Re: XEN MTRR

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Sun, Jun 03, 2012 at 05:31:32PM +1000, aorchis@gmail.com wrote:
> Hi Jeremy and Konrad,

CC-ing xen-devel.

> 
> Basically the driver NVIDIA provided is a binary blob and recent
> versions does not work with the PAT layout of XEN so it falls back to
> MTRR to provide write combining (please correct me if I'm wrong).

OK? Which is still OK. Are you using a v3.4 kernel with an up-to-date
NVidia driver? I've had reports that it works OK.

> However there is no MTRR support on XEN so the driver hard crashed my
> machine (I can't ssh into the box anymore).
> 
> Moreover, there is problem with the open source driver 'nouveau' for
> NVIDIA card  (also has something to do with PAT layout of XEN) which
> causes memory corruption.

Huh? Can you point me to a bugzilla please? There was a corruption
issue where you can pass in 'nopat' on the command line.

> 
> I found several patches for XEN which supposedly provide basic MTRR
> support for XEN however there is still no /proc/mtrr. Jeremy, can you
> tell me if you had been able to get /proc/mtrr on XEN dom0?

> 
> Thanks for your time.
> 
> Damien.
> 
> 
> On Sun, Jun 3, 2012 at 5:37 AM, Jeremy Fitzhardinge <jeremy@goop.org> wrote:
> > On 06/02/2012 03:13 AM, aorchis@gmail.com wrote:
> >> Hi Jeremy,
> >>
> >> Is there any way I can get back MTRR support in XEN in 3.0 kernel? To
> >> make a long story short, NVIDIA binary driver rejects PAT in XEN and
> >> it falls back to using MTRR but MTRR in XEN was taken out a long time
> >> ago so now there's no way to get the NVIDIA binary blob running under
> >> a linux XEN dom0. I was about to tear my hair out looking for
> >> solutions high and low.
> >
> > Hi!
> >
> > Firstly, Konrad is probably the person you should send this to these
> > days, since I'm not managing to get much Xen stuff done.
> >
> > Secondly, hm.  Unfortunately, the changes we did have to integrate Xen's
> > MTRR machinery with Linux have been solidly rejected by the upstream
> > maintainers several times, so I think its unlikely that they will ever
> > make it into the mainline kernel.  And it doesn't seem to have really
> > made a difference because PAT does subsume MTRR for at least all the in
> > kernel users, as far as I know.
> >
> > What do you mean by "[the] NVIDIA binary driver rejects PAT in XEN and
> > it falls back to using MTRR"?  Why does the Nvidia driver reject PAT?
> > Perhaps addressing that would be a more profitable way of getting this
> > working.  In the past we've talked about changing Xen's PAT mapping to
> > match the kernel's (or make it configurable), but for now we're
> > remapping between the PAT schemes in the pte pvops.  If the NVIDIA
> > driver is using that mechanism to set ptes (as it must to get anywhere
> > in a pvops kernel), then it should be fine with the remapping.  Or its
> > possible they're having problems with reading a pte back and mapping
> > from Xen->Linux PAT formats, which is a problem some of the in-kernel
> > drivers also had.  Konrad, how did that turn out in the end?

Attic. I've turned it off since we had corruption issues (the WC didn't
turn back into WB b/c of page_attr using the pte_flag instead of pte_var).
Peter was talking about some software PAT lookup code but I hadn't
focused on that. There is also some performance numbers to run and collect.