[PATCH] cpufreq_conservative: make for_each_cpu() safe
[powerpc.git] / drivers / cpufreq / cpufreq_conservative.c
index 39543a2..7498f25 100644 (file)
 #include <linux/jiffies.h>
 #include <linux/kernel_stat.h>
 #include <linux/percpu.h>
-
+#include <linux/mutex.h>
 /*
  * dbs is used in this file as a shortform for demandbased switching
  * It helps to keep variable names smaller, simpler
  */
 
 #define DEF_FREQUENCY_UP_THRESHOLD             (80)
-#define MIN_FREQUENCY_UP_THRESHOLD             (0)
-#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 
  * 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   (100000)
-#define DEF_SAMPLING_DOWN_FACTOR               (5)
+#define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER   (1000)
+#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);
@@ -71,7 +70,7 @@ static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
 
 static unsigned int dbs_enable;        /* number of CPUs using this policy */
 
-static DECLARE_MUTEX   (dbs_sem);
+static DEFINE_MUTEX    (dbs_mutex);
 static DECLARE_WORK    (dbs_work, do_dbs_timer, NULL);
 
 struct dbs_tuners {
@@ -136,12 +135,12 @@ static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
        unsigned int input;
        int ret;
        ret = sscanf (buf, "%u", &input);
-       if (ret != 1 )
+       if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
                return -EINVAL;
 
-       down(&dbs_sem);
+       mutex_lock(&dbs_mutex);
        dbs_tuners_ins.sampling_down_factor = input;
-       up(&dbs_sem);
+       mutex_unlock(&dbs_mutex);
 
        return count;
 }
@@ -153,14 +152,14 @@ static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
        int ret;
        ret = sscanf (buf, "%u", &input);
 
-       down(&dbs_sem);
+       mutex_lock(&dbs_mutex);
        if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
-               up(&dbs_sem);
+               mutex_unlock(&dbs_mutex);
                return -EINVAL;
        }
 
        dbs_tuners_ins.sampling_rate = input;
-       up(&dbs_sem);
+       mutex_unlock(&dbs_mutex);
 
        return count;
 }
@@ -172,16 +171,15 @@ static ssize_t store_up_threshold(struct cpufreq_policy *unused,
        int ret;
        ret = sscanf (buf, "%u", &input);
 
-       down(&dbs_sem);
-       if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || 
-                       input < MIN_FREQUENCY_UP_THRESHOLD ||
+       mutex_lock(&dbs_mutex);
+       if (ret != 1 || input > 100 || input < 0 ||
                        input <= dbs_tuners_ins.down_threshold) {
-               up(&dbs_sem);
+               mutex_unlock(&dbs_mutex);
                return -EINVAL;
        }
 
        dbs_tuners_ins.up_threshold = input;
-       up(&dbs_sem);
+       mutex_unlock(&dbs_mutex);
 
        return count;
 }
@@ -193,16 +191,15 @@ static ssize_t store_down_threshold(struct cpufreq_policy *unused,
        int ret;
        ret = sscanf (buf, "%u", &input);
 
-       down(&dbs_sem);
-       if (ret != 1 || input > MAX_FREQUENCY_DOWN_THRESHOLD || 
-                       input < MIN_FREQUENCY_DOWN_THRESHOLD ||
+       mutex_lock(&dbs_mutex);
+       if (ret != 1 || input > 100 || input < 0 ||
                        input >= dbs_tuners_ins.up_threshold) {
-               up(&dbs_sem);
+               mutex_unlock(&dbs_mutex);
                return -EINVAL;
        }
 
        dbs_tuners_ins.down_threshold = input;
-       up(&dbs_sem);
+       mutex_unlock(&dbs_mutex);
 
        return count;
 }
@@ -222,9 +219,9 @@ static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
        if ( input > 1 )
                input = 1;
        
-       down(&dbs_sem);
+       mutex_lock(&dbs_mutex);
        if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
-               up(&dbs_sem);
+               mutex_unlock(&dbs_mutex);
                return count;
        }
        dbs_tuners_ins.ignore_nice = input;
@@ -236,7 +233,7 @@ static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
                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);
+       mutex_unlock(&dbs_mutex);
 
        return count;
 }
@@ -257,9 +254,9 @@ static ssize_t store_freq_step(struct cpufreq_policy *policy,
        
        /* no need to test here if freq_step is zero as the user might actually
         * want this, they would be crazy though :) */
-       down(&dbs_sem);
+       mutex_lock(&dbs_mutex);
        dbs_tuners_ins.freq_step = input;
-       up(&dbs_sem);
+       mutex_unlock(&dbs_mutex);
 
        return count;
 }
@@ -297,31 +294,40 @@ 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 tmp_idle_ticks, total_idle_ticks;
        unsigned int freq_step;
        unsigned int freq_down_sampling_rate;
-       static int down_skip[NR_CPUS];
-       static int requested_freq[NR_CPUS];
-       static unsigned short init_flag = 0;
-       struct cpu_dbs_info_s *this_dbs_info;
-       struct cpu_dbs_info_s *dbs_info;
-
+       static unsigned short down_skip[NR_CPUS];
+       static unsigned int requested_freq[NR_CPUS];
+       static unsigned int init_flag = NR_CPUS;
+       struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
        struct cpufreq_policy *policy;
-       unsigned int j;
 
-       this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
        if (!this_dbs_info->enable)
                return;
 
-       policy = this_dbs_info->cur_policy;
-
-       if ( init_flag == 0 ) {
-               for_each_online_cpu(j) {
-                       dbs_info = &per_cpu(cpu_dbs_info, j);
-                       requested_freq[j] = dbs_info->cur_policy->cur;
+       if ( init_flag != 0 ) {
+               for_each_cpu(init_flag) {
+                       down_skip[init_flag] = 0;
+                       /* I doubt a CPU exists with a freq of 0hz :) */
+                       requested_freq[init_flag] = 0;
                }
-               init_flag = 1;
+               init_flag = 0;
        }
        
+       /*
+        * If its a freshly initialised cpu we setup requested_freq.  This
+        * check could be avoided if we did not care about a first time
+        * stunted increase in CPU speed when there is a load.  I feel we
+        * should be initialising this to something.  The removal of a CPU
+        * is not a problem, after a short time the CPU should settle down
+        * to a 'natural' frequency.
+        */
+       if (requested_freq[cpu] == 0)
+               requested_freq[cpu] = this_dbs_info->cur_policy->cur;
+
+       policy = this_dbs_info->cur_policy;
+
        /* 
         * The default safe range is 20% to 80% 
         * Every sampling_rate, we check
@@ -337,37 +343,27 @@ static void dbs_check_cpu(int cpu)
         */
 
        /* Check for frequency increase */
-
        idle_ticks = UINT_MAX;
-       for_each_cpu_mask(j, policy->cpus) {
-               unsigned int tmp_idle_ticks, total_idle_ticks;
-               struct cpu_dbs_info_s *j_dbs_info;
 
-               j_dbs_info = &per_cpu(cpu_dbs_info, j);
-               /* Check for frequency increase */
-               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;
-
-               if (tmp_idle_ticks < idle_ticks)
-                       idle_ticks = tmp_idle_ticks;
-       }
+       /* Check for frequency increase */
+       total_idle_ticks = get_cpu_idle_time(cpu);
+       tmp_idle_ticks = total_idle_ticks -
+               this_dbs_info->prev_cpu_idle_up;
+       this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
+
+       if (tmp_idle_ticks < idle_ticks)
+               idle_ticks = tmp_idle_ticks;
 
        /* Scale idle ticks by 100 and compare with up and down ticks */
        idle_ticks *= 100;
        up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
-               usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
+                       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;
+               this_dbs_info->prev_cpu_idle_down =
+                       this_dbs_info->prev_cpu_idle_up;
 
-                       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 (requested_freq[cpu] == policy->max)
                        return;
@@ -392,20 +388,14 @@ static void dbs_check_cpu(int cpu)
        if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
                return;
 
-       idle_ticks = UINT_MAX;
-       for_each_cpu_mask(j, policy->cpus) {
-               unsigned int tmp_idle_ticks, total_idle_ticks;
-               struct cpu_dbs_info_s *j_dbs_info;
-
-               j_dbs_info = &per_cpu(cpu_dbs_info, j);
-               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;
+       /* Check for frequency decrease */
+       total_idle_ticks = this_dbs_info->prev_cpu_idle_up;
+       tmp_idle_ticks = total_idle_ticks -
+               this_dbs_info->prev_cpu_idle_down;
+       this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
 
-               if (tmp_idle_ticks < idle_ticks)
-                       idle_ticks = tmp_idle_ticks;
-       }
+       if (tmp_idle_ticks < idle_ticks)
+               idle_ticks = tmp_idle_ticks;
 
        /* Scale idle ticks by 100 and compare with up and down ticks */
        idle_ticks *= 100;
@@ -414,12 +404,14 @@ static void dbs_check_cpu(int cpu)
        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);
+               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 we are already at the lowest speed then break out early
                 * or if we 'cannot' reduce the speed as the user might want
-                * freq_step to be zero */
+                * freq_step to be zero
+                */
                if (requested_freq[cpu] == policy->min
                                || dbs_tuners_ins.freq_step == 0)
                        return;
@@ -434,9 +426,8 @@ static void dbs_check_cpu(int cpu)
                if (requested_freq[cpu] < policy->min)
                        requested_freq[cpu] = policy->min;
 
-               __cpufreq_driver_target(policy,
-                       requested_freq[cpu],
-                       CPUFREQ_RELATION_H);
+               __cpufreq_driver_target(policy, requested_freq[cpu],
+                               CPUFREQ_RELATION_H);
                return;
        }
 }
@@ -444,12 +435,12 @@ static void dbs_check_cpu(int cpu)
 static void do_dbs_timer(void *data)
 { 
        int i;
-       down(&dbs_sem);
+       mutex_lock(&dbs_mutex);
        for_each_online_cpu(i)
                dbs_check_cpu(i);
        schedule_delayed_work(&dbs_work, 
                        usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
-       up(&dbs_sem);
+       mutex_unlock(&dbs_mutex);
 } 
 
 static inline void dbs_timer_init(void)
@@ -487,13 +478,13 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
                if (this_dbs_info->enable) /* Already enabled */
                        break;
                 
-               down(&dbs_sem);
+               mutex_lock(&dbs_mutex);
                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->cur_policy = policy;
                
-                       j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
+                       j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(cpu);
                        j_dbs_info->prev_cpu_idle_down
                                = j_dbs_info->prev_cpu_idle_up;
                }
@@ -507,13 +498,16 @@ 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 = 10 * 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;
                        dbs_tuners_ins.freq_step = 5;
@@ -521,11 +515,11 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
                        dbs_timer_init();
                }
                
-               up(&dbs_sem);
+               mutex_unlock(&dbs_mutex);
                break;
 
        case CPUFREQ_GOV_STOP:
-               down(&dbs_sem);
+               mutex_lock(&dbs_mutex);
                this_dbs_info->enable = 0;
                sysfs_remove_group(&policy->kobj, &dbs_attr_group);
                dbs_enable--;
@@ -536,12 +530,12 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
                if (dbs_enable == 0) 
                        dbs_timer_exit();
                
-               up(&dbs_sem);
+               mutex_unlock(&dbs_mutex);
 
                break;
 
        case CPUFREQ_GOV_LIMITS:
-               down(&dbs_sem);
+               mutex_lock(&dbs_mutex);
                if (policy->max < this_dbs_info->cur_policy->cur)
                        __cpufreq_driver_target(
                                        this_dbs_info->cur_policy,
@@ -550,7 +544,7 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
                        __cpufreq_driver_target(
                                        this_dbs_info->cur_policy,
                                        policy->min, CPUFREQ_RELATION_L);
-               up(&dbs_sem);
+               mutex_unlock(&dbs_mutex);
                break;
        }
        return 0;