Re: [PATCH v4 2/2] arm64: kvm: Introduce MTE VCPU feature

[Andrew Jones]

On Wed, Nov 18, 2020 at 04:50:01PM +0000, Catalin Marinas wrote:
> On Wed, Nov 18, 2020 at 04:01:20PM +0000, Steven Price wrote:
> > On 17/11/2020 16:07, Catalin Marinas wrote:
> > > On Mon, Oct 26, 2020 at 03:57:27PM +0000, Steven Price wrote:
> > > > diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
> > > > index 19aacc7d64de..38fe25310ca1 100644
> > > > --- a/arch/arm64/kvm/mmu.c
> > > > +++ b/arch/arm64/kvm/mmu.c
> > > > @@ -862,6 +862,26 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, 
> > > > phys_addr_t fault_ipa,
> > > >         if (vma_pagesize == PAGE_SIZE && !force_pte)
> > > >                 vma_pagesize = transparent_hugepage_adjust(memslot, hva,
> > > >                                                            &pfn, 
> > > > &fault_ipa);
> > > > +
> > > > +       /*
> > > > +        * The otherwise redundant test for system_supports_mte() 
> > > > allows the
> > > > +        * code to be compiled out when CONFIG_ARM64_MTE is not present.
> > > > +        */
> > > > +       if (system_supports_mte() && kvm->arch.mte_enabled && 
> > > > pfn_valid(pfn)) {
> > > > +               /*
> > > > +                * VM will be able to see the page's tags, so we must 
> > > > ensure
> > > > +                * they have been initialised.
> > > > +                */
> > > > +               struct page *page = pfn_to_page(pfn);
> > > > +               long i, nr_pages = compound_nr(page);
> > > > +
> > > > +               /* if PG_mte_tagged is set, tags have already been 
> > > > initialised */
> > > > +               for (i = 0; i < nr_pages; i++, page++) {
> > > > +                       if (!test_and_set_bit(PG_mte_tagged, 
> > > > &page->flags))
> > > > +                               mte_clear_page_tags(page_address(page));
> > > > +               }
> > > > +       }
> > > 
> > > If this page was swapped out and mapped back in, where does the
> > > restoring from swap happen?
> > 
> > Restoring from swap happens above this in the call to gfn_to_pfn_prot()
> 
> Looking at the call chain, gfn_to_pfn_prot() ends up with
> get_user_pages() using the current->mm (the VMM) and that does a
> set_pte_at(), presumably restoring the tags. Does this mean that all
> memory mapped by the VMM in user space should have PROT_MTE set?
> Otherwise we don't take the mte_sync_tags() path in set_pte_at() and no
> tags restored from swap (we do save them since when they were mapped,
> PG_mte_tagged was set).
> 
> So I think the code above should be similar to mte_sync_tags(), even
> calling a common function, but I'm not sure where to get the swap pte
> from.
> 
> An alternative is to only enable HCR_EL2.ATA and MTE in guest if the vmm
> mapped the memory with PROT_MTE.

This is a very reasonable alternative. The VMM must be aware of whether
the guest may use MTE anyway. Asking it to map the memory with PROT_MTE
when it wants to offer the guest that option is a reasonable requirement.
If the memory is not mapped as such, then the host kernel shouldn't assume
MTE may be used by the guest, and it should even enforce that it is not
(by not enabling the feature).

Thanks,
drew

> 
> Yet another option is to always call mte_sync_tags() from set_pte_at()
> and defer the pte_tagged() or is_swap_pte() checks to the MTE code.
> 
> -- 
> Catalin
>