Hey Will,
On Wednesday 22 Jul 2020 at 14:15:10 (+0100), Will Deacon wrote:
> diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
> index 8c0035cab6b6..69dc36d1d486 100644
> --- a/arch/arm64/kvm/mmu.c
> +++ b/arch/arm64/kvm/mmu.c
> @@ -1326,7 +1326,7 @@ static bool stage2_get_leaf_entry(struct kvm *kvm, phys_addr_t addr,
> return true;
> }
>
> -static bool stage2_is_exec(struct kvm *kvm, phys_addr_t addr)
> +static bool stage2_is_exec(struct kvm *kvm, phys_addr_t addr, unsigned long sz)
> {
> pud_t *pudp;
> pmd_t *pmdp;
> @@ -1338,9 +1338,9 @@ static bool stage2_is_exec(struct kvm *kvm, phys_addr_t addr)
> return false;
>
> if (pudp)
> - return kvm_s2pud_exec(pudp);
> + return sz == PUD_SIZE && kvm_s2pud_exec(pudp);
> else if (pmdp)
> - return kvm_s2pmd_exec(pmdp);
> + return sz == PMD_SIZE && kvm_s2pmd_exec(pmdp);
> else
> return kvm_s2pte_exec(ptep);
This wants a 'sz == PAGE_SIZE' check, otherwise you'll happily inherit
the exec flag when a PTE has exec rights while you create a block
mapping on top.
Also, I think it should be safe to make the PMD and PUD case more
permissive, as 'sz <= PMD_SIZE' for instance, as the icache
invalidation shouldn't be an issue there? That probably doesn't matter
all that much though.
> }
> @@ -1958,7 +1958,8 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
> * execute permissions, and we preserve whatever we have.
> */
> needs_exec = exec_fault ||
> - (fault_status == FSC_PERM && stage2_is_exec(kvm, fault_ipa));
> + (fault_status == FSC_PERM &&
> + stage2_is_exec(kvm, fault_ipa, vma_pagesize));
>
> if (vma_pagesize == PUD_SIZE) {
> pud_t new_pud = kvm_pfn_pud(pfn, mem_type);
> --
> 2.28.0.rc0.105.gf9edc3c819-goog
>
FWIW, I reproduced the issue with a dummy guest accessing memory just
the wrong way, and toggling dirty logging at the right moment. And this
patch + my suggestion above seems to cure things. So, with the above
applied:
Reviewed-by: Quentin Perret <qperret@xxxxxxxxxx>
Tested-by: Quentin Perret <qperret@xxxxxxxxxx>
Thanks,
Quentin
