Re: Use of barriers in pvclock ABI

[Avi Kivity <avi@qumranet.com>]

Glauber Costa wrote:
> On Mon, Aug 11, 2008 at 4:08 AM, Gerd Hoffmann <kraxel@redhat.com> wrote:
>   
>>  Hi,
>>
>>     
>>> However, the pvclock_clocksource_read() implementation is
>>> over-engineered, because it checks for an odd version and uses very
>>> strong rmb() barriers (which generates either an "lfence" or "lock add
>>> $0, (%esp)").
>>>
>>> If we're happy to guarantee as an ABI issue that the record will never
>>> be updated cross-cpu, then we can make the barriers simply barrier() and
>>> just check for (src->version != dst->version).
>>>
>>> Is that OK with you, or do you want to leave open the possibility of
>>> doing cross-cpu time updates?
>>>       
>> Due to the TSC being involved here I don't expect cross-cpu time updates
>> will ever happen.  IMHO it is fine to change that.
>>     
>
> Okay for guest vcpu, but what about physical cpus?
>
>   
> IIRC, the checks are there, and so strict, to account for the
> possiblity of the vcpu to be migrated to another cpu in the middle of
> the
>   


Migration implies an smp barrier (but not a compiler barrier, of 
course).  As to the the version number oddness check, I don't see how it 
can be necessary.


-- 
error compiling committee.c: too many arguments to function