]> granicus.if.org Git - postgresql/blobdiff - src/backend/executor/execMain.c
Add a Gather executor node.
[postgresql] / src / backend / executor / execMain.c
index 73b6ebd5ec240f916a4766b3ca53ea80ae09ad15..37b7bbd413b3c629d95d2e27f0e9ea01adfebf04 100644 (file)
@@ -56,6 +56,7 @@
 #include "utils/acl.h"
 #include "utils/lsyscache.h"
 #include "utils/memutils.h"
+#include "utils/rls.h"
 #include "utils/snapmgr.h"
 #include "utils/tqual.h"
 
@@ -81,13 +82,29 @@ static void ExecutePlan(EState *estate, PlanState *planstate,
                        ScanDirection direction,
                        DestReceiver *dest);
 static bool ExecCheckRTEPerms(RangeTblEntry *rte);
+static bool ExecCheckRTEPermsModified(Oid relOid, Oid userid,
+                                                 Bitmapset *modifiedCols,
+                                                 AclMode requiredPerms);
 static void ExecCheckXactReadOnly(PlannedStmt *plannedstmt);
-static char *ExecBuildSlotValueDescription(TupleTableSlot *slot,
+static char *ExecBuildSlotValueDescription(Oid reloid,
+                                                         TupleTableSlot *slot,
                                                          TupleDesc tupdesc,
+                                                         Bitmapset *modifiedCols,
                                                          int maxfieldlen);
 static void EvalPlanQualStart(EPQState *epqstate, EState *parentestate,
                                  Plan *planTree);
 
+/*
+ * Note that GetUpdatedColumns() also exists in commands/trigger.c.  There does
+ * not appear to be any good header to put it into, given the structures that
+ * it uses, so we let them be duplicated.  Be sure to update both if one needs
+ * to be changed, however.
+ */
+#define GetInsertedColumns(relinfo, estate) \
+       (rt_fetch((relinfo)->ri_RangeTableIndex, (estate)->es_range_table)->insertedCols)
+#define GetUpdatedColumns(relinfo, estate) \
+       (rt_fetch((relinfo)->ri_RangeTableIndex, (estate)->es_range_table)->updatedCols)
+
 /* end of local decls */
 
 
@@ -135,8 +152,20 @@ standard_ExecutorStart(QueryDesc *queryDesc, int eflags)
        /*
         * If the transaction is read-only, we need to check if any writes are
         * planned to non-temporary tables.  EXPLAIN is considered read-only.
+        *
+        * Don't allow writes in parallel mode.  Supporting UPDATE and DELETE
+        * would require (a) storing the combocid hash in shared memory, rather
+        * than synchronizing it just once at the start of parallelism, and (b) an
+        * alternative to heap_update()'s reliance on xmax for mutual exclusion.
+        * INSERT may have no such troubles, but we forbid it to simplify the
+        * checks.
+        *
+        * We have lower-level defenses in CommandCounterIncrement and elsewhere
+        * against performing unsafe operations in parallel mode, but this gives a
+        * more user-friendly error message.
         */
-       if (XactReadOnly && !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
+       if ((XactReadOnly || IsInParallelMode()) &&
+               !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
                ExecCheckXactReadOnly(queryDesc->plannedstmt);
 
        /*
@@ -214,6 +243,11 @@ standard_ExecutorStart(QueryDesc *queryDesc, int eflags)
        if (!(eflags & (EXEC_FLAG_SKIP_TRIGGERS | EXEC_FLAG_EXPLAIN_ONLY)))
                AfterTriggerBeginQuery();
 
+       /* Enter parallel mode, if required by the query. */
+       if (queryDesc->plannedstmt->parallelModeNeeded &&
+               !(eflags & EXEC_FLAG_EXPLAIN_ONLY))
+               EnterParallelMode();
+
        MemoryContextSwitchTo(oldcontext);
 }
 
@@ -313,6 +347,9 @@ standard_ExecutorRun(QueryDesc *queryDesc,
                                        direction,
                                        dest);
 
+       /* Allow nodes to release or shut down resources. */
+       (void) ExecShutdownNode(queryDesc->planstate);
+
        /*
         * shutdown tuple receiver, if we started it
         */
@@ -445,6 +482,11 @@ standard_ExecutorEnd(QueryDesc *queryDesc)
         */
        MemoryContextSwitchTo(oldcontext);
 
+       /* Exit parallel mode, if it was required by the query. */
+       if (queryDesc->plannedstmt->parallelModeNeeded &&
+               !(estate->es_top_eflags & EXEC_FLAG_EXPLAIN_ONLY))
+               ExitParallelMode();
+
        /*
         * Release EState and per-query memory context.  This should release
         * everything the executor has allocated.
@@ -547,7 +589,6 @@ ExecCheckRTEPerms(RangeTblEntry *rte)
        AclMode         remainingPerms;
        Oid                     relOid;
        Oid                     userid;
-       int                     col;
 
        /*
         * Only plain-relation RTEs need to be checked here.  Function RTEs are
@@ -585,6 +626,8 @@ ExecCheckRTEPerms(RangeTblEntry *rte)
        remainingPerms = requiredPerms & ~relPerms;
        if (remainingPerms != 0)
        {
+               int                     col = -1;
+
                /*
                 * If we lack any permissions that exist only as relation permissions,
                 * we can fail straight away.
@@ -613,7 +656,6 @@ ExecCheckRTEPerms(RangeTblEntry *rte)
                                        return false;
                        }
 
-                       col = -1;
                        while ((col = bms_next_member(rte->selectedCols, col)) >= 0)
                        {
                                /* bit #s are offset by FirstLowInvalidHeapAttributeNumber */
@@ -636,61 +678,85 @@ ExecCheckRTEPerms(RangeTblEntry *rte)
                }
 
                /*
-                * Basically the same for the mod columns, with either INSERT or
-                * UPDATE privilege as specified by remainingPerms.
+                * Basically the same for the mod columns, for both INSERT and UPDATE
+                * privilege as specified by remainingPerms.
                 */
-               remainingPerms &= ~ACL_SELECT;
-               if (remainingPerms != 0)
-               {
-                       /*
-                        * When the query doesn't explicitly change any columns, allow the
-                        * query if we have permission on any column of the rel.  This is
-                        * to handle SELECT FOR UPDATE as well as possible corner cases in
-                        * INSERT and UPDATE.
-                        */
-                       if (bms_is_empty(rte->modifiedCols))
-                       {
-                               if (pg_attribute_aclcheck_all(relOid, userid, remainingPerms,
-                                                                                         ACLMASK_ANY) != ACLCHECK_OK)
-                                       return false;
-                       }
+               if (remainingPerms & ACL_INSERT && !ExecCheckRTEPermsModified(relOid,
+                                                                                                                                         userid,
+                                                                                                                  rte->insertedCols,
+                                                                                                                                ACL_INSERT))
+                       return false;
 
-                       col = -1;
-                       while ((col = bms_next_member(rte->modifiedCols, col)) >= 0)
-                       {
-                               /* bit #s are offset by FirstLowInvalidHeapAttributeNumber */
-                               AttrNumber      attno = col + FirstLowInvalidHeapAttributeNumber;
+               if (remainingPerms & ACL_UPDATE && !ExecCheckRTEPermsModified(relOid,
+                                                                                                                                         userid,
+                                                                                                                       rte->updatedCols,
+                                                                                                                                ACL_UPDATE))
+                       return false;
+       }
+       return true;
+}
 
-                               if (attno == InvalidAttrNumber)
-                               {
-                                       /* whole-row reference can't happen here */
-                                       elog(ERROR, "whole-row update is not implemented");
-                               }
-                               else
-                               {
-                                       if (pg_attribute_aclcheck(relOid, attno, userid,
-                                                                                         remainingPerms) != ACLCHECK_OK)
-                                               return false;
-                               }
-                       }
+/*
+ * ExecCheckRTEPermsModified
+ *             Check INSERT or UPDATE access permissions for a single RTE (these
+ *             are processed uniformly).
+ */
+static bool
+ExecCheckRTEPermsModified(Oid relOid, Oid userid, Bitmapset *modifiedCols,
+                                                 AclMode requiredPerms)
+{
+       int                     col = -1;
+
+       /*
+        * When the query doesn't explicitly update any columns, allow the query
+        * if we have permission on any column of the rel.  This is to handle
+        * SELECT FOR UPDATE as well as possible corner cases in UPDATE.
+        */
+       if (bms_is_empty(modifiedCols))
+       {
+               if (pg_attribute_aclcheck_all(relOid, userid, requiredPerms,
+                                                                         ACLMASK_ANY) != ACLCHECK_OK)
+                       return false;
+       }
+
+       while ((col = bms_next_member(modifiedCols, col)) >= 0)
+       {
+               /* bit #s are offset by FirstLowInvalidHeapAttributeNumber */
+               AttrNumber      attno = col + FirstLowInvalidHeapAttributeNumber;
+
+               if (attno == InvalidAttrNumber)
+               {
+                       /* whole-row reference can't happen here */
+                       elog(ERROR, "whole-row update is not implemented");
+               }
+               else
+               {
+                       if (pg_attribute_aclcheck(relOid, attno, userid,
+                                                                         requiredPerms) != ACLCHECK_OK)
+                               return false;
                }
        }
        return true;
 }
 
 /*
- * Check that the query does not imply any writes to non-temp tables.
+ * Check that the query does not imply any writes to non-temp tables;
+ * unless we're in parallel mode, in which case don't even allow writes
+ * to 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.
+ * tables just as we do in parallel mode; 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)
 {
        ListCell   *l;
 
-       /* Fail if write permissions are requested on any non-temp table */
+       /*
+        * Fail if write permissions are requested in parallel mode for table
+        * (temp or non-temp), otherwise fail for any non-temp table.
+        */
        foreach(l, plannedstmt->rtable)
        {
                RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
@@ -706,6 +772,9 @@ ExecCheckXactReadOnly(PlannedStmt *plannedstmt)
 
                PreventCommandIfReadOnly(CreateCommandTag((Node *) plannedstmt));
        }
+
+       if (plannedstmt->commandType != CMD_SELECT || plannedstmt->hasModifyingCTE)
+               PreventCommandIfParallelMode(CreateCommandTag((Node *) plannedstmt));
 }
 
 
@@ -802,21 +871,26 @@ InitPlan(QueryDesc *queryDesc, int eflags)
                if (rc->isParent)
                        continue;
 
+               /* get relation's OID (will produce InvalidOid if subquery) */
+               relid = getrelid(rc->rti, rangeTable);
+
+               /*
+                * If you change the conditions under which rel locks are acquired
+                * here, be sure to adjust ExecOpenScanRelation to match.
+                */
                switch (rc->markType)
                {
                        case ROW_MARK_EXCLUSIVE:
                        case ROW_MARK_NOKEYEXCLUSIVE:
                        case ROW_MARK_SHARE:
                        case ROW_MARK_KEYSHARE:
-                               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 ... */
+                               /* no physical table access is required */
                                relation = NULL;
                                break;
                        default:
@@ -831,12 +905,16 @@ InitPlan(QueryDesc *queryDesc, int eflags)
 
                erm = (ExecRowMark *) palloc(sizeof(ExecRowMark));
                erm->relation = relation;
+               erm->relid = relid;
                erm->rti = rc->rti;
                erm->prti = rc->prti;
                erm->rowmarkId = rc->rowmarkId;
                erm->markType = rc->markType;
+               erm->strength = rc->strength;
                erm->waitPolicy = rc->waitPolicy;
+               erm->ermActive = false;
                ItemPointerSetInvalid(&(erm->curCtid));
+               erm->ermExtra = NULL;
                estate->es_rowMarks = lappend(estate->es_rowMarks, erm);
        }
 
@@ -1080,6 +1158,8 @@ CheckValidResultRel(Relation resultRel, CmdType operation)
 static void
 CheckValidRowMarkRel(Relation rel, RowMarkType markType)
 {
+       FdwRoutine *fdwroutine;
+
        switch (rel->rd_rel->relkind)
        {
                case RELKIND_RELATION:
@@ -1115,11 +1195,13 @@ CheckValidRowMarkRel(Relation rel, RowMarkType markType)
                                                          RelationGetRelationName(rel))));
                        break;
                case RELKIND_FOREIGN_TABLE:
-                       /* Should not get here; planner should have used ROW_MARK_COPY */
-                       ereport(ERROR,
-                                       (errcode(ERRCODE_WRONG_OBJECT_TYPE),
-                                        errmsg("cannot lock rows in foreign table \"%s\"",
-                                                       RelationGetRelationName(rel))));
+                       /* Okay only if the FDW supports it */
+                       fdwroutine = GetFdwRoutineForRelation(rel, false);
+                       if (fdwroutine->RefetchForeignRow == NULL)
+                               ereport(ERROR,
+                                               (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+                                                errmsg("cannot lock rows in foreign table \"%s\"",
+                                                               RelationGetRelationName(rel))));
                        break;
                default:
                        ereport(ERROR,
@@ -1595,6 +1677,9 @@ ExecConstraints(ResultRelInfo *resultRelInfo,
        Relation        rel = resultRelInfo->ri_RelationDesc;
        TupleDesc       tupdesc = RelationGetDescr(rel);
        TupleConstr *constr = tupdesc->constr;
+       Bitmapset  *modifiedCols;
+       Bitmapset  *insertedCols;
+       Bitmapset  *updatedCols;
 
        Assert(constr);
 
@@ -1607,15 +1692,25 @@ ExecConstraints(ResultRelInfo *resultRelInfo,
                {
                        if (tupdesc->attrs[attrChk - 1]->attnotnull &&
                                slot_attisnull(slot, attrChk))
+                       {
+                               char       *val_desc;
+
+                               insertedCols = GetInsertedColumns(resultRelInfo, estate);
+                               updatedCols = GetUpdatedColumns(resultRelInfo, estate);
+                               modifiedCols = bms_union(insertedCols, updatedCols);
+                               val_desc = ExecBuildSlotValueDescription(RelationGetRelid(rel),
+                                                                                                                slot,
+                                                                                                                tupdesc,
+                                                                                                                modifiedCols,
+                                                                                                                64);
+
                                ereport(ERROR,
                                                (errcode(ERRCODE_NOT_NULL_VIOLATION),
                                                 errmsg("null value in column \"%s\" violates not-null constraint",
                                                          NameStr(tupdesc->attrs[attrChk - 1]->attname)),
-                                                errdetail("Failing row contains %s.",
-                                                                  ExecBuildSlotValueDescription(slot,
-                                                                                                                                tupdesc,
-                                                                                                                                64)),
+                                                val_desc ? errdetail("Failing row contains %s.", val_desc) : 0,
                                                 errtablecol(rel, attrChk)));
+                       }
                }
        }
 
@@ -1624,25 +1719,42 @@ ExecConstraints(ResultRelInfo *resultRelInfo,
                const char *failed;
 
                if ((failed = ExecRelCheck(resultRelInfo, slot, estate)) != NULL)
+               {
+                       char       *val_desc;
+
+                       insertedCols = GetInsertedColumns(resultRelInfo, estate);
+                       updatedCols = GetUpdatedColumns(resultRelInfo, estate);
+                       modifiedCols = bms_union(insertedCols, updatedCols);
+                       val_desc = ExecBuildSlotValueDescription(RelationGetRelid(rel),
+                                                                                                        slot,
+                                                                                                        tupdesc,
+                                                                                                        modifiedCols,
+                                                                                                        64);
                        ereport(ERROR,
                                        (errcode(ERRCODE_CHECK_VIOLATION),
                                         errmsg("new row for relation \"%s\" violates check constraint \"%s\"",
                                                        RelationGetRelationName(rel), failed),
-                                        errdetail("Failing row contains %s.",
-                                                          ExecBuildSlotValueDescription(slot,
-                                                                                                                        tupdesc,
-                                                                                                                        64)),
+                         val_desc ? errdetail("Failing row contains %s.", val_desc) : 0,
                                         errtableconstraint(rel, failed)));
+               }
        }
 }
 
 /*
  * ExecWithCheckOptions -- check that tuple satisfies any WITH CHECK OPTIONs
+ * of the specified kind.
+ *
+ * Note that this needs to be called multiple times to ensure that all kinds of
+ * WITH CHECK OPTIONs are handled (both those from views which have the WITH
+ * CHECK OPTION set and from row level security policies).  See ExecInsert()
+ * and ExecUpdate().
  */
 void
-ExecWithCheckOptions(ResultRelInfo *resultRelInfo,
+ExecWithCheckOptions(WCOKind kind, ResultRelInfo *resultRelInfo,
                                         TupleTableSlot *slot, EState *estate)
 {
+       Relation        rel = resultRelInfo->ri_RelationDesc;
+       TupleDesc       tupdesc = RelationGetDescr(rel);
        ExprContext *econtext;
        ListCell   *l1,
                           *l2;
@@ -1663,24 +1775,87 @@ ExecWithCheckOptions(ResultRelInfo *resultRelInfo,
                WithCheckOption *wco = (WithCheckOption *) lfirst(l1);
                ExprState  *wcoExpr = (ExprState *) lfirst(l2);
 
+               /*
+                * Skip any WCOs which are not the kind we are looking for at this
+                * time.
+                */
+               if (wco->kind != kind)
+                       continue;
+
                /*
                 * WITH CHECK OPTION checks are intended to ensure that the new tuple
                 * is visible (in the case of a view) or that it passes the
-                * 'with-check' policy (in the case of row security).
-                * If the qual evaluates to NULL or FALSE, then the new tuple won't be
-                * included in the view or doesn't pass the 'with-check' policy for the
-                * table.  We need ExecQual to return FALSE for NULL to handle the view
-                * case (the opposite of what we do above for CHECK constraints).
+                * 'with-check' policy (in the case of row security). If the qual
+                * evaluates to NULL or FALSE, then the new tuple won't be included in
+                * the view or doesn't pass the 'with-check' policy for the table.  We
+                * need ExecQual to return FALSE for NULL to handle the view case (the
+                * opposite of what we do above for CHECK constraints).
                 */
                if (!ExecQual((List *) wcoExpr, econtext, false))
-                       ereport(ERROR,
-                                       (errcode(ERRCODE_WITH_CHECK_OPTION_VIOLATION),
-                                errmsg("new row violates WITH CHECK OPTION for \"%s\"",
-                                               wco->viewname),
-                                        errdetail("Failing row contains %s.",
-                                                          ExecBuildSlotValueDescription(slot,
-                                                       RelationGetDescr(resultRelInfo->ri_RelationDesc),
-                                                                                                                        64))));
+               {
+                       char       *val_desc;
+                       Bitmapset  *modifiedCols;
+                       Bitmapset  *insertedCols;
+                       Bitmapset  *updatedCols;
+
+                       switch (wco->kind)
+                       {
+                                       /*
+                                        * For WITH CHECK OPTIONs coming from views, we might be
+                                        * able to provide the details on the row, depending on
+                                        * the permissions on the relation (that is, if the user
+                                        * could view it directly anyway).  For RLS violations, we
+                                        * don't include the data since we don't know if the user
+                                        * should be able to view the tuple as as that depends on
+                                        * the USING policy.
+                                        */
+                               case WCO_VIEW_CHECK:
+                                       insertedCols = GetInsertedColumns(resultRelInfo, estate);
+                                       updatedCols = GetUpdatedColumns(resultRelInfo, estate);
+                                       modifiedCols = bms_union(insertedCols, updatedCols);
+                                       val_desc = ExecBuildSlotValueDescription(RelationGetRelid(rel),
+                                                                                                                        slot,
+                                                                                                                        tupdesc,
+                                                                                                                        modifiedCols,
+                                                                                                                        64);
+
+                                       ereport(ERROR,
+                                                       (errcode(ERRCODE_WITH_CHECK_OPTION_VIOLATION),
+                                         errmsg("new row violates WITH CHECK OPTION for \"%s\"",
+                                                        wco->relname),
+                                                        val_desc ? errdetail("Failing row contains %s.",
+                                                                                                 val_desc) : 0));
+                                       break;
+                               case WCO_RLS_INSERT_CHECK:
+                               case WCO_RLS_UPDATE_CHECK:
+                                       if (wco->polname != NULL)
+                                               ereport(ERROR,
+                                                               (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
+                                                        errmsg("new row violates row level security policy \"%s\" for \"%s\"",
+                                                                       wco->polname, wco->relname)));
+                                       else
+                                               ereport(ERROR,
+                                                               (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
+                                                        errmsg("new row violates row level security policy for \"%s\"",
+                                                                       wco->relname)));
+                                       break;
+                               case WCO_RLS_CONFLICT_CHECK:
+                                       if (wco->polname != NULL)
+                                               ereport(ERROR,
+                                                               (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
+                                                        errmsg("new row violates row level security policy \"%s\" (USING expression) for \"%s\"",
+                                                                       wco->polname, wco->relname)));
+                                       else
+                                               ereport(ERROR,
+                                                               (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
+                                                        errmsg("new row violates row level security policy (USING expression) for \"%s\"",
+                                                                       wco->relname)));
+                                       break;
+                               default:
+                                       elog(ERROR, "unrecognized WCO kind: %u", wco->kind);
+                                       break;
+                       }
+               }
        }
 }
 
@@ -1696,25 +1871,64 @@ ExecWithCheckOptions(ResultRelInfo *resultRelInfo,
  * dropped columns.  We used to use the slot's tuple descriptor to decode the
  * data, but the slot's descriptor doesn't identify dropped columns, so we
  * now need to be passed the relation's descriptor.
+ *
+ * Note that, like BuildIndexValueDescription, if the user does not have
+ * permission to view any of the columns involved, a NULL is returned.  Unlike
+ * BuildIndexValueDescription, if the user has access to view a subset of the
+ * column involved, that subset will be returned with a key identifying which
+ * columns they are.
  */
 static char *
-ExecBuildSlotValueDescription(TupleTableSlot *slot,
+ExecBuildSlotValueDescription(Oid reloid,
+                                                         TupleTableSlot *slot,
                                                          TupleDesc tupdesc,
+                                                         Bitmapset *modifiedCols,
                                                          int maxfieldlen)
 {
        StringInfoData buf;
+       StringInfoData collist;
        bool            write_comma = false;
+       bool            write_comma_collist = false;
        int                     i;
+       AclResult       aclresult;
+       bool            table_perm = false;
+       bool            any_perm = false;
 
-       /* Make sure the tuple is fully deconstructed */
-       slot_getallattrs(slot);
+       /*
+        * Check if RLS is enabled and should be active for the relation; if so,
+        * then don't return anything.  Otherwise, go through normal permission
+        * checks.
+        */
+       if (check_enable_rls(reloid, InvalidOid, true) == RLS_ENABLED)
+               return NULL;
 
        initStringInfo(&buf);
 
        appendStringInfoChar(&buf, '(');
 
+       /*
+        * Check if the user has permissions to see the row.  Table-level SELECT
+        * allows access to all columns.  If the user does not have table-level
+        * SELECT then we check each column and include those the user has SELECT
+        * rights on.  Additionally, we always include columns the user provided
+        * data for.
+        */
+       aclresult = pg_class_aclcheck(reloid, GetUserId(), ACL_SELECT);
+       if (aclresult != ACLCHECK_OK)
+       {
+               /* Set up the buffer for the column list */
+               initStringInfo(&collist);
+               appendStringInfoChar(&collist, '(');
+       }
+       else
+               table_perm = any_perm = true;
+
+       /* Make sure the tuple is fully deconstructed */
+       slot_getallattrs(slot);
+
        for (i = 0; i < tupdesc->natts; i++)
        {
+               bool            column_perm = false;
                char       *val;
                int                     vallen;
 
@@ -1722,46 +1936,112 @@ ExecBuildSlotValueDescription(TupleTableSlot *slot,
                if (tupdesc->attrs[i]->attisdropped)
                        continue;
 
-               if (slot->tts_isnull[i])
-                       val = "null";
-               else
+               if (!table_perm)
                {
-                       Oid                     foutoid;
-                       bool            typisvarlena;
+                       /*
+                        * No table-level SELECT, so need to make sure they either have
+                        * SELECT rights on the column or that they have provided the data
+                        * for the column.  If not, omit this column from the error
+                        * message.
+                        */
+                       aclresult = pg_attribute_aclcheck(reloid, tupdesc->attrs[i]->attnum,
+                                                                                         GetUserId(), ACL_SELECT);
+                       if (bms_is_member(tupdesc->attrs[i]->attnum - FirstLowInvalidHeapAttributeNumber,
+                                                         modifiedCols) || aclresult == ACLCHECK_OK)
+                       {
+                               column_perm = any_perm = true;
 
-                       getTypeOutputInfo(tupdesc->attrs[i]->atttypid,
-                                                         &foutoid, &typisvarlena);
-                       val = OidOutputFunctionCall(foutoid, slot->tts_values[i]);
-               }
+                               if (write_comma_collist)
+                                       appendStringInfoString(&collist, ", ");
+                               else
+                                       write_comma_collist = true;
 
-               if (write_comma)
-                       appendStringInfoString(&buf, ", ");
-               else
-                       write_comma = true;
+                               appendStringInfoString(&collist, NameStr(tupdesc->attrs[i]->attname));
+                       }
+               }
 
-               /* truncate if needed */
-               vallen = strlen(val);
-               if (vallen <= maxfieldlen)
-                       appendStringInfoString(&buf, val);
-               else
+               if (table_perm || column_perm)
                {
-                       vallen = pg_mbcliplen(val, vallen, maxfieldlen);
-                       appendBinaryStringInfo(&buf, val, vallen);
-                       appendStringInfoString(&buf, "...");
+                       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 (write_comma)
+                               appendStringInfoString(&buf, ", ");
+                       else
+                               write_comma = true;
+
+                       /* 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, "...");
+                       }
                }
        }
 
+       /* If we end up with zero columns being returned, then return NULL. */
+       if (!any_perm)
+               return NULL;
+
        appendStringInfoChar(&buf, ')');
 
+       if (!table_perm)
+       {
+               appendStringInfoString(&collist, ") = ");
+               appendStringInfoString(&collist, buf.data);
+
+               return collist.data;
+       }
+
        return buf.data;
 }
 
 
+/*
+ * ExecUpdateLockMode -- find the appropriate UPDATE tuple lock mode for a
+ * given ResultRelInfo
+ */
+LockTupleMode
+ExecUpdateLockMode(EState *estate, ResultRelInfo *relinfo)
+{
+       Bitmapset  *keyCols;
+       Bitmapset  *updatedCols;
+
+       /*
+        * Compute lock mode to use.  If columns that are part of the key have not
+        * been modified, then we can use a weaker lock, allowing for better
+        * concurrency.
+        */
+       updatedCols = GetUpdatedColumns(relinfo, estate);
+       keyCols = RelationGetIndexAttrBitmap(relinfo->ri_RelationDesc,
+                                                                                INDEX_ATTR_BITMAP_KEY);
+
+       if (bms_overlap(keyCols, updatedCols))
+               return LockTupleExclusive;
+
+       return LockTupleNoKeyExclusive;
+}
+
 /*
  * ExecFindRowMark -- find the ExecRowMark struct for given rangetable index
+ *
+ * If no such struct, either return NULL or throw error depending on missing_ok
  */
 ExecRowMark *
-ExecFindRowMark(EState *estate, Index rti)
+ExecFindRowMark(EState *estate, Index rti, bool missing_ok)
 {
        ListCell   *lc;
 
@@ -1772,8 +2052,9 @@ ExecFindRowMark(EState *estate, Index rti)
                if (erm->rti == rti)
                        return erm;
        }
-       elog(ERROR, "failed to find ExecRowMark for rangetable index %u", rti);
-       return NULL;                            /* keep compiler quiet */
+       if (!missing_ok)
+               elog(ERROR, "failed to find ExecRowMark for rangetable index %u", rti);
+       return NULL;
 }
 
 /*
@@ -1792,21 +2073,9 @@ ExecBuildAuxRowMark(ExecRowMark *erm, List *targetlist)
        aerm->rowmark = erm;
 
        /* Look up the resjunk columns associated with this rowmark */
-       if (erm->relation)
+       if (erm->markType != ROW_MARK_COPY)
        {
-               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 */
+               /* need ctid for all methods other than COPY */
                snprintf(resname, sizeof(resname), "ctid%u", erm->rowmarkId);
                aerm->ctidAttNo = ExecFindJunkAttributeInTlist(targetlist,
                                                                                                           resname);
@@ -1815,8 +2084,7 @@ ExecBuildAuxRowMark(ExecRowMark *erm, List *targetlist)
        }
        else
        {
-               Assert(erm->markType == ROW_MARK_COPY);
-
+               /* need wholerow if COPY */
                snprintf(resname, sizeof(resname), "wholerow%u", erm->rowmarkId);
                aerm->wholeAttNo = ExecFindJunkAttributeInTlist(targetlist,
                                                                                                                resname);
@@ -1824,6 +2092,16 @@ ExecBuildAuxRowMark(ExecRowMark *erm, List *targetlist)
                        elog(ERROR, "could not find junk %s column", resname);
        }
 
+       /* 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);
+       }
+
        return aerm;
 }
 
@@ -1974,8 +2252,9 @@ EvalPlanQualFetch(EState *estate, Relation relation, int lockmode,
                         * recycled and reused for an unrelated tuple.  This implies that
                         * the latest version of the row was deleted, so we need do
                         * nothing.  (Should be safe to examine xmin without getting
-                        * buffer's content lock, since xmin never changes in an existing
-                        * tuple.)
+                        * buffer's content lock.  We assume reading a TransactionId to be
+                        * atomic, and Xmin never changes in an existing tuple, except to
+                        * invalid or frozen, and neither of those can match priorXmax.)
                         */
                        if (!TransactionIdEquals(HeapTupleHeaderGetXmin(tuple.t_data),
                                                                         priorXmax))
@@ -1999,19 +2278,19 @@ EvalPlanQualFetch(EState *estate, Relation relation, int lockmode,
                                {
                                        case LockWaitBlock:
                                                XactLockTableWait(SnapshotDirty.xmax,
-                                                                                 relation, &tuple.t_data->t_ctid,
+                                                                                 relation, &tuple.t_self,
                                                                                  XLTW_FetchUpdated);
                                                break;
                                        case LockWaitSkip:
                                                if (!ConditionalXactLockTableWait(SnapshotDirty.xmax))
-                                                       return NULL; /* skip instead of waiting */
+                                                       return NULL;            /* skip instead of waiting */
                                                break;
                                        case LockWaitError:
                                                if (!ConditionalXactLockTableWait(SnapshotDirty.xmax))
                                                        ereport(ERROR,
                                                                        (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
                                                                         errmsg("could not obtain lock on row in relation \"%s\"",
-                                                                                       RelationGetRelationName(relation))));
+                                                                               RelationGetRelationName(relation))));
                                                break;
                                }
                                continue;               /* loop back to repeat heap_fetch */
@@ -2056,11 +2335,12 @@ EvalPlanQualFetch(EState *estate, Relation relation, int lockmode,
                                         * case, so as to avoid the "Halloween problem" of
                                         * repeated update attempts.  In the latter case it might
                                         * be sensible to fetch the updated tuple instead, but
-                                        * doing so would require changing heap_lock_tuple as well
-                                        * as heap_update and heap_delete to not complain about
-                                        * updating "invisible" tuples, which seems pretty scary.
-                                        * So for now, treat the tuple as deleted and do not
-                                        * process.
+                                        * doing so would require changing heap_update and
+                                        * heap_delete to not complain about updating "invisible"
+                                        * tuples, which seems pretty scary (heap_lock_tuple will
+                                        * not complain, but few callers expect
+                                        * HeapTupleInvisible, and we're not one of them).  So for
+                                        * now, treat the tuple as deleted and do not process.
                                         */
                                        ReleaseBuffer(buffer);
                                        return NULL;
@@ -2075,6 +2355,9 @@ EvalPlanQualFetch(EState *estate, Relation relation, int lockmode,
                                                ereport(ERROR,
                                                                (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
                                                                 errmsg("could not serialize access due to concurrent update")));
+
+                                       /* Should not encounter speculative tuple on recheck */
+                                       Assert(!HeapTupleHeaderIsSpeculative(tuple.t_data));
                                        if (!ItemPointerEquals(&hufd.ctid, &tuple.t_self))
                                        {
                                                /* it was updated, so look at the updated version */
@@ -2090,6 +2373,9 @@ EvalPlanQualFetch(EState *estate, Relation relation, int lockmode,
                                        ReleaseBuffer(buffer);
                                        return NULL;
 
+                               case HeapTupleInvisible:
+                                       elog(ERROR, "attempted to lock invisible tuple");
+
                                default:
                                        ReleaseBuffer(buffer);
                                        elog(ERROR, "unrecognized heap_lock_tuple status: %u",
@@ -2182,7 +2468,7 @@ EvalPlanQualInit(EPQState *epqstate, EState *estate,
 /*
  * EvalPlanQualSetPlan -- set or change subplan of an EPQState.
  *
- * We need this so that ModifyTuple can deal with multiple subplans.
+ * We need this so that ModifyTable can deal with multiple subplans.
  */
 void
 EvalPlanQualSetPlan(EPQState *epqstate, Plan *subplan, List *auxrowmarks)
@@ -2256,31 +2542,32 @@ EvalPlanQualFetchRowMarks(EPQState *epqstate)
                /* clear any leftover test tuple for this rel */
                EvalPlanQualSetTuple(epqstate, erm->rti, NULL);
 
-               if (erm->relation)
+               /* if child rel, must check whether it produced this row */
+               if (erm->rti != erm->prti)
                {
-                       Buffer          buffer;
+                       Oid                     tableoid;
 
-                       Assert(erm->markType == ROW_MARK_REFERENCE);
+                       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 child rel, must check whether it produced this row */
-                       if (erm->rti != erm->prti)
+                       Assert(OidIsValid(erm->relid));
+                       if (tableoid != erm->relid)
                        {
-                               Oid                     tableoid;
+                               /* this child is inactive right now */
+                               continue;
+                       }
+               }
 
-                               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 (erm->markType == ROW_MARK_REFERENCE)
+               {
+                       HeapTuple       copyTuple;
 
-                               if (tableoid != RelationGetRelid(erm->relation))
-                               {
-                                       /* this child is inactive right now */
-                                       continue;
-                               }
-                       }
+                       Assert(erm->relation != NULL);
 
                        /* fetch the tuple's ctid */
                        datum = ExecGetJunkAttribute(epqstate->origslot,
@@ -2289,17 +2576,50 @@ EvalPlanQualFetchRowMarks(EPQState *epqstate)
                        /* 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");
+                       /* fetch requests on foreign tables must be passed to their FDW */
+                       if (erm->relation->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
+                       {
+                               FdwRoutine *fdwroutine;
+                               bool            updated = false;
+
+                               fdwroutine = GetFdwRoutineForRelation(erm->relation, false);
+                               /* this should have been checked already, but let's be safe */
+                               if (fdwroutine->RefetchForeignRow == NULL)
+                                       ereport(ERROR,
+                                                       (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+                                                  errmsg("cannot lock rows in foreign table \"%s\"",
+                                                                 RelationGetRelationName(erm->relation))));
+                               copyTuple = fdwroutine->RefetchForeignRow(epqstate->estate,
+                                                                                                                 erm,
+                                                                                                                 datum,
+                                                                                                                 &updated);
+                               if (copyTuple == NULL)
+                                       elog(ERROR, "failed to fetch tuple for EvalPlanQual recheck");
+
+                               /*
+                                * Ideally we'd insist on updated == false here, but that
+                                * assumes that FDWs can track that exactly, which they might
+                                * not be able to.  So just ignore the flag.
+                                */
+                       }
+                       else
+                       {
+                               /* ordinary table, fetch the tuple */
+                               Buffer          buffer;
 
-                       /* successful, copy and store tuple */
-                       EvalPlanQualSetTuple(epqstate, erm->rti,
-                                                                heap_copytuple(&tuple));
-                       ReleaseBuffer(buffer);
+                               tuple.t_self = *((ItemPointer) DatumGetPointer(datum));
+                               if (!heap_fetch(erm->relation, SnapshotAny, &tuple, &buffer,
+                                                               false, NULL))
+                                       elog(ERROR, "failed to fetch tuple for EvalPlanQual recheck");
+
+                               /* successful, copy tuple */
+                               copyTuple = heap_copytuple(&tuple);
+                               ReleaseBuffer(buffer);
+                       }
+
+                       /* store tuple */
+                       EvalPlanQualSetTuple(epqstate, erm->rti, copyTuple);
                }
                else
                {
@@ -2318,9 +2638,11 @@ EvalPlanQualFetchRowMarks(EPQState *epqstate)
 
                        /* build a temporary HeapTuple control structure */
                        tuple.t_len = HeapTupleHeaderGetDatumLength(td);
-                       ItemPointerSetInvalid(&(tuple.t_self));
-                       tuple.t_tableOid = InvalidOid;
                        tuple.t_data = td;
+                       /* relation might be a foreign table, if so provide tableoid */
+                       tuple.t_tableOid = erm->relid;
+                       /* also copy t_ctid in case there's valid data there */
+                       tuple.t_self = td->t_ctid;
 
                        /* copy and store tuple */
                        EvalPlanQualSetTuple(epqstate, erm->rti,