Skip to contentSkip to conversation

cpufreq: Fix show()/store() issue for hotplugging

PATCH In review 'Guanjun' 2 participants 4 messages Oct 19, 2022 – Oct 21, 2022 CPU frequency scaling framework

lore
Share & tools
lore
author
#1 lore raw

PATCH 1/1 cpufreq: Fix show()/store() issue for hotplugging offline CPU

In reply to 'Guanjun'

From: Zelin Deng <[email protected]>

After brought one CPU offline, lscpu returned failure:

lscpu: cannot read /sys/devices/system/cpu/cpu64/cpufreq/cpuinfo_max_freq: Device or resource busy

which had blocked all outputs of lscpu.

This is not the case mentioned in commit d4627a287e251, as the policy had been created successfully but is inactive due to CPU gets offline. To fix this issue, just add an addtional check whether CPU is online or not.

Fixes d4627a287e251 ("cpufreq: Abort show()/store() for half-initialized policies")
---
 drivers/cpufreq/cpufreq.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)
drivers/cpufreq/cpufreq.c +12 -4
index 69b3d61852ac..aa238ba7d2fe 100644
@@ -956,8 +956,12 @@ static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
 		return -EIO;
 
 	down_read(&policy->rwsem);
-	if (likely(!policy_is_inactive(policy)))
-		ret = fattr->show(policy, buf);
+	if (unlikely(policy_is_inactive(policy) && cpu_online(policy->cpu)))
+		goto err;
+
+	ret = fattr->show(policy, buf);
+
+err:
 	up_read(&policy->rwsem);
 
 	return ret;
@@ -974,8 +978,12 @@ static ssize_t store(struct kobject *kobj, struct attribute *attr,
 		return -EIO;
 
 	down_write(&policy->rwsem);
-	if (likely(!policy_is_inactive(policy)))
-		ret = fattr->store(policy, buf, count);
+	if (unlikely(policy_is_inactive(policy) && cpu_online(policy->cpu)))
+		goto err;
+
+	ret = fattr->store(policy, buf, count);
+
+err:
 	up_write(&policy->rwsem);
 
 	return ret;
Signature · 1 line
2.32.0.GIT

'Guanjun'

author
#0 lore raw

cover cpufreq: Fix show()/store() issue for hotplugging

From: Zelin Deng <[email protected]>

Hello maintainers,

After I brought one CPU (64) offline, I got failure when I used lscpu:
lscpu: cannot read /sys/devices/system/cpu/cpu64/cpufreq/cpuinfo_max_freq: Device or resource busy

I found in commit d4627a287e251, policy_is_inactive() check was added to avoid the case in which policy initialization failed. However it didn't consider the situation in which policy has been created successfully but is not inactive: CPU is hotplugging offline.

This patch just add an addtional check whether CPU is online or not to fix the issue.

Zelin Deng (1):
cpufreq: Fix show()/store() issue for hotplugging offline CPU

 drivers/cpufreq/cpufreq.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)
Signature · 1 line
2.32.0.GIT
#2 lore raw

Re: [PATCH 1/1] cpufreq: Fix show()/store() issue for hotplugging offline CPU

In reply to 'Guanjun'
Quoted 8 lines · Guanjun

From: Zelin Deng <[email protected]>

After brought one CPU offline, lscpu returned failure:

lscpu: cannot read /sys/devices/system/cpu/cpu64/cpufreq/cpuinfo_max_freq: Device or resource busy

which had blocked all outputs of lscpu.

OK, so the policy->cpus mask is empty and -EBUSY is returned.

What's wrong?

Quoted 2 lines

This is not the case mentioned in commit d4627a287e251, as the policy had been created successfully but is inactive due to CPU gets offline.

Yes, that's when policy_is_inactive(policy) returns "true" IIUC.

Quoted 2 lines

To fix this issue, just add an addtional check whether CPU is online or not.

Which is racy.

Please explain the problem in the first place.

Quoted 44 lines
Fixes d4627a287e251 ("cpufreq: Abort show()/store() for half-initialized policies")
---
 drivers/cpufreq/cpufreq.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)
drivers/cpufreq/cpufreq.c +12 -5
index 69b3d61852ac..aa238ba7d2fe 100644
@@ -956,8 +956,12 @@ static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
                return -EIO;
 
        down_read(&policy->rwsem);
-       if (likely(!policy_is_inactive(policy)))
-               ret = fattr->show(policy, buf);
+       if (unlikely(policy_is_inactive(policy) && cpu_online(policy->cpu)))
+               goto err;
+
+       ret = fattr->show(policy, buf);
+
+err:
        up_read(&policy->rwsem);
 
        return ret;
@@ -974,8 +978,12 @@ static ssize_t store(struct kobject *kobj, struct attribute *attr,
                return -EIO;
 
        down_write(&policy->rwsem);
-       if (likely(!policy_is_inactive(policy)))
-               ret = fattr->store(policy, buf, count);
+       if (unlikely(policy_is_inactive(policy) && cpu_online(policy->cpu)))
+               goto err;
+
+       ret = fattr->store(policy, buf, count);
+
+err:
        up_write(&policy->rwsem);
 
        return ret;
--

2.32.0.GIT

author
#3 lore raw

Re: [PATCH 1/1] cpufreq: Fix show()/store() issue for hotplugging offline CPU

Hi Rafael,

Quoted 8 lines

2022年10月19日 下午7:47,Rafael J. Wysocki <[email protected]> 写道:

Quoted 5 lines · Guanjun

From: Zelin Deng <[email protected]>

After brought one CPU offline, lscpu returned failure:

Quoted 7 lines
Quoted 3 lines

lscpu: cannot read /sys/devices/system/cpu/cpu64/cpufreq/cpuinfo_max_freq: Device or resource busy

which had blocked all outputs of lscpu.

OK, so the policy->cpus mask is empty and -EBUSY is returned.

What's wrong?

Here is all right.
The problem is that when I offline one cpu manually and lscpu will fail.
The reproduce process is as follows:
1. lscpu (success)
2. echo 0 > /sys/devices/system/cpu/cpu63/online (offline cpu63)
3. lscpu (fail, and print the error message, “lscpu: cannot read /sys/devices/system/cpu/cpu64/cpufreq/cpuinfo_max_freq: Device or resource busy”)

I think this failure doesn’t make sense.

Maybe I should make the commit message more readable.

Thanks,
Guanjun

Quoted 56 lines
Quoted 3 lines

This is not the case mentioned in commit d4627a287e251, as the policy had been created successfully but is inactive due to CPU gets offline.

Yes, that's when policy_is_inactive(policy) returns "true" IIUC.

Quoted 2 lines

To fix this issue, just add an addtional check whether CPU is online or not.

Which is racy.

Please explain the problem in the first place.

Quoted 43 lines
Fixes d4627a287e251 ("cpufreq: Abort show()/store() for half-initialized policies")

---
drivers/cpufreq/cpufreq.c | 16 ++++++++++++----

1 file changed, 12 insertions(+), 4 deletions(-)
drivers/cpufreq/cpufreq.c +12 -5
index 69b3d61852ac..aa238ba7d2fe 100644
@@ -956,8 +956,12 @@ static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
               return -EIO;
 
       down_read(&policy->rwsem);
-       if (likely(!policy_is_inactive(policy)))
-               ret = fattr->show(policy, buf);
+       if (unlikely(policy_is_inactive(policy) && cpu_online(policy->cpu)))
+               goto err;
+
+       ret = fattr->show(policy, buf);
+
+err:
       up_read(&policy->rwsem);
 
       return ret;
@@ -974,8 +978,12 @@ static ssize_t store(struct kobject *kobj, struct attribute *attr,
               return -EIO;
 
       down_write(&policy->rwsem);
-       if (likely(!policy_is_inactive(policy)))
-               ret = fattr->store(policy, buf, count);
+       if (unlikely(policy_is_inactive(policy) && cpu_online(policy->cpu)))
+               goto err;
+
+       ret = fattr->store(policy, buf, count);
+
+err:
       up_write(&policy->rwsem);
 
       return ret;
--

2.32.0.GIT