Re: Driverfs updates

[John Alvord <jalvo@mbay.net>]

On Tue, 9 Jul 2002 09:56:55 -0700 (PDT), Patrick Mochel
<mochel@osdl.org> wrote:

>
>On Tue, 9 Jul 2002, Keith Owens wrote:
>
>> On Mon, 8 Jul 2002 11:41:52 -0700 (PDT), 
>> Patrick Mochel <mochel@osdl.org> wrote:
>> >- Add struct module * owner field to struct device_driver
>> >- Change {get,put}_driver to use it
>> 
>> struct device_driver * get_driver(struct device_driver * drv)
>> {
>>         if (drv && drv->owner)
>>                 if (!try_inc_mod_count(drv->owner))
>>                         return NULL;
>>         return drv;
>> }
>> 
>> is racy.  The module can be unloaded after if (drv->owner) and before
>> try_inc_mod_count.  To prevent that race, drv itself must be locked
>> around calls to get_driver().
>> 
>> The "normal" method is to have a high level lock that controls the drv
>> list and to take that lock in the register and unregister routines and
>> around the call to try_inc_mod_count.  drv->bus->lock is no good,
>> anything that relies on reading drv without a lock or module reference
>> count is racy.  I suggest you add a global driverfs_lock.
>
>This race really sucks. 
>
>Adding a high level lock is no big deal, but I don't think it will solve 
>the problem. Hopefully you can educate me a bit more. 
>
>If you add a driver_lock, you might have something like:
>
>	struct device_driver * d = NULL;
>
>	spin_lock(&driver_lock);
>	if (drv && drv->owner)
>		if (try_inc_mod_count(drv->owner))
>			d = drv;
>
>	spin_unlock(&driver_lock):
>	return d;
>
>...but, what if someone has unloaded the module before you get to the if 
>statement? The memory for the module has been freed, including drv itself. 
>
>How do you protect against that? The simplest solutions, given the current 
>infrastructure, are:
>
>- The BKL
>- Not allowing module unload
>- Ignoring it, and hoping it goes away
>
>None of those solutions are ideal, though I don't have any bright ideas 
>off the top of my head.

The only idea I can see is to have a single kernel-thread process
which would do each load/unload request serially on a single
processor. 

john alvord