ceph_fsync race with reconnect?

[Jeff Layton <jlayton@xxxxxxxxxx>]

I've been working on a patchset to add inline write support to kcephfs,
and have run across a potential race in fsync. I could use someone to
sanity check me though since I don't have a great grasp of the MDS
session handling:

ceph_fsync() calls try_flush_caps() to flush the dirty metadata back to
the MDS when Fw caps are flushed back.  try_flush_caps does this,
however:

                if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
                        spin_unlock(&ci->i_ceph_lock);
                        goto out;
                }

...at that point, try_flush_caps will return 0, and set *ptid to 0 on
the way out. ceph_fsync won't see that Fw is still dirty at that point
and won't wait, returning without flushing metadata.

Am I missing something that prevents this? I can open a tracker bug for
this if it is a problem, but I wanted to be sure it was a bug before I
did so.

Thanks,
-- 
Jeff Layton <jlayton@xxxxxxxxxx>