Re: [PATCH v12 7/8] KVM: arm64: ioctl to fetch/store tags in a guest

[Catalin Marinas]

On Thu, May 27, 2021 at 08:50:30AM +0100, Steven Price wrote:
> On 24/05/2021 19:11, Catalin Marinas wrote:
> > I had some (random) thoughts on how to make things simpler, maybe. I
> > think most of these races would have been solved if we required PROT_MTE
> > in the VMM but this has an impact on the VMM if it wants to use MTE
> > itself. If such requirement was in place, all KVM needed to do is check
> > PG_mte_tagged.
> > 
> > So what we actually need is a set_pte_at() in the VMM to clear the tags
> > and set PG_mte_tagged. Currently, we only do this if the memory type is
> > tagged (PROT_MTE) but it's not strictly necessary.
> > 
> > As an optimisation for normal programs, we don't want to do this all the
> > time but the visible behaviour wouldn't change (well, maybe for ptrace
> > slightly). However, it doesn't mean we couldn't for a VMM, with an
> > opt-in via prctl(). This would add a MMCF_MTE_TAG_INIT bit (couldn't
> > think of a better name) to mm_context_t.flags and set_pte_at() would
> > behave as if the pte was tagged without actually mapping the memory in
> > user space as tagged (protection flags not changed). Pages that don't
> > support tagging are still safe, just some unnecessary ignored tag
> > writes. This would need to be set before the mmap() for the guest
> > memory.
> > 
> > If we want finer-grained control we'd have to store this information in
> > the vma flags, in addition to VM_MTE (e.g. VM_MTE_TAG_INIT) but without
> > affecting the actual memory type. The easiest would be another pte bit,
> > though we are short on them. A more intrusive (not too bad) approach is
> > to introduce a set_pte_at_vma() and read the flags directly in the arch
> > code. In most places where set_pte_at() is called on a user mm, the vma
> > is also available.
> > 
> > Anyway, I'm not saying we go this route, just thinking out loud, get
> > some opinions.
> 
> Does get_user_pages() actually end up calling set_pte_at() normally?

Not always, at least as how it's called from hva_to_pfn(). My reading of
the get_user_page_fast_only() is that it doesn't touch the pte, just
walks the page tables and pins the page. Of course, it expects a valid
pte to have been set in the VMM already, otherwise it doesn't pin any
page and the caller falls back to the slow path.

The slow path, get_user_pages_unlocked(), passes FOLL_TOUCH and
set_pte_at() will be called either in follow_pfn_pte() if it was valid
or via faultin_page() -> handle_mm_fault().

> If not then on the normal user_mem_abort() route although we can
> easily check VM_MTE_TAG_INIT there's no obvious place to hook in to
> ensure that the pages actually allocated have the PG_mte_tagged flag.

I don't think it helps if we checked such vma flag in user_mem_abort(),
we'd still have the race with set_pte_at() on the page flags. What I was
trying to avoid is touching the page flags in too many places, so
deferring this always to set_pte_at() in the VMM.

> I'm also not sure how well this would work with the MMU notifiers path
> in KVM. With MMU notifiers (i.e. the VMM replacing a page in the
> memslot) there's not even an obvious hook to enforce the VMA flag. So I
> think we'd end up with something like the sanitise_mte_tags() function
> to at least check that the PG_mte_tagged flag is set on the pages
> (assuming that the trigger for the MMU notifier has done the
> corresponding set_pte_at()). Admittedly this might close the current
> race documented there.

If we kept this check to the VMM set_pte_at(), I think we can ignore the
notifiers.

> It also feels wrong to me to tie this to a process with prctl(), it
> seems much more normal to implement this as a new mprotect() flag as
> this is really a memory property not a process property. And I think
> we'll find some scary corner cases if we try to associate everything
> back to a process - although I can't instantly think of anything that
> will actually break.

I agree, tying it to the process looks wrong, only that it's less
intrusive. I don't think it would break anything, only potential
performance regression. A process would still need to pass PROT_MTE to
be able to get tag checking. That's basically what I had in an early MTE
implementation with clear_user_page() always zeroing the tags.

I agree with you that a vma flag would be better but it's more
complicated without an additional pte bit. We could also miss some
updates as mprotect() for example checks for pte_same() before calling
set_pte_at() (it would need to check the updated vma flags).

I'll review the latest series but I'm tempted to move the logic in
santise_mte_tags() to mte.c and take the big lock in there if
PG_mte_tagged is not already set. If we hit performance issues, we can
optimise this later to have the page flag set already on creation (new
PROT flag, prctl etc.).

-- 
Catalin