot.xmax;
 784
 785                 if (TransactionIdIsValid(xwait) &&
 786                         (waitMode == CEOUC_WAIT ||
 787                          (waitMode == CEOUC_LIVELOCK_PREVENTING_WAIT &&
 788                           DirtySnapshot.speculativeToken &&
 789                           TransactionIdPrecedes(GetCurrentTransactionId(), xwait))))
 790                 {
 791                         ctid_wait = tup->t_data->t_ctid;
 792                         reason_wait = indexInfo->ii_ExclusionOps ?
 793                                 XLTW_RecheckExclusionConstr : XLTW_InsertIndex;
 794                         index_endscan(index_scan);
 795                         if (DirtySnapshot.speculativeToken)
 796                                 SpeculativeInsertionWait(DirtySnapshot.xmin,
 797                                                                                  DirtySnapshot.speculativeToken);
 798                         else
 799                                 XactLockTableWait(xwait, heap, &ctid_wait, reason_wait);
 800                         goto retry;
 801                 }
 802
 803                 /*
 804                  * We have a definite conflict (or a potential one, but the caller
 805                  * didn't want to wait).  Return it to caller, or report it.
 806                  */
 807                 if (violationOK)
 808                 {
 809                         conflict = true;
 810                         if (conflictTid)
 811                                 *conflictTid = tup->t_self;
 812                         break;
 813                 }
 814
 815                 error_new = BuildIndexValueDescription(index, values, isnull);
 816                 error_existing = BuildIndexValueDescription(index, existing_values,
 817                                                                                                         existing_isnull);
 818                 if (newIndex)
 819                         ereport(ERROR,
 820                                         (errcode(ERRCODE_EXCLUSION_VIOLATION),
 821                                          errmsg("could not create exclusion constraint \"%s\"",
 822                                                         RelationGetRelationName(index)),
 823                                          error_new && error_existing ?
 824                                          errdetail("Key %s conflicts with key %s.",
 825                                                            error_new, error_existing) :
 826                                          errdetail("Key conflicts exist."),
 827                                          errtableconstraint(heap,
 828                                                                                 RelationGetRelationName(index))));
 829                 else
 830                         ereport(ERROR,
 831                                         (errcode(ERRCODE_EXCLUSION_VIOLATION),
 832                                          errmsg("conflicting key value violates exclusion constraint \"%s\"",
 833                                                         RelationGetRelationName(index)),
 834                                          error_new && error_existing ?
 835                                          errdetail("Key %s conflicts with existing key %s.",
 836                                                            error_new, error_existing) :
 837                                          errdetail("Key conflicts with existing key."),
 838                                          errtableconstraint(heap,
 839                                                                                 RelationGetRelationName(index))));
 840         }
 841
 842         index_endscan(index_scan);
 843
 844         /*
 845          * Ordinarily, at this point the search should have found the originally
 846          * inserted tuple (if any), unless we exited the loop early because of
 847          * conflict.  However, it is possible to define exclusion constraints for
 848          * which that wouldn't be true --- for instance, if the operator is <>. So
 849          * we no longer complain if found_self is still false.
 850          */
 851
 852         econtext->ecxt_scantuple = save_scantuple;
 853
 854         ExecDropSingleTupleTableSlot(existing_slot);
 855
 856         return !conflict;
 857 }
 858
 859 /*
 860  * Check for violation of an exclusion constraint
 861  *
 862  * This is a dumbed down version of check_exclusion_or_unique_constraint
 863  * for external callers. They don't need all the special modes.
 864  */
 865 void
 866 check_exclusion_constraint(Relation heap, Relation index,
 867                                                    IndexInfo *indexInfo,
 868                                                    ItemPointer tupleid,
 869                                                    Datum *values, bool *isnull,
 870                                                    EState *estate, bool newIndex)
 871 {
 872         (void) check_exclusion_or_unique_constraint(heap, index, indexInfo, tupleid,
 873                                                                                                 values, isnull,
 874                                                                                                 estate, newIndex,
 875                                                                                                 CEOUC_WAIT, false, NULL);
 876 }
 877
 878 /*
 879  * Check existing tuple's index values to see if it really matches the
 880  * exclusion condition against the new_values.  Returns true if conflict.
 881  */
 882 static bool
 883 index_recheck_constraint(Relation index, Oid *constr_procs,
 884                                                  Datum *existing_values, bool *existing_isnull,
 885                                                  Datum *new_values)
 886 {
 887         int                     index_natts = index->rd_index->indnatts;
 888         int                     i;
 889
 890         for (i = 0; i < index_natts; i++)
 891         {
 892                 /* Assume the exclusion operators are strict */
 893                 if (existing_isnull[i])
 894                         return false;
 895
 896                 if (!DatumGetBool(OidFunctionCall2Coll(constr_procs[i],
 897                                                                                            index->rd_indcollation[i],
 898                                                                                            existing_values[i],
 899                                                                                            new_values[i])))
 900                         return false;
 901         }
 902
 903         return true;
 904 }