On Mon, Aug 19, 2019 at 02:23:48PM +0100, Mark Rutland wrote: > On Mon, Aug 19, 2019 at 01:56:26PM +0100, Will Deacon wrote: > > On Mon, Aug 19, 2019 at 07:44:20PM +0800, Walter Wu wrote: > > > __arm_v7s_unmap() call iopte_deref() to translate pyh_to_virt address, > > > but it will modify pointer tag into 0xff, so there is a false positive. > > > > > > When enable tag-based kasan, phys_to_virt() function need to rewrite > > > its original pointer tag in order to avoid kasan report an incorrect > > > memory corruption. > > > > Hmm. Which tree did you see this on? We've recently queued a load of fixes > > in this area, but I /thought/ they were only needed after the support for > > 52-bit virtual addressing in the kernel. > > I'm seeing similar issues in the virtio blk code (splat below), atop of > the arm64 for-next/core branch. I think this is a latent issue, and > people are only just starting to test with KASAN_SW_TAGS. > > It looks like the virtio blk code will round-trip a SLUB-allocated pointer > from > virt->page->virt, losing the per-object tag in the process. > > Our page_to_virt() seems to get a per-page tag, but this only makes > sense if you're dealing with the page allocator, rather than something > like SLUB which carves a page into smaller objects giving each object a > distinct tag. > > Any round-trip of a pointer from SLUB is going to lose the per-object > tag.
Urgh, I wonder how this is supposed to work? If we end up having to check the KASAN shadow for *_to_virt(), then why do we need to store anything in the page flags at all? Andrey? Will
