cpufreq: schedutil: Avoid missing updates for one-CPU policies
authorRafael J. Wysocki <rafael.j.wysocki@intel.com>
Wed, 23 May 2018 09:47:45 +0000 (11:47 +0200)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Thu, 24 May 2018 08:21:18 +0000 (10:21 +0200)
Commit 152db033d775 (schedutil: Allow cpufreq requests to be made
even when kthread kicked) made changes to prevent utilization updates
from being discarded during processing a previous request, but it
left a small window in which that still can happen in the one-CPU
policy case.  Namely, updates coming in after setting work_in_progress
in sugov_update_commit() and clearing it in sugov_work() will still
be dropped due to the work_in_progress check in sugov_update_single().

To close that window, rearrange the code so as to acquire the update
lock around the deferred update branch in sugov_update_single()
and drop the work_in_progress check from it.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Juri Lelli <juri.lelli@redhat.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org>
kernel/sched/cpufreq_schedutil.c

index 178946e3639348260fa6b42ddbc47fd384327f1c..fd76497efeb17619357e0acbe6ba0f5aed4989ac 100644 (file)
@@ -100,25 +100,41 @@ static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
        return delta_ns >= sg_policy->freq_update_delay_ns;
 }
 
-static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
-                               unsigned int next_freq)
+static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
+                                  unsigned int next_freq)
 {
-       struct cpufreq_policy *policy = sg_policy->policy;
-
        if (sg_policy->next_freq == next_freq)
-               return;
+               return false;
 
        sg_policy->next_freq = next_freq;
        sg_policy->last_freq_update_time = time;
 
-       if (policy->fast_switch_enabled) {
-               next_freq = cpufreq_driver_fast_switch(policy, next_freq);
-               if (!next_freq)
-                       return;
+       return true;
+}
+
+static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
+                             unsigned int next_freq)
+{
+       struct cpufreq_policy *policy = sg_policy->policy;
+
+       if (!sugov_update_next_freq(sg_policy, time, next_freq))
+               return;
+
+       next_freq = cpufreq_driver_fast_switch(policy, next_freq);
+       if (!next_freq)
+               return;
 
-               policy->cur = next_freq;
-               trace_cpu_frequency(next_freq, smp_processor_id());
-       } else if (!sg_policy->work_in_progress) {
+       policy->cur = next_freq;
+       trace_cpu_frequency(next_freq, smp_processor_id());
+}
+
+static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
+                                 unsigned int next_freq)
+{
+       if (!sugov_update_next_freq(sg_policy, time, next_freq))
+               return;
+
+       if (!sg_policy->work_in_progress) {
                sg_policy->work_in_progress = true;
                irq_work_queue(&sg_policy->irq_work);
        }
@@ -363,13 +379,6 @@ static void sugov_update_single(struct update_util_data *hook, u64 time,
 
        ignore_dl_rate_limit(sg_cpu, sg_policy);
 
-       /*
-        * For slow-switch systems, single policy requests can't run at the
-        * moment if update is in progress, unless we acquire update_lock.
-        */
-       if (sg_policy->work_in_progress)
-               return;
-
        if (!sugov_should_update_freq(sg_policy, time))
                return;
 
@@ -391,7 +400,18 @@ static void sugov_update_single(struct update_util_data *hook, u64 time,
                sg_policy->cached_raw_freq = 0;
        }
 
-       sugov_update_commit(sg_policy, time, next_f);
+       /*
+        * This code runs under rq->lock for the target CPU, so it won't run
+        * concurrently on two different CPUs for the same target and it is not
+        * necessary to acquire the lock in the fast switch case.
+        */
+       if (sg_policy->policy->fast_switch_enabled) {
+               sugov_fast_switch(sg_policy, time, next_f);
+       } else {
+               raw_spin_lock(&sg_policy->update_lock);
+               sugov_deferred_update(sg_policy, time, next_f);
+               raw_spin_unlock(&sg_policy->update_lock);
+       }
 }
 
 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
@@ -435,7 +455,11 @@ sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 
        if (sugov_should_update_freq(sg_policy, time)) {
                next_f = sugov_next_freq_shared(sg_cpu, time);
-               sugov_update_commit(sg_policy, time, next_f);
+
+               if (sg_policy->policy->fast_switch_enabled)
+                       sugov_fast_switch(sg_policy, time, next_f);
+               else
+                       sugov_deferred_update(sg_policy, time, next_f);
        }
 
        raw_spin_unlock(&sg_policy->update_lock);
@@ -450,11 +474,11 @@ static void sugov_work(struct kthread_work *work)
        /*
         * Hold sg_policy->update_lock shortly to handle the case where:
         * incase sg_policy->next_freq is read here, and then updated by
-        * sugov_update_shared just before work_in_progress is set to false
+        * sugov_deferred_update() just before work_in_progress is set to false
         * here, we may miss queueing the new update.
         *
         * Note: If a work was queued after the update_lock is released,
-        * sugov_work will just be called again by kthread_work code; and the
+        * sugov_work() will just be called again by kthread_work code; and the
         * request will be proceed before the sugov thread sleeps.
         */
        raw_spin_lock_irqsave(&sg_policy->update_lock, flags);