Re: [RFC] locking: rwbase: Take care of ordering guarantee for fastpath reader

[Boqun Feng]

On Wed, Sep 08, 2021 at 02:14:28PM +0200, Peter Zijlstra wrote:
> On Wed, Sep 08, 2021 at 01:51:24PM +0200, Peter Zijlstra wrote:
> > On Wed, Sep 01, 2021 at 11:06:27PM +0800, Boqun Feng wrote:
> > @@ -201,23 +207,30 @@ static int __sched rwbase_write_lock(struct rwbase_rt *rwb,
> >  {
> >  	struct rt_mutex_base *rtm = &rwb->rtmutex;
> >  	unsigned long flags;
> > +	int readers;
> >  
> >  	/* Take the rtmutex as a first step */
> >  	if (rwbase_rtmutex_lock_state(rtm, state))
> >  		return -EINTR;
> >  
> >  	/* Force readers into slow path */
> > -	atomic_sub(READER_BIAS, &rwb->readers);
> > +	readers = atomic_sub_return_relaxed(READER_BIAS, &rwb->readers);
> 
> Hurmph... the above really begs for something like
> 
> 	if (!readers)
> 		return 0;
> 


I don't think we can return early here, don't we need to set WRITER_BIAS
to grab the write lock? And we can only do that with ->wait_lock held,
otherwise we race with the slowpath of readers.

Regards,
Boqun

> But then we needs that _acquire() thing again :/
>