Re: [PATCH v3 0/5] KVM: optimize userspace exits with a new ioctl

[Paolo Bonzini <pbonzini@redhat.com>]

On 14/08/2015 12:08, Radim Krčmář wrote:
> v3:
>  * acked by Christian [1/5]
>  * use ioctl argument directly (unsigned long as flags) [4/5]
>  * precisely #ifdef arch-specific ioctls [5/5]
> v2:
>  * move request_exits debug counter patch right after introduction of
>    KVM_REQ_EXIT [3/5]
>  * use vcpu ioctl instead of vm one [4/5]
>  * shrink kvm_user_exit from 64 to 32 bytes [4/5]
>  * new [5/5]
> 
> QEMU uses SIGUSR1 to force a userspace exit and also to queue an early
> exit before calling VCPU_RUN -- the signal is blocked in user space and
> temporarily unblocked in VCPU_RUN.
> The temporal unblocking by sigprocmask() in kvm_arch_vcpu_ioctl_run()
> takes a shared siglock, which leads to cacheline bouncing in NUMA
> systems.
> 
> This series allows the same with a new request bit and VM IOCTL that
> marks and kicks target VCPU, hence no need to unblock.
> 
> inl_from_{pmtimer,qemu} vmexit benchmark from kvm-unit-tests shows ~5%
> speedup for 1-4 VCPUs (300-2000 saved cycles) without noticeably
> regressing kernel VM exits.
> (Paolo did a quick run of older version of this series on a NUMA system
>  and the speedup was around 35% when utilizing more nodes.)
> 
> Radim Krčmář (5):
>   KVM: add kvm_has_request wrapper
>   KVM: add KVM_REQ_EXIT request for userspace exit
>   KVM: x86: add request_exits debug counter
>   KVM: add KVM_USER_EXIT vcpu ioctl for userspace exit
>   KVM: refactor asynchronous vcpu ioctl dispatch
> 
>  Documentation/virtual/kvm/api.txt | 25 +++++++++++++++++++++++++
>  arch/x86/include/asm/kvm_host.h   |  1 +
>  arch/x86/kvm/vmx.c                |  4 ++--
>  arch/x86/kvm/x86.c                | 23 +++++++++++++++++++++++
>  include/linux/kvm_host.h          | 15 +++++++++++++--
>  include/uapi/linux/kvm.h          |  4 ++++
>  virt/kvm/kvm_main.c               | 15 ++++++++++-----
>  7 files changed, 78 insertions(+), 9 deletions(-)
> 

Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>

... however, we still need to decide what to do about machine-check
exceptions before enabling the capability, otherwise we'd need a new
KVM_CAP_USER_EXIT_MCE capability in the future.  So I'm holding up the
patches for now.

Paolo