I think I have 8xx working...
joakim.tjernlund at transmode.se
Thu Oct 15 16:42:05 EST 2009
Benjamin Herrenschmidt <benh at kernel.crashing.org> wrote on 15/10/2009 03:12:50:
> On Wed, 2009-10-14 at 18:08 -0700, Rex Feany wrote:
> > Thus spake Benjamin Herrenschmidt (benh at kernel.crashing.org):
> > > On Wed, 2009-10-14 at 17:41 -0700, Rex Feany wrote:
> > > > The biggest problem for me turned out to be the MMU context IDs being
> > > > clamped to 32 when the 8xx only has 16. With this, things are a bit more
> > > > stable :)
> > >
> > > This is with Joakim patches or without ?
> > with just one: adding tlbil_va() to do_page_fault().
> Ah cool, at least that keeps my sanity, I didn't get how it could work
> at all without that bit :-)
> Scott, Joakim: It will be interesting to figure out exactly what is the
> condition where that is necessary. It definitely should make the one
> in set_pte_filter() or ptep_set_access_flags_filter() unnecessary.
> Joakim, you mentioned increased amount of TLB errors or misses when
> doing that too often, that sounds a bit weird and worth investigating
I didn't say that, did I? More like:
if I don't do tlbil_va at all I get a lot of extra/duplicate TLB errors
for the same address. Adding the patch makes these go away.
I guess one could do tlbil_va unconditionally but I didn't
see any improvements from doing that, but this was all on 2.4
> Finally we could implement preload from update_mmu_cache() but that's a
> different matter alltogether.
More information about the Linuxppc-dev