Re: [Y2038] [RFC PATCH v2 3/4] ppdev: add compat ioctl

[Arnd Bergmann <arnd@arndb.de>]

On Wednesday 08 July 2015 13:17:18 John Stultz wrote:
> On Mon, Jun 29, 2015 at 7:23 AM, Bamvor Zhang Jian
> <bamvor.zhangjian@linaro.org> wrote:
> > Add compat ioctl in ppdev in order to solve the y2038 issue in
> > later patch.
> > This patch simply add pp_do_ioctl to compat_ioctl, because I found
> > that all the ioctl access the arg as a pointer.
> >
> > Signed-off-by: Bamvor Zhang Jian <bamvor.zhangjian@linaro.org>

I just saw this mail fly by when you replied, but I guess it would
have been better to reply when the original mail came.

The description above makes no sense: The problem for compat ioctl
is not whether the argument is a pointer or not, but rather what
data structure it points to. In this case, we already know that 
it is /not/ compatible between 32-bit and 64-bit user space, because
at least two commands need special handling for the timespec
argument, which gets added in patch 4 of the series.

This means patches 3 and 4 have to be swapped in order to allow
bisection and not introduce a bug when only this one gets applied
but patch 4 is missing.

Moreover, all other ioctl commands that are handled in pp_ioctl
need to be checked regarding what their arguments are, including
data structures pointed to by the arguments (recursively, if there
are again pointers in those structures).

> >         unsigned int minor = iminor(inode);
> > @@ -744,6 +750,9 @@ static const struct file_operations pp_fops = {
> >         .write          = pp_write,
> >         .poll           = pp_poll,
> >         .unlocked_ioctl = pp_ioctl,
> > +#ifdef CONFIG_COMPAT
> > +       .compat_ioctl   = pp_compat_ioctl,
> > +#endif

The #ifdef here is not necessary, but will cause a warning on kernels 
that do not define CONFIG_COMPAT, in particular all 32-bit ones.

> Does adding this patch w/o the following patch break 32bit apps using
> this on 64bit kernels?

Without the patch, those apps will all get -EINVAL from the ioctl
handler. With the patch, the kernel actually performs the ioctl
that was requested, but that may use the wrong data structure.

	Arnd