> Why is there any upper limit at all? The limits should be either
> removed or bumped to even higher value, IMHO.

The only argument I can think of for a limit is so that someone doesn't
try and start 100,000,000 of them. (I am impressed that there is hardware
out there running BSD that can actually make use of 256 of them.)

> Since the worker threads only call nfssvc() in loop, this would be
> reasonable to do. Alternatively, we could just use normal userland
> threads instead of forked processes there, saving MMU switches too.

Actually, they only call nfssvc() once and then remain in the kernel
until killed off via a signal or system crash. (There was old code
in there that had the nfsds pop up to userland to handle Kerberos
tickets, but I don't think anyone except me ever used it.) My newer
nfs code (for V4 etc.) uses an upcall for authentication instead,
so the new nfsd won't ever return to userland.

In summary, sounds like a good idea to me, rick