Re: [RFC PATCH 04/19] cpufreq: bring data structures close to their locks

[Peter Zijlstra]

On Tue, Jan 12, 2016 at 12:03:39AM +0100, Rafael J. Wysocki wrote:
> On Monday, January 11, 2016 11:05:28 PM Peter Zijlstra wrote:
> > On Mon, Jan 11, 2016 at 05:35:45PM +0000, Juri Lelli wrote:
> > > +/**
> > > + * The "cpufreq driver" - the arch- or hardware-dependent low
> > > + * level driver of CPUFreq support, and its spinlock (cpufreq_driver_lock).
> > > + * This lock also protects cpufreq_cpu_data array and cpufreq_policy_list.
> > > + */
> > > +static struct cpufreq_driver *cpufreq_driver;
> > > +static DEFINE_PER_CPU(struct cpufreq_policy *, cpufreq_cpu_data);
> > >  static LIST_HEAD(cpufreq_policy_list);
> > > +static DEFINE_RWLOCK(cpufreq_driver_lock);
> > 
> > Part of my suggestion was to fold the per-cpu data of cpufreq_cpu_data
> > into struct cpufreq_driver.
> > 
> > That way each cpufreq_driver will have its own copy and there'd be only
> > the one global pointer to swizzle. Something very well suited to RCU.
> 
> Well, I'm not really sure reworking all that is necessary.
> 
> What we need is to be able to call something analogous to dbs_timer_handler()
> from the scheduler and a driver callback from there (if present).  For that,
> it should be sufficient to have a pointer to that callback (that may be set
> upon driver registration) protected by RCU (or should that be sched RCU
> rather?) if I'm not missing anything.

But such a callback will invariably want to use the per-cpu state. And
now you have two pointers, one for the driver and one for the per-cpu
state. Keeping that in sync is a pain.

Moving the per-cpu data into the driver solves that trivially.