Re: [BISECTED] Linux 3.12.7 introduces page map handling regression

[Mel Gorman <mgorman@suse.de>]

On Wed, Jan 22, 2014 at 01:39:32PM -0500, Rik van Riel wrote:
> On 01/22/2014 01:24 PM, Linus Torvalds wrote:
> >On Wed, Jan 22, 2014 at 10:07 AM, Rik van Riel <riel@redhat.com> wrote:
> >>
> >>The difference between a numa pte and a protnone pte is
> >>the VMA permissions.
> >
> >If that is indeed the only difference, then we should damn well get
> >rid of that f*cking stupid _PAGE_NUMA name entirely.
> >
> >It's misleading crap. Really. Just do a quick grep for that bit, and
> >you see just *how* confused people are about it:
> >
> >    #define _PAGE_NUMA _PAGE_PROTNONE
> >    ...
> >         if ((pte_flags(a) & (_PAGE_PROTNONE | _PAGE_NUMA)) &
> >
> >think about it. Just *THINK* about how broken that code is. The whole
> >thing is a disaster. _PAGE_NUMA must die. It's shit.
> 
> The reason things are this way is that we were
> not sure whether we can indeed use _PAGE_PROTNONE
> for NUMA balancing on all architectures.
> 

Power is not using _PAGE_PROTNONE to trap NUMA hinting faults because they
do not have that bit. Instead they reuse _PAGE_COHERENT with various tricks,
patches are in -next.

92c08a0d522c7e62c01a63e42597f0c2b02c4245 powerpc/mm: Use HPTE constants when updating hpte bits
c8c06f5a0dde0fed260c54d550962187f266ed0d powerpc/mm: Free up _PAGE_COHERENCE for numa fault use later
8937ba48dcf62b5cdf7abb93652914af16756f50 powerpc/mm: Only check for _PAGE_PRESENT in set_pte/pmd functions
c34a51ce49b40b9667cd7f5cc2e40475af8b4c3d powerpc/mm: Enable _PAGE_NUMA for book3s

As confusing as _PAGE_NUMA is, the intent was to express support in an
architecture-independent manner. If we had started with _PAGE_PROTNONE
then it would still be ambiguous -- are we interested in NUMA hinting
information or is this really PROTNONE protection? The power people would
then have had to add something like _PAGE_NUMA anyway when they had no
_PAGE_PROTNONE bit or define _PAGE_PROTNONE as _PAGE_COHERENT which is
just a different type of misleading.

> If we are sure that _PAGE_PROTNONE can be used
> everywhere, I agree we should get rid of the whole
> _PAGE_NUMA naming, and replace that ambiguous
> code with some comments and documentation instead.

We are sure that _PAGE_PROTNONE cannot be used everywhere.

-- 
Mel Gorman
SUSE Labs