Re: KVM: Questions and comments on make_all_cpus_request

[Marcelo Tosatti <mtosatti@redhat.com>]

On Tue, Jan 15, 2013 at 07:43:40PM -0500, Christoffer Dall wrote:
> Hi KVM guys,
> 
> I've had a bit of challenges figuring out the exact functinality and
> synchronization of vcpu->requests and friends.  In lack of a better
> method, I wrote some comments as a patch.
> 
> I think this code really deserves some explaining, as it is really hard
> to understand otherwise.  Unfortunately, I wasn't able to write down
> concise and exact comments, but I hope someone else feels up to the
> challenge.
> 
> Let me know if I just got this completely wrong and upside down.
> 
> Thanks,
> Christoffer
> ---
>  include/linux/kvm_host.h |    4 ++++
>  virt/kvm/kvm_main.c      |   29 +++++++++++++++++++++++++++--
>  2 files changed, 31 insertions(+), 2 deletions(-)
> 
> diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> index cbe0d68..25deef8 100644
> --- a/include/linux/kvm_host.h
> +++ b/include/linux/kvm_host.h
> @@ -252,6 +252,10 @@ struct kvm_vcpu {
>  	struct kvm_vcpu_arch arch;
>  };
>  
> +/*
> + * XXX: Could we explain what we're trying to achieve? Is this an
> + * optimization as to not send multiple IPIs?
> + */
>  static inline int kvm_vcpu_exiting_guest_mode(struct kvm_vcpu *vcpu)
>  {
>  	return cmpxchg(&vcpu->mode, IN_GUEST_MODE, EXITING_GUEST_MODE);
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index e45c20c..ccc292d 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -165,7 +165,18 @@ static void ack_flush(void *_completed)
>  {
>  }
>  
> -static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
> +/**
> + * make_all_cpus_request - place request on vcpus
> + * @kvm: KVM Struct
> + * @req: Request to make on the VCPU
> + *
> + * This function places a request on a VCPU and ensures that the VCPU request
> + * is handled before returning from the function, if the VCPU is in guest
> + * mode (or exiting, or reading shadow page tables?).
> + *
> + * Returns true if at least of the vcpus were sent an IPI and responded to it,
> + */
> +static bool make_all_vcpus_request(struct kvm *kvm, unsigned int req)
>  {
>  	int i, cpu, me;
>  	cpumask_var_t cpus;
> @@ -179,9 +190,19 @@ static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
>  		kvm_make_request(req, vcpu);
>  		cpu = vcpu->cpu;
>  
> +		/*
> +		 * Is the following really true? Can we have an example of the
> +		 * race that would otherwise happen? Doesn't the make_request
> +		 * pair against the IPI and interrupt disabling, not the mode?
> +		 */
>  		/* Set ->requests bit before we read ->mode */
>  		smp_mb();

cpu0						cpu1
					
make_all_reqs
					
mode = vcpu->mode = OUTSIDE_GUEST_MODE
						vcpu->mode = IN_GUEST_MODE
						if (vcpu->requests)
set_bit(KVM_REQ_y, vcpu)
						
no kick, its outside guest mode