Re: [PATCH] usb: ulpi: call put_device if device_register fails

[Greg Kroah-Hartman <gregkh@linuxfoundation.org>]

On Wed, Jul 22, 2015 at 09:04:40PM -0500, Felipe Balbi wrote:
> On Wed, Jul 22, 2015 at 02:39:34PM -0700, Greg Kroah-Hartman wrote:
> > On Tue, Jun 23, 2015 at 01:57:38PM +0300, Heikki Krogerus wrote:
> > > On Fri, Jun 19, 2015 at 01:12:36AM +0800, ChengYi He wrote:
> > > > put_device is required to release the last reference to the device.
> > > > 
> > > > Signed-off-by: ChengYi He <chengyihetaipei@gmail.com>
> > > > ---
> > > >  drivers/usb/common/ulpi.c | 4 +++-
> > > >  1 file changed, 3 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/drivers/usb/common/ulpi.c b/drivers/usb/common/ulpi.c
> > > > index 0e6f968..bd25bdb 100644
> > > > --- a/drivers/usb/common/ulpi.c
> > > > +++ b/drivers/usb/common/ulpi.c
> > > > @@ -184,8 +184,10 @@ static int ulpi_register(struct device *dev, struct ulpi *ulpi)
> > > >  	request_module("ulpi:v%04xp%04x", ulpi->id.vendor, ulpi->id.product);
> > > >  
> > > >  	ret = device_register(&ulpi->dev);
> > > > -	if (ret)
> > > > +	if (ret) {
> > > > +		put_device(&ulpi->dev);
> > > 
> > > If device_register returns failure, put_device has already been
> > > called. Check device_add in drivers/base/core.c.
> > 
> > Yes, please read the function, which says:
> >  * NOTE: _Never_ directly free @dev after calling this function, even
> >  * if it returned an error! Always use put_device() to give up your
> >  * reference instead.
> > 
> > But, the problem is that the ulpi core doesn't "own" that struct device.
> > It comes from elsewhere.  It comes from somewhere deep down in the dw3
> > core, which is where I lost the path.  Something needs to be fixed in
> > dwc3_probe() to properly clean up the device if it fails, which is not
> > happening right now.
> > 
> > So this patch would actually cause much bigger problems than fixing
> > anything, so it's wrong, but for a different reason than you are talking
> > about here.
> > 
> > And ugh, the ulpi and dwc code binding together, what a mess, horrid...
> 
> any suggestions ? DWC *is* the one implementing the bus. If there's a
> better way, we can certainly shuffle code around.

As dwc is the only thing using the bus, why is it drivers/usb/core/ ?

And the error path here is broken, the bus should be creating the device
(i.e. no subsystem should ever be registering a device it did not
create), so that it can properly clean things up when stuff goes wrong.

The whole subsys_init() is also a bad feeling that it's not architected
correctly, that shouldn't be needed, which is why I never took that
patch.  Just noticed it came in through yours, I wanted it "broken" so
it would be fixed "properly" and not papered over like this.

thanks,

greg k-h