Re: [Qemu-devel] [PATCH 1/1 V5 tuning] kernel/kvm: introduce KVM_SET_LINT1 and fix improper nmi emulation

[Lai Jiangshan <laijs@cn.fujitsu.com>]

On 10/14/2011 08:07 PM, Jan Kiszka wrote:
> On 2011-10-14 13:59, Sasha Levin wrote:
>> On Fri, 2011-10-14 at 17:51 +0800, Lai Jiangshan wrote:
>>> Currently, NMI interrupt is blindly sent to all the vCPUs when NMI
>>> button event happens. This doesn't properly emulate real hardware on
>>> which NMI button event triggers LINT1. Because of this, NMI is sent to
>>> the processor even when LINT1 is masked in LVT. For example, this
>>> causes the problem that kdump initiated by NMI sometimes doesn't work
>>> on KVM, because kdump assumes NMI is masked on CPUs other than CPU0.
>>>
>>> With this patch, we introduce introduce KVM_SET_LINT1,
>>> and we can use KVM_SET_LINT1 to correctly emulate NMI button
>>> without change the old KVM_NMI behavior.
>>>
>>> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
>>> Reported-by: Kenji Kaneshige <kaneshige.kenji@jp.fujitsu.com>
>>> ---
>>
>> It could use a documentation update as well.
>>
>>>  arch/x86/kvm/irq.h   |    1 +
>>>  arch/x86/kvm/lapic.c |    7 +++++++
>>>  arch/x86/kvm/x86.c   |    8 ++++++++
>>>  include/linux/kvm.h  |    3 +++
>>>  4 files changed, 19 insertions(+), 0 deletions(-)
>>> diff --git a/arch/x86/kvm/irq.h b/arch/x86/kvm/irq.h
>>> index 53e2d08..0c96315 100644
>>> --- a/arch/x86/kvm/irq.h
>>> +++ b/arch/x86/kvm/irq.h
>>> @@ -95,6 +95,7 @@ void kvm_pic_reset(struct kvm_kpic_state *s);
>>>  void kvm_inject_pending_timer_irqs(struct kvm_vcpu *vcpu);
>>>  void kvm_inject_apic_timer_irqs(struct kvm_vcpu *vcpu);
>>>  void kvm_apic_nmi_wd_deliver(struct kvm_vcpu *vcpu);
>>> +void kvm_apic_lint1_deliver(struct kvm_vcpu *vcpu);
>>>  void __kvm_migrate_apic_timer(struct kvm_vcpu *vcpu);
>>>  void __kvm_migrate_pit_timer(struct kvm_vcpu *vcpu);
>>>  void __kvm_migrate_timers(struct kvm_vcpu *vcpu);
>>> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
>>> index 57dcbd4..87fe36a 100644
>>> --- a/arch/x86/kvm/lapic.c
>>> +++ b/arch/x86/kvm/lapic.c
>>> @@ -1039,6 +1039,13 @@ void kvm_apic_nmi_wd_deliver(struct kvm_vcpu *vcpu)
>>>  		kvm_apic_local_deliver(apic, APIC_LVT0);
>>>  }
>>>  
>>> +void kvm_apic_lint1_deliver(struct kvm_vcpu *vcpu)
>>> +{
>>> +	struct kvm_lapic *apic = vcpu->arch.apic;
>>> +
>>> +	kvm_apic_local_deliver(apic, APIC_LVT1);
>>> +}
>>> +
>>>  static struct kvm_timer_ops lapic_timer_ops = {
>>>  	.is_periodic = lapic_is_periodic,
>>>  };
>>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>>> index 84a28ea..fccd094 100644
>>> --- a/arch/x86/kvm/x86.c
>>> +++ b/arch/x86/kvm/x86.c
>>> @@ -2077,6 +2077,7 @@ int kvm_dev_ioctl_check_extension(long ext)
>>>  	case KVM_CAP_XSAVE:
>>>  	case KVM_CAP_ASYNC_PF:
>>>  	case KVM_CAP_GET_TSC_KHZ:
>>> +	case KVM_CAP_SET_LINT1:
>>>  		r = 1;
>>>  		break;
>>>  	case KVM_CAP_COALESCED_MMIO:
>>> @@ -3264,6 +3265,13 @@ long kvm_arch_vcpu_ioctl(struct file *filp,
>>>  
>>>  		goto out;
>>>  	}
>>> +	case KVM_SET_LINT1: {
>>> +		r = -EINVAL;
>>> +		if (!irqchip_in_kernel(vcpu->kvm))
>>> +			goto out;
>>> +		r = 0;
>>> +		kvm_apic_lint1_deliver(vcpu);
>>
>> We simply ignore the return value of kvm_apic_local_deliver() and assume
>> it always works. why?
>>
> 
> Hmm, I suddenly realized that we switched from enhancing the KVM_NMI
> IOCTL to adding KVM_SET_LINT1 - what motivated this?

Enhancing the KVM_NMI directly fixes the problem and matches the
real hard ware more, but it changes API bahavior.(we preferred to this one)

>From the previous mails, I found you and Avi prefer to SET_LINT1
which keep old behavior and it is also OK for us.
But I found it is hard to be implemented before, and I switched
this one when you told me the clue.


> 
> ( Maybe we should let the kernel part settle first before iterating
> through user space changes. )
> 


Yes, you are right, we should settle the kernel-site at first,
But I need you and Avi's suggestions.

Thanks,
Lai