Hi,
I found the cause of the frequent address exceptions on my A5000 under 2.2.
What was happening was that the pte for a data page was disappearing but the
mm area remained intact. This meant that the next time the process faulted in
that page the kernel thought it was an anonymous mapping and provided a zero
page instead. This tends to lead to a NULL dereference pretty quickly.
The root cause is an assumption in copy_page_range that pagetable structures
are naturally aligned. Because 26-bit machines have fake pagetables that
aren't used by the hardware, this isn't necessarily true. The logic that
works out how large an area is by masking address bits was getting confused.
Things now seem a lot more stable. I'll include a fix for this in my next
kernel patch.
p.
unsubscribe: body of `unsubscribe linux-arm' to [EMAIL PROTECTED]
