[PATCH 02/16] KVM: arm64: Don't issue CMOs when the physical address is invalid

[Quentin Perret]

On Thursday 15 Jul 2021 at 17:31:45 (+0100), Marc Zyngier wrote:
> Make sure we don't issue CMOs when mapping something that
> is not a memory address in the S2 page tables.
> 
> Signed-off-by: Marc Zyngier <maz at kernel.org>
> ---
>  arch/arm64/kvm/hyp/pgtable.c | 16 ++++++++++------
>  1 file changed, 10 insertions(+), 6 deletions(-)
> 
> diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c
> index 05321f4165e3..a5874ebd0354 100644
> --- a/arch/arm64/kvm/hyp/pgtable.c
> +++ b/arch/arm64/kvm/hyp/pgtable.c
> @@ -619,12 +619,16 @@ static int stage2_map_walker_try_leaf(u64 addr, u64 end, u32 level,
>  	}
>  
>  	/* Perform CMOs before installation of the guest stage-2 PTE */
> -	if (mm_ops->dcache_clean_inval_poc && stage2_pte_cacheable(pgt, new))
> -		mm_ops->dcache_clean_inval_poc(kvm_pte_follow(new, mm_ops),
> -						granule);
> -
> -	if (mm_ops->icache_inval_pou && stage2_pte_executable(new))
> -		mm_ops->icache_inval_pou(kvm_pte_follow(new, mm_ops), granule);
> +	if (kvm_phys_is_valid(phys)) {
> +		if (mm_ops->dcache_clean_inval_poc &&
> +		    stage2_pte_cacheable(pgt, new))
> +			mm_ops->dcache_clean_inval_poc(kvm_pte_follow(new,
> +								      mm_ops),
> +						       granule);
> +		if (mm_ops->icache_inval_pou && stage2_pte_executable(new))
> +			mm_ops->icache_inval_pou(kvm_pte_follow(new, mm_ops),
> +						 granule);
> +	}

Hrmpf so this makes me realize we have a problem here, not really caused
by your patch though.

Specifically, calling kvm_pgtable_stage2_set_owner() can lead to
overriding valid mappings with invalid mappings, which is effectively an
unmap operation. In this case we should issue CMOs when unmapping a
cacheable page to ensure it is clean to the PoC, like the
kvm_pgtable_stage2_unmap() does.

Note that you patch is already an improvement over the current state of
things, because calling stage2_pte_cacheable(pgt, new),
kvm_pte_follow(new, mm_ops) and friends is bogus when 'new' is invalid
...

Thanks,
Quentin