inplace230-index_update_stats-io-before-buflock-v1.patch

Author:     Noah Misch <noah@leadboat.com>
Commit:     Noah Misch <noah@leadboat.com>

    

diff --git a/src/backend/catalog/index.c b/src/backend/catalog/index.c
index 74d0f30..bef1af6 100644
--- a/src/backend/catalog/index.c
+++ b/src/backend/catalog/index.c
@@ -2806,6 +2806,9 @@ index_update_stats(Relation rel,
 				   bool hasindex,
 				   double reltuples)
 {
+	bool		update_stats;
+	BlockNumber relpages;
+	BlockNumber relallvisible;
 	Oid			relid = RelationGetRelid(rel);
 	Relation	pg_class;
 	ScanKeyData key[1];
@@ -2815,6 +2818,42 @@ index_update_stats(Relation rel,
 	bool		dirty;
 
 	/*
+	 * As a special hack, if we are dealing with an empty table and the
+	 * existing reltuples is -1, we leave that alone.  This ensures that
+	 * creating an index as part of CREATE TABLE doesn't cause the table to
+	 * prematurely look like it's been vacuumed.  The final rd_rel may be
+	 * different from rel->rd_rel due to e.g. commit of concurrent GRANT, but
+	 * the commands that change reltuples take locks conflicting with ours.
+	 * (Even if a command changed reltuples under a weaker lock, this affects
+	 * only statistics for an empty table.)
+	 */
+	if (reltuples == 0 && rel->rd_rel->reltuples < 0)
+		reltuples = -1;
+
+	/*
+	 * Don't update statistics during binary upgrade, because the indexes are
+	 * created before the data is moved into place.
+	 */
+	update_stats = reltuples >= 0 && !IsBinaryUpgrade;
+
+	/*
+	 * Finish I/O and visibility map buffer locks before
+	 * systable_inplace_update_begin() locks the pg_class buffer.  The final
+	 * rd_rel may be different from rel->rd_rel due to e.g. commit of
+	 * concurrent GRANT, but no command changes a relkind from non-index to
+	 * index.  (Even if one did, relallvisible doesn't break functionality.)
+	 */
+	if (update_stats)
+	{
+		relpages = RelationGetNumberOfBlocks(rel);
+
+		if (rel->rd_rel->relkind != RELKIND_INDEX)
+			visibilitymap_count(rel, &relallvisible, NULL);
+		else					/* don't bother for indexes */
+			relallvisible = 0;
+	}
+
+	/*
 	 * We always update the pg_class row using a non-transactional,
 	 * overwrite-in-place update.  There are several reasons for this:
 	 *
@@ -2858,15 +2897,6 @@ index_update_stats(Relation rel,
 	/* Should this be a more comprehensive test? */
 	Assert(rd_rel->relkind != RELKIND_PARTITIONED_INDEX);
 
-	/*
-	 * As a special hack, if we are dealing with an empty table and the
-	 * existing reltuples is -1, we leave that alone.  This ensures that
-	 * creating an index as part of CREATE TABLE doesn't cause the table to
-	 * prematurely look like it's been vacuumed.
-	 */
-	if (reltuples == 0 && rd_rel->reltuples < 0)
-		reltuples = -1;
-
 	/* Apply required updates, if any, to copied tuple */
 
 	dirty = false;
@@ -2876,20 +2906,8 @@ index_update_stats(Relation rel,
 		dirty = true;
 	}
 
-	/*
-	 * Avoid updating statistics during binary upgrade, because the indexes
-	 * are created before the data is moved into place.
-	 */
-	if (reltuples >= 0 && !IsBinaryUpgrade)
+	if (update_stats)
 	{
-		BlockNumber relpages = RelationGetNumberOfBlocks(rel);
-		BlockNumber relallvisible;
-
-		if (rd_rel->relkind != RELKIND_INDEX)
-			visibilitymap_count(rel, &relallvisible, NULL);
-		else					/* don't bother for indexes */
-			relallvisible = 0;
-
 		if (rd_rel->relpages != (int32) relpages)
 		{
 			rd_rel->relpages = (int32) relpages;