Re: [PATCH v8 04/11] iommu: Add sva iommu_domain support

[Baolu Lu]

On 2022/6/17 15:43, Tian, Kevin wrote:
> > From: Baolu Lu
> > Sent: Friday, June 10, 2022 3:16 PM
> > > > +#define __IOMMU_DOMAIN_HOST_VA	(1U << 5)  /* Host CPU virtual
> > address */
> > > Do you mean general CPU VA? or Host CPU VA, I'm reading the latter as
> > 2nd
> > > stage?
> >
> > Host CPU VA. In the near future, we will add another flag _GUEST_VA, so
> > that the shared page table types are distiguished.
>
> How does the kernel knows that an user page table translates guest VA?
> IMHO I don't think the kernel should care about it except managing
> all the aspects related to the user page table itself...

Okay.

> > > > +
> > > >    /*
> > > >     * This are the possible domain-types
> > > >     *
> > > > @@ -86,15 +89,24 @@ struct iommu_domain_geometry {
> > > >    #define IOMMU_DOMAIN_DMA_FQ
> > 	(__IOMMU_DOMAIN_PAGING |	\
> > > >    				 __IOMMU_DOMAIN_DMA_API |	\
> > > >    				 __IOMMU_DOMAIN_DMA_FQ)
> > > > +#define IOMMU_DOMAIN_SVA	(__IOMMU_DOMAIN_SHARED |
> > 	\
> > > > +				 __IOMMU_DOMAIN_HOST_VA)
> > >
> > > Doesn't shared automatically mean CPU VA? Do we need another flag?
> >
> > Yes. Shared means CPU VA, but there're many types. Besides above two, we
> > also see the shared KVM/EPT.
>
> Will the two sharing scenarios share any common code? If not then
> having a separate flag bit is meaningless.

So far, I haven't seen the need for common code. I've ever thought about
the common notifier callback for page table entry update of SVA and KVM.
But there has been no feasible plan.

> It might be more straightforward to be:
>
> #define IOMMU_DOMAIN_SVA	__IOMMU_DOMAIN_SVA
> #define IOMMU_DOMAIN_KVM __IOMMU_DOMAIN_KVM
> #define IOMMU_DOMAIN_USER __IOMMU_DOMAIN_USER

I am okay with this and we can add some shared bits later if we need to
consolidate any code.

--
Best regards,
baolu