I have been searching for a bug causing 'page invalid' panics lately.
It only showed up on KA650 systems, and after digging deeper into the
problem I found the following:

A pagefault in user page tables causes a trap to be taken through the
'translation invalid' vector, even if the protection for the page table
mapping in system page table are 'no access' and should therefore go
through the 'access violation' vector. Section 5.7 in 'VAX Architecture
Reference Manual' states that "If both faults occur, then the Access
Violation fault takes precedence". This seems to be true for all but
the KA650 CPUs.

Does anyone have any ideas about? Someone who has worked at DEC? 
I have made a solution that seems to work; stress-testing a
MV3900 overnight now.

-- Ragge