On Tuesday 02 Oct 2018 at 14:30:31 (+0200), Peter Zijlstra wrote:
> On Wed, Sep 12, 2018 at 10:12:58AM +0100, Quentin Perret wrote:
> > +/**
> > + * em_register_perf_domain() - Register the Energy Model of a performance
> > domain
> > + * @span : Mask of CPUs in the performance domain
> > + * @nr_states : Number of capacity states to register
> > + * @cb : Callback functions providing the data of the Energy
> > Model
> > + *
> > + * Create Energy Model tables for a performance domain using the callbacks
> > + * defined in cb.
> > + *
> > + * If multiple clients register the same performance domain, all but the
> > first
> > + * registration will be ignored.
> > + *
> > + * Return 0 on success
> > + */
> > +int em_register_perf_domain(cpumask_t *span, unsigned int nr_states,
> > + struct em_data_callback *cb)
> > +{
> > + unsigned long cap, prev_cap = 0;
> > + struct em_perf_domain *pd;
> > + int cpu, ret = 0;
> > +
> > + if (!span || !nr_states || !cb)
> > + return -EINVAL;
> > +
> > + /*
> > + * Use a mutex to serialize the registration of performance domains and
> > + * let the driver-defined callback functions sleep.
> > + */
> > + mutex_lock(&em_pd_mutex);
> > +
> > + for_each_cpu(cpu, span) {
> > + /* Make sure we don't register again an existing domain. */
> > + if (READ_ONCE(per_cpu(em_data, cpu))) {
> > + ret = -EEXIST;
> > + goto unlock;
> > + }
> > +
> > + /*
> > + * All CPUs of a domain must have the same micro-architecture
> > + * since they all share the same table.
> > + */
> > + cap = arch_scale_cpu_capacity(NULL, cpu);
> > + if (prev_cap && prev_cap != cap) {
> > + pr_err("CPUs of %*pbl must have the same capacity\n",
> > + cpumask_pr_args(span));
> > + ret = -EINVAL;
> > + goto unlock;
> > + }
> > + prev_cap = cap;
> > + }
> > +
> > + /* Create the performance domain and add it to the Energy Model. */
> > + pd = em_create_pd(span, nr_states, cb);
> > + if (!pd) {
> > + ret = -EINVAL;
> > + goto unlock;
> > + }
> > +
> > + for_each_cpu(cpu, span)
> > + WRITE_ONCE(per_cpu(em_data, cpu), pd);
>
> It's not immediately obvious to me why this doesn't need to be
> smp_store_release(). The moment you publish that pointer, it can be
> read, right?
>
> Even if you never again change the pointer value, you want to ensure the
> content of pd is stable before pd itself is observable, right?
So, I figured the mutex already gives me some of that. I mean, AFAIU it
should guarantee that concurrent callers to em_register_perf_domain are
serialized correctly.
For example, if I have two concurrent calls (let's name them A and B) to
em_register_perf_domain(), and say A takes the mutex first, then B
should be guaranteed to always see the totality of the update that A
made to the per_cpu table. Is that right ?
If the above is correct, then it's pretty much all I can do, I think ...
In the case of concurrent readers and writers to em_data, the
smp_store_release() call still doesn't give me the guarantee that the
per_cpu table is stable since em_cpu_get() is lock-free ...
If I want to be sure the per_cpu thing is stable from em_cpu_get() then
I can add a mutex_lock/unlock there too, but even then I won't need the
smp_store_release(), I think. Or maybe I got confused again ?
Thanks,
Quentin
