Re: 2.6.30-rc6: Reported regressions from 2.6.29

[Greg KH <gregkh@suse.de>]

On Sat, May 16, 2009 at 07:13:59PM -0700, Linus Torvalds wrote:
> 
> 
> On Sun, 17 May 2009, Kay Sievers wrote:
> > 
> > This makes the oops in the driver-core, caused by the rtc driver
> > unregister, go away. The original issue is also fixed in the rtc driver
> > itself.
> 
> I don't think this is sufficient.
> 
> > --- a/drivers/base/driver.c
> > +++ b/drivers/base/driver.c
> > @@ -257,6 +257,8 @@ EXPORT_SYMBOL_GPL(driver_register);
> >   */
> >  void driver_unregister(struct device_driver *drv)
> >  {
> > +	if (!drv || !drv->p)
> > +		return;
> >  	driver_remove_groups(drv, drv->groups);
> >  	bus_remove_driver(drv);
> >  }
> 
> Ok, fine so far, but look at "driver_register()".
> 
> It will set drv->p, but then not unset it if it fails! (For a certain 
> class of failures)
> 
> So for a certain failure pattern, drv->p will point to some stale value. 
> Should we not clear drv->p in the "out_unregister" patch?
> 
> To confuse the thing more, there are actually "half-way failures" that 
> _succeed_ in driver registration, but then return an error code. See that 
> whole
> 
> 
> 	kobject_uevent(&priv->kobj, KOBJ_ADD);
> 	return error;
> 
> case in the "success" path driver_register(). We may return an error 
> despite the fact that we actually attached the driver to bus, but 
> "add_bind_files()" failed. A caller would be understandable very unhappy.
> 
> So I suspect we should do something like the appended (in addition to your 
> patch). Comments?

That looks good.  I'll add the WARN_ON that Ingo pointed out, and merge
this with Kay's patch.  Give me a few hours to wake up...

thanks,

greg k-h