]> granicus.if.org Git - postgresql/blobdiff - src/backend/executor/execMain.c
Prevent adding relations to a concurrently dropped schema.
[postgresql] / src / backend / executor / execMain.c
index d03ad0941848b2d4129c3a210b9cb0692d06d485..422f737e82dcb1202d3f9e927fe87f06f6b33c7b 100644 (file)
@@ -6,27 +6,32 @@
  * 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- *     The old ExecutorMain() has been replaced by ExecutorStart(),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) 1996-2009, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1996-2012, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
  *
  * IDENTIFICATION
- *       $PostgreSQL: pgsql/src/backend/executor/execMain.c,v 1.333 2009/10/12 18:10:41 tgl Exp $
+ *       src/backend/executor/execMain.c
  *
  *-------------------------------------------------------------------------
  */
 #include "commands/tablespace.h"
 #include "commands/trigger.h"
 #include "executor/execdebug.h"
-#include "executor/instrument.h"
+#include "mb/pg_wchar.h"
 #include "miscadmin.h"
 #include "optimizer/clauses.h"
 #include "parser/parse_clause.h"
 #include "parser/parsetree.h"
 #include "storage/bufmgr.h"
 #include "storage/lmgr.h"
+#include "storage/smgr.h"
+#include "tcop/utility.h"
 #include "utils/acl.h"
+#include "utils/builtins.h"
 #include "utils/lsyscache.h"
 #include "utils/memutils.h"
 #include "utils/snapmgr.h"
 #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;
 
-typedef struct evalPlanQual
-{
-       Index           rti;
-       EState     *estate;
-       PlanState  *planstate;
-       PlanState  *origplanstate;
-       TupleTableSlot *resultslot;
-       struct evalPlanQual *next;      /* stack of active PlanQual plans */
-       struct evalPlanQual *free;      /* list of free PlanQual plans */
-} evalPlanQual;
+/* Hook for plugin to get control in ExecCheckRTPerms() */
+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,
@@ -81,13 +84,12 @@ static void ExecutePlan(EState *estate, PlanState *planstate,
                        long numberTuples,
                        ScanDirection direction,
                        DestReceiver *dest);
-static void EndEvalPlanQual(EState *estate);
-static void ExecCheckRTPerms(List *rangeTable);
-static void ExecCheckRTEPerms(RangeTblEntry *rte);
+static bool ExecCheckRTEPerms(RangeTblEntry *rte);
 static void ExecCheckXactReadOnly(PlannedStmt *plannedstmt);
-static void EvalPlanQualStart(evalPlanQual *epq, EState *estate,
-                                                         Plan *planTree, evalPlanQual *priorepq);
-static void EvalPlanQualStop(evalPlanQual *epq);
+static char *ExecBuildSlotValueDescription(TupleTableSlot *slot,
+                                                                                  int maxfieldlen);
+static void EvalPlanQualStart(EPQState *epqstate, EState *parentestate,
+                                 Plan *planTree);
 static void OpenIntoRel(QueryDesc *queryDesc);
 static void CloseIntoRel(QueryDesc *queryDesc);
 static void intorel_startup(DestReceiver *self, int operation, TupleDesc typeinfo);
@@ -104,8 +106,8 @@ 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.
  *
@@ -155,7 +157,8 @@ standard_ExecutorStart(QueryDesc *queryDesc, int eflags)
        oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 
        /*
-        * Fill in parameters, if any, from queryDesc
+        * Fill in external parameters, if any, from queryDesc; and allocate
+        * workspace for internal parameters
         */
        estate->es_param_list_info = queryDesc->params;
 
@@ -169,10 +172,24 @@ 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:
@@ -192,13 +209,21 @@ standard_ExecutorStart(QueryDesc *queryDesc, int eflags)
         */
        estate->es_snapshot = RegisterSnapshot(queryDesc->snapshot);
        estate->es_crosscheck_snapshot = RegisterSnapshot(queryDesc->crosscheck_snapshot);
-       estate->es_instrument = queryDesc->doInstrument;
+       estate->es_top_eflags = eflags;
+       estate->es_instrument = queryDesc->instrument_options;
 
        /*
         * Initialize the plan state tree
         */
        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);
 }
 
@@ -255,13 +280,14 @@ 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);
 
@@ -283,6 +309,13 @@ standard_ExecutorRun(QueryDesc *queryDesc,
        if (sendTuples)
                (*dest->rStartup) (dest, operation, queryDesc->tupDesc);
 
+       /*
+        * if it's CREATE TABLE AS ... WITH NO DATA, skip plan execution
+        */
+       if (estate->es_select_into &&
+               queryDesc->plannedstmt->intoClause->skipData)
+               direction = NoMovementScanDirection;
+
        /*
         * run plan
         */
@@ -307,6 +340,68 @@ standard_ExecutorRun(QueryDesc *queryDesc,
        MemoryContextSwitchTo(oldcontext);
 }
 
+/* ----------------------------------------------------------------
+ *             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
  *
@@ -341,6 +436,14 @@ 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
         */
@@ -407,7 +510,7 @@ ExecutorRewind(QueryDesc *queryDesc)
        /*
         * rescan plan
         */
-       ExecReScan(queryDesc->planstate, NULL);
+       ExecReScan(queryDesc->planstate);
 
        MemoryContextSwitchTo(oldcontext);
 }
@@ -416,23 +519,42 @@ ExecutorRewind(QueryDesc *queryDesc)
 /*
  * ExecCheckRTPerms
  *             Check access permissions for all relations listed in a range table.
+ *
+ * Returns true if permissions are adequate.  Otherwise, throws an appropriate
+ * error if ereport_on_violation is true, or simply returns false otherwise.
  */
-static void
-ExecCheckRTPerms(List *rangeTable)
+bool
+ExecCheckRTPerms(List *rangeTable, bool ereport_on_violation)
 {
        ListCell   *l;
+       bool            result = true;
 
        foreach(l, rangeTable)
        {
-               ExecCheckRTEPerms((RangeTblEntry *) lfirst(l));
+               RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
+
+               result = ExecCheckRTEPerms(rte);
+               if (!result)
+               {
+                       Assert(rte->rtekind == RTE_RELATION);
+                       if (ereport_on_violation)
+                               aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
+                                                          get_rel_name(rte->relid));
+                       return false;
+               }
        }
+
+       if (ExecutorCheckPerms_hook)
+               result = (*ExecutorCheckPerms_hook) (rangeTable,
+                                                                                        ereport_on_violation);
+       return result;
 }
 
 /*
  * ExecCheckRTEPerms
  *             Check access permissions for a single RTE.
  */
-static void
+static bool
 ExecCheckRTEPerms(RangeTblEntry *rte)
 {
        AclMode         requiredPerms;
@@ -449,14 +571,14 @@ ExecCheckRTEPerms(RangeTblEntry *rte)
         * Join, subquery, and special RTEs need no checks.
         */
        if (rte->rtekind != RTE_RELATION)
-               return;
+               return true;
 
        /*
         * No work if requiredPerms is empty.
         */
        requiredPerms = rte->requiredPerms;
        if (requiredPerms == 0)
-               return;
+               return true;
 
        relOid = rte->relid;
 
@@ -484,8 +606,7 @@ ExecCheckRTEPerms(RangeTblEntry *rte)
                 * we can fail straight away.
                 */
                if (remainingPerms & ~(ACL_SELECT | ACL_INSERT | ACL_UPDATE))
-                       aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
-                                                  get_rel_name(relOid));
+                       return false;
 
                /*
                 * Check to see if we have the needed privileges at column level.
@@ -505,8 +626,7 @@ ExecCheckRTEPerms(RangeTblEntry *rte)
                        {
                                if (pg_attribute_aclcheck_all(relOid, userid, ACL_SELECT,
                                                                                          ACLMASK_ANY) != ACLCHECK_OK)
-                                       aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
-                                                                  get_rel_name(relOid));
+                                       return false;
                        }
 
                        tmpset = bms_copy(rte->selectedCols);
@@ -519,15 +639,13 @@ ExecCheckRTEPerms(RangeTblEntry *rte)
                                        /* Whole-row reference, must have priv on all cols */
                                        if (pg_attribute_aclcheck_all(relOid, userid, ACL_SELECT,
                                                                                                  ACLMASK_ALL) != ACLCHECK_OK)
-                                               aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
-                                                                          get_rel_name(relOid));
+                                               return false;
                                }
                                else
                                {
-                                       if (pg_attribute_aclcheck(relOid, col, userid, ACL_SELECT)
-                                               != ACLCHECK_OK)
-                                               aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
-                                                                          get_rel_name(relOid));
+                                       if (pg_attribute_aclcheck(relOid, col, userid,
+                                                                                         ACL_SELECT) != ACLCHECK_OK)
+                                               return false;
                                }
                        }
                        bms_free(tmpset);
@@ -550,8 +668,7 @@ ExecCheckRTEPerms(RangeTblEntry *rte)
                        {
                                if (pg_attribute_aclcheck_all(relOid, userid, remainingPerms,
                                                                                          ACLMASK_ANY) != ACLCHECK_OK)
-                                       aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
-                                                                  get_rel_name(relOid));
+                                       return false;
                        }
 
                        tmpset = bms_copy(rte->modifiedCols);
@@ -566,19 +683,23 @@ ExecCheckRTEPerms(RangeTblEntry *rte)
                                }
                                else
                                {
-                                       if (pg_attribute_aclcheck(relOid, col, userid, remainingPerms)
-                                               != ACLCHECK_OK)
-                                               aclcheck_error(ACLCHECK_NO_PRIV, ACL_KIND_CLASS,
-                                                                          get_rel_name(relOid));
+                                       if (pg_attribute_aclcheck(relOid, col, userid,
+                                                                                         remainingPerms) != ACLCHECK_OK)
+                                               return false;
                                }
                        }
                        bms_free(tmpset);
                }
        }
+       return true;
 }
 
 /*
  * Check that the query does not imply any writes to non-temp tables.
+ *
+ * Note: in a Hot Standby slave this would need to reject writes to temp
+ * tables as well; but an HS slave can't have created any temp tables
+ * in the first place, so no need to check that.
  */
 static void
 ExecCheckXactReadOnly(PlannedStmt *plannedstmt)
@@ -588,10 +709,11 @@ ExecCheckXactReadOnly(PlannedStmt *plannedstmt)
        /*
         * CREATE TABLE AS or SELECT INTO?
         *
-        * XXX should we allow this if the destination is temp?
+        * XXX should we allow this if the destination is temp?  Considering that
+        * it would still require catalog changes, probably not.
         */
        if (plannedstmt->intoClause != NULL)
-               goto fail;
+               PreventCommandIfReadOnly(CreateCommandTag((Node *) plannedstmt));
 
        /* Fail if write permissions are requested on any non-temp table */
        foreach(l, plannedstmt->rtable)
@@ -607,15 +729,8 @@ ExecCheckXactReadOnly(PlannedStmt *plannedstmt)
                if (isTempNamespace(get_rel_namespace(rte->relid)))
                        continue;
 
-               goto fail;
+               PreventCommandIfReadOnly(CreateCommandTag((Node *) plannedstmt));
        }
-
-       return;
-
-fail:
-       ereport(ERROR,
-                       (errcode(ERRCODE_READ_ONLY_SQL_TRANSACTION),
-                        errmsg("transaction is read-only")));
 }
 
 
@@ -642,18 +757,19 @@ InitPlan(QueryDesc *queryDesc, int eflags)
        /*
         * Do permissions checks
         */
-       ExecCheckRTPerms(rangeTable);
+       ExecCheckRTPerms(rangeTable, true);
 
        /*
         * initialize the node's execution state
         */
        estate->es_range_table = rangeTable;
+       estate->es_plannedstmt = plannedstmt;
 
        /*
         * initialize result relation stuff, and open/lock the result rels.
         *
-        * We must do this before initializing the plan tree, else we might
-        * try to do a lock upgrade if a result rel is also a source rel.
+        * We must do this before initializing the plan tree, else we might try to
+        * do a lock upgrade if a result rel is also a source rel.
         */
        if (plannedstmt->resultRelations)
        {
@@ -676,7 +792,6 @@ InitPlan(QueryDesc *queryDesc, int eflags)
                        InitResultRelInfo(resultRelInfo,
                                                          resultRelation,
                                                          resultRelationIndex,
-                                                         operation,
                                                          estate->es_instrument);
                        resultRelInfo++;
                }
@@ -697,13 +812,13 @@ InitPlan(QueryDesc *queryDesc, int eflags)
 
        /*
         * Similarly, we have to lock relations selected FOR UPDATE/FOR SHARE
-        * before we initialize the plan tree, else we'd be risking lock
-        * upgrades.  While we are at it, build the ExecRowMark list.
+        * before we initialize the plan tree, else we'd be risking lock upgrades.
+        * While we are at it, build the ExecRowMark list.
         */
        estate->es_rowMarks = NIL;
        foreach(l, plannedstmt->rowMarks)
        {
-               RowMarkClause *rc = (RowMarkClause *) lfirst(l);
+               PlanRowMark *rc = (PlanRowMark *) lfirst(l);
                Oid                     relid;
                Relation        relation;
                ExecRowMark *erm;
@@ -712,18 +827,38 @@ InitPlan(QueryDesc *queryDesc, int eflags)
                if (rc->isParent)
                        continue;
 
-               relid = getrelid(rc->rti, rangeTable);
-               relation = heap_open(relid, RowShareLock);
+               switch (rc->markType)
+               {
+                       case ROW_MARK_EXCLUSIVE:
+                       case ROW_MARK_SHARE:
+                               relid = getrelid(rc->rti, rangeTable);
+                               relation = heap_open(relid, RowShareLock);
+                               break;
+                       case ROW_MARK_REFERENCE:
+                               relid = getrelid(rc->rti, rangeTable);
+                               relation = heap_open(relid, AccessShareLock);
+                               break;
+                       case ROW_MARK_COPY:
+                               /* there's no real table here ... */
+                               relation = NULL;
+                               break;
+                       default:
+                               elog(ERROR, "unrecognized markType: %d", rc->markType);
+                               relation = NULL;        /* keep compiler quiet */
+                               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->forUpdate = rc->forUpdate;
+               erm->markType = rc->markType;
                erm->noWait = rc->noWait;
-               /* remaining fields are filled during LockRows plan node init */
-               erm->ctidAttNo = InvalidAttrNumber;
-               erm->toidAttNo = InvalidAttrNumber;
                ItemPointerSetInvalid(&(erm->curCtid));
                estate->es_rowMarks = lappend(estate->es_rowMarks, erm);
        }
@@ -745,12 +880,13 @@ InitPlan(QueryDesc *queryDesc, int eflags)
         */
        estate->es_tupleTable = NIL;
        estate->es_trig_tuple_slot = NULL;
+       estate->es_trig_oldtup_slot = NULL;
+       estate->es_trig_newtup_slot = NULL;
 
        /* mark EvalPlanQual not active */
-       estate->es_plannedstmt = plannedstmt;
-       estate->es_evalPlanQual = NULL;
-       estate->es_evTupleNull = NULL;
-       estate->es_evTuple = NULL;
+       estate->es_epqTuple = NULL;
+       estate->es_epqTupleSet = NULL;
+       estate->es_epqScanDone = NULL;
 
        /*
         * Initialize private state information for each SubPlan.  We must do this
@@ -797,8 +933,8 @@ InitPlan(QueryDesc *queryDesc, int eflags)
        tupType = ExecGetResultType(planstate);
 
        /*
-        * Initialize the junk filter if needed.  SELECT queries need a
-        * filter if there are any junk attrs in the top-level tlist.
+        * Initialize the junk filter if needed.  SELECT queries need a filter if
+        * there are any junk attrs in the top-level tlist.
         */
        if (operation == CMD_SELECT)
        {
@@ -845,20 +981,21 @@ 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,
-                                 bool doInstrument)
+CheckValidResultRel(Relation resultRel, CmdType operation)
 {
-       /*
-        * Check valid relkind ... parser and/or planner should have noticed this
-        * already, but let's make sure.
-        */
-       switch (resultRelationDesc->rd_rel->relkind)
+       TriggerDesc *trigDesc = resultRel->trigdesc;
+
+       switch (resultRel->rd_rel->relkind)
        {
                case RELKIND_RELATION:
                        /* OK */
@@ -867,29 +1004,125 @@ 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)
+                       {
+                               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(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(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(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);
+                                       break;
+                       }
+                       break;
+               case RELKIND_FOREIGN_TABLE:
                        ereport(ERROR,
                                        (errcode(ERRCODE_WRONG_OBJECT_TYPE),
-                                        errmsg("cannot change view \"%s\"",
-                                                       RelationGetRelationName(resultRelationDesc))));
+                                        errmsg("cannot change foreign table \"%s\"",
+                                                       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;
@@ -905,29 +1138,20 @@ InitResultRelInfo(ResultRelInfo *resultRelInfo,
 
                resultRelInfo->ri_TrigFunctions = (FmgrInfo *)
                        palloc0(n * sizeof(FmgrInfo));
-               if (doInstrument)
-                       resultRelInfo->ri_TrigInstrument = InstrAlloc(n);
-               else
-                       resultRelInfo->ri_TrigInstrument = NULL;
+               resultRelInfo->ri_TrigWhenExprs = (List **)
+                       palloc0(n * sizeof(List *));
+               if (instrument_options)
+                       resultRelInfo->ri_TrigInstrument = InstrAlloc(n, instrument_options);
        }
        else
        {
                resultRelInfo->ri_TrigFunctions = NULL;
+               resultRelInfo->ri_TrigWhenExprs = NULL;
                resultRelInfo->ri_TrigInstrument = NULL;
        }
        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);
 }
 
 /*
@@ -977,26 +1201,29 @@ 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;
 }
 
@@ -1057,6 +1284,46 @@ ExecContextForcesOids(PlanState *planstate, bool *hasoids)
        return false;
 }
 
+/* ----------------------------------------------------------------
+ *             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
  *
@@ -1076,12 +1343,6 @@ ExecEndPlan(PlanState *planstate, EState *estate)
        int                     i;
        ListCell   *l;
 
-       /*
-        * shut down any PlanQual processing we were doing
-        */
-       if (estate->es_evalPlanQual != NULL)
-               EndEvalPlanQual(estate);
-
        /*
         * shut down the node-type-specific query processing
         */
@@ -1099,9 +1360,9 @@ ExecEndPlan(PlanState *planstate, EState *estate)
 
        /*
         * destroy the executor's tuple table.  Actually we only care about
-        * releasing buffer pins and tupdesc refcounts; there's no need to
-        * pfree the TupleTableSlots, since the containing memory context
-        * is about to go away anyway.
+        * releasing buffer pins and tupdesc refcounts; there's no need to pfree
+        * the TupleTableSlots, since the containing memory context is about to go
+        * away anyway.
         */
        ExecResetTupleTable(estate->es_tupleTable, false);
 
@@ -1133,9 +1394,10 @@ ExecEndPlan(PlanState *planstate, EState *estate)
         */
        foreach(l, estate->es_rowMarks)
        {
-               ExecRowMark *erm = lfirst(l);
+               ExecRowMark *erm = (ExecRowMark *) lfirst(l);
 
-               heap_close(erm->relation, NoLock);
+               if (erm->relation)
+                       heap_close(erm->relation, NoLock);
        }
 }
 
@@ -1205,8 +1467,8 @@ ExecutePlan(EState *estate,
                        slot = ExecFilterJunk(estate->es_junkFilter, slot);
 
                /*
-                * If we are supposed to send the tuple somewhere, do so.
-                * (In practice, this is probably always the case at this point.)
+                * If we are supposed to send the tuple somewhere, do so. (In
+                * practice, this is probably always the case at this point.)
                 */
                if (sendTuples)
                        (*dest->receiveSlot) (slot, dest);
@@ -1314,7 +1576,9 @@ ExecConstraints(ResultRelInfo *resultRelInfo,
                                ereport(ERROR,
                                                (errcode(ERRCODE_NOT_NULL_VIOLATION),
                                                 errmsg("null value in column \"%s\" violates not-null constraint",
-                                               NameStr(rel->rd_att->attrs[attrChk - 1]->attname))));
+                                               NameStr(rel->rd_att->attrs[attrChk - 1]->attname)),
+                                                errdetail("Failing row contains %s.",
+                                                                  ExecBuildSlotValueDescription(slot, 64))));
                }
        }
 
@@ -1326,19 +1590,159 @@ ExecConstraints(ResultRelInfo *resultRelInfo,
                        ereport(ERROR,
                                        (errcode(ERRCODE_CHECK_VIOLATION),
                                         errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
-                                                       RelationGetRelationName(rel), failed)));
+                                                       RelationGetRelationName(rel), failed),
+                                        errdetail("Failing row contains %s.",
+                                                          ExecBuildSlotValueDescription(slot, 64))));
        }
 }
 
 /*
- * Check a modified tuple to see if we want to process its updated version
- * under READ COMMITTED rules.
+ * ExecBuildSlotValueDescription -- construct a string representing a tuple
+ *
+ * This is intentionally very similar to BuildIndexValueDescription, but
+ * unlike that function, we truncate long field values.  That seems necessary
+ * here since heap field values could be very long, whereas index entries
+ * typically aren't so wide.
+ */
+static char *
+ExecBuildSlotValueDescription(TupleTableSlot *slot, int maxfieldlen)
+{
+       StringInfoData buf;
+       TupleDesc       tupdesc = slot->tts_tupleDescriptor;
+       int                     i;
+
+       /* Make sure the tuple is fully deconstructed */
+       slot_getallattrs(slot);
+
+       initStringInfo(&buf);
+
+       appendStringInfoChar(&buf, '(');
+
+       for (i = 0; i < tupdesc->natts; i++)
+       {
+               char       *val;
+               int                     vallen;
+
+               if (slot->tts_isnull[i])
+                       val = "null";
+               else
+               {
+                       Oid                     foutoid;
+                       bool            typisvarlena;
+
+                       getTypeOutputInfo(tupdesc->attrs[i]->atttypid,
+                                                         &foutoid, &typisvarlena);
+                       val = OidOutputFunctionCall(foutoid, slot->tts_values[i]);
+               }
+
+               if (i > 0)
+                       appendStringInfoString(&buf, ", ");
+
+               /* truncate if needed */
+               vallen = strlen(val);
+               if (vallen <= maxfieldlen)
+                       appendStringInfoString(&buf, val);
+               else
+               {
+                       vallen = pg_mbcliplen(val, vallen, maxfieldlen);
+                       appendBinaryStringInfo(&buf, val, vallen);
+                       appendStringInfoString(&buf, "...");
+               }
+       }
+
+       appendStringInfoChar(&buf, ')');
+
+       return buf.data;
+}
+
+
+/*
+ * 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.
  *
  * See backend/executor/README for some info about how this works.
+ */
+
+
+/*
+ * Check a modified tuple to see if we want to process its updated version
+ * under READ COMMITTED rules.
  *
- *     estate - executor state data
+ *     estate - outer executor state data
+ *     epqstate - state for EvalPlanQual rechecking
+ *     relation - table containing tuple
  *     rti - rangetable index of table containing tuple
- *     subplanstate - portion of plan tree that needs to be re-evaluated
  *     *tid - t_ctid from the outdated tuple (ie, next updated version)
  *     priorXmax - t_xmax from the outdated tuple
  *
@@ -1349,19 +1753,20 @@ ExecConstraints(ResultRelInfo *resultRelInfo,
  * NULL if we determine we shouldn't process the row.
  */
 TupleTableSlot *
-EvalPlanQual(EState *estate, Index rti,
-                        PlanState *subplanstate,
+EvalPlanQual(EState *estate, EPQState *epqstate,
+                        Relation relation, Index rti,
                         ItemPointer tid, TransactionId priorXmax)
 {
        TupleTableSlot *slot;
        HeapTuple       copyTuple;
 
-       Assert(rti != 0);
+       Assert(rti > 0);
 
        /*
-        * Get the updated version of the row; if fail, return NULL.
+        * Get and lock the updated version of the row; if fail, return NULL.
         */
-       copyTuple = EvalPlanQualFetch(estate, rti, tid, priorXmax);
+       copyTuple = EvalPlanQualFetch(estate, relation, LockTupleExclusive,
+                                                                 tid, priorXmax);
 
        if (copyTuple == NULL)
                return NULL;
@@ -1373,52 +1778,42 @@ EvalPlanQual(EState *estate, Index rti,
        *tid = copyTuple->t_self;
 
        /*
-        * Need to run a recheck subquery.      Find or create a PQ stack entry.
+        * Need to run a recheck subquery.      Initialize or reinitialize EPQ state.
         */
-       EvalPlanQualPush(estate, rti, subplanstate);
+       EvalPlanQualBegin(epqstate, estate);
 
        /*
-        * free old RTE' tuple, if any, and store target tuple where relation's
-        * scan node will see it
+        * Free old test tuple, if any, and store new tuple where relation's scan
+        * node will see it
         */
-       EvalPlanQualSetTuple(estate, rti, copyTuple);
+       EvalPlanQualSetTuple(epqstate, rti, copyTuple);
 
        /*
-        * Run the EPQ query, but just for one tuple.
+        * Fetch any non-locked source rows
         */
-       slot = EvalPlanQualNext(estate);
+       EvalPlanQualFetchRowMarks(epqstate);
 
        /*
-        * If we got a result, we must copy it out of the EPQ query's local
-        * context before we shut down the EPQ query.
+        * Run the EPQ query.  We assume it will return at most one tuple.
         */
-       if (TupIsNull(slot))
-               slot = NULL;                    /* in case we got back an empty slot */
-       else
-       {
-               TupleDesc tupdesc = CreateTupleDescCopy(slot->tts_tupleDescriptor);
-               evalPlanQual *epq = estate->es_evalPlanQual;
-
-               if (epq->resultslot == NULL)
-               {
-                       epq->resultslot = ExecInitExtraTupleSlot(estate);
-                       ExecSetSlotDescriptor(epq->resultslot, tupdesc);
-               }
-               else
-               {
-                       TupleDesc oldtupdesc = epq->resultslot->tts_tupleDescriptor;
-
-                       ExecSetSlotDescriptor(epq->resultslot, tupdesc);
-                       FreeTupleDesc(oldtupdesc);
-               }
+       slot = EvalPlanQualNext(epqstate);
 
-               slot = ExecCopySlot(epq->resultslot, slot);
-       }
+       /*
+        * If we got a tuple, force the slot to materialize the tuple so that it
+        * is not dependent on any local state in the EPQ query (in particular,
+        * it's highly likely that the slot contains references to any pass-by-ref
+        * datums that may be present in copyTuple).  As with the next step, this
+        * is to guard against early re-use of the EPQ query.
+        */
+       if (!TupIsNull(slot))
+               (void) ExecMaterializeSlot(slot);
 
        /*
-        * Shut it down ...
+        * Clear out the test tuple.  This is needed in case the EPQ query is
+        * re-used to test a tuple for a different relation.  (Not clear that can
+        * really happen, but let's be safe.)
         */
-       EvalPlanQualPop(estate, subplanstate);
+       EvalPlanQualSetTuple(epqstate, rti, NULL);
 
        return slot;
 }
@@ -1427,55 +1822,29 @@ EvalPlanQual(EState *estate, Index rti,
  * Fetch a copy of the newest version of an outdated tuple
  *
  *     estate - executor state data
- *     rti - rangetable index of table containing tuple
+ *     relation - table containing tuple
+ *     lockmode - requested tuple lock mode
  *     *tid - t_ctid from the outdated tuple (ie, next updated version)
  *     priorXmax - t_xmax from the outdated tuple
  *
  * Returns a palloc'd copy of the newest tuple version, or NULL if we find
  * that there is no newest version (ie, the row was deleted not updated).
+ * If successful, we have locked the newest tuple version, so caller does not
+ * need to worry about it changing anymore.
  *
- * XXX this does not lock the new row version ... wouldn't it be better if
- * it did?  As-is, caller might have to repeat all its work.
+ * Note: properly, lockmode should be declared as enum LockTupleMode,
+ * but we use "int" to avoid having to include heapam.h in executor.h.
  */
 HeapTuple
-EvalPlanQualFetch(EState *estate, Index rti,
+EvalPlanQualFetch(EState *estate, Relation relation, int lockmode,
                                  ItemPointer tid, TransactionId priorXmax)
 {
        HeapTuple       copyTuple = NULL;
-       Relation        relation;
        HeapTupleData tuple;
        SnapshotData SnapshotDirty;
 
-       Assert(rti != 0);
-
-       /*
-        * Find relation containing target tuple --- must be either a result
-        * relation of the query, or a SELECT FOR UPDATE target
-        */
-       if (estate->es_result_relation_info != NULL &&
-               estate->es_result_relation_info->ri_RangeTableIndex == rti)
-               relation = estate->es_result_relation_info->ri_RelationDesc;
-       else
-       {
-               ListCell   *l;
-
-               relation = NULL;
-               foreach(l, estate->es_rowMarks)
-               {
-                       ExecRowMark *erm = lfirst(l);
-
-                       if (erm->rti == rti)
-                       {
-                               relation = erm->relation;
-                               break;
-                       }
-               }
-               if (relation == NULL)
-                       elog(ERROR, "could not find RowMark for RT index %u", rti);
-       }
-
        /*
-        * fetch tid tuple
+        * fetch target tuple
         *
         * Loop here to deal with updated or busy tuples
         */
@@ -1487,6 +1856,10 @@ EvalPlanQualFetch(EState *estate, Index rti,
 
                if (heap_fetch(relation, &SnapshotDirty, &tuple, &buffer, true, NULL))
                {
+                       HTSU_Result test;
+                       ItemPointerData update_ctid;
+                       TransactionId update_xmax;
+
                        /*
                         * If xmin isn't what we're expecting, the slot must have been
                         * recycled and reused for an unrelated tuple.  This implies that
@@ -1535,6 +1908,51 @@ EvalPlanQualFetch(EState *estate, Index rti,
                                return NULL;
                        }
 
+                       /*
+                        * This is a live tuple, so now try to lock it.
+                        */
+                       test = heap_lock_tuple(relation, &tuple, &buffer,
+                                                                  &update_ctid, &update_xmax,
+                                                                  estate->es_output_cid,
+                                                                  lockmode, false);
+                       /* We now have two pins on the buffer, get rid of one */
+                       ReleaseBuffer(buffer);
+
+                       switch (test)
+                       {
+                               case HeapTupleSelfUpdated:
+                                       /* treat it as deleted; do not process */
+                                       ReleaseBuffer(buffer);
+                                       return NULL;
+
+                               case HeapTupleMayBeUpdated:
+                                       /* successfully locked */
+                                       break;
+
+                               case HeapTupleUpdated:
+                                       ReleaseBuffer(buffer);
+                                       if (IsolationUsesXactSnapshot())
+                                               ereport(ERROR,
+                                                               (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
+                                                                errmsg("could not serialize access due to concurrent update")));
+                                       if (!ItemPointerEquals(&update_ctid, &tuple.t_self))
+                                       {
+                                               /* it was updated, so look at the updated version */
+                                               tuple.t_self = update_ctid;
+                                               /* updated row should have xmin matching this xmax */
+                                               priorXmax = update_xmax;
+                                               continue;
+                                       }
+                                       /* tuple was deleted, so give up */
+                                       return NULL;
+
+                               default:
+                                       ReleaseBuffer(buffer);
+                                       elog(ERROR, "unrecognized heap_lock_tuple status: %u",
+                                                test);
+                                       return NULL;    /* keep compiler quiet */
+                       }
+
                        /*
                         * We got tuple - now copy it for use by recheck query.
                         */
@@ -1570,7 +1988,7 @@ EvalPlanQualFetch(EState *estate, Index rti,
                 * mean that the row was updated or deleted by either a committed xact
                 * or our own xact.  If it was deleted, we can ignore it; if it was
                 * updated then chain up to the next version and repeat the whole
-                * test.
+                * process.
                 *
                 * As above, it should be safe to examine xmax and t_ctid without the
                 * buffer content lock, because they can't be changing.
@@ -1597,362 +2015,406 @@ EvalPlanQualFetch(EState *estate, Index rti,
 }
 
 /*
- * Push a new level of EPQ state, and prepare to execute the given subplan
+ * EvalPlanQualInit -- initialize during creation of a plan state node
+ * that might need to invoke EPQ processing.
+ *
+ * Note: subplan/auxrowmarks can be NULL/NIL if they will be set later
+ * with EvalPlanQualSetPlan.
  */
 void
-EvalPlanQualPush(EState *estate, Index rti, PlanState *subplanstate)
+EvalPlanQualInit(EPQState *epqstate, EState *estate,
+                                Plan *subplan, List *auxrowmarks, int epqParam)
 {
-       evalPlanQual *epq;
-       bool            endNode;
-
-       Assert(rti != 0);
+       /* Mark the EPQ state inactive */
+       epqstate->estate = NULL;
+       epqstate->planstate = NULL;
+       epqstate->origslot = NULL;
+       /* ... and remember data that EvalPlanQualBegin will need */
+       epqstate->plan = subplan;
+       epqstate->arowMarks = auxrowmarks;
+       epqstate->epqParam = epqParam;
+}
 
-       epq = estate->es_evalPlanQual;
-       endNode = true;
+/*
+ * EvalPlanQualSetPlan -- set or change subplan of an EPQState.
+ *
+ * We need this so that ModifyTuple can deal with multiple subplans.
+ */
+void
+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;
+       /* The rowmarks depend on the plan, too */
+       epqstate->arowMarks = auxrowmarks;
+}
 
-       if (epq != NULL && epq->rti == 0)
-       {
-               /* Top PQ stack entry is idle, so re-use it */
-               Assert(epq->next == NULL);
-               epq->rti = rti;
-               endNode = false;
-       }
+/*
+ * Install one test tuple into EPQ state, or clear test tuple if tuple == NULL
+ *
+ * NB: passed tuple must be palloc'd; it may get freed later
+ */
+void
+EvalPlanQualSetTuple(EPQState *epqstate, Index rti, HeapTuple tuple)
+{
+       EState     *estate = epqstate->estate;
 
-       /*
-        * If this is request for another RTE - Ra, - then we have to check wasn't
-        * PlanQual requested for Ra already and if so then Ra' row was updated
-        * again and we have to re-start old execution for Ra and forget all what
-        * we done after Ra was suspended. Cool? -:))
-        */
-       if (epq != NULL && epq->rti != rti &&
-               epq->estate->es_evTuple[rti - 1] != NULL)
-       {
-               do
-               {
-                       evalPlanQual *oldepq;
-
-                       /* stop execution */
-                       EvalPlanQualStop(epq);
-                       /* pop previous PlanQual from the stack */
-                       oldepq = epq->next;
-                       Assert(oldepq && oldepq->rti != 0);
-                       /* push current PQ to freePQ stack */
-                       oldepq->free = epq;
-                       epq = oldepq;
-                       estate->es_evalPlanQual = epq;
-               } while (epq->rti != rti);
-       }
+       Assert(rti > 0);
 
        /*
-        * If we are requested for another RTE then we have to suspend execution
-        * of current PlanQual and start execution for new one.
+        * free old test tuple, if any, and store new tuple where relation's scan
+        * node will see it
         */
-       if (epq == NULL || epq->rti != rti)
-       {
-               /* try to reuse plan used previously */
-               evalPlanQual *newepq = (epq != NULL) ? epq->free : NULL;
-
-               if (newepq == NULL)             /* first call or freePQ stack is empty */
-               {
-                       newepq = (evalPlanQual *) palloc0(sizeof(evalPlanQual));
-                       newepq->free = NULL;
-                       newepq->estate = NULL;
-                       newepq->planstate = NULL;
-                       newepq->origplanstate = NULL;
-                       newepq->resultslot = NULL;
-               }
-               else
-               {
-                       /* recycle previously used PlanQual */
-                       Assert(newepq->estate == NULL);
-                       epq->free = NULL;
-               }
-               /* push current PQ to the stack */
-               newepq->next = epq;
-               epq = newepq;
-               estate->es_evalPlanQual = epq;
-               epq->rti = rti;
-               endNode = false;
-       }
+       if (estate->es_epqTuple[rti - 1] != NULL)
+               heap_freetuple(estate->es_epqTuple[rti - 1]);
+       estate->es_epqTuple[rti - 1] = tuple;
+       estate->es_epqTupleSet[rti - 1] = true;
+}
 
-       Assert(epq->rti == rti);
-       Assert(estate->es_evalPlanQual == epq);
+/*
+ * Fetch back the current test tuple (if any) for the specified RTI
+ */
+HeapTuple
+EvalPlanQualGetTuple(EPQState *epqstate, Index rti)
+{
+       EState     *estate = epqstate->estate;
 
-       /*
-        * Ok - we're requested for the same RTE.  Unfortunately we still have to
-        * end and restart execution of the plan, because ExecReScan wouldn't
-        * ensure that upper plan nodes would reset themselves.  We could make
-        * that work if insertion of the target tuple were integrated with the
-        * Param mechanism somehow, so that the upper plan nodes know that their
-        * children's outputs have changed.
-        *
-        * Note that the stack of free evalPlanQual nodes is quite useless at the
-        * moment, since it only saves us from pallocing/releasing the
-        * evalPlanQual nodes themselves.  But it will be useful once we implement
-        * ReScan instead of end/restart for re-using PlanQual nodes.
-        */
-       if (endNode)
-       {
-               /* stop execution */
-               EvalPlanQualStop(epq);
-       }
+       Assert(rti > 0);
 
-       /*
-        * Initialize new recheck query.
-        *
-        * Note: if we were re-using PlanQual plans via ExecReScan, we'd need to
-        * instead copy down changeable state from the top plan (including
-        * es_result_relation_info) and reset locally changeable
-        * state in the epq (including es_param_exec_vals, es_evTupleNull).
-        */
-       epq->origplanstate = subplanstate;
-       EvalPlanQualStart(epq, estate, subplanstate->plan, epq->next);
+       return estate->es_epqTuple[rti - 1];
 }
 
 /*
- * Install one test tuple into current EPQ level
+ * Fetch the current row values for any non-locked relations that need
+ * to be scanned by an EvalPlanQual operation. origslot must have been set
+ * to contain the current result row (top-level row) that we need to recheck.
  */
 void
-EvalPlanQualSetTuple(EState *estate, Index rti, HeapTuple tuple)
+EvalPlanQualFetchRowMarks(EPQState *epqstate)
 {
-       evalPlanQual *epq = estate->es_evalPlanQual;
-       EState     *epqstate;
+       ListCell   *l;
 
-       Assert(rti != 0);
+       Assert(epqstate->origslot != NULL);
 
-       /*
-        * free old RTE' tuple, if any, and store target tuple where relation's
-        * scan node will see it
-        */
-       epqstate = epq->estate;
-       if (epqstate->es_evTuple[rti - 1] != NULL)
-               heap_freetuple(epqstate->es_evTuple[rti - 1]);
-       epqstate->es_evTuple[rti - 1] = tuple;
+       foreach(l, epqstate->arowMarks)
+       {
+               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);
+
+               if (erm->relation)
+               {
+                       Buffer          buffer;
+
+                       Assert(erm->markType == ROW_MARK_REFERENCE);
+
+                       /* if child rel, must check whether it produced this row */
+                       if (erm->rti != erm->prti)
+                       {
+                               Oid                     tableoid;
+
+                               datum = ExecGetJunkAttribute(epqstate->origslot,
+                                                                                        aerm->toidAttNo,
+                                                                                        &isNull);
+                               /* non-locked rels could be on the inside of outer joins */
+                               if (isNull)
+                                       continue;
+                               tableoid = DatumGetObjectId(datum);
+
+                               if (tableoid != RelationGetRelid(erm->relation))
+                               {
+                                       /* this child is inactive right now */
+                                       continue;
+                               }
+                       }
+
+                       /* fetch the tuple's ctid */
+                       datum = ExecGetJunkAttribute(epqstate->origslot,
+                                                                                aerm->ctidAttNo,
+                                                                                &isNull);
+                       /* non-locked rels could be on the inside of outer joins */
+                       if (isNull)
+                               continue;
+                       tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
+
+                       /* okay, fetch the tuple */
+                       if (!heap_fetch(erm->relation, SnapshotAny, &tuple, &buffer,
+                                                       false, NULL))
+                               elog(ERROR, "failed to fetch tuple for EvalPlanQual recheck");
+
+                       /* successful, copy and store tuple */
+                       EvalPlanQualSetTuple(epqstate, erm->rti,
+                                                                heap_copytuple(&tuple));
+                       ReleaseBuffer(buffer);
+               }
+               else
+               {
+                       HeapTupleHeader td;
+
+                       Assert(erm->markType == ROW_MARK_COPY);
+
+                       /* fetch the whole-row Var for the relation */
+                       datum = ExecGetJunkAttribute(epqstate->origslot,
+                                                                                aerm->wholeAttNo,
+                                                                                &isNull);
+                       /* non-locked rels could be on the inside of outer joins */
+                       if (isNull)
+                               continue;
+                       td = DatumGetHeapTupleHeader(datum);
+
+                       /* build a temporary HeapTuple control structure */
+                       tuple.t_len = HeapTupleHeaderGetDatumLength(td);
+                       ItemPointerSetInvalid(&(tuple.t_self));
+                       tuple.t_tableOid = InvalidOid;
+                       tuple.t_data = td;
+
+                       /* copy and store tuple */
+                       EvalPlanQualSetTuple(epqstate, erm->rti,
+                                                                heap_copytuple(&tuple));
+               }
+       }
 }
 
 /*
  * Fetch the next row (if any) from EvalPlanQual testing
+ *
+ * (In practice, there should never be more than one row...)
  */
 TupleTableSlot *
-EvalPlanQualNext(EState *estate)
+EvalPlanQualNext(EPQState *epqstate)
 {
-       evalPlanQual *epq = estate->es_evalPlanQual;
        MemoryContext oldcontext;
        TupleTableSlot *slot;
 
-       Assert(epq->rti != 0);
-
-       oldcontext = MemoryContextSwitchTo(epq->estate->es_query_cxt);
-       slot = ExecProcNode(epq->planstate);
+       oldcontext = MemoryContextSwitchTo(epqstate->estate->es_query_cxt);
+       slot = ExecProcNode(epqstate->planstate);
        MemoryContextSwitchTo(oldcontext);
 
        return slot;
 }
 
 /*
- * Shut down and pop the specified level of EvalPlanQual machinery,
- * plus any levels nested within it
+ * Initialize or reset an EvalPlanQual state tree
  */
 void
-EvalPlanQualPop(EState *estate, PlanState *subplanstate)
+EvalPlanQualBegin(EPQState *epqstate, EState *parentestate)
 {
-       evalPlanQual *epq = estate->es_evalPlanQual;
+       EState     *estate = epqstate->estate;
 
-       for (;;)
+       if (estate == NULL)
        {
-               PlanState *epqplanstate = epq->origplanstate;
-               evalPlanQual *oldepq;
-
-               Assert(epq->rti != 0);
-
-               /* stop execution */
-               EvalPlanQualStop(epq);
-               epq->origplanstate = NULL;
-               /* pop old PQ from the stack */
-               oldepq = epq->next;
-               if (oldepq == NULL)
-               {
-                       /* this is the first (oldest) PQ - mark as free */
-                       epq->rti = 0;
-                       break;
-               }
-               Assert(oldepq->rti != 0);
-               /* push current PQ to freePQ stack */
-               oldepq->free = epq;
-               epq = oldepq;
-               estate->es_evalPlanQual = epq;
-               if (epqplanstate == subplanstate)
-                       break;
+               /* First time through, so create a child EState */
+               EvalPlanQualStart(epqstate, parentestate, epqstate->plan);
        }
-}
-
-static void
-EndEvalPlanQual(EState *estate)
-{
-       evalPlanQual *epq = estate->es_evalPlanQual;
-
-       if (epq->rti == 0)                      /* plans already shutdowned */
+       else
        {
-               Assert(epq->next == NULL);
-               return;
-       }
+               /*
+                * We already have a suitable child EPQ tree, so just reset it.
+                */
+               int                     rtsize = list_length(parentestate->es_range_table);
+               PlanState  *planstate = epqstate->planstate;
 
-       for (;;)
-       {
-               evalPlanQual *oldepq;
-
-               /* stop execution */
-               EvalPlanQualStop(epq);
-               epq->origplanstate = NULL;
-               /* pop old PQ from the stack */
-               oldepq = epq->next;
-               if (oldepq == NULL)
+               MemSet(estate->es_epqScanDone, 0, rtsize * sizeof(bool));
+
+               /* Recopy current values of parent parameters */
+               if (parentestate->es_plannedstmt->nParamExec > 0)
                {
-                       /* this is the first (oldest) PQ - mark as free */
-                       epq->rti = 0;
-                       break;
+                       int                     i = parentestate->es_plannedstmt->nParamExec;
+
+                       while (--i >= 0)
+                       {
+                               /* copy value if any, but not execPlan link */
+                               estate->es_param_exec_vals[i].value =
+                                       parentestate->es_param_exec_vals[i].value;
+                               estate->es_param_exec_vals[i].isnull =
+                                       parentestate->es_param_exec_vals[i].isnull;
+                       }
                }
-               Assert(oldepq->rti != 0);
-               /* push current PQ to freePQ stack */
-               oldepq->free = epq;
-               epq = oldepq;
-               estate->es_evalPlanQual = epq;
+
+               /*
+                * Mark child plan tree as needing rescan at all scan nodes.  The
+                * first ExecProcNode will take care of actually doing the rescan.
+                */
+               planstate->chgParam = bms_add_member(planstate->chgParam,
+                                                                                        epqstate->epqParam);
        }
 }
 
 /*
- * Start execution of one level of PlanQual.
+ * Start execution of an EvalPlanQual plan tree.
  *
  * This is a cut-down version of ExecutorStart(): we copy some state from
  * the top-level estate rather than initializing it fresh.
  */
 static void
-EvalPlanQualStart(evalPlanQual *epq, EState *estate, Plan *planTree,
-                                 evalPlanQual *priorepq)
+EvalPlanQualStart(EPQState *epqstate, EState *parentestate, Plan *planTree)
 {
-       EState     *epqstate;
+       EState     *estate;
        int                     rtsize;
        MemoryContext oldcontext;
        ListCell   *l;
 
-       rtsize = list_length(estate->es_range_table);
+       rtsize = list_length(parentestate->es_range_table);
 
-       epq->estate = epqstate = CreateExecutorState();
+       epqstate->estate = estate = CreateExecutorState();
 
-       oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
+       oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 
        /*
-        * The epqstates share the top query's copy of unchanging state such as
+        * Child EPQ EStates share the parent's copy of unchanging state such as
         * the snapshot, rangetable, result-rel info, and external Param info.
         * They need their own copies of local state, including a tuple table,
         * es_param_exec_vals, etc.
         */
-       epqstate->es_direction = ForwardScanDirection;
-       epqstate->es_snapshot = estate->es_snapshot;
-       epqstate->es_crosscheck_snapshot = estate->es_crosscheck_snapshot;
-       epqstate->es_range_table = estate->es_range_table;
-       epqstate->es_junkFilter = estate->es_junkFilter;
-       epqstate->es_output_cid = estate->es_output_cid;
-       epqstate->es_result_relations = estate->es_result_relations;
-       epqstate->es_num_result_relations = estate->es_num_result_relations;
-       epqstate->es_result_relation_info = estate->es_result_relation_info;
+       estate->es_direction = ForwardScanDirection;
+       estate->es_snapshot = parentestate->es_snapshot;
+       estate->es_crosscheck_snapshot = parentestate->es_crosscheck_snapshot;
+       estate->es_range_table = parentestate->es_range_table;
+       estate->es_plannedstmt = parentestate->es_plannedstmt;
+       estate->es_junkFilter = parentestate->es_junkFilter;
+       estate->es_output_cid = parentestate->es_output_cid;
+       estate->es_result_relations = parentestate->es_result_relations;
+       estate->es_num_result_relations = parentestate->es_num_result_relations;
+       estate->es_result_relation_info = parentestate->es_result_relation_info;
        /* es_trig_target_relations must NOT be copied */
-       epqstate->es_param_list_info = estate->es_param_list_info;
-       if (estate->es_plannedstmt->nParamExec > 0)
-               epqstate->es_param_exec_vals = (ParamExecData *)
-                       palloc0(estate->es_plannedstmt->nParamExec * sizeof(ParamExecData));
-       epqstate->es_rowMarks = estate->es_rowMarks;
-       epqstate->es_instrument = estate->es_instrument;
-       epqstate->es_select_into = estate->es_select_into;
-       epqstate->es_into_oids = estate->es_into_oids;
-       epqstate->es_plannedstmt = estate->es_plannedstmt;
-
-       /*
-        * Each epqstate must have its own es_evTupleNull state, but all the stack
-        * entries share es_evTuple state.      This allows sub-rechecks to inherit
-        * the value being examined by an outer recheck.
-        */
-       epqstate->es_evTupleNull = (bool *) palloc0(rtsize * sizeof(bool));
-       if (priorepq == NULL)
-               /* first PQ stack entry */
-               epqstate->es_evTuple = (HeapTuple *)
-                       palloc0(rtsize * sizeof(HeapTuple));
+       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
+        * param workspace has to be local state, but we copy the initial values
+        * from the parent, so as to have access to any param values that were
+        * already set from other parts of the parent's plan tree.
+        */
+       estate->es_param_list_info = parentestate->es_param_list_info;
+       if (parentestate->es_plannedstmt->nParamExec > 0)
+       {
+               int                     i = parentestate->es_plannedstmt->nParamExec;
+
+               estate->es_param_exec_vals = (ParamExecData *)
+                       palloc0(i * sizeof(ParamExecData));
+               while (--i >= 0)
+               {
+                       /* copy value if any, but not execPlan link */
+                       estate->es_param_exec_vals[i].value =
+                               parentestate->es_param_exec_vals[i].value;
+                       estate->es_param_exec_vals[i].isnull =
+                               parentestate->es_param_exec_vals[i].isnull;
+               }
+       }
+
+       /*
+        * Each EState must have its own es_epqScanDone state, but if we have
+        * nested EPQ checks they should share es_epqTuple arrays.      This allows
+        * sub-rechecks to inherit the values being examined by an outer recheck.
+        */
+       estate->es_epqScanDone = (bool *) palloc0(rtsize * sizeof(bool));
+       if (parentestate->es_epqTuple != NULL)
+       {
+               estate->es_epqTuple = parentestate->es_epqTuple;
+               estate->es_epqTupleSet = parentestate->es_epqTupleSet;
+       }
        else
-               /* later stack entries share the same storage */
-               epqstate->es_evTuple = priorepq->estate->es_evTuple;
+       {
+               estate->es_epqTuple = (HeapTuple *)
+                       palloc0(rtsize * sizeof(HeapTuple));
+               estate->es_epqTupleSet = (bool *)
+                       palloc0(rtsize * sizeof(bool));
+       }
 
        /*
-        * Each epqstate also has its own tuple table.
+        * Each estate also has its own tuple table.
         */
-       epqstate->es_tupleTable = NIL;
+       estate->es_tupleTable = NIL;
 
        /*
         * Initialize private state information for each SubPlan.  We must do this
         * before running ExecInitNode on the main query tree, since
-        * 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.
+        * 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.  (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(epqstate->es_subplanstates == NIL);
-       foreach(l, estate->es_plannedstmt->subplans)
+       Assert(estate->es_subplanstates == NIL);
+       foreach(l, parentestate->es_plannedstmt->subplans)
        {
                Plan       *subplan = (Plan *) lfirst(l);
                PlanState  *subplanstate;
 
-               subplanstate = ExecInitNode(subplan, epqstate, 0);
+               /* Don't initialize ModifyTable subplans, per comment above */
+               if (IsA(subplan, ModifyTable))
+                       subplanstate = NULL;
+               else
+                       subplanstate = ExecInitNode(subplan, estate, 0);
 
-               epqstate->es_subplanstates = lappend(epqstate->es_subplanstates,
-                                                                                        subplanstate);
+               estate->es_subplanstates = lappend(estate->es_subplanstates,
+                                                                                  subplanstate);
        }
 
        /*
-        * Initialize the private state information for all the nodes in the
-        * part of the plan tree we need to run.  This opens files, allocates
-        * storage and leaves us ready to start processing tuples.
+        * Initialize the private state information for all the nodes in the part
+        * of the plan tree we need to run.  This opens files, allocates storage
+        * and leaves us ready to start processing tuples.
         */
-       epq->planstate = ExecInitNode(planTree, epqstate, 0);
+       epqstate->planstate = ExecInitNode(planTree, estate, 0);
 
        MemoryContextSwitchTo(oldcontext);
 }
 
 /*
- * End execution of one level of PlanQual.
+ * EvalPlanQualEnd -- shut down at termination of parent plan state node,
+ * or if we are done with the current EPQ child.
  *
  * This is a cut-down version of ExecutorEnd(); basically we want to do most
  * of the normal cleanup, but *not* close result relations (which we are
  * just sharing from the outer query). We do, however, have to close any
  * trigger target relations that got opened, since those are not shared.
+ * (There probably shouldn't be any of the latter, but just in case...)
  */
-static void
-EvalPlanQualStop(evalPlanQual *epq)
+void
+EvalPlanQualEnd(EPQState *epqstate)
 {
-       EState     *epqstate = epq->estate;
+       EState     *estate = epqstate->estate;
        MemoryContext oldcontext;
        ListCell   *l;
 
-       oldcontext = MemoryContextSwitchTo(epqstate->es_query_cxt);
+       if (estate == NULL)
+               return;                                 /* idle, so nothing to do */
+
+       oldcontext = MemoryContextSwitchTo(estate->es_query_cxt);
 
-       ExecEndNode(epq->planstate);
+       ExecEndNode(epqstate->planstate);
 
-       foreach(l, epqstate->es_subplanstates)
+       foreach(l, estate->es_subplanstates)
        {
                PlanState  *subplanstate = (PlanState *) lfirst(l);
 
                ExecEndNode(subplanstate);
        }
 
-       /* throw away the per-epqstate tuple table completely */
-       ExecResetTupleTable(epqstate->es_tupleTable, true);
-       epqstate->es_tupleTable = NIL;
-
-       if (epqstate->es_evTuple[epq->rti - 1] != NULL)
-       {
-               heap_freetuple(epqstate->es_evTuple[epq->rti - 1]);
-               epqstate->es_evTuple[epq->rti - 1] = NULL;
-       }
+       /* throw away the per-estate tuple table */
+       ExecResetTupleTable(estate->es_tupleTable, false);
 
-       foreach(l, epqstate->es_trig_target_relations)
+       /* close any trigger target relations attached to this EState */
+       foreach(l, estate->es_trig_target_relations)
        {
                ResultRelInfo *resultRelInfo = (ResultRelInfo *) lfirst(l);
 
@@ -1963,10 +2425,12 @@ EvalPlanQualStop(evalPlanQual *epq)
 
        MemoryContextSwitchTo(oldcontext);
 
-       FreeExecutorState(epqstate);
+       FreeExecutorState(estate);
 
-       epq->estate = NULL;
-       epq->planstate = NULL;
+       /* Mark EPQState idle */
+       epqstate->estate = NULL;
+       epqstate->planstate = NULL;
+       epqstate->origslot = NULL;
 }
 
 
@@ -1981,6 +2445,7 @@ typedef struct
 {
        DestReceiver pub;                       /* publicly-known function pointers */
        EState     *estate;                     /* EState we are working with */
+       DestReceiver *origdest;         /* QueryDesc's original receiver */
        Relation        rel;                    /* Relation to write to */
        int                     hi_options;             /* heap_insert performance options */
        BulkInsertState bistate;        /* bulk insert state */
@@ -1998,38 +2463,93 @@ OpenIntoRel(QueryDesc *queryDesc)
 {
        IntoClause *into = queryDesc->plannedstmt->intoClause;
        EState     *estate = queryDesc->estate;
+       TupleDesc       intoTupDesc = queryDesc->tupDesc;
        Relation        intoRelationDesc;
        char       *intoName;
        Oid                     namespaceId;
        Oid                     tablespaceId;
        Datum           reloptions;
-       AclResult       aclresult;
        Oid                     intoRelationId;
-       TupleDesc       tupdesc;
        DR_intorel *myState;
+       RangeTblEntry  *rte;
+       AttrNumber              attnum;
        static char *validnsps[] = HEAP_RELOPT_NAMESPACES;
 
        Assert(into);
 
+       /*
+        * XXX This code needs to be kept in sync with DefineRelation(). Maybe we
+        * should try to use that function instead.
+        */
+
        /*
         * Check consistency of arguments
         */
-       if (into->onCommit != ONCOMMIT_NOOP && !into->rel->istemp)
+       if (into->onCommit != ONCOMMIT_NOOP
+               && into->rel->relpersistence != RELPERSISTENCE_TEMP)
                ereport(ERROR,
                                (errcode(ERRCODE_INVALID_TABLE_DEFINITION),
                                 errmsg("ON COMMIT can only be used on temporary tables")));
 
+       {
+               AclResult aclresult;
+               int i;
+
+               for (i = 0; i < intoTupDesc->natts; i++)
+               {
+                       Oid atttypid = intoTupDesc->attrs[i]->atttypid;
+
+                       aclresult = pg_type_aclcheck(atttypid, GetUserId(), ACL_USAGE);
+                       if (aclresult != ACLCHECK_OK)
+                               aclcheck_error(aclresult, ACL_KIND_TYPE,
+                                                          format_type_be(atttypid));
+               }
+       }
+
        /*
-        * Find namespace to create in, check its permissions
+        * If a column name list was specified in CREATE TABLE AS, override the
+        * column names derived from the query.  (Too few column names are OK, too
+        * many are not.)  It would probably be all right to scribble directly on
+        * the query's result tupdesc, but let's be safe and make a copy.
+        */
+       if (into->colNames)
+       {
+               ListCell   *lc;
+
+               intoTupDesc = CreateTupleDescCopy(intoTupDesc);
+               attnum = 1;
+               foreach(lc, into->colNames)
+               {
+                       char       *colname = strVal(lfirst(lc));
+
+                       if (attnum > intoTupDesc->natts)
+                               ereport(ERROR,
+                                               (errcode(ERRCODE_SYNTAX_ERROR),
+                                                errmsg("CREATE TABLE AS specifies too many column names")));
+                       namestrcpy(&(intoTupDesc->attrs[attnum - 1]->attname), colname);
+                       attnum++;
+               }
+       }
+
+       /*
+        * Find namespace to create in, check its permissions, lock it against
+        * concurrent drop, and mark into->rel as RELPERSISTENCE_TEMP if the
+        * selected namespace is temporary.
         */
        intoName = into->rel->relname;
-       namespaceId = RangeVarGetCreationNamespace(into->rel);
+       namespaceId = RangeVarGetAndCheckCreationNamespace(into->rel, NoLock,
+                                                                                                          NULL);
 
-       aclresult = pg_namespace_aclcheck(namespaceId, GetUserId(),
-                                                                         ACL_CREATE);
-       if (aclresult != ACLCHECK_OK)
-               aclcheck_error(aclresult, ACL_KIND_NAMESPACE,
-                                          get_namespace_name(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.
+        */
+       if (into->rel->relpersistence == RELPERSISTENCE_TEMP
+               && InSecurityRestrictedOperation())
+               ereport(ERROR,
+                               (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
+                                errmsg("cannot create temporary table within security-restricted operation")));
 
        /*
         * Select tablespace to use.  If not specified, use default tablespace
@@ -2037,16 +2557,11 @@ OpenIntoRel(QueryDesc *queryDesc)
         */
        if (into->tableSpaceName)
        {
-               tablespaceId = get_tablespace_oid(into->tableSpaceName);
-               if (!OidIsValid(tablespaceId))
-                       ereport(ERROR,
-                                       (errcode(ERRCODE_UNDEFINED_OBJECT),
-                                        errmsg("tablespace \"%s\" does not exist",
-                                                       into->tableSpaceName)));
+               tablespaceId = get_tablespace_oid(into->tableSpaceName, false);
        }
        else
        {
-               tablespaceId = GetDefaultTablespace(into->rel->istemp);
+               tablespaceId = GetDefaultTablespace(into->rel->relpersistence);
                /* note InvalidOid is OK in this case */
        }
 
@@ -2072,19 +2587,19 @@ 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,
                                                                                          tablespaceId,
                                                                                          InvalidOid,
                                                                                          InvalidOid,
+                                                                                         InvalidOid,
                                                                                          GetUserId(),
-                                                                                         tupdesc,
+                                                                                         intoTupDesc,
                                                                                          NIL,
                                                                                          RELKIND_RELATION,
+                                                                                         into->rel->relpersistence,
+                                                                                         false,
                                                                                          false,
                                                                                          true,
                                                                                          0,
@@ -2092,8 +2607,7 @@ OpenIntoRel(QueryDesc *queryDesc)
                                                                                          reloptions,
                                                                                          true,
                                                                                          allowSystemTableMods);
-
-       FreeTupleDesc(tupdesc);
+       Assert(intoRelationId != InvalidOid);
 
        /*
         * Advance command counter so that the newly-created relation's catalog
@@ -2115,32 +2629,49 @@ OpenIntoRel(QueryDesc *queryDesc)
 
        (void) heap_reloptions(RELKIND_TOASTVALUE, reloptions, true);
 
-       AlterTableCreateToastTable(intoRelationId, InvalidOid, reloptions, false);
+       AlterTableCreateToastTable(intoRelationId, reloptions);
 
        /*
         * And open the constructed table for writing.
         */
        intoRelationDesc = heap_open(intoRelationId, AccessExclusiveLock);
 
+       /*
+        * Check INSERT permission on the constructed table.
+        */
+       rte = makeNode(RangeTblEntry);
+       rte->rtekind = RTE_RELATION;
+       rte->relid = intoRelationId;
+       rte->relkind = RELKIND_RELATION;
+       rte->requiredPerms = ACL_INSERT;
+
+       for (attnum = 1; attnum <= intoTupDesc->natts; attnum++)
+               rte->modifiedCols = bms_add_member(rte->modifiedCols,
+                               attnum - FirstLowInvalidHeapAttributeNumber);
+
+       ExecCheckRTPerms(list_make1(rte), true);
+
        /*
         * Now replace the query's DestReceiver with one for SELECT INTO
         */
-       queryDesc->dest = CreateDestReceiver(DestIntoRel);
-       myState = (DR_intorel *) queryDesc->dest;
+       myState = (DR_intorel *) CreateDestReceiver(DestIntoRel);
        Assert(myState->pub.mydest == DestIntoRel);
        myState->estate = estate;
+       myState->origdest = queryDesc->dest;
        myState->rel = intoRelationDesc;
 
+       queryDesc->dest = (DestReceiver *) myState;
+
        /*
-        * We can skip WAL-logging the insertions, unless PITR is in use.  We can
-        * skip the FSM in any case.
+        * We can skip WAL-logging the insertions, unless PITR or streaming
+        * replication is in use. We can skip the FSM in any case.
         */
        myState->hi_options = HEAP_INSERT_SKIP_FSM |
-               (XLogArchivingActive() ? 0 : HEAP_INSERT_SKIP_WAL);
+               (XLogIsNeeded() ? 0 : HEAP_INSERT_SKIP_WAL);
        myState->bistate = GetBulkInsertState();
 
-       /* Not using WAL requires rd_targblock be initially invalid */
-       Assert(intoRelationDesc->rd_targblock == InvalidBlockNumber);
+       /* Not using WAL requires smgr_targblock be initially invalid */
+       Assert(RelationGetTargetBlock(intoRelationDesc) == InvalidBlockNumber);
 }
 
 /*
@@ -2151,8 +2682,11 @@ CloseIntoRel(QueryDesc *queryDesc)
 {
        DR_intorel *myState = (DR_intorel *) queryDesc->dest;
 
-       /* OpenIntoRel might never have gotten called */
-       if (myState && myState->pub.mydest == DestIntoRel && myState->rel)
+       /*
+        * OpenIntoRel might never have gotten called, and we also want to guard
+        * against double destruction.
+        */
+       if (myState && myState->pub.mydest == DestIntoRel)
        {
                FreeBulkInsertState(myState->bistate);
 
@@ -2163,7 +2697,11 @@ CloseIntoRel(QueryDesc *queryDesc)
                /* close rel, but keep lock until commit */
                heap_close(myState->rel, NoLock);
 
-               myState->rel = NULL;
+               /* restore the receiver belonging to executor's caller */
+               queryDesc->dest = myState->origdest;
+
+               /* might as well invoke my destructor */
+               intorel_destroy((DestReceiver *) myState);
        }
 }