On 09/03/2017 10:40, Wanpeng Li wrote:
> 2017-03-09 9:23 GMT+08:00 Wanpeng Li <kernellwp@xxxxxxxxx>:
>> 2016-12-20 0:17 GMT+08:00 Paolo Bonzini <pbonzini@xxxxxxxxxx>:
>>> Since bf9f6ac8d749 ("KVM: Update Posted-Interrupts Descriptor when vCPU
>>> is blocked", 2015-09-18) the posted interrupt descriptor is checked
>>> unconditionally for PIR.ON. Therefore we don't need KVM_REQ_EVENT to
>>> trigger the scan and, if NMIs or SMIs are not involved, we can avoid
>>> the complicated event injection path.
>>>
>>> Calling kvm_vcpu_kick if PIR.ON=1 is also useless, though it has been
>>> there since APICv was introduced.
>>>
>>> However, without the KVM_REQ_EVENT safety net KVM needs to be much
>>> more careful about races between vmx_deliver_posted_interrupt and
>>> vcpu_enter_guest. First, the IPI for posted interrupts may be issued
>>> between setting vcpu->mode = IN_GUEST_MODE and disabling interrupts.
>>> If that happens, kvm_trigger_posted_interrupt returns true, but
>>> smp_kvm_posted_intr_ipi doesn't do anything about it. The guest is
>>> entered with PIR.ON, but the posted interrupt IPI has not been sent
>>> and the interrupt is only delivered to the guest on the next vmentry
>>> (if any). To fix this, disable interrupts before setting vcpu->mode.
>>> This ensures that the IPI is delayed until the guest enters non-root mode;
>>> it is then trapped by the processor causing the interrupt to be injected.
>>>
>>> Second, the IPI may be issued between
>>>
>>> kvm_x86_ops->hwapic_irr_update(vcpu,
>>> kvm_lapic_find_highest_irr(vcpu));
>>>
>>> and vcpu->mode = IN_GUEST_MODE. In this case, kvm_vcpu_kick is called
>>> but it (correctly) doesn't do anything because it sees vcpu->mode ==
>>> OUTSIDE_GUEST_MODE. Again, the guest is entered with PIR.ON but no
>>> posted interrupt IPI is pending; this time, the fix for this is to move
>>> the RVI update after IN_GUEST_MODE.
>>>
>>> Both issues were previously masked by the liberal usage of KVM_REQ_EVENT.
>>> In both race scenarios KVM_REQ_EVENT would cancel guest entry, resulting
>>> in another vmentry which would inject the interrupt.
>>>
>>> This saves about 300 cycles on the self_ipi_* tests of vmexit.flat.
>>>
>>> Signed-off-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>
>>> ---
>>> arch/x86/kvm/lapic.c | 11 ++++-------
>>> arch/x86/kvm/vmx.c | 8 +++++---
>>> arch/x86/kvm/x86.c | 44 +++++++++++++++++++++++++-------------------
>>> 3 files changed, 34 insertions(+), 29 deletions(-)
>>>
>>> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
>>> index f644dd1dbe71..5ea94b622e88 100644
>>> --- a/arch/x86/kvm/lapic.c
>>> +++ b/arch/x86/kvm/lapic.c
>>> @@ -385,12 +385,8 @@ int __kvm_apic_update_irr(u32 *pir, void *regs)
>>> int kvm_apic_update_irr(struct kvm_vcpu *vcpu, u32 *pir)
>>> {
>>> struct kvm_lapic *apic = vcpu->arch.apic;
>>> - int max_irr;
>>>
>>> - max_irr = __kvm_apic_update_irr(pir, apic->regs);
>>> -
>>> - kvm_make_request(KVM_REQ_EVENT, vcpu);
>>> - return max_irr;
>>> + return __kvm_apic_update_irr(pir, apic->regs);
>>> }
>>> EXPORT_SYMBOL_GPL(kvm_apic_update_irr);
>>>
>>> @@ -423,9 +419,10 @@ static inline void apic_clear_irr(int vec, struct kvm_lapic *apic)
>>> vcpu = apic->vcpu;
>>>
>>> if (unlikely(vcpu->arch.apicv_active)) {
>>> - /* try to update RVI */
>>> + /* need to update RVI */
>>> apic_clear_vector(vec, apic->regs + APIC_IRR);
>>> - kvm_make_request(KVM_REQ_EVENT, vcpu);
>>> + kvm_x86_ops->hwapic_irr_update(vcpu,
>>> + apic_find_highest_irr(apic));
>>> } else {
>>> apic->irr_pending = false;
>>> apic_clear_vector(vec, apic->regs + APIC_IRR);
>>> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
>>> index 27e40b180242..3dd4fad35a3e 100644
>>> --- a/arch/x86/kvm/vmx.c
>>> +++ b/arch/x86/kvm/vmx.c
>>> @@ -5062,9 +5062,11 @@ static void vmx_deliver_posted_interrupt(struct kvm_vcpu *vcpu, int vector)
>>> if (pi_test_and_set_pir(vector, &vmx->pi_desc))
>>> return;
>>>
>>> - r = pi_test_and_set_on(&vmx->pi_desc);
>>> - kvm_make_request(KVM_REQ_EVENT, vcpu);
>>> - if (r || !kvm_vcpu_trigger_posted_interrupt(vcpu))
>>> + /* If a previous notification has sent the IPI, nothing to do. */
>>> + if (pi_test_and_set_on(&vmx->pi_desc))
>>> + return;
>>> +
>>> + if (!kvm_vcpu_trigger_posted_interrupt(vcpu))
>>> kvm_vcpu_kick(vcpu);
>>> }
>>>
>>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>>> index c666414adc1d..725473ba6dd3 100644
>>> --- a/arch/x86/kvm/x86.c
>>> +++ b/arch/x86/kvm/x86.c
>>> @@ -6710,19 +6710,6 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
>>> kvm_hv_process_stimers(vcpu);
>>> }
>>>
>>> - /*
>>> - * KVM_REQ_EVENT is not set when posted interrupts are set by
>>> - * VT-d hardware, so we have to update RVI unconditionally.
>>> - */
>>> - if (kvm_lapic_enabled(vcpu)) {
>>> - /*
>>> - * Update architecture specific hints for APIC
>>> - * virtual interrupt delivery.
>>> - */
>>> - if (kvm_x86_ops->sync_pir_to_irr)
>>> - kvm_x86_ops->sync_pir_to_irr(vcpu);
>>> - }
>>> -
>>> if (kvm_check_request(KVM_REQ_EVENT, vcpu) || req_int_win) {
>>> ++vcpu->stat.req_event;
>>> kvm_apic_accept_events(vcpu);
>>> @@ -6767,20 +6754,39 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
>>> kvm_x86_ops->prepare_guest_switch(vcpu);
>>> if (vcpu->fpu_active)
>>> kvm_load_guest_fpu(vcpu);
>>> +
>>> + /*
>>> + * Disabling IRQs before setting IN_GUEST_MODE. Posted interrupt
>>> + * IPI are then delayed after guest entry, which ensures that they
>>> + * result in virtual interrupt delivery.
>>> + */
>>> + local_irq_disable();
>>> vcpu->mode = IN_GUEST_MODE;
>>>
>>> srcu_read_unlock(&vcpu->kvm->srcu, vcpu->srcu_idx);
>>>
>>> /*
>>> - * We should set ->mode before check ->requests,
>>> - * Please see the comment in kvm_make_all_cpus_request.
>>> - * This also orders the write to mode from any reads
>>> - * to the page tables done while the VCPU is running.
>>> - * Please see the comment in kvm_flush_remote_tlbs.
>>> + * 1) We should set ->mode before checking ->requests. Please see
>>> + * the comment in kvm_make_all_cpus_request.
>>> + *
>>> + * 2) For APICv, we should set ->mode before checking PIR.ON. This
>>> + * pairs with the memory barrier implicit in pi_test_and_set_on
>>> + * (see vmx_deliver_posted_interrupt).
>>> + *
>>> + * 3) This also orders the write to mode from any reads to the page
>>> + * tables done while the VCPU is running. Please see the comment
>>> + * in kvm_flush_remote_tlbs.
>>> */
>>> smp_mb__after_srcu_read_unlock();
>>>
>>> - local_irq_disable();
>>
>> The local_irq_disable() movement is unnecessary if you move sync_pir_to_irr.
>
> In addition, this movement will increase the time of irq disable to
> some degree. Do you think I can send a patch to revert it?
The difference is a few dozen hundred clock cycles, I don't think it
matters. Also, a posted interrupt sent to the host while IN_GUEST_MODE
is more expensive than one sent while the processor is in non-root mode.
All in all, I think it's preferrable to keep the local_irq_disable here.
Your observation seems correct though.
Paolo
> Regards,
> Wanpeng Li
>
>>
>> - IPI after vcpu->mode = IN_GUEST_MODE and interrupt disable, PI is
>> successfully.
>> - IPI between vcpu->mode = IN_GUEST_MODE and interrupt disable, the
>> sync_ir_to_irr will catch the PIR and set RVI.
>>
>> Regards,
>> Wanpeng Li
>>
>>> + if (kvm_lapic_enabled(vcpu)) {
>>> + /*
>>> + * This handles the case where a posted interrupt was
>>> + * notified with kvm_vcpu_kick.
>>> + */
>>> + if (kvm_x86_ops->sync_pir_to_irr)
>>> + kvm_x86_ops->sync_pir_to_irr(vcpu);
>>> + }
>>>
>>> if (vcpu->mode == EXITING_GUEST_MODE || vcpu->requests
>>> || need_resched() || signal_pending(current)) {
>>> --
>>> 1.8.3.1
>>>
