(In reply to Andy Furniss from comment #27)
> Created attachment 118665 [details]
> valley vm fault dump
> 
> Aha, here it is.

I am getting similar results using check_vm. However, since the hang usually
doesn't happen after the first vm fail (it did actually for me a few times, but
most of the time it didn't) I modified the mesa code to disable closing the
program once a vm fault happened. Unfortunately this doesn't really provide me
with any interesting information - they have all complained about the paging
fault happening in a VERTEX_BUFFER with an identical, or close to identical
address.