Re: [PATCH 05/10] nohz: New tick dependency mask

[Peter Zijlstra]

On Mon, Aug 03, 2015 at 03:55:34PM +0200, Frederic Weisbecker wrote:
> On Mon, Aug 03, 2015 at 03:29:58PM +0200, Peter Zijlstra wrote:
> > On Mon, Aug 03, 2015 at 03:09:39PM +0200, Frederic Weisbecker wrote:
> > 
> > > > That doesn't make any sense:
> > > > 
> > > >   tick_nohz_set_tick_dependency_this_cpu();
> > > > 
> > > > (shees, you're nowhere near lazy enough, that's insane to type) is
> > > > almost identical to:
> > > > 
> > > >   tick_nohz_set_tick_dependency_cpu(.cpu = smp_processor_id());
> > > > 
> > > > The only difference is a _very_ slight reduction in cost for computing
> > > > the per-cpu offset.
> > > 
> > > But the local one must be NMI-safe. Now I can do:
> > > 
> > >     if (cpu == smp_processor_id())
> > >         tick_nohz_full_kick() // NMI-safe
> > >     else
> > >         tick_nohz_full_kick_cpu(cpu); // not NMI-safe.
> > 
> > Urgh, I missed that. But yes, I suppose that's ok seeing how we result
> > in a smaller interface.
> > 
> > I was going to say that with a bit of luck GCC could optimize it, but
> > its not inline so no it cannot.
> 
> I might inline all these set_dep() things to introduce static keys on these
> APIs.. But the kick itself will remain real calls.

Sure, but first check if GCC will optimize:

static inline void foo(int cpu)
{
	if (cpu == smp_processor_id())
		bar1();
	else
		bar2();
}


	foo(smp_processor_id());

Into a direct call to bar1(), if not see if we can make it so. If not,
there's no point in inlining at all.

> Ok how about tick_nohz_set_dep_nmi() so that we know exactly what's the purpose
> here. Still a long function name but it's clear.

Only for the set, if you really care about it. The alternative is
WARN_ON(in_nmi() && cpu != smp_processor_id()) or somesuch.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/