Re: [Linux-kernel-mentees] [PATCH v3] media/v4l2-core: Fix kernel-infoleak in video_put_user()

[Arnd Bergmann <arnd@arndb.de>]

On Thu, Jul 30, 2020 at 10:38 AM Andy Shevchenko
<andriy.shevchenko@linux.intel.com> wrote:
> On Thu, Jul 30, 2020 at 10:15:24AM +0200, Arnd Bergmann wrote:
> > On Thu, Jul 30, 2020 at 10:07 AM Bartosz Golaszewski <brgl@bgdev.pl> wrote:
> > >
> > > On Tue, Jul 28, 2020 at 3:58 PM Arnd Bergmann <arnd@arndb.de> wrote:
> > > >
> > > > On Tue, Jul 28, 2020 at 3:06 PM Dan Carpenter <dan.carpenter@oracle.com> wrote:
> > > > Something like
> > > >
> > > > static int lineevent_put_data(void __user *uptr, struct gpioevent_data *ge)
> > > > {
> > > > #ifdef __x86_64__
> > > >         /* i386 has no padding after 'id' */
> > > >         if (in_ia32_syscall()) {
> > > >                 struct {
> > > >                         compat_u64      timestamp __packed;
> > > >                         u32             id;
> > > >                 } compat_ge = { ge->timestamp, ge->id };
> > > >
> > > >                 if (copy_to_user(uptr, &compat_ge, sizeof(compat_ge)))
> > > >                         return -EFAULT;
> > > >
> > > >                 return sizeof(compat_ge);
> > > >         }
> > > > #endif
> > > >
> > > >         if (copy_to_user(uptr, ge, sizeof(*ge))
> > > >                 return -EFAULT;
> > > >
> > > >         return sizeof(*ge);
> > > > }
> > > >
> > > >        Arnd
> > >
> > > Hi Arnd,
> > >
> > > Andy actually had a patch for that but since this isn't a regression
> > > (it never worked), we decided to leave it as it is and get it right in
> > > v2 API.
> >
> > I would argue that it needs to be fixed anyway, unless you also want
> > to remove the v1 interface for native mode. If this works on 32-bit
> > kernels, on 64-bit kernels with 64-bit user space and on compat
> > 32-bit user space on 64-bit non-x86 architectures, I see no reason
> > to leave it broken specifically on x86 compat user space. There are
> > still reasons to use 32-bit x86 user space for low-memory machines
> > even though native i386 kernels are getting increasingly silly.
>
> It was possible to "fix" (mitigate to some extent) before libgpiod got support
> for several events in a request. Now it seems to be impossible to fix. AFAIU we
> must discard any request to more than one event in it.

Any reason why the workaround I suggested above would not work?
The in_ia32_syscall() check should be completely reliable in telling whether
we are called from read() by an ia32 task or not, and we use the same
logic for input_event, which has a similar problem (on all compat architectures,
not just x86).

> However I'm not an expert in compat IOCTL code (you are :-) and perhaps you may
> provide ideas better than mine.

What makes this interface tricky is that this is actually a read() call, not
ioctl() which is usually easier because it encodes the data length in the
command code. As far as I could tell from skimming the interface, the
ioctls are actually fine here.

    Arnd