RE: [PATCH v4 09/12] iommu/vt-d: Add iotlb flush for nested domain

[Tian, Kevin]

> From: Nicolin Chen <nicolinc@xxxxxxxxxx>
> Sent: Wednesday, August 9, 2023 1:42 AM
> 
> On Tue, Aug 08, 2023 at 09:34:03AM -0300, Jason Gunthorpe wrote:
> > On Mon, Aug 07, 2023 at 08:12:37PM -0700, Nicolin Chen wrote:
> > > On Mon, Aug 07, 2023 at 03:08:29PM +0000, Liu, Yi L wrote:
> > > > > > From: Liu, Yi L <yi.l.liu@xxxxxxxxx>
> > > > > > Sent: Monday, July 24, 2023 7:14 PM
> > > > > >
> > > > > > +static int intel_nested_cache_invalidate_user(struct
> iommu_domain
> > > > > > *domain,
> > > > > > +                                         void *user_data)
> > > > > > +{
> > > > > > +   struct iommu_hwpt_vtd_s1_invalidate_desc *req = user_data;
> > > > > > +   struct iommu_hwpt_vtd_s1_invalidate *inv_info = user_data;
> > > > > > +   struct dmar_domain *dmar_domain = to_dmar_domain(domain);
> > > > > > +   unsigned int entry_size = inv_info->entry_size;
> > > > > > +   u64 uptr = inv_info->inv_data_uptr;
> > > > > > +   u64 nr_uptr = inv_info->entry_nr_uptr;
> > > > > > +   struct device_domain_info *info;
> > > > > > +   u32 entry_nr, index;
> > > > > > +   unsigned long flags;
> > > > > > +   int ret = 0;
> > > > > > +
> > > > > > +   if (get_user(entry_nr, (uint32_t __user
> *)u64_to_user_ptr(nr_uptr)))
> > > > > > +           return -EFAULT;
> > > > > > +
> > > > > > +   for (index = 0; index < entry_nr; index++) {
> > > > > > +           ret = copy_struct_from_user(req, sizeof(*req),
> > > > > > +                                       u64_to_user_ptr(uptr + index *
> > > > > > entry_size),
> > > > > > +                                       entry_size);
> > > > >
> > > > > If continuing this direction then the driver should also check minsz etc.
> > > > > for struct iommu_hwpt_vtd_s1_invalidate and
> iommu_hwpt_vtd_s1_invalidate_desc
> > > > > since they are uAPI and subject to change.
> > > >
> > > > Then needs to define size in the uapi data structure, and copy size first
> and
> > > > check minsz before going forward. How about the structures for hwpt
> alloc
> > > > like struct iommu_hwpt_vtd_s1? Should check minsz for them as well?
> > >
> > > Assuming that every uAPI data structure needs a min_size, we can
> > > either add a structure holding all min_sizes like iommufd main.c
> > > or have another xx_min_len in iommu_/domain_ops.
> >
> > If driver is doing the copy it is OK that driver does the min_size
> > check too
> 
> Ah, just realized my reply above was missing a context..
> 
> Yi and I are having a concern that the core iommu_hpwt_alloc()
> and iommu_hwpt_cache_invalidate(), in the nesting series, copy
> data without checking the min_sizes. So, we are trying to add
> the missing piece into the next version but not sure which way
> could be optimal.
> 
> It probably makes sense to add cache_invalidate_user_min_len
> next to the existing cache_invalidate_user_data_len. For the
> iommu_hwpt_alloc, we are missing a data_len, as the core just
> uses sizeof(union iommu_domain_user_data) when calling the
> copy_struct_from_user().
> 
> Perhaps we could add two pairs of data_len/min_len in the ops
> structs:
> 	// iommu_ops
> 	const size_t domain_alloc_user_data_len; // for sanity&copy
> 	const size_t domain_alloc_user_min_len; // for sanity only
> 	// iommu_domain_ops
> 	const size_t cache_invalidate_user_data_len; // for sanity&copy
> 	const size_t cache_invalidate_user_min_len; // for sanity only
> 

What about creating a simple array to track type specific len in
iommufd instead of adding more fields to iommu/domain_ops?
anyway it's iommufd doing the copy and all the type specific
structures are already defined in the uapi header.

and a similar example already exists in union ucmd_buffer which
includes type specific structures to avoid memory copy...