On Mon, 24 May 1999, Jonathan Stone wrote:

> >Thoughts?
> 
> Sounds very elegant. but weren't stackable filesystems also meant to
> support putting the acutal filesystem code into userspace (or via
> userspace)? Like the Solaris inkernel support for automount points?

Yes, and that's not affected one way or the other with this change. :-)
For nodes involved in such a filesystem, v_vnlock would point to v_lock in
the vnode, and the lock/unlock/islocked routines would work differently.

> In such cases we really need to go the full RPC route like we do now
> (or rather, like nullfs doesn't), *and* enforce the VOP_LOOKUP- style
> calls acquire locks in the same order as VOP_LOCK.

Actually we'd need VOP_LOCK to do things bottom-up rather than the top
down it does now. The way these calls work, the VOP_LOOKUP-style calls
will always lock the lower node then the upper one (in these calls the
lower vnode comes into existance as a result of the call on the lower
level, so the upper level(s) have to come in after the fact).

I think the "full RPC route" would only be needed by heavier weight fs's
than nullfs.

> Do we have  any such cases now? In future?

None that I'm aware of, and I'm willing to cross that bridge when we come
to it. :-)

Take care,

Bill