Re: uhid: broken interface: 32/64-bit compatibility

[Johan Hedberg <johan.hedberg-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>]

Hi David,

On Fri, Feb 15, 2013, David Herrmann wrote:
> On Fri, Feb 15, 2013 at 12:29 PM, Kirill A. Shutemov
> <kirill.shutemov-VuQAYsv1563Yd54FQh9/CA@public.gmane.org> wrote:
> > Hi David and all,
> >
> > There's claim in uhid.h that the interface is "compatible even between
> > architectures". But it obviously is not true: struct uhid_create_req
> > contains pointer which breaks everything.
> >
> > The easy way to demonstrate the issue is compile uhid-example.c with -m32
> > and try to run it on 64 bit kernel. Creating of the device will fail.
> 
> Indeed, we missed that. We should probably also notify the HIDP
> developers as "struct hidp_connadd_req" suffers from the same
> problems. (CC'ed)
> 
> > I don't see an easy way to fix this. Few options:
> >
> > 1. Replace the pointer with u64. It will fix the issue, but it breaks ABI
> >    which is never a good idea. Not sure how many users interface already
> >    has.
> 
> The only users I am aware of is an HID debugging tool and experimental
> HoG Bluetooth support (bluez). Maybe Marcel or Johan can comment
> whether this is already used by bluez-5? If it is, then we shouldn't
> break ABI and go with #2+#3. Otherwise, I think changing to u64 should
> be ok.
> On the other hand, it would break any future build for older stable
> kernels so not breaking ABI is probably the best idea. Any comments? I
> can add a COMPAT fix and a comment to fix this in the next version of
> UHID_CREATE.

The HoG code in BlueZ 5 does indeed use this API and it's also not
anymore behind any kind of experimental flag (i.e. it is an officially
supported feature).

Johan