Re: [PATCH v2 4/5] KVM: add KVM_USER_EXIT vcpu ioctl for userspace exit

[Avi Kivity <avi.kivity@gmail.com>]

On 08/17/2015 04:15 PM, Paolo Bonzini wrote:
>
> On 16/08/2015 13:27, Avi Kivity wrote:
>> On 08/05/2015 07:33 PM, Radim Krčmář wrote:
>>> The guest can use KVM_USER_EXIT instead of a signal-based exiting to
>>> userspace.  Availability depends on KVM_CAP_USER_EXIT.
>>> Only x86 is implemented so far.
>>>
>>> Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
>>> ---
>>>    v2:
>>>     * use vcpu ioctl instead of vm one [4/5]
>>>     * shrink kvm_user_exit from 64 to 32 bytes [4/5]
>>>
>>>    Documentation/virtual/kvm/api.txt | 30 ++++++++++++++++++++++++++++++
>>>    arch/x86/kvm/x86.c                | 24 ++++++++++++++++++++++++
>>>    include/uapi/linux/kvm.h          |  7 +++++++
>>>    virt/kvm/kvm_main.c               |  5 +++--
>>>    4 files changed, 64 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/Documentation/virtual/kvm/api.txt
>>> b/Documentation/virtual/kvm/api.txt
>>> index 3c714d43a717..c5844f0b8e7c 100644
>>> --- a/Documentation/virtual/kvm/api.txt
>>> +++ b/Documentation/virtual/kvm/api.txt
>>> @@ -3020,6 +3020,36 @@ Returns: 0 on success, -1 on error
>>>      Queues an SMI on the thread's vcpu.
>>>    +
>>> +4.97 KVM_USER_EXIT
>>> +
>>> +Capability: KVM_CAP_USER_EXIT
>>> +Architectures: x86
>>> +Type: vcpu ioctl
>>> +Parameters: struct kvm_user_exit (in)
>>> +Returns: 0 on success,
>>> +         -EFAULT if the parameter couldn't be read,
>>> +         -EINVAL if 'reserved' is not zeroed,
>>> +
>>> +struct kvm_user_exit {
>>> +    __u8 reserved[32];
>>> +};
>>> +
>>> +The ioctl is asynchronous to VCPU execution and can be issued from
>>> all threads.
>>> +format
>> This breaks an invariant of vcpu ioctls, and also forces a cacheline
>> bounce when we fget() the vcpu fd.
> KVM_USER_EXIT in practice should be so rare (at least with in-kernel
> LAPIC) that I don't think this matters.  KVM_USER_EXIT is relatively
> uninteresting, it only exists to provide an alternative to signals that
> doesn't require expensive atomics on each and every KVM_RUN. :(

Ah, so the idea is to remove the cost of changing the signal mask?

Yes, although it looks like a thread-local operation, it takes a 
process-wide lock.

I expect most user wakeups are via irqfd, so indeed the performance of 
KVM_USER_EXIT is uninteresting.