Re: [RFC PATCH, x86]: Disable CPA cache flush for selfsnoop targets

[Uros Bizjak]

On Mon, Jul 15, 2019 at 8:41 PM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>
> On Mon, 15 Jul 2019, Andi Kleen wrote:
> > Uros Bizjak <ubizjak@xxxxxxxxx> writes:
> >
> > > Recent patch [1] disabled a self-snoop feature on a list of processor
> > > models with a known errata, so we are confident that the feature
> > > should work on remaining models also for other purposes than to speed
> > > up MTRR programming.
> >
> > MTRR is very different than TLBs.
> >
> > >From my understanding not flushing with PAT is only safe everywhere when
> > the memory is only used for coherent devices (like the Internal GPU on
> > Intel CPUs). We don't have any infrastructure to track or enforce
> > this unfortunately.
>
> Right, we don't know where the PAT invocation comes from and whether they
> are safe to omit flushing the cache. The module load code would be one
> obvious candidate.
>
> But unless there is some really worthwhile speedup, e.g. for boot, then
> adding some flag to let CPA know about the safe 'no flush' operation might
> be not worth it.

For the reference, FreeBSD implements this approach, later changed to
use pmap_invalidate_cache_range ifunc (that calls
pmap_invalidate_cache_range_selfsnoop for targets with self-snoop
capability) and pmap_force_invalidate_cache_range [1]. The full
cross-referenced source is at [2].

[1] https://reviews.freebsd.org/D16736
[2] http://fxr.watson.org/fxr/source/amd64/amd64/pmap.c

Uros.