[CPUFREQ] Eliminate the recent lockdep warnings in cpufreq
authorvenkatesh.pallipadi@intel.com <venkatesh.pallipadi@intel.com>
Fri, 3 Jul 2009 00:08:30 +0000 (17:08 -0700)
committerDave Jones <davej@redhat.com>
Tue, 7 Jul 2009 01:38:27 +0000 (21:38 -0400)
Commit b14893a62c73af0eca414cfed505b8c09efc613c although it was very
much needed to properly cleanup ondemand timer, opened-up a can of worms
related to locking dependencies in cpufreq.

Patch here defines the need for dbs_mutex and cleans up its usage in
ondemand governor. This also resolves the lockdep warnings reported here

http://lkml.indiana.edu/hypermail/linux/kernel/0906.1/01925.html
http://lkml.indiana.edu/hypermail/linux/kernel/0907.0/00820.html

and few others..

Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Dave Jones <davej@redhat.com>
drivers/cpufreq/cpufreq.c
drivers/cpufreq/cpufreq_conservative.c
drivers/cpufreq/cpufreq_ondemand.c

index 6e2ec0b..c7fe16e 100644 (file)
@@ -1070,8 +1070,6 @@ static int __cpufreq_remove_dev(struct sys_device *sys_dev)
        spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
 #endif
 
-       unlock_policy_rwsem_write(cpu);
-
        if (cpufreq_driver->target)
                __cpufreq_governor(data, CPUFREQ_GOV_STOP);
 
@@ -1088,6 +1086,8 @@ static int __cpufreq_remove_dev(struct sys_device *sys_dev)
        if (cpufreq_driver->exit)
                cpufreq_driver->exit(data);
 
+       unlock_policy_rwsem_write(cpu);
+
        free_cpumask_var(data->related_cpus);
        free_cpumask_var(data->cpus);
        kfree(data);
index 7fc58af..58889f2 100644 (file)
@@ -70,15 +70,10 @@ 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. It also
+ * serializes governor limit_change with do_dbs_timer. We do not want
+ * do_dbs_timer to run when user is changing the governor or limits.
  */
 static DEFINE_MUTEX(dbs_mutex);
 
@@ -488,18 +483,17 @@ static void do_dbs_timer(struct work_struct *work)
 
        delay -= jiffies % delay;
 
-       if (lock_policy_rwsem_write(cpu) < 0)
-               return;
+       mutex_lock(&dbs_mutex);
 
        if (!dbs_info->enable) {
-               unlock_policy_rwsem_write(cpu);
+               mutex_unlock(&dbs_mutex);
                return;
        }
 
        dbs_check_cpu(dbs_info);
 
        queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay);
-       unlock_policy_rwsem_write(cpu);
+       mutex_unlock(&dbs_mutex);
 }
 
 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
@@ -590,15 +584,16 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
                                        &dbs_cpufreq_notifier_block,
                                        CPUFREQ_TRANSITION_NOTIFIER);
                }
-               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);
                dbs_enable--;
 
index 1911d17..246ae14 100644 (file)
@@ -78,15 +78,10 @@ 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. It also
+ * serializes governor limit_change with do_dbs_timer. We do not want
+ * do_dbs_timer to run when user is changing the governor or limits.
  */
 static DEFINE_MUTEX(dbs_mutex);
 
@@ -494,11 +489,10 @@ static void do_dbs_timer(struct work_struct *work)
 
        delay -= jiffies % delay;
 
-       if (lock_policy_rwsem_write(cpu) < 0)
-               return;
+       mutex_lock(&dbs_mutex);
 
        if (!dbs_info->enable) {
-               unlock_policy_rwsem_write(cpu);
+               mutex_unlock(&dbs_mutex);
                return;
        }
 
@@ -517,7 +511,7 @@ 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_mutex);
 }
 
 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
@@ -598,14 +592,15 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
                                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);
                dbs_enable--;
                mutex_unlock(&dbs_mutex);