Re: What to do on cdev_add failure

From: Jean Delvare
Date: Fri Jul 22 2016 - 05:18:33 EST

Hi Greg,

On Thu, 14 Jul 2016 15:47:35 +0900, Greg KH wrote:
> On Wed, Jul 13, 2016 at 03:46:14PM +0200, Jean Delvare wrote:
> > Hi all,
> >
> > I am currently working on the i2c-dev driver, which has just been
> > converted to the non-ancestral cdev API. As I am cleaning up the
> > driver, I would like to switch from static cdev initialization
> > (cdev_init) to dynamic allocation (cdev_alloc.)
> >
> > While I was looking at other drivers to figure out how to deal with
> > error cases, I found that different drivers do different things if
> > cdev_add fails after cdev_alloc was called successfully. I guess some
> > of them are right, others are wrong, and I'd like to know which is
> > which ;-)
> >
> > * char/virtio_console.c, s390/char/tape_class.c, s390/char/vmur.c,
> > infiniband/.../qib_file_ops.c, fuse/cuse.c, scsi/sg.c and scsi/st.g
> > are calling cdev_del(cdev).
> > * v4l2-core/v4l2-dev.c is calling kfree(cdev).
> > * s390/char/vmlogrdr.c, uio/uio.c, tty/ty_io.c and __register_chrdev()
> > are calling kobject_put(&cdev->kobj). The former explicitly says "no
> > cdev_del here!" in a comment.
> >
> > My gut feeling is that kobject_put(&cdev->kobj) is correct, even though
> > it feels strange to have to use a low-level function to clean-up after
> > a higher level API call.
> >
> > If cdev_del(cdev) is also correct (and as I read the code it could be,
> > iff calling kobj_unmap() is a no-op if kobj_map() failed - is it the
> > case?), then it should be clearly documented as such, as it is
> > counter-intuitive (to me, at least.)
> >
> > Anyone wants to comment on this?
> >
> > On top of this, another thing looks strange to me. cdev_add() only gets
> > the parent kobj on success. However the release methods
> > (cdev_default_release and cdev_dynamic_release) will put the parent
> > kobj unconditionally. So it looks to me that we are over-putting the
> > parent whenever cdev_add() fails. OTOH I can't see where the parent is
> > set. If it is NULL then all these get and put are no-ops to start with
> > and it no longer matters. But why would we be doing that?
> >
> > Then again, what do I know about kobj black magic...
> It's worse than you think, the kobject in a cdev is not a "real"
> kobject. Well, it's kind of real, but it's only there to be used for
> the kmap logic. I have a 10+ year old TODO item here that says "remove
> kobj from cdev" that I really should get to one of these days.

I did figure that out actually.

> Anyone that touches the kobj outside of the cdev core code is probably
> wrong, it's "funny" that both uio and tty do that, the maintainer of
> that code must be lazy... :)

Or just as confused as myself.

> Let me look into what the "correct" thing to do here is, I used to know
> it, need some time to refresh my memory...

Did you find?

> And the cdev interface has what, 4 different ways it can be used?
> Another of my TODO items is to fix it all up to only use it one way, or
> maybe just 2 as it does have the ability to make driver code pretty
> small if you use it in unique ways...

Jean Delvare
SUSE L3 Support