Re: [PATCH v7 03/10] iommu/sva: Add iommu_sva_domain support

[Baolu Lu]

Hi Robin,

On 2022/5/24 22:36, Robin Murphy wrote:
> On 2022-05-19 08:20, Lu Baolu wrote:
> [...]
> > diff --git a/drivers/iommu/iommu-sva-lib.c 
> > b/drivers/iommu/iommu-sva-lib.c
> > index 106506143896..210c376f6043 100644
> > --- a/drivers/iommu/iommu-sva-lib.c
> > +++ b/drivers/iommu/iommu-sva-lib.c
> > @@ -69,3 +69,51 @@ struct mm_struct *iommu_sva_find(ioasid_t pasid)
> >       return ioasid_find(&iommu_sva_pasid, pasid, __mmget_not_zero);
> >   }
> >   EXPORT_SYMBOL_GPL(iommu_sva_find);
> > +
> > +/*
> > + * IOMMU SVA driver-oriented interfaces
> > + */
> > +struct iommu_domain *
> > +iommu_sva_alloc_domain(struct bus_type *bus, struct mm_struct *mm)
>
> Argh, please no new bus-based external interfaces! Domain allocation 
> needs to resolve to the right IOMMU instance to solve a number of 
> issues, and cleaning up existing users of iommu_domain_alloc() to 
> prepare for that is already hard enough. This is arguably even more 
> relevant here than for other domain types, since SVA support is more 
> likely to depend on specific features that can vary between IOMMU 
> instances even with the same driver. Please make the external interface 
> take a struct device, then resolve the ops through dev->iommu.
>
> Further nit: the naming inconsistency bugs me a bit - 
> iommu_sva_domain_alloc() seems more natural. Also I'd question the 
> symmetry vs. usability dichotomy of whether we *really* want two 
> different free functions for a struct iommu_domain pointer, where any 
> caller which might mix SVA and non-SVA usage then has to remember how 
> they allocated any particular domain :/
>
> > +{
> > +    struct iommu_sva_domain *sva_domain;
> > +    struct iommu_domain *domain;
> > +
> > +    if (!bus->iommu_ops || !bus->iommu_ops->sva_domain_ops)
> > +        return ERR_PTR(-ENODEV);
> > +
> > +    sva_domain = kzalloc(sizeof(*sva_domain), GFP_KERNEL);
> > +    if (!sva_domain)
> > +        return ERR_PTR(-ENOMEM);
> > +
> > +    mmgrab(mm);
> > +    sva_domain->mm = mm;
> > +
> > +    domain = &sva_domain->domain;
> > +    domain->type = IOMMU_DOMAIN_SVA;
> > +    domain->ops = bus->iommu_ops->sva_domain_ops;
>
> I'd have thought it would be logical to pass IOMMU_DOMAIN_SVA to the 
> normal domain_alloc call, so that driver-internal stuff like context 
> descriptors can be still be hung off the domain as usual (rather than 
> all drivers having to implement some extra internal lookup mechanism to 
> handle all the SVA domain ops), but that's something we're free to come 

Agreed with above comments. Thanks! I will post an additional patch
for review later.

> back and change later. FWIW I'd just stick the mm pointer in struct 
> iommu_domain, in a union with the fault handler stuff and/or iova_cookie 
> - those are mutually exclusive with SVA, right?

"iova_cookie" is mutually exclusive with SVA, but I am not sure about
the fault handler stuff.

Did you mean @handler and @handler_token staffs below?

struct iommu_domain {
        unsigned type;
        const struct iommu_domain_ops *ops;
        unsigned long pgsize_bitmap;    /* Bitmap of page sizes in use */
        iommu_fault_handler_t handler;
        void *handler_token;
        struct iommu_domain_geometry geometry;
        struct iommu_dma_cookie *iova_cookie;
};

Is it only for DMA domains? From the point view of IOMMU faults, it
seems to be generic.

Best regards,
baolu