cpufreq: Abort show/store for half initialized policy

PATCH In review Schspa Shi 3 participants 5 messages May 16, 2022 – May 17, 2022 CPU frequency scaling framework

Share & tools

Schspa Shi

author

2022-05-16 03:03 UTC #0 lore raw

PATCH v6 1/2 cpufreq: Abort show/store for half initialized policy

If policy initialization fails after the sysfs files are created, there is a possibility that we may end up running show()/store() callbacks for half initialized policies, which may have unpredictable outcomes.

Abort show/store in such a case by making sure the policy is active. Also inactivate the policy on such failures.

Signed-off-by Schspa Shi <[email protected]>

---

Changelog:
v1 -> v2:
- Fix bad critical region enlarge which causes uninitialized unlock.
- Move cpumask_clear(policy->cpus); before out_offline_policy
v2 -> v3:
- Remove the missed down_write() before
cpumask_and(policy->cpus, policy->cpus, cpu_online_mask);
v3 -> v4:
- Seprate to two patchs.
- Add policy_is_inactive check before sysfs access
v4 -> v5:
- Change the commit message as Viresh advised.
- Initialize ret to -EBUSY to get rid of the else part.
v5 -> v6:
- Change up_write(&policy->rwsem); to after out_free_policy;
- Fixes: https://lore.kernel.org/all/20220515095313.GE10578@xsang-OptiPlex-9020/

Signed-off-by Schspa Shi <[email protected]>

---
 drivers/cpufreq/cpufreq.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

drivers/cpufreq/cpufreq.c +7 -4

index 80f535cc8a75..ba73be6f0490 100644
@@ -947,13 +947,14 @@ static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
 {
 	struct cpufreq_policy *policy = to_policy(kobj);
 	struct freq_attr *fattr = to_attr(attr);
-	ssize_t ret;
+	ssize_t ret = -EBUSY;
 
 	if (!fattr->show)
 		return -EIO;
 
 	down_read(&policy->rwsem);
-	ret = fattr->show(policy, buf);
+	if (likely(!policy_is_inactive(policy)))
+		ret = fattr->show(policy, buf);
 	up_read(&policy->rwsem);
 
 	return ret;

@@ -964,7 +965,7 @@ static ssize_t store(struct kobject *kobj, struct attribute *attr,
 {
 	struct cpufreq_policy *policy = to_policy(kobj);
 	struct freq_attr *fattr = to_attr(attr);
-	ssize_t ret = -EINVAL;
+	ssize_t ret = -EBUSY;
 
 	if (!fattr->store)
 		return -EIO;

@@ -978,7 +979,8 @@ static ssize_t store(struct kobject *kobj, struct attribute *attr,
 
 	if (cpu_online(policy->cpu)) {
 		down_write(&policy->rwsem);
-		ret = fattr->store(policy, buf, count);
+		if (likely(!policy_is_inactive(policy)))
+			ret = fattr->store(policy, buf, count);
 		up_write(&policy->rwsem);
 	}

@@ -1533,6 +1535,7 @@ static int cpufreq_online(unsigned int cpu)
 	for_each_cpu(j, policy->real_cpus)
 		remove_cpu_dev_symlink(policy, get_cpu_device(j));
 
+	cpumask_clear(policy->cpus);
 	up_write(&policy->rwsem);
 
 out_offline_policy:

Signature · 1 line

2.29.0

Schspa Shi

author

2022-05-16 03:12 UTC #1 lore raw

PATCH v6 2/2 cpufreq: make interface functions and lock holding state clear

cpufreq_offline() calls offline() and exit() under the policy rwsem But they are called outside the rwsem in cpufreq_online().

This patch move the offline(), exit(), online(), init() to be inside of policy rwsem to achieve a clear lock relationship.

All the init() online() implement only initialize policy object without holding this lock and won't call cpufreq APIs need to hold this lock.

Signed-off-by Schspa Shi <[email protected]>

---
 drivers/cpufreq/cpufreq.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

drivers/cpufreq/cpufreq.c +3 -3

index ba73be6f0490..ccb774e02934 100644
@@ -1339,12 +1339,12 @@ static int cpufreq_online(unsigned int cpu)
 		down_write(&policy->rwsem);
 		policy->cpu = cpu;
 		policy->governor = NULL;
-		up_write(&policy->rwsem);
 	} else {
 		new_policy = true;
 		policy = cpufreq_policy_alloc(cpu);
 		if (!policy)
 			return -ENOMEM;
+		down_write(&policy->rwsem);
 	}
 
 	if (!new_policy && cpufreq_driver->online) {

@@ -1384,7 +1384,6 @@ static int cpufreq_online(unsigned int cpu)
 		cpumask_copy(policy->related_cpus, policy->cpus);
 	}
 
-	down_write(&policy->rwsem);
 	/*
 	 * affected cpus must always be the one, which are online. We aren't
 	 * managing offline cpus here.

@@ -1536,7 +1535,6 @@ static int cpufreq_online(unsigned int cpu)
 		remove_cpu_dev_symlink(policy, get_cpu_device(j));
 
 	cpumask_clear(policy->cpus);
-	up_write(&policy->rwsem);
 
 out_offline_policy:
 	if (cpufreq_driver->offline)

@@ -1547,6 +1545,8 @@ static int cpufreq_online(unsigned int cpu)
 		cpufreq_driver->exit(policy);
 
 out_free_policy:
+	up_write(&policy->rwsem);
+
 	cpufreq_policy_free(policy);
 	return ret;
 }

Signature · 1 line

2.29.0

Viresh Kumar

maintainer Acked-by on v6 2/2

2022-05-16 10:25 UTC #2 lore raw

Re: [PATCH v6 2/2] cpufreq: make interface functions and lock holding state clear

Quoted 7 lines · Schspa Shi

cpufreq_offline() calls offline() and exit() under the policy rwsem But they are called outside the rwsem in cpufreq_online().
This patch move the offline(), exit(), online(), init() to be inside of policy rwsem to achieve a clear lock relationship.

Quoted 7 lines

All the init() online() implement only initialize policy object without holding this lock and won't call cpufreq APIs need to hold this lock.
Signed-off-by Schspa Shi <[email protected]>
---
 drivers/cpufreq/cpufreq.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

Acked-by Viresh Kumar <[email protected]>

Viresh Kumar

maintainer Acked-by on v6 1/2

2022-05-16 10:25 UTC #3 lore raw

In reply to Schspa Shi

Quoted 10 lines · Schspa Shi

If policy initialization fails after the sysfs files are created, there is a possibility that we may end up running show()/store() callbacks for half initialized policies, which may have unpredictable outcomes.
Abort show/store in such a case by making sure the policy is active. Also inactivate the policy on such failures.
Signed-off-by Schspa Shi <[email protected]>

Acked-by Viresh Kumar <[email protected]>

Rafael J. Wysocki

maintainer

2022-05-17 19:41 UTC #4 lore raw

Quoted 8 lines · Viresh Kumar

Quoted 6 lines · Schspa Shi
If policy initialization fails after the sysfs files are created, there is a possibility that we may end up running show()/store() callbacks for half initialized policies, which may have unpredictable outcomes.

Quoted 6 lines

Quoted 4 lines
Abort show/store in such a case by making sure the policy is active. Also inactivate the policy on such failures.
Signed-off-by Schspa Shi <[email protected]>
Acked-by Viresh Kumar <[email protected]>

Applied along with the [2/2], with some adjustments in the subject and changelog, as 5.19 material.

Thanks!