[PATCH] ondemand governor use new cpufreq rwsem locking in work callback

[PATCH] ondemand governor use new cpufreq rwsem locking in work callback

[Venkatesh Pallipadi]

Eliminate flush_workqueue in cpufreq_governor(STOP) callpath. Using flush
there has a deadlock potential as in 

http://uwsg.iu.edu/hypermail/linux/kernel/0611.3/1223.html

Also, cleanup the locking issues with do_dbs_timer delayed_work callback.
As it changes the CPU frequency using __cpufreq_target, it needs to have
policy_rwsem in write mode, which also protects it from hot plug.

Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>

Index: linux-2.6.20-rc-mm/drivers/cpufreq/cpufreq_ondemand.c
===================================================================
--- linux-2.6.20-rc-mm.orig/drivers/cpufreq/cpufreq_ondemand.c
+++ linux-2.6.20-rc-mm/drivers/cpufreq/cpufreq_ondemand.c
@@ -437,8 +437,14 @@ static void do_dbs_timer(struct work_str
 
 	delay -= jiffies % delay;
 
-	if (!dbs_info->enable)
+	if (lock_policy_rwsem_write(cpu) < 0)
 		return;
+
+	if (!dbs_info->enable) {
+		unlock_policy_rwsem_write(cpu);
+		return;
+	}
+
 	/* Common NORMAL_SAMPLE setup */
 	dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 	if (!dbs_tuners_ins.powersave_bias ||
@@ -455,6 +461,7 @@ static void do_dbs_timer(struct work_str
 	                        	CPUFREQ_RELATION_H);
 	}
 	queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
+	unlock_policy_rwsem_write(cpu);
 }
 
 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
@@ -463,6 +470,7 @@ static inline void dbs_timer_init(struct
 	int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 	delay -= jiffies % delay;
 
+ 	dbs_info->enable = 1;
 	ondemand_powersave_bias_init();
 	dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 	INIT_DELAYED_WORK_NAR(&dbs_info->work, do_dbs_timer);
@@ -474,7 +482,6 @@ static inline void dbs_timer_exit(struct
 {
 	dbs_info->enable = 0;
 	cancel_delayed_work(&dbs_info->work);
-	flush_workqueue(kondemand_wq);
 }
 
 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
@@ -503,21 +510,9 @@ static int cpufreq_governor_dbs(struct c
 
 		mutex_lock(&dbs_mutex);
 		dbs_enable++;
-		if (dbs_enable == 1) {
-			kondemand_wq = create_workqueue("kondemand");
-			if (!kondemand_wq) {
-				printk(KERN_ERR
-					 "Creation of kondemand failed\n");
-				dbs_enable--;
-				mutex_unlock(&dbs_mutex);
-				return -ENOSPC;
-			}
-		}
 
 		rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
 		if (rc) {
-			if (dbs_enable == 1)
-				destroy_workqueue(kondemand_wq);
 			dbs_enable--;
 			mutex_unlock(&dbs_mutex);
 			return rc;
@@ -532,7 +527,6 @@ static int cpufreq_governor_dbs(struct c
 			j_dbs_info->prev_cpu_wall = get_jiffies_64();
 		}
 		this_dbs_info->cpu = cpu;
-		this_dbs_info->enable = 1;
 		/*
 		 * Start the timerschedule work, when this governor
 		 * is used for first time
@@ -562,9 +556,6 @@ static int cpufreq_governor_dbs(struct c
 		dbs_timer_exit(this_dbs_info);
 		sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 		dbs_enable--;
-		if (dbs_enable == 0)
-			destroy_workqueue(kondemand_wq);
-
 		mutex_unlock(&dbs_mutex);
 
 		break;
@@ -593,12 +584,18 @@ static struct cpufreq_governor cpufreq_g
 
 static int __init cpufreq_gov_dbs_init(void)
 {
+	kondemand_wq = create_workqueue("kondemand");
+	if (!kondemand_wq) {
+		printk(KERN_ERR "Creation of kondemand failed\n");
+		return -EFAULT;
+	}
 	return cpufreq_register_governor(&cpufreq_gov_dbs);
 }
 
 static void __exit cpufreq_gov_dbs_exit(void)
 {
 	cpufreq_unregister_governor(&cpufreq_gov_dbs);
+	destroy_workqueue(kondemand_wq);
 }
 
 
@@ -610,3 +607,4 @@ MODULE_LICENSE("GPL");
 
 module_init(cpufreq_gov_dbs_init);
 module_exit(cpufreq_gov_dbs_exit);
+

Re: [PATCH] cpufreq lock rewrite bugfix - Fix locking in cpufreq_get

[Venkatesh Pallipadi]

Incremental bugfix to previous 3 patch patchset of cpufreq lock rewrite.

There was one code path in cpufreq_get, that was using the write lock
in place of read and also potential recursive lock with sysfs
interface of cpuinfo_cur_freq.

Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>

Index: linux-2.6.20-rc-mm/drivers/cpufreq/cpufreq.c
===================================================================
--- linux-2.6.20-rc-mm.orig/drivers/cpufreq/cpufreq.c
+++ linux-2.6.20-rc-mm/drivers/cpufreq/cpufreq.c
@@ -101,6 +101,7 @@ EXPORT_SYMBOL_GPL(unlock_policy_rwsem_wr
 
 /* internal prototypes */
 static int __cpufreq_governor(struct cpufreq_policy *policy, unsigned int event);
+static unsigned int __cpufreq_get(unsigned int cpu);
 static void handle_update(struct work_struct *work);
 
 /**
@@ -488,7 +489,7 @@ store_one(scaling_max_freq,max);
 static ssize_t show_cpuinfo_cur_freq (struct cpufreq_policy * policy,
 							char *buf)
 {
-	unsigned int cur_freq = cpufreq_get(policy->cpu);
+	unsigned int cur_freq = __cpufreq_get(policy->cpu);
 	if (!cur_freq)
 		return sprintf(buf, "<unknown>");
 	return sprintf(buf, "%u\n", cur_freq);
@@ -1074,25 +1075,13 @@ unsigned int cpufreq_quick_get(unsigned 
 EXPORT_SYMBOL(cpufreq_quick_get);
 
 
-/**
- * cpufreq_get - get the current CPU frequency (in kHz)
- * @cpu: CPU number
- *
- * Get the CPU current (static) CPU frequency
- */
-unsigned int cpufreq_get(unsigned int cpu)
+static unsigned int __cpufreq_get(unsigned int cpu)
 {
-	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
+	struct cpufreq_policy *policy = cpufreq_cpu_data[cpu];
 	unsigned int ret_freq = 0;
 
-	if (!policy)
-		return 0;
-
 	if (!cpufreq_driver->get)
-		goto out;
-
-	if (unlikely(lock_policy_rwsem_write(cpu)))
-		goto out;
+		return (ret_freq);
 
 	ret_freq = cpufreq_driver->get(cpu);
 
@@ -1106,11 +1095,32 @@ unsigned int cpufreq_get(unsigned int cp
 		}
 	}
 
-	unlock_policy_rwsem_write(cpu);
+	return (ret_freq);
+}
+
+/**
+ * cpufreq_get - get the current CPU frequency (in kHz)
+ * @cpu: CPU number
+ *
+ * Get the CPU current (static) CPU frequency
+ */
+unsigned int cpufreq_get(unsigned int cpu)
+{
+	unsigned int ret_freq = 0;
+	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
+
+	if (!policy)
+		goto out;
+
+	if (unlikely(lock_policy_rwsem_read(cpu)))
+		goto out;
+
+	ret_freq = __cpufreq_get(cpu);
+
+	unlock_policy_rwsem_read(cpu);
 
 out:
 	cpufreq_cpu_put(policy);
-
 	return (ret_freq);
 }
 EXPORT_SYMBOL(cpufreq_get);