[PATCHv3 7/7] drm/msm/a6xx: Add support for using system cache(LLC)

[Jordan Crouse]

On Fri, Jul 03, 2020 at 09:04:49AM -0700, Rob Clark wrote:
> On Fri, Jul 3, 2020 at 7:53 AM Sai Prakash Ranjan
> <saiprakash.ranjan at codeaurora.org> wrote:
> >
> > Hi Will,
> >
> > On 2020-07-03 19:07, Will Deacon wrote:
> > > On Mon, Jun 29, 2020 at 09:22:50PM +0530, Sai Prakash Ranjan wrote:
> > >> diff --git a/drivers/gpu/drm/msm/msm_iommu.c
> > >> b/drivers/gpu/drm/msm/msm_iommu.c
> > >> index f455c597f76d..bd1d58229cc2 100644
> > >> --- a/drivers/gpu/drm/msm/msm_iommu.c
> > >> +++ b/drivers/gpu/drm/msm/msm_iommu.c
> > >> @@ -218,6 +218,9 @@ static int msm_iommu_map(struct msm_mmu *mmu,
> > >> uint64_t iova,
> > >>              iova |= GENMASK_ULL(63, 49);
> > >>
> > >>
> > >> +    if (mmu->features & MMU_FEATURE_USE_SYSTEM_CACHE)
> > >> +            prot |= IOMMU_SYS_CACHE_ONLY;
> > >
> > > Given that I think this is the only user of IOMMU_SYS_CACHE_ONLY, then
> > > it
> > > looks like it should actually be a property on the domain because we
> > > never
> > > need to configure it on a per-mapping basis within a domain, and
> > > therefore
> > > it shouldn't be exposed by the IOMMU API as a prot flag.
> > >
> > > Do you agree?
> > >
> >
> > GPU being the only user is for now, but there are other clients which
> > can use this.
> > Plus how do we set the memory attributes if we do not expose this as
> > prot flag?
> 
> It does appear that the downstream kgsl driver sets this for basically
> all mappings.. well there is some conditional stuff around
> DOMAIN_ATTR_USE_LLC_NWA but it seems based on the property of the
> domain.  (Jordan may know more about what that is about.)  But looks
> like there are a lot of different paths into iommu_map in kgsl so I
> might have missed something.

Downstream does set it universally. There are some theoretical use cases
where it might be beneficial to set it on a per-mapping basis with a bunch
of hinting from userspace and nobody has tried to characterize this on real
hardware so it is not clear to me if it is worth it.

I think a domain wide attribute works for now but if a compelling per-mapping
use case does comes down the pipeline we need to have a backup in mind -
possibly a prot flag to disable NWA?

Jordan

> Assuming there isn't some case where we specifically don't want to use
> the system cache for some mapping, I think it could be a domain
> attribute that sets an io_pgtable_cfg::quirks flag
> 
> BR,
> -R

-- 
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project