Re: [PATCH v2] KVM: x86: avoid redundant REQ_EVENT

From: Radim KrÄmÃÅ
Date: Thu Dec 15 2016 - 09:31:04 EST

2016-12-15 10:18+0300, Roman Kagan:
> On Wed, Dec 14, 2016 at 11:29:33PM +0100, Paolo Bonzini wrote:
>> On 14/12/2016 11:59, Denis Plotnikov wrote:
>> >
>> > if ((exit_reason == EXIT_REASON_EXTERNAL_INTERRUPT)
>> > && nested_exit_intr_ack_set(vcpu)) {
>> > - int irq = kvm_cpu_get_interrupt(vcpu);
>> > + int irq = kvm_cpu_get_interrupt(vcpu, true);
>> > WARN_ON(irq < 0);
>> I think this is not needed, because all nested vmexits end with a KVM_REQ_EVENT:

I also think that it can safely be false and we could drop the parameter
from kvm_cpu_get_interrupt().

(We have injected the highest priority interrupt and put it into ISR,
raising PPR again to its level, so there should be nothing to do in
KVM_REQ_EVENT due to any TPR changes.)

>> /*
>> * the KVM_REQ_EVENT optimization bit is only on for one entry, and if
>> * we did not inject a still-pending event to L1 now because of
>> * nested_run_pending, we need to re-enable this bit.
>> */
>> if (vmx->nested.nested_run_pending)
>> kvm_make_request(KVM_REQ_EVENT, vcpu);
> IIRC .nested_run_pending indicates we're emulating vmlaunch/vmresume and
> should not vmexit to L1, so this is not exactly "all nested vmexits"...
>> This would allow you to always pass false from kvm_cpu_get_interrupt to
>> kvm_get_apic_interrupt. Not sure if the additional complication in vmx.c
>> is worth the simplification in lapic.c. Radim, second opinion? :)

This patch goes for a minimal change in the non-nested case, so I would
leave nVMX optimizations for another patch.

One useless round of KVM_REQ_EVENT is not going change nested
performance by much and it is not the only thing we could improve wrt.
TPR ... I would just leave it for now and take care of it when we
* don't to update PPR at all with APICv -- it is already correct
* drop the KVM_REQ_EVENT with flex priority, because lower TPR cannot
unmask an interrupt