Re: [PATCH v7 1/3] iommufd: Add data structure for Intel VT-d stage-1 cache invalidation

[Nicolin Chen <nicolinc@nvidia.com>]

On Wed, Jan 03, 2024 at 08:02:04PM -0400, Jason Gunthorpe wrote:
> On Wed, Jan 03, 2024 at 12:18:35PM -0800, Nicolin Chen wrote:
> > > The driver would have to create it and there would be some driver
> > > specific enclosing struct to go with it
> > > 
> > > Perhaps device_ids goes in the driver specific struct, I don't know.
> > 
> > +struct iommufd_viommu {
> > +	struct iommufd_object obj;
> > +	struct iommufd_ctx *ictx;
> > +	struct iommu_device *iommu_dev;
> > +	struct iommufd_hwpt_paging *hwpt;	/* maybe unneeded */
> > +
> > +	int vmid;
> > +
> > +	union iommu_driver_user_data {
> > +		struct iommu_driver_user_vtd;
> > +		struct iommu_driver_user_arm_smmuv3;
> > +		struct iommu_driver_user_amd_viommu;
> > +	};
> 
> Not like that, in the usual container_of way

How about this:

// iommu.h
@@ -490,6 +490,16 @@ struct iommu_ops {
+       /* User space instance allocation and freeing by the iommu driver */
+       struct iommu_device_user *(*iommu_alloc_user)(struct iommu_device *iommu);
+       void (*iommu_free_user)(struct iommu_device_user *iommu_user);
+       int (*iommu_user_set_dev_id)(struct iommu_device_user *iommu_user,
+                                    struct device *dev, u64 dev_id);
+       int (*iommu_user_unset_dev_id)(struct iommu_device_user *iommu_user,
+                                      struct device *dev);
+       int (*iommu_cache_invalidate_usewr)(struct iommu_device_user *iommu_user,
+                                           struct iommu_user_data_array *array);
...
+/**
+ * struct iommu_device_user - IOMMU core representation of one IOMMU virtual
+ *                            instance
+ * @iommu_dev: Underlying IOMMU hardware instance
+ * @id: Virtual instance ID, e.g. a vmid
+ */
+struct iommu_device_user {
+       struct iommu_device *iommu_dev;
+       struct xarray virtual_ids;
+       u32 id;
+};

// iommufd_private.h
+struct iommufd_viommu {
+       struct iommufd_object obj;
+       struct iommufd_ctx *ictx;
+       struct iommu_device *iommu_dev;
+       struct iommu_device_user *iommu_user;
+       struct iommufd_hwpt_paging *hwpt;
+};
+
+int iommufd_viommu_alloc_ioctl(struct iommufd_ucmd *ucmd);
+void iommufd_viommu_destroy(struct iommufd_object *obj);

So iommu_alloc_user op allocates a driver private structure that
contains an iommu_device_user and returns &priv->iommu_user.

The set/unset_dev_id ops probably need a type argument if there
can be a multi-xarray case, then the virtual_ids xarray should
be moved to the driver private structure too?

Also, a 64-bit virtual id in the uAPI was suggested previously.
But xarray is limited to a 32-bit index? Should we compromise
the uAPI to 32-bit or is there an alternative for a 64-bit id
lookup?

Thanks
Nicolin