When a filesystem is unmounted, we currently call fsnotify_sb_delete()
before evict_inodes(), which means that fsnotify_unmount_inodes()
must iterate over all inodes on the superblock, even though it will
only act on inodes with a refcount. This is inefficient and can lead
to livelocks as it iterates over many unrefcounted inodes.
However, since fsnotify_sb_delete() and evict_inodes() are working
on orthogonal sets of inodes (fsnotify_sb_delete() only cares about nonzero
refcount, and evict_inodes() only cares about zero refcount), we can swap
the order of the calls. The fsnotify call will then have a much smaller
list to walk (any refcounted inodes).
This should speed things up overall, and avoid livelocks in
fsnotify_unmount_inodes().
Signed-off-by: Eric Sandeen <sandeen@xxxxxxxxxx>
---
I just did basic sanity testing here, but AFAIK there is no *notify
test suite, so I'm not sure how to really give this a workout.
diff --git a/fs/notify/fsnotify.c b/fs/notify/fsnotify.c
index ac9eb273e28c..426f03b6e660 100644
--- a/fs/notify/fsnotify.c
+++ b/fs/notify/fsnotify.c
@@ -57,6 +57,10 @@ static void fsnotify_unmount_inodes(struct super_block *sb)
* doing an __iget/iput with SB_ACTIVE clear would actually
* evict all inodes with zero i_count from icache which is
* unnecessarily violent and may in fact be illegal to do.
+ *
+ * However, we should have been called /after/ evict_inodes
+ * removed all zero refcount inodes, in any case. Test to
+ * be sure.
*/
if (!atomic_read(&inode->i_count)) {
spin_unlock(&inode->i_lock);
diff --git a/fs/super.c b/fs/super.c
index cfadab2cbf35..cd352530eca9 100644
--- a/fs/super.c
+++ b/fs/super.c
@@ -448,10 +448,12 @@ void generic_shutdown_super(struct super_block *sb)
sync_filesystem(sb);
sb->s_flags &= ~SB_ACTIVE;
- fsnotify_sb_delete(sb);
cgroup_writeback_umount();
+ /* evict all inodes with zero refcount */
evict_inodes(sb);
+ /* only nonzero refcount inodes can have marks */
+ fsnotify_sb_delete(sb);
if (sb->s_dio_done_wq) {
destroy_workqueue(sb->s_dio_done_wq);
