Fix bugs in RETURNING in cross-partition UPDATE cases.

If the source and destination partitions don't have identical
rowtypes (for example, one has dropped columns the other lacks),
then the planSlot contents will be different because of that.
If the query has a RETURNING list that tries to return resjunk
columns out of the planSlot, that is columns from tables that
were joined to the target table, we'd get errors or wrong answers.
That's because we used the RETURNING list generated for the
destination partition, which expects a planSlot matching that
partition's subplan.

The most practical fix seems to be to convert the updated destination
tuple back to the source partition's rowtype, and then apply the
RETURNING list generated for the source partition.  This avoids making
fragile assumptions about whether the per-subpartition subplans
generated all the resjunk columns in the same order.

This has been broken since v11 introduced cross-partition UPDATE.
The lack of field complaints shows that non-identical partitions
aren't a common case; therefore, don't stress too hard about
making the conversion efficient.

There's no such bug in HEAD, because commit 86dc90056 got rid of
per-target-relation variance in the contents of the planSlot.
Hence, patch v11-v13 only.

Amit Langote and Etsuro Fujita, small changes by me

Discussion: https://postgr.es/m/CA+HiwqE_UK1jTSNrjb8mpTdivzd3dum6mK--xqKq0Y9VmfwWQA@mail.gmail.com

src/backend/executor/nodeModifyTable.c

@@ -149,34 +149,40 @@ ExecCheckPlanOutput(Relation resultRel, List *targetList)
 /*
  * ExecProcessReturning --- evaluate a RETURNING list
  *
- * resultRelInfo: current result rel
+ * projectReturning: the projection to evaluate
+ * resultRelOid: result relation's OID
  * tupleSlot: slot holding tuple actually inserted/updated/deleted
  * planSlot: slot holding tuple returned by top subplan node
  *
+ * In cross-partition UPDATE cases, projectReturning and planSlot are as
+ * for the source partition, and tupleSlot must conform to that.  But
+ * resultRelOid is for the destination partition.
+ *
  * Note: If tupleSlot is NULL, the FDW should have already provided econtext's
  * scan tuple.
  *
  * Returns a slot holding the result tuple
  */
 static TupleTableSlot *
-ExecProcessReturning(ResultRelInfo *resultRelInfo,
+ExecProcessReturning(ProjectionInfo *projectReturning,
+					 Oid resultRelOid,
 					 TupleTableSlot *tupleSlot,
 					 TupleTableSlot *planSlot)
 {
-	ProjectionInfo *projectReturning = resultRelInfo->ri_projectReturning;
 	ExprContext *econtext = projectReturning->pi_exprContext;
 
 	/* Make tuple and any needed join variables available to ExecProject */
 	if (tupleSlot)
 		econtext->ecxt_scantuple = tupleSlot;
+	else
+		Assert(econtext->ecxt_scantuple);
 	econtext->ecxt_outertuple = planSlot;
 
 	/*
-	 * RETURNING expressions might reference the tableoid column, so
-	 * reinitialize tts_tableOid before evaluating them.
+	 * RETURNING expressions might reference the tableoid column, so be sure
+	 * we expose the desired OID, ie that of the real target relation.
 	 */
-	econtext->ecxt_scantuple->tts_tableOid =
-		RelationGetRelid(resultRelInfo->ri_RelationDesc);
+	econtext->ecxt_scantuple->tts_tableOid = resultRelOid;
 
 	/* Compute the RETURNING expressions */
 	return ExecProject(projectReturning);
@@ -368,6 +374,16 @@ ExecComputeStoredGenerated(EState *estate, TupleTableSlot *slot, CmdType cmdtype
  *		For INSERT, we have to insert the tuple into the target relation
  *		and insert appropriate tuples into the index relations.
  *
+ *		slot contains the new tuple value to be stored.
+ *		planSlot is the output of the ModifyTable's subplan; we use it
+ *		to access "junk" columns that are not going to be stored.
+ *		In a cross-partition UPDATE, srcSlot is the slot that held the
+ *		updated tuple for the source relation; otherwise it's NULL.
+ *
+ *		returningRelInfo is the resultRelInfo for the source relation of a
+ *		cross-partition UPDATE; otherwise it's the current result relation.
+ *		We use it to process RETURNING lists, for reasons explained below.
+ *
  *		Returns RETURNING result if any, otherwise NULL.
  * ----------------------------------------------------------------
  */
@@ -375,6 +391,8 @@ static TupleTableSlot *
 ExecInsert(ModifyTableState *mtstate,
 		   TupleTableSlot *slot,
 		   TupleTableSlot *planSlot,
+		   TupleTableSlot *srcSlot,
+		   ResultRelInfo *returningRelInfo,
 		   EState *estate,
 		   bool canSetTag)
 {
@@ -677,8 +695,45 @@ ExecInsert(ModifyTableState *mtstate,
 		ExecWithCheckOptions(WCO_VIEW_CHECK, resultRelInfo, slot, estate);
 
 	/* Process RETURNING if present */
-	if (resultRelInfo->ri_projectReturning)
-		result = ExecProcessReturning(resultRelInfo, slot, planSlot);
+	if (returningRelInfo->ri_projectReturning)
+	{
+		/*
+		 * In a cross-partition UPDATE with RETURNING, we have to use the
+		 * source partition's RETURNING list, because that matches the output
+		 * of the planSlot, while the destination partition might have
+		 * different resjunk columns.  This means we have to map the
+		 * destination tuple back to the source's format so we can apply that
+		 * RETURNING list.  This is expensive, but it should be an uncommon
+		 * corner case, so we won't spend much effort on making it fast.
+		 *
+		 * We assume that we can use srcSlot to hold the re-converted tuple.
+		 * Note that in the common case where the child partitions both match
+		 * the root's format, previous optimizations will have resulted in
+		 * slot and srcSlot being identical, cueing us that there's nothing to
+		 * do here.
+		 */
+		if (returningRelInfo != resultRelInfo && slot != srcSlot)
+		{
+			Relation	srcRelationDesc = returningRelInfo->ri_RelationDesc;
+			AttrMap    *map;
+
+			map = build_attrmap_by_name_if_req(RelationGetDescr(resultRelationDesc),
+											   RelationGetDescr(srcRelationDesc));
+			if (map)
+			{
+				TupleTableSlot *origSlot = slot;
+
+				slot = execute_attr_map_slot(map, slot, srcSlot);
+				slot->tts_tid = origSlot->tts_tid;
+				slot->tts_tableOid = origSlot->tts_tableOid;
+				free_attrmap(map);
+			}
+		}
+
+		result = ExecProcessReturning(returningRelInfo->ri_projectReturning,
+									  RelationGetRelid(resultRelationDesc),
+									  slot, planSlot);
+	}
 
 	return result;
 }
@@ -1027,7 +1082,9 @@ ldelete:;
 			}
 		}
 
-		rslot = ExecProcessReturning(resultRelInfo, slot, planSlot);
+		rslot = ExecProcessReturning(resultRelInfo->ri_projectReturning,
+									 RelationGetRelid(resultRelationDesc),
+									 slot, planSlot);
 
 		/*
 		 * Before releasing the target tuple again, make sure rslot has a
@@ -1203,6 +1260,7 @@ lreplace:;
 		{
 			bool		tuple_deleted;
 			TupleTableSlot *ret_slot;
+			TupleTableSlot *orig_slot = slot;
 			TupleTableSlot *epqslot = NULL;
 			PartitionTupleRouting *proute = mtstate->mt_partition_tuple_routing;
 			int			map_index;
@@ -1309,6 +1367,7 @@ lreplace:;
 										   mtstate->rootResultRelInfo, slot);
 
 			ret_slot = ExecInsert(mtstate, slot, planSlot,
+								  orig_slot, resultRelInfo,
 								  estate, canSetTag);
 
 			/* Revert ExecPrepareTupleRouting's node change. */
@@ -1505,7 +1564,9 @@ lreplace:;
 
 	/* Process RETURNING if present */
 	if (resultRelInfo->ri_projectReturning)
-		return ExecProcessReturning(resultRelInfo, slot, planSlot);
+		return ExecProcessReturning(resultRelInfo->ri_projectReturning,
+									RelationGetRelid(resultRelationDesc),
+									slot, planSlot);
 
 	return NULL;
 }
@@ -2154,7 +2215,9 @@ ExecModifyTable(PlanState *pstate)
 			 * ExecProcessReturning by IterateDirectModify, so no need to
 			 * provide it here.
 			 */
-			slot = ExecProcessReturning(resultRelInfo, NULL, planSlot);
+			slot = ExecProcessReturning(resultRelInfo->ri_projectReturning,
+										RelationGetRelid(resultRelInfo->ri_RelationDesc),
+										NULL, planSlot);
 
 			estate->es_result_relation_info = saved_resultRelInfo;
 			return slot;
@@ -2244,6 +2307,7 @@ ExecModifyTable(PlanState *pstate)
 					slot = ExecPrepareTupleRouting(node, estate, proute,
 												   resultRelInfo, slot);
 				slot = ExecInsert(node, slot, planSlot,
+								  NULL, estate->es_result_relation_info,
 								  estate, node->canSetTag);
 				/* Revert ExecPrepareTupleRouting's state change. */
 				if (proute)

src/test/regress/expected/update.out

@@ -445,6 +445,46 @@ UPDATE range_parted set c = 95 WHERE a = 'b' and b > 10 and c > 100 returning (r
  part_c_1_100   | b | 17 |  95 | 19 | 
 (6 rows)
 
+-- The following tests computing RETURNING when the source and the destination
+-- partitions of a UPDATE row movement operation have different tuple
+-- descriptors, which has been shown to be problematic in the cases where the
+-- RETURNING targetlist contains non-target relation attributes that are
+-- computed by referring to the source partition plan's output tuple.  Also,
+-- a trigger on the destination relation may change the tuple, which must be
+-- reflected in the RETURNING output, so we test that too.
+CREATE TABLE part_c_1_c_20 (LIKE range_parted);
+ALTER TABLE part_c_1_c_20 DROP a, DROP b, ADD a text, ADD b bigint;
+ALTER TABLE range_parted ATTACH PARTITION part_c_1_c_20 FOR VALUES FROM ('c', 1) TO ('c', 20);
+CREATE FUNCTION trigfunc () RETURNS TRIGGER LANGUAGE plpgsql as $$ BEGIN NEW.e := 'in trigfunc()'; RETURN NEW; END; $$;
+CREATE TRIGGER trig BEFORE INSERT ON part_c_1_c_20 FOR EACH ROW EXECUTE FUNCTION trigfunc();
+UPDATE range_parted r set a = 'c' FROM (VALUES ('a', 1), ('a', 10), ('b', 12)) s(x, y) WHERE s.x = r.a AND s.y = r.b RETURNING tableoid::regclass, *;
+   tableoid    | a | b  |  c  | d |       e       | x | y  
+---------------+---+----+-----+---+---------------+---+----
+ part_c_1_c_20 | c |  1 |   1 | 1 | in trigfunc() | a |  1
+ part_c_1_c_20 | c | 10 | 200 | 1 | in trigfunc() | a | 10
+ part_c_1_c_20 | c | 12 |  96 | 1 | in trigfunc() | b | 12
+(3 rows)
+
+DROP TRIGGER trig ON part_c_1_c_20;
+DROP FUNCTION trigfunc;
+:init_range_parted;
+CREATE FUNCTION trigfunc () RETURNS TRIGGER LANGUAGE plpgsql as $$ BEGIN RETURN NULL; END; $$;
+CREATE TRIGGER trig BEFORE INSERT ON part_c_1_c_20 FOR EACH ROW EXECUTE FUNCTION trigfunc();
+UPDATE range_parted r set a = 'c' FROM (VALUES ('a', 1), ('a', 10), ('b', 12)) s(x, y) WHERE s.x = r.a AND s.y = r.b RETURNING tableoid::regclass, *;
+ tableoid | a | b | c | d | e | x | y 
+----------+---+---+---+---+---+---+---
+(0 rows)
+
+:show_data;
+   partname   | a | b  |  c  | d  | e 
+--------------+---+----+-----+----+---
+ part_c_1_100 | b | 13 |  97 |  2 | 
+ part_d_15_20 | b | 15 | 105 | 16 | 
+ part_d_15_20 | b | 17 | 105 | 19 | 
+(3 rows)
+
+DROP TABLE part_c_1_c_20;
+DROP FUNCTION trigfunc;
 -- Transition tables with update row movement
 :init_range_parted;
 CREATE FUNCTION trans_updatetrigfunc() RETURNS trigger LANGUAGE plpgsql AS

src/test/regress/sql/update.sql

@@ -236,6 +236,31 @@ DROP VIEW upview;
 UPDATE range_parted set c = 95 WHERE a = 'b' and b > 10 and c > 100 returning (range_parted), *;
 :show_data;
 
+-- The following tests computing RETURNING when the source and the destination
+-- partitions of a UPDATE row movement operation have different tuple
+-- descriptors, which has been shown to be problematic in the cases where the
+-- RETURNING targetlist contains non-target relation attributes that are
+-- computed by referring to the source partition plan's output tuple.  Also,
+-- a trigger on the destination relation may change the tuple, which must be
+-- reflected in the RETURNING output, so we test that too.
+CREATE TABLE part_c_1_c_20 (LIKE range_parted);
+ALTER TABLE part_c_1_c_20 DROP a, DROP b, ADD a text, ADD b bigint;
+ALTER TABLE range_parted ATTACH PARTITION part_c_1_c_20 FOR VALUES FROM ('c', 1) TO ('c', 20);
+CREATE FUNCTION trigfunc () RETURNS TRIGGER LANGUAGE plpgsql as $$ BEGIN NEW.e := 'in trigfunc()'; RETURN NEW; END; $$;
+CREATE TRIGGER trig BEFORE INSERT ON part_c_1_c_20 FOR EACH ROW EXECUTE FUNCTION trigfunc();
+UPDATE range_parted r set a = 'c' FROM (VALUES ('a', 1), ('a', 10), ('b', 12)) s(x, y) WHERE s.x = r.a AND s.y = r.b RETURNING tableoid::regclass, *;
+
+DROP TRIGGER trig ON part_c_1_c_20;
+DROP FUNCTION trigfunc;
+
+:init_range_parted;
+CREATE FUNCTION trigfunc () RETURNS TRIGGER LANGUAGE plpgsql as $$ BEGIN RETURN NULL; END; $$;
+CREATE TRIGGER trig BEFORE INSERT ON part_c_1_c_20 FOR EACH ROW EXECUTE FUNCTION trigfunc();
+UPDATE range_parted r set a = 'c' FROM (VALUES ('a', 1), ('a', 10), ('b', 12)) s(x, y) WHERE s.x = r.a AND s.y = r.b RETURNING tableoid::regclass, *;
+:show_data;
+
+DROP TABLE part_c_1_c_20;
+DROP FUNCTION trigfunc;
 
 -- Transition tables with update row movement
 :init_range_parted;