Re: [PATCH] KVM: nVMX: Skip IBPB when switching between vmcs01 and vmcs02

[Alexander Graf <graf@amazon.com>]

On 30.04.20 22:41, Sean Christopherson wrote:
> 
> Skip the Indirect Branch Prediction Barrier that is triggered on a VMCS
> switch when running with spectre_v2_user=on/auto if the switch is
> between two VMCSes in the same guest, i.e. between vmcs01 and vmcs02.
> The IBPB is intended to prevent one guest from attacking another, which
> is unnecessary in the nested case as it's the same guest from KVM's
> perspective.
> 
> This all but eliminates the overhead observed for nested VMX transitions
> when running with CONFIG_RETPOLINE=y and spectre_v2_user=on/auto, which
> can be significant, e.g. roughly 3x on current systems.
> 
> Reported-by: Alexander Graf <graf@amazon.com>
> Cc: KarimAllah Raslan <karahmed@amazon.de>
> Cc: stable@vger.kernel.org
> Fixes: 15d45071523d ("KVM/x86: Add IBPB support")
> Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>

I can confirm that with kvm-unit-test's vmcall benchmark, the patch does 
make a big difference:

   BEFORE: vmcall 33488
   AFTER:  vmcall 14898

So we're at least getting a good chunk of performance back :)

> ---
>   arch/x86/kvm/vmx/nested.c |  2 +-
>   arch/x86/kvm/vmx/vmx.c    | 12 ++++++++----
>   arch/x86/kvm/vmx/vmx.h    |  3 ++-
>   3 files changed, 11 insertions(+), 6 deletions(-)
> 
> diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
> index 2c36f3f53108..1a02bdfeeb2b 100644
> --- a/arch/x86/kvm/vmx/nested.c
> +++ b/arch/x86/kvm/vmx/nested.c
> @@ -303,7 +303,7 @@ static void vmx_switch_vmcs(struct kvm_vcpu *vcpu, struct loaded_vmcs *vmcs)
>          cpu = get_cpu();
>          prev = vmx->loaded_vmcs;
>          vmx->loaded_vmcs = vmcs;
> -       vmx_vcpu_load_vmcs(vcpu, cpu);
> +       vmx_vcpu_load_vmcs(vcpu, cpu, prev);
>          vmx_sync_vmcs_host_state(vmx, prev);
>          put_cpu();
> 
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index 3ab6ca6062ce..818dd8ba5e9f 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -1311,10 +1311,12 @@ static void vmx_vcpu_pi_load(struct kvm_vcpu *vcpu, int cpu)
>                  pi_set_on(pi_desc);
>   }
> 
> -void vmx_vcpu_load_vmcs(struct kvm_vcpu *vcpu, int cpu)
> +void vmx_vcpu_load_vmcs(struct kvm_vcpu *vcpu, int cpu,
> +                       struct loaded_vmcs *buddy)
>   {
>          struct vcpu_vmx *vmx = to_vmx(vcpu);
>          bool already_loaded = vmx->loaded_vmcs->cpu == cpu;
> +       struct vmcs *prev;
> 
>          if (!already_loaded) {
>                  loaded_vmcs_clear(vmx->loaded_vmcs);
> @@ -1333,10 +1335,12 @@ void vmx_vcpu_load_vmcs(struct kvm_vcpu *vcpu, int cpu)
>                  local_irq_enable();
>          }
> 
> -       if (per_cpu(current_vmcs, cpu) != vmx->loaded_vmcs->vmcs) {
> +       prev = per_cpu(current_vmcs, cpu);
> +       if (prev != vmx->loaded_vmcs->vmcs) {
>                  per_cpu(current_vmcs, cpu) = vmx->loaded_vmcs->vmcs;
>                  vmcs_load(vmx->loaded_vmcs->vmcs);
> -               indirect_branch_prediction_barrier();
> +               if (!buddy || buddy->vmcs != prev)
> +                       indirect_branch_prediction_barrier();

I fail to understand the logic here though. What exactly are you trying 
to catch? We only do the barrier when the current_vmcs as loaded by 
vmx_vcpu_load_vmcs is different from the vmcs of the context that was 
issuing the vmcs load.

Isn't this a really complicated way to say "Don't flush for nested"? Why 
not just make it explicit and pass in a bool that says "nested = true" 
from vmx_switch_vmcs()? Is there any case I'm missing where that would 
be unsafe?


Thanks,

Alex




Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss
Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
Sitz: Berlin
Ust-ID: DE 289 237 879