Messages in this thread | | | Date | Sun, 26 Oct 2014 11:56:08 -0700 | Subject | Re: fs: lockup on rename_mutex in fs/dcache.c:1035 | From | Linus Torvalds <> |
| |
On Sat, Oct 25, 2014 at 8:57 PM, Al Viro <viro@zeniv.linux.org.uk> wrote: > [context for Linus] > > Fuzzer has triggered deadlock in d_walk() with rename_lock taken twice. > AFAICS, the plausible scenario is > (child->d_flags & DCACHE_DENTRY_KILLED) || > triggering while ascending to parent, on the pass with rename_lock already > held exclusive. In that case we go to rename_retry and either return without > unlocking rename_lock, or try to take in exclusive one more time, again > without unlocking it first.
Your patch looks fine, and I don't think we can livelock - because we always set 'seq' to 1 if we retry, and that causes us to get the exclusive lock, so we'd better not then enter the retry loop again. Although I guess not only renames cause it - looks like we get to that mis-named "rename_retry" for a deletion too according to the comment. Although I'm not sure that comment is correct - I don't think we change d_parent for deletes, do we?
So at a quick glance, the one-liner patch looks fine. That said, I really detest how that code is written. Especially if the rename_retry really can only happen once, I get the feeling that we would be *much* better off doing this explicitly with a wrapper function, and do something like
void d_walk(,..) { /* Try non-exclusive first */ seq = read_seqbegin(lock); retry = __d_walk(.. seq); if (retry) { read_seqlock_excl(lock); // lock->seqcount is now guaranteed stable retry = __d_walk(.. lock->seqcount); read_sequnlock_excl(lock); WARN_ON_ONCE(retry, "Really?"); } }
or whatever. But maybe I'm missing some reason why the above is just crazy, and I'm a moron. Entirely possible.
Linus
| |