Re: [GIT PULL] xen /proc/mtrr implementation

[ebiederm@xmission.com (Eric W. Biederman)]

Jesse Barnes <jbarnes@virtuousgeek.org> writes:

> On Fri, 15 May 2009 16:49:12 -0700
> Jeremy Fitzhardinge <jeremy@goop.org> wrote:
>
>> Eric W. Biederman wrote:
>> >> /proc/mtrr is in wide use today.  It may be planned for
>> >> obsolescence, but there's no way you can claim its obsolete today
>> >> (my completely up-to-date F10 X server is using it, for example).
>> >> We don't break oldish usermode ABIs in new kernels.
>> >>     
>> >
>> > Sure it is.  There is a better newer replacement.  It is taking a
>> > while to get userspace transitioned but that is different.
>> > Honestly I am puzzled why that it but whatever.
>> >   
>> 
>> There's no mention in feature-removal-schedule.txt.

I don't know that it makes sense to remove mtrrs but it certainly
doesn't make sense to use them if you can avoid it.

>> >> Besides, the MTRR code is also a kernel-internal API, used by DRM
>> >> and other drivers to configure the system MTRR state.  Those
>> >> drivers will either perform badly or outright fail if they can't
>> >> set the appropriate cachability properties. That is not obsolete
>> >> in any way. 
>> >
>> > There are about 5 of them so let's fix them.
>> >   
>> 
>> Well, I count at least 30+, but anyway.

Wow.  We had a lot of those slip in.  Definitely time to fix the
drivers.

>> > With PAT we are in a much better position both for portability and
>> > for flexibility.
>> >   
>> 
>> PAT is relatively recent, and even more recently bug-free.  There are 
>> many people with processors which can't or won't do PAT; what's the
>> plan to support them?  Just hit them with a performance regression?
>> Or wrap MTRR in some other API?

PPro is roughly when PAT came out.  I remember discussing this a while
ago and the conclusion was that there are very few systems with MTRRs
that don't have a usable PAT implementation.  I expect many of those
systems are on their last legs today.

>> Sure, when available.  We're sorting out the details for Xen, but
>> even then it may not be available, either because we're running on an
>> old version of Xen, or because some other guest is using PAT
>> differently.

There are only 3 states that are interesting.  WB UC and WC.  Since
Xen controls the page tables anyway.  I expect it can even remap
it feels like it.

>> But I honestly don't understand the hostility towards 120 lines of
>> code to make an interface (albeit legacy/deprecated/whatever) behave
>> in an expected way.

> FWIW I think supporting the MTRR API in Xen makes sense.  There's a lot
> of old code out there that wants it; would be nice if it mostly worked,
> especially at such a minimal cost.  It's taken awhile to get PAT going
> (and there are still issues here and there) so having the MTRR stuffa
> available is awfully nice.

I won't argue that having MTRRs when you can makes sense.  It is a bit
weird in a vitalized system.  At a practical level there are an
increasing number of systems for which MTRRs are unusable because the
BIOS sets up overlapping mtrrs.  With cheap entry level systems
shipping with 4G I expect it is becoming a majority of systems.

Eric