OSDN Git Service

Fix bugs in relpersistence handling during table creation.
[pg-rex/syncrep.git] / src / backend / executor / execMain.c
index 17c7ea2..eacd863 100644 (file)
@@ -6,20 +6,25 @@
  * INTERFACE ROUTINES
  *     ExecutorStart()
  *     ExecutorRun()
+ *     ExecutorFinish()
  *     ExecutorEnd()
  *
- *     The old ExecutorMain() has been replaced by ExecutorStart(),
- *     ExecutorRun() and ExecutorEnd()
- *
- *     These three procedures are the external interfaces to the executor.
+ *     These four procedures are the external interface to the executor.
  *     In each case, the query descriptor is required as an argument.
  *
- *     ExecutorStart() must be called at the beginning of execution of any
- *     query plan and ExecutorEnd() should always be called at the end of
- *     execution of a plan.
+ *     ExecutorStart must be called at the beginning of execution of any
+ *     query plan and ExecutorEnd must always be called at the end of
+ *     execution of a plan (unless it is aborted due to error).
  *
  *     ExecutorRun accepts direction and count arguments that specify whether
  *     the plan is to be executed forwards, backwards, and for how many tuples.
+ *     In some cases ExecutorRun may be called multiple times to process all
+ *     the tuples for a plan.  It is also acceptable to stop short of executing
+ *     the whole plan (but only if it is a SELECT).
+ *
+ *     ExecutorFinish must be called after the final ExecutorRun call and
+ *     before ExecutorEnd.  This can be omitted only in case of EXPLAIN,
+ *     which should also omit ExecutorRun.
  *
  * Portions Copyright (c) 1996-2011, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
 #include "utils/tqual.h"
 
 
-/* Hooks for plugins to get control in ExecutorStart/Run/End() */
+/* Hooks for plugins to get control in ExecutorStart/Run/Finish/End */
 ExecutorStart_hook_type ExecutorStart_hook = NULL;
 ExecutorRun_hook_type ExecutorRun_hook = NULL;
+ExecutorFinish_hook_type ExecutorFinish_hook = NULL;
 ExecutorEnd_hook_type ExecutorEnd_hook = NULL;
 
 /* Hook for plugin to get control in ExecCheckRTPerms() */
@@ -68,6 +74,8 @@ ExecutorCheckPerms_hook_type ExecutorCheckPerms_hook = NULL;
 
 /* decls for local routines only used within this module */
 static void InitPlan(QueryDesc *queryDesc, int eflags);
+static void CheckValidRowMarkRel(Relation rel, RowMarkType markType);
+static void ExecPostprocessPlan(EState *estate);
 static void ExecEndPlan(PlanState *planstate, EState *estate);
 static void ExecutePlan(EState *estate, PlanState *planstate,
                        CmdType operation,
@@ -95,8 +103,8 @@ static void intorel_destroy(DestReceiver *self);
  *             This routine must be called at the beginning of any execution of any
  *             query plan
  *
- * Takes a QueryDesc previously created by CreateQueryDesc (it's not real
- * clear why we bother to separate the two functions, but...). The tupDesc
+ * Takes a QueryDesc previously created by CreateQueryDesc (which is separate
+ * only because some places use QueryDescs for utility commands).  The tupDesc
  * field of the QueryDesc is filled in to describe the tuples that will be
  * returned, and the internal fields (estate and planstate) are set up.
  *
@@ -161,10 +169,24 @@ standard_ExecutorStart(QueryDesc *queryDesc, int eflags)
        switch (queryDesc->operation)
        {
                case CMD_SELECT:
-                       /* SELECT INTO and SELECT FOR UPDATE/SHARE need to mark tuples */
+
+                       /*
+                        * SELECT INTO, SELECT FOR UPDATE/SHARE and modifying CTEs need to
+                        * mark tuples
+                        */
                        if (queryDesc->plannedstmt->intoClause != NULL ||
-                               queryDesc->plannedstmt->rowMarks != NIL)
+                               queryDesc->plannedstmt->rowMarks != NIL ||
+                               queryDesc->plannedstmt->hasModifyingCTE)
                                estate->es_output_cid = GetCurrentCommandId(true);
+
+                       /*
+                        * A SELECT without modifying CTEs can't possibly queue triggers,
+                        * so force skip-triggers mode. This is just a marginal efficiency
+                        * hack, since AfterTriggerBeginQuery/AfterTriggerEndQuery aren't
+                        * all that expensive, but we might as well do it.
+                        */
+                       if (!queryDesc->plannedstmt->hasModifyingCTE)
+                               eflags |= EXEC_FLAG_SKIP_TRIGGERS;
                        break;
 
                case CMD_INSERT:
@@ -184,6 +206,7 @@ standard_ExecutorStart(QueryDesc *queryDesc, int eflags)
         */
        estate->es_snapshot = RegisterSnapshot(queryDesc->snapshot);
        estate->es_crosscheck_snapshot = RegisterSnapshot(queryDesc->crosscheck_snapshot);
+       estate->es_top_eflags = eflags;
        estate->es_instrument = queryDesc->instrument_options;
 
        /*
@@ -191,6 +214,13 @@ standard_ExecutorStart(QueryDesc *queryDesc, int eflags)
         */
        InitPlan(queryDesc, eflags);
 
+       /*
+        * Set up an AFTER-trigger statement context, unless told not to, or
+        * unless it's EXPLAIN-only mode (when ExecutorFinish won't be called).
+        */
+       if (!(eflags & (EXEC_FLAG_SKIP_TRIGGERS | EXEC_FLAG_EXPLAIN_ONLY)))
+               AfterTriggerBeginQuery();
+
        MemoryContextSwitchTo(oldcontext);
 }
 
@@ -247,13 +277,14 @@ standard_ExecutorRun(QueryDesc *queryDesc,
        estate = queryDesc->estate;
 
        Assert(estate != NULL);
+       Assert(!(estate->es_top_eflags & EXEC_FLAG_EXPLAIN_ONLY));
 
        /*
         * Switch into per-query memory context
         */
        oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 
-       /* Allow instrumentation of ExecutorRun overall runtime */
+       /* Allow instrumentation of Executor overall runtime */
        if (queryDesc->totaltime)
                InstrStartNode(queryDesc->totaltime);
 
@@ -300,6 +331,68 @@ standard_ExecutorRun(QueryDesc *queryDesc,
 }
 
 /* ----------------------------------------------------------------
+ *             ExecutorFinish
+ *
+ *             This routine must be called after the last ExecutorRun call.
+ *             It performs cleanup such as firing AFTER triggers.      It is
+ *             separate from ExecutorEnd because EXPLAIN ANALYZE needs to
+ *             include these actions in the total runtime.
+ *
+ *             We provide a function hook variable that lets loadable plugins
+ *             get control when ExecutorFinish is called.      Such a plugin would
+ *             normally call standard_ExecutorFinish().
+ *
+ * ----------------------------------------------------------------
+ */
+void
+ExecutorFinish(QueryDesc *queryDesc)
+{
+       if (ExecutorFinish_hook)
+               (*ExecutorFinish_hook) (queryDesc);
+       else
+               standard_ExecutorFinish(queryDesc);
+}
+
+void
+standard_ExecutorFinish(QueryDesc *queryDesc)
+{
+       EState     *estate;
+       MemoryContext oldcontext;
+
+       /* sanity checks */
+       Assert(queryDesc != NULL);
+
+       estate = queryDesc->estate;
+
+       Assert(estate != NULL);
+       Assert(!(estate->es_top_eflags & EXEC_FLAG_EXPLAIN_ONLY));
+
+       /* This should be run once and only once per Executor instance */
+       Assert(!estate->es_finished);
+
+       /* Switch into per-query memory context */
+       oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
+
+       /* Allow instrumentation of Executor overall runtime */
+       if (queryDesc->totaltime)
+               InstrStartNode(queryDesc->totaltime);
+
+       /* Run ModifyTable nodes to completion */
+       ExecPostprocessPlan(estate);
+
+       /* Execute queued AFTER triggers, unless told not to */
+       if (!(estate->es_top_eflags & EXEC_FLAG_SKIP_TRIGGERS))
+               AfterTriggerEndQuery(estate);
+
+       if (queryDesc->totaltime)
+               InstrStopNode(queryDesc->totaltime, 0);
+
+       MemoryContextSwitchTo(oldcontext);
+
+       estate->es_finished = true;
+}
+
+/* ----------------------------------------------------------------
  *             ExecutorEnd
  *
  *             This routine must be called at the end of execution of any
@@ -334,6 +427,14 @@ standard_ExecutorEnd(QueryDesc *queryDesc)
        Assert(estate != NULL);
 
        /*
+        * Check that ExecutorFinish was called, unless in EXPLAIN-only mode. This
+        * Assert is needed because ExecutorFinish is new as of 9.1, and callers
+        * might forget to call it.
+        */
+       Assert(estate->es_finished ||
+                  (estate->es_top_eflags & EXEC_FLAG_EXPLAIN_ONLY));
+
+       /*
         * Switch into per-query memory context to run ExecEndPlan
         */
        oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
@@ -420,7 +521,7 @@ ExecCheckRTPerms(List *rangeTable, bool ereport_on_violation)
 
        foreach(l, rangeTable)
        {
-               RangeTblEntry  *rte = (RangeTblEntry *) lfirst(l);
+               RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 
                result = ExecCheckRTEPerms(rte);
                if (!result)
@@ -434,8 +535,8 @@ ExecCheckRTPerms(List *rangeTable, bool ereport_on_violation)
        }
 
        if (ExecutorCheckPerms_hook)
-               result = (*ExecutorCheckPerms_hook)(rangeTable,
-                                                                                       ereport_on_violation);
+               result = (*ExecutorCheckPerms_hook) (rangeTable,
+                                                                                        ereport_on_violation);
        return result;
 }
 
@@ -681,7 +782,6 @@ InitPlan(QueryDesc *queryDesc, int eflags)
                        InitResultRelInfo(resultRelInfo,
                                                          resultRelation,
                                                          resultRelationIndex,
-                                                         operation,
                                                          estate->es_instrument);
                        resultRelInfo++;
                }
@@ -738,15 +838,17 @@ InitPlan(QueryDesc *queryDesc, int eflags)
                                break;
                }
 
+               /* Check that relation is a legal target for marking */
+               if (relation)
+                       CheckValidRowMarkRel(relation, rc->markType);
+
                erm = (ExecRowMark *) palloc(sizeof(ExecRowMark));
                erm->relation = relation;
                erm->rti = rc->rti;
                erm->prti = rc->prti;
+               erm->rowmarkId = rc->rowmarkId;
                erm->markType = rc->markType;
                erm->noWait = rc->noWait;
-               erm->ctidAttNo = rc->ctidAttNo;
-               erm->toidAttNo = rc->toidAttNo;
-               erm->wholeAttNo = rc->wholeAttNo;
                ItemPointerSetInvalid(&(erm->curCtid));
                estate->es_rowMarks = lappend(estate->es_rowMarks, erm);
        }
@@ -868,24 +970,21 @@ InitPlan(QueryDesc *queryDesc, int eflags)
 }
 
 /*
- * Initialize ResultRelInfo data for one result relation
+ * Check that a proposed result relation is a legal target for the operation
+ *
+ * In most cases parser and/or planner should have noticed this already, but
+ * let's make sure.  In the view case we do need a test here, because if the
+ * view wasn't rewritten by a rule, it had better have an INSTEAD trigger.
+ *
+ * Note: when changing this function, you probably also need to look at
+ * CheckValidRowMarkRel.
  */
 void
-InitResultRelInfo(ResultRelInfo *resultRelInfo,
-                                 Relation resultRelationDesc,
-                                 Index resultRelationIndex,
-                                 CmdType operation,
-                                 int instrument_options)
+CheckValidResultRel(Relation resultRel, CmdType operation)
 {
-       TriggerDesc     *trigDesc = resultRelationDesc->trigdesc;
+       TriggerDesc *trigDesc = resultRel->trigdesc;
 
-       /*
-        * Check valid relkind ... in most cases parser and/or planner should have
-        * noticed this already, but let's make sure.  In the view case we do need
-        * a test here, because if the view wasn't rewritten by a rule, it had
-        * better have an INSTEAD trigger.
-        */
-       switch (resultRelationDesc->rd_rel->relkind)
+       switch (resultRel->rd_rel->relkind)
        {
                case RELKIND_RELATION:
                        /* OK */
@@ -894,13 +993,13 @@ InitResultRelInfo(ResultRelInfo *resultRelInfo,
                        ereport(ERROR,
                                        (errcode(ERRCODE_WRONG_OBJECT_TYPE),
                                         errmsg("cannot change sequence \"%s\"",
-                                                       RelationGetRelationName(resultRelationDesc))));
+                                                       RelationGetRelationName(resultRel))));
                        break;
                case RELKIND_TOASTVALUE:
                        ereport(ERROR,
                                        (errcode(ERRCODE_WRONG_OBJECT_TYPE),
                                         errmsg("cannot change TOAST relation \"%s\"",
-                                                       RelationGetRelationName(resultRelationDesc))));
+                                                       RelationGetRelationName(resultRel))));
                        break;
                case RELKIND_VIEW:
                        switch (operation)
@@ -908,26 +1007,26 @@ InitResultRelInfo(ResultRelInfo *resultRelInfo,
                                case CMD_INSERT:
                                        if (!trigDesc || !trigDesc->trig_insert_instead_row)
                                                ereport(ERROR,
-                                                               (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
-                                                                errmsg("cannot insert into view \"%s\"",
-                                                                               RelationGetRelationName(resultRelationDesc)),
-                                                                errhint("You need an unconditional ON INSERT DO INSTEAD rule or an INSTEAD OF INSERT trigger.")));
+                                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
+                                                  errmsg("cannot insert into view \"%s\"",
+                                                                 RelationGetRelationName(resultRel)),
+                                                  errhint("You need an unconditional ON INSERT DO INSTEAD rule or an INSTEAD OF INSERT trigger.")));
                                        break;
                                case CMD_UPDATE:
                                        if (!trigDesc || !trigDesc->trig_update_instead_row)
                                                ereport(ERROR,
-                                                               (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
-                                                                errmsg("cannot update view \"%s\"",
-                                                                               RelationGetRelationName(resultRelationDesc)),
-                                                                errhint("You need an unconditional ON UPDATE DO INSTEAD rule or an INSTEAD OF UPDATE trigger.")));
+                                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
+                                                  errmsg("cannot update view \"%s\"",
+                                                                 RelationGetRelationName(resultRel)),
+                                                  errhint("You need an unconditional ON UPDATE DO INSTEAD rule or an INSTEAD OF UPDATE trigger.")));
                                        break;
                                case CMD_DELETE:
                                        if (!trigDesc || !trigDesc->trig_delete_instead_row)
                                                ereport(ERROR,
-                                                               (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
-                                                                errmsg("cannot delete from view \"%s\"",
-                                                                               RelationGetRelationName(resultRelationDesc)),
-                                                                errhint("You need an unconditional ON DELETE DO INSTEAD rule or an INSTEAD OF DELETE trigger.")));
+                                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
+                                                  errmsg("cannot delete from view \"%s\"",
+                                                                 RelationGetRelationName(resultRel)),
+                                                  errhint("You need an unconditional ON DELETE DO INSTEAD rule or an INSTEAD OF DELETE trigger.")));
                                        break;
                                default:
                                        elog(ERROR, "unrecognized CmdType: %d", (int) operation);
@@ -938,17 +1037,81 @@ InitResultRelInfo(ResultRelInfo *resultRelInfo,
                        ereport(ERROR,
                                        (errcode(ERRCODE_WRONG_OBJECT_TYPE),
                                         errmsg("cannot change foreign table \"%s\"",
-                                                       RelationGetRelationName(resultRelationDesc))));
+                                                       RelationGetRelationName(resultRel))));
                        break;
                default:
                        ereport(ERROR,
                                        (errcode(ERRCODE_WRONG_OBJECT_TYPE),
                                         errmsg("cannot change relation \"%s\"",
-                                                       RelationGetRelationName(resultRelationDesc))));
+                                                       RelationGetRelationName(resultRel))));
                        break;
        }
+}
 
-       /* OK, fill in the node */
+/*
+ * Check that a proposed rowmark target relation is a legal target
+ *
+ * In most cases parser and/or planner should have noticed this already, but
+ * they don't cover all cases.
+ */
+static void
+CheckValidRowMarkRel(Relation rel, RowMarkType markType)
+{
+       switch (rel->rd_rel->relkind)
+       {
+               case RELKIND_RELATION:
+                       /* OK */
+                       break;
+               case RELKIND_SEQUENCE:
+                       /* Must disallow this because we don't vacuum sequences */
+                       ereport(ERROR,
+                                       (errcode(ERRCODE_WRONG_OBJECT_TYPE),
+                                        errmsg("cannot lock rows in sequence \"%s\"",
+                                                       RelationGetRelationName(rel))));
+                       break;
+               case RELKIND_TOASTVALUE:
+                       /* We could allow this, but there seems no good reason to */
+                       ereport(ERROR,
+                                       (errcode(ERRCODE_WRONG_OBJECT_TYPE),
+                                        errmsg("cannot lock rows in TOAST relation \"%s\"",
+                                                       RelationGetRelationName(rel))));
+                       break;
+               case RELKIND_VIEW:
+                       /* Should not get here */
+                       ereport(ERROR,
+                                       (errcode(ERRCODE_WRONG_OBJECT_TYPE),
+                                        errmsg("cannot lock rows in view \"%s\"",
+                                                       RelationGetRelationName(rel))));
+                       break;
+               case RELKIND_FOREIGN_TABLE:
+                       /* Perhaps we can support this someday, but not today */
+                       ereport(ERROR,
+                                       (errcode(ERRCODE_WRONG_OBJECT_TYPE),
+                                        errmsg("cannot lock rows in foreign table \"%s\"",
+                                                       RelationGetRelationName(rel))));
+                       break;
+               default:
+                       ereport(ERROR,
+                                       (errcode(ERRCODE_WRONG_OBJECT_TYPE),
+                                        errmsg("cannot lock rows in relation \"%s\"",
+                                                       RelationGetRelationName(rel))));
+                       break;
+       }
+}
+
+/*
+ * Initialize ResultRelInfo data for one result relation
+ *
+ * Caution: before Postgres 9.1, this function included the relkind checking
+ * that's now in CheckValidResultRel, and it also did ExecOpenIndices if
+ * appropriate.  Be sure callers cover those needs.
+ */
+void
+InitResultRelInfo(ResultRelInfo *resultRelInfo,
+                                 Relation resultRelationDesc,
+                                 Index resultRelationIndex,
+                                 int instrument_options)
+{
        MemSet(resultRelInfo, 0, sizeof(ResultRelInfo));
        resultRelInfo->type = T_ResultRelInfo;
        resultRelInfo->ri_RangeTableIndex = resultRelationIndex;
@@ -957,7 +1120,7 @@ InitResultRelInfo(ResultRelInfo *resultRelInfo,
        resultRelInfo->ri_IndexRelationDescs = NULL;
        resultRelInfo->ri_IndexRelationInfo = NULL;
        /* make a copy so as not to depend on relcache info not changing... */
-       resultRelInfo->ri_TrigDesc = CopyTriggerDesc(trigDesc);
+       resultRelInfo->ri_TrigDesc = CopyTriggerDesc(resultRelationDesc->trigdesc);
        if (resultRelInfo->ri_TrigDesc)
        {
                int                     n = resultRelInfo->ri_TrigDesc->numtriggers;
@@ -978,16 +1141,6 @@ InitResultRelInfo(ResultRelInfo *resultRelInfo,
        resultRelInfo->ri_ConstraintExprs = NULL;
        resultRelInfo->ri_junkFilter = NULL;
        resultRelInfo->ri_projectReturning = NULL;
-
-       /*
-        * If there are indices on the result relation, open them and save
-        * descriptors in the result relation info, so that we can add new index
-        * entries for the tuples we add/update.  We need not do this for a
-        * DELETE, however, since deletion doesn't affect indexes.
-        */
-       if (resultRelationDesc->rd_rel->relhasindex &&
-               operation != CMD_DELETE)
-               ExecOpenIndices(resultRelInfo);
 }
 
 /*
@@ -1037,26 +1190,29 @@ ExecGetTriggerResultRel(EState *estate, Oid relid)
        /*
         * Open the target relation's relcache entry.  We assume that an
         * appropriate lock is still held by the backend from whenever the trigger
-        * event got queued, so we need take no new lock here.
+        * event got queued, so we need take no new lock here.  Also, we need not
+        * recheck the relkind, so no need for CheckValidResultRel.
         */
        rel = heap_open(relid, NoLock);
 
        /*
-        * Make the new entry in the right context.  Currently, we don't need any
-        * index information in ResultRelInfos used only for triggers, so tell
-        * InitResultRelInfo it's a DELETE.
+        * Make the new entry in the right context.
         */
        oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
        rInfo = makeNode(ResultRelInfo);
        InitResultRelInfo(rInfo,
                                          rel,
                                          0,            /* dummy rangetable index */
-                                         CMD_DELETE,
                                          estate->es_instrument);
        estate->es_trig_target_relations =
                lappend(estate->es_trig_target_relations, rInfo);
        MemoryContextSwitchTo(oldcontext);
 
+       /*
+        * Currently, we don't need any index information in ResultRelInfos used
+        * only for triggers, so no need to call ExecOpenIndices.
+        */
+
        return rInfo;
 }
 
@@ -1118,6 +1274,46 @@ ExecContextForcesOids(PlanState *planstate, bool *hasoids)
 }
 
 /* ----------------------------------------------------------------
+ *             ExecPostprocessPlan
+ *
+ *             Give plan nodes a final chance to execute before shutdown
+ * ----------------------------------------------------------------
+ */
+static void
+ExecPostprocessPlan(EState *estate)
+{
+       ListCell   *lc;
+
+       /*
+        * Make sure nodes run forward.
+        */
+       estate->es_direction = ForwardScanDirection;
+
+       /*
+        * Run any secondary ModifyTable nodes to completion, in case the main
+        * query did not fetch all rows from them.      (We do this to ensure that
+        * such nodes have predictable results.)
+        */
+       foreach(lc, estate->es_auxmodifytables)
+       {
+               PlanState  *ps = (PlanState *) lfirst(lc);
+
+               for (;;)
+               {
+                       TupleTableSlot *slot;
+
+                       /* Reset the per-output-tuple exprcontext each time */
+                       ResetPerTupleExprContext(estate);
+
+                       slot = ExecProcNode(ps);
+
+                       if (TupIsNull(slot))
+                               break;
+               }
+       }
+}
+
+/* ----------------------------------------------------------------
  *             ExecEndPlan
  *
  *             Cleans up the query plan -- closes files and frees up storage
@@ -1387,6 +1583,77 @@ ExecConstraints(ResultRelInfo *resultRelInfo,
 
 
 /*
+ * ExecFindRowMark -- find the ExecRowMark struct for given rangetable index
+ */
+ExecRowMark *
+ExecFindRowMark(EState *estate, Index rti)
+{
+       ListCell   *lc;
+
+       foreach(lc, estate->es_rowMarks)
+       {
+               ExecRowMark *erm = (ExecRowMark *) lfirst(lc);
+
+               if (erm->rti == rti)
+                       return erm;
+       }
+       elog(ERROR, "failed to find ExecRowMark for rangetable index %u", rti);
+       return NULL;                            /* keep compiler quiet */
+}
+
+/*
+ * ExecBuildAuxRowMark -- create an ExecAuxRowMark struct
+ *
+ * Inputs are the underlying ExecRowMark struct and the targetlist of the
+ * input plan node (not planstate node!).  We need the latter to find out
+ * the column numbers of the resjunk columns.
+ */
+ExecAuxRowMark *
+ExecBuildAuxRowMark(ExecRowMark *erm, List *targetlist)
+{
+       ExecAuxRowMark *aerm = (ExecAuxRowMark *) palloc0(sizeof(ExecAuxRowMark));
+       char            resname[32];
+
+       aerm->rowmark = erm;
+
+       /* Look up the resjunk columns associated with this rowmark */
+       if (erm->relation)
+       {
+               Assert(erm->markType != ROW_MARK_COPY);
+
+               /* if child rel, need tableoid */
+               if (erm->rti != erm->prti)
+               {
+                       snprintf(resname, sizeof(resname), "tableoid%u", erm->rowmarkId);
+                       aerm->toidAttNo = ExecFindJunkAttributeInTlist(targetlist,
+                                                                                                                  resname);
+                       if (!AttributeNumberIsValid(aerm->toidAttNo))
+                               elog(ERROR, "could not find junk %s column", resname);
+               }
+
+               /* always need ctid for real relations */
+               snprintf(resname, sizeof(resname), "ctid%u", erm->rowmarkId);
+               aerm->ctidAttNo = ExecFindJunkAttributeInTlist(targetlist,
+                                                                                                          resname);
+               if (!AttributeNumberIsValid(aerm->ctidAttNo))
+                       elog(ERROR, "could not find junk %s column", resname);
+       }
+       else
+       {
+               Assert(erm->markType == ROW_MARK_COPY);
+
+               snprintf(resname, sizeof(resname), "wholerow%u", erm->rowmarkId);
+               aerm->wholeAttNo = ExecFindJunkAttributeInTlist(targetlist,
+                                                                                                               resname);
+               if (!AttributeNumberIsValid(aerm->wholeAttNo))
+                       elog(ERROR, "could not find junk %s column", resname);
+       }
+
+       return aerm;
+}
+
+
+/*
  * EvalPlanQual logic --- recheck modified tuple(s) to see if we want to
  * process the updated version under READ COMMITTED rules.
  *
@@ -1676,11 +1943,13 @@ EvalPlanQualFetch(EState *estate, Relation relation, int lockmode,
 /*
  * EvalPlanQualInit -- initialize during creation of a plan state node
  * that might need to invoke EPQ processing.
- * Note: subplan can be NULL if it will be set later with EvalPlanQualSetPlan.
+ *
+ * Note: subplan/auxrowmarks can be NULL/NIL if they will be set later
+ * with EvalPlanQualSetPlan.
  */
 void
 EvalPlanQualInit(EPQState *epqstate, EState *estate,
-                                Plan *subplan, int epqParam)
+                                Plan *subplan, List *auxrowmarks, int epqParam)
 {
        /* Mark the EPQ state inactive */
        epqstate->estate = NULL;
@@ -1688,7 +1957,7 @@ EvalPlanQualInit(EPQState *epqstate, EState *estate,
        epqstate->origslot = NULL;
        /* ... and remember data that EvalPlanQualBegin will need */
        epqstate->plan = subplan;
-       epqstate->rowMarks = NIL;
+       epqstate->arowMarks = auxrowmarks;
        epqstate->epqParam = epqParam;
 }
 
@@ -1698,25 +1967,14 @@ EvalPlanQualInit(EPQState *epqstate, EState *estate,
  * We need this so that ModifyTuple can deal with multiple subplans.
  */
 void
-EvalPlanQualSetPlan(EPQState *epqstate, Plan *subplan)
+EvalPlanQualSetPlan(EPQState *epqstate, Plan *subplan, List *auxrowmarks)
 {
        /* If we have a live EPQ query, shut it down */
        EvalPlanQualEnd(epqstate);
        /* And set/change the plan pointer */
        epqstate->plan = subplan;
-}
-
-/*
- * EvalPlanQualAddRowMark -- add an ExecRowMark that EPQ needs to handle.
- *
- * Currently, only non-locking RowMarks are supported.
- */
-void
-EvalPlanQualAddRowMark(EPQState *epqstate, ExecRowMark *erm)
-{
-       if (RowMarkRequiresRowShareLock(erm->markType))
-               elog(ERROR, "EvalPlanQual doesn't support locking rowmarks");
-       epqstate->rowMarks = lappend(epqstate->rowMarks, erm);
+       /* The rowmarks depend on the plan, too */
+       epqstate->arowMarks = auxrowmarks;
 }
 
 /*
@@ -1766,13 +2024,17 @@ EvalPlanQualFetchRowMarks(EPQState *epqstate)
 
        Assert(epqstate->origslot != NULL);
 
-       foreach(l, epqstate->rowMarks)
+       foreach(l, epqstate->arowMarks)
        {
-               ExecRowMark *erm = (ExecRowMark *) lfirst(l);
+               ExecAuxRowMark *aerm = (ExecAuxRowMark *) lfirst(l);
+               ExecRowMark *erm = aerm->rowmark;
                Datum           datum;
                bool            isNull;
                HeapTupleData tuple;
 
+               if (RowMarkRequiresRowShareLock(erm->markType))
+                       elog(ERROR, "EvalPlanQual doesn't support locking rowmarks");
+
                /* clear any leftover test tuple for this rel */
                EvalPlanQualSetTuple(epqstate, erm->rti, NULL);
 
@@ -1788,7 +2050,7 @@ EvalPlanQualFetchRowMarks(EPQState *epqstate)
                                Oid                     tableoid;
 
                                datum = ExecGetJunkAttribute(epqstate->origslot,
-                                                                                        erm->toidAttNo,
+                                                                                        aerm->toidAttNo,
                                                                                         &isNull);
                                /* non-locked rels could be on the inside of outer joins */
                                if (isNull)
@@ -1804,7 +2066,7 @@ EvalPlanQualFetchRowMarks(EPQState *epqstate)
 
                        /* fetch the tuple's ctid */
                        datum = ExecGetJunkAttribute(epqstate->origslot,
-                                                                                erm->ctidAttNo,
+                                                                                aerm->ctidAttNo,
                                                                                 &isNull);
                        /* non-locked rels could be on the inside of outer joins */
                        if (isNull)
@@ -1829,7 +2091,7 @@ EvalPlanQualFetchRowMarks(EPQState *epqstate)
 
                        /* fetch the whole-row Var for the relation */
                        datum = ExecGetJunkAttribute(epqstate->origslot,
-                                                                                erm->wholeAttNo,
+                                                                                aerm->wholeAttNo,
                                                                                 &isNull);
                        /* non-locked rels could be on the inside of outer joins */
                        if (isNull)
@@ -1952,9 +2214,11 @@ EvalPlanQualStart(EPQState *epqstate, EState *parentestate, Plan *planTree)
        estate->es_result_relation_info = parentestate->es_result_relation_info;
        /* es_trig_target_relations must NOT be copied */
        estate->es_rowMarks = parentestate->es_rowMarks;
+       estate->es_top_eflags = parentestate->es_top_eflags;
        estate->es_instrument = parentestate->es_instrument;
        estate->es_select_into = parentestate->es_select_into;
        estate->es_into_oids = parentestate->es_into_oids;
+       /* es_auxmodifytables must NOT be copied */
 
        /*
         * The external param list is simply shared from parent.  The internal
@@ -2009,7 +2273,11 @@ EvalPlanQualStart(EPQState *epqstate, EState *parentestate, Plan *planTree)
         * ExecInitSubPlan expects to be able to find these entries. Some of the
         * SubPlans might not be used in the part of the plan tree we intend to
         * run, but since it's not easy to tell which, we just initialize them
-        * all.
+        * all.  (However, if the subplan is headed by a ModifyTable node, then it
+        * must be a data-modifying CTE, which we will certainly not need to
+        * re-run, so we can skip initializing it.      This is just an efficiency
+        * hack; it won't skip data-modifying CTEs for which the ModifyTable node
+        * is not at the top.)
         */
        Assert(estate->es_subplanstates == NIL);
        foreach(l, parentestate->es_plannedstmt->subplans)
@@ -2017,7 +2285,11 @@ EvalPlanQualStart(EPQState *epqstate, EState *parentestate, Plan *planTree)
                Plan       *subplan = (Plan *) lfirst(l);
                PlanState  *subplanstate;
 
-               subplanstate = ExecInitNode(subplan, estate, 0);
+               /* Don't initialize ModifyTable subplans, per comment above */
+               if (IsA(subplan, ModifyTable))
+                       subplanstate = NULL;
+               else
+                       subplanstate = ExecInitNode(subplan, estate, 0);
 
                estate->es_subplanstates = lappend(estate->es_subplanstates,
                                                                                   subplanstate);
@@ -2121,9 +2393,7 @@ OpenIntoRel(QueryDesc *queryDesc)
        Oid                     namespaceId;
        Oid                     tablespaceId;
        Datum           reloptions;
-       AclResult       aclresult;
        Oid                     intoRelationId;
-       TupleDesc       tupdesc;
        DR_intorel *myState;
        static char *validnsps[] = HEAP_RELOPT_NAMESPACES;
 
@@ -2144,6 +2414,13 @@ OpenIntoRel(QueryDesc *queryDesc)
                                 errmsg("ON COMMIT can only be used on temporary tables")));
 
        /*
+        * Find namespace to create in, check its permissions
+        */
+       intoName = into->rel->relname;
+       namespaceId = RangeVarGetAndCheckCreationNamespace(into->rel);
+       RangeVarAdjustRelationPersistence(into->rel, namespaceId);
+
+       /*
         * Security check: disallow creating temp tables from security-restricted
         * code.  This is needed because calling code might not expect untrusted
         * tables to appear in pg_temp at the front of its search path.
@@ -2155,18 +2432,6 @@ OpenIntoRel(QueryDesc *queryDesc)
                                 errmsg("cannot create temporary table within security-restricted operation")));
 
        /*
-        * Find namespace to create in, check its permissions
-        */
-       intoName = into->rel->relname;
-       namespaceId = RangeVarGetCreationNamespace(into->rel);
-
-       aclresult = pg_namespace_aclcheck(namespaceId, GetUserId(),
-                                                                         ACL_CREATE);
-       if (aclresult != ACLCHECK_OK)
-               aclcheck_error(aclresult, ACL_KIND_NAMESPACE,
-                                          get_namespace_name(namespaceId));
-
-       /*
         * Select tablespace to use.  If not specified, use default tablespace
         * (which may in turn default to database's default).
         */
@@ -2202,9 +2467,6 @@ OpenIntoRel(QueryDesc *queryDesc)
                                                                         false);
        (void) heap_reloptions(RELKIND_RELATION, reloptions, true);
 
-       /* Copy the tupdesc because heap_create_with_catalog modifies it */
-       tupdesc = CreateTupleDescCopy(queryDesc->tupDesc);
-
        /* Now we can actually create the new relation */
        intoRelationId = heap_create_with_catalog(intoName,
                                                                                          namespaceId,
@@ -2213,7 +2475,7 @@ OpenIntoRel(QueryDesc *queryDesc)
                                                                                          InvalidOid,
                                                                                          InvalidOid,
                                                                                          GetUserId(),
-                                                                                         tupdesc,
+                                                                                         queryDesc->tupDesc,
                                                                                          NIL,
                                                                                          RELKIND_RELATION,
                                                                                          into->rel->relpersistence,
@@ -2224,12 +2486,9 @@ OpenIntoRel(QueryDesc *queryDesc)
                                                                                          into->onCommit,
                                                                                          reloptions,
                                                                                          true,
-                                                                                         allowSystemTableMods,
-                                                                                         false);
+                                                                                         allowSystemTableMods);
        Assert(intoRelationId != InvalidOid);
 
-       FreeTupleDesc(tupdesc);
-
        /*
         * Advance command counter so that the newly-created relation's catalog
         * tuples will be visible to heap_open.