Re: /hurd/tmpfs problem

[Roland McGrath]

I don't see that there is anywhere that explicitly states that the
translator connection protocol requires holding the underlying node send
right in order to stay attached as the active translator on that node.  Nor
is there an unavoidable need for that requirement.  The parent filesystem
could just get dead-name notification on the fsys control port or reap
stale control ports at convenient times.  

But it seems like a reasonable convention to set in the protocol, and it
avoids needing to implement any of that.  It just ought to be documented in
the fsys.defs comments.

I've changed tmpfs to hold onto its realnode reference.