Skip to contentSkip to conversation

Take care of cpufreq lockdep issues

PATCH In review [email protected] 2 participants 7 messages Jun 25, 2009 CPU frequency scaling framework

lore
Share & tools
lore
#0 lore raw

cover Take care of cpufreq lockdep issues

Since recent chanegs to ondemand and conservative governor, there have been multiple reports of lockdep issues in cpufreq. Patch series takes care of these problems.

#1 lore raw

PATCH 1/3 cpufreq: remove rwsem lock from CPUFREQ_GOV_STOP call (second call site)

From: Mathieu Desnoyers <mathieu.desnoyers-scC8bbJcJLCw5LPnMra/[email protected]>

commit 42a06f2166f2f6f7bf04f32b4e823eacdceafdc9

Missed a call site for CPUFREQ_GOV_STOP to remove the rwlock taken around the teardown. To make a long story short, the rwlock write-lock causes a circular dependency with cancel_delayed_work_sync(), because the timer handler takes the write lock.

Note that all callers to __cpufreq_set_policy are taking the rwsem. All sysfs callers (writers) hold the write rwsem at the earliest sysfs calling stage.

However, the rwlock write-lock is not needed upon governor stop.

Change :
- Added comment from Venki at lock definition site.

---
 drivers/cpufreq/cpufreq.c |   13 ++++++++++++-
 1 files changed, 12 insertions(+), 1 deletions(-)
drivers/cpufreq/cpufreq.c +12 -1
index 6e2ec0b..728656c 100644
@@ -61,6 +61,8 @@ static DEFINE_SPINLOCK(cpufreq_driver_lock);
  *   are concerned with are online after they get the lock.
  * - Governor routines that can be called in cpufreq hotplug path should not
  *   take this sem as top level hotplug notifier handler takes this.
+ * - Lock should not be held across
+ *     __cpufreq_governor(data, CPUFREQ_GOV_STOP);
  */
 static DEFINE_PER_CPU(int, policy_cpu);
 static DEFINE_PER_CPU(struct rw_semaphore, cpu_policy_rwsem);
@@ -1697,8 +1699,17 @@ static int __cpufreq_set_policy(struct cpufreq_policy *data,
 			dprintk("governor switch\n");
 
 			/* end old governor */
-			if (data->governor)
+			if (data->governor) {
+				/*
+				 * Need to release the rwsem around governor
+				 * stop due to lock dependency between
+				 * cancel_delayed_work_sync and the read lock
+				 * taken in the delayed work handler.
+				 */
+				unlock_policy_rwsem_write(data->cpu);
 				__cpufreq_governor(data, CPUFREQ_GOV_STOP);
+				lock_policy_rwsem_write(data->cpu);
+			}
 
 			/* start new governor */
 			data->governor = policy->governor;
Signature · 3 lines
1.6.0.6

-- 
#2 lore raw

PATCH 2/3 cpufreq: Define dbs_mutex purpose and cleanup its usage

In reply to [email protected]

Commit b14893a62c73af0eca414cfed505b8c09efc613c although it was very much needed to cleanup ondemand timer cleanly, openup 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

---
 drivers/cpufreq/cpufreq_ondemand.c |   37 +++++++++++++++--------------------
 1 files changed, 16 insertions(+), 21 deletions(-)
drivers/cpufreq/cpufreq_ondemand.c +16 -21
index 1911d17..b2d2106 100644
@@ -78,15 +78,14 @@ 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 also serializes dbs_enable usage in CPUFREQ_GOV_START
+ * and CPUFREQ_GOV_STOP.
+ *
+ * dbs_mutex should be always held after lock_policy_rwsem whenever needed.
+ * 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.
  */
 static DEFINE_MUTEX(dbs_mutex);
 
@@ -240,12 +239,10 @@ 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, min_sampling_rate);
 	mutex_unlock(&dbs_mutex);
 
@@ -258,14 +255,12 @@ static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 	unsigned int input;
 	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);
 
@@ -324,8 +319,8 @@ static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
 
 	mutex_lock(&dbs_mutex);
 	dbs_tuners_ins.powersave_bias = input;
-	ondemand_powersave_bias_init();
 	mutex_unlock(&dbs_mutex);
+	ondemand_powersave_bias_init();
 
 	return count;
 }
@@ -598,14 +593,16 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 				max(min_sampling_rate,
 				    latency * LATENCY_MULTIPLIER);
 		}
+		mutex_unlock(&dbs_mutex);
+
 		dbs_timer_init(this_dbs_info);
 
-		mutex_unlock(&dbs_mutex);
 		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);
@@ -613,14 +610,12 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 		break;
 
 	case CPUFREQ_GOV_LIMITS:
-		mutex_lock(&dbs_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);
 		break;
 	}
 	return 0;
Signature · 3 lines
1.6.0.6

-- 
#3 lore raw

Re: [patch 2/3] cpufreq: Define dbs_mutex purpose and cleanup its usage

Quoted 87 lines · * [email protected]…

Commit b14893a62c73af0eca414cfed505b8c09efc613c although it was very much needed to cleanup ondemand timer cleanly, openup 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

---
 drivers/cpufreq/cpufreq_ondemand.c |   37 +++++++++++++++--------------------
 1 files changed, 16 insertions(+), 21 deletions(-)
drivers/cpufreq/cpufreq_ondemand.c +12 -17
index 1911d17..b2d2106 100644
@@ -78,15 +78,14 @@ 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 also serializes dbs_enable usage in CPUFREQ_GOV_START
+ * and CPUFREQ_GOV_STOP.
+ *
+ * dbs_mutex should be always held after lock_policy_rwsem whenever needed.
+ * 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.
  */
 static DEFINE_MUTEX(dbs_mutex);
 
@@ -240,12 +239,10 @@ 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, min_sampling_rate);
 	mutex_unlock(&dbs_mutex);
 
@@ -258,14 +255,12 @@ static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 	unsigned int input;
 	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);
 
@@ -324,8 +319,8 @@ static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
 
 	mutex_lock(&dbs_mutex);
 	dbs_tuners_ins.powersave_bias = input;
-	ondemand_powersave_bias_init();
 	mutex_unlock(&dbs_mutex);
+	ondemand_powersave_bias_init();
 
 	return count;
 }
@@ -598,14 +593,16 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 				max(min_sampling_rate,
 				    latency * LATENCY_MULTIPLIER);
 		}

drivers/cpufreq/cpufreq_ondemand.c static int cpufreq_governor_dbs(struct cpufreq_policy *policy, · line 603

Quoted 10 lines
drivers/cpufreq/cpufreq_ondemand.c +2 -2
+		mutex_unlock(&dbs_mutex);
+
 		dbs_timer_init(this_dbs_info);
 
-		mutex_unlock(&dbs_mutex);
 		break;
 
 	case CPUFREQ_GOV_STOP:
-		mutex_lock(&dbs_mutex);
 		dbs_timer_exit(this_dbs_info);

Hrm, so.. how do we protect against concurrent :

CPUFREQ_GOV_START/CPUFREQ_GOV_STOP now ?

Mathieu

Quoted 25 lines
drivers/cpufreq/cpufreq_ondemand.c +2 -2
+
+		mutex_lock(&dbs_mutex);
 		sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 		dbs_enable--;
 		mutex_unlock(&dbs_mutex);
@@ -613,14 +610,12 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 		break;
 
 	case CPUFREQ_GOV_LIMITS:
-		mutex_lock(&dbs_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);
 		break;
 	}
 	return 0;
Signature · 3 lines
1.6.0.6

-- 
Signature · 2 lines
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68
#4 lore raw

Re: [patch 2/3] cpufreq: Define dbs_mutex purpose and cleanup its usage

Quoted 22 lines · Mathieu Desnoyers
Quoted 21 lines · * [email protected] ([email protected]) wrote:

Commit b14893a62c73af0eca414cfed505b8c09efc613c although it was very much needed to cleanup ondemand timer cleanly, openup 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

drivers/cpufreq/cpufreq_ondemand.c +2 -1
@@ -598,14 +593,16 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 				max(min_sampling_rate,
 				    latency * LATENCY_MULTIPLIER);
 		}
+		mutex_unlock(&dbs_mutex);
+
 		dbs_timer_init(this_dbs_info);
 
-		mutex_unlock(&dbs_mutex);
 		break;
 
Quoted 7 lines

drivers/cpufreq/cpufreq_ondemand.c static int cpufreq_governor_dbs(struct cpufreq_policy *policy, · line 603

Quoted 3 lines
drivers/cpufreq/cpufreq_ondemand.c +0 -1
 	case CPUFREQ_GOV_STOP:
-		mutex_lock(&dbs_mutex);
 		dbs_timer_exit(this_dbs_info);

Hrm, so.. how do we protect against concurrent :

CPUFREQ_GOV_START/CPUFREQ_GOV_STOP now ?

concurrent _START _STOP across CPUs does not matter for timer_init and timer_exit. On same CPU, there cannot be two concurrent _START as upper level cpufreq will have policy_rwsem in write mode. I cannot think of a flow where _START and _STOP on same CPU is possible.

However two concurrent _STOP for same CPU is still possible, as we are releasing the rwsem lock before STOP callback. "Back to drawing board" time to figure this all out..

Thanks,
Venki

#5 lore raw

Re: [patch 2/3] cpufreq: Define dbs_mutex purpose and cleanup its usage

Quoted 29 lines · * Pallipadi, Venkatesh ([email protected]…
Quoted 28 lines · Mathieu Desnoyers
Quoted 14 lines · * [email protected]…

Commit b14893a62c73af0eca414cfed505b8c09efc613c although it was very much needed to cleanup ondemand timer cleanly, openup 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

drivers/cpufreq/cpufreq_ondemand.c +0 -0
@@ -598,14 +593,16 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 				max(min_sampling_rate,
 				    latency * LATENCY_MULTIPLIER);
 		}

drivers/cpufreq/cpufreq_ondemand.c static int cpufreq_governor_dbs(struct cpufreq_policy *policy, · line 603

Quoted 10 lines
drivers/cpufreq/cpufreq_ondemand.c +2 -2
+		mutex_unlock(&dbs_mutex);
+
 		dbs_timer_init(this_dbs_info);
 
-		mutex_unlock(&dbs_mutex);
 		break;
 
 	case CPUFREQ_GOV_STOP:
-		mutex_lock(&dbs_mutex);
 		dbs_timer_exit(this_dbs_info);

Hrm, so.. how do we protect against concurrent :

Quoted 4 lines
Quoted 1 line

CPUFREQ_GOV_START/CPUFREQ_GOV_STOP now ?

concurrent _START _STOP across CPUs does not matter for timer_init and timer_exit.

Given those are per-cpu anyway I guess. Hopefully it works OK with CPU hotplug.

Quoted 2 lines

On same CPU, there cannot be two concurrent _START as upper level cpufreq will have policy_rwsem in write mode.

Agreed.

Quoted 3 lines

I cannot think of a
flow where _START and _STOP on same CPU is possible.

_STOP is not protected by any mutex now. So it could be preempted, and then a _START executed, and there is your race.

Quoted 3 lines

However two concurrent _STOP for same CPU is still possible, as we are releasing the rwsem lock before STOP callback. "Back to drawing board" time to figure this all out..

I fear that it is indeed the case. If you can come up with a document explaining the expected interactions between :

- cpu hotplug
- policy lock
- cpufreq driver lock
- timer lock

that would be awesome. :)

Mathieu

Quoted 4 lines

Thanks,
Venki

Signature · 2 lines
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68
#6 lore raw

PATCH 3/3 cpufreq: Define dbs_mutex purpose and cleanup its usage conservative gov

In reply to [email protected]

Commit b253d2b2d28ead6fed012feb54694b3d0562839a although it was very much needed to cleanup ondemand timer cleanly, openup a can of worms related to locking dependencies in cpufreq.

Patch here defines the need for dbs_mutex and cleans up its usage in conservative governor. This also resolves the lockdep warnings in conservative governor which would be similar to one reported here

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

---
 drivers/cpufreq/cpufreq_conservative.c |   26 ++++++++++++--------------
 1 files changed, 12 insertions(+), 14 deletions(-)
drivers/cpufreq/cpufreq_conservative.c +12 -14
index 7fc58af..581d057 100644
@@ -70,15 +70,14 @@ 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 also serializes dbs_enable usage in CPUFREQ_GOV_START
+ * and CPUFREQ_GOV_STOP.
+ *
+ * dbs_mutex should be always held after lock_policy_rwsem whenever needed.
+ * 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.
  */
 static DEFINE_MUTEX(dbs_mutex);
 
@@ -590,15 +589,16 @@ 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--;
 
@@ -616,7 +616,6 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 		break;
 
 	case CPUFREQ_GOV_LIMITS:
-		mutex_lock(&dbs_mutex);
 		if (policy->max < this_dbs_info->cur_policy->cur)
 			__cpufreq_driver_target(
 					this_dbs_info->cur_policy,
@@ -625,7 +624,6 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 			__cpufreq_driver_target(
 					this_dbs_info->cur_policy,
 					policy->min, CPUFREQ_RELATION_L);
-		mutex_unlock(&dbs_mutex);
 
 		break;
 	}
Signature · 3 lines
1.6.0.6

--