Re: [PATCHv3] iommu/arm-smmu: Optimize ->tlb_flush_walk() for qcom implementation

From: Will Deacon
Date: Tue Aug 10 2021 - 14:10:37 EST


On Tue, Aug 03, 2021 at 11:09:17AM +0530, Sai Prakash Ranjan wrote:
> On 2021-08-02 21:13, Will Deacon wrote:
> > On Wed, Jun 23, 2021 at 07:12:01PM +0530, Sai Prakash Ranjan wrote:
> > > diff --git a/drivers/iommu/arm/arm-smmu/arm-smmu.c
> > > b/drivers/iommu/arm/arm-smmu/arm-smmu.c
> > > index d3c6f54110a5..f3845e822565 100644
> > > --- a/drivers/iommu/arm/arm-smmu/arm-smmu.c
> > > +++ b/drivers/iommu/arm/arm-smmu/arm-smmu.c
> > > @@ -341,6 +341,12 @@ static void arm_smmu_tlb_add_page_s1(struct
> > > iommu_iotlb_gather *gather,
> > > ARM_SMMU_CB_S1_TLBIVAL);
> > > }
> > >
> > > +static void arm_smmu_tlb_inv_walk_impl_s1(unsigned long iova,
> > > size_t size,
> > > + size_t granule, void *cookie)
> > > +{
> > > + arm_smmu_tlb_inv_context_s1(cookie);
> > > +}
> > > +
> > > static void arm_smmu_tlb_inv_walk_s2(unsigned long iova, size_t size,
> > > size_t granule, void *cookie)
> > > {
> > > @@ -388,6 +394,12 @@ static const struct iommu_flush_ops
> > > arm_smmu_s1_tlb_ops = {
> > > .tlb_add_page = arm_smmu_tlb_add_page_s1,
> > > };
> > >
> > > +const struct iommu_flush_ops arm_smmu_s1_tlb_impl_ops = {
> > > + .tlb_flush_all = arm_smmu_tlb_inv_context_s1,
> > > + .tlb_flush_walk = arm_smmu_tlb_inv_walk_impl_s1,
> > > + .tlb_add_page = arm_smmu_tlb_add_page_s1,
> > > +};
> >
> > Hmm, dunno about this. Wouldn't it be a lot cleaner if the
> > tlb_flush_walk
> > callbacks just did the right thing based on the smmu_domain (maybe in
> > the
> > arm_smmu_cfg?) rather than having an entirely new set of ops just
> > because
> > they're const and you can't overide the bit you want?
> >
> > I don't think there's really an awful lot qcom-specific about the
> > principle
> > here -- there's a trade-off between over-invalidation and invalidation
> > latency. That happens on the CPU as well.
> >
>
> Sorry didn't understand, based on smmu_domain what? How do we make
> this implementation specific? Do you mean something like a quirk?
> The reason we didn't make this common was because nvidia folks weren't
> so happy with that, you can find the discussion in this thread [1].
>
> [1] https://lore.kernel.org/lkml/20210609145315.25750-1-saiprakash.ranjan@xxxxxxxxxxxxxx/

The ->tlb_flush_walk() callbacks take a 'void *cookie' which, for this
driver, is a 'struct arm_smmu_domain *'. From that, you can get to the
'struct arm_smmu_cfg' which could have something as coarse as:

bool flush_walk_prefer_tlbiasid;

which you can set when you initialise the domain (maybe in the
->init_context callback?). It shouldn't affect anybody else.

Will