Re: [PATCH 1/6] KVM: Check for pending events before attempting injection

[Marcelo Tosatti <mtosatti@redhat.com>]

On Tue, Jul 27, 2010 at 04:19:35PM +0300, Avi Kivity wrote:
> Instead of blindly attempting to inject an event before each guest entry,
> check for a possible event first in vcpu->requests.  Sites that can trigger
> event injection are modified to set KVM_REQ_EVENT:
> 
> - interrupt, nmi window opening
> - ppr updates
> - i8259 output changes
> - local apic irr changes
> - rflags updates
> - gif flag set
> - event set on exit
> 
> This improves non-injecting entry performance, and sets the stage for
> non-atomic injection.
> 
> Signed-off-by: Avi Kivity <avi@redhat.com>
> ---
>  arch/x86/kvm/i8259.c     |    1 +
>  arch/x86/kvm/lapic.c     |   12 ++++++++++--
>  arch/x86/kvm/svm.c       |    8 +++++++-
>  arch/x86/kvm/vmx.c       |    6 ++++++
>  arch/x86/kvm/x86.c       |   35 ++++++++++++++++++++++++++---------
>  include/linux/kvm_host.h |    1 +
>  6 files changed, 51 insertions(+), 12 deletions(-)
> 

> @@ -4731,17 +4737,19 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
>  		goto out;
>  	}
>  
> -	inject_pending_event(vcpu);
> +	if (kvm_check_request(KVM_REQ_EVENT, vcpu)) {
> +		inject_pending_event(vcpu);
>  
> -	/* enable NMI/IRQ window open exits if needed */
> -	if (vcpu->arch.nmi_pending)
> -		kvm_x86_ops->enable_nmi_window(vcpu);
> -	else if (kvm_cpu_has_interrupt(vcpu) || req_int_win)
> -		kvm_x86_ops->enable_irq_window(vcpu);
> +		/* enable NMI/IRQ window open exits if needed */
> +		if (vcpu->arch.nmi_pending)
> +			kvm_x86_ops->enable_nmi_window(vcpu);
> +		else if (kvm_cpu_has_interrupt(vcpu) || req_int_win)
> +			kvm_x86_ops->enable_irq_window(vcpu);

Problem is it might not be possible to inject the event signalled by
KVM_REQ_EVENT, say an interrupt from an irqchip, if there is an event
that needs reinjection (or an exception).

Perhaps moving atomic_set(&vcpu->guest_mode, 1) up to preemptible
section is safe, because kvm_vcpu_kick can only IPI stale vcpu->cpu
while preemption is enabled. In that case, it will hit

if (!atomic_read(&vcpu->guest_mode)

later.

Although the KVM_REQ_EVENT idea is nice. Can you think of a way 
to fix the issue?