From: Wanpeng Li <[email protected]>
kvm_load_guest_fpu() and kvm_put_guest_fpu() each consume more than 14us
observed by ftrace, the qemu userspace FPU is swapped out for the guest
FPU context for the duration of the KVM_RUN ioctl even if sleeping AP,
we shouldn't load/put guest FPU context for this case especially for
serverless scenario which sensitives to boot time.
Signed-off-by: Wanpeng Li <[email protected]>
---
arch/x86/kvm/x86.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 5de2006..080ffa4 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -8680,7 +8680,6 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
vcpu_load(vcpu);
kvm_sigset_activate(vcpu);
- kvm_load_guest_fpu(vcpu);
if (unlikely(vcpu->arch.mp_state == KVM_MP_STATE_UNINITIALIZED)) {
if (kvm_run->immediate_exit) {
@@ -8718,12 +8717,14 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
}
}
+ kvm_load_guest_fpu(vcpu);
+
if (unlikely(vcpu->arch.complete_userspace_io)) {
int (*cui)(struct kvm_vcpu *) = vcpu->arch.complete_userspace_io;
vcpu->arch.complete_userspace_io = NULL;
r = cui(vcpu);
if (r <= 0)
- goto out;
+ goto out_fpu;
} else
WARN_ON(vcpu->arch.pio.count || vcpu->mmio_needed);
@@ -8732,8 +8733,9 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
else
r = vcpu_run(vcpu);
-out:
+out_fpu:
kvm_put_guest_fpu(vcpu);
+out:
if (vcpu->run->kvm_valid_regs)
store_regs(vcpu);
post_kvm_run_save(vcpu);
--
2.7.4
On Tue, Mar 10, 2020 at 03:01:19PM +0800, Wanpeng Li wrote:
> From: Wanpeng Li <[email protected]>
>
> kvm_load_guest_fpu() and kvm_put_guest_fpu() each consume more than 14us
> observed by ftrace, the qemu userspace FPU is swapped out for the guest
> FPU context for the duration of the KVM_RUN ioctl even if sleeping AP,
> we shouldn't load/put guest FPU context for this case especially for
> serverless scenario which sensitives to boot time.
>
> Signed-off-by: Wanpeng Li <[email protected]>
> ---
> arch/x86/kvm/x86.c | 8 +++++---
> 1 file changed, 5 insertions(+), 3 deletions(-)
>
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 5de2006..080ffa4 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -8680,7 +8680,6 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
>
> vcpu_load(vcpu);
> kvm_sigset_activate(vcpu);
> - kvm_load_guest_fpu(vcpu);
>
> if (unlikely(vcpu->arch.mp_state == KVM_MP_STATE_UNINITIALIZED)) {
> if (kvm_run->immediate_exit) {
> @@ -8718,12 +8717,14 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
> }
> }
>
> + kvm_load_guest_fpu(vcpu);
Ugh, so this isn't safe on MPX capable CPUs, kvm_apic_accept_events() can
trigger kvm_vcpu_reset() with @init_event=true and try to unload guest_fpu.
We could hack around that issue, but it'd be ugly, and I'm also concerned
that calling vmx_vcpu_reset() without guest_fpu loaded will be problematic
in the future with all the things that are getting managed by XSAVE.
> +
> if (unlikely(vcpu->arch.complete_userspace_io)) {
> int (*cui)(struct kvm_vcpu *) = vcpu->arch.complete_userspace_io;
> vcpu->arch.complete_userspace_io = NULL;
> r = cui(vcpu);
> if (r <= 0)
> - goto out;
> + goto out_fpu;
> } else
> WARN_ON(vcpu->arch.pio.count || vcpu->mmio_needed);
>
> @@ -8732,8 +8733,9 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
> else
> r = vcpu_run(vcpu);
>
> -out:
> +out_fpu:
> kvm_put_guest_fpu(vcpu);
> +out:
> if (vcpu->run->kvm_valid_regs)
> store_regs(vcpu);
> post_kvm_run_save(vcpu);
> --
> 2.7.4
>
On Wed, 11 Mar 2020 at 00:01, Sean Christopherson
<[email protected]> wrote:
>
> On Tue, Mar 10, 2020 at 03:01:19PM +0800, Wanpeng Li wrote:
> > From: Wanpeng Li <[email protected]>
> >
> > kvm_load_guest_fpu() and kvm_put_guest_fpu() each consume more than 14us
> > observed by ftrace, the qemu userspace FPU is swapped out for the guest
> > FPU context for the duration of the KVM_RUN ioctl even if sleeping AP,
> > we shouldn't load/put guest FPU context for this case especially for
> > serverless scenario which sensitives to boot time.
> >
> > Signed-off-by: Wanpeng Li <[email protected]>
> > ---
> > arch/x86/kvm/x86.c | 8 +++++---
> > 1 file changed, 5 insertions(+), 3 deletions(-)
> >
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index 5de2006..080ffa4 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -8680,7 +8680,6 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
> >
> > vcpu_load(vcpu);
> > kvm_sigset_activate(vcpu);
> > - kvm_load_guest_fpu(vcpu);
> >
> > if (unlikely(vcpu->arch.mp_state == KVM_MP_STATE_UNINITIALIZED)) {
> > if (kvm_run->immediate_exit) {
> > @@ -8718,12 +8717,14 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
> > }
> > }
> >
> > + kvm_load_guest_fpu(vcpu);
>
> Ugh, so this isn't safe on MPX capable CPUs, kvm_apic_accept_events() can
> trigger kvm_vcpu_reset() with @init_event=true and try to unload guest_fpu.
Right.
>
> We could hack around that issue, but it'd be ugly, and I'm also concerned
> that calling vmx_vcpu_reset() without guest_fpu loaded will be problematic
> in the future with all the things that are getting managed by XSAVE.
>
> > +
> > if (unlikely(vcpu->arch.complete_userspace_io)) {
> > int (*cui)(struct kvm_vcpu *) = vcpu->arch.complete_userspace_io;
> > vcpu->arch.complete_userspace_io = NULL;
> > r = cui(vcpu);
> > if (r <= 0)
> > - goto out;
> > + goto out_fpu;
> > } else
> > WARN_ON(vcpu->arch.pio.count || vcpu->mmio_needed);
> >
> > @@ -8732,8 +8733,9 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
> > else
> > r = vcpu_run(vcpu);
> >
> > -out:
> > +out_fpu:
> > kvm_put_guest_fpu(vcpu);
> > +out:
> > if (vcpu->run->kvm_valid_regs)
> > store_regs(vcpu);
> > post_kvm_run_save(vcpu);
> > --
> > 2.7.4
> >