On 6/21/19 10:38 AM, Marc Zyngier wrote:
> Starting a S2 MMU search from the beginning all the time means that
> we're potentially nuking a useful context (like we'd potentially
> have on a !VHE KVM guest).
>
> Instead, let's always start the search from the point *after* the
> last allocated context. This should ensure that alternating between
> two EL1 contexts will not result in nuking the whole S2 each time.
>
> lookup_s2_mmu now has a chance to provide a hit.
>
> Signed-off-by: Marc Zyngier <marc.zyngier@xxxxxxx>
> ---
> arch/arm64/include/asm/kvm_host.h | 1 +
> arch/arm64/kvm/nested.c | 14 ++++++++++++--
> 2 files changed, 13 insertions(+), 2 deletions(-)
>
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index b71a7a237f95..b7c44adcdbf3 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -123,6 +123,7 @@ struct kvm_arch {
> */
> struct kvm_s2_mmu *nested_mmus;
> size_t nested_mmus_size;
> + int nested_mmus_next;
For consistency, shouldn't nested_mmus_next be zero initialized in
kvm_init_nested (arch/arm64/kvm/nested.c), like nested_mmus and
nested_mmus_size? Not a big deal either way, since struct kvm is allocated using
vzalloc.
> really
> /* VTCR_EL2 value for this VM */
> u64 vtcr;
> diff --git a/arch/arm64/kvm/nested.c b/arch/arm64/kvm/nested.c
> index 09afafbdc8fe..214d59019935 100644
> --- a/arch/arm64/kvm/nested.c
> +++ b/arch/arm64/kvm/nested.c
> @@ -363,14 +363,24 @@ static struct kvm_s2_mmu *get_s2_mmu_nested(struct kvm_vcpu *vcpu)
> if (s2_mmu)
> goto out;
>
> - for (i = 0; i < kvm->arch.nested_mmus_size; i++) {
> - s2_mmu = &kvm->arch.nested_mmus[i];
> + /*
> + * Make sure we don't always search from the same point, or we
> + * will always reuse a potentially active context, leaving
> + * free contexts unused.
> + */
> + for (i = kvm->arch.nested_mmus_next;
> + i < (kvm->arch.nested_mmus_size + kvm->arch.nested_mmus_next);
> + i++) {
> + s2_mmu = &kvm->arch.nested_mmus[i % kvm->arch.nested_mmus_size];
>
> if (atomic_read(&s2_mmu->refcnt) == 0)
> break;
> }
> BUG_ON(atomic_read(&s2_mmu->refcnt)); /* We have struct MMUs to spare */
>
> + /* Set the scene for the next search */
> + kvm->arch.nested_mmus_next = (i + 1) % kvm->arch.nested_mmus_size;
> +
> if (kvm_s2_mmu_valid(s2_mmu)) {
> /* Clear the old state */
> kvm_unmap_stage2_range(s2_mmu, 0, kvm_phys_size(kvm));
