On Tue, Oct 02, 2018 at 03:40:28PM +0100, Quentin Perret wrote:
> On Tuesday 02 Oct 2018 at 16:29:24 (+0200), Peter Zijlstra wrote:
> > On Tue, Oct 02, 2018 at 03:05:23PM +0100, Quentin Perret wrote:
> > > On Tuesday 02 Oct 2018 at 15:48:57 (+0200), Peter Zijlstra wrote:
> > > > +/**
> > > > + * em_cpu_get() - Return the performance domain for a CPU
> > > > + * @cpu : CPU to find the performance domain for
> > > > + *
> > > > + * Return: the performance domain to which 'cpu' belongs, or NULL if
> > > > it doesn't
> > > > + * exist.
> > > > + */
> > > > +struct em_perf_domain *em_cpu_get(int cpu)
> > > > +{
> > > > + return READ_ONCE(per_cpu(em_data, cpu));
> > > > +}
> > > > +EXPORT_SYMBOL_GPL(em_cpu_get);
> > > >
> > > > But your read side doesn't take, not is required to take em_pd_mutex.
> > > >
> > > > At that point, the mutex_unlock() doesn't guarantee anything.
> > > >
> > > > A CPU observing the em_data store, doesn't need to observe the store
> > > > that filled the data structure it points to.
> > >
> > > Right but even if I add the smp_store_release(), I can still have a
> > > CPU observing em_data while another is in the process of updating it.
> > > So, if smp_store_release() doesn't guarantee that readers will see a
> > > complete update, do I actually get something interesting from it ?
> > > (That's not a rhetorical question, I'm actually wondering :-)
> >
> > I thought the update would fail if em_data was already set.
> >
> > That is, you can only set this thing up _once_ and then you'll have to
> > forever live with it.
> >
> > Or did I read that wrong?
>
> No no, that's correct. em_data is populated once and kept as-is
> forever.
>
> What I was trying to say is, when em_data is being populated for the
> first time, nothing prevents a reader from using em_cpu_get()
> concurrently. And in this case, it doesn't matter if you use
> smp_store_release() or not, the reader might see the table half-updated.
>
> So, basically, smp_store_release() doesn't guarantee that readers won't
> see a half-baked em_data. That's the point I'm trying to make at least :-)
An example might help clarify this: here is a scenario I can _imagine,
based on your description.
CPU0 (em_register_perf_domain()) CPU1 (reader)
[...] my_pd = READ_ONCE(per_cpu(em_data, 1));
/* em_cpu_get() */
pd->table = table if (my_pd)
WRITE_ONCE(per_cpu(em_data, 1), pd); my_table = my_pd->table; /*
process, dereference, ... my_table */
In this scenario, we'd like CPU1 to see CPU0's store to ->table (as well
as the stores to table[]) _if CPU1 sees CPU0's store to em_data (that is,
if my_pd != NULL).
This guarantee does not hold with the WRITE_ONCE(), because CPU0 could
propagate the store to ->table and the store to em_data out-of-order.
The smp_store_release(), together with the address dependency headed by
the READ_ONCE(), provides this guarantee (and more...).
(Enclosing the reader into an em_pd_mutex critical section would also
provide this guarantee, but I can imagine a few arguments for not using
a mutex... ;-) ).
The question, I guess, is whether you want such a guarantee.
Andrea
