Andrew, I have test this patch for a while now and none of the users has seen
the "busy inodes" message for a while now. Can you please apply and test it in
-mm?

This is an updated version of the patch which adresses some issues that came
up during discussion. Although sb->prunes usually is 0, I'm testing it now
before calling wake_up(). Besides that, the shrink_dcache_parent() is only
waiting for prunes if we are called through generic_shutdown_super() when
sb->s_root is NULL.

Original patch description:

Kirill Korotaev <dev@sw.ru> discovered a race between shrink_dcache_parent()
and shrink_dcache_memory() which leads to "Busy inodes after unmount".
When unmounting a file system shrink_dcache_parent() is racing against a
possible shrink_dcache_memory(). This might lead to the situation that
shrink_dcache_parent() is returning too early. In this situation the
super_block is destroyed before shrink_dcache_memory() could put the inode.

This patch fixes the problem through introducing a prunes counter which is
incremented when a dentry is pruned but the corresponding inoded isn't put
yet.When the prunes counter is not null, shrink_dcache_parent() is waiting and
restarting its work.

Signed-off-by: Jan Blunck <jblunck@suse.de>