[CPUFREQ] fix (utter) cpufreq_add_dev mess
authorMathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Fri, 3 Jul 2009 15:25:16 +0000 (11:25 -0400)
committerDave Jones <davej@redhat.com>
Tue, 7 Jul 2009 01:38:28 +0000 (21:38 -0400)
OK, I've tried to clean it up the best I could, but please test this with
concurrent cpu hotplug and cpufreq add/remove in loops. I'm sure we will make
other interesting findings.

This is step one of fixing the overall locking dependency mess in cpufreq.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
CC: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
CC: rjw@sisk.pl
CC: mingo@elte.hu
CC: Shaohua Li <shaohua.li@intel.com>
CC: Pekka Enberg <penberg@cs.helsinki.fi>
CC: Dave Young <hidave.darkstar@gmail.com>
CC: "Rafael J. Wysocki" <rjw@sisk.pl>
CC: Rusty Russell <rusty@rustcorp.com.au>
CC: sven.wegener@stealer.net
CC: cpufreq@vger.kernel.org
CC: Thomas Renninger <trenn@suse.de>
Signed-off-by: Dave Jones <davej@redhat.com>
drivers/cpufreq/cpufreq.c

index c7fe16e0474b738f9f60260caf0480d1054365a3..c668ac855f71cf7793e05246b153ce308f9f415d 100644 (file)
@@ -761,6 +761,10 @@ static struct kobj_type ktype_cpufreq = {
  * cpufreq_add_dev - add a CPU device
  *
  * Adds the cpufreq interface for a CPU device.
+ *
+ * The Oracle says: try running cpufreq registration/unregistration concurrently
+ * with with cpu hotplugging and all hell will break loose. Tried to clean this
+ * mess up, but more thorough testing is needed. - Mathieu
  */
 static int cpufreq_add_dev(struct sys_device *sys_dev)
 {
@@ -804,15 +808,12 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
                goto nomem_out;
        }
        if (!alloc_cpumask_var(&policy->cpus, GFP_KERNEL)) {
-               kfree(policy);
                ret = -ENOMEM;
-               goto nomem_out;
+               goto err_free_policy;
        }
        if (!zalloc_cpumask_var(&policy->related_cpus, GFP_KERNEL)) {
-               free_cpumask_var(policy->cpus);
-               kfree(policy);
                ret = -ENOMEM;
-               goto nomem_out;
+               goto err_free_cpumask;
        }
 
        policy->cpu = cpu;
@@ -820,7 +821,8 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
 
        /* Initially set CPU itself as the policy_cpu */
        per_cpu(policy_cpu, cpu) = cpu;
-       lock_policy_rwsem_write(cpu);
+       ret = (lock_policy_rwsem_write(cpu) < 0);
+       WARN_ON(ret);
 
        init_completion(&policy->kobj_unregister);
        INIT_WORK(&policy->update, handle_update);
@@ -833,7 +835,7 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
        ret = cpufreq_driver->init(policy);
        if (ret) {
                dprintk("initialization failed\n");
-               goto err_out;
+               goto err_unlock_policy;
        }
        policy->user_policy.min = policy->min;
        policy->user_policy.max = policy->max;
@@ -858,15 +860,21 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
                /* Check for existing affected CPUs.
                 * They may not be aware of it due to CPU Hotplug.
                 */
-               managed_policy = cpufreq_cpu_get(j);            /* FIXME: Where is this released?  What about error paths? */
+               managed_policy = cpufreq_cpu_get(j);
                if (unlikely(managed_policy)) {
 
                        /* Set proper policy_cpu */
                        unlock_policy_rwsem_write(cpu);
                        per_cpu(policy_cpu, cpu) = managed_policy->cpu;
 
-                       if (lock_policy_rwsem_write(cpu) < 0)
-                               goto err_out_driver_exit;
+                       if (lock_policy_rwsem_write(cpu) < 0) {
+                               /* Should not go through policy unlock path */
+                               if (cpufreq_driver->exit)
+                                       cpufreq_driver->exit(policy);
+                               ret = -EBUSY;
+                               cpufreq_cpu_put(managed_policy);
+                               goto err_free_cpumask;
+                       }
 
                        spin_lock_irqsave(&cpufreq_driver_lock, flags);
                        cpumask_copy(managed_policy->cpus, policy->cpus);
@@ -877,12 +885,14 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
                        ret = sysfs_create_link(&sys_dev->kobj,
                                                &managed_policy->kobj,
                                                "cpufreq");
-                       if (ret)
-                               goto err_out_driver_exit;
-
-                       cpufreq_debug_enable_ratelimit();
-                       ret = 0;
-                       goto err_out_driver_exit; /* call driver->exit() */
+                       if (!ret)
+                               cpufreq_cpu_put(managed_policy);
+                       /*
+                        * Success. We only needed to be added to the mask.
+                        * Call driver->exit() because only the cpu parent of
+                        * the kobj needed to call init().
+                        */
+                       goto out_driver_exit; /* call driver->exit() */
                }
        }
 #endif
@@ -892,25 +902,25 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
        ret = kobject_init_and_add(&policy->kobj, &ktype_cpufreq, &sys_dev->kobj,
                                   "cpufreq");
        if (ret)
-               goto err_out_driver_exit;
+               goto out_driver_exit;
 
        /* set up files for this cpu device */
        drv_attr = cpufreq_driver->attr;
        while ((drv_attr) && (*drv_attr)) {
                ret = sysfs_create_file(&policy->kobj, &((*drv_attr)->attr));
                if (ret)
-                       goto err_out_driver_exit;
+                       goto err_out_kobj_put;
                drv_attr++;
        }
        if (cpufreq_driver->get) {
                ret = sysfs_create_file(&policy->kobj, &cpuinfo_cur_freq.attr);
                if (ret)
-                       goto err_out_driver_exit;
+                       goto err_out_kobj_put;
        }
        if (cpufreq_driver->target) {
                ret = sysfs_create_file(&policy->kobj, &scaling_cur_freq.attr);
                if (ret)
-                       goto err_out_driver_exit;
+                       goto err_out_kobj_put;
        }
 
        spin_lock_irqsave(&cpufreq_driver_lock, flags);
@@ -928,12 +938,14 @@ static int cpufreq_add_dev(struct sys_device *sys_dev)
                        continue;
 
                dprintk("CPU %u already managed, adding link\n", j);
-               cpufreq_cpu_get(cpu);
+               managed_policy = cpufreq_cpu_get(cpu);
                cpu_sys_dev = get_cpu_sysdev(j);
                ret = sysfs_create_link(&cpu_sys_dev->kobj, &policy->kobj,
                                        "cpufreq");
-               if (ret)
+               if (ret) {
+                       cpufreq_cpu_put(managed_policy);
                        goto err_out_unregister;
+               }
        }
 
        policy->governor = NULL; /* to assure that the starting sequence is
@@ -965,17 +977,20 @@ err_out_unregister:
                per_cpu(cpufreq_cpu_data, j) = NULL;
        spin_unlock_irqrestore(&cpufreq_driver_lock, flags);
 
+err_out_kobj_put:
        kobject_put(&policy->kobj);
        wait_for_completion(&policy->kobj_unregister);
 
-err_out_driver_exit:
+out_driver_exit:
        if (cpufreq_driver->exit)
                cpufreq_driver->exit(policy);
 
-err_out:
+err_unlock_policy:
        unlock_policy_rwsem_write(cpu);
+err_free_cpumask:
+       free_cpumask_var(policy->cpus);
+err_free_policy:
        kfree(policy);
-
 nomem_out:
        module_put(cpufreq_driver->owner);
 module_out: