Re: [PATCH] KVM: x86: fix xen guest panic due to lack of KVM_REQ_EVENT

[Wanpeng Li <wanpeng.li@linux.intel.com>]

On Mon, Aug 25, 2014 at 11:16:16AM +0200, Paolo Bonzini wrote:
>Il 25/08/2014 11:08, Wanpeng Li ha scritto:
>> Hi Paolo,
>> On Mon, Aug 25, 2014 at 11:01:07AM +0200, Paolo Bonzini wrote:
>>> Il 25/08/2014 09:58, Wanpeng Li ha scritto:
>>>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>>>> index c10408e..b7c0073 100644
>>>> --- a/arch/x86/kvm/x86.c
>>>> +++ b/arch/x86/kvm/x86.c
>>>> @@ -4928,6 +4928,8 @@ static void toggle_interruptibility(struct kvm_vcpu *vcpu, u32 mask)
>>>>  		if (!mask)
>>>>  			kvm_make_request(KVM_REQ_EVENT, vcpu);
>>>>  	}
>>>> +	if (!(int_shadow || mask))
>>>> +		kvm_make_request(KVM_REQ_EVENT, vcpu);
>>>>  }
>>>>  
>>>>  static void inject_emulated_exception(struct kvm_vcpu *vcpu)
>>>
>>> No, this patch undoes the optimization in the buggy patch.
>>>
>>> A KVM_REQ_EVENT must be missing somewhere else.
>>>
>> 
>> Could you give some tips in order that I can figure it out?
>
>I have no idea right now (I was planning to debug it this week).
>
>(BTW, look at the original commit that introduced KVM_REQ_EVENT --
>https://git.kernel.org/cgit/virt/kvm/kvm.git/commit/?id=3842d135 -- and
>compare the patch and the commit message.  You can see that it was added
>to the emulator because it is a "place that can set EFLAGS" and this
>idea is preserved in the buggy patch).
>

>From xen codes which report panic:

check_timer 
	timer_irq_works

	local_save_flags(flags);  => pushf;pop
	local_irq_enable();       => sti 
	delay xxxx 
	local_irq_restore(flags); => pushfq;andq;orq;popfq 

Regards,
Wanpeng Li 

>The important thing is that (despite Xen being involved) this is not
>related to nested virtualization.  So I would first of all try to see if
>some module parameter makes it go away (apicv and unrestricted mode
>especially), then capture a trace of the panic.  At least this is how I
>was planning to start... :)
>
>Paolo