Ian Dall <Ian.Dall@dsto.defence.gov.au> writes:

> I'd work on the assumption that it is a problem in machine dependent
> code since no one else has reported it. Vnodes are used by mmap code
> which interacts with pmap, so maybe something in there?

This was a bit vague. I was also writing from memory. I was thinking
mmap, because of the frequency of occurrance goes up with little
memory.  Of course, low physical memory might stress other things as
well.  Also, I reasoned that the filesystem code was all being
examined from the point of view of locking problems, but maybe the
memory mapping code (see uvm_mmap.c) had been neglected. There is also
the possibility that our MD pmap code is interacting somehow with uvm
to cause a problem with the mmap code. I don't know if the unified
buffer cache is tied up in all this.

> Maybe a grep
> for "vnode" in the pc532 part of the tree would be a start (I'm sure I
> did this but don't remember what I found).

This was a red herring. The only use of vnode structures seems to be
to create coredumps.

Ian