[PATCH 01/17] leds: create a trigger for CPU activity

[jamie@jamieiles.com (Jamie Iles)]

Hi Linus,

On Wed, Aug 03, 2011 at 05:22:02PM +0200, Linus Walleij wrote:
> On Wed, Aug 3, 2011 at 12:28 PM, Jamie Iles <jamie@jamieiles.com> wrote:
> > Hi Bryan,
> > [...]
> >> +static DEFINE_PER_CPU(struct ledtrig_cpu_data *, ledtrig_cpu_triggers);
> > [...]
> >> +static void ledtrig_cpu_activate_cpu(void *data)
> >> +{
> >> + ? ? struct ledtrig_cpu_data *cpu_data;
> >> + ? ? struct led_classdev *led = data;
> >> + ? ? int my_cpu = smp_processor_id();
> >> + ? ? unsigned long target_cpu = (unsigned long) led->trigger_data;
> >> +
> >> + ? ? if (target_cpu != my_cpu)
> >> + ? ? ? ? ? ? return;
> >> +
> >> + ? ? cpu_data = kzalloc(sizeof(*cpu_data), GFP_KERNEL);
> >> + ? ? if (!cpu_data)
> >> + ? ? ? ? ? ? return;
> >> +
> >> + ? ? dev_info(led->dev, "led %s indicate activity on CPU %d\n",
> >> + ? ? ? ? ? ? ?led->name, my_cpu);
> >> +
> >> + ? ? cpu_data->led = led;
> >> + ? ? __get_cpu_var(ledtrig_cpu_triggers) = cpu_data;
> >> +}
> >> +
> >> +static void ledtrig_cpu_activate(struct led_classdev *led)
> >> +{
> >> + ? ? on_each_cpu(ledtrig_cpu_activate_cpu, led, 1);
> >> +}
> >> +
> >> +static void ledtrig_cpu_deactivate(struct led_classdev *led)
> >> +{
> >> + ? ? struct ledtrig_cpu_data *cpu_data = led->trigger_data;
> >> +
> >> + ? ? kfree(cpu_data);
> >> +}
> >
> > Is this deactivation correct? ?My (limited) understanding of the smp api
> > is that we'll allocate the ledtrig_cpu_data for each CPU and store it in
> > the ledtrig_cpu_triggers pointers.
> 
> At this point you have a handle to a specific LED and that one is tied
> to a certain CPU aldready, since it is a member of
> struct ledtrig_cpu_data. The LED is CPU-local by nature already since
> it is looked up in the code being called from the ARM kernel,
> i.e. the activate/deactivate pairs get called once per CPU.

OK, I'm fine with that.
> 
> > So shouldn't this be doing a
> > __get_cpu_var(ledtrig_cpu_triggers) and a kfree() on that (and setting
> > to NULL)?
> 
> You could do that but why? There is no way the pointer value
> can make any harm. If activate() is called again it will be overwritten
> with the new pointer.

Yes, but can't ledtrig_cpu() potentially still be called after 
deactivation, then the:

	if (!trigdata)
		return;

would be incorrect?  My main point though was that it looks like 
led->trigger_data is the processor id and kfree() is being called on 
that.

> > Also, where does led->trigger_data get assigned with the cpu id?
> 
> The activate() function is called on every core and if the CPU ID
> is not equal to the assigned CPU it just exits, here:
> 
> +     int my_cpu = smp_processor_id();
> +     unsigned long target_cpu = (unsigned long) led->trigger_data;
> +
> +     if (target_cpu != my_cpu)
> +             return;

I understand that bit, but I can't see anything that actually assigns 
led->trigger_data in the first place.

Jamie