Re: [RFC] Input: Remove unsafe device module references

[Greg KH <gregkh@suse.de>]

On Tue, Nov 01, 2011 at 10:52:10AM -0700, Dmitry Torokhov wrote:
> On Tue, Nov 01, 2011 at 10:01:56AM -0700, Greg KH wrote:
> > On Tue, Nov 01, 2011 at 04:41:40PM +0100, David Herrmann wrote:
> > > Hi Dmitry and Greg
> > > 
> > > It doesn't make sense to take a reference to our own module. When we call
> > > module_put(THIS_MODULE) we cannot make sure that our module is still alive when
> > > this function returns. Therefore, module_put() will return to invalid memory and
> > > our input_dev_release() function is no longer available.
> > > 
> > > It would be interesting if Greg could elaborate what else we could do to replace
> > > this module-refcount as it is definitely needed here. However, "struct device"
> > > doesn't provide an owner field so there is no way for us to let the device core
> > > keep a reference to our module.
> > 
> > For a bus module, yes, this is needed, so don't remove these calls, it's
> > wrong to do so.
> 
> Strictly speaking, David is right, there is a race condition here.
> However since we do module_put() as very last operation of
> input_dev_release() it is extremely hard to trigger this race.
> 
> Until we have a better way of pinning the bus (or class) implementation
> in memory we should keep __module_get/module_put in input core.

I agree, that's fine for a bus to do, as long as you are aware of how it
is working.

greg k-h