Re: sem_lock() vs qspinlocks

[Davidlohr Bueso <dave@stgolabs.net>]

On Sat, 21 May 2016, Peter Zijlstra wrote:

>On Fri, May 20, 2016 at 05:48:39PM -0700, Davidlohr Bueso wrote:
>> On Fri, 20 May 2016, Linus Torvalds wrote:
>>
>>
>> >Oh, I definitely agree on the stable part, and yes, the "splt things
>> >up" model should come later if people agree that it's a good thing.
>>
>> The backporting part is quite nice, yes, but ultimately I think I prefer
>> Linus' suggestion making things explicit, as opposed to consulting the spinlock
>> implying barriers. I also hate to have an smp_mb() (particularly for spin_is_locked)
>> given that we are not optimizing for the common case (regular mutual excl).
>
>I'm confused; we _are_ optimizing for the common case. spin_is_locked()
>is very unlikely to be used. And arguably should be used less in favour
>of lockdep_assert_held().

Indeed we are.

But by 'common case' I was really thinking about spin_is_locked() vs spin_wait_unlock().
The former being the more common of the two, and the one which mostly will _not_ be used
for lock correctness purposes, hence it doesn't need that new smp_mb. Hence allowing users
to explicitly set the ordering needs (ie spin_lock_synchronize()) seems like the better
long term alternative. otoh, with your approach all such bugs are automatically fixed :)

Thanks,
Davidlohr