Re: Regression in next with ext4 oops

From: Al Viro
Date: Tue Oct 04 2016 - 10:59:27 EST

On Tue, Oct 04, 2016 at 10:02:31AM -0400, Theodore Ts'o wrote:
> On Tue, Oct 04, 2016 at 11:00:41AM +0200, Jan Kara wrote:
> > Never seen this but I suspect it is a fallout from Al's directory locking
> > changes. In particular ext4_htree_fill_tree() builds rb-tree of found
> > directory entries in file->private_data (and generally modifies the
> > structure stored there) but after Al's changes we don't have exclusive
> > access to struct file if I'm right so if two processes end up calling
> > getdents() for the same 'struct file' we are doomed.
> I haven't seen it either, and I've been doing a lot of testing on the
> ext4 test branch. So I'm guessing Tony has the only reliable repro
> for the problem at the moment. That being said, it shouldn't be that
> hard to create a test case for this and add it to xfstests.
> I'm pretty sure Jan is right about this, though, but it would be great
> to a get a quick confirmation from Tony if at all possible.

Jan is wrong - we do have per-struct-file serialization for getdents() It might be a race between getdents() on *different* struct
file for the same directory, but ->private_data is not a problem.