On Tue, 14 Jul 2015, Gregory Farnum wrote: > The following dentry. When it gets unlinked it also gets removed from the > xlist we're using to traverse the directory contents. We could add some kind > of refcounting to try and make it do a NULL or similat but I don't think > anything like that exists yet. Yeah that sounds likely. I think that's the fix, though.. and I have some recollection of conditions where we favor making the dentry null instead of removing it, so hopefully it's not to involved. It's probably a better behavior anyway for caching reasons, in case the directory isn't complete? > (Sorry this is probably going to get bounced from the list, I'm on my phone > and haven't found an html-free way to send from it...) My reply should go through at least :) sage > > > On Tue, Jul 14, 2015, 7:36 PM Sage Weil <sweil@xxxxxxxxxx> wrote: > On Tue, 14 Jul 2015, Gregory Farnum wrote: > > I spent a bunch of today looking at > http://tracker.ceph.com/issues/12297. > > > > Long story short: the workload is doing a readdir at the same > time as > > it's unlinking files. The readdir functions (in this case, > > _readdir_cache_cb) drop the client_lock each time they invoke > the > > callback (for obvious reasons). There is some effort in > > _readdir_cache_cb to try and keep the iterator valid (we check > on each > > loop that we aren't at end; we increment the iterator before > dropping > > the lock), but it's not sufficient. > > > > Is there supposed to be something preventing this kind of > race? If not > > I can work something out in the code but I've not done much > work in > > that bit and there are enough pieces that I wonder if I'm > missing some > > other issue. > > What is the race you're worried about? Unlinking the file that > we're > doing the callback on, or the one that follows it (where the > iterator now > points)? > > My guess is that in this case unlink should see that there is a > reference > on the dentry and should make it NULL instead of unlinking it > from the > directory entirely... > > sage > > >
