Fix incorrect assertion in dnode_dirty_l1range

The db_dirtycnt of an EVICTING dbuf is always 0. However, it still appears in the dn_dbufs tree. If we call dnode_dirty_l1range on a range that contains an EVICTING dbuf, we will attempt to mark it dirty (which will fail because it's EVICTING, resulting in a new dbuf being created and dirtied). Later, in ZFS_DEBUG mode, we assert that all the dbufs in the range are dirty. If the EVICTING dbuf is still present, this will trip the assertion erroneously. Reviewed-by: Matt Ahrens <[email protected]> Reviewed-by: Richard Elling <[email protected]> Reviewed-by: Sara Hartse <[email protected]> Reviewed-by: Brian Behlendorf <[email protected]> Signed-off-by: Paul Dagnelie <[email protected]> Closes openzfs#8745
allanjude · Jun 7, 2019 · f0f04ac · f0f04ac
1 parent 987f50e
commit f0f04ac
Showing 1 changed file with 2 additions and 1 deletion.
diff --git a/module/zfs/dnode.c b/module/zfs/dnode.c
@@ -1969,7 +1969,8 @@ dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
 	for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
 		if (db->db_level != 1 || db->db_blkid >= end_blkid)
 			break;
-		ASSERT(db->db_dirtycnt > 0);
+		if (db->db_state != DB_EVICTING)
+			ASSERT(db->db_dirtycnt > 0);
 	}
 #endif
 	mutex_exit(&dn->dn_dbufs_mtx);