Re: [PATCH v7 1/2] nohz: add tick_nohz_full_add_cpus_to() and _remove_cpus_from() APIs

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, Apr 13, 2015 at 08:49:20PM -0400, Chris Metcalf wrote:
> On 4/13/2015 8:33 PM, Frederic Weisbecker wrote:
> >On Fri, Apr 10, 2015 at 04:53:51PM -0400, Chris Metcalf wrote:
> >>The "removes_cpus_from" API is useful, for example, to modify a cpumask
> >>to avoid the nohz cores so that interrupts aren't sent to them.
> >>
> >>Likewise the "add_cpus_to" API is useful to modify a cpumask indicating
> >>some special nohz-type functionality so that the nohz cores are
> >>automatically added to that set.
> >>
> >>Signed-off-by: Chris Metcalf <cmetcalf@ezchip.com>
> >>---
> >>[...]
> >>
> >>diff --git a/include/linux/tick.h b/include/linux/tick.h
> >>index 9c085dc12ae9..8d1754c0f694 100644
> >>--- a/include/linux/tick.h
> >>+++ b/include/linux/tick.h
> >>@@ -186,6 +186,18 @@ static inline bool tick_nohz_full_cpu(int cpu)
> >>  	return cpumask_test_cpu(cpu, tick_nohz_full_mask);
> >>  }
> >>+static inline void tick_nohz_full_add_cpus_to(struct cpumask *mask)
> >>+{
> >>+	if (tick_nohz_full_enabled())
> >>+		cpumask_or(mask, mask, tick_nohz_full_mask);
> >>+}
> >>+
> >>+static inline void tick_nohz_full_remove_cpus_from(struct cpumask *mask)
> >>+{
> >>+	if (tick_nohz_full_enabled())
> >>+		cpumask_andnot(mask, mask, tick_nohz_full_mask);
> >I would prefer that you don't introduce new APIs if they aren't used in your
> >patchset. It seems that's the case for tick_nohz_full_remove_cpus_from().
> 
> Yes, it's used in a tile-tree patch to remove nohz_full cpus from the set that
> take interrupts from the tilegx network driver.
> 
> I can certainly make this patchset have just the _add_cpus_to() variant,
> and the other patchset have just the _remove_cpus_from() variant.
> It seemed to make sense to include them together as a matched set,
> but doing it the way you suggest makes an equal if different amount of sense.
> 
> >Also we have housekeeping_affine() that affines a task to CPUs outside the
> >range of nohz full. In case you still need tick_nohz_full_remove_cpus_from()
> >in a further patchset, housekeeping_affine_cpumask() would extend the existing
> >naming.
> 
> I like housekeeping_affine(), but it overwrites the affinity mask of
> the task.  So I would expect housekeeping_affine_mask() to overwrite
> the specified argument cpumask, which it doesn't in my definition.
> 
> I don't know that I can think of a good name in the "housekeeping_xxx"
> namespace that feels better than the one I proposed.  In context of the
> proposed client of the API so far, it's:
> 
>        if (!network_cpus_init()) {
>                network_cpus_map = *cpu_online_mask;
> tick_nohz_full_remove_cpus_from(&network_cpus_map);
>        }
> 
> If housekeeping_mask were defined for non-nohz_full I could just use
> it unconditionally here.  Alternately I could just put in an #ifdef for
> CONFIG_NO_HZ_FULL and either use cpu_only_mask or housekeeping_mask
> to initialize network_cpus_map, which dodges the bullet of creating
> an acceptable API name here for the moment.
> 
> Frederic, what's your thought/preference?

Right, the more I look into this whole, the more I think we should perhaps
make tick_nohz_full_cpumask a read-only wide visible cpumask like we do
for the cpumask in kernel/cpu.c defined as "const struct cpumask *const tick_nohz_full_mask"
and do the cpumask operations with it.

> 
> -- 
> Chris Metcalf, EZChip Semiconductor
> http://www.ezchip.com
>