hwmon: (gpio-fan) fix pm_runtime imbalance for alarm-only fans
PATCH 1/1 hwmon: (gpio-fan) fix pm_runtime imbalance for alarm-only fans
pm_runtime_enable() runs unconditionally at the end of probe, but the devm action that calls pm_runtime_disable() is registered only when control GPIOs are present. An alarm-only fan (no control GPIOs, alarm GPIO optional) is a valid DT configuration; on unbind the runtime PM core warns about the missing disable.
Use devm_pm_runtime_enable() so disable is automatic on device release, and drop the redundant pm_runtime_disable() from gpio_fan_stop().
--- drivers/hwmon/gpio-fan.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/hwmon/gpio-fan.c +3 -3
@@ -524,8 +524,6 @@ static void gpio_fan_stop(void *data)mutex_lock (&fan_data->lock);set_fan_speed (data,0 );mutex_unlock (&fan_data->lock); - -pm_runtime_disable (fan_data->dev); }static int gpio_fan_probe (struct platform_device *pdev)
@@ -581,7 +579,9 @@ static int gpio_fan_probe(struct platform_device *pdev) }pm_runtime_set_suspended (&pdev->dev); -pm_runtime_enable (&pdev->dev); + err =devm_pm_runtime_enable (&pdev->dev); +if (err) +return err;/* If current GPIO state is active, mark RPM as active as well */ if (fan_data->speed_index >0 ) {int ret;
Signature · 1 line
2.25.1
[email protected]
AI reviewThank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [High] Converting to `devm_pm_runtime_enable()` and placing it at the end of `probe()` incorrectly…
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] Converting to devm_pm_runtime_enable() and placing it at the end of probe() incorrectly reverses the teardown order, causing PM runtime to be disabled too early. This leads to a regulator enable count leak and exposes a sysfs race condition on driver unbind. --
commit 728a875c515df8244bd554b338f4f008d88b3b9e
Author: Cong Nguyen <[email protected]>
hwmon: (gpio-fan) fix pm_runtime imbalance for alarm-only fans
This commit changes the driver to use devm_pm_runtime_enable() during probe and removes the manual pm_runtime_disable() from gpio_fan_stop(). This ensures PM runtime is automatically disabled on unbind, fixing an imbalance for alarm-only fans.
Quoted 14 lines
drivers/hwmon/gpio-fan.c +0
-2@@ -524,8 +524,6 @@ static void gpio_fan_stop(void *data)mutex_lock (&fan_data->lock);set_fan_speed (data,0 );mutex_unlock (&fan_data->lock); - -pm_runtime_disable (fan_data->dev); }static int gpio_fan_probe (struct platform_device *pdev)
drivers/hwmon/gpio-fan.c static int gpio_fan_probe(struct platform_device *pdev) · line 584
Quoted 7 lines
drivers/hwmon/gpio-fan.c +3
-1@@ -581,7 +579,9 @@ static int gpio_fan_probe(struct platform_device *pdev) }pm_runtime_set_suspended (&pdev->dev); -pm_runtime_enable (&pdev->dev); + err =devm_pm_runtime_enable (&pdev->dev); +if (err) +return err;
[Severity: High]
Does moving the pm_runtime_disable() action to the end of probe reverse the teardown order in a way that breaks unbind?
Because devm unwinds in LIFO order, registering devm_pm_runtime_enable() at the end of gpio_fan_probe() means pm_runtime_disable() will execute first during unbind. The cleanup action for gpio_fan_stop() is registered earlier in probe via devm_add_action_or_reset(), so it will execute after PM runtime has already been disabled.
When gpio_fan_stop() calls set_fan_speed(..., 0), it invokes pm_runtime_put_sync(). Since PM runtime is already disabled, this will fail with -EACCES. This bypasses the runtime suspend callback, leaving the hardware regulator permanently enabled and causing a regulator enable count leak.
Quoted 3 lines
drivers/hwmon/gpio-fan.c +0
-0/* If current GPIO state is active, mark RPM as active as well */ if (fan_data->speed_index >0 ) {int ret;
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Guenter Roeck
maintainerQuoted 46 lines · On 8/28/26 04:01, [email protected] wrote:
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] Converting todevm_pm_runtime_enable()and placing it at the end ofprobe()incorrectly reverses the teardown order, causing PM runtime to be disabled too early. This leads to a regulator enable count leak and exposes a sysfs race condition on driver unbind. --commit
728a875c515df8244bd554b338f4f008d88b3b9e
Author: Cong Nguyen <[email protected]>hwmon: (gpio-fan) fix
pm_runtimeimbalance for alarm-only fansThis commit changes the driver to use
devm_pm_runtime_enable()during probe and removes the manualpm_runtime_disable()fromgpio_fan_stop(). This ensures PM runtime is automatically disabled on unbind, fixing an imbalance for alarm-only fans.Quoted 14 lines
drivers/hwmon/gpio-fan.c +0
-2@@ -524,8 +524,6 @@ static void gpio_fan_stop(void *data)mutex_lock (&fan_data->lock);set_fan_speed (data,0 );mutex_unlock (&fan_data->lock); - -pm_runtime_disable (fan_data->dev); }static int gpio_fan_probe (struct platform_device *pdev)drivers/hwmon/gpio-fan.c static int gpio_fan_probe(struct platform_device *pdev) · line 584
Quoted 7 lines
drivers/hwmon/gpio-fan.c +3
-1@@ -581,7 +579,9 @@ static int gpio_fan_probe(struct platform_device *pdev) }pm_runtime_set_suspended (&pdev->dev); -pm_runtime_enable (&pdev->dev); + err =devm_pm_runtime_enable (&pdev->dev); +if (err) +return err;[Severity: High]
Does moving thepm_runtime_disable()action to the end of probe reverse the teardown order in a way that breaks unbind?Because devm unwinds in LIFO order, registering
devm_pm_runtime_enable()at the end ofgpio_fan_probe()meanspm_runtime_disable()will execute first during unbind. The cleanup action forgpio_fan_stop()is registered earlier in probe viadevm_add_action_or_reset(), so it will execute after PM runtime has already been disabled.
Quoted 6 lines
When
gpio_fan_stop()callsset_fan_speed(..., 0), it invokespm_runtime_put_sync(). Since PM runtime is already disabled, this will fail with-EACCES. This bypasses the runtime suspend callback, leaving the hardware regulator permanently enabled and causing a regulator enable count leak.
Seems to me that this will need a more comprehensive fix.
Thanks,
Guenter
Quoted 4 lines
Quoted 3 lines
drivers/hwmon/gpio-fan.c +0
-0/* If current GPIO state is active, mark RPM as active as well */ if (fan_data->speed_index >0 ) {int ret;
Quoted 48 lines · Guenter Roeck
Quoted 46 lines · On 8/28/26 04:01, [email protected] wrote:
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] Converting todevm_pm_runtime_enable()and placing it at the end ofprobe()incorrectly reverses the teardown order, causing PM runtime to be disabled too early. This leads to a regulator enable count leak and exposes a sysfs race condition on driver unbind. --commit
728a875c515df8244bd554b338f4f008d88b3b9e
Author: Cong Nguyen <[email protected]>hwmon: (gpio-fan) fix
pm_runtimeimbalance for alarm-only fansThis commit changes the driver to use
devm_pm_runtime_enable()during probe and removes the manualpm_runtime_disable()fromgpio_fan_stop(). This ensures PM runtime is automatically disabled on unbind, fixing an imbalance for alarm-only fans.Quoted 14 lines
drivers/hwmon/gpio-fan.c +0
-2@@ -524,8 +524,6 @@ static void gpio_fan_stop(void *data)mutex_lock (&fan_data->lock);set_fan_speed (data,0 );mutex_unlock (&fan_data->lock); - -pm_runtime_disable (fan_data->dev); }static int gpio_fan_probe (struct platform_device *pdev)drivers/hwmon/gpio-fan.c static int gpio_fan_probe(struct platform_device *pdev) · line 584
Quoted 7 lines
drivers/hwmon/gpio-fan.c +3
-1@@ -581,7 +579,9 @@ static int gpio_fan_probe(struct platform_device *pdev) }pm_runtime_set_suspended (&pdev->dev); -pm_runtime_enable (&pdev->dev); + err =devm_pm_runtime_enable (&pdev->dev); +if (err) +return err;[Severity: High]
Does moving thepm_runtime_disable()action to the end of probe reverse the teardown order in a way that breaks unbind?Because devm unwinds in LIFO order, registering
devm_pm_runtime_enable()at the end ofgpio_fan_probe()meanspm_runtime_disable()will execute first during unbind. The cleanup action forgpio_fan_stop()is registered earlier in probe viadevm_add_action_or_reset(), so it will execute after PM runtime has already been disabled.
Quoted 7 lines
Quoted 6 lines
When
gpio_fan_stop()callsset_fan_speed(..., 0), it invokespm_runtime_put_sync(). Since PM runtime is already disabled, this will fail with-EACCES. This bypasses the runtime suspend callback, leaving the hardware regulator permanently enabled and causing a regulator enable count leak.Seems to me that this will need a more comprehensive fix.
Confirmed, Sashiko's right. v1 registered devm_pm_runtime_enable() after gpio_fan_stop()'s devm action, so LIFO teardown ran pm_runtime_disable() first -- gpio_fan_stop()'s pm_runtime_put_sync() got -EACCES and never reached the regulator disable.
v2 moves devm_pm_runtime_enable() earlier so gpio_fan_stop() runs first again. Sending shortly.
thanks,
Cong
Quoted 9 lines
Thanks,
GuenterQuoted 4 lines
Quoted 3 lines
drivers/hwmon/gpio-fan.c +0
-0/* If current GPIO state is active, mark RPM as active as well */ if (fan_data->speed_index >0 ) {int ret;
PATCH v2 1/1 hwmon: (gpio-fan) fix pm_runtime imbalance for alarm-only fans
In reply to Cong Nguyenpm_runtime_enable() runs unconditionally in probe, but the devm cleanup that calls pm_runtime_disable() is only registered when control GPIOs are present. Alarm-only fans never get it, so unbind warns about the missing disable.
Switch to devm_pm_runtime_enable(), registered before gpio_fan_stop()'s devm action so LIFO teardown runs gpio_fan_stop() first. gpio_fan_stop() calls set_fan_speed(0) -> pm_runtime_put_sync(), which needs PM still enabled to reach gpio_fan_runtime_suspend() and turn the regulator off; registering enable after gpio_fan_stop() (as v1 did) reverses that order and leaves the regulator on.
---
Changes in v2:
- v1's devm_pm_runtime_enable() was registered after gpio_fan_stop()'s devm action, reversing LIFO teardown order and leaving the regulator enabled on unbind (Guenter/Sashiko). Moved it earlier so gpio_fan_stop() registers first.
drivers/hwmon/gpio-fan.c | 14 ++++++++++---- 1 file changed, 10 insertions(+), 4 deletions(-)
drivers/hwmon/gpio-fan.c +10 -4
@@ -524,8 +524,6 @@ static void gpio_fan_stop(void *data)mutex_lock (&fan_data->lock);set_fan_speed (data,0 );mutex_unlock (&fan_data->lock); - -pm_runtime_disable (fan_data->dev); }static int gpio_fan_probe (struct platform_device *pdev)
@@ -553,6 +551,16 @@ static int gpio_fan_probe(struct platform_device *pdev)return dev_err_probe (dev,PTR_ERR (fan_data->supply),"Failed to get fan-supply" ); +/* +* Register before gpio_fan_stop()'s devm action: LIFO teardown must +* run gpio_fan_stop() (needs PM enabled to disable the regulator) +* before this disables PM. +*/ +pm_runtime_set_suspended (&pdev->dev); + err =devm_pm_runtime_enable (&pdev->dev); +if (err) +return err; +/* Configure control GPIOs if available. */ if (fan_data->gpios && fan_data->num_gpios >0 ) {if (!fan_data->speed || fan_data->num_speed <=1 )
@@ -580,8 +588,6 @@ static int gpio_fan_probe(struct platform_device *pdev)return err; } -pm_runtime_set_suspended (&pdev->dev); -pm_runtime_enable (&pdev->dev);/* If current GPIO state is active, mark RPM as active as well */ if (fan_data->speed_index >0 ) {int ret;
Signature · 1 line
2.25.1
[email protected]
AI reviewThank you for your contribution! Sashiko AI review found 7 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 7 potential issue(s) to consider:
New issues:
- [High] Moving devm_pm_runtime_enable() to the start of probe opens a critical race window causing a permanent PM usage count leak.
Pre-existing issues:
- [Critical] Missing cancel_work_sync() for alarm_work leads to Use-After-Free.
- [High] Missing synchronize_irq() in runtime suspend callback for shared IRQ.
- [High] Probe error paths trigger pm_runtime_put_sync() on an unresumed device, causing a usage count underflow and preventing the regulator from disabling.
- [High] State mismatch and PM usage counter underflow on pm_runtime_put_sync() failure.
- [High] Missing lock acquisition in gpio_fan_shutdown() leads to data races and unbalanced runtime PM.
- [High] Alarm IRQ handler unconditionally returns IRQ_NONE, breaking the shared interrupt line.
--
commit 7341755fa2914776aa77e1819998064fadc8592b
Author: Cong Nguyen <[email protected]>
hwmon: (gpio-fan) fix pm_runtime imbalance for alarm-only fans
This commit addresses a runtime PM imbalance in alarm-only fans where pm_runtime_disable() was skipped on teardown. It replaces pm_runtime_enable() with devm_pm_runtime_enable() and moves it earlier in probe to ensure LIFO teardown runs gpio_fan_stop() before disabling PM.
Quoted 4 lines
drivers/hwmon/gpio-fan.c +0
-0
[ ... ]
drivers/hwmon/gpio-fan.c static int gpio_fan_probe(struct platform_device *pdev) · line 560
Quoted 11 lines
drivers/hwmon/gpio-fan.c +7
-0@@ -553,6 +551,16 @@ static int gpio_fan_probe(struct platform_device *pdev)return dev_err_probe (dev,PTR_ERR (fan_data->supply),"Failed to get fan-supply" ); +/* +* Register before gpio_fan_stop()'s devm action: LIFO teardown must +* run gpio_fan_stop() (needs PM enabled to disable the regulator) +* before this disables PM. +*/ +pm_runtime_set_suspended (&pdev->dev); + err =devm_pm_runtime_enable (&pdev->dev);
[Severity: High]
Does moving this earlier open a race condition that could permanently leak a PM usage count?
If a concurrent sysfs write occurs after hwmon registration exposes sysfs files (like pwm1, fan1_target) but before probe finishes, set_fan_speed() will successfully call pm_runtime_resume_and_get().
This increments the usage count to 1 and sets fan_data->speed_index > 0.
When probe reaches its end, it checks if fan_data->speed_index > 0 and unconditionally calls pm_runtime_resume_and_get() again.
Will this double-increment the counter, preventing the device from ever entering runtime suspend?
Previously, PM was disabled during this window, causing concurrent sysfs access to fail cleanly with -EACCES.
[ ... ]
[Severity: High]
This is a pre-existing issue, but could the devm teardown cause a PM usage count underflow during probe error paths?
In gpio_fan_probe(), the teardown action is registered before the final PM reference is acquired:
err = devm_add_action_or_reset(dev, gpio_fan_stop, fan_data);
If probe fails after fan_ctrl_init() sets speed_index > 0, but before pm_runtime_resume_and_get() is called at the end of probe, the PM usage count is 0.
The devm rollback calls gpio_fan_stop() which calls set_fan_speed(0). This invokes pm_runtime_put_sync(), decrementing the usage count from 0 to -1.
Because the count underflows, does the PM core (rpm_suspend) abort with an error and leave the regulator powered on despite the teardown?
[Severity: High]
This is a pre-existing issue, but are we missing a synchronize_irq() in the runtime suspend callback?
The PM Subsystem Concurrency and Locking guidelines dictate that drivers using IRQF_SHARED must call synchronize_irq() before powering down hardware.
In fan_alarm_init(), the alarm interrupt is requested with IRQF_SHARED:
devm_request_irq(dev, alarm_irq, fan_alarm_irq_handler,
IRQF_SHARED, "GPIO fan alarm", fan_data);
However, in gpio_fan_runtime_suspend(), the hardware regulator is disabled without first synchronizing the interrupt:
if (fan_data->supply)
ret = regulator_disable(fan_data->supply);Could this race with a shared IRQ executing mid-flight when the device enters runtime suspend, causing spurious interrupts or data races?
[Severity: Critical]
This is a pre-existing issue, but can the lack of cancel_work_sync() for alarm_work lead to a Use-After-Free?
In fan_alarm_init(), the work is initialized and the IRQ is requested via devm_request_irq():
INIT_WORK(&fan_data->alarm_work, fan_alarm_notify);
return devm_request_irq(dev, alarm_irq, fan_alarm_irq_handler,
IRQF_SHARED, "GPIO fan alarm", fan_data);On driver unbind or probe failure, devm frees the IRQ, but there is no call to cancel_work_sync() to wait for or cancel already-queued work.
Devm then frees the fan_data structure.
If the asynchronous workqueue eventually executes fan_alarm_notify(), it dereferences the freed memory:
sysfs_notify(&fan_data->hwmon_dev->kobj, NULL, "fan1_alarm");
Would a devm_add_action_or_reset() be needed here to cancel the work?
[Severity: High]
This is a pre-existing issue, but does set_fan_speed() cause a state mismatch and PM counter underflow if pm_runtime_put_sync() fails?
In set_fan_speed():
ret = pm_runtime_put_sync(fan_data->dev);
if (ret < 0 && ret != -ENOSYS)
return ret;
fan_data->speed_index = speed_index;pm_runtime_put_sync() unconditionally drops the PM usage count even on error. If it returns an error, the function returns early without updating fan_data->speed_index to 0.
Because speed_index incorrectly remains > 0, any subsequent attempt to stop the fan will invoke pm_runtime_put_sync() again.
Could this lead to multiple decrements of the PM usage counter for a single PM reference?
[Severity: High]
This is a pre-existing issue, but might gpio_fan_shutdown() race due to missing lock acquisition?
static void gpio_fan_shutdown(struct platform_device *pdev)
{
struct gpio_fan_data *fan_data = platform_get_drvdata(pdev);
if (fan_data->gpios)
set_fan_speed(fan_data, 0);
}set_fan_speed() explicitly requires fan_data->lock to be held to safely read and update fan_data->speed_index.
If system shutdown races with a concurrent sysfs write or thermal callback, both threads might issue set_fan_speed(0) simultaneously.
Could this issue pm_runtime_put_sync() twice and underflow the runtime PM usage count?
[Severity: High]
This is a pre-existing issue, but does the alarm IRQ handler break the shared interrupt line by unconditionally returning IRQ_NONE?
static irqreturn_t fan_alarm_irq_handler(int irq, void *dev_id)
{
struct gpio_fan_data *fan_data = dev_id;
schedule_work(&fan_data->alarm_work);
return IRQ_NONE;
}When requested with IRQF_SHARED, returning IRQ_NONE informs the core IRQ subsystem that the interrupt was unhandled.
After enough unhandled triggers, won't the kernel's spurious interrupt detection permanently disable the shared interrupt line, causing a denial of service for any other device sharing this IRQ?
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Guenter Roeck
maintainerQuoted 12 lines · Cong Nguyen
pm_runtime_enable()runs unconditionally in probe, but the devm cleanup that callspm_runtime_disable()is only registered when control GPIOs are present. Alarm-only fans never get it, so unbind warns about the missing disable.Switch to
devm_pm_runtime_enable(), registered before gpio_fan_stop()'s devm action so LIFO teardown runsgpio_fan_stop()first.gpio_fan_stop()callsset_fan_speed(0)->pm_runtime_put_sync(), which needs PM still enabled to reachgpio_fan_runtime_suspend()and turn the regulator off; registering enable aftergpio_fan_stop()(as v1 did) reverses that order and leaves the regulator on.
Quoted 7 lines
Fixes 0d01110e6356 ("hwmon: (gpio-fan) Add regulator support")Reported-by Guenter Roeck <[email protected]>Assisted-by Claude:claude-opus-4Signed-off-by Cong Nguyen <[email protected]>---
Another instance of a new patch version sent as reply to a previous version.
It is against guidance in Documentation/process/submitting-patches.rst, yet it proliferates, and more and more people send new patch revisions this way. Where is this suggested ?
Thanks,
Guenter
Quoted 22 lines · Guenter Roeck
Quoted 12 lines · Cong Nguyen
pm_runtime_enable()runs unconditionally in probe, but the devm cleanup that callspm_runtime_disable()is only registered when control GPIOs are present. Alarm-only fans never get it, so unbind warns about the missing disable.Switch to
devm_pm_runtime_enable(), registered before gpio_fan_stop()'s devm action so LIFO teardown runsgpio_fan_stop()first.gpio_fan_stop()callsset_fan_speed(0)->pm_runtime_put_sync(), which needs PM still enabled to reachgpio_fan_runtime_suspend()and turn the regulator off; registering enable aftergpio_fan_stop()(as v1 did) reverses that order and leaves the regulator on.Quoted 7 lines
Fixes 0d01110e6356 ("hwmon: (gpio-fan) Add regulator support")Reported-by Guenter Roeck <[email protected]>Assisted-by Claude:claude-opus-4Signed-off-by Cong Nguyen <[email protected]>---
Quoted 6 lines
Another instance of a new patch version sent as reply to a previous version.
It is against guidance in
Documentation/process/submitting-patches.rst, yet it proliferates, and more and more people send new patch revisions this way. Where is this suggested ?
Fair, this was my mistake -- won't thread rerolls under the previous version again. v3 just sent standalone.
Quoted 4 lines
Thanks,
Guenter