[PATCH v3 12/26] compat_ioctl: move more drivers to compat_ptr_ioctl

[Arnd Bergmann]

On Thu, Apr 25, 2019 at 11:25 PM Johannes Berg
<johannes at sipsolutions.net> wrote:
> On Thu, 2019-04-25 at 17:55 +0200, Arnd Bergmann wrote:
> > On Thu, Apr 25, 2019 at 5:35 PM Al Viro <viro at zeniv.linux.org.uk> wrote:
> > >
> > > On Thu, Apr 25, 2019 at 12:21:53PM -0300, Mauro Carvalho Chehab wrote:
> > >
> > > > If I understand your patch description well, using compat_ptr_ioctl
> > > > only works if the driver is not for s390, right?
> > >
> > > No; s390 is where "oh, just set ->compat_ioctl same as ->unlocked_ioctl
> > > and be done with that; compat_ptr() is a no-op anyway" breaks.  IOW,
> > > s390 is the reason for having compat_ptr_ioctl() in the first place;
> > > that thing works on all biarch architectures, as long as all stuff
> > > handled by ->ioctl() takes pointer to arch-independent object as
> > > argument.  IOW,
> > >         argument ignored => OK
> > >         any arithmetical type => no go, compat_ptr() would bugger it
> > >         pointer to int => OK
> > >         pointer to string => OK
> > >         pointer to u64 => OK
> > >         pointer to struct {u64 addr; char s[11];} => OK
> >
> > To be extra pedantic, the 'struct {u64 addr; char s[11];} '
> > case is also broken on x86, because sizeof (obj) is smaller
> > on i386, even though the location of the members are
> > the same. i.e. you can copy_from_user() this
>
> Actually, you can't even do that because the struct might sit at the end
> of a page and then you'd erroneously fault in this case.
>
> We had this a while ago with struct ifreq, see commit 98406133dd and its
> parents.

Yes, you are right. Very rare to hit with real-life code, but easily
reproduced by intentionally hitting it and clearly a bug.

As the saying goes

  | the difference between "always works" and "almost always works"
  | is called data corruption

here the difference is an -EFAULT.

      Arnd