[PATCH] Revert "hwmon: (ibmpex) fix use-after-free in high/low store"

[PATCH] Revert "hwmon: (ibmpex) fix use-after-free in high/low store"

[Guenter Roeck]

This reverts commit 6946c726c3f4c36f0f049e6f97e88c510b15f65d.

Jean Delvare points out that the patch does not completely
fix the reported problem, that it in fact introduces a
(new) race condition, and that it may actually not be needed in
the first place.

Various AI reviews agree. Specific and relevant AI feedback:

"
This reordering sets the driver data to NULL before removing the sensor
attributes in the loop below.

ibmpex_show_sensor() retrieves this driver data via dev_get_drvdata() but
does not check if it is NULL before dereferencing it to access
data->sensors[].

If a userspace process reads a sensor file (like temp1_input) while this
delete function is running, could it race with the dev_set_drvdata(...,
NULL) call here and crash in ibmpex_show_sensor()?

Would it be safer to keep the original order where device_remove_file() is
called before clearing the driver data? device_remove_file() should wait
for any active sysfs callbacks to complete, which might already prevent the
use-after-free this patch intends to fix.
"

Revert the offending patch. If it can be shown that the originally reported
alleged race condition does indeed exist, it can always be re-introduced
with a complete fix.

Reported-by: Jean Delvare <jdelvare@suse.de>
Closes: https://lore.kernel.org/linux-hwmon/20260121095342.73e723cb@endymion/
Cc: Jean Delvare <jdelvare@suse.de>
Cc: Junrui Luo <moonafterrain@outlook.com>
Fixes: 6946c726c3f4 ("hwmon: (ibmpex) fix use-after-free in high/low store")
Signed-off-by: Guenter Roeck <linux@roeck-us.net>
---
 drivers/hwmon/ibmpex.c | 9 ++-------
 1 file changed, 2 insertions(+), 7 deletions(-)

diff --git a/drivers/hwmon/ibmpex.c b/drivers/hwmon/ibmpex.c
index 129f3a9e8fe9..228c5f6c6f38 100644
--- a/drivers/hwmon/ibmpex.c
+++ b/drivers/hwmon/ibmpex.c
@@ -277,9 +277,6 @@ static ssize_t ibmpex_high_low_store(struct device *dev,
 {
 	struct ibmpex_bmc_data *data = dev_get_drvdata(dev);
 
-	if (!data)
-		return -ENODEV;
-
 	ibmpex_reset_high_low_data(data);
 
 	return count;
@@ -511,9 +508,6 @@ static void ibmpex_bmc_delete(struct ibmpex_bmc_data *data)
 {
 	int i, j;
 
-	hwmon_device_unregister(data->hwmon_dev);
-	dev_set_drvdata(data->bmc_device, NULL);
-
 	device_remove_file(data->bmc_device,
 			   &sensor_dev_attr_reset_high_low.dev_attr);
 	device_remove_file(data->bmc_device, &dev_attr_name.attr);
@@ -527,7 +521,8 @@ static void ibmpex_bmc_delete(struct ibmpex_bmc_data *data)
 		}
 
 	list_del(&data->list);
-
+	dev_set_drvdata(data->bmc_device, NULL);
+	hwmon_device_unregister(data->hwmon_dev);
 	ipmi_destroy_user(data->user);
 	kfree(data->sensors);
 	kfree(data);
-- 
2.45.2

Re: [PATCH] Revert "hwmon: (ibmpex) fix use-after-free in high/low store"

[Jean Delvare]

Hi Guenter,

On Sat, 2026-02-07 at 08:21 -0800, Guenter Roeck wrote:
> This reverts commit 6946c726c3f4c36f0f049e6f97e88c510b15f65d.
> 
> Jean Delvare points out that the patch does not completely
> fix the reported problem, that it in fact introduces a
> (new) race condition, and that it may actually not be needed in
> the first place.
> 
> Various AI reviews agree. Specific and relevant AI feedback:
> 
> "
> This reordering sets the driver data to NULL before removing the sensor
> attributes in the loop below.
> 
> ibmpex_show_sensor() retrieves this driver data via dev_get_drvdata() but
> does not check if it is NULL before dereferencing it to access
> data->sensors[].
> 
> If a userspace process reads a sensor file (like temp1_input) while this
> delete function is running, could it race with the dev_set_drvdata(...,
> NULL) call here and crash in ibmpex_show_sensor()?
> 
> Would it be safer to keep the original order where device_remove_file() is
> called before clearing the driver data? device_remove_file() should wait
> for any active sysfs callbacks to complete, which might already prevent the
> use-after-free this patch intends to fix.
> "
> 
> Revert the offending patch. If it can be shown that the originally reported
> alleged race condition does indeed exist, it can always be re-introduced
> with a complete fix.
> 
> Reported-by: Jean Delvare <jdelvare@suse.de>
> Closes: https://lore.kernel.org/linux-hwmon/20260121095342.73e723cb@endymion/
> Cc: Jean Delvare <jdelvare@suse.de>
> Cc: Junrui Luo <moonafterrain@outlook.com>
> Fixes: 6946c726c3f4 ("hwmon: (ibmpex) fix use-after-free in high/low store")
> Signed-off-by: Guenter Roeck <linux@roeck-us.net>
> ---
>  drivers/hwmon/ibmpex.c | 9 ++-------
>  1 file changed, 2 insertions(+), 7 deletions(-)
> 
> diff --git a/drivers/hwmon/ibmpex.c b/drivers/hwmon/ibmpex.c
> index 129f3a9e8fe9..228c5f6c6f38 100644
> --- a/drivers/hwmon/ibmpex.c
> +++ b/drivers/hwmon/ibmpex.c
> @@ -277,9 +277,6 @@ static ssize_t ibmpex_high_low_store(struct device *dev,
>  {
>  	struct ibmpex_bmc_data *data = dev_get_drvdata(dev);
>  
> -	if (!data)
> -		return -ENODEV;
> -
>  	ibmpex_reset_high_low_data(data);
>  
>  	return count;
> @@ -511,9 +508,6 @@ static void ibmpex_bmc_delete(struct ibmpex_bmc_data *data)
>  {
>  	int i, j;
>  
> -	hwmon_device_unregister(data->hwmon_dev);
> -	dev_set_drvdata(data->bmc_device, NULL);
> -
>  	device_remove_file(data->bmc_device,
>  			   &sensor_dev_attr_reset_high_low.dev_attr);
>  	device_remove_file(data->bmc_device, &dev_attr_name.attr);
> @@ -527,7 +521,8 @@ static void ibmpex_bmc_delete(struct ibmpex_bmc_data *data)
>  		}
>  
>  	list_del(&data->list);
> -
> +	dev_set_drvdata(data->bmc_device, NULL);
> +	hwmon_device_unregister(data->hwmon_dev);
>  	ipmi_destroy_user(data->user);
>  	kfree(data->sensors);
>  	kfree(data);

Reviewed-by: Jean Delvare <jdelvare@suse.de>

Thanks,
-- 
Jean Delvare
SUSE L3 Support