ceph: fix inode removal from snap realm when racing with migration
authorSage Weil <sage@newdream.net>
Tue, 9 Mar 2010 20:59:08 +0000 (12:59 -0800)
committerSage Weil <sage@newdream.net>
Sun, 21 Mar 2010 04:33:08 +0000 (21:33 -0700)
When an inode was dropped while being migrated between two MDSs,
i_cap_exporting_issued was non-zero such that issue caps were non-zero and
__ceph_is_any_caps(ci) was true.  This prevented the inode from being
removed from the snap realm, even as it was dropped from the cache.

Fix this by dropping any residual i_snap_realm ref in destroy_inode.

Signed-off-by: Sage Weil <sage@newdream.net>
fs/ceph/inode.c

index 7abe1aed819b080de138a17ec16f8d7d9c7d9523..aca82d55cc53882a5fc3af08955e56a9f317e8cd 100644 (file)
@@ -378,6 +378,22 @@ void ceph_destroy_inode(struct inode *inode)
 
        ceph_queue_caps_release(inode);
 
+       /*
+        * we may still have a snap_realm reference if there are stray
+        * caps in i_cap_exporting_issued or i_snap_caps.
+        */
+       if (ci->i_snap_realm) {
+               struct ceph_mds_client *mdsc =
+                       &ceph_client(ci->vfs_inode.i_sb)->mdsc;
+               struct ceph_snap_realm *realm = ci->i_snap_realm;
+
+               dout(" dropping residual ref to snap realm %p\n", realm);
+               spin_lock(&realm->inodes_with_caps_lock);
+               list_del_init(&ci->i_snap_realm_item);
+               spin_unlock(&realm->inodes_with_caps_lock);
+               ceph_put_snap_realm(mdsc, realm);
+       }
+
        kfree(ci->i_symlink);
        while ((n = rb_first(&ci->i_fragtree)) != NULL) {
                frag = rb_entry(n, struct ceph_inode_frag, node);