ret = -EINVAL;
                        goto free_policy_dbs_info;
                }
-               dbs_data->usage_count++;
                policy_dbs->dbs_data = dbs_data;
                policy->governor_data = policy_dbs;
+
+               mutex_lock(&dbs_data->mutex);
+               dbs_data->usage_count++;
+               list_add(&policy_dbs->list, &dbs_data->policy_dbs_list);
+               mutex_unlock(&dbs_data->mutex);
+
                return 0;
        }
 
                goto free_policy_dbs_info;
        }
 
-       dbs_data->usage_count = 1;
+       INIT_LIST_HEAD(&dbs_data->policy_dbs_list);
        mutex_init(&dbs_data->mutex);
 
        ret = gov->init(dbs_data, !policy->governor->initialized);
        if (!have_governor_per_policy())
                gov->gdbs_data = dbs_data;
 
-       policy_dbs->dbs_data = dbs_data;
        policy->governor_data = policy_dbs;
 
+       policy_dbs->dbs_data = dbs_data;
+       dbs_data->usage_count = 1;
+       list_add(&policy_dbs->list, &dbs_data->policy_dbs_list);
+
        gov->kobj_type.sysfs_ops = &governor_sysfs_ops;
        ret = kobject_init_and_add(&dbs_data->kobj, &gov->kobj_type,
                                   get_governor_parent_kobj(policy),
        struct dbs_governor *gov = dbs_governor_of(policy);
        struct policy_dbs_info *policy_dbs = policy->governor_data;
        struct dbs_data *dbs_data = policy_dbs->dbs_data;
+       int count;
 
        /* State should be equivalent to INIT */
        if (policy_dbs->policy)
                return -EBUSY;
 
-       if (!--dbs_data->usage_count) {
+       mutex_lock(&dbs_data->mutex);
+       list_del(&policy_dbs->list);
+       count = --dbs_data->usage_count;
+       mutex_unlock(&dbs_data->mutex);
+
+       if (!count) {
                kobject_put(&dbs_data->kobj);
 
                policy->governor_data = NULL;
 
  * @new_rate: new sampling rate
  *
  * If new rate is smaller than the old, simply updating
- * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
+ * dbs.sampling_rate might not be appropriate. For example, if the
  * original sampling_rate was 1 second and the requested new sampling rate is 10
  * ms because the user needs immediate reaction from ondemand governor, but not
  * sure if higher frequency will be required or not, then, the governor may
  * change the sampling rate too late; up to 1 second later. Thus, if we are
  * reducing the sampling rate, we need to make the new value effective
  * immediately.
+ *
+ * On the other hand, if new rate is larger than the old, then we may evaluate
+ * the load too soon, and it might we worth updating sample_delay_ns then as
+ * well.
+ *
+ * This must be called with dbs_data->mutex held, otherwise traversing
+ * policy_dbs_list isn't safe.
  */
 static void update_sampling_rate(struct dbs_data *dbs_data,
                unsigned int new_rate)
 {
-       struct cpumask cpumask;
-       int cpu;
+       struct policy_dbs_info *policy_dbs;
 
        dbs_data->sampling_rate = new_rate = max(new_rate,
                        dbs_data->min_sampling_rate);
 
        /*
-        * Lock governor so that governor start/stop can't execute in parallel.
+        * We are operating under dbs_data->mutex and so the list and its
+        * entries can't be freed concurrently.
         */
-       mutex_lock(&dbs_data_mutex);
-
-       cpumask_copy(&cpumask, cpu_online_mask);
-
-       for_each_cpu(cpu, &cpumask) {
-               struct cpufreq_policy *policy;
-               struct od_cpu_dbs_info_s *dbs_info;
-               struct cpu_dbs_info *cdbs;
-               struct policy_dbs_info *policy_dbs;
-
-               dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
-               cdbs = &dbs_info->cdbs;
-               policy_dbs = cdbs->policy_dbs;
-
+       list_for_each_entry(policy_dbs, &dbs_data->policy_dbs_list, list) {
+               mutex_lock(&policy_dbs->timer_mutex);
                /*
-                * A valid policy_dbs and policy_dbs->policy means governor
-                * hasn't stopped or exited yet.
+                * On 32-bit architectures this may race with the
+                * sample_delay_ns read in dbs_update_util_handler(), but that
+                * really doesn't matter.  If the read returns a value that's
+                * too big, the sample will be skipped, but the next invocation
+                * of dbs_update_util_handler() (when the update has been
+                * completed) will take a sample.  If the returned value is too
+                * small, the sample will be taken immediately, but that isn't a
+                * problem, as we want the new rate to take effect immediately
+                * anyway.
+                *
+                * If this runs in parallel with dbs_work_handler(), we may end
+                * up overwriting the sample_delay_ns value that it has just
+                * written, but the difference should not be too big and it will
+                * be corrected next time a sample is taken, so it shouldn't be
+                * significant.
                 */
-               if (!policy_dbs || !policy_dbs->policy)
-                       continue;
-
-               policy = policy_dbs->policy;
-
-               /* clear all CPUs of this policy */
-               cpumask_andnot(&cpumask, &cpumask, policy->cpus);
-
-               /*
-                * Update sampling rate for CPUs whose policy is governed by
-                * dbs_data. In case of governor_per_policy, only a single
-                * policy will be governed by dbs_data, otherwise there can be
-                * multiple policies that are governed by the same dbs_data.
-                */
-               if (dbs_data == policy_dbs->dbs_data) {
-                       mutex_lock(&policy_dbs->timer_mutex);
-                       /*
-                        * On 32-bit architectures this may race with the
-                        * sample_delay_ns read in dbs_update_util_handler(),
-                        * but that really doesn't matter.  If the read returns
-                        * a value that's too big, the sample will be skipped,
-                        * but the next invocation of dbs_update_util_handler()
-                        * (when the update has been completed) will take a
-                        * sample.  If the returned value is too small, the
-                        * sample will be taken immediately, but that isn't a
-                        * problem, as we want the new rate to take effect
-                        * immediately anyway.
-                        *
-                        * If this runs in parallel with dbs_work_handler(), we
-                        * may end up overwriting the sample_delay_ns value that
-                        * it has just written, but the difference should not be
-                        * too big and it will be corrected next time a sample
-                        * is taken, so it shouldn't be significant.
-                        */
-                       gov_update_sample_delay(policy_dbs, new_rate);
-                       mutex_unlock(&policy_dbs->timer_mutex);
-               }
+               gov_update_sample_delay(policy_dbs, new_rate);
+               mutex_unlock(&policy_dbs->timer_mutex);
        }
-
-       mutex_unlock(&dbs_data_mutex);
 }
 
 static ssize_t store_sampling_rate(struct dbs_data *dbs_data, const char *buf,