On Fri, Sep 13, 2013 at 4:00 PM, Al Viro <viro@zeniv.linux.org.uk> wrote:
\>
> It is right - for one thing, we are holding the lock on that LRU list,
> so list_lru_del() would deadlock right there.  For another, the same
> list_lru_walk (OK, list_lru_walk_node()) will do ->nr_items decrement
> when we return LRU_REMOVED to it, so we don't want to do it twice.
> Plain list_del_init() is correct here.

Yes. And I found the opposite bug in one place: when we are collecting
dentries by walking the parents etc, we do *not* hold the global RCU
lock, so we cannot use the "d_lru_shrink_list()" thing after all. It's
correct as far as the internal logic of fs/dcache.c goes, but it
violates the global LRU list rules. So I replaced that with a
dentry_lru_del() followed by a d_shrink_add() instead.

Updated patch attached.

                        Linus