Re: [PATCH] Add Documentation/kvm/msr.txt

[Zachary Amsden <zamsden@redhat.com>]

On 05/27/2010 02:10 PM, Glauber Costa wrote:
> On Thu, May 27, 2010 at 11:02:35AM -1000, Zachary Amsden wrote:
>    
>> On 05/27/2010 10:36 AM, Glauber Costa wrote:
>>      
>>> On Thu, May 27, 2010 at 10:13:12AM -1000, Zachary Amsden wrote:
>>>        
>>>> On 05/27/2010 06:02 AM, Glauber Costa wrote:
>>>>          
>>>>> On Thu, May 27, 2010 at 11:15:43AM +0300, Avi Kivity wrote:
>>>>>            
>>>>>> On 05/26/2010 09:04 PM, Glauber Costa wrote:
>>>>>>              
>>>>>>> This patch adds a file that documents the usage of KVM-specific
>>>>>>> MSRs.
>>>>>>>
>>>>>>>                
>>>>>> Looks good.  A few comments:
>>>>>>
>>>>>>              
>>>>>>> +
>>>>>>> +Custom MSR list
>>>>>>> +--------
>>>>>>> +
>>>>>>> +The current supported Custom MSR list is:
>>>>>>> +
>>>>>>> +MSR_KVM_WALL_CLOCK:  0x11
>>>>>>> +
>>>>>>> +	data: physical address of a memory area.
>>>>>>>                
>>>>>> Which must be in guest RAM (i.e., don't point it somewhere random
>>>>>> and expect the hypervisor to allocate it for you).
>>>>>>
>>>>>> Must be aligned to 4 bytes (we don't enforce it though).
>>>>>>              
>>>>> I don't see the reason for it.
>>>>>
>>>>> If this is a requirement, our own implementation
>>>>> is failing to meet it.
>>>>>            
>>>> It's so the atomic write actually is atomic.
>>>>          
>>> Which atomic write? This is the wallclock, we do no atomic writes for
>>> querying it. Not to confuse with system time (the other msr).
>>>
>>>        
>>>> Stating a 4 -byte
>>>> alignment requirement prevents the wall clock from crossing a page
>>>> boundary.
>>>>          
>>> Yes, but why require it?
>>>
>>> reading the wallclock is not a hot path for anybody, is usually done
>>> just once, and crossing a page boundary here does not pose any correctness
>>> issue.
>>>        
>> Little-endian non-atomic page crossing writes will write the small
>> part of the wallclock first, so another CPU may observe the
>> following wallclock sequence:
>>
>> 0x01ff .. 0x0100 .. 0x0200
>>
>> Big-endian writes also have similar failure:
>>
>> 0x01ff .. 0x02ff .. 0x0200
>>
>> This won't happen if there is a single instruction write of the wall
>> clock word.
>>      
> We already specify that users can only trust the value of the wallclock
> after we have an even version field.
> When we start the update, and during the time of all writes to it,
> it is odd, and thus, invalid.
> The ABI guarantees to the guest that we'll only bump version
> after we're done updating.
>    

I guess there is not a point.