Re: [PATCH 01/18] led-triggers: create a trigger for CPU activity

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 30 Mar 2012 19:58:16 +0800
Bryan Wu <bryan.wu@canonical.com> wrote:

> Attempting to consolidate the ARM LED code, this removes the
> custom RealView LED trigger code to turn LEDs on and off in
> response to CPU activity and replace it with a standard trigger.
> 
> (bryan.wu@canonical.com:
> It moves arch/arm/kernel/leds.c syscore stubs into this trigger.
> It also provides ledtrig_cpu trigger event stub in <linux/leds.h>.
> Although it was inspired by ARM work, it can be used in other arch.)
> 
>
> ...
>
> +#include <linux/percpu.h>
> +#include <linux/syscore_ops.h>
> +#include "leds.h"
> +
> +#define MAX_NAME_LEN	8

The kernel already has at least eight different definitions of
MAX_NAME_LEN.  I guess a ninth won't hurt ;)

>
> ...
>
> +static void __init ledtrig_cpu_register(void)
> +{
> +	int cpuid = smp_processor_id();
> +	struct led_trigger *trig;
> +	char *name = __get_cpu_var(trig_name);
> +
> +	snprintf(name, MAX_NAME_LEN, "cpu%d", cpuid);
> +	led_trigger_register_simple(name, &trig);
> +
> +	pr_info("LED trigger %s indicate activity on CPU %d\n",
> +		trig->name, cpuid);
> +
> +	__get_cpu_var(cpu_trig) = trig;
> +}
> +
> +static void __exit ledtrig_cpu_unregister(void)
> +{
> +	struct led_trigger *trig = __get_cpu_var(cpu_trig);
> +	char *name = __get_cpu_var(trig_name);
> +
> +	led_trigger_unregister_simple(trig);
> +	__get_cpu_var(cpu_trig) = NULL;
> +	memset(name, 0, MAX_NAME_LEN);
> +}
> +
> +static int __init ledtrig_cpu_init(void)
> +{
> +	int cpu;
> +
> +	for_each_possible_cpu(cpu)
> +		ledtrig_cpu_register();
> +
> +	register_syscore_ops(&ledtrig_cpu_syscore_ops);
> +
> +	return 0;
> +}
> +module_init(ledtrig_cpu_init);

This all looks horridly broken.  We call ledtrig_cpu_register() once
for each CPU, but we call it on the same CPU each time, and it uses
smp_processor_id() which a) is wrong and b) will emit a runtime
preemption-is-enabled warning.

I'm assuming this wasn't tested on SMP.  It needs to be, please.

Also, the code uses for_each_possible_cpu, ignoring CPU hotplug. 
That's probably justifiable for this small storage size and not-hotpath
code, but the decision should be given prominence and justified in the
changelog or, better, in code comments please.

>
> ...
>

The patchset is basically an ARM thing.  Is there some ARM tree via
which we can get it merged?