On Wed, Oct 7, 2020 at 9:00 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote: > > On Wed, Oct 07, 2020 at 08:10:34PM +0200, Daniel Vetter wrote: > > On Wed, Oct 7, 2020 at 7:36 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote: > > > > > > On Wed, Oct 07, 2020 at 06:44:24PM +0200, Daniel Vetter wrote: > > > > Way back it was a reasonable assumptions that iomem mappings never > > > > change the pfn range they point at. But this has changed: > > > > > > > > - gpu drivers dynamically manage their memory nowadays, invalidating > > > > ptes with unmap_mapping_range when buffers get moved > > > > > > > > - contiguous dma allocations have moved from dedicated carvetouts to > > > > cma regions. This means if we miss the unmap the pfn might contain > > > > pagecache or anon memory (well anything allocated with GFP_MOVEABLE) > > > > > > > > - even /dev/mem now invalidates mappings when the kernel requests that > > > > iomem region when CONFIG_IO_STRICT_DEVMEM is set, see 3234ac664a87 > > > > ("/dev/mem: Revoke mappings when a driver claims the region") > > > > > > > > Accessing pfns obtained from ptes without holding all the locks is > > > > therefore no longer a good idea. > > > > > > > > Unfortunately there's some users where this is not fixable (like v4l > > > > userptr of iomem mappings) or involves a pile of work (vfio type1 > > > > iommu). For now annotate these as unsafe and splat appropriately. > > > > > > > > This patch adds an unsafe_follow_pfn, which later patches will then > > > > roll out to all appropriate places. > > > > > > > > Signed-off-by: Daniel Vetter <daniel.vetter@xxxxxxxxx> > > > > Cc: Jason Gunthorpe <jgg@xxxxxxxx> > > > > Cc: Kees Cook <keescook@xxxxxxxxxxxx> > > > > Cc: Dan Williams <dan.j.williams@xxxxxxxxx> > > > > Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> > > > > Cc: John Hubbard <jhubbard@xxxxxxxxxx> > > > > Cc: Jérôme Glisse <jglisse@xxxxxxxxxx> > > > > Cc: Jan Kara <jack@xxxxxxx> > > > > Cc: Dan Williams <dan.j.williams@xxxxxxxxx> > > > > Cc: linux-mm@xxxxxxxxx > > > > Cc: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx > > > > Cc: linux-samsung-soc@xxxxxxxxxxxxxxx > > > > Cc: linux-media@xxxxxxxxxxxxxxx > > > > Cc: kvm@xxxxxxxxxxxxxxx > > > > include/linux/mm.h | 2 ++ > > > > mm/memory.c | 32 +++++++++++++++++++++++++++++++- > > > > mm/nommu.c | 17 +++++++++++++++++ > > > > security/Kconfig | 13 +++++++++++++ > > > > 4 files changed, 63 insertions(+), 1 deletion(-) > > > > > > Makes sense to me. > > > > > > I wonder if we could change the original follow_pfn to require the > > > ptep and then lockdep_assert_held() it against the page table lock? > > > > The safe variant with the pagetable lock is follow_pte_pmd. The only > > way to make follow_pfn safe is if you have an mmu notifier and > > corresponding retry logic. That is not covered by lockdep (it would > > splat if we annotate the retry side), so I'm not sure how you'd check > > for that? > > Right OK. > > > Checking for ptep lock doesn't work here, since the one leftover safe > > user of this (kvm) doesn't need that at all, because it has the mmu > > notifier. > > Ah, so a better name and/or function kdoc for follow_pfn is probably a > good iead in this patch as well. I did change that already to mention that you need an mmu notifier, and that follow_pte_pmd respectively unsafe_follow_pfn are the alternatives. Do you want more or something else here? Note that I left the kerneldoc for the nommu.c case unchanged, since without an mmu all bets are off anyway. > > So I think we're as good as it gets, since I really have no idea how > > to make sure follow_pfn callers do have an mmu notifier registered. > > Yah, can't be done. Most mmu notifier users should be using > hmm_range_fault anyhow, kvm is really very special here. We could pass an mmu notifier to follow_pfn and check that it has a registration for vma->vm_mm, but that feels like overkill when kvm is the only legit user for this. > > I've followed the few other CONFIG_STRICT_FOO I've seen, which are all > > explicit enables and default to "do not break uapi, damn the > > (security) bugs". Which is I think how this should be done. It is in > > the security section though, so hopefully competent distros will > > enable this all. > > I thought the strict ones were more general and less clear security > worries, not bugs like this. > > This is "allow a user triggerable use after free bug to exist in the > kernel" Since at most you get at GFP_MOVEABLE stuff I'm not sure you can use this to pull the kernel over the table. Maybe best way is if you get a gpu pagetable somehow into your pfn and then use that to access abitrary stuff, but there's still an iommu. I think leveraging this is going to be very tricky, and pretty much has to be device or driver specific somehow. -Daniel -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch