]> pilppa.org Git - linux-2.6-omap-h63xx.git/blobdiff - drivers/cpufreq/cpufreq_ondemand.c
Merge git://git.kernel.org/pub/scm/linux/kernel/git/sam/kbuild
[linux-2.6-omap-h63xx.git] / drivers / cpufreq / cpufreq_ondemand.c
index 6dc83808e5901102dae1d7aed2bb7111504a8120..e69fd8dd1f1cb7bf6042a878b2545a714c7df133 100644 (file)
  */
 
 #define DEF_FREQUENCY_UP_THRESHOLD             (80)
-#define MIN_FREQUENCY_UP_THRESHOLD             (0)
+#define MIN_FREQUENCY_UP_THRESHOLD             (11)
 #define MAX_FREQUENCY_UP_THRESHOLD             (100)
 
-#define DEF_FREQUENCY_DOWN_THRESHOLD           (20)
-#define MIN_FREQUENCY_DOWN_THRESHOLD           (0)
-#define MAX_FREQUENCY_DOWN_THRESHOLD           (100)
-
 /* 
  * The polling frequency of this governor depends on the capability of 
  * the processor. Default polling frequency is 1000 times the transition
  * All times here are in uS.
  */
 static unsigned int                            def_sampling_rate;
-#define MIN_SAMPLING_RATE                      (def_sampling_rate / 2)
+#define MIN_SAMPLING_RATE_RATIO                        (2)
+/* for correct statistics, we need at least 10 ticks between each measure */
+#define MIN_STAT_SAMPLING_RATE                 (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
+#define MIN_SAMPLING_RATE                      (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
 #define MAX_SAMPLING_RATE                      (500 * def_sampling_rate)
 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER   (1000)
-#define DEF_SAMPLING_DOWN_FACTOR               (10)
+#define DEF_SAMPLING_DOWN_FACTOR               (1)
+#define MAX_SAMPLING_DOWN_FACTOR               (10)
 #define TRANSITION_LATENCY_LIMIT               (10 * 1000)
 
 static void do_dbs_timer(void *data);
@@ -77,16 +77,23 @@ struct dbs_tuners {
        unsigned int            sampling_rate;
        unsigned int            sampling_down_factor;
        unsigned int            up_threshold;
-       unsigned int            down_threshold;
        unsigned int            ignore_nice;
 };
 
 static struct dbs_tuners dbs_tuners_ins = {
        .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
-       .down_threshold         = DEF_FREQUENCY_DOWN_THRESHOLD,
        .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
 };
 
+static inline unsigned int get_cpu_idle_time(unsigned int cpu)
+{
+       return  kstat_cpu(cpu).cpustat.idle +
+               kstat_cpu(cpu).cpustat.iowait +
+               ( dbs_tuners_ins.ignore_nice ?
+                 kstat_cpu(cpu).cpustat.nice :
+                 0);
+}
+
 /************************** sysfs interface ************************/
 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
 {
@@ -115,8 +122,7 @@ static ssize_t show_##file_name                                             \
 show_one(sampling_rate, sampling_rate);
 show_one(sampling_down_factor, sampling_down_factor);
 show_one(up_threshold, up_threshold);
-show_one(down_threshold, down_threshold);
-show_one(ignore_nice, ignore_nice);
+show_one(ignore_nice_load, ignore_nice);
 
 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, 
                const char *buf, size_t count)
@@ -127,6 +133,9 @@ static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
        if (ret != 1 )
                return -EINVAL;
 
+       if (input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
+               return -EINVAL;
+
        down(&dbs_sem);
        dbs_tuners_ins.sampling_down_factor = input;
        up(&dbs_sem);
@@ -162,8 +171,7 @@ static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 
        down(&dbs_sem);
        if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || 
-                       input < MIN_FREQUENCY_UP_THRESHOLD ||
-                       input <= dbs_tuners_ins.down_threshold) {
+                       input < MIN_FREQUENCY_UP_THRESHOLD) {
                up(&dbs_sem);
                return -EINVAL;
        }
@@ -174,28 +182,7 @@ static ssize_t store_up_threshold(struct cpufreq_policy *unused,
        return count;
 }
 
-static ssize_t store_down_threshold(struct cpufreq_policy *unused, 
-               const char *buf, size_t count)
-{
-       unsigned int input;
-       int ret;
-       ret = sscanf (buf, "%u", &input);
-
-       down(&dbs_sem);
-       if (ret != 1 || input > MAX_FREQUENCY_DOWN_THRESHOLD || 
-                       input < MIN_FREQUENCY_DOWN_THRESHOLD ||
-                       input >= dbs_tuners_ins.up_threshold) {
-               up(&dbs_sem);
-               return -EINVAL;
-       }
-
-       dbs_tuners_ins.down_threshold = input;
-       up(&dbs_sem);
-
-       return count;
-}
-
-static ssize_t store_ignore_nice(struct cpufreq_policy *policy,
+static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
                const char *buf, size_t count)
 {
        unsigned int input;
@@ -218,16 +205,10 @@ static ssize_t store_ignore_nice(struct cpufreq_policy *policy,
        dbs_tuners_ins.ignore_nice = input;
 
        /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
-       for_each_cpu_mask(j, policy->cpus) {
+       for_each_online_cpu(j) {
                struct cpu_dbs_info_s *j_dbs_info;
                j_dbs_info = &per_cpu(cpu_dbs_info, j);
-               j_dbs_info->cur_policy = policy;
-
-               j_dbs_info->prev_cpu_idle_up =
-                       kstat_cpu(j).cpustat.idle +
-                       kstat_cpu(j).cpustat.iowait +
-                       ( !dbs_tuners_ins.ignore_nice
-                         ? kstat_cpu(j).cpustat.nice : 0 );
+               j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
                j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
        }
        up(&dbs_sem);
@@ -242,8 +223,7 @@ __ATTR(_name, 0644, show_##_name, store_##_name)
 define_one_rw(sampling_rate);
 define_one_rw(sampling_down_factor);
 define_one_rw(up_threshold);
-define_one_rw(down_threshold);
-define_one_rw(ignore_nice);
+define_one_rw(ignore_nice_load);
 
 static struct attribute * dbs_attributes[] = {
        &sampling_rate_max.attr,
@@ -251,8 +231,7 @@ static struct attribute * dbs_attributes[] = {
        &sampling_rate.attr,
        &sampling_down_factor.attr,
        &up_threshold.attr,
-       &down_threshold.attr,
-       &ignore_nice.attr,
+       &ignore_nice_load.attr,
        NULL
 };
 
@@ -265,9 +244,8 @@ static struct attribute_group dbs_attr_group = {
 
 static void dbs_check_cpu(int cpu)
 {
-       unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
-       unsigned int total_idle_ticks;
-       unsigned int freq_down_step;
+       unsigned int idle_ticks, up_idle_ticks, total_ticks;
+       unsigned int freq_next;
        unsigned int freq_down_sampling_rate;
        static int down_skip[NR_CPUS];
        struct cpu_dbs_info_s *this_dbs_info;
@@ -281,44 +259,25 @@ static void dbs_check_cpu(int cpu)
 
        policy = this_dbs_info->cur_policy;
        /* 
-        * The default safe range is 20% to 80% 
-        * Every sampling_rate, we check
-        *      - If current idle time is less than 20%, then we try to 
-        *        increase frequency
-        * Every sampling_rate*sampling_down_factor, we check
-        *      - If current idle time is more than 80%, then we try to
-        *        decrease frequency
+        * Every sampling_rate, we check, if current idle time is less
+        * than 20% (default), then we try to increase frequency
+        * Every sampling_rate*sampling_down_factor, we look for a the lowest
+        * frequency which can sustain the load while keeping idle time over
+        * 30%. If such a frequency exist, we try to decrease to this frequency.
         *
         * Any frequency increase takes it to the maximum frequency. 
         * Frequency reduction happens at minimum steps of 
-        * 5% of max_frequency 
+        * 5% (default) of current frequency 
         */
 
        /* Check for frequency increase */
-       total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
-               kstat_cpu(cpu).cpustat.iowait;
-         /* consider 'nice' tasks as 'idle' time too if required */
-         if (dbs_tuners_ins.ignore_nice == 0)
-               total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
-       idle_ticks = total_idle_ticks -
-               this_dbs_info->prev_cpu_idle_up;
-       this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
-       
-
+       idle_ticks = UINT_MAX;
        for_each_cpu_mask(j, policy->cpus) {
-               unsigned int tmp_idle_ticks;
+               unsigned int tmp_idle_ticks, total_idle_ticks;
                struct cpu_dbs_info_s *j_dbs_info;
 
-               if (j == cpu)
-                       continue;
-
                j_dbs_info = &per_cpu(cpu_dbs_info, j);
-               /* Check for frequency increase */
-               total_idle_ticks = kstat_cpu(j).cpustat.idle +
-                       kstat_cpu(j).cpustat.iowait;
-                 /* consider 'nice' too? */
-                 if (dbs_tuners_ins.ignore_nice == 0)
-                          total_idle_ticks += kstat_cpu(j).cpustat.nice;
+               total_idle_ticks = get_cpu_idle_time(j);
                tmp_idle_ticks = total_idle_ticks -
                        j_dbs_info->prev_cpu_idle_up;
                j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
@@ -333,14 +292,20 @@ static void dbs_check_cpu(int cpu)
                        usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 
        if (idle_ticks < up_idle_ticks) {
+               down_skip[cpu] = 0;
+               for_each_cpu_mask(j, policy->cpus) {
+                       struct cpu_dbs_info_s *j_dbs_info;
+
+                       j_dbs_info = &per_cpu(cpu_dbs_info, j);
+                       j_dbs_info->prev_cpu_idle_down = 
+                                       j_dbs_info->prev_cpu_idle_up;
+               }
                /* if we are already at full speed then break out early */
                if (policy->cur == policy->max)
                        return;
                
                __cpufreq_driver_target(policy, policy->max, 
                        CPUFREQ_RELATION_H);
-               down_skip[cpu] = 0;
-               this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
                return;
        }
 
@@ -349,29 +314,14 @@ static void dbs_check_cpu(int cpu)
        if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
                return;
 
-       total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
-               kstat_cpu(cpu).cpustat.iowait;
-         /* consider 'nice' too? */
-         if (dbs_tuners_ins.ignore_nice == 0)
-                 total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
-       idle_ticks = total_idle_ticks -
-               this_dbs_info->prev_cpu_idle_down;
-       this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
-
+       idle_ticks = UINT_MAX;
        for_each_cpu_mask(j, policy->cpus) {
-               unsigned int tmp_idle_ticks;
+               unsigned int tmp_idle_ticks, total_idle_ticks;
                struct cpu_dbs_info_s *j_dbs_info;
 
-               if (j == cpu)
-                       continue;
-
                j_dbs_info = &per_cpu(cpu_dbs_info, j);
-               /* Check for frequency increase */
-               total_idle_ticks = kstat_cpu(j).cpustat.idle +
-                       kstat_cpu(j).cpustat.iowait;
-                 /* consider 'nice' too? */
-                 if (dbs_tuners_ins.ignore_nice == 0)
-                       total_idle_ticks += kstat_cpu(j).cpustat.nice;
+               /* Check for frequency decrease */
+               total_idle_ticks = j_dbs_info->prev_cpu_idle_up;
                tmp_idle_ticks = total_idle_ticks -
                        j_dbs_info->prev_cpu_idle_down;
                j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
@@ -380,31 +330,27 @@ static void dbs_check_cpu(int cpu)
                        idle_ticks = tmp_idle_ticks;
        }
 
-       /* Scale idle ticks by 100 and compare with up and down ticks */
-       idle_ticks *= 100;
        down_skip[cpu] = 0;
+       /* if we cannot reduce the frequency anymore, break out early */
+       if (policy->cur == policy->min)
+               return;
 
+       /* Compute how many ticks there are between two measurements */
        freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
                dbs_tuners_ins.sampling_down_factor;
-       down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
-                       usecs_to_jiffies(freq_down_sampling_rate);
-
-       if (idle_ticks > down_idle_ticks ) {
-               /* if we are already at the lowest speed then break out early */
-               if (policy->cur == policy->min)
-                       return;
-               
-               freq_down_step = (5 * policy->max) / 100;
+       total_ticks = usecs_to_jiffies(freq_down_sampling_rate);
 
-               /* max freq cannot be less than 100. But who knows.... */
-               if (unlikely(freq_down_step == 0))
-                       freq_down_step = 5;
+       /*
+        * The optimal frequency is the frequency that is the lowest that
+        * can support the current CPU usage without triggering the up
+        * policy. To be safe, we focus 10 points under the threshold.
+        */
+       freq_next = ((total_ticks - idle_ticks) * 100) / total_ticks;
+       freq_next = (freq_next * policy->cur) / 
+                       (dbs_tuners_ins.up_threshold - 10);
 
-               __cpufreq_driver_target(policy,
-                       policy->cur - freq_down_step, 
-                       CPUFREQ_RELATION_H);
-               return;
-       }
+       if (freq_next <= ((policy->cur * 95) / 100))
+               __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_L);
 }
 
 static void do_dbs_timer(void *data)
@@ -459,11 +405,7 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
                        j_dbs_info = &per_cpu(cpu_dbs_info, j);
                        j_dbs_info->cur_policy = policy;
                
-                       j_dbs_info->prev_cpu_idle_up = 
-                               kstat_cpu(j).cpustat.idle +
-                               kstat_cpu(j).cpustat.iowait +
-                               ( !dbs_tuners_ins.ignore_nice
-                                 ? kstat_cpu(j).cpustat.nice : 0 );
+                       j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
                        j_dbs_info->prev_cpu_idle_down
                                = j_dbs_info->prev_cpu_idle_up;
                }
@@ -477,13 +419,16 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
                if (dbs_enable == 1) {
                        unsigned int latency;
                        /* policy latency is in nS. Convert it to uS first */
+                       latency = policy->cpuinfo.transition_latency / 1000;
+                       if (latency == 0)
+                               latency = 1;
 
-                       latency = policy->cpuinfo.transition_latency;
-                       if (latency < 1000)
-                               latency = 1000;
-
-                       def_sampling_rate = (latency / 1000) *
+                       def_sampling_rate = latency *
                                        DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
+
+                       if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
+                               def_sampling_rate = MIN_STAT_SAMPLING_RATE;
+
                        dbs_tuners_ins.sampling_rate = def_sampling_rate;
                        dbs_tuners_ins.ignore_nice = 0;