Re: [PATCH v5 1/4] sched/fair: add util_est on top of PELT

[Peter Zijlstra]

On Wed, Mar 07, 2018 at 11:47:11AM +0000, Patrick Bellasi wrote:
> On 06-Mar 20:02, Peter Zijlstra wrote:
> > On Thu, Feb 22, 2018 at 05:01:50PM +0000, Patrick Bellasi wrote:
> > > +struct util_est {
> > > +	unsigned int			enqueued;
> > > +	unsigned int			ewma;
> > > +#define UTIL_EST_WEIGHT_SHIFT		2
> > > +};
> > 
> > > +	ue = READ_ONCE(p->se.avg.util_est);
> > 
> > > +	WRITE_ONCE(p->se.avg.util_est, ue);
> > 
> > That is actually quite dodgy... and relies on the fact that we have the
> > 8 byte case in __write_once_size() and __read_once_size()
> > unconditionally. It then further relies on the compiler DTRT for 32bit
> > platforms, which is generating 2 32bit loads/stores.
> >
> > The advantage is of course that it will use single u64 loads/stores
> > where available.
> 
> Yes, that's mainly an "optimization" for 64bit targets... but perhaps
> the benefits are negligible.
> 
> Do you prefer to keep more "under control" the generated code by using
> two {READ,WRITE}_ONCEs?
> 
> IMO here we can also go with just the WRITE_ONCEs. I don't see a case
> for the compiler to mangle load/store. While the WRITE_ONCE are still
> required to sync with non rq-lock serialized code.
> But... maybe I'm missing something... ?

I'm not sure we rely on READ/WRITE_ONCE() of 64bit variables on 32bit
targets to be sane anywhere else (we could be, I just dont know).

I suspect it all works as expected... but its a tad tricky.