Re: [PATCH 1/7] sched: Add static_key for asymmetric cpu capacity optimizations

[Quentin Perret]

Hi Morten,

On Friday 16 Feb 2018 at 15:41:01 (+0000), Morten Rasmussen wrote:
> On Fri, Feb 16, 2018 at 02:47:04PM +0100, Peter Zijlstra wrote:
> > On Thu, Feb 15, 2018 at 04:20:48PM +0000, Morten Rasmussen wrote:
> > > +static void update_asym_cpucapacity(int cpu)
> > > +{
> > > +	if (!static_branch_unlikely(&sched_asym_cpucapacity) &&
> > > +	    lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY))
> > > +		static_branch_enable(&sched_asym_cpucapacity);
> > > +}
> > 
> > That looks odd, why not just:
> > 
> > 	if (lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY))
> > 		static_branch_enable(&sched_asym_cpucapacity);
> 
> I actually had that initially and then I misread the implementation of
> static_key_enable() as if it trigger the WARN_ON_ONCE() condition if I
> enable an already enabled static key. But I see now that it should be
> safe to do.

AFAIU it should be safe, but without your check you'll have to go through
cpus_read_lock()/unlock() every time a CPU is hotplugged. There is probably
no good reason to re-do that again and again if the state of the key
never changes. I don't know how expensive those lock/unlock operations are,
but I bet they are more expensive than testing static_branch_unlikely() :)

> 
> > ? possibly with:
> > 
> > 	else
> > 		static_branch_disable(&sched_asym_cpucapacity);
> > 
> > if you want to play funny games :-)
> 
> I thought about that too. It could make certain hotplug scenarios even
> more expensive. I think we want the sched_asym_cpucapacity code to behave
> even if SD_ASYM_CPUCAPACITY isn't set anywhere, so the static key would
> be permanently from the point we detect asymmetry and leave it set. This
> would be in line with how the smt static key works.

Thanks !
Quentin