Skip to contentSkip to conversation

cpufreq: fix locking order in store_local_boost to prevent deadlock

PATCH In review Seyediman Seyedarab 3 participants 6 messages Apr 30, 2025 – May 10, 2025 CPU frequency scaling framework

lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 cpufreq: fix locking order in store_local_boost to prevent deadlock

Lockdep reports a possible circular locking dependency[1] when writing to /sys/devices/system/cpu/cpufreq/policyN/boost, triggered by power-profiles-daemon at boot.

store_local_boost() used to acquire cpu_hotplug_lock *after* the policy lock had already been taken by the store() handler. However, the expected locking hierarchy is to acquire cpu_hotplug_lock before the policy guard. This inverted lock order creates a *theoretical* deadlock possibility.

Acquire cpu_hotplug_lock in the store() handler *only* for the local_boost attribute, before entering the policy guard block, and remove the cpus_read_lock/unlock() calls from store_local_boost(). Also switch from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

[1]
======================================================
WARNING: possible circular locking dependency detected 6.15.0-rc4-debug #28 Not tainted
------------------------------------------------------
power-profiles-/596 is trying to acquire lock:
ffffffffb147e910 (cpu_hotplug_lock){++++}-{0:0}, at: store_local_boost+0x6a/0xd0

but task is already holding lock:
ffff9eaa48377b80 (&policy->rwsem){++++}-{4:4}, at: store+0x37/0x90

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #2 (&policy->rwsem){++++}-{4:4}:

        down_write+0x29/0xb0
        cpufreq_online+0x841/0xa00
        cpufreq_add_dev+0x71/0x80
        subsys_interface_register+0x14b/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #1 (subsys mutex#3){+.+.}-{4:4}:

        __mutex_lock+0xc2/0x930
        subsys_interface_register+0x83/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #0 (cpu_hotplug_lock){++++}-{0:0}:

        __lock_acquire+0x1087/0x17e0
        lock_acquire.part.0+0x66/0x1b0
        cpus_read_lock+0x2a/0xc0
        store_local_boost+0x6a/0xd0
        store+0x50/0x90
        kernfs_fop_write_iter+0x135/0x200
        vfs_write+0x2ab/0x540
        ksys_write+0x6c/0xe0
        do_syscall_64+0xbb/0x1d0
        entry_SYSCALL_64_after_hwframe+0x56/0x5e

---
Changes in v3:
- Rebased over PM tree's linux-next branch
- Added a comment to explain why this piece of code is required
- Switched from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

Changes in v2:
- Restrict cpu_hotplug_lock acquisition to only
the local_boost attribute in store() handler.

Regards,
Seyediman

 drivers/cpufreq/cpufreq.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)
drivers/cpufreq/cpufreq.c +16 -7
index 21fa733a2..b349adbeb 100644
@@ -622,10 +622,7 @@ static ssize_t store_local_boost(struct cpufreq_policy *policy,
 	if (!policy->boost_supported)
 		return -EINVAL;
 
-	cpus_read_lock();
 	ret = policy_set_boost(policy, enable);
-	cpus_read_unlock();
-
 	if (!ret)
 		return count;
 
@@ -1006,16 +1003,28 @@ static ssize_t store(struct kobject *kobj, struct attribute *attr,
 {
 	struct cpufreq_policy *policy = to_policy(kobj);
 	struct freq_attr *fattr = to_attr(attr);
+	int ret = -EBUSY;
 
 	if (!fattr->store)
 		return -EIO;
 
-	guard(cpufreq_policy_write)(policy);
+	/*
+	 * store_local_boost() requires cpu_hotplug_lock to be held, and must be
+	 * called with that lock acquired *before* taking policy->rwsem to avoid
+	 * lock ordering violations.
+	 */
+	if (fattr == &local_boost)
+		cpus_read_lock();
 
-	if (likely(!policy_is_inactive(policy)))
-		return fattr->store(policy, buf, count);
+	scoped_guard(cpufreq_policy_write, policy) {
+		if (likely(!policy_is_inactive(policy)))
+			ret = fattr->store(policy, buf, count);
+	}
 
-	return -EBUSY;
+	if (fattr == &local_boost)
+		cpus_read_unlock();
+
+	return ret;
 }
 
 static void cpufreq_sysfs_release(struct kobject *kobj)
Signature · 1 line
2.49.0
maintainer Acked-by on v3 1/1
#1 lore raw
Quoted 123 lines · Seyediman Seyedarab

Lockdep reports a possible circular locking dependency[1] when writing to /sys/devices/system/cpu/cpufreq/policyN/boost, triggered by power-profiles-daemon at boot.

store_local_boost() used to acquire cpu_hotplug_lock *after* the policy lock had already been taken by the store() handler. However, the expected locking hierarchy is to acquire cpu_hotplug_lock before the policy guard. This inverted lock order creates a *theoretical* deadlock possibility.

Acquire cpu_hotplug_lock in the store() handler *only* for the local_boost attribute, before entering the policy guard block, and remove the cpus_read_lock/unlock() calls from store_local_boost(). Also switch from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

[1]
======================================================
WARNING: possible circular locking dependency detected 6.15.0-rc4-debug #28 Not tainted
------------------------------------------------------
power-profiles-/596 is trying to acquire lock:
ffffffffb147e910 (cpu_hotplug_lock){++++}-{0:0}, at: store_local_boost+0x6a/0xd0

but task is already holding lock:
ffff9eaa48377b80 (&policy->rwsem){++++}-{4:4}, at: store+0x37/0x90

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #2 (&policy->rwsem){++++}-{4:4}:

        down_write+0x29/0xb0
        cpufreq_online+0x841/0xa00
        cpufreq_add_dev+0x71/0x80
        subsys_interface_register+0x14b/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #1 (subsys mutex#3){+.+.}-{4:4}:

        __mutex_lock+0xc2/0x930
        subsys_interface_register+0x83/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #0 (cpu_hotplug_lock){++++}-{0:0}:

        __lock_acquire+0x1087/0x17e0
        lock_acquire.part.0+0x66/0x1b0
        cpus_read_lock+0x2a/0xc0
        store_local_boost+0x6a/0xd0
        store+0x50/0x90
        kernfs_fop_write_iter+0x135/0x200
        vfs_write+0x2ab/0x540
        ksys_write+0x6c/0xe0
        do_syscall_64+0xbb/0x1d0
        entry_SYSCALL_64_after_hwframe+0x56/0x5e

---
Changes in v3:
- Rebased over PM tree's linux-next branch
- Added a comment to explain why this piece of code is required
- Switched from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

Changes in v2:
- Restrict cpu_hotplug_lock acquisition to only
the local_boost attribute in store() handler.

Regards,
Seyediman

 drivers/cpufreq/cpufreq.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)
drivers/cpufreq/cpufreq.c +9 -6
index 21fa733a2..b349adbeb 100644
@@ -622,10 +622,7 @@ static ssize_t store_local_boost(struct cpufreq_policy *policy,
 	if (!policy->boost_supported)
 		return -EINVAL;
 
-	cpus_read_lock();
 	ret = policy_set_boost(policy, enable);
-	cpus_read_unlock();
-
 	if (!ret)
 		return count;
 
@@ -1006,16 +1003,28 @@ static ssize_t store(struct kobject *kobj, struct attribute *attr,
 {
 	struct cpufreq_policy *policy = to_policy(kobj);
 	struct freq_attr *fattr = to_attr(attr);
+	int ret = -EBUSY;
 
 	if (!fattr->store)
 		return -EIO;
 
-	guard(cpufreq_policy_write)(policy);
+	/*
+	 * store_local_boost() requires cpu_hotplug_lock to be held, and must be
+	 * called with that lock acquired *before* taking policy->rwsem to avoid
+	 * lock ordering violations.
+	 */
+	if (fattr == &local_boost)
+		cpus_read_lock();
 
-	if (likely(!policy_is_inactive(policy)))
-		return fattr->store(policy, buf, count);
+	scoped_guard(cpufreq_policy_write, policy) {

drivers/cpufreq/cpufreq.c static ssize_t store(struct kobject *kobj, struct attribute *attr, · line 1028

Quoted 10 lines
drivers/cpufreq/cpufreq.c +7 -1
+		if (likely(!policy_is_inactive(policy)))
+			ret = fattr->store(policy, buf, count);
+	}
 
-	return -EBUSY;
+	if (fattr == &local_boost)
+		cpus_read_unlock();
+
+	return ret;
 }
Signature · 1 line
viresh
#2 lore raw
Quoted 128 lines · Viresh Kumar
Quoted 127 lines · Seyediman Seyedarab

Lockdep reports a possible circular locking dependency[1] when writing to /sys/devices/system/cpu/cpufreq/policyN/boost, triggered by power-profiles-daemon at boot.

store_local_boost() used to acquire cpu_hotplug_lock *after* the policy lock had already been taken by the store() handler. However, the expected locking hierarchy is to acquire cpu_hotplug_lock before the policy guard. This inverted lock order creates a *theoretical* deadlock possibility.

Acquire cpu_hotplug_lock in the store() handler *only* for the local_boost attribute, before entering the policy guard block, and remove the cpus_read_lock/unlock() calls from store_local_boost(). Also switch from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

[1]
======================================================
WARNING: possible circular locking dependency detected 6.15.0-rc4-debug #28 Not tainted
------------------------------------------------------
power-profiles-/596 is trying to acquire lock:
ffffffffb147e910 (cpu_hotplug_lock){++++}-{0:0}, at: store_local_boost+0x6a/0xd0

but task is already holding lock:
ffff9eaa48377b80 (&policy->rwsem){++++}-{4:4}, at: store+0x37/0x90

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #2 (&policy->rwsem){++++}-{4:4}:

        down_write+0x29/0xb0
        cpufreq_online+0x841/0xa00
        cpufreq_add_dev+0x71/0x80
        subsys_interface_register+0x14b/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #1 (subsys mutex#3){+.+.}-{4:4}:

        __mutex_lock+0xc2/0x930
        subsys_interface_register+0x83/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #0 (cpu_hotplug_lock){++++}-{0:0}:

        __lock_acquire+0x1087/0x17e0
        lock_acquire.part.0+0x66/0x1b0
        cpus_read_lock+0x2a/0xc0
        store_local_boost+0x6a/0xd0
        store+0x50/0x90
        kernfs_fop_write_iter+0x135/0x200
        vfs_write+0x2ab/0x540
        ksys_write+0x6c/0xe0
        do_syscall_64+0xbb/0x1d0
        entry_SYSCALL_64_after_hwframe+0x56/0x5e

---
Changes in v3:
- Rebased over PM tree's linux-next branch
- Added a comment to explain why this piece of code is required
- Switched from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

Changes in v2:
- Restrict cpu_hotplug_lock acquisition to only
the local_boost attribute in store() handler.

Regards,
Seyediman

 drivers/cpufreq/cpufreq.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)
drivers/cpufreq/cpufreq.c +12 -6
index 21fa733a2..b349adbeb 100644
@@ -622,10 +622,7 @@ static ssize_t store_local_boost(struct cpufreq_policy *policy,
 	if (!policy->boost_supported)
 		return -EINVAL;
 
-	cpus_read_lock();
 	ret = policy_set_boost(policy, enable);
-	cpus_read_unlock();
-
 	if (!ret)
 		return count;
 
@@ -1006,16 +1003,28 @@ static ssize_t store(struct kobject *kobj, struct attribute *attr,
 {
 	struct cpufreq_policy *policy = to_policy(kobj);
 	struct freq_attr *fattr = to_attr(attr);
+	int ret = -EBUSY;
 
 	if (!fattr->store)
 		return -EIO;
 
-	guard(cpufreq_policy_write)(policy);
+	/*
+	 * store_local_boost() requires cpu_hotplug_lock to be held, and must be
+	 * called with that lock acquired *before* taking policy->rwsem to avoid
+	 * lock ordering violations.
+	 */
+	if (fattr == &local_boost)
+		cpus_read_lock();
 
-	if (likely(!policy_is_inactive(policy)))
-		return fattr->store(policy, buf, count);
+	scoped_guard(cpufreq_policy_write, policy) {
+		if (likely(!policy_is_inactive(policy)))
+			ret = fattr->store(policy, buf, count);
+	}
 
Quoted 11 lines

drivers/cpufreq/cpufreq.c static ssize_t store(struct kobject *kobj, struct attribute *attr, · line 1028

Quoted 6 lines
drivers/cpufreq/cpufreq.c +4 -1
-	return -EBUSY;
+	if (fattr == &local_boost)
+		cpus_read_unlock();
+
+	return ret;
 }
Signature · 1 line
viresh

Hi there,

Just following up to see if there's anything you'd like me to change or address in the patch before it can move forward. Please let me know if any updates are needed.

Regards,
Seyediman

#3 lore raw
Quoted 139 lines · Seyediman Seyedarab
Quoted 137 lines · Viresh Kumar
Quoted 123 lines · Seyediman Seyedarab

Lockdep reports a possible circular locking dependency[1] when writing to /sys/devices/system/cpu/cpufreq/policyN/boost, triggered by power-profiles-daemon at boot.

store_local_boost() used to acquire cpu_hotplug_lock *after* the policy lock had already been taken by the store() handler. However, the expected locking hierarchy is to acquire cpu_hotplug_lock before the policy guard. This inverted lock order creates a *theoretical* deadlock possibility.

Acquire cpu_hotplug_lock in the store() handler *only* for the local_boost attribute, before entering the policy guard block, and remove the cpus_read_lock/unlock() calls from store_local_boost(). Also switch from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

[1]
======================================================
WARNING: possible circular locking dependency detected 6.15.0-rc4-debug #28 Not tainted
------------------------------------------------------
power-profiles-/596 is trying to acquire lock:
ffffffffb147e910 (cpu_hotplug_lock){++++}-{0:0}, at: store_local_boost+0x6a/0xd0

but task is already holding lock:
ffff9eaa48377b80 (&policy->rwsem){++++}-{4:4}, at: store+0x37/0x90

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #2 (&policy->rwsem){++++}-{4:4}:

        down_write+0x29/0xb0
        cpufreq_online+0x841/0xa00
        cpufreq_add_dev+0x71/0x80
        subsys_interface_register+0x14b/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #1 (subsys mutex#3){+.+.}-{4:4}:

        __mutex_lock+0xc2/0x930
        subsys_interface_register+0x83/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #0 (cpu_hotplug_lock){++++}-{0:0}:

        __lock_acquire+0x1087/0x17e0
        lock_acquire.part.0+0x66/0x1b0
        cpus_read_lock+0x2a/0xc0
        store_local_boost+0x6a/0xd0
        store+0x50/0x90
        kernfs_fop_write_iter+0x135/0x200
        vfs_write+0x2ab/0x540
        ksys_write+0x6c/0xe0
        do_syscall_64+0xbb/0x1d0
        entry_SYSCALL_64_after_hwframe+0x56/0x5e

---
Changes in v3:
- Rebased over PM tree's linux-next branch
- Added a comment to explain why this piece of code is required
- Switched from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

Changes in v2:
- Restrict cpu_hotplug_lock acquisition to only
the local_boost attribute in store() handler.

Regards,
Seyediman

 drivers/cpufreq/cpufreq.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)
drivers/cpufreq/cpufreq.c +9 -6
index 21fa733a2..b349adbeb 100644
@@ -622,10 +622,7 @@ static ssize_t store_local_boost(struct cpufreq_policy *policy,
    if (!policy->boost_supported)
            return -EINVAL;
 
-   cpus_read_lock();
    ret = policy_set_boost(policy, enable);
-   cpus_read_unlock();
-
    if (!ret)
            return count;
 
@@ -1006,16 +1003,28 @@ static ssize_t store(struct kobject *kobj, struct attribute *attr,
 {
    struct cpufreq_policy *policy = to_policy(kobj);
    struct freq_attr *fattr = to_attr(attr);
+   int ret = -EBUSY;
 
    if (!fattr->store)
            return -EIO;
 
-   guard(cpufreq_policy_write)(policy);
+   /*
+    * store_local_boost() requires cpu_hotplug_lock to be held, and must be
+    * called with that lock acquired *before* taking policy->rwsem to avoid
+    * lock ordering violations.
+    */
+   if (fattr == &local_boost)
+           cpus_read_lock();
 
-   if (likely(!policy_is_inactive(policy)))
-           return fattr->store(policy, buf, count);
+   scoped_guard(cpufreq_policy_write, policy) {

drivers/cpufreq/cpufreq.c static ssize_t store(struct kobject *kobj, struct attribute *attr, · line 1028

Quoted 10 lines
drivers/cpufreq/cpufreq.c +7 -1
+           if (likely(!policy_is_inactive(policy)))
+                   ret = fattr->store(policy, buf, count);
+   }
 
-   return -EBUSY;
+   if (fattr == &local_boost)
+           cpus_read_unlock();
+
+   return ret;
 }
Quoted 8 lines
Quoted 2 lines

--
viresh

Hi there,

Just following up to see if there's anything you'd like me to change or address in the patch before it can move forward. Please let me know if any updates are needed.

I'm kind of wondering why local_boost needs cpus_read_lock() at all. Holding the policy rwsem blocks CPU online/offline already for this policy.

Is that because ->set_boost() may need to synchronize with the other policies?

#4 lore raw
Quoted 150 lines · Rafael J. Wysocki
Quoted 139 lines · Seyediman Seyedarab
Quoted 137 lines · Viresh Kumar
Quoted 123 lines · Seyediman Seyedarab

Lockdep reports a possible circular locking dependency[1] when writing to /sys/devices/system/cpu/cpufreq/policyN/boost, triggered by power-profiles-daemon at boot.

store_local_boost() used to acquire cpu_hotplug_lock *after* the policy lock had already been taken by the store() handler. However, the expected locking hierarchy is to acquire cpu_hotplug_lock before the policy guard. This inverted lock order creates a *theoretical* deadlock possibility.

Acquire cpu_hotplug_lock in the store() handler *only* for the local_boost attribute, before entering the policy guard block, and remove the cpus_read_lock/unlock() calls from store_local_boost(). Also switch from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

[1]
======================================================
WARNING: possible circular locking dependency detected 6.15.0-rc4-debug #28 Not tainted
------------------------------------------------------
power-profiles-/596 is trying to acquire lock:
ffffffffb147e910 (cpu_hotplug_lock){++++}-{0:0}, at: store_local_boost+0x6a/0xd0

but task is already holding lock:
ffff9eaa48377b80 (&policy->rwsem){++++}-{4:4}, at: store+0x37/0x90

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #2 (&policy->rwsem){++++}-{4:4}:

        down_write+0x29/0xb0
        cpufreq_online+0x841/0xa00
        cpufreq_add_dev+0x71/0x80
        subsys_interface_register+0x14b/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #1 (subsys mutex#3){+.+.}-{4:4}:

        __mutex_lock+0xc2/0x930
        subsys_interface_register+0x83/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #0 (cpu_hotplug_lock){++++}-{0:0}:

        __lock_acquire+0x1087/0x17e0
        lock_acquire.part.0+0x66/0x1b0
        cpus_read_lock+0x2a/0xc0
        store_local_boost+0x6a/0xd0
        store+0x50/0x90
        kernfs_fop_write_iter+0x135/0x200
        vfs_write+0x2ab/0x540
        ksys_write+0x6c/0xe0
        do_syscall_64+0xbb/0x1d0
        entry_SYSCALL_64_after_hwframe+0x56/0x5e

---
Changes in v3:
- Rebased over PM tree's linux-next branch
- Added a comment to explain why this piece of code is required
- Switched from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

Changes in v2:
- Restrict cpu_hotplug_lock acquisition to only
the local_boost attribute in store() handler.

Regards,
Seyediman

 drivers/cpufreq/cpufreq.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)
drivers/cpufreq/cpufreq.c +9 -6
index 21fa733a2..b349adbeb 100644
@@ -622,10 +622,7 @@ static ssize_t store_local_boost(struct cpufreq_policy *policy,
    if (!policy->boost_supported)
            return -EINVAL;
 
-   cpus_read_lock();
    ret = policy_set_boost(policy, enable);
-   cpus_read_unlock();
-
    if (!ret)
            return count;
 
@@ -1006,16 +1003,28 @@ static ssize_t store(struct kobject *kobj, struct attribute *attr,
 {
    struct cpufreq_policy *policy = to_policy(kobj);
    struct freq_attr *fattr = to_attr(attr);
+   int ret = -EBUSY;
 
    if (!fattr->store)
            return -EIO;
 
-   guard(cpufreq_policy_write)(policy);
+   /*
+    * store_local_boost() requires cpu_hotplug_lock to be held, and must be
+    * called with that lock acquired *before* taking policy->rwsem to avoid
+    * lock ordering violations.
+    */
+   if (fattr == &local_boost)
+           cpus_read_lock();
 
-   if (likely(!policy_is_inactive(policy)))
-           return fattr->store(policy, buf, count);
+   scoped_guard(cpufreq_policy_write, policy) {

drivers/cpufreq/cpufreq.c static ssize_t store(struct kobject *kobj, struct attribute *attr, · line 1028

Quoted 10 lines
drivers/cpufreq/cpufreq.c +7 -1
+           if (likely(!policy_is_inactive(policy)))
+                   ret = fattr->store(policy, buf, count);
+   }
 
-   return -EBUSY;
+   if (fattr == &local_boost)
+           cpus_read_unlock();
+
+   return ret;
 }
Quoted 8 lines
Quoted 2 lines

--
viresh

Hi there,

Just following up to see if there's anything you'd like me to change or address in the patch before it can move forward. Please let me know if any updates are needed.

Quoted 5 lines

I'm kind of wondering why local_boost needs cpus_read_lock() at all. Holding the policy rwsem blocks CPU online/offline already for this policy.

Is that because ->set_boost() may need to synchronize with the other policies?

IOW, what can go wrong if the cpus_read_lock() locking is dropped from there altogether?

#5 lore raw
Quoted 151 lines · Rafael J. Wysocki
Quoted 150 lines · Rafael J. Wysocki
Quoted 147 lines · Seyediman Seyedarab
Quoted 137 lines · Viresh Kumar
Quoted 123 lines · Seyediman Seyedarab

Lockdep reports a possible circular locking dependency[1] when writing to /sys/devices/system/cpu/cpufreq/policyN/boost, triggered by power-profiles-daemon at boot.

store_local_boost() used to acquire cpu_hotplug_lock *after* the policy lock had already been taken by the store() handler. However, the expected locking hierarchy is to acquire cpu_hotplug_lock before the policy guard. This inverted lock order creates a *theoretical* deadlock possibility.

Acquire cpu_hotplug_lock in the store() handler *only* for the local_boost attribute, before entering the policy guard block, and remove the cpus_read_lock/unlock() calls from store_local_boost(). Also switch from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

[1]
======================================================
WARNING: possible circular locking dependency detected 6.15.0-rc4-debug #28 Not tainted
------------------------------------------------------
power-profiles-/596 is trying to acquire lock:
ffffffffb147e910 (cpu_hotplug_lock){++++}-{0:0}, at: store_local_boost+0x6a/0xd0

but task is already holding lock:
ffff9eaa48377b80 (&policy->rwsem){++++}-{4:4}, at: store+0x37/0x90

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #2 (&policy->rwsem){++++}-{4:4}:

        down_write+0x29/0xb0
        cpufreq_online+0x841/0xa00
        cpufreq_add_dev+0x71/0x80
        subsys_interface_register+0x14b/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #1 (subsys mutex#3){+.+.}-{4:4}:

        __mutex_lock+0xc2/0x930
        subsys_interface_register+0x83/0x170
        cpufreq_register_driver+0x154/0x250
        amd_pstate_register_driver+0x36/0x70
        amd_pstate_init+0x1e7/0x270
        do_one_initcall+0x67/0x2c0
        kernel_init_freeable+0x230/0x270
        kernel_init+0x15/0x130
        ret_from_fork+0x2c/0x50
        ret_from_fork_asm+0x11/0x20

-> #0 (cpu_hotplug_lock){++++}-{0:0}:

        __lock_acquire+0x1087/0x17e0
        lock_acquire.part.0+0x66/0x1b0
        cpus_read_lock+0x2a/0xc0
        store_local_boost+0x6a/0xd0
        store+0x50/0x90
        kernfs_fop_write_iter+0x135/0x200
        vfs_write+0x2ab/0x540
        ksys_write+0x6c/0xe0
        do_syscall_64+0xbb/0x1d0
        entry_SYSCALL_64_after_hwframe+0x56/0x5e

---
Changes in v3:
- Rebased over PM tree's linux-next branch
- Added a comment to explain why this piece of code is required
- Switched from guard() to scoped_guard() to allow explicitly wrapping the policy guard inside the cpu_hotplug_lock critical section.

Changes in v2:
- Restrict cpu_hotplug_lock acquisition to only
the local_boost attribute in store() handler.

Regards,
Seyediman

 drivers/cpufreq/cpufreq.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)
drivers/cpufreq/cpufreq.c +9 -6
index 21fa733a2..b349adbeb 100644
@@ -622,10 +622,7 @@ static ssize_t store_local_boost(struct cpufreq_policy *policy,
    if (!policy->boost_supported)
            return -EINVAL;
 
-   cpus_read_lock();
    ret = policy_set_boost(policy, enable);
-   cpus_read_unlock();
-
    if (!ret)
            return count;
 
@@ -1006,16 +1003,28 @@ static ssize_t store(struct kobject *kobj, struct attribute *attr,
 {
    struct cpufreq_policy *policy = to_policy(kobj);
    struct freq_attr *fattr = to_attr(attr);
+   int ret = -EBUSY;
 
    if (!fattr->store)
            return -EIO;
 
-   guard(cpufreq_policy_write)(policy);
+   /*
+    * store_local_boost() requires cpu_hotplug_lock to be held, and must be
+    * called with that lock acquired *before* taking policy->rwsem to avoid
+    * lock ordering violations.
+    */
+   if (fattr == &local_boost)
+           cpus_read_lock();
 
-   if (likely(!policy_is_inactive(policy)))
-           return fattr->store(policy, buf, count);
+   scoped_guard(cpufreq_policy_write, policy) {

drivers/cpufreq/cpufreq.c static ssize_t store(struct kobject *kobj, struct attribute *attr, · line 1028

Quoted 10 lines
drivers/cpufreq/cpufreq.c +7 -1
+           if (likely(!policy_is_inactive(policy)))
+                   ret = fattr->store(policy, buf, count);
+   }
 
-   return -EBUSY;
+   if (fattr == &local_boost)
+           cpus_read_unlock();
+
+   return ret;
 }
Quoted 2 lines

--
viresh

Hi there,

Just following up to see if there's anything you'd like me to change or address in the patch before it can move forward. Please let me know if any updates are needed.

Quoted 8 lines
Quoted 5 lines

I'm kind of wondering why local_boost needs cpus_read_lock() at all. Holding the policy rwsem blocks CPU online/offline already for this policy.

Is that because ->set_boost() may need to synchronize with the other policies?

IOW, what can go wrong if the cpus_read_lock() locking is dropped from there altogether?

I think ->set_boost() being per-policy makes cpus_read_lock() unnecessary here. Since we already hold the policy lock, any topology changes involving this policy should be blocked. And because we're not iterating over all CPUs or policies to set boost, we don't need to worry about CPU hotplug synchronization in this case.

Regards,
Seyediman