[PATCH v3] Avoid memory barrier in read_seqcount() through load acquire

[Christoph Lameter (Ampere)]

On Tue, 17 Sep 2024, Will Deacon wrote:

> > +config ARCH_HAS_ACQUIRE_RELEASE
> > +	bool
> > +	help
> > +	  Setting ARCH_HAS_ACQUIRE_RELEASE indicates that the architecture
> > +	  supports load acquire and release. Typically these are more effective
> > +	  than memory barriers. Code will prefer the use of load acquire and
> > +	  store release over memory barriers if this option is enabled.
> > +
>
> Unsurprisingly, I'd be in favour of making this unconditional rather than
> adding a new Kconfig option. Would that actually hurt any architectures
> where we care about the last few shreds of performance?

Other arches do not have acquire / release and will create additional
barriers in the fallback implementation of smp_load_acquire. So it needs
to be an arch config option.

> > +	if (USE_COND_LOAD_ACQUIRE)
> > +		return smp_cond_load_acquire((unsigned int *)&s->sequence, (s->sequence & 1) == 0);
>
> This looks wrong to me.
>
> The conditional expression passed to smp_cond_load_acquire() should be
> written in terms of 'VAL', otherwise you're introducing an additional
> non-atomic access to the sequence counter.

Hmmm... The compiler seems to have optimized that out. Will use VAL in
next rollup.