On Sat, Oct 28, 2023 at 03:45:13PM +0530, Amit Kapila wrote: > Yes, we need it to exclude any concurrent in-progress scans that could > return incorrect tuples during bucket squeeze operation.
Thanks. So I assume that we should just set REGBUF_NO_CHANGE when the primary and write buffers are the same and there are no tuples to move. Say with something like the attached? -- Michael
diff --git a/src/backend/access/hash/hashovfl.c b/src/backend/access/hash/hashovfl.c
index 9d1ff20b92..9928068179 100644
--- a/src/backend/access/hash/hashovfl.c
+++ b/src/backend/access/hash/hashovfl.c
@@ -647,6 +647,7 @@ _hash_freeovflpage(Relation rel, Buffer bucketbuf, Buffer ovflbuf,
xl_hash_squeeze_page xlrec;
XLogRecPtr recptr;
int i;
+ int wbuf_flags;
xlrec.prevblkno = prevblkno;
xlrec.nextblkno = nextblkno;
@@ -668,7 +669,15 @@ _hash_freeovflpage(Relation rel, Buffer bucketbuf, Buffer ovflbuf,
XLogRegisterBuffer(0, bucketbuf, flags);
}
- XLogRegisterBuffer(1, wbuf, REGBUF_STANDARD);
+ /*
+ * If the bucket buffer and the primary buffer are the same, a cleanup
+ * lock is still required on the write buffer even if there no tuples
+ * to move, so it needs to be registered in this case.
+ */
+ wbuf_flags = REGBUF_STANDARD;
+ if (xlrec.is_prim_bucket_same_wrt && xlrec.ntups == 0)
+ wbuf_flags |= REGBUF_NO_CHANGE;
+ XLogRegisterBuffer(1, wbuf, wbuf_flags);
if (xlrec.ntups > 0)
{
XLogRegisterBufData(1, (char *) itup_offsets,
signature.asc
Description: PGP signature
