]> granicus.if.org Git - postgresql/blob - src/backend/executor/execMain.c
Change the return value of HeapTupleSatisfiesUpdate() to be an enum,
[postgresql] / src / backend / executor / execMain.c
1 /*-------------------------------------------------------------------------
2  *
3  * execMain.c
4  *        top level executor interface routines
5  *
6  * INTERFACE ROUTINES
7  *      ExecutorStart()
8  *      ExecutorRun()
9  *      ExecutorEnd()
10  *
11  *      The old ExecutorMain() has been replaced by ExecutorStart(),
12  *      ExecutorRun() and ExecutorEnd()
13  *
14  *      These three procedures are the external interfaces to the executor.
15  *      In each case, the query descriptor is required as an argument.
16  *
17  *      ExecutorStart() must be called at the beginning of execution of any
18  *      query plan and ExecutorEnd() should always be called at the end of
19  *      execution of a plan.
20  *
21  *      ExecutorRun accepts direction and count arguments that specify whether
22  *      the plan is to be executed forwards, backwards, and for how many tuples.
23  *
24  * Portions Copyright (c) 1996-2005, PostgreSQL Global Development Group
25  * Portions Copyright (c) 1994, Regents of the University of California
26  *
27  *
28  * IDENTIFICATION
29  *        $PostgreSQL: pgsql/src/backend/executor/execMain.c,v 1.243 2005/03/20 23:40:25 neilc Exp $
30  *
31  *-------------------------------------------------------------------------
32  */
33 #include "postgres.h"
34
35 #include "access/heapam.h"
36 #include "catalog/heap.h"
37 #include "catalog/namespace.h"
38 #include "commands/tablecmds.h"
39 #include "commands/trigger.h"
40 #include "executor/execdebug.h"
41 #include "executor/execdefs.h"
42 #include "miscadmin.h"
43 #include "optimizer/clauses.h"
44 #include "optimizer/var.h"
45 #include "parser/parsetree.h"
46 #include "utils/acl.h"
47 #include "utils/guc.h"
48 #include "utils/lsyscache.h"
49
50
51 typedef struct execRowMark
52 {
53         Relation        relation;
54         Index           rti;
55         char            resname[32];
56 } execRowMark;
57
58 typedef struct evalPlanQual
59 {
60         Index           rti;
61         EState     *estate;
62         PlanState  *planstate;
63         struct evalPlanQual *next;      /* stack of active PlanQual plans */
64         struct evalPlanQual *free;      /* list of free PlanQual plans */
65 } evalPlanQual;
66
67 /* decls for local routines only used within this module */
68 static void InitPlan(QueryDesc *queryDesc, bool explainOnly);
69 static void initResultRelInfo(ResultRelInfo *resultRelInfo,
70                                   Index resultRelationIndex,
71                                   List *rangeTable,
72                                   CmdType operation);
73 static TupleTableSlot *ExecutePlan(EState *estate, PlanState *planstate,
74                         CmdType operation,
75                         long numberTuples,
76                         ScanDirection direction,
77                         DestReceiver *dest);
78 static void ExecSelect(TupleTableSlot *slot,
79                    DestReceiver *dest,
80                    EState *estate);
81 static void ExecInsert(TupleTableSlot *slot, ItemPointer tupleid,
82                    EState *estate);
83 static void ExecDelete(TupleTableSlot *slot, ItemPointer tupleid,
84                    EState *estate);
85 static void ExecUpdate(TupleTableSlot *slot, ItemPointer tupleid,
86                    EState *estate);
87 static TupleTableSlot *EvalPlanQualNext(EState *estate);
88 static void EndEvalPlanQual(EState *estate);
89 static void ExecCheckRTEPerms(RangeTblEntry *rte);
90 static void ExecCheckXactReadOnly(Query *parsetree);
91 static void EvalPlanQualStart(evalPlanQual *epq, EState *estate,
92                                   evalPlanQual *priorepq);
93 static void EvalPlanQualStop(evalPlanQual *epq);
94
95 /* end of local decls */
96
97
98 /* ----------------------------------------------------------------
99  *              ExecutorStart
100  *
101  *              This routine must be called at the beginning of any execution of any
102  *              query plan
103  *
104  * Takes a QueryDesc previously created by CreateQueryDesc (it's not real
105  * clear why we bother to separate the two functions, but...).  The tupDesc
106  * field of the QueryDesc is filled in to describe the tuples that will be
107  * returned, and the internal fields (estate and planstate) are set up.
108  *
109  * If explainOnly is true, we are not actually intending to run the plan,
110  * only to set up for EXPLAIN; so skip unwanted side-effects.
111  *
112  * NB: the CurrentMemoryContext when this is called will become the parent
113  * of the per-query context used for this Executor invocation.
114  * ----------------------------------------------------------------
115  */
116 void
117 ExecutorStart(QueryDesc *queryDesc, bool explainOnly)
118 {
119         EState     *estate;
120         MemoryContext oldcontext;
121
122         /* sanity checks: queryDesc must not be started already */
123         Assert(queryDesc != NULL);
124         Assert(queryDesc->estate == NULL);
125
126         /*
127          * If the transaction is read-only, we need to check if any writes are
128          * planned to non-temporary tables.
129          */
130         if (XactReadOnly && !explainOnly)
131                 ExecCheckXactReadOnly(queryDesc->parsetree);
132
133         /*
134          * Build EState, switch into per-query memory context for startup.
135          */
136         estate = CreateExecutorState();
137         queryDesc->estate = estate;
138
139         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
140
141         /*
142          * Fill in parameters, if any, from queryDesc
143          */
144         estate->es_param_list_info = queryDesc->params;
145
146         if (queryDesc->plantree->nParamExec > 0)
147                 estate->es_param_exec_vals = (ParamExecData *)
148                         palloc0(queryDesc->plantree->nParamExec * sizeof(ParamExecData));
149
150         /*
151          * Copy other important information into the EState
152          */
153         estate->es_snapshot = queryDesc->snapshot;
154         estate->es_crosscheck_snapshot = queryDesc->crosscheck_snapshot;
155         estate->es_instrument = queryDesc->doInstrument;
156
157         /*
158          * Initialize the plan state tree
159          */
160         InitPlan(queryDesc, explainOnly);
161
162         MemoryContextSwitchTo(oldcontext);
163 }
164
165 /* ----------------------------------------------------------------
166  *              ExecutorRun
167  *
168  *              This is the main routine of the executor module. It accepts
169  *              the query descriptor from the traffic cop and executes the
170  *              query plan.
171  *
172  *              ExecutorStart must have been called already.
173  *
174  *              If direction is NoMovementScanDirection then nothing is done
175  *              except to start up/shut down the destination.  Otherwise,
176  *              we retrieve up to 'count' tuples in the specified direction.
177  *
178  *              Note: count = 0 is interpreted as no portal limit, i.e., run to
179  *              completion.
180  *
181  * ----------------------------------------------------------------
182  */
183 TupleTableSlot *
184 ExecutorRun(QueryDesc *queryDesc,
185                         ScanDirection direction, long count)
186 {
187         EState     *estate;
188         CmdType         operation;
189         DestReceiver *dest;
190         TupleTableSlot *result;
191         MemoryContext oldcontext;
192
193         /* sanity checks */
194         Assert(queryDesc != NULL);
195
196         estate = queryDesc->estate;
197
198         Assert(estate != NULL);
199
200         /*
201          * Switch into per-query memory context
202          */
203         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
204
205         /*
206          * extract information from the query descriptor and the query
207          * feature.
208          */
209         operation = queryDesc->operation;
210         dest = queryDesc->dest;
211
212         /*
213          * startup tuple receiver
214          */
215         estate->es_processed = 0;
216         estate->es_lastoid = InvalidOid;
217
218         (*dest->rStartup) (dest, operation, queryDesc->tupDesc);
219
220         /*
221          * run plan
222          */
223         if (direction == NoMovementScanDirection)
224                 result = NULL;
225         else
226                 result = ExecutePlan(estate,
227                                                          queryDesc->planstate,
228                                                          operation,
229                                                          count,
230                                                          direction,
231                                                          dest);
232
233         /*
234          * shutdown receiver
235          */
236         (*dest->rShutdown) (dest);
237
238         MemoryContextSwitchTo(oldcontext);
239
240         return result;
241 }
242
243 /* ----------------------------------------------------------------
244  *              ExecutorEnd
245  *
246  *              This routine must be called at the end of execution of any
247  *              query plan
248  * ----------------------------------------------------------------
249  */
250 void
251 ExecutorEnd(QueryDesc *queryDesc)
252 {
253         EState     *estate;
254         MemoryContext oldcontext;
255
256         /* sanity checks */
257         Assert(queryDesc != NULL);
258
259         estate = queryDesc->estate;
260
261         Assert(estate != NULL);
262
263         /*
264          * Switch into per-query memory context to run ExecEndPlan
265          */
266         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
267
268         ExecEndPlan(queryDesc->planstate, estate);
269
270         /*
271          * Must switch out of context before destroying it
272          */
273         MemoryContextSwitchTo(oldcontext);
274
275         /*
276          * Release EState and per-query memory context.  This should release
277          * everything the executor has allocated.
278          */
279         FreeExecutorState(estate);
280
281         /* Reset queryDesc fields that no longer point to anything */
282         queryDesc->tupDesc = NULL;
283         queryDesc->estate = NULL;
284         queryDesc->planstate = NULL;
285 }
286
287 /* ----------------------------------------------------------------
288  *              ExecutorRewind
289  *
290  *              This routine may be called on an open queryDesc to rewind it
291  *              to the start.
292  * ----------------------------------------------------------------
293  */
294 void
295 ExecutorRewind(QueryDesc *queryDesc)
296 {
297         EState     *estate;
298         MemoryContext oldcontext;
299
300         /* sanity checks */
301         Assert(queryDesc != NULL);
302
303         estate = queryDesc->estate;
304
305         Assert(estate != NULL);
306
307         /* It's probably not sensible to rescan updating queries */
308         Assert(queryDesc->operation == CMD_SELECT);
309
310         /*
311          * Switch into per-query memory context
312          */
313         oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
314
315         /*
316          * rescan plan
317          */
318         ExecReScan(queryDesc->planstate, NULL);
319
320         MemoryContextSwitchTo(oldcontext);
321 }
322
323
324 /*
325  * ExecCheckRTPerms
326  *              Check access permissions for all relations listed in a range table.
327  */
328 void
329 ExecCheckRTPerms(List *rangeTable)
330 {
331         ListCell   *l;
332
333         foreach(l, rangeTable)
334         {
335                 RangeTblEntry *rte = lfirst(l);
336
337                 ExecCheckRTEPerms(rte);
338         }
339 }
340
341 /*
342  * ExecCheckRTEPerms
343  *              Check access permissions for a single RTE.
344  */
345 static void
346 ExecCheckRTEPerms(RangeTblEntry *rte)
347 {
348         AclMode         requiredPerms;
349         Oid                     relOid;
350         AclId           userid;
351
352         /*
353          * If it's a subquery, recursively examine its rangetable.
354          */
355         if (rte->rtekind == RTE_SUBQUERY)
356         {
357                 ExecCheckRTPerms(rte->subquery->rtable);
358                 return;
359         }
360
361         /*
362          * Otherwise, only plain-relation RTEs need to be checked here.
363          * Function RTEs are checked by init_fcache when the function is
364          * prepared for execution. Join and special RTEs need no checks.
365          */
366         if (rte->rtekind != RTE_RELATION)
367                 return;
368
369         /*
370          * No work if requiredPerms is empty.
371          */
372         requiredPerms = rte->requiredPerms;
373         if (requiredPerms == 0)
374                 return;
375
376         relOid = rte->relid;
377
378         /*
379          * userid to check as: current user unless we have a setuid
380          * indication.
381          *
382          * Note: GetUserId() is presently fast enough that there's no harm in
383          * calling it separately for each RTE.  If that stops being true, we
384          * could call it once in ExecCheckRTPerms and pass the userid down
385          * from there.  But for now, no need for the extra clutter.
386          */
387         userid = rte->checkAsUser ? rte->checkAsUser : GetUserId();
388
389         /*
390          * We must have *all* the requiredPerms bits, so use aclmask not
391          * aclcheck.
392          */
393         if (pg_class_aclmask(relOid, userid, requiredPerms, ACLMASK_ALL)
394                 != requiredPerms)
395                 aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
396                                            get_rel_name(relOid));
397 }
398
399 /*
400  * Check that the query does not imply any writes to non-temp tables.
401  */
402 static void
403 ExecCheckXactReadOnly(Query *parsetree)
404 {
405         ListCell   *l;
406
407         /*
408          * CREATE TABLE AS or SELECT INTO?
409          *
410          * XXX should we allow this if the destination is temp?
411          */
412         if (parsetree->into != NULL)
413                 goto fail;
414
415         /* Fail if write permissions are requested on any non-temp table */
416         foreach(l, parsetree->rtable)
417         {
418                 RangeTblEntry *rte = lfirst(l);
419
420                 if (rte->rtekind == RTE_SUBQUERY)
421                 {
422                         ExecCheckXactReadOnly(rte->subquery);
423                         continue;
424                 }
425
426                 if (rte->rtekind != RTE_RELATION)
427                         continue;
428
429                 if ((rte->requiredPerms & (~ACL_SELECT)) == 0)
430                         continue;
431
432                 if (isTempNamespace(get_rel_namespace(rte->relid)))
433                         continue;
434
435                 goto fail;
436         }
437
438         return;
439
440 fail:
441         ereport(ERROR,
442                         (errcode(ERRCODE_READ_ONLY_SQL_TRANSACTION),
443                          errmsg("transaction is read-only")));
444 }
445
446
447 /* ----------------------------------------------------------------
448  *              InitPlan
449  *
450  *              Initializes the query plan: open files, allocate storage
451  *              and start up the rule manager
452  * ----------------------------------------------------------------
453  */
454 static void
455 InitPlan(QueryDesc *queryDesc, bool explainOnly)
456 {
457         CmdType         operation = queryDesc->operation;
458         Query      *parseTree = queryDesc->parsetree;
459         Plan       *plan = queryDesc->plantree;
460         EState     *estate = queryDesc->estate;
461         PlanState  *planstate;
462         List       *rangeTable;
463         Relation        intoRelationDesc;
464         bool            do_select_into;
465         TupleDesc       tupType;
466
467         /*
468          * Do permissions checks.  It's sufficient to examine the query's top
469          * rangetable here --- subplan RTEs will be checked during
470          * ExecInitSubPlan().
471          */
472         ExecCheckRTPerms(parseTree->rtable);
473
474         /*
475          * get information from query descriptor
476          */
477         rangeTable = parseTree->rtable;
478
479         /*
480          * initialize the node's execution state
481          */
482         estate->es_range_table = rangeTable;
483
484         /*
485          * if there is a result relation, initialize result relation stuff
486          */
487         if (parseTree->resultRelation != 0 && operation != CMD_SELECT)
488         {
489                 List       *resultRelations = parseTree->resultRelations;
490                 int                     numResultRelations;
491                 ResultRelInfo *resultRelInfos;
492
493                 if (resultRelations != NIL)
494                 {
495                         /*
496                          * Multiple result relations (due to inheritance)
497                          * parseTree->resultRelations identifies them all
498                          */
499                         ResultRelInfo *resultRelInfo;
500                         ListCell   *l;
501
502                         numResultRelations = list_length(resultRelations);
503                         resultRelInfos = (ResultRelInfo *)
504                                 palloc(numResultRelations * sizeof(ResultRelInfo));
505                         resultRelInfo = resultRelInfos;
506                         foreach(l, resultRelations)
507                         {
508                                 initResultRelInfo(resultRelInfo,
509                                                                   lfirst_int(l),
510                                                                   rangeTable,
511                                                                   operation);
512                                 resultRelInfo++;
513                         }
514                 }
515                 else
516                 {
517                         /*
518                          * Single result relation identified by
519                          * parseTree->resultRelation
520                          */
521                         numResultRelations = 1;
522                         resultRelInfos = (ResultRelInfo *) palloc(sizeof(ResultRelInfo));
523                         initResultRelInfo(resultRelInfos,
524                                                           parseTree->resultRelation,
525                                                           rangeTable,
526                                                           operation);
527                 }
528
529                 estate->es_result_relations = resultRelInfos;
530                 estate->es_num_result_relations = numResultRelations;
531                 /* Initialize to first or only result rel */
532                 estate->es_result_relation_info = resultRelInfos;
533         }
534         else
535         {
536                 /*
537                  * if no result relation, then set state appropriately
538                  */
539                 estate->es_result_relations = NULL;
540                 estate->es_num_result_relations = 0;
541                 estate->es_result_relation_info = NULL;
542         }
543
544         /*
545          * Detect whether we're doing SELECT INTO.  If so, set the es_into_oids
546          * flag appropriately so that the plan tree will be initialized with
547          * the correct tuple descriptors.
548          */
549         do_select_into = false;
550
551         if (operation == CMD_SELECT && parseTree->into != NULL)
552         {
553                 do_select_into = true;
554                 estate->es_select_into = true;
555                 estate->es_into_oids = parseTree->intoHasOids;
556         }
557
558         /*
559          * Have to lock relations selected for update
560          */
561         estate->es_rowMark = NIL;
562         if (parseTree->rowMarks != NIL)
563         {
564                 ListCell   *l;
565
566                 foreach(l, parseTree->rowMarks)
567                 {
568                         Index           rti = lfirst_int(l);
569                         Oid                     relid = getrelid(rti, rangeTable);
570                         Relation        relation;
571                         execRowMark *erm;
572
573                         relation = heap_open(relid, RowShareLock);
574                         erm = (execRowMark *) palloc(sizeof(execRowMark));
575                         erm->relation = relation;
576                         erm->rti = rti;
577                         snprintf(erm->resname, sizeof(erm->resname), "ctid%u", rti);
578                         estate->es_rowMark = lappend(estate->es_rowMark, erm);
579                 }
580         }
581
582         /*
583          * initialize the executor "tuple" table.  We need slots for all the
584          * plan nodes, plus possibly output slots for the junkfilter(s). At
585          * this point we aren't sure if we need junkfilters, so just add slots
586          * for them unconditionally.
587          */
588         {
589                 int                     nSlots = ExecCountSlotsNode(plan);
590
591                 if (parseTree->resultRelations != NIL)
592                         nSlots += list_length(parseTree->resultRelations);
593                 else
594                         nSlots += 1;
595                 estate->es_tupleTable = ExecCreateTupleTable(nSlots);
596         }
597
598         /* mark EvalPlanQual not active */
599         estate->es_topPlan = plan;
600         estate->es_evalPlanQual = NULL;
601         estate->es_evTupleNull = NULL;
602         estate->es_evTuple = NULL;
603         estate->es_useEvalPlan = false;
604
605         /*
606          * initialize the private state information for all the nodes in the
607          * query tree.  This opens files, allocates storage and leaves us
608          * ready to start processing tuples.
609          */
610         planstate = ExecInitNode(plan, estate);
611
612         /*
613          * Get the tuple descriptor describing the type of tuples to return.
614          * (this is especially important if we are creating a relation with
615          * "SELECT INTO")
616          */
617         tupType = ExecGetResultType(planstate);
618
619         /*
620          * Initialize the junk filter if needed.  SELECT and INSERT queries
621          * need a filter if there are any junk attrs in the tlist.      INSERT and
622          * SELECT INTO also need a filter if the plan may return raw disk
623          * tuples (else heap_insert will be scribbling on the source
624          * relation!). UPDATE and DELETE always need a filter, since there's
625          * always a junk 'ctid' attribute present --- no need to look first.
626          */
627         {
628                 bool            junk_filter_needed = false;
629                 ListCell   *tlist;
630
631                 switch (operation)
632                 {
633                         case CMD_SELECT:
634                         case CMD_INSERT:
635                                 foreach(tlist, plan->targetlist)
636                                 {
637                                         TargetEntry *tle = (TargetEntry *) lfirst(tlist);
638
639                                         if (tle->resdom->resjunk)
640                                         {
641                                                 junk_filter_needed = true;
642                                                 break;
643                                         }
644                                 }
645                                 if (!junk_filter_needed &&
646                                         (operation == CMD_INSERT || do_select_into) &&
647                                         ExecMayReturnRawTuples(planstate))
648                                         junk_filter_needed = true;
649                                 break;
650                         case CMD_UPDATE:
651                         case CMD_DELETE:
652                                 junk_filter_needed = true;
653                                 break;
654                         default:
655                                 break;
656                 }
657
658                 if (junk_filter_needed)
659                 {
660                         /*
661                          * If there are multiple result relations, each one needs its
662                          * own junk filter.  Note this is only possible for
663                          * UPDATE/DELETE, so we can't be fooled by some needing a
664                          * filter and some not.
665                          */
666                         if (parseTree->resultRelations != NIL)
667                         {
668                                 PlanState **appendplans;
669                                 int                     as_nplans;
670                                 ResultRelInfo *resultRelInfo;
671                                 int                     i;
672
673                                 /* Top plan had better be an Append here. */
674                                 Assert(IsA(plan, Append));
675                                 Assert(((Append *) plan)->isTarget);
676                                 Assert(IsA(planstate, AppendState));
677                                 appendplans = ((AppendState *) planstate)->appendplans;
678                                 as_nplans = ((AppendState *) planstate)->as_nplans;
679                                 Assert(as_nplans == estate->es_num_result_relations);
680                                 resultRelInfo = estate->es_result_relations;
681                                 for (i = 0; i < as_nplans; i++)
682                                 {
683                                         PlanState  *subplan = appendplans[i];
684                                         JunkFilter *j;
685
686                                         j = ExecInitJunkFilter(subplan->plan->targetlist,
687                                                                                    resultRelInfo->ri_RelationDesc->rd_att->tdhasoid,
688                                                                                    ExecAllocTableSlot(estate->es_tupleTable));
689                                         resultRelInfo->ri_junkFilter = j;
690                                         resultRelInfo++;
691                                 }
692
693                                 /*
694                                  * Set active junkfilter too; at this point ExecInitAppend
695                                  * has already selected an active result relation...
696                                  */
697                                 estate->es_junkFilter =
698                                         estate->es_result_relation_info->ri_junkFilter;
699                         }
700                         else
701                         {
702                                 /* Normal case with just one JunkFilter */
703                                 JunkFilter *j;
704
705                                 j = ExecInitJunkFilter(planstate->plan->targetlist,
706                                                                            tupType->tdhasoid,
707                                                           ExecAllocTableSlot(estate->es_tupleTable));
708                                 estate->es_junkFilter = j;
709                                 if (estate->es_result_relation_info)
710                                         estate->es_result_relation_info->ri_junkFilter = j;
711
712                                 /* For SELECT, want to return the cleaned tuple type */
713                                 if (operation == CMD_SELECT)
714                                         tupType = j->jf_cleanTupType;
715                         }
716                 }
717                 else
718                         estate->es_junkFilter = NULL;
719         }
720
721         /*
722          * If doing SELECT INTO, initialize the "into" relation.  We must wait
723          * till now so we have the "clean" result tuple type to create the new
724          * table from.
725          *
726          * If EXPLAIN, skip creating the "into" relation.
727          */
728         intoRelationDesc = NULL;
729
730         if (do_select_into && !explainOnly)
731         {
732                 char       *intoName;
733                 Oid                     namespaceId;
734                 AclResult       aclresult;
735                 Oid                     intoRelationId;
736                 TupleDesc       tupdesc;
737
738                 /*
739                  * find namespace to create in, check permissions
740                  */
741                 intoName = parseTree->into->relname;
742                 namespaceId = RangeVarGetCreationNamespace(parseTree->into);
743
744                 aclresult = pg_namespace_aclcheck(namespaceId, GetUserId(),
745                                                                                   ACL_CREATE);
746                 if (aclresult != ACLCHECK_OK)
747                         aclcheck_error(aclresult, ACL_KIND_NAMESPACE,
748                                                    get_namespace_name(namespaceId));
749
750                 /*
751                  * have to copy tupType to get rid of constraints
752                  */
753                 tupdesc = CreateTupleDescCopy(tupType);
754
755                 intoRelationId = heap_create_with_catalog(intoName,
756                                                                                                   namespaceId,
757                                                                                                   InvalidOid,
758                                                                                                   tupdesc,
759                                                                                                   RELKIND_RELATION,
760                                                                                                   false,
761                                                                                                   true,
762                                                                                                   0,
763                                                                                                   ONCOMMIT_NOOP,
764                                                                                                   allowSystemTableMods);
765
766                 FreeTupleDesc(tupdesc);
767
768                 /*
769                  * Advance command counter so that the newly-created relation's
770                  * catalog tuples will be visible to heap_open.
771                  */
772                 CommandCounterIncrement();
773
774                 /*
775                  * If necessary, create a TOAST table for the into relation. Note
776                  * that AlterTableCreateToastTable ends with
777                  * CommandCounterIncrement(), so that the TOAST table will be
778                  * visible for insertion.
779                  */
780                 AlterTableCreateToastTable(intoRelationId, true);
781
782                 /*
783                  * And open the constructed table for writing.
784                  */
785                 intoRelationDesc = heap_open(intoRelationId, AccessExclusiveLock);
786         }
787
788         estate->es_into_relation_descriptor = intoRelationDesc;
789
790         queryDesc->tupDesc = tupType;
791         queryDesc->planstate = planstate;
792 }
793
794 /*
795  * Initialize ResultRelInfo data for one result relation
796  */
797 static void
798 initResultRelInfo(ResultRelInfo *resultRelInfo,
799                                   Index resultRelationIndex,
800                                   List *rangeTable,
801                                   CmdType operation)
802 {
803         Oid                     resultRelationOid;
804         Relation        resultRelationDesc;
805
806         resultRelationOid = getrelid(resultRelationIndex, rangeTable);
807         resultRelationDesc = heap_open(resultRelationOid, RowExclusiveLock);
808
809         switch (resultRelationDesc->rd_rel->relkind)
810         {
811                 case RELKIND_SEQUENCE:
812                         ereport(ERROR,
813                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
814                                          errmsg("cannot change sequence \"%s\"",
815                                                   RelationGetRelationName(resultRelationDesc))));
816                         break;
817                 case RELKIND_TOASTVALUE:
818                         ereport(ERROR,
819                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
820                                          errmsg("cannot change TOAST relation \"%s\"",
821                                                   RelationGetRelationName(resultRelationDesc))));
822                         break;
823                 case RELKIND_VIEW:
824                         ereport(ERROR,
825                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
826                                          errmsg("cannot change view \"%s\"",
827                                                   RelationGetRelationName(resultRelationDesc))));
828                         break;
829         }
830
831         MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
832         resultRelInfo->type = T_ResultRelInfo;
833         resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
834         resultRelInfo->ri_RelationDesc = resultRelationDesc;
835         resultRelInfo->ri_NumIndices = 0;
836         resultRelInfo->ri_IndexRelationDescs = NULL;
837         resultRelInfo->ri_IndexRelationInfo = NULL;
838         /* make a copy so as not to depend on relcache info not changing... */
839         resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
840         resultRelInfo->ri_TrigFunctions = NULL;
841         resultRelInfo->ri_ConstraintExprs = NULL;
842         resultRelInfo->ri_junkFilter = NULL;
843
844         /*
845          * If there are indices on the result relation, open them and save
846          * descriptors in the result relation info, so that we can add new
847          * index entries for the tuples we add/update.  We need not do this
848          * for a DELETE, however, since deletion doesn't affect indexes.
849          */
850         if (resultRelationDesc->rd_rel->relhasindex &&
851                 operation != CMD_DELETE)
852                 ExecOpenIndices(resultRelInfo);
853 }
854
855 /*
856  *              ExecContextForcesOids
857  *
858  * This is pretty grotty: when doing INSERT, UPDATE, or SELECT INTO,
859  * we need to ensure that result tuples have space for an OID iff they are
860  * going to be stored into a relation that has OIDs.  In other contexts
861  * we are free to choose whether to leave space for OIDs in result tuples
862  * (we generally don't want to, but we do if a physical-tlist optimization
863  * is possible).  This routine checks the plan context and returns TRUE if the
864  * choice is forced, FALSE if the choice is not forced.  In the TRUE case,
865  * *hasoids is set to the required value.
866  *
867  * One reason this is ugly is that all plan nodes in the plan tree will emit
868  * tuples with space for an OID, though we really only need the topmost node
869  * to do so.  However, node types like Sort don't project new tuples but just
870  * return their inputs, and in those cases the requirement propagates down
871  * to the input node.  Eventually we might make this code smart enough to
872  * recognize how far down the requirement really goes, but for now we just
873  * make all plan nodes do the same thing if the top level forces the choice.
874  *
875  * We assume that estate->es_result_relation_info is already set up to
876  * describe the target relation.  Note that in an UPDATE that spans an
877  * inheritance tree, some of the target relations may have OIDs and some not.
878  * We have to make the decisions on a per-relation basis as we initialize
879  * each of the child plans of the topmost Append plan.
880  *
881  * SELECT INTO is even uglier, because we don't have the INTO relation's
882  * descriptor available when this code runs; we have to look aside at a
883  * flag set by InitPlan().
884  */
885 bool
886 ExecContextForcesOids(PlanState *planstate, bool *hasoids)
887 {
888         if (planstate->state->es_select_into)
889         {
890                 *hasoids = planstate->state->es_into_oids;
891                 return true;
892         }
893         else
894         {
895                 ResultRelInfo *ri = planstate->state->es_result_relation_info;
896
897                 if (ri != NULL)
898                 {
899                         Relation        rel = ri->ri_RelationDesc;
900
901                         if (rel != NULL)
902                         {
903                                 *hasoids = rel->rd_rel->relhasoids;
904                                 return true;
905                         }
906                 }
907         }
908
909         return false;
910 }
911
912 /* ----------------------------------------------------------------
913  *              ExecEndPlan
914  *
915  *              Cleans up the query plan -- closes files and frees up storage
916  *
917  * NOTE: we are no longer very worried about freeing storage per se
918  * in this code; FreeExecutorState should be guaranteed to release all
919  * memory that needs to be released.  What we are worried about doing
920  * is closing relations and dropping buffer pins.  Thus, for example,
921  * tuple tables must be cleared or dropped to ensure pins are released.
922  * ----------------------------------------------------------------
923  */
924 void
925 ExecEndPlan(PlanState *planstate, EState *estate)
926 {
927         ResultRelInfo *resultRelInfo;
928         int                     i;
929         ListCell   *l;
930
931         /*
932          * shut down any PlanQual processing we were doing
933          */
934         if (estate->es_evalPlanQual != NULL)
935                 EndEvalPlanQual(estate);
936
937         /*
938          * shut down the node-type-specific query processing
939          */
940         ExecEndNode(planstate);
941
942         /*
943          * destroy the executor "tuple" table.
944          */
945         ExecDropTupleTable(estate->es_tupleTable, true);
946         estate->es_tupleTable = NULL;
947
948         /*
949          * close the result relation(s) if any, but hold locks until xact
950          * commit.
951          */
952         resultRelInfo = estate->es_result_relations;
953         for (i = estate->es_num_result_relations; i > 0; i--)
954         {
955                 /* Close indices and then the relation itself */
956                 ExecCloseIndices(resultRelInfo);
957                 heap_close(resultRelInfo->ri_RelationDesc, NoLock);
958                 resultRelInfo++;
959         }
960
961         /*
962          * close the "into" relation if necessary, again keeping lock
963          */
964         if (estate->es_into_relation_descriptor != NULL)
965                 heap_close(estate->es_into_relation_descriptor, NoLock);
966
967         /*
968          * close any relations selected FOR UPDATE, again keeping locks
969          */
970         foreach(l, estate->es_rowMark)
971         {
972                 execRowMark *erm = lfirst(l);
973
974                 heap_close(erm->relation, NoLock);
975         }
976 }
977
978 /* ----------------------------------------------------------------
979  *              ExecutePlan
980  *
981  *              processes the query plan to retrieve 'numberTuples' tuples in the
982  *              direction specified.
983  *
984  *              Retrieves all tuples if numberTuples is 0
985  *
986  *              result is either a slot containing the last tuple in the case
987  *              of a SELECT or NULL otherwise.
988  *
989  * Note: the ctid attribute is a 'junk' attribute that is removed before the
990  * user can see it
991  * ----------------------------------------------------------------
992  */
993 static TupleTableSlot *
994 ExecutePlan(EState *estate,
995                         PlanState *planstate,
996                         CmdType operation,
997                         long numberTuples,
998                         ScanDirection direction,
999                         DestReceiver *dest)
1000 {
1001         JunkFilter *junkfilter;
1002         TupleTableSlot *slot;
1003         ItemPointer tupleid = NULL;
1004         ItemPointerData tuple_ctid;
1005         long            current_tuple_count;
1006         TupleTableSlot *result;
1007
1008         /*
1009          * initialize local variables
1010          */
1011         slot = NULL;
1012         current_tuple_count = 0;
1013         result = NULL;
1014
1015         /*
1016          * Set the direction.
1017          */
1018         estate->es_direction = direction;
1019
1020         /*
1021          * Process BEFORE EACH STATEMENT triggers
1022          */
1023         switch (operation)
1024         {
1025                 case CMD_UPDATE:
1026                         ExecBSUpdateTriggers(estate, estate->es_result_relation_info);
1027                         break;
1028                 case CMD_DELETE:
1029                         ExecBSDeleteTriggers(estate, estate->es_result_relation_info);
1030                         break;
1031                 case CMD_INSERT:
1032                         ExecBSInsertTriggers(estate, estate->es_result_relation_info);
1033                         break;
1034                 default:
1035                         /* do nothing */
1036                         break;
1037         }
1038
1039         /*
1040          * Loop until we've processed the proper number of tuples from the
1041          * plan.
1042          */
1043
1044         for (;;)
1045         {
1046                 /* Reset the per-output-tuple exprcontext */
1047                 ResetPerTupleExprContext(estate);
1048
1049                 /*
1050                  * Execute the plan and obtain a tuple
1051                  */
1052 lnext:  ;
1053                 if (estate->es_useEvalPlan)
1054                 {
1055                         slot = EvalPlanQualNext(estate);
1056                         if (TupIsNull(slot))
1057                                 slot = ExecProcNode(planstate);
1058                 }
1059                 else
1060                         slot = ExecProcNode(planstate);
1061
1062                 /*
1063                  * if the tuple is null, then we assume there is nothing more to
1064                  * process so we just return null...
1065                  */
1066                 if (TupIsNull(slot))
1067                 {
1068                         result = NULL;
1069                         break;
1070                 }
1071
1072                 /*
1073                  * if we have a junk filter, then project a new tuple with the
1074                  * junk removed.
1075                  *
1076                  * Store this new "clean" tuple in the junkfilter's resultSlot.
1077                  * (Formerly, we stored it back over the "dirty" tuple, which is
1078                  * WRONG because that tuple slot has the wrong descriptor.)
1079                  *
1080                  * Also, extract all the junk information we need.
1081                  */
1082                 if ((junkfilter = estate->es_junkFilter) != NULL)
1083                 {
1084                         Datum           datum;
1085                         bool            isNull;
1086
1087                         /*
1088                          * extract the 'ctid' junk attribute.
1089                          */
1090                         if (operation == CMD_UPDATE || operation == CMD_DELETE)
1091                         {
1092                                 if (!ExecGetJunkAttribute(junkfilter,
1093                                                                                   slot,
1094                                                                                   "ctid",
1095                                                                                   &datum,
1096                                                                                   &isNull))
1097                                         elog(ERROR, "could not find junk ctid column");
1098
1099                                 /* shouldn't ever get a null result... */
1100                                 if (isNull)
1101                                         elog(ERROR, "ctid is NULL");
1102
1103                                 tupleid = (ItemPointer) DatumGetPointer(datum);
1104                                 tuple_ctid = *tupleid;  /* make sure we don't free the
1105                                                                                  * ctid!! */
1106                                 tupleid = &tuple_ctid;
1107                         }
1108                         else if (estate->es_rowMark != NIL)
1109                         {
1110                                 ListCell   *l;
1111
1112                 lmark:  ;
1113                                 foreach(l, estate->es_rowMark)
1114                                 {
1115                                         execRowMark *erm = lfirst(l);
1116                                         Buffer          buffer;
1117                                         HeapTupleData tuple;
1118                                         TupleTableSlot *newSlot;
1119                                         HTSU_Result             test;
1120
1121                                         if (!ExecGetJunkAttribute(junkfilter,
1122                                                                                           slot,
1123                                                                                           erm->resname,
1124                                                                                           &datum,
1125                                                                                           &isNull))
1126                                                 elog(ERROR, "could not find junk \"%s\" column",
1127                                                          erm->resname);
1128
1129                                         /* shouldn't ever get a null result... */
1130                                         if (isNull)
1131                                                 elog(ERROR, "\"%s\" is NULL", erm->resname);
1132
1133                                         tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
1134                                         test = heap_mark4update(erm->relation, &tuple, &buffer,
1135                                                                                         estate->es_snapshot->curcid);
1136                                         ReleaseBuffer(buffer);
1137                                         switch (test)
1138                                         {
1139                                                 case HeapTupleSelfUpdated:
1140                                                         /* treat it as deleted; do not process */
1141                                                         goto lnext;
1142
1143                                                 case HeapTupleMayBeUpdated:
1144                                                         break;
1145
1146                                                 case HeapTupleUpdated:
1147                                                         if (IsXactIsoLevelSerializable)
1148                                                                 ereport(ERROR,
1149                                                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1150                                                                                  errmsg("could not serialize access due to concurrent update")));
1151                                                         if (!(ItemPointerEquals(&(tuple.t_self),
1152                                                                   (ItemPointer) DatumGetPointer(datum))))
1153                                                         {
1154                                                                 newSlot = EvalPlanQual(estate, erm->rti, &(tuple.t_self));
1155                                                                 if (!(TupIsNull(newSlot)))
1156                                                                 {
1157                                                                         slot = newSlot;
1158                                                                         estate->es_useEvalPlan = true;
1159                                                                         goto lmark;
1160                                                                 }
1161                                                         }
1162
1163                                                         /*
1164                                                          * if tuple was deleted or PlanQual failed for
1165                                                          * updated tuple - we must not return this
1166                                                          * tuple!
1167                                                          */
1168                                                         goto lnext;
1169
1170                                                 default:
1171                                                         elog(ERROR, "unrecognized heap_mark4update status: %u",
1172                                                                  test);
1173                                                         return (NULL);
1174                                         }
1175                                 }
1176                         }
1177
1178                         /*
1179                          * Finally create a new "clean" tuple with all junk attributes
1180                          * removed
1181                          */
1182                         slot = ExecFilterJunk(junkfilter, slot);
1183                 }
1184
1185                 /*
1186                  * now that we have a tuple, do the appropriate thing with it..
1187                  * either return it to the user, add it to a relation someplace,
1188                  * delete it from a relation, or modify some of its attributes.
1189                  */
1190                 switch (operation)
1191                 {
1192                         case CMD_SELECT:
1193                                 ExecSelect(slot,        /* slot containing tuple */
1194                                                    dest,        /* destination's tuple-receiver obj */
1195                                                    estate);
1196                                 result = slot;
1197                                 break;
1198
1199                         case CMD_INSERT:
1200                                 ExecInsert(slot, tupleid, estate);
1201                                 result = NULL;
1202                                 break;
1203
1204                         case CMD_DELETE:
1205                                 ExecDelete(slot, tupleid, estate);
1206                                 result = NULL;
1207                                 break;
1208
1209                         case CMD_UPDATE:
1210                                 ExecUpdate(slot, tupleid, estate);
1211                                 result = NULL;
1212                                 break;
1213
1214                         default:
1215                                 elog(ERROR, "unrecognized operation code: %d",
1216                                          (int) operation);
1217                                 result = NULL;
1218                                 break;
1219                 }
1220
1221                 /*
1222                  * check our tuple count.. if we've processed the proper number
1223                  * then quit, else loop again and process more tuples.  Zero
1224                  * numberTuples means no limit.
1225                  */
1226                 current_tuple_count++;
1227                 if (numberTuples && numberTuples == current_tuple_count)
1228                         break;
1229         }
1230
1231         /*
1232          * Process AFTER EACH STATEMENT triggers
1233          */
1234         switch (operation)
1235         {
1236                 case CMD_UPDATE:
1237                         ExecASUpdateTriggers(estate, estate->es_result_relation_info);
1238                         break;
1239                 case CMD_DELETE:
1240                         ExecASDeleteTriggers(estate, estate->es_result_relation_info);
1241                         break;
1242                 case CMD_INSERT:
1243                         ExecASInsertTriggers(estate, estate->es_result_relation_info);
1244                         break;
1245                 default:
1246                         /* do nothing */
1247                         break;
1248         }
1249
1250         /*
1251          * here, result is either a slot containing a tuple in the case of a
1252          * SELECT or NULL otherwise.
1253          */
1254         return result;
1255 }
1256
1257 /* ----------------------------------------------------------------
1258  *              ExecSelect
1259  *
1260  *              SELECTs are easy.. we just pass the tuple to the appropriate
1261  *              print function.  The only complexity is when we do a
1262  *              "SELECT INTO", in which case we insert the tuple into
1263  *              the appropriate relation (note: this is a newly created relation
1264  *              so we don't need to worry about indices or locks.)
1265  * ----------------------------------------------------------------
1266  */
1267 static void
1268 ExecSelect(TupleTableSlot *slot,
1269                    DestReceiver *dest,
1270                    EState *estate)
1271 {
1272         /*
1273          * insert the tuple into the "into relation"
1274          *
1275          * XXX this probably ought to be replaced by a separate destination
1276          */
1277         if (estate->es_into_relation_descriptor != NULL)
1278         {
1279                 HeapTuple       tuple;
1280
1281                 tuple = ExecCopySlotTuple(slot);
1282                 heap_insert(estate->es_into_relation_descriptor, tuple,
1283                                         estate->es_snapshot->curcid);
1284                 /* we know there are no indexes to update */
1285                 heap_freetuple(tuple);
1286                 IncrAppended();
1287         }
1288
1289         /*
1290          * send the tuple to the destination
1291          */
1292         (*dest->receiveSlot) (slot, dest);
1293         IncrRetrieved();
1294         (estate->es_processed)++;
1295 }
1296
1297 /* ----------------------------------------------------------------
1298  *              ExecInsert
1299  *
1300  *              INSERTs are trickier.. we have to insert the tuple into
1301  *              the base relation and insert appropriate tuples into the
1302  *              index relations.
1303  * ----------------------------------------------------------------
1304  */
1305 static void
1306 ExecInsert(TupleTableSlot *slot,
1307                    ItemPointer tupleid,
1308                    EState *estate)
1309 {
1310         HeapTuple       tuple;
1311         ResultRelInfo *resultRelInfo;
1312         Relation        resultRelationDesc;
1313         int                     numIndices;
1314         Oid                     newId;
1315
1316         /*
1317          * get the heap tuple out of the tuple table slot, making sure
1318          * we have a writable copy
1319          */
1320         tuple = ExecMaterializeSlot(slot);
1321
1322         /*
1323          * get information on the (current) result relation
1324          */
1325         resultRelInfo = estate->es_result_relation_info;
1326         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1327
1328         /* BEFORE ROW INSERT Triggers */
1329         if (resultRelInfo->ri_TrigDesc &&
1330           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_INSERT] > 0)
1331         {
1332                 HeapTuple       newtuple;
1333
1334                 newtuple = ExecBRInsertTriggers(estate, resultRelInfo, tuple);
1335
1336                 if (newtuple == NULL)   /* "do nothing" */
1337                         return;
1338
1339                 if (newtuple != tuple)  /* modified by Trigger(s) */
1340                 {
1341                         /*
1342                          * Insert modified tuple into tuple table slot, replacing the
1343                          * original.  We assume that it was allocated in per-tuple
1344                          * memory context, and therefore will go away by itself. The
1345                          * tuple table slot should not try to clear it.
1346                          */
1347                         ExecStoreTuple(newtuple, slot, InvalidBuffer, false);
1348                         tuple = newtuple;
1349                 }
1350         }
1351
1352         /*
1353          * Check the constraints of the tuple
1354          */
1355         if (resultRelationDesc->rd_att->constr)
1356                 ExecConstraints(resultRelInfo, slot, estate);
1357
1358         /*
1359          * insert the tuple
1360          */
1361         newId = heap_insert(resultRelationDesc, tuple,
1362                                                 estate->es_snapshot->curcid);
1363
1364         IncrAppended();
1365         (estate->es_processed)++;
1366         estate->es_lastoid = newId;
1367         setLastTid(&(tuple->t_self));
1368
1369         /*
1370          * process indices
1371          *
1372          * Note: heap_insert adds a new tuple to a relation.  As a side effect,
1373          * the tupleid of the new tuple is placed in the new tuple's t_ctid
1374          * field.
1375          */
1376         numIndices = resultRelInfo->ri_NumIndices;
1377         if (numIndices > 0)
1378                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1379
1380         /* AFTER ROW INSERT Triggers */
1381         ExecARInsertTriggers(estate, resultRelInfo, tuple);
1382 }
1383
1384 /* ----------------------------------------------------------------
1385  *              ExecDelete
1386  *
1387  *              DELETE is like UPDATE, we delete the tuple and its
1388  *              index tuples.
1389  * ----------------------------------------------------------------
1390  */
1391 static void
1392 ExecDelete(TupleTableSlot *slot,
1393                    ItemPointer tupleid,
1394                    EState *estate)
1395 {
1396         ResultRelInfo *resultRelInfo;
1397         Relation        resultRelationDesc;
1398         ItemPointerData ctid;
1399         HTSU_Result     result;
1400
1401         /*
1402          * get information on the (current) result relation
1403          */
1404         resultRelInfo = estate->es_result_relation_info;
1405         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1406
1407         /* BEFORE ROW DELETE Triggers */
1408         if (resultRelInfo->ri_TrigDesc &&
1409           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_DELETE] > 0)
1410         {
1411                 bool            dodelete;
1412
1413                 dodelete = ExecBRDeleteTriggers(estate, resultRelInfo, tupleid,
1414                                                                                 estate->es_snapshot->curcid);
1415
1416                 if (!dodelete)                  /* "do nothing" */
1417                         return;
1418         }
1419
1420         /*
1421          * delete the tuple
1422          *
1423          * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check that
1424          * the row to be deleted is visible to that snapshot, and throw a can't-
1425          * serialize error if not.  This is a special-case behavior needed for
1426          * referential integrity updates in serializable transactions.
1427          */
1428 ldelete:;
1429         result = heap_delete(resultRelationDesc, tupleid,
1430                                                  &ctid,
1431                                                  estate->es_snapshot->curcid,
1432                                                  estate->es_crosscheck_snapshot,
1433                                                  true /* wait for commit */ );
1434         switch (result)
1435         {
1436                 case HeapTupleSelfUpdated:
1437                         /* already deleted by self; nothing to do */
1438                         return;
1439
1440                 case HeapTupleMayBeUpdated:
1441                         break;
1442
1443                 case HeapTupleUpdated:
1444                         if (IsXactIsoLevelSerializable)
1445                                 ereport(ERROR,
1446                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1447                                                  errmsg("could not serialize access due to concurrent update")));
1448                         else if (!(ItemPointerEquals(tupleid, &ctid)))
1449                         {
1450                                 TupleTableSlot *epqslot = EvalPlanQual(estate,
1451                                                            resultRelInfo->ri_RangeTableIndex, &ctid);
1452
1453                                 if (!TupIsNull(epqslot))
1454                                 {
1455                                         *tupleid = ctid;
1456                                         goto ldelete;
1457                                 }
1458                         }
1459                         /* tuple already deleted; nothing to do */
1460                         return;
1461
1462                 default:
1463                         elog(ERROR, "unrecognized heap_delete status: %u", result);
1464                         return;
1465         }
1466
1467         IncrDeleted();
1468         (estate->es_processed)++;
1469
1470         /*
1471          * Note: Normally one would think that we have to delete index tuples
1472          * associated with the heap tuple now..
1473          *
1474          * ... but in POSTGRES, we have no need to do this because the vacuum
1475          * daemon automatically opens an index scan and deletes index tuples
1476          * when it finds deleted heap tuples. -cim 9/27/89
1477          */
1478
1479         /* AFTER ROW DELETE Triggers */
1480         ExecARDeleteTriggers(estate, resultRelInfo, tupleid);
1481 }
1482
1483 /* ----------------------------------------------------------------
1484  *              ExecUpdate
1485  *
1486  *              note: we can't run UPDATE queries with transactions
1487  *              off because UPDATEs are actually INSERTs and our
1488  *              scan will mistakenly loop forever, updating the tuple
1489  *              it just inserted..      This should be fixed but until it
1490  *              is, we don't want to get stuck in an infinite loop
1491  *              which corrupts your database..
1492  * ----------------------------------------------------------------
1493  */
1494 static void
1495 ExecUpdate(TupleTableSlot *slot,
1496                    ItemPointer tupleid,
1497                    EState *estate)
1498 {
1499         HeapTuple       tuple;
1500         ResultRelInfo *resultRelInfo;
1501         Relation        resultRelationDesc;
1502         ItemPointerData ctid;
1503         HTSU_Result     result;
1504         int                     numIndices;
1505
1506         /*
1507          * abort the operation if not running transactions
1508          */
1509         if (IsBootstrapProcessingMode())
1510                 elog(ERROR, "cannot UPDATE during bootstrap");
1511
1512         /*
1513          * get the heap tuple out of the tuple table slot, making sure
1514          * we have a writable copy
1515          */
1516         tuple = ExecMaterializeSlot(slot);
1517
1518         /*
1519          * get information on the (current) result relation
1520          */
1521         resultRelInfo = estate->es_result_relation_info;
1522         resultRelationDesc = resultRelInfo->ri_RelationDesc;
1523
1524         /* BEFORE ROW UPDATE Triggers */
1525         if (resultRelInfo->ri_TrigDesc &&
1526           resultRelInfo->ri_TrigDesc->n_before_row[TRIGGER_EVENT_UPDATE] > 0)
1527         {
1528                 HeapTuple       newtuple;
1529
1530                 newtuple = ExecBRUpdateTriggers(estate, resultRelInfo,
1531                                                                                 tupleid, tuple,
1532                                                                                 estate->es_snapshot->curcid);
1533
1534                 if (newtuple == NULL)   /* "do nothing" */
1535                         return;
1536
1537                 if (newtuple != tuple)  /* modified by Trigger(s) */
1538                 {
1539                         /*
1540                          * Insert modified tuple into tuple table slot, replacing the
1541                          * original.  We assume that it was allocated in per-tuple
1542                          * memory context, and therefore will go away by itself. The
1543                          * tuple table slot should not try to clear it.
1544                          */
1545                         ExecStoreTuple(newtuple, slot, InvalidBuffer, false);
1546                         tuple = newtuple;
1547                 }
1548         }
1549
1550         /*
1551          * Check the constraints of the tuple
1552          *
1553          * If we generate a new candidate tuple after EvalPlanQual testing, we
1554          * must loop back here and recheck constraints.  (We don't need to
1555          * redo triggers, however.      If there are any BEFORE triggers then
1556          * trigger.c will have done mark4update to lock the correct tuple, so
1557          * there's no need to do them again.)
1558          */
1559 lreplace:;
1560         if (resultRelationDesc->rd_att->constr)
1561                 ExecConstraints(resultRelInfo, slot, estate);
1562
1563         /*
1564          * replace the heap tuple
1565          *
1566          * Note: if es_crosscheck_snapshot isn't InvalidSnapshot, we check that
1567          * the row to be updated is visible to that snapshot, and throw a can't-
1568          * serialize error if not.  This is a special-case behavior needed for
1569          * referential integrity updates in serializable transactions.
1570          */
1571         result = heap_update(resultRelationDesc, tupleid, tuple,
1572                                                  &ctid,
1573                                                  estate->es_snapshot->curcid,
1574                                                  estate->es_crosscheck_snapshot,
1575                                                  true /* wait for commit */ );
1576         switch (result)
1577         {
1578                 case HeapTupleSelfUpdated:
1579                         /* already deleted by self; nothing to do */
1580                         return;
1581
1582                 case HeapTupleMayBeUpdated:
1583                         break;
1584
1585                 case HeapTupleUpdated:
1586                         if (IsXactIsoLevelSerializable)
1587                                 ereport(ERROR,
1588                                                 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1589                                                  errmsg("could not serialize access due to concurrent update")));
1590                         else if (!(ItemPointerEquals(tupleid, &ctid)))
1591                         {
1592                                 TupleTableSlot *epqslot = EvalPlanQual(estate,
1593                                                            resultRelInfo->ri_RangeTableIndex, &ctid);
1594
1595                                 if (!TupIsNull(epqslot))
1596                                 {
1597                                         *tupleid = ctid;
1598                                         slot = ExecFilterJunk(estate->es_junkFilter, epqslot);
1599                                         tuple = ExecMaterializeSlot(slot);
1600                                         goto lreplace;
1601                                 }
1602                         }
1603                         /* tuple already deleted; nothing to do */
1604                         return;
1605
1606                 default:
1607                         elog(ERROR, "unrecognized heap_update status: %u", result);
1608                         return;
1609         }
1610
1611         IncrReplaced();
1612         (estate->es_processed)++;
1613
1614         /*
1615          * Note: instead of having to update the old index tuples associated
1616          * with the heap tuple, all we do is form and insert new index tuples.
1617          * This is because UPDATEs are actually DELETEs and INSERTs and index
1618          * tuple deletion is done automagically by the vacuum daemon. All we
1619          * do is insert new index tuples.  -cim 9/27/89
1620          */
1621
1622         /*
1623          * process indices
1624          *
1625          * heap_update updates a tuple in the base relation by invalidating it
1626          * and then inserting a new tuple to the relation.      As a side effect,
1627          * the tupleid of the new tuple is placed in the new tuple's t_ctid
1628          * field.  So we now insert index tuples using the new tupleid stored
1629          * there.
1630          */
1631
1632         numIndices = resultRelInfo->ri_NumIndices;
1633         if (numIndices > 0)
1634                 ExecInsertIndexTuples(slot, &(tuple->t_self), estate, false);
1635
1636         /* AFTER ROW UPDATE Triggers */
1637         ExecARUpdateTriggers(estate, resultRelInfo, tupleid, tuple);
1638 }
1639
1640 static const char *
1641 ExecRelCheck(ResultRelInfo *resultRelInfo,
1642                          TupleTableSlot *slot, EState *estate)
1643 {
1644         Relation        rel = resultRelInfo->ri_RelationDesc;
1645         int                     ncheck = rel->rd_att->constr->num_check;
1646         ConstrCheck *check = rel->rd_att->constr->check;
1647         ExprContext *econtext;
1648         MemoryContext oldContext;
1649         List       *qual;
1650         int                     i;
1651
1652         /*
1653          * If first time through for this result relation, build expression
1654          * nodetrees for rel's constraint expressions.  Keep them in the
1655          * per-query memory context so they'll survive throughout the query.
1656          */
1657         if (resultRelInfo->ri_ConstraintExprs == NULL)
1658         {
1659                 oldContext = MemoryContextSwitchTo(estate->es_query_cxt);
1660                 resultRelInfo->ri_ConstraintExprs =
1661                         (List **) palloc(ncheck * sizeof(List *));
1662                 for (i = 0; i < ncheck; i++)
1663                 {
1664                         /* ExecQual wants implicit-AND form */
1665                         qual = make_ands_implicit(stringToNode(check[i].ccbin));
1666                         resultRelInfo->ri_ConstraintExprs[i] = (List *)
1667                                 ExecPrepareExpr((Expr *) qual, estate);
1668                 }
1669                 MemoryContextSwitchTo(oldContext);
1670         }
1671
1672         /*
1673          * We will use the EState's per-tuple context for evaluating
1674          * constraint expressions (creating it if it's not already there).
1675          */
1676         econtext = GetPerTupleExprContext(estate);
1677
1678         /* Arrange for econtext's scan tuple to be the tuple under test */
1679         econtext->ecxt_scantuple = slot;
1680
1681         /* And evaluate the constraints */
1682         for (i = 0; i < ncheck; i++)
1683         {
1684                 qual = resultRelInfo->ri_ConstraintExprs[i];
1685
1686                 /*
1687                  * NOTE: SQL92 specifies that a NULL result from a constraint
1688                  * expression is not to be treated as a failure.  Therefore, tell
1689                  * ExecQual to return TRUE for NULL.
1690                  */
1691                 if (!ExecQual(qual, econtext, true))
1692                         return check[i].ccname;
1693         }
1694
1695         /* NULL result means no error */
1696         return NULL;
1697 }
1698
1699 void
1700 ExecConstraints(ResultRelInfo *resultRelInfo,
1701                                 TupleTableSlot *slot, EState *estate)
1702 {
1703         Relation        rel = resultRelInfo->ri_RelationDesc;
1704         TupleConstr *constr = rel->rd_att->constr;
1705
1706         Assert(constr);
1707
1708         if (constr->has_not_null)
1709         {
1710                 int                     natts = rel->rd_att->natts;
1711                 int                     attrChk;
1712
1713                 for (attrChk = 1; attrChk <= natts; attrChk++)
1714                 {
1715                         if (rel->rd_att->attrs[attrChk - 1]->attnotnull &&
1716                                 slot_attisnull(slot, attrChk))
1717                                 ereport(ERROR,
1718                                                 (errcode(ERRCODE_NOT_NULL_VIOLATION),
1719                                                  errmsg("null value in column \"%s\" violates not-null constraint",
1720                                         NameStr(rel->rd_att->attrs[attrChk - 1]->attname))));
1721                 }
1722         }
1723
1724         if (constr->num_check > 0)
1725         {
1726                 const char *failed;
1727
1728                 if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
1729                         ereport(ERROR,
1730                                         (errcode(ERRCODE_CHECK_VIOLATION),
1731                                          errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
1732                                                         RelationGetRelationName(rel), failed)));
1733         }
1734 }
1735
1736 /*
1737  * Check a modified tuple to see if we want to process its updated version
1738  * under READ COMMITTED rules.
1739  *
1740  * See backend/executor/README for some info about how this works.
1741  */
1742 TupleTableSlot *
1743 EvalPlanQual(EState *estate, Index rti, ItemPointer tid)
1744 {
1745         evalPlanQual *epq;
1746         EState     *epqstate;
1747         Relation        relation;
1748         HeapTupleData tuple;
1749         HeapTuple       copyTuple = NULL;
1750         bool            endNode;
1751
1752         Assert(rti != 0);
1753
1754         /*
1755          * find relation containing target tuple
1756          */
1757         if (estate->es_result_relation_info != NULL &&
1758                 estate->es_result_relation_info->ri_RangeTableIndex == rti)
1759                 relation = estate->es_result_relation_info->ri_RelationDesc;
1760         else
1761         {
1762                 ListCell   *l;
1763
1764                 relation = NULL;
1765                 foreach(l, estate->es_rowMark)
1766                 {
1767                         if (((execRowMark *) lfirst(l))->rti == rti)
1768                         {
1769                                 relation = ((execRowMark *) lfirst(l))->relation;
1770                                 break;
1771                         }
1772                 }
1773                 if (relation == NULL)
1774                         elog(ERROR, "could not find RowMark for RT index %u", rti);
1775         }
1776
1777         /*
1778          * fetch tid tuple
1779          *
1780          * Loop here to deal with updated or busy tuples
1781          */
1782         tuple.t_self = *tid;
1783         for (;;)
1784         {
1785                 Buffer          buffer;
1786
1787                 if (heap_fetch(relation, SnapshotDirty, &tuple, &buffer, false, NULL))
1788                 {
1789                         TransactionId xwait = SnapshotDirty->xmax;
1790
1791                         /* xmin should not be dirty... */
1792                         if (TransactionIdIsValid(SnapshotDirty->xmin))
1793                                 elog(ERROR, "t_xmin is uncommitted in tuple to be updated");
1794
1795                         /*
1796                          * If tuple is being updated by other transaction then we have
1797                          * to wait for its commit/abort.
1798                          */
1799                         if (TransactionIdIsValid(xwait))
1800                         {
1801                                 ReleaseBuffer(buffer);
1802                                 XactLockTableWait(xwait);
1803                                 continue;
1804                         }
1805
1806                         /*
1807                          * We got tuple - now copy it for use by recheck query.
1808                          */
1809                         copyTuple = heap_copytuple(&tuple);
1810                         ReleaseBuffer(buffer);
1811                         break;
1812                 }
1813
1814                 /*
1815                  * Oops! Invalid tuple. Have to check is it updated or deleted.
1816                  * Note that it's possible to get invalid SnapshotDirty->tid if
1817                  * tuple updated by this transaction. Have we to check this ?
1818                  */
1819                 if (ItemPointerIsValid(&(SnapshotDirty->tid)) &&
1820                         !(ItemPointerEquals(&(tuple.t_self), &(SnapshotDirty->tid))))
1821                 {
1822                         /* updated, so look at the updated copy */
1823                         tuple.t_self = SnapshotDirty->tid;
1824                         continue;
1825                 }
1826
1827                 /*
1828                  * Deleted or updated by this transaction; forget it.
1829                  */
1830                 return NULL;
1831         }
1832
1833         /*
1834          * For UPDATE/DELETE we have to return tid of actual row we're
1835          * executing PQ for.
1836          */
1837         *tid = tuple.t_self;
1838
1839         /*
1840          * Need to run a recheck subquery.      Find or create a PQ stack entry.
1841          */
1842         epq = estate->es_evalPlanQual;
1843         endNode = true;
1844
1845         if (epq != NULL && epq->rti == 0)
1846         {
1847                 /* Top PQ stack entry is idle, so re-use it */
1848                 Assert(!(estate->es_useEvalPlan) && epq->next == NULL);
1849                 epq->rti = rti;
1850                 endNode = false;
1851         }
1852
1853         /*
1854          * If this is request for another RTE - Ra, - then we have to check
1855          * wasn't PlanQual requested for Ra already and if so then Ra' row was
1856          * updated again and we have to re-start old execution for Ra and
1857          * forget all what we done after Ra was suspended. Cool? -:))
1858          */
1859         if (epq != NULL && epq->rti != rti &&
1860                 epq->estate->es_evTuple[rti - 1] != NULL)
1861         {
1862                 do
1863                 {
1864                         evalPlanQual *oldepq;
1865
1866                         /* stop execution */
1867                         EvalPlanQualStop(epq);
1868                         /* pop previous PlanQual from the stack */
1869                         oldepq = epq->next;
1870                         Assert(oldepq && oldepq->rti != 0);
1871                         /* push current PQ to freePQ stack */
1872                         oldepq->free = epq;
1873                         epq = oldepq;
1874                         estate->es_evalPlanQual = epq;
1875                 } while (epq->rti != rti);
1876         }
1877
1878         /*
1879          * If we are requested for another RTE then we have to suspend
1880          * execution of current PlanQual and start execution for new one.
1881          */
1882         if (epq == NULL || epq->rti != rti)
1883         {
1884                 /* try to reuse plan used previously */
1885                 evalPlanQual *newepq = (epq != NULL) ? epq->free : NULL;
1886
1887                 if (newepq == NULL)             /* first call or freePQ stack is empty */
1888                 {
1889                         newepq = (evalPlanQual *) palloc0(sizeof(evalPlanQual));
1890                         newepq->free = NULL;
1891                         newepq->estate = NULL;
1892                         newepq->planstate = NULL;
1893                 }
1894                 else
1895                 {
1896                         /* recycle previously used PlanQual */
1897                         Assert(newepq->estate == NULL);
1898                         epq->free = NULL;
1899                 }
1900                 /* push current PQ to the stack */
1901                 newepq->next = epq;
1902                 epq = newepq;
1903                 estate->es_evalPlanQual = epq;
1904                 epq->rti = rti;
1905                 endNode = false;
1906         }
1907
1908         Assert(epq->rti == rti);
1909
1910         /*
1911          * Ok - we're requested for the same RTE.  Unfortunately we still have
1912          * to end and restart execution of the plan, because ExecReScan
1913          * wouldn't ensure that upper plan nodes would reset themselves.  We
1914          * could make that work if insertion of the target tuple were
1915          * integrated with the Param mechanism somehow, so that the upper plan
1916          * nodes know that their children's outputs have changed.
1917          *
1918          * Note that the stack of free evalPlanQual nodes is quite useless at the
1919          * moment, since it only saves us from pallocing/releasing the
1920          * evalPlanQual nodes themselves.  But it will be useful once we
1921          * implement ReScan instead of end/restart for re-using PlanQual
1922          * nodes.
1923          */
1924         if (endNode)
1925         {
1926                 /* stop execution */
1927                 EvalPlanQualStop(epq);
1928         }
1929
1930         /*
1931          * Initialize new recheck query.
1932          *
1933          * Note: if we were re-using PlanQual plans via ExecReScan, we'd need to
1934          * instead copy down changeable state from the top plan (including
1935          * es_result_relation_info, es_junkFilter) and reset locally
1936          * changeable state in the epq (including es_param_exec_vals,
1937          * es_evTupleNull).
1938          */
1939         EvalPlanQualStart(epq, estate, epq->next);
1940
1941         /*
1942          * free old RTE' tuple, if any, and store target tuple where
1943          * relation's scan node will see it
1944          */
1945         epqstate = epq->estate;
1946         if (epqstate->es_evTuple[rti - 1] != NULL)
1947                 heap_freetuple(epqstate->es_evTuple[rti - 1]);
1948         epqstate->es_evTuple[rti - 1] = copyTuple;
1949
1950         return EvalPlanQualNext(estate);
1951 }
1952
1953 static TupleTableSlot *
1954 EvalPlanQualNext(EState *estate)
1955 {
1956         evalPlanQual *epq = estate->es_evalPlanQual;
1957         MemoryContext oldcontext;
1958         TupleTableSlot *slot;
1959
1960         Assert(epq->rti != 0);
1961
1962 lpqnext:;
1963         oldcontext = MemoryContextSwitchTo(epq->estate->es_query_cxt);
1964         slot = ExecProcNode(epq->planstate);
1965         MemoryContextSwitchTo(oldcontext);
1966
1967         /*
1968          * No more tuples for this PQ. Continue previous one.
1969          */
1970         if (TupIsNull(slot))
1971         {
1972                 evalPlanQual *oldepq;
1973
1974                 /* stop execution */
1975                 EvalPlanQualStop(epq);
1976                 /* pop old PQ from the stack */
1977                 oldepq = epq->next;
1978                 if (oldepq == NULL)
1979                 {
1980                         /* this is the first (oldest) PQ - mark as free */
1981                         epq->rti = 0;
1982                         estate->es_useEvalPlan = false;
1983                         /* and continue Query execution */
1984                         return (NULL);
1985                 }
1986                 Assert(oldepq->rti != 0);
1987                 /* push current PQ to freePQ stack */
1988                 oldepq->free = epq;
1989                 epq = oldepq;
1990                 estate->es_evalPlanQual = epq;
1991                 goto lpqnext;
1992         }
1993
1994         return (slot);
1995 }
1996
1997 static void
1998 EndEvalPlanQual(EState *estate)
1999 {
2000         evalPlanQual *epq = estate->es_evalPlanQual;
2001
2002         if (epq->rti == 0)                      /* plans already shutdowned */
2003         {
2004                 Assert(epq->next == NULL);
2005                 return;
2006         }
2007
2008         for (;;)
2009         {
2010                 evalPlanQual *oldepq;
2011
2012                 /* stop execution */
2013                 EvalPlanQualStop(epq);
2014                 /* pop old PQ from the stack */
2015                 oldepq = epq->next;
2016                 if (oldepq == NULL)
2017                 {
2018                         /* this is the first (oldest) PQ - mark as free */
2019                         epq->rti = 0;
2020                         estate->es_useEvalPlan = false;
2021                         break;
2022                 }
2023                 Assert(oldepq->rti != 0);
2024                 /* push current PQ to freePQ stack */
2025                 oldepq->free = epq;
2026                 epq = oldepq;
2027                 estate->es_evalPlanQual = epq;
2028         }
2029 }
2030
2031 /*
2032  * Start execution of one level of PlanQual.
2033  *
2034  * This is a cut-down version of ExecutorStart(): we copy some state from
2035  * the top-level estate rather than initializing it fresh.
2036  */
2037 static void
2038 EvalPlanQualStart(evalPlanQual *epq, EState *estate, evalPlanQual *priorepq)
2039 {
2040         EState     *epqstate;
2041         int                     rtsize;
2042         MemoryContext oldcontext;
2043
2044         rtsize = list_length(estate->es_range_table);
2045
2046         epq->estate = epqstate = CreateExecutorState();
2047
2048         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2049
2050         /*
2051          * The epqstates share the top query's copy of unchanging state such
2052          * as the snapshot, rangetable, result-rel info, and external Param
2053          * info. They need their own copies of local state, including a tuple
2054          * table, es_param_exec_vals, etc.
2055          */
2056         epqstate->es_direction = ForwardScanDirection;
2057         epqstate->es_snapshot = estate->es_snapshot;
2058         epqstate->es_crosscheck_snapshot = estate->es_crosscheck_snapshot;
2059         epqstate->es_range_table = estate->es_range_table;
2060         epqstate->es_result_relations = estate->es_result_relations;
2061         epqstate->es_num_result_relations = estate->es_num_result_relations;
2062         epqstate->es_result_relation_info = estate->es_result_relation_info;
2063         epqstate->es_junkFilter = estate->es_junkFilter;
2064         epqstate->es_into_relation_descriptor = estate->es_into_relation_descriptor;
2065         epqstate->es_param_list_info = estate->es_param_list_info;
2066         if (estate->es_topPlan->nParamExec > 0)
2067                 epqstate->es_param_exec_vals = (ParamExecData *)
2068                         palloc0(estate->es_topPlan->nParamExec * sizeof(ParamExecData));
2069         epqstate->es_rowMark = estate->es_rowMark;
2070         epqstate->es_instrument = estate->es_instrument;
2071         epqstate->es_select_into = estate->es_select_into;
2072         epqstate->es_into_oids = estate->es_into_oids;
2073         epqstate->es_topPlan = estate->es_topPlan;
2074
2075         /*
2076          * Each epqstate must have its own es_evTupleNull state, but all the
2077          * stack entries share es_evTuple state.  This allows sub-rechecks to
2078          * inherit the value being examined by an outer recheck.
2079          */
2080         epqstate->es_evTupleNull = (bool *) palloc0(rtsize * sizeof(bool));
2081         if (priorepq == NULL)
2082                 /* first PQ stack entry */
2083                 epqstate->es_evTuple = (HeapTuple *)
2084                         palloc0(rtsize * sizeof(HeapTuple));
2085         else
2086                 /* later stack entries share the same storage */
2087                 epqstate->es_evTuple = priorepq->estate->es_evTuple;
2088
2089         epqstate->es_tupleTable =
2090                 ExecCreateTupleTable(estate->es_tupleTable->size);
2091
2092         epq->planstate = ExecInitNode(estate->es_topPlan, epqstate);
2093
2094         MemoryContextSwitchTo(oldcontext);
2095 }
2096
2097 /*
2098  * End execution of one level of PlanQual.
2099  *
2100  * This is a cut-down version of ExecutorEnd(); basically we want to do most
2101  * of the normal cleanup, but *not* close result relations (which we are
2102  * just sharing from the outer query).
2103  */
2104 static void
2105 EvalPlanQualStop(evalPlanQual *epq)
2106 {
2107         EState     *epqstate = epq->estate;
2108         MemoryContext oldcontext;
2109
2110         oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
2111
2112         ExecEndNode(epq->planstate);
2113
2114         ExecDropTupleTable(epqstate->es_tupleTable, true);
2115         epqstate->es_tupleTable = NULL;
2116
2117         if (epqstate->es_evTuple[epq->rti - 1] != NULL)
2118         {
2119                 heap_freetuple(epqstate->es_evTuple[epq->rti - 1]);
2120                 epqstate->es_evTuple[epq->rti - 1] = NULL;
2121         }
2122
2123         MemoryContextSwitchTo(oldcontext);
2124
2125         FreeExecutorState(epqstate);
2126
2127         epq->estate = NULL;
2128         epq->planstate = NULL;
2129 }