Re: [PATCH] anobjrmap 1/6 objrmap

[Andrea Arcangeli <andrea@suse.de>]

On Sat, Mar 20, 2004 at 06:03:41AM -0800, William Lee Irwin III wrote:
> On Sat, Mar 20, 2004 at 01:30:09PM +0100, Andrea Arcangeli wrote:
> > I'm working on my code yes, I think my code is finished, I prefer my
> > design for the various reasons explained in the other emails (you don't
> > swap so you can't appreciate the benefits, you only have to check that
> > performs as well as Hugh's code).
> > Hugh's and your code is unstable in objrmap, you can find the details in
> > the email I sent to Hugh, mine is stable (running such simulation for a
> > few days just fine on 4-way xeon, without my objrmap fixes it live locks
> > as soon as it hits swap).
> > You find my anon_vma in 2.6.5-rc1aa2, it's rock solid, just apply the
> > whole patch and compare it with your other below results. thanks.
> 
> There's an outstanding issue that's biting people on ppc64, which is
> that arch/ppc64/mm/tlb.c uses the mm pointer and virtual addresses that
> used to be put into page->mapping and page->index respectively for
> pagetable pages to assist updates to the inverted pagetable. Without
> leaving that assignment and invalidation of page->mapping and page->index
> in place or converting ppc64 to other methods of carrying out its
> inverted pagetable updates, ppc64 (e.g. G5 Macs) support is broken.

agreed, I was talking all archs but ppc of course when I said rock
solid. (it's not yet applied to ppc at this time), after reading Hugh's
and Paul's comments it should be easily fixable (those pages are not the
ones mapped to userspace, so their page->mapping and page->index can be
reused). I'll try to find a cross compiler soon to fix it. The other
archs seems already working fine.