Re: [Xen-devel] Re: [PATCH 3/5] x86/pvclock: add vsyscall implementation

[Avi Kivity <avi@redhat.com>]

On 10/07/2009 11:51 PM, Jeremy Fitzhardinge wrote:
> On 10/07/09 14:37, Avi Kivity wrote:
>    
>> If the guest does a RMW on the version, but the host does not (copying
>> it from somewhere else), then the guest RMW can be lost.
>>
>> Looking at the code, that's what kvm does:
>>
>>      vcpu->hv_clock.version += 2;
>>
>>      shared_kaddr = kmap_atomic(vcpu->time_page, KM_USER0);
>>
>>      memcpy(shared_kaddr + vcpu->time_offset,&vcpu->hv_clock,
>>             sizeof(vcpu->hv_clock));
>>
>> so a guest-side ++version can be lost.
>>      
> I see, yes.  The Xen code explicitly reads back the guest version and
> increments that (I realize now that's what you meant by guest-private
> version).  If you were to have a second version number it would have to
> be separated as well to avoid being overwritten by the hypervisor.
>    

Yes.  We have the space since a cacheline is 64 bytes (minimum) vs 32 
bytes of pvclock data.

-- 
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.