OK, I've built a DDB kernel and have identified that when the system hangs
it does so because everyone is getting stuck in getnewbuf() (vfs_bio.c) The
only sleep there is on needbuffer. And brelse should wake up the process if
there was a free buffer but it appears as if all the buffers have been lost
(aka buffer leak somewhere else in the kernel) since this happens
relatively quickly when doing diskio I'm guessing that somewhere in the
disk stuff there is a leak. (not the vfs layer since that would affect nfs
as well.) I'm using vmstat -m to track memory usage and I'll let you know
what I find out.
--Chuck