Re: [Xen-devel] RIP MTRR - status update for upcoming v4.2

[Toshi Kani <toshi.kani@hp.com>]

On Fri, 2015-08-07 at 17:08 -0600, Toshi Kani wrote:
> On Fri, 2015-08-07 at 15:23 -0700, Luis R. Rodriguez wrote:
> > On Fri, Aug 7, 2015 at 2:56 PM, Toshi Kani <toshi.kani@hp.com> wrote:
> > > On Fri, 2015-08-07 at 13:25 -0700, Luis R. Rodriguez wrote:
> > > > On Thu, Aug 6, 2015 at 3:58 PM, Toshi Kani <toshi.kani@hp.com> 
> > > > wrote:
> > > > > On Thu, 2015-08-06 at 12:53 -0700, Luis R. Rodriguez wrote:
> > > > > > On Fri, Jun 12, 2015 at 9:58 AM, Toshi Kani <toshi.kani@hp.com> 
> > > > > > wrote:
>  :
> > > > > 
> > > > > No, there is no OS support necessary to use MTRR.  After firmware 
> > > > > sets it up, CPUs continue to use it without any OS support.  I 
> > > > > think the Linux change you are referring is to obsolete legacy
> > > > > interfaces that modify the MTRR setup.  I agree that Linux should 
> > > > > not modify MTRR.
> > > > 
> > > > Its a bit more than that though. Since you agree that the OS can 
> > > > live without MTRR code I was hoping to then see if we can fold out 
> > > > PAT Linux code from under the MTRR dependency on Linux and make PAT 
> > > > a first class citizen, maybe at least for x86-64. Right now you can
> > > > only get PAT support on Linux if you have MTRR code, but I'd like to 
> > > > see if instead we can rip MTRR code out completely under its own 
> > > > Kconfig and let it start rotting away.
> > > > 
> > > > Code-wise the only issue I saw was that PAT code also relies on
> > > > mtrr_type_lookup(), see pat_x_mtrr_type(), but other than this I 
> > > > found no other obvious issues.
> > > 
> > > We can rip of the MTTR code that modifies the MTRR setup, but not
> > > mtrr_type_lookup().  This function provides necessary checks per 
> > > documented in commit 7f0431e3dc89 as follows.
> > > 
> > >     1) reserve_memtype() tracks an effective memory type in case
> > >        a request type is WB (ex. /dev/mem blindly uses WB). Missing
> > >        to track with its effective type causes a subsequent request
> > >        to map the same range with the effective type to fail.
> > > 
> > >     2) pud_set_huge() and pmd_set_huge() check if a requested range
> > >        has any overlap with MTRRs. Missing to detect an overlap may
> > >        cause a performance penalty or undefined behavior.
> > > 
> > > mtrr_type_lookup() is still admittedly awkward, but I do not think we 
> > > have an immediate issue in PAT code calling it.  I do not think it 
> > > makes 
> > > PAT code a second class citizen.
> > 
> > OK since we know that if MTRR set up code ends up disabled and would
> > return MTRR_TYPE_INVALID what if we just static inline this for the
> > no-MTRR Kconfig build option immediately, and only then have the full
> > blown implementation for the case where MTRR Kconfig option is
> > enabled?
> 
> Yes, the MTRR code could be disabled by Kconfig with such inline stubs as
> long as the kernel is built specifically for a particular platform with 
> MTRR disabled, such as Xen guest kernel.

Noticed that we do have CONFIG_MTRR and mtrr_type_lookup() inline stub
returns MTRR_INVALID.

-Toshi