cpufreq: governor: Quit work-handlers early if governor is stopped
authorViresh Kumar <viresh.kumar@linaro.org>
Thu, 29 Oct 2015 02:38:38 +0000 (08:08 +0530)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Mon, 2 Nov 2015 01:07:11 +0000 (02:07 +0100)
gov_queue_work() acquires cpufreq_governor_lock to allow
cpufreq_governor_stop() to drain delayed work items possibly scheduled
on CPUs that share the policy with a CPU being taken offline.

However, the same goal may be achieved in a more straightforward way if
the policy pointer in the struct cpu_dbs_info matching the policy CPU is
reset upfront by cpufreq_governor_stop() under the timer_mutex belonging
to it and checked against NULL, under the same lock, at the beginning of
dbs_timer().

In that case every instance of dbs_timer() run for a struct cpu_dbs_info
sharing the policy pointer in question after cpufreq_governor_stop() has
started will notice that that pointer is NULL and bail out immediately
without queuing up any new work items.  In turn, gov_cancel_work()
called by cpufreq_governor_stop() before destroying timer_mutex will
wait for all of the delayed work items currently running on the CPUs
sharing the policy to drop the mutex, so it may be destroyed safely.

Make cpufreq_governor_stop() and dbs_timer() work as described and
modify gov_queue_work() so it does not acquire cpufreq_governor_lock any
more.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/cpufreq/cpufreq_governor.c

index 11258c4c1b175be37827c201a8d7a422a5d3eda3..b260576ddb129c9bf485a12a954d0b1db30e08c5 100644 (file)
@@ -171,10 +171,6 @@ void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
 {
        int i;
 
-       mutex_lock(&cpufreq_governor_lock);
-       if (!policy->governor_enabled)
-               goto out_unlock;
-
        if (!all_cpus) {
                /*
                 * Use raw_smp_processor_id() to avoid preemptible warnings.
@@ -188,9 +184,6 @@ void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
                for_each_cpu(i, policy->cpus)
                        __gov_queue_work(i, dbs_data, delay);
        }
-
-out_unlock:
-       mutex_unlock(&cpufreq_governor_lock);
 }
 EXPORT_SYMBOL_GPL(gov_queue_work);
 
@@ -229,13 +222,24 @@ static void dbs_timer(struct work_struct *work)
        struct cpu_dbs_info *cdbs = container_of(work, struct cpu_dbs_info,
                                                 dwork.work);
        struct cpu_common_dbs_info *shared = cdbs->shared;
-       struct cpufreq_policy *policy = shared->policy;
-       struct dbs_data *dbs_data = policy->governor_data;
+       struct cpufreq_policy *policy;
+       struct dbs_data *dbs_data;
        unsigned int sampling_rate, delay;
        bool modify_all = true;
 
        mutex_lock(&shared->timer_mutex);
 
+       policy = shared->policy;
+
+       /*
+        * Governor might already be disabled and there is no point continuing
+        * with the work-handler.
+        */
+       if (!policy)
+               goto unlock;
+
+       dbs_data = policy->governor_data;
+
        if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
                struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
 
@@ -252,6 +256,7 @@ static void dbs_timer(struct work_struct *work)
        delay = dbs_data->cdata->gov_dbs_timer(cdbs, dbs_data, modify_all);
        gov_queue_work(dbs_data, policy, delay, modify_all);
 
+unlock:
        mutex_unlock(&shared->timer_mutex);
 }
 
@@ -478,9 +483,17 @@ static int cpufreq_governor_stop(struct cpufreq_policy *policy,
        if (!shared || !shared->policy)
                return -EBUSY;
 
+       /*
+        * Work-handler must see this updated, as it should not proceed any
+        * further after governor is disabled. And so timer_mutex is taken while
+        * updating this value.
+        */
+       mutex_lock(&shared->timer_mutex);
+       shared->policy = NULL;
+       mutex_unlock(&shared->timer_mutex);
+
        gov_cancel_work(dbs_data, policy);
 
-       shared->policy = NULL;
        mutex_destroy(&shared->timer_mutex);
        return 0;
 }