From: Pallipadi, Venkatesh on

Some comments/questions inlined below.

On Tue, 2010-01-26 at 17:06 -0800, Mike Chan wrote:
> Allow lower priority threads to scale frequency to specified nice_max_freq.
> This allows low priority threads to operate at the most efficient
> power/performance frequency.
>
> Often the highest and lowest cpu speeds do not provide the the optimal
> performance/power ratios. Latency requirements for normal and high priority
> threads require the maximum speed that are not always optimal power wise
> inorder to satisfy the requirements.
>
> To enable set nice_max_freq (to a speed lower than the scaling_max_freq).
>
> The governor will first attempt to scale the cpu to policy->max (default)
> only using normal and high priority threads. It will ignore nice threads.
> If the load is high enough without nice threads then ondemand will scale to
> the max speed and exit.
>
> If load for normal and high priority threads are not high enough to increase
> the cpu speed, check again including the load from nice threads. Only scale
> to the nice_max_freq specified.
>
> Previous behavior is maintained by setting the values below:
>
> + When nice_max_freq is set to 0, behavior is the current default
> (nice is counted for load).
>
> + When nice_max_freq is set to scaling_min_freq, the behavior is the same
> as the original ignore_nice_load == 1. Which counts all nice threads as
> idle time when computing cpu load.
>
> *** v2 ***
> + The ignore_nice_load sysfs still behaves the same as before (0/1) and is
> kept around for legacy support. Userspace scripts should now use
> nice_max_freq.
>
> *** v3 ***
> Fixed: Properly scale freq if powersave_bias is enabled.
>
> Signed-off-by: Mike Chan <mike(a)android.com>
> ---
> drivers/cpufreq/cpufreq_ondemand.c | 156 ++++++++++++++++++++++++++----------
> 1 files changed, 113 insertions(+), 43 deletions(-)
>
> diff --git a/drivers/cpufreq/cpufreq_ondemand.c b/drivers/cpufreq/cpufreq_ondemand.c
> index 516d0fe..92727c4 100644
> --- a/drivers/cpufreq/cpufreq_ondemand.c
> +++ b/drivers/cpufreq/cpufreq_ondemand.c
> @@ -1,5 +1,4 @@
> -/*
> - * drivers/cpufreq/cpufreq_ondemand.c
> +/* * drivers/cpufreq/cpufreq_ondemand.c
> *
> * Copyright (C) 2001 Russell King
> * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi(a)intel.com>.

Above hunk is by accident?

> @@ -108,11 +107,13 @@ static struct dbs_tuners {
> unsigned int down_differential;
> unsigned int ignore_nice;
> unsigned int powersave_bias;
> + unsigned int nice_max_freq;
> } dbs_tuners_ins = {
> .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
> .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
> .ignore_nice = 0,
> .powersave_bias = 0,
> + .nice_max_freq = 0,
> };
>
> static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
> @@ -251,6 +252,7 @@ static ssize_t show_##file_name \
> show_one(sampling_rate, sampling_rate);
> show_one(up_threshold, up_threshold);
> show_one(ignore_nice_load, ignore_nice);
> +show_one(nice_max_freq, nice_max_freq);
> show_one(powersave_bias, powersave_bias);
>
> /*** delete after deprecation time ***/
> @@ -318,27 +320,60 @@ static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
> return count;
> }
>
> +/*
> + * Preserve ignore_nice_load behavior, if enabled do not allow low priority
> + * threads to scale beyond the minimum frequency.
> + */
> static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
> const char *buf, size_t count)
> {
> - unsigned int input;
> - int ret;
> -
> + unsigned long input;
> unsigned int j;
>
> - ret = sscanf(buf, "%u", &input);
> - if (ret != 1)
> + printk_once(KERN_INFO "CPUFREQ: ondemand ignore_nice_load"
> + "sysfs file is deprecated - use nice_max_freq instead");

usage of nice_max_freq should be added to
Documentation/cpu-freq/governors.txt

Also, ignore_nice if being deprecated, add it to
Documentation/feature-removal-schedule.txt

> +
> + if (strict_strtoul(buf, 10, &input) < 0)
> return -EINVAL;
>
> if (input > 1)
> input = 1;
>
> mutex_lock(&dbs_mutex);
> - if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
> + dbs_tuners_ins.ignore_nice = input;
> +
> + for_each_online_cpu(j) {
> + struct cpufreq_policy *policy;
> + struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, j);
> + policy = dbs_info->cur_policy;
> +
> + if (input && policy->min < dbs_tuners_ins.nice_max_freq)
> + dbs_tuners_ins.nice_max_freq = policy->min;
> + else if (!input && policy->max > dbs_tuners_ins.nice_max_freq)
> + dbs_tuners_ins.nice_max_freq = policy->max;
> + }
> + mutex_unlock(&dbs_mutex);
> +
> + return count;
> +}
> +
> +static ssize_t store_nice_max_freq(struct kobject *a, struct attribute *b,
> + const char *buf, size_t count)
> +{
> + unsigned long input;
> + int ret;
> +
> + unsigned int j;
> +
> + if (strict_strtoul(buf, 10, &input) < 0)
> + return -EINVAL;
> +
> + mutex_lock(&dbs_mutex);
> + if (input == dbs_tuners_ins.nice_max_freq) { /* nothing to do */
> mutex_unlock(&dbs_mutex);
> return count;
> }
> - dbs_tuners_ins.ignore_nice = input;
> + dbs_tuners_ins.nice_max_freq = input;
>

Don't we need some sanity checks for user provided value here? Or we are
assuming that underlying freq lookup will do the sane thing when user
gives freq out of range.

> /* we need to re-evaluate prev_cpu_idle */
> for_each_online_cpu(j) {
> @@ -346,9 +381,7 @@ static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
> dbs_info = &per_cpu(od_cpu_dbs_info, j);
> dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
> &dbs_info->prev_cpu_wall);
> - if (dbs_tuners_ins.ignore_nice)
> - dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
> -
> + dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
> }
> mutex_unlock(&dbs_mutex);
>
> @@ -383,6 +416,7 @@ __ATTR(_name, 0644, show_##_name, store_##_name)
> define_one_rw(sampling_rate);
> define_one_rw(up_threshold);
> define_one_rw(ignore_nice_load);
> +define_one_rw(nice_max_freq);
> define_one_rw(powersave_bias);
>
> static struct attribute *dbs_attributes[] = {
> @@ -391,6 +425,7 @@ static struct attribute *dbs_attributes[] = {
> &sampling_rate.attr,
> &up_threshold.attr,
> &ignore_nice_load.attr,
> + &nice_max_freq.attr,
> &powersave_bias.attr,
> NULL
> };
> @@ -457,6 +492,8 @@ static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
> static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
> {
> unsigned int max_load_freq;
> + unsigned int max_ignore_nice_load_freq;
> + unsigned int down_load_freq;
>
> struct cpufreq_policy *policy;
> unsigned int j;
> @@ -477,12 +514,14 @@ static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
> */
>
> /* Get Absolute Load - in terms of freq */
> - max_load_freq = 0;
> + max_load_freq = max_ignore_nice_load_freq = 0;
>
> for_each_cpu(j, policy->cpus) {
> struct cpu_dbs_info_s *j_dbs_info;
> cputime64_t cur_wall_time, cur_idle_time;
> + cputime64_t cur_nice;
> unsigned int idle_time, wall_time;
> + unsigned long cur_nice_jiffies;
> unsigned int load, load_freq;
> int freq_avg;
>
> @@ -498,43 +537,57 @@ static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
> j_dbs_info->prev_cpu_idle);
> j_dbs_info->prev_cpu_idle = cur_idle_time;
>
> - if (dbs_tuners_ins.ignore_nice) {
> - cputime64_t cur_nice;
> - unsigned long cur_nice_jiffies;
> -
> - cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
> - j_dbs_info->prev_cpu_nice);
> - /*
> - * Assumption: nice time between sampling periods will
> - * be less than 2^32 jiffies for 32 bit sys
> - */
> - cur_nice_jiffies = (unsigned long)
> - cputime64_to_jiffies64(cur_nice);
> -
> - j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
> - idle_time += jiffies_to_usecs(cur_nice_jiffies);
> - }
> -
> if (unlikely(!wall_time || wall_time < idle_time))
> continue;
>
> - load = 100 * (wall_time - idle_time) / wall_time;
> -
> freq_avg = __cpufreq_driver_getavg(policy, j);
> if (freq_avg <= 0)
> freq_avg = policy->cur;
>
> + /* Calculate load with with idle */
> + load = 100 * (wall_time - idle_time) / wall_time;
> load_freq = load * freq_avg;
> if (load_freq > max_load_freq)
> max_load_freq = load_freq;
> +
> + cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
> + j_dbs_info->prev_cpu_nice);
> + /*
> + * Assumption: nice time between sampling periods will
> + * be less than 2^32 jiffies for 32 bit sys
> + */
> + cur_nice_jiffies =
> + (unsigned long) cputime64_to_jiffies64(cur_nice);
> +
> + j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
> + idle_time += jiffies_to_usecs(cur_nice_jiffies);
> +
> + if (unlikely(!wall_time || wall_time < idle_time))
> + continue;
> +
> + /* Calculate load with without idle */
> + load = 100 * (wall_time - idle_time) / wall_time;
> + load_freq = load * freq_avg;
> + if (load_freq > max_ignore_nice_load_freq)
> + max_ignore_nice_load_freq = load_freq;
> }
>
> - /* Check for frequency increase */
> - if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
> + /* Check for frequency increase ignoring nice, scale to max */
> + if (max_ignore_nice_load_freq >
> + dbs_tuners_ins.up_threshold * policy->cur) {
> dbs_freq_increase(policy, policy->max);
> return;
> }
>
> + /*
> + * If we failed to increase frequency, check again including nice load.
> + * This time only scale to the specified maximum speed for nice loads.
> + */
> + if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
> + dbs_freq_increase(policy, dbs_tuners_ins.nice_max_freq);
> + return;
> + }
> +
> /* Check for frequency decrease */
> /* if we cannot reduce the frequency anymore, break out early */
> if (policy->cur == policy->min)
> @@ -545,14 +598,31 @@ 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 (max_load_freq <
> - (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
> - policy->cur) {
> + down_load_freq = (dbs_tuners_ins.up_threshold -
> + dbs_tuners_ins.down_differential) * policy->cur;
> +
> + /* First attempt to scale down ignoring low priority threads */
> + if (max_ignore_nice_load_freq < down_load_freq) {
> unsigned int freq_next;
> - freq_next = max_load_freq /
> + freq_next = max_ignore_nice_load_freq /
> + (dbs_tuners_ins.up_threshold -
> + dbs_tuners_ins.down_differential);
> +
> + /*
> + * If freq_next is below nice_max, recalculate frequency
> + * factoring in nice threads. We do not want to cripple
> + * nice threads.
> + */
> + if (freq_next < dbs_tuners_ins.nice_max_freq &&
> + max_load_freq < down_load_freq) {
> + freq_next = max_load_freq /
> (dbs_tuners_ins.up_threshold -
> dbs_tuners_ins.down_differential);
>
> + if (freq_next > dbs_tuners_ins.nice_max_freq)
> + freq_next = dbs_tuners_ins.nice_max_freq;
> + }
> +
> if (!dbs_tuners_ins.powersave_bias) {
> __cpufreq_driver_target(policy, freq_next,
> CPUFREQ_RELATION_L);
> @@ -641,13 +711,13 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
> struct cpu_dbs_info_s *j_dbs_info;
> j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
> j_dbs_info->cur_policy = policy;
> -
> j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
> &j_dbs_info->prev_cpu_wall);
> - if (dbs_tuners_ins.ignore_nice) {
> - j_dbs_info->prev_cpu_nice =
> - kstat_cpu(j).cpustat.nice;
> - }
> + j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
> +
> + /* Take the largest policy->max frequency */
> + if (dbs_tuners_ins.nice_max_freq < policy->max)
> + dbs_tuners_ins.nice_max_freq = policy->max;

I did not follow this change. Is it setting the default nice_max_freq to
policy->max? What happens if user sets nice_max_freq then changes
governor on one CPU to say perf and back to ondemand.

> }
> this_dbs_info->cpu = cpu;
> ondemand_powersave_bias_init_cpu(cpu);


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: Andrew Morton on
On Mon, 25 Jan 2010 17:47:15 -0800
Mike Chan <mike(a)android.com> wrote:

> Allow lower priority threads to scale frequency to specified nice_max_freq.
> This allows low priority threads to operate at the most efficient
> power/performance frequency.
>
> Often the highest and lowest cpu speeds do not provide the the optimal
> performance/power ratios. Latency requirements for normal and high priority
> threads require the maximum speed that are not always optimal power wise
> inorder to satisfy the requirements.
>
> To enable set nice_max_freq (to a speed lower than the scaling_max_freq).
>
> The governor will first attempt to scale the cpu to policy->max (default)
> only using normal and high priority threads. It will ignore nice threads.
> If the load is high enough without nice threads then ondemand will scale to
> the max speed and exit.
>
> If load for normal and high priority threads are not high enough to increase
> the cpu speed, check again including the load from nice threads. Only scale
> to the nice_max_freq specified.
>
> Previous behavior is maintained by setting the values below:
>
> + When nice_max_freq is set to 0, behavior is the current default
> (nice is counted for load).
>
> + When nice_max_freq is set to scaling_min_freq, the behavior is the same
> as the original ignore_nice_load == 1. Which counts all nice threads as
> idle time when computing cpu load.
>
> *** v2 ***
> + The ignore_nice_load sysfs still behaves the same as before (0/1) and is
> kept around for legacy support. Userspace scripts should now use
> nice_max_freq.

The patch conflicts a bit with a change which is pending in linux-next:

--- linux-2.6.33-rc5/drivers/cpufreq/cpufreq_ondemand.c 2009-12-03 12:12:09.000000000 -0800
+++ 25/drivers/cpufreq/cpufreq_ondemand.c 2010-01-27 16:11:18.000000000 -0800
@@ -554,6 +554,9 @@ static void dbs_check_cpu(struct cpu_dbs
(dbs_tuners_ins.up_threshold -
dbs_tuners_ins.down_differential);

+ if (freq_next < policy->min)
+ freq_next = policy->min;
+
if (!dbs_tuners_ins.powersave_bias) {
__cpufreq_driver_target(policy, freq_next,
CPUFREQ_RELATION_L);

You might want to check that - there might be functional interactions.

> index 3dcf126..2a5a414 100644
> --- a/drivers/cpufreq/cpufreq_ondemand.c
> +++ b/drivers/cpufreq/cpufreq_ondemand.c
> @@ -108,11 +108,13 @@ static struct dbs_tuners {
> unsigned int down_differential;
> unsigned int ignore_nice;
> unsigned int powersave_bias;
> + unsigned int nice_max_freq;
> } dbs_tuners_ins = {
> .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
> .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
> .ignore_nice = 0,
> .powersave_bias = 0,
> + .nice_max_freq = 0,
> };

The initialisation to zero is unneeded and unidiomatic. It'd be better
to remove the other two.

> static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
> @@ -251,6 +253,7 @@ static ssize_t show_##file_name \
> show_one(sampling_rate, sampling_rate);
> show_one(up_threshold, up_threshold);
> show_one(ignore_nice_load, ignore_nice);
> +show_one(nice_max_freq, nice_max_freq);
> show_one(powersave_bias, powersave_bias);
>
> /*** delete after deprecation time ***/
> @@ -318,10 +321,48 @@ static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
> return count;
> }
>
> +/*
> + * Preserve ignore_nice_load behavior, if enabled do not allow low priority
> + * threads to scale beyond the minimum frequency.
> + */
> static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
> const char *buf, size_t count)
> {
> unsigned int input;
> + unsigned int j;
> +
> + printk_once(KERN_INFO "CPUFREQ: ondemand ignore_nice_load"
> + "sysfs file is deprecated - use nice_max_freq instead");

This printk will come out wrong: "ondemand ignore_nice_loadsysfs file"

> + if (sscanf(buf, "%u", &input) != 1)
> + return -EINVAL;

This will treat input of the form "42foo" as a valid number, which is
sloppy. Use strict_strtoul() to fix.

> + if (input > 1)
> + input = 1;

So inputs which aren't 0 or 1 are invalid. It'd be better to fail,
rather than to silently modify-and-accept?

> + mutex_lock(&dbs_mutex);
> + dbs_tuners_ins.ignore_nice = input;
> +
> + for_each_online_cpu(j) {
> + struct cpufreq_policy *policy;
> + struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, j);
> + policy = dbs_info->cur_policy;
> +
> +
> + if (input && policy->min < dbs_tuners_ins.nice_max_freq)
> + dbs_tuners_ins.nice_max_freq = policy->min;
> + else if (!input && policy->max > dbs_tuners_ins.nice_max_freq)
> + dbs_tuners_ins.nice_max_freq = policy->max;
> + }
> + mutex_unlock(&dbs_mutex);

What prevents a CPU from going offline while this loop is executing?

> + return count;
> +}
> +
>
> ...
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/