> There was a *lot* more NFS activity while the machine was ``hung'', but
> this could have been due to something stuck in a loop dumping core and
> restarting (there were inetd.core and getty.core files, for example).

The VM improvements keep the page daemon and possibly other kernel functions
going instead of deadlocking, and it looks like they succeeded in that one
thing for you.

Try disabling sys_mlock() in /sys/vm/vm_mmap.c by having it do nothing
except return 0.

Ross Harvey