There are actually rare races where this is possible (e.g. if a new open
intervenes between the read of i_writecount and the fi_fds).
Signed-off-by: J. Bruce Fields <bfields@redhat.com>