Dominik,
Le 25/03/2010 18:44, Dominik Brodowski a écrit :
> From: Dominik Brodowski <linux@xxxxxxxxxxxxxxxxxxxx>
> Date: Thu, 25 Mar 2010 18:13:20 +0100
> Subject: [PATCH] cpufreq: use average load in conservative governor
>
> Instead of using the load of the last CPU in a package, use the
> average of all CPUs in a package.
Don't we want to check the max load rather than the average, for a user
running a single threaded job wants the CPU to jump to max speed?
>
> Reported-by: Jean-Christian Goussard <jeanchristian.goussard@xxxxxx>
> Signed-off-by: Dominik Brodowski <linux@xxxxxxxxxxxxxxxxxxxx>
>
> diff --git a/drivers/cpufreq/cpufreq_conservative.c b/drivers/cpufreq/cpufreq_conservative.c
> index 599a40b..e0be6bd 100644
> --- a/drivers/cpufreq/cpufreq_conservative.c
> +++ b/drivers/cpufreq/cpufreq_conservative.c
> @@ -444,6 +444,7 @@ static struct attribute_group dbs_attr_group_old = {
> static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
> {
> unsigned int load = 0;
> + unsigned int max_load = 0;
> unsigned int freq_target;
>
> struct cpufreq_policy *policy;
> @@ -500,8 +501,9 @@ static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
> if (unlikely(!wall_time || wall_time < idle_time))
> continue;
>
> - load = 100 * (wall_time - idle_time) / wall_time;
> + load += 100 * (wall_time - idle_time) / wall_time;
> }
> + load = load / cpumask_weight(policy->cpus);
>
> /*
> * break out if we 'cannot' reduce the speed as the user might
> @@ -511,7 +513,7 @@ static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
> return;
>
> /* Check for frequency increase */
> - if (load > dbs_tuners_ins.up_threshold) {
> + if (max_load > dbs_tuners_ins.up_threshold) {
max_load is still 0 here.
"load" now holds the average load but as I said, I think the max load should
be used:
- load = 100 * (wall_time - idle_time) / wall_time;
+ if (load > max_load)
+ max_load = load;
}
> this_dbs_info->down_skip = 0;
>
> /* if we are already at full speed then break out early */
> @@ -538,7 +540,7 @@ static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
> * can support the current CPU usage without triggering the up
> * policy. To be safe, we focus 10 points under the threshold.
> */
> - if (load < (dbs_tuners_ins.down_threshold - 10)) {
> + if (max_load < (dbs_tuners_ins.down_threshold - 10)) {
> freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
>
> this_dbs_info->requested_freq -= freq_target;
>
Jean-Christian
--
To unsubscribe from this list: send the line "unsubscribe cpufreq" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
