Re: [PATCH v1] lib: cpu_rmap: avoid flushing all workqueues

[Josh Triplett <josh@joshtriplett.org>]

On Fri, Dec 28, 2012 at 10:18:11AM -0800, David Decotigny wrote:
> Thank you, Josh,
> 
> A few comments below, and the revised version shortly.

Responses below.

> On Thu, Dec 27, 2012 at 8:04 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> > On Thu, Dec 27, 2012 at 11:24:34AM -0800, David Decotigny wrote:
> >> In some cases, free_irq_cpu_rmap() is called while holding a lock
> >> (eg. rtnl). This can lead to deadlocks, because it invokes
> >> flush_scheduled_work() which ends up waiting for whole system
> >> workqueue to flush, but some pending works might try to acquire the
> >> lock we are already holding.
> >>
> >> This commit uses reference-counting to replace
> >> irq_run_affinity_notifiers(). It also removes
> >> irq_run_affinity_notifiers() altogether.
> >>
> >> Signed-off-by: David Decotigny <decot@googlers.com>
> >
> > A couple of comments below; with those addressed,
> > Reviewed-by: Josh Triplett <josh@joshtriplett.org>
> >
> >> --- a/lib/cpu_rmap.c
> >> +++ b/lib/cpu_rmap.c
> >> @@ -230,16 +256,23 @@ irq_cpu_rmap_notify(struct irq_affinity_notify *notify, const cpumask_t *mask)
> >>               pr_warning("irq_cpu_rmap_notify: update failed: %d\n", rc);
> >>  }
> >>
> >> +/**
> >> + * irq_cpu_rmap_release - reclaiming callback for IRQ subsystem
> >> + * @ref: kref to struct irq_affinity_notify passed by irq/manage.c
> >> + */
> >>  static void irq_cpu_rmap_release(struct kref *ref)
> >>  {
> >>       struct irq_glue *glue =
> >>               container_of(ref, struct irq_glue, notify.kref);
> >> +     struct cpu_rmap *rmap = glue->rmap;
> >> +
> >>       kfree(glue);
> >> +     kref_put(&rmap->refcount, reclaim_cpu_rmap);
> >
> > Likewise, but also, why not call free_cpu_rmap(glue->rmap) before
> > kfree(glue) so you don't need the local copy?
> 
> I prefer to keep this kref_put here. I believe that calling something
> named "free_cpu_rmap" here might be misleading. It's code sharing vs.
> what we actually need to do, even though both are equivalent... for
> now.

If calling something named free_cpu_rmap feels wrong here, perhaps you
should call it cpu_rmap_put or cpu_rmap_unref or similar instead, since
it doesn't actually free unless the refcount goes to zero.  Then you
could have irq_cpu_rmap_release calling cpu_rmap_put, which feels more
natural.  But in any case, I think you should avoid having multiple
instances of the full call to kref_put on a cpu_rmap.

> For the order, it was deliberate, to have some kind of symmetry with
> kfree/kref_put in the error path we have in next function
> (irq_cpu_rmap_add). I reversed the order in that next function to
> avoid this unneeded local variable here. New ordering makes more sense
> anyways.

Ah, I see; makes sense to me.

> >>  }
> >>
> >>  /**
> >>   * irq_cpu_rmap_add - add an IRQ to a CPU affinity reverse-map
> >> - * @rmap: The reverse-map
> >> + * @rmap: The per-IRQ reverse-map
> >>   * @irq: The IRQ number
> >>   *
> >>   * This adds an IRQ affinity notifier that will update the reverse-map
> >> @@ -259,9 +292,12 @@ int irq_cpu_rmap_add(struct cpu_rmap *rmap, int irq)
> >>       glue->notify.release = irq_cpu_rmap_release;
> >>       glue->rmap = rmap;
> >>       glue->index = cpu_rmap_add(rmap, glue);
> >> +     kref_get(&rmap->refcount);
> >>       rc = irq_set_affinity_notifier(irq, &glue->notify);
> >> -     if (rc)
> >> +     if (rc) {
> >>               kfree(glue);
> >> +             kref_put(&rmap->refcount, reclaim_cpu_rmap);
> >
> > Likewise.
> 
> I prefer to leave the explicit kref_put here too.

In this case, for symmetry with kref_get?

Would it help to add a cpu_rmap_get, along with cpu_rmap_put?

static inline struct cpu_rmap *cpu_rmap_get(struct cpu_rmap *rmap)
{
    kref_get(&rmap->refcount);
    return rmap;
}

...
    glue->rmap = cpu_rmap_get(rmap);
...

> Next version soon, after some re-testing.

Thanks.

- Josh Triplett