On Tue, Feb 04, 2025, James Houghton wrote:
> spte_has_volatile_bits() is now a misnomer, as the an SPTE can have its
> Accessed bit set or cleared without the mmu_lock held, but the state of
> the Accessed bit is not checked in spte_has_volatile_bits().
> Even if a caller uses spte_needs_atomic_write(), Accessed bit
> information may still be lost, but that is already tolerated, as the TLB
> is not invalidated after the Accessed bit is cleared.
>
> Signed-off-by: James Houghton <jthoughton@xxxxxxxxxx>
> ---
...
> diff --git a/arch/x86/kvm/mmu/spte.h b/arch/x86/kvm/mmu/spte.h
> index 59746854c0af..4c290ae9a02a 100644
> --- a/arch/x86/kvm/mmu/spte.h
> +++ b/arch/x86/kvm/mmu/spte.h
> @@ -519,7 +519,7 @@ static inline u64 get_mmio_spte_generation(u64 spte)
> return gen;
> }
>
> -bool spte_has_volatile_bits(u64 spte);
> +bool spte_needs_atomic_write(u64 spte);
>
> bool make_spte(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp,
> const struct kvm_memory_slot *slot,
> diff --git a/arch/x86/kvm/mmu/tdp_iter.h b/arch/x86/kvm/mmu/tdp_iter.h
> index 05e9d678aac9..b54123163efc 100644
> --- a/arch/x86/kvm/mmu/tdp_iter.h
> +++ b/arch/x86/kvm/mmu/tdp_iter.h
> @@ -55,7 +55,7 @@ static inline bool kvm_tdp_mmu_spte_need_atomic_write(u64 old_spte, int level)
> {
> return is_shadow_present_pte(old_spte) &&
> is_last_spte(old_spte, level) &&
> - spte_has_volatile_bits(old_spte);
> + spte_needs_atomic_write(old_spte);
Unless you object, I'll change this to spte_needs_atomic_update(), and tweak
kvm_tdp_mmu_spte_need_atomic_write() accordingly. "write" was a bad choice by
me. It's not just the store/write that needs to be atomic, it's the entire
read-modify-write. E.g. KVM needs to preserve the existing value, but for many
flows, it's even more important that KVM's snapshot of the old SPTE is accurate.
> }
>
> static inline u64 kvm_tdp_mmu_write_spte(tdp_ptep_t sptep, u64 old_spte,
> --
> 2.48.1.362.g079036d154-goog
>
