On 21/02/18 22:59, Jordan Crouse wrote:
> +int iommu_sva_alloc_pasid(struct iommu_domain *domain, struct device *dev)
> +{
> +     int ret, pasid;
> +     struct io_pasid *io_pasid;
> +
> +     if (!domain->ops->pasid_alloc || !domain->ops->pasid_free)
> +             return -ENODEV;
> +
> +     io_pasid = kzalloc(sizeof(*io_pasid), GFP_KERNEL);
> +     if (!io_pasid)
> +             return -ENOMEM;
> +
> +     io_pasid->domain = domain;
> +     io_pasid->base.type = IO_TYPE_PASID;
> +
> +     idr_preload(GFP_KERNEL);
> +     spin_lock(&iommu_sva_lock);
> +     pasid = idr_alloc_cyclic(&iommu_pasid_idr, &io_pasid->base,
> +             1, (1 << 31), GFP_ATOMIC);

To be usable by other IOMMUs, this should restrict the PASID range to what
the IOMMU and the device support like io_mm_alloc(). In your case 31 bits,
but with PCI PASID it depends on the PASID capability and the SMMU
SubstreamID range.

For this reason I think device drivers should call iommu_sva_device_init()
once, even for the alloc_pasid() API. For SMMUv2 I guess it will be a NOP,
but other IOMMUs will allocate PASID tables and enable features in the
device. In addition, knowing that all users of the API call
iommu_sva_device_init()/shutdown() could allow us to allocate and enable
stuff lazily in the future.

It would also allow a given device driver to use both
iommu_sva_pasid_alloc() and iommu_sva_bind() at the same time. So that the
driver can assigns contexts to userspace and still use some of them for

> +int iommu_sva_map(int pasid, unsigned long iova,
> +           phys_addr_t paddr, size_t size, int prot)

It would be nice to factor iommu_map(), since this logic for map, map_sg
and unmap should be the same regardless of the PASID argument.

For example
- iommu_sva_map(domain, pasid, ...)
- iommu_map(domain, ...)

both call
- __iommu_map(domain, pasid, ...)

which calls either
- ops->map(domain, ...)
- ops->sva_map(domain, pasid, ...)

> @@ -347,6 +353,15 @@ struct iommu_ops {
>       int (*page_response)(struct iommu_domain *domain, struct device *dev,
>                            struct page_response_msg *msg);
> +     int (*pasid_alloc)(struct iommu_domain *domain, struct device *dev,
> +             int pasid);
> +     int (*sva_map)(struct iommu_domain *domain, int pasid,
> +                    unsigned long iova, phys_addr_t paddr, size_t size,
> +                    int prot);
> +     size_t (*sva_unmap)(struct iommu_domain *domain, int pasid,
> +                         unsigned long iova, size_t size);
> +     void (*pasid_free)(struct iommu_domain *domain, int pasid);
> +

Hmm, now IOMMU has the following ops:

* mm_alloc(): allocates a shared mm structure
* mm_attach(): writes the entry in the PASID table
* mm_detach(): removes the entry from the PASID table and invalidates it
* mm_free(): free shared mm
* pasid_alloc(): allocates a pasid structure (which I usually call
"private mm") and write the entry in the PASID table (or call
install_pasid() for SMMUv2)
* pasid_free(): remove from the PASID table (or call remove_pasid()) and
free the pasid structure.

Splitting mm_alloc and mm_attach is necessary because the io_mm in my case
can be shared between devices (allocated once, attached multiple times).
In your case a PASID is private to one device so only one callback is
needed. However mm_alloc+mm_attach will do roughly the same as
pasid_alloc, so to reduce redundancy in iommu_ops, maybe we could reuse
mm_alloc and mm_attach for the private PASID case?

iommu mailing list

Reply via email to