On Thu, Apr 24, 2025 at 10:40:49AM -0300, Jason Gunthorpe wrote: > On Wed, Apr 23, 2025 at 11:51:53PM -0700, Nicolin Chen wrote: > > On Wed, Apr 23, 2025 at 08:13:33PM -0300, Jason Gunthorpe wrote: > > > On Wed, Apr 23, 2025 at 11:31:29AM -0700, Nicolin Chen wrote: > > > > > > > > It also needs to act like a mdev and lock down the part of the IOAS > > > > > that provides that memory so the pin can't be released and UAF things. > > > > > > > > If I capture this correctly, the GPA->PA mapping is already done > > > > at the IOAS level for the S2 HWPT/domain, i.e. pages are already > > > > pinned. So we just need to a pair of for-driver APIs to validate > > > > the contiguity and refcount pages calling iopt_area_add_access(). > > > > > > Yes, adding an access is the key thing, the access will give you a > > > page list which you can validate, but it also provides a way to > > > synchronize if a hostile userspace does an unmap. > > > > The new APIs are very like iommufd_access_pin/unpin_pages(). But > > to reduce the amount of code that we have to share with driver.o, > > I added a smaller iopt_area_get/put_access() that gets an access > > and increases/decreases the refcounts only. > > Maybe the access should be obtained by the core code to avoid the > driver.o bloating? All the cmdq types need a memory buffer, right? > Perhaps that should just be generalized Yes. AMD just confirmed that they needed a similar setup. I think we can do that! > > Meanwhile, I am thinking if we could use the known S2 domain to > > translate the GPAs to PAs for the contiguity test, which feels a > > little cleaner to do in an IOMMU driver v.s. with a page list? > > You still need the access, and the access already generates a page > list.. Right. I was thinking it could save a few lines that fetches the page list, but then that would need another around of translation, which is unnecessary. Thanks Nicolin
