Re: vmstat: On demand vmstat workers V5

From: Paul E. McKenney
Date: Thu May 29 2014 - 12:29:30 EST


On Thu, May 29, 2014 at 04:26:05PM +0200, Frederic Weisbecker wrote:
> On Thu, May 29, 2014 at 09:07:44AM -0500, Christoph Lameter wrote:
> > On Thu, 29 May 2014, Frederic Weisbecker wrote:
> >
> > > The cpumasks in cpu.c are special as they are the base of the cpumask_var_t
> > > definition. They are necessary to define nr_cpu_bits which is the base of
> > > cpumask_var_t allocations. As such they must stay lower level and defined
> > > on top of NR_CPUS.
> > >
> > > But most other cases don't need that huge static bitmap. I actually haven't
> > > seen any other struct cpumask than isn't based on cpumask_var_t.
> >
> > Well yes and I am tying directly into that scheme there in cpu.c to
> > display the active vmstat threads in sysfs. so its the same.
>
> I don't think so. Or is there something in vmstat that cpumask_var_t
> definition depends upon?
>
> >
> > > Please post it on a new thread so it gets noticed by others.
> >
> > Ok. Will do when we got agreement on the cpumask issue.
> >
> > I would like to have some way to display the activities on cpus in /sysfs
> > like I have done here with the active vmstat workers.
> >
> > What I think we need is display cpumasks for
> >
> > 1. Cpus where the tick is currently off
> > 2. Cpus that have dynticks enabled.
> > 3. Cpus that are idle
>
> You should find all that in /proc/timer_list
>
> Now for CPUs that have full dynticks enabled, we probably need something
> in sysfs. We could dump the nohz cpumask somewhere. For now you can only grep
> the dmesg
>
> > 4. Cpus that are used for RCU.
>
> So, you mean those that aren't in extended grace period (between rcu_user_enter()/exit
> or rcu_idle_enter/exit)?
>
> Paul?

We are clearly going to have to be very careful to avoid cache thrashing,
so methods that update a CPU mask on each transition, as the saying goes,
"need not apply".

So we need a function like __rcu_is_watching(), but that takes the
CPU number as an argument. Something like the following:

include/linux/rcutree.h:

bool rcu_is_watching_cpu(int cpu);

kernel/rcu/tree.c:

bool rcu_is_watching_cpu(int cpu)
{
return atomic_read(per_cpu(&rcu_dynticks.dynticks), cpu) & 0x1;
}

include/linux/rcutiny.h:

static inline bool rcu_is_watching_cpu(int cpu)
{
return true;
}

This could then be invoked from the appropriate sysfs or /proc setup.

Thanx, Paul

--
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/