Re: [PATCH v8 04/25] iommu/arm-smmu-v3: Move IDR parsing to common functions

[Mostafa Saleh <smostafa@google.com>]

On Wed, Sep 23, 2026 at 08:52:57AM -0300, Jason Gunthorpe wrote:
> On Wed, Sep 23, 2026 at 10:09:06AM +0000, Mostafa Saleh wrote:
> > > > +	if (coherent && !disable_msipolling &&
> > > > +	    smmu->features & ARM_SMMU_FEAT_MSI)
> > > > +		smmu->options |= ARM_SMMU_OPT_MSIPOLL;
> > > 
> > > Will pKVM ever use MSIPOLL?
> > 
> > No, this version does not support MSI and hides it.
> > And this check can not be moved because disable_msipolling is a
> > module_param.
> 
> Could you move the above if into the shared area and have the callers
> mask off options if they don't want it?

Sure, I can move it and pass disable_polling as an arg which the
hypervisor will set to true.

> 
> > > > +	if (smmu->features & ARM_SMMU_FEAT_HYP &&
> > > > +	    cpus_have_cap(ARM64_HAS_VIRT_HOST_EXTN))
> > > > +		smmu->features |= ARM_SMMU_FEAT_E2H;
> > > 
> > > Why is ARM64_HAS_VIRT_HOST_EXTN left behind?
> > > 
> > 
> > cpus_have_cap() can not be used in the hypervisor.
> > Also, ARM_SMMU_FEAT_E2H is not exactly FEAT_HYP. As it defines the
> > world the translation lives in based on the kernel EL.
> 
> But don't you need this FEAT to issue invalidations correctly?
> pkvm never issues a S1 invalidation?

No, STRW is enforced to zero (NS-EL1) with pKVM as the kernel must be
running in EL1.

pKVM will issue S1 CMDQ_OP_TLBI_NH_ALL to invalidate all of S1
after S2 changes and invalidations.

> 
> > With pKVM at EL2 ARM64_HAS_VIRT_HOST_EXTN is always true anyway.
> 
> > And the hypervisor never owns a page table itself, so it never
> > checks this feature.
> 
> ? I thought the whole point of this was for pkvm to own a S2?

By own I mean "it's own translation regime" as if the hypervisor wants
to do DMA itself (synonym to EL2-S1 translation) in that case the
hypervisor will have a different STRW and use the E2H stuff. But that
never happens, the hypervsior populates S2 to enforce isolation of the
host but it never uses it itself.

> 
> Maybe same as I suggested for MSIPOLL, set the FEAT and mask it later
> if it can't be used. If KVM implicitly knows it has
> ARM64_HAS_VIRT_HOST_EXTN then it should operate with ARM_SMMU_FEAT_E2H
> set.
> 

As I replied to Nicolin, the problem here is not the FEAT being used or
not, it is that cpus_have_cap() can not be used in the hypervisor.

I think this can be replaced with cpus_have_final_cap() which should be
the same and then move this code also.

> > Otherwise, I think we can move this check and use cpus_have_final_cap()
> > instead as it can be used in the hypervisor.
> 
> Or that, I don't know the difference

I will try that for the next version.

Thanks,
Mostafa

> 
> Jason