Merge branch 'fix/misc' into for-linus
[pandora-kernel.git] / drivers / cpufreq / cpufreq_ondemand.c
index e741c33..d6ba142 100644 (file)
@@ -32,6 +32,7 @@
 #define DEF_FREQUENCY_UP_THRESHOLD             (80)
 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL      (3)
 #define MICRO_FREQUENCY_UP_THRESHOLD           (95)
+#define MICRO_FREQUENCY_MIN_SAMPLE_RATE                (10000)
 #define MIN_FREQUENCY_UP_THRESHOLD             (11)
 #define MAX_FREQUENCY_UP_THRESHOLD             (100)
 
  * this governor will not work.
  * All times here are in uS.
  */
-static unsigned int def_sampling_rate;
 #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)
-/* Above MIN_SAMPLING_RATE will vanish with its sysfs file soon
- * Define the minimal settable sampling rate to the greater of:
- *   - "HW transition latency" * 100 (same as default sampling / 10)
- *   - MIN_STAT_SAMPLING_RATE
- * To avoid that userspace shoots itself.
-*/
-static unsigned int minimum_sampling_rate(void)
-{
-       return max(def_sampling_rate / 10, MIN_STAT_SAMPLING_RATE);
-}
 
-/* This will also vanish soon with removing sampling_rate_max */
-#define MAX_SAMPLING_RATE                      (500 * def_sampling_rate)
+static unsigned int min_sampling_rate;
+
 #define LATENCY_MULTIPLIER                     (1000)
+#define MIN_LATENCY_MULTIPLIER                 (100)
 #define TRANSITION_LATENCY_LIMIT               (10 * 1000 * 1000)
 
 static void do_dbs_timer(struct work_struct *work);
@@ -84,23 +70,21 @@ struct cpu_dbs_info_s {
        unsigned int freq_lo_jiffies;
        unsigned int freq_hi_jiffies;
        int cpu;
-       unsigned int enable:1,
-               sample_type:1;
+       unsigned int sample_type:1;
+       /*
+        * percpu mutex that serializes governor limit change with
+        * do_dbs_timer invocation. We do not want do_dbs_timer to run
+        * when user is changing the governor or limits.
+        */
+       struct mutex timer_mutex;
 };
 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
 
 static unsigned int dbs_enable;        /* number of CPUs using this policy */
 
 /*
- * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
- * lock and dbs_mutex. cpu_hotplug lock should always be held before
- * dbs_mutex. If any function that can potentially take cpu_hotplug lock
- * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
- * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
- * is recursive for the same process. -Venki
- * DEADLOCK ALERT! (2) : do_dbs_timer() must not take the dbs_mutex, because it
- * would deadlock with cancel_delayed_work_sync(), which is needed for proper
- * raceless workqueue teardown.
+ * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
+ * different CPUs. It protects dbs_enable in governor start/stop.
  */
 static DEFINE_MUTEX(dbs_mutex);
 
@@ -206,41 +190,32 @@ static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
        return freq_hi;
 }
 
+static void ondemand_powersave_bias_init_cpu(int cpu)
+{
+       struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu);
+       dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
+       dbs_info->freq_lo = 0;
+}
+
 static void ondemand_powersave_bias_init(void)
 {
        int i;
        for_each_online_cpu(i) {
-               struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
-               dbs_info->freq_table = cpufreq_frequency_get_table(i);
-               dbs_info->freq_lo = 0;
+               ondemand_powersave_bias_init_cpu(i);
        }
 }
 
 /************************** sysfs interface ************************/
 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
 {
-       static int print_once;
-
-       if (!print_once) {
-               printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
-                      "sysfs file is deprecated - used by: %s\n",
-                      current->comm);
-               print_once = 1;
-       }
-       return sprintf(buf, "%u\n", MAX_SAMPLING_RATE);
+       printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
+              "sysfs file is deprecated - used by: %s\n", current->comm);
+       return sprintf(buf, "%u\n", -1U);
 }
 
 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
 {
-       static int print_once;
-
-       if (!print_once) {
-               printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_min "
-                      "sysfs file is deprecated - used by: %s\n",
-                      current->comm);
-               print_once = 1;
-       }
-       return sprintf(buf, "%u\n", MIN_SAMPLING_RATE);
+       return sprintf(buf, "%u\n", min_sampling_rate);
 }
 
 #define define_one_ro(_name)           \
@@ -268,13 +243,11 @@ static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
        unsigned int input;
        int ret;
        ret = sscanf(buf, "%u", &input);
+       if (ret != 1)
+               return -EINVAL;
 
        mutex_lock(&dbs_mutex);
-       if (ret != 1) {
-               mutex_unlock(&dbs_mutex);
-               return -EINVAL;
-       }
-       dbs_tuners_ins.sampling_rate = max(input, minimum_sampling_rate());
+       dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
        mutex_unlock(&dbs_mutex);
 
        return count;
@@ -287,13 +260,12 @@ static ssize_t store_up_threshold(struct cpufreq_policy *unused,
        int ret;
        ret = sscanf(buf, "%u", &input);
 
-       mutex_lock(&dbs_mutex);
        if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
                        input < MIN_FREQUENCY_UP_THRESHOLD) {
-               mutex_unlock(&dbs_mutex);
                return -EINVAL;
        }
 
+       mutex_lock(&dbs_mutex);
        dbs_tuners_ins.up_threshold = input;
        mutex_unlock(&dbs_mutex);
 
@@ -391,9 +363,6 @@ static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
        struct cpufreq_policy *policy;
        unsigned int j;
 
-       if (!this_dbs_info->enable)
-               return;
-
        this_dbs_info->freq_lo = 0;
        policy = this_dbs_info->cur_policy;
 
@@ -521,14 +490,7 @@ static void do_dbs_timer(struct work_struct *work)
        int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 
        delay -= jiffies % delay;
-
-       if (lock_policy_rwsem_write(cpu) < 0)
-               return;
-
-       if (!dbs_info->enable) {
-               unlock_policy_rwsem_write(cpu);
-               return;
-       }
+       mutex_lock(&dbs_info->timer_mutex);
 
        /* Common NORMAL_SAMPLE setup */
        dbs_info->sample_type = DBS_NORMAL_SAMPLE;
@@ -545,7 +507,7 @@ static void do_dbs_timer(struct work_struct *work)
                        dbs_info->freq_lo, CPUFREQ_RELATION_H);
        }
        queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
-       unlock_policy_rwsem_write(cpu);
+       mutex_unlock(&dbs_info->timer_mutex);
 }
 
 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
@@ -554,8 +516,6 @@ static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
        int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
        delay -= jiffies % delay;
 
-       dbs_info->enable = 1;
-       ondemand_powersave_bias_init();
        dbs_info->sample_type = DBS_NORMAL_SAMPLE;
        INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
        queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
@@ -564,7 +524,6 @@ static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 
 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 {
-       dbs_info->enable = 0;
        cancel_delayed_work_sync(&dbs_info->work);
 }
 
@@ -583,19 +542,15 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
                if ((!cpu_online(cpu)) || (!policy->cur))
                        return -EINVAL;
 
-               if (this_dbs_info->enable) /* Already enabled */
-                       break;
-
                mutex_lock(&dbs_mutex);
-               dbs_enable++;
 
                rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
                if (rc) {
-                       dbs_enable--;
                        mutex_unlock(&dbs_mutex);
                        return rc;
                }
 
+               dbs_enable++;
                for_each_cpu(j, policy->cpus) {
                        struct cpu_dbs_info_s *j_dbs_info;
                        j_dbs_info = &per_cpu(cpu_dbs_info, j);
@@ -609,6 +564,8 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
                        }
                }
                this_dbs_info->cpu = cpu;
+               ondemand_powersave_bias_init_cpu(cpu);
+               mutex_init(&this_dbs_info->timer_mutex);
                /*
                 * Start the timerschedule work, when this governor
                 * is used for first time
@@ -619,36 +576,38 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
                        latency = policy->cpuinfo.transition_latency / 1000;
                        if (latency == 0)
                                latency = 1;
-
-                       def_sampling_rate =
-                               max(latency * LATENCY_MULTIPLIER,
-                                   MIN_STAT_SAMPLING_RATE);
-
-                       dbs_tuners_ins.sampling_rate = def_sampling_rate;
+                       /* Bring kernel and HW constraints together */
+                       min_sampling_rate = max(min_sampling_rate,
+                                       MIN_LATENCY_MULTIPLIER * latency);
+                       dbs_tuners_ins.sampling_rate =
+                               max(min_sampling_rate,
+                                   latency * LATENCY_MULTIPLIER);
                }
-               dbs_timer_init(this_dbs_info);
-
                mutex_unlock(&dbs_mutex);
+
+               dbs_timer_init(this_dbs_info);
                break;
 
        case CPUFREQ_GOV_STOP:
-               mutex_lock(&dbs_mutex);
                dbs_timer_exit(this_dbs_info);
+
+               mutex_lock(&dbs_mutex);
                sysfs_remove_group(&policy->kobj, &dbs_attr_group);
+               mutex_destroy(&this_dbs_info->timer_mutex);
                dbs_enable--;
                mutex_unlock(&dbs_mutex);
 
                break;
 
        case CPUFREQ_GOV_LIMITS:
-               mutex_lock(&dbs_mutex);
+               mutex_lock(&this_dbs_info->timer_mutex);
                if (policy->max < this_dbs_info->cur_policy->cur)
                        __cpufreq_driver_target(this_dbs_info->cur_policy,
                                policy->max, CPUFREQ_RELATION_H);
                else if (policy->min > this_dbs_info->cur_policy->cur)
                        __cpufreq_driver_target(this_dbs_info->cur_policy,
                                policy->min, CPUFREQ_RELATION_L);
-               mutex_unlock(&dbs_mutex);
+               mutex_unlock(&this_dbs_info->timer_mutex);
                break;
        }
        return 0;
@@ -678,6 +637,16 @@ static int __init cpufreq_gov_dbs_init(void)
                dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
                dbs_tuners_ins.down_differential =
                                        MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
+               /*
+                * In no_hz/micro accounting case we set the minimum frequency
+                * not depending on HZ, but fixed (very low). The deferred
+                * timer might skip some samples if idle/sleeping as needed.
+               */
+               min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
+       } else {
+               /* For correct statistics, we need 10 ticks for each measure */
+               min_sampling_rate =
+                       MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
        }
 
        kondemand_wq = create_workqueue("kondemand");