Paolo Bonzini <pbonzini@xxxxxxxxxx> writes: > On 30/04/20 10:40, Vitaly Kuznetsov wrote: >>> I think in that case >>> kvm_check_async_pf_completion will refuse to make progress. >>> You need to make this bit stateful (e.g. 1 = async PF in progress, 0 = >>> not in progress), and check that for page ready notifications instead of >>> EFLAGS.IF. >>> This probably means that; >>> >>> - it might be simpler to move it to the vector MSR >> I didn't want to merge 'ACK' with the vector MSR as it forces the guest >> to remember the setting. It doesn't matter at all for Linux as we >> hardcode the interrupt number but I can imaging an OS assigning IRQ >> numbers dynamically, it'll need to keep record to avoid doing rdmsr. > > I would expect that it needs to keep it in a global variable anyway, but > yes this is a good point. You can also keep the ACK MSR and store the > pending bit in the other MSR, kind of like you have separate ISR and EOI > registers in the LAPIC. > Honestly I was inspired by Hyper-V's HV_X64_MSR_EOM MSR as the protocol we're trying to come up with here is very similar to HV messaging) I'm not exactly sure why we need the pending bit after we drop #PF. When we call kvm_check_async_pf_completion() from MSR_KVM_ASYNC_PF_ACK write it will (in case there are page ready events in the queue) check if the slot is empty, put one there and raise IRQ regardless of guest's current state. It may or may not get injected immediately but we don't care. The second invocation of kvm_check_async_pf_completion() from vcpu_run() will just go away. I'm probably just missing something, will think of it again while working on v1, it seems nobody is against the idea in general. Thanks! -- Vitaly
