Re: [patch] PCI device matching fix

[Patrick Mochel]

> Yes, you're basically right, whether you use the refcount in struct driver 
> or the module use count doesn't make any difference. Except for one thing:
> You cannot call module_exit() when the module count is > 0. Since only 
> then unregister_driver() is called, there's no way to get the use count to 
> zero and thus unload the module. One could of course change the policy to 
> call unregister_driver() not from module_exit(), but e.g. by
> "echo remove > /driversfs/../my_driver", but it's surely not that I'm 
> suggesting this.

We can keep the same policy as modules now - keep the usage count at 0 
when not in use. That way the module can unload at any time. module_exit() 
calls driver_unregister(), which, down the line, calls the drivers' 
remove() for each device attached to it. 

That seems a lot cleaner than anything so far; we leverage the existing
module infrastructure as much as possible. I'll see about codifying the
concept in the next day or so and sending it out..

	-pat