Re: [PATCH 17/28] KVM: x86/mmu: Terminate yield-friendly walk if invalid root observed

[Ben Gardon <bgardon@google.com>]

On Fri, Nov 19, 2021 at 8:51 PM Sean Christopherson <seanjc@google.com> wrote:
>
> Stop walking TDP MMU roots if the previous root was invalidated while the
> caller yielded (dropped mmu_lock).  Any effect that the caller wishes to
> be recognized by a root must be made visible before walking the list of
> roots.  Because all roots are invalided when any root is invalidated, if
> the previous root was invalidated, then all roots on the list when the
> caller first started the walk also were invalidated, and any valid roots
> on the list must have been added after the invalidation event.
>
> Signed-off-by: Sean Christopherson <seanjc@google.com>
> ---
>  arch/x86/kvm/mmu/tdp_mmu.c | 19 ++++++++++++-------
>  1 file changed, 12 insertions(+), 7 deletions(-)
>
> diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
> index d9524b387221..cc8d021a1ba5 100644
> --- a/arch/x86/kvm/mmu/tdp_mmu.c
> +++ b/arch/x86/kvm/mmu/tdp_mmu.c
> @@ -140,16 +140,21 @@ static struct kvm_mmu_page *tdp_mmu_next_root(struct kvm *kvm,
>         lockdep_assert_held(&kvm->mmu_lock);
>
>         /*
> -        * Restart the walk if the previous root was invalidated, which can
> -        * happen if the caller drops mmu_lock when yielding.  Restarting the
> -        * walke is necessary because invalidating a root also removes it from
> -        * tdp_mmu_roots.  Restarting is safe and correct because invalidating
> -        * a root is done if and only if _all_ roots are invalidated, i.e. any
> -        * root on tdp_mmu_roots was added _after_ the invalidation event.
> +        * Terminate the walk if the previous root was invalidated, which can
> +        * happen if the caller yielded and dropped mmu_lock.  Because invalid
> +        * roots are removed from tdp_mmu_roots with mmu_lock held for write,
> +        * if the previous root was invalidated, then the invalidation occurred
> +        * after this walk started.  And because _all_ roots are invalidated
> +        * during an invalidation event, any root on tdp_mmu_roots was created
> +        * after the invalidation.  Lastly, any state change being made by the
> +        * caller must be effected before updating SPTEs, otherwise vCPUs could
> +        * simply create new SPTEs with the old state.  Thus, if the previous
> +        * root was invalidated, all valid roots are guaranteed to have been
> +        * created after the desired state change and don't need to be updated.
>          */
>         if (prev_root && prev_root->role.invalid) {
>                 kvm_tdp_mmu_put_root(kvm, prev_root, shared);
> -               prev_root = NULL;
> +               return NULL;
>         }

Another option might be to walk the list FIFO so that the newly
created roots are encountered, but since they start empty, there might
not be any work to do.

>
>         /*
> --
> 2.34.0.rc2.393.gf8c9666880-goog
>