Re: [PATCH 01/11] watchdog: WatchDog Timer Driver Core - Add basic framework

[Lars-Peter Clausen <lars@metafoo.de>]

On 07/22/2011 09:32 PM, Wim Van Sebroeck wrote:
> Hi Lars-Peter,
> 
>> On 07/11/2011 04:19 PM, Wim Van Sebroeck wrote:
>>> +
>>> +int watchdog_dev_unregister(struct watchdog_device *watchdog)
>>> +{
>>> +	/* Check that a watchdog device was registered in the past */
>>> +	if (!test_bit(0, &watchdog_dev_busy) || !wdd)
>>> +		return -ENODEV;
>>> +
>>> +	/* We can only unregister the watchdog device that was registered */
>>> +	if (watchdog != wdd) {
>>> +		pr_err("%s: watchdog was not registered as /dev/watchdog.\n",
>>> +			watchdog->info->identity);
>>> +		return -ENODEV;
>>> +	}
>>> +
>>> +	/* Unregister the miscdevice */
>>> +	misc_deregister(&watchdog_miscdev);
>>> +	wdd = NULL;
>>> +	clear_bit(0, &watchdog_dev_busy);
>>> +	return 0;
>>> +}
>>
>> What happens if the watchdog gets unregistered if the device is still opened?
>> Even though if you'd check wdd for not being NULL in the file callbacks there
>> is still a chance for races if the devices is unregistered at the same time as
>> the callback is running. You'd either need a big lock to protect from having a
>> file callback and unregister running concurrently or add ref-counting to the
>> watchdog_device, the later best done by embedding a struct device and using the
>> device driver model.
> 
> You cannot unload the watchdog-drivers module if /dev/watchdog is still open.
> So if the watchdog_unregister function is in the exit function of the module
> then we are safe. But I think you have a point if that is not the case.
> Solution would be to return an error when the watchdog_unregister_device routine
> is called and the WDOG_DEV_OPEN bit is set. Will create an extra patch for that.
> 

The problem is, that this doesn't fit nicely into the linux device driver
model, because it doesn't allow the removal of a device to fail. So you'll
still end of with undefined behavior.

- Lars