Ceph

I was able to reproduce this.
ESTALE shows up in the messages log.
I don't see much in the mds log.

At this point, the fs is pretty much borked until I go into the ceph direct mount and perform lookup operations.
I turned mds and kernel logging on just before running the operation I new would produce the ESTALE.
I'm not sure how much I'd have to do to go from scratch as this is after copying a source tree to the filesystem with scp, then startup up the nfs export, then mounting, then attempting to build. That encounters an ESTALE and I continue to get the ESTALE after I've restarted everything from ceph on up (without umount/mount of the NFS client).

In this case, the NFS client is holding filehandles while the ceph/NFS server restart.

I don't have things set up to reproduce/test this easily, but it looks like this is the problem. Can you give it a go?

diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index 79743d1..5e0e8d1 100644
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -1571,11 +1571,11 @@ static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
                r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
                dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
                     *ppath);
-       } else if (rpath) {
+       } else if (rpath || rino) {
                *ino = rino;
                *ppath = rpath;
                *pathlen = strlen(rpath);
-               dout(" path %.*s\n", *pathlen, rpath);
+               dout(" path #%llx/%.*s\n", rino, *pathlen, rpath);
        }

        return r;

pushed this to for-linus branch.

going to assume the above fixed it until we hear otherwise :)