[Freedreno] [PATCH 5/5] drm/msm: Skip tlbinv on unmap from non-current pgtables

[Rob Clark]

On Wed, Aug 24, 2022 at 10:46 AM Akhil P Oommen
<quic_akhilpo at quicinc.com> wrote:
>
> On 8/21/2022 11:49 PM, Rob Clark wrote:
> > From: Rob Clark <robdclark at chromium.org>
> >
> > We can rely on the tlbinv done by CP_SMMU_TABLE_UPDATE in this case.
> >
> > Signed-off-by: Rob Clark <robdclark at chromium.org>
> > ---
> >   drivers/gpu/drm/msm/adreno/a6xx_gpu.c |  6 ++++++
> >   drivers/gpu/drm/msm/msm_iommu.c       | 29 +++++++++++++++++++++++++++
> >   2 files changed, 35 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> > index c8ad8aeca777..1ba0ed629549 100644
> > --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> > +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> > @@ -1180,6 +1180,12 @@ static int hw_init(struct msm_gpu *gpu)
> >       /* Always come up on rb 0 */
> >       a6xx_gpu->cur_ring = gpu->rb[0];
> >
> > +     /*
> > +      * Note, we cannot assume anything about the state of the SMMU when
> > +      * coming back from power collapse, so force a CP_SMMU_TABLE_UPDATE
> > +      * on the first submit.  Also, msm_iommu_pagetable_unmap() relies on
> > +      * this behavior.
> > +      */
> >       gpu->cur_ctx_seqno = 0;
> >
> >       /* Enable the SQE_to start the CP engine */
> > diff --git a/drivers/gpu/drm/msm/msm_iommu.c b/drivers/gpu/drm/msm/msm_iommu.c
> > index 94c8c09980d1..218074a58081 100644
> > --- a/drivers/gpu/drm/msm/msm_iommu.c
> > +++ b/drivers/gpu/drm/msm/msm_iommu.c
> > @@ -45,8 +45,37 @@ static int msm_iommu_pagetable_unmap(struct msm_mmu *mmu, u64 iova,
> >               size -= 4096;
> >       }
> >
> > +     /*
> > +      * A CP_SMMU_TABLE_UPDATE is always sent for the first
> > +      * submit after resume, and that does a TLB invalidate.
> > +      * So we can skip that if the device is not currently
> > +      * powered.
> > +      */
> > +     if (!pm_runtime_get_if_in_use(pagetable->parent->dev))
> > +             goto out;
> > +
> > +     /*
> > +      * If we are not the current pgtables, we can rely on the
> > +      * TLB invalidate done by CP_SMMU_TABLE_UPDATE.
> > +      *
> > +      * We'll always be racing with the GPU updating ttbr0,
> > +      * but there are only two cases:
> > +      *
> > +      *  + either we are not the the current pgtables and there
> > +      *    will be a tlbinv done by the GPU before we are again
> > +      *
> > +      *  + or we are.. there might have already been a tblinv
> > +      *    if we raced with the GPU, but we have to assume the
> > +      *    worse and do the tlbinv
> > +      */
> > +     if (adreno_smmu->get_ttbr0(adreno_smmu->cookie) != pagetable->ttbr)
> > +             goto out_put;
> > +
> >       adreno_smmu->tlb_inv_by_id(adreno_smmu->cookie, pagetable->asid);
> >
> > +out_put:
> > +     pm_runtime_put(pagetable->parent->dev);
> > +out:
> >       return (unmapped == size) ? 0 : -EINVAL;
> >   }
> >
> Asking because it is a *security issue* if we get this wrong:
> 1. Is there any measure benefit with this patch? I believe tlb
> invalidation doesn't contribute much to the unmap latency.

It turned out to not make a huge difference.. although I expect the
part about skipping the inv when runtime suspended is still useful
from a power standpoint (but don't have a great setup to measure that)

BR,
-R

> 2. We at least should insert a full memory barrier before reading the
> ttbr0 register to ensure that everything we did prior to that is visible
> to smmu. But then I guess the cost of the full barrier would be similar
> to the tlb invalidation.
>
> Because it could lead to security issues or other very hard to debug
> issues, I would prefer this optimization only if there is a significant
> measurable gain.
>
> -Akhil.
>