Nicholas Piggin <npiggin@xxxxxxxxx> writes:
> As explained in the comment, there is no need to flush TLBs on all
> threads in a core when a vcpu moves between threads in the same core.
>
> Thread migrations can be a significant proportion of vcpu migrations,
> so this can help reduce the TLB flushing and IPI traffic.
>
> Signed-off-by: Nicholas Piggin <npiggin@xxxxxxxxx>
> ---
> I believe we can do this and have the TLB coherency correct as per
> the architecture, but would appreciate someone else verifying my
> thinking.
>
> Thanks,
> Nick
>
> arch/powerpc/kvm/book3s_hv.c | 28 ++++++++++++++++++++++++++--
> 1 file changed, 26 insertions(+), 2 deletions(-)
>
> diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c
> index 752daf43f780..53d0cbfe5933 100644
> --- a/arch/powerpc/kvm/book3s_hv.c
> +++ b/arch/powerpc/kvm/book3s_hv.c
> @@ -2584,7 +2584,7 @@ static void kvmppc_release_hwthread(int cpu)
> tpaca->kvm_hstate.kvm_split_mode = NULL;
> }
>
> -static void radix_flush_cpu(struct kvm *kvm, int cpu, struct kvm_vcpu *vcpu)
> +static void radix_flush_cpu(struct kvm *kvm, int cpu, bool core, struct kvm_vcpu *vcpu)
> {
Can we rename 'core' to something like 'core_sched' or 'within_core'
> struct kvm_nested_guest *nested = vcpu->arch.nested;
> cpumask_t *cpu_in_guest;
> @@ -2599,6 +2599,14 @@ static void radix_flush_cpu(struct kvm *kvm, int cpu, struct kvm_vcpu *vcpu)
> cpu_in_guest = &kvm->arch.cpu_in_guest;
> }
>
and do
if (core_sched) {
> + if (!core) {
> + cpumask_set_cpu(cpu, need_tlb_flush);
> + smp_mb();
> + if (cpumask_test_cpu(cpu, cpu_in_guest))
> + smp_call_function_single(cpu, do_nothing, NULL, 1);
> + return;
> + }
> +
> cpu = cpu_first_thread_sibling(cpu);
> for (i = 0; i < threads_per_core; ++i)
> cpumask_set_cpu(cpu + i, need_tlb_flush);
> @@ -2655,7 +2663,23 @@ static void kvmppc_prepare_radix_vcpu(struct kvm_vcpu *vcpu, int pcpu)
> if (prev_cpu < 0)
> return; /* first run */
>
> - radix_flush_cpu(kvm, prev_cpu, vcpu);
> + /*
> + * If changing cores, all threads on the old core should
> + * flush, because TLBs can be shared between threads. More
> + * precisely, the thread we previously ran on should be
> + * flushed, and the thread to first run a vcpu on the old
> + * core should flush, but we don't keep enough information
> + * around to track that, so we flush all.
> + *
> + * If changing threads in the same core, only the old thread
> + * need be flushed.
> + */
> + if (cpu_first_thread_sibling(prev_cpu) !=
> + cpu_first_thread_sibling(pcpu))
> + radix_flush_cpu(kvm, prev_cpu, true, vcpu);
> + else
> + radix_flush_cpu(kvm, prev_cpu, false, vcpu);
> +
> }
> }
>
> --
> 2.23.0
