RE: [Xen-devel] [PATCH] X86 MCE: Add SRAR handler

["Liu, Jinsong" <jinsong.liu@xxxxxxxxx>]

Jan Beulich wrote:
>>>> On 30.09.11 at 10:21, "Liu, Jinsong" <jinsong.liu@xxxxxxxxx> wrote:
>> Jan Beulich wrote:
>>>>>> On 30.09.11 at 09:44, "Liu, Jinsong" <jinsong.liu@xxxxxxxxx>
>>>>>> wrote: 
>>>> Jan Beulich wrote:
>>>>>>>> On 30.09.11 at 04:51, "Jiang, Yunhong"
>>>>>>>> <yunhong.jiang@xxxxxxxxx> wrote:
>>>>>>> This made me look at the current source, and there I see in
>>>>>>> mce_urgent_action() 
>>>>>>> 
>>>>>>>     if ( !(gstatus & MCG_STATUS_RIPV) && !guest_mode(regs))
>>>>>>> return -1; 
>>>>>>> 
>>>>>>> which I think should say ... _EIPV and use || instead. Thoughts?
>>>>>> 
>>>>>> I think this code means, if the error happens in hypervisor mode
>>>>>> (i.e. !guest_mode()), and RIPV indicate the RIP in stack can't be
>>>>>> restarted, we have to panic.
>>>>> 
>>>>> Then the guest_mode() check still lacks an extra check of EIPV,
>>>>> like 
>>>>> 
>>>>>      if ( !(gstatus & MCG_STATUS_RIPV) &&
>>>>>           (!(gstatus & MCG_STATUS_EIPV) || !guest_mode(regs)))
>>>>> return -1; 
>>>>> 
>>>> 
>>>> That would be overkilled.
>>>> Considering instruction fetch error occur at guest context,
>>>> hypervisor deliver to guest to handle the error is perfer, not
>>>> panic all system.
>>> 
>>> Even if it was hypervisor code that got prefetched while still
>>> executing guest code (which ought to be possible at least
>>> across a syscall/sysenter instruction)?
>>> 
>> 
>> Executing guest code will not satisfy the check
>> if ( !(gstatus & MCG_STATUS_RIPV) && !guest_mode(regs))      return -1;
>> so it would not panic system.
> 
> Exactly. But it should when the prefetch was to hypervisor code.
> 

Wouldn't processor refresh instruction prefetch queue under such case?

Thanks,
Jinsong
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel