Re: [PATCH v4 0/3] qspinlock: Introducing a 4-byte queue spinlock

[Waiman Long <waiman.long@hp.com>]

On 02/19/2014 03:17 PM, Linus Torvalds wrote:
> On Wed, Feb 19, 2014 at 11:24 AM, Waiman Long<waiman.long@hp.com>  wrote:
>> We do have a CONFIG_PARAVIRT macro that turns on or off PV support. The
>> queue spinlock can be easily changed into an unfair lock which allows lock
>> stealing. We could have a config option to make it unfair in the PARAVIRT
>> environment, but I don't think Linus like the idea of an unfair lock.
> I could care less for the paravirt case. As long as the native case
> ends up being sane (even when CONFIG_PARAVIRT is set at compile time),
> I'm fine.
>
> When actually running in a paravirtualized environment, locks are
> always going to have problems.
>
>                    Linus

I think we could implement 2 versions of _raw_spin_lock. The primary one 
is fair and the other one is unfair with a jump label that jump from the 
fair version to the unfair version. At boot time, the kernel can check 
to see it is really running in a PV environment and then activate the 
jump label to use the unfair version.

Now the key is how to detect if a kernel is really running in a PV 
environment. I need to ask some virtualization experts on that.

-Longman