@@ -1073,6 +1073,7 @@ static int watchdog_cdev_register(struct watchdog_device *wdd)
/* Fill in the data structures */
cdev_init(&wd_data->cdev, &watchdog_fops);
+ wd_data->cdev.owner = wdd->ops->owner;
/* Add the device */
err = cdev_device_add(&wd_data->cdev, &wd_data->dev);
@@ -1087,8 +1088,6 @@ static int watchdog_cdev_register(struct watchdog_device *wdd)
return err;
}
- wd_data->cdev.owner = wdd->ops->owner;
-
/* Record time of most recent heartbeat as 'just before now'. */
wd_data->last_hw_keepalive = ktime_sub(ktime_get(), 1);
watchdog_set_open_deadline(wd_data);
When the new watchdog character device is registered, it becomes available for opening. This creates a race where userspace may open the device before the character device's owner is set. This results in an imbalance in module_get calls as the cdev_get in cdev_open will not increment the reference count on the watchdog driver module. This causes problems when the watchdog character device is released as the module loader's reference will also be released. This makes it impossible to open the watchdog device later on as it now appears that the module is being unloaded. The open will fail with -ENXIO from chrdev_open. The legacy watchdog device will fail with -EBUSY from the try_module_get in watchdog_open because it's module owner is the watchdog core module so it can still be opened but it will fail to get a refcount on the underlying watchdog device driver. Fixes: 72139dfa2464 ("watchdog: Fix the race between the release of watchdog_core_data and cdev") Signed-off-by: Curtis Klein <curtis.klein@hpe.com> --- drivers/watchdog/watchdog_dev.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)