Prevent pgstats from getting confused when relkind of a relation changes

anarazel · anarazel · commit c6a60471a1a5 · 2022-12-02T18:07:47.000-08:00
When the relkind of a relache entry changes, because a table is converted into a view, pgstats can get confused in 15+, leading to crashes or assertion failures. For HEAD, Tom fixed this in b23cd18, by removing support for converting a table to a view, removing the source of the inconsistency. This commit just adds an assertion that a relcache entry's relkind does not change, just in case we end up with another case of that in the future. As there's no cases of changing relkind anymore, we can't add a test that that's handled correctly. For 15, fix the problem by not maintaining the association with the old pgstat entry when the relkind changes during a relcache invalidation processing. In that case the pgstat entry needs to be unlinked first, to avoid PgStat_TableStatus->relation getting out of sync. Also add a test reproducing the issues. No known problem exists in 11-14, so just add the test there. Reported-by: vignesh C <vignesh21@gmail.com> Author: Andres Freund <andres@anarazel.de> Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us> Discussion: https://postgr.es/m/CALDaNm2yXz+zOtv7y5zBd5WKT8O0Ld3YxikuU3dcyCvxF7gypA@mail.gmail.com Discussion: https://postgr.es/m/CALDaNm3oZA-8Wbps2Jd1g5_Gjrr-x3YWrJPek-mF5Asrrvz2Dg@mail.gmail.com Backpatch: 15-
diff --git a/src/backend/utils/cache/relcache.c b/src/backend/utils/cache/relcache.c
@@ -2626,6 +2626,7 @@ RelationClearRelation(Relation relation, bool rebuild)
 		bool		keep_rules;
 		bool		keep_policies;
 		bool		keep_partkey;
+		bool		keep_pgstats;
 
 		/* Build temporary entry, but don't link it into hashtable */
 		newrel = RelationBuildDesc(save_relid, false);
@@ -2667,6 +2668,21 @@ RelationClearRelation(Relation relation, bool rebuild)
 		/* partkey is immutable once set up, so we can always keep it */
 		keep_partkey = (relation->rd_partkey != NULL);
 
+		/*
+		 * Keep stats pointers, except when the relkind changes (e.g. when
+		 * converting tables into views). Different kinds of relations might
+		 * have different types of stats.
+		 *
+		 * If we don't want to keep the stats, unlink the stats and relcache
+		 * entry (and do so before entering the "critical section"
+		 * below). This is important because otherwise
+		 * PgStat_TableStatus->relation would get out of sync with
+		 * relation->pgstat_info.
+		 */
+		keep_pgstats = relation->rd_rel->relkind == newrel->rd_rel->relkind;
+		if (!keep_pgstats)
+			pgstat_unlink_relation(relation);
+
 		/*
 		 * Perform swapping of the relcache entry contents.  Within this
 		 * process the old entry is momentarily invalid, so there *must* be no
@@ -2720,9 +2736,14 @@ RelationClearRelation(Relation relation, bool rebuild)
 			SWAPFIELD(RowSecurityDesc *, rd_rsdesc);
 		/* toast OID override must be preserved */
 		SWAPFIELD(Oid, rd_toastoid);
+
 		/* pgstat_info / enabled must be preserved */
-		SWAPFIELD(struct PgStat_TableStatus *, pgstat_info);
-		SWAPFIELD(bool, pgstat_enabled);
+		if (keep_pgstats)
+		{
+			SWAPFIELD(struct PgStat_TableStatus *, pgstat_info);
+			SWAPFIELD(bool, pgstat_enabled);
+		}
+
 		/* preserve old partition key if we have one */
 		if (keep_partkey)
 		{
diff --git a/src/test/regress/expected/create_view.out b/src/test/regress/expected/create_view.out
@@ -2122,6 +2122,29 @@ select pg_get_viewdef('tt26v', true);
     FROM ( VALUES (1,2,3)) v(x, y, z);
 (1 row)
 
+-- Test that changing the relkind of a relcache entry doesn't cause
+-- trouble. Prior instances of where it did:
+-- CALDaNm2yXz+zOtv7y5zBd5WKT8O0Ld3YxikuU3dcyCvxF7gypA@mail.gmail.com
+-- CALDaNm3oZA-8Wbps2Jd1g5_Gjrr-x3YWrJPek-mF5Asrrvz2Dg@mail.gmail.com
+CREATE TABLE tt26(c int);
+BEGIN;
+CREATE TABLE tt27(c int);
+SAVEPOINT q;
+CREATE RULE "_RETURN" AS ON SELECT TO tt27 DO INSTEAD SELECT * FROM tt26;
+SELECT * FROM tt27;
+ c 
+---
+(0 rows)
+
+ROLLBACK TO q;
+CREATE RULE "_RETURN" AS ON SELECT TO tt27 DO INSTEAD SELECT * FROM tt26;
+ROLLBACK;
+BEGIN;
+CREATE TABLE tt28(c int);
+CREATE RULE "_RETURN" AS ON SELECT TO tt28 DO INSTEAD SELECT * FROM tt26;
+CREATE RULE "_RETURN" AS ON SELECT TO tt28 DO INSTEAD SELECT * FROM tt26;
+ERROR:  "tt28" is already a view
+ROLLBACK;
 -- clean up all the random objects we made above
 DROP SCHEMA temp_view_test CASCADE;
 NOTICE:  drop cascades to 27 other objects
@@ -2153,7 +2176,7 @@ drop cascades to view aliased_view_2
 drop cascades to view aliased_view_3
 drop cascades to view aliased_view_4
 DROP SCHEMA testviewschm2 CASCADE;
-NOTICE:  drop cascades to 77 other objects
+NOTICE:  drop cascades to 78 other objects
 DETAIL:  drop cascades to table t1
 drop cascades to view temporal1
 drop cascades to view temporal2
@@ -2231,3 +2254,4 @@ drop cascades to view tt23v
 drop cascades to view tt24v
 drop cascades to view tt25v
 drop cascades to view tt26v
+drop cascades to table tt26
diff --git a/src/test/regress/sql/create_view.sql b/src/test/regress/sql/create_view.sql
@@ -781,6 +781,29 @@ select x + y + z as c1,
 from (values(1,2,3)) v(x,y,z);
 select pg_get_viewdef('tt26v', true);
 
+
+-- Test that changing the relkind of a relcache entry doesn't cause
+-- trouble. Prior instances of where it did:
+-- CALDaNm2yXz+zOtv7y5zBd5WKT8O0Ld3YxikuU3dcyCvxF7gypA@mail.gmail.com
+-- CALDaNm3oZA-8Wbps2Jd1g5_Gjrr-x3YWrJPek-mF5Asrrvz2Dg@mail.gmail.com
+CREATE TABLE tt26(c int);
+
+BEGIN;
+CREATE TABLE tt27(c int);
+SAVEPOINT q;
+CREATE RULE "_RETURN" AS ON SELECT TO tt27 DO INSTEAD SELECT * FROM tt26;
+SELECT * FROM tt27;
+ROLLBACK TO q;
+CREATE RULE "_RETURN" AS ON SELECT TO tt27 DO INSTEAD SELECT * FROM tt26;
+ROLLBACK;
+
+BEGIN;
+CREATE TABLE tt28(c int);
+CREATE RULE "_RETURN" AS ON SELECT TO tt28 DO INSTEAD SELECT * FROM tt26;
+CREATE RULE "_RETURN" AS ON SELECT TO tt28 DO INSTEAD SELECT * FROM tt26;
+ROLLBACK;
+
+
 -- clean up all the random objects we made above
 DROP SCHEMA temp_view_test CASCADE;
 DROP SCHEMA testviewschm2 CASCADE;
