Skip to contentSkip to conversation

ACPI: platform-profile: Fix possible deadlock in platform_profile_remove()

PATCH In review Hans de Goede 3 participants 4 messages Jan 25, 2021 – Jan 27, 2021

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover ACPI: platform-profile: Fix possible deadlock in platform_profile_remove()

Hi Rafael,

Sorry for throwing another patch into the mix for the acpi-platform branch. I always build my local kernels with lockdep enabled and while testing an unrelated acpi_thinkpad patch I did "rmmod acpi_thinkpad" and lockdep pointed out this potential deadlock.

The fix is simple enough, although it does rely on the assumption that drivers will only call platform_profile_remove() after they have first successfully called platform_profile_register().

See the actual patch for more details.

Regards,

Hans

Hans de Goede (1):
ACPI: platform-profile: Fix possible deadlock in platform_profile_remove()

 drivers/acpi/platform_profile.c | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)
#1 lore raw

PATCH 1/1 ACPI: platform-profile: Fix possible deadlock in platform_profile_remove()

After a rmmod thinkpad_acpi, lockdep pointed out this possible deadlock:

Our _show and _store sysfs attr functions get called with the kn->active lock held for the sysfs attr and then take the profile_lock. sysfs_remove_group() also takes the kn->active lock for the sysfs attr, so if we call it with the profile_lock held, then we get an ABBA deadlock.

platform_profile_remove() must only be called by drivers which have first *successfully* called platform_profile_register(). Anything else is a driver bug. So the check for cur_profile being set before calling sysfs_remove_group() is not necessary and it can be dropped.

It is safe to call sysfs_remove_group() without holding the profile_lock since the attr-group group cannot be re-added until after we clear cur_profile.

Change platform_profile_remove() to only hold the profile_lock while clearing the cur_profile, fixing the deadlock.

---
 drivers/acpi/platform_profile.c | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)
drivers/acpi/platform_profile.c +2 -6
index 80e9df427eb8..4a59c5993bde 100644
@@ -164,13 +164,9 @@ EXPORT_SYMBOL_GPL(platform_profile_register);
 
 int platform_profile_remove(void)
 {
-	mutex_lock(&profile_lock);
-	if (!cur_profile) {
-		mutex_unlock(&profile_lock);
-		return -ENODEV;
-	}
-
 	sysfs_remove_group(acpi_kobj, &platform_profile_group);
+
+	mutex_lock(&profile_lock);
 	cur_profile = NULL;
 	mutex_unlock(&profile_lock);
 	return 0;
Signature · 1 line
2.29.2
#2 lore raw
Quoted 49 lines

-----Original Message-----
From: Hans de Goede <[email protected]>
Sent: Monday, January 25, 2021 2:09 PM

After a rmmod thinkpad_acpi, lockdep pointed out this possible deadlock:

Our _show and _store sysfs attr functions get called with the kn->active lock held for the sysfs attr and then take the profile_lock. sysfs_remove_group() also takes the kn->active lock for the sysfs attr, so if we call it with the profile_lock held, then we get an ABBA deadlock.

platform_profile_remove() must only be called by drivers which have first *successfully* called platform_profile_register(). Anything else is a driver bug. So the check for cur_profile being set before calling sysfs_remove_group() is not necessary and it can be dropped.

It is safe to call sysfs_remove_group() without holding the profile_lock since the attr-group group cannot be re-added until after we clear cur_profile.

Change platform_profile_remove() to only hold the profile_lock while clearing the cur_profile, fixing the deadlock.

---
 drivers/acpi/platform_profile.c | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)
drivers/acpi/platform_profile.c +2 -7
index 80e9df427eb8..4a59c5993bde 100644
@@ -164,13 +164,9 @@ EXPORT_SYMBOL_GPL(platform_profile_register);
 
 int platform_profile_remove(void)
 {
-	mutex_lock(&profile_lock);
-	if (!cur_profile) {
-		mutex_unlock(&profile_lock);
-		return -ENODEV;
-	}
-
 	sysfs_remove_group(acpi_kobj, &platform_profile_group);
+
+	mutex_lock(&profile_lock);
 	cur_profile = NULL;
 	mutex_unlock(&profile_lock);
 	return 0;
--
Quoted 1 line

2.29.2

Thanks Hans - good catch! Looks good to me.
Mark

#3 lore raw
In reply to Hans de Goede
Quoted 13 lines · Hans de Goede

After a rmmod thinkpad_acpi, lockdep pointed out this possible deadlock:

Our _show and _store sysfs attr functions get called with the kn->active lock held for the sysfs attr and then take the profile_lock. sysfs_remove_group() also takes the kn->active lock for the sysfs attr, so if we call it with the profile_lock held, then we get an ABBA deadlock.

platform_profile_remove() must only be called by drivers which have first *successfully* called platform_profile_register(). Anything else is a driver bug. So the check for cur_profile being set before calling sysfs_remove_group() is not necessary and it can be dropped.

Quoted 8 lines

It is safe to call sysfs_remove_group() without holding the profile_lock since the attr-group group cannot be re-added until after we clear cur_profile.

Change platform_profile_remove() to only hold the profile_lock while clearing the cur_profile, fixing the deadlock.

Applied on top of the previous platform-profile material, thanks!

Quoted 27 lines
---
 drivers/acpi/platform_profile.c | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)
drivers/acpi/platform_profile.c +2 -7
index 80e9df427eb8..4a59c5993bde 100644
@@ -164,13 +164,9 @@ EXPORT_SYMBOL_GPL(platform_profile_register);
 
 int platform_profile_remove(void)
 {
-       mutex_lock(&profile_lock);
-       if (!cur_profile) {
-               mutex_unlock(&profile_lock);
-               return -ENODEV;
-       }
-
        sysfs_remove_group(acpi_kobj, &platform_profile_group);
+
+       mutex_lock(&profile_lock);
        cur_profile = NULL;
        mutex_unlock(&profile_lock);
        return 0;
--

2.29.2