Benjamin Herrenschmidt wrote:

On Mon, 2009-10-05 at 14:28 -0500, Scott Wood wrote:

quoted

Yes, it hits ptep_set_access_flags() and adding _tlbil_va there helps (I
didn't put it in the filter function, because that doesn't take address as a
parameter).  I'd misread your suggestion as referring to the various changes
to set_pte_filter() that were posted.

As for unconditionally invalidating in set_pte_filter(), that doesn't boot
for me unless I check for kernel addresses -- I guess we populate page
tables that overlap the pinned large page region? 

Good point. I think we do on 8xx. Does doing it in set_pte_filter()
(with the kernel check) makes any difference ? faster ? slower ? no
visible change ?

ptep_set_access_flags() is enough to make the worst of it go away.  I'm 
having another intermittent problem that seems to happen more often 
without the change in set_pte_filter(), but I've yet to narrow down 
what's actually going on there.

-Scott