Hi,
On Fri, Jul 24, 2020 at 02:31:19PM -0500, Eric W. Biederman wrote:
> Michael,
>
> As the original author of NS_GET_OWNER_UID can you take a look at this?
This is a gentle reminder that my patch hasn't been applied,
the problem reported by Ákos Uzonyi hasn't been fixed,
and the example in ioctl_ns(2) manual page doesn't work
when e.g. it's compiled with -m32 on a 64-bit kernel.
> "Dmitry V. Levin" <ldv@xxxxxxxxxxxx> writes:
> > On Fri, Jul 24, 2020 at 11:20:26AM +0200, Arnd Bergmann wrote:
> >> On Fri, Jul 24, 2020 at 2:12 AM Dmitry V. Levin wrote:
> >> >
> >> > According to Documentation/driver-api/ioctl.rst, in order to support
> >> > 32-bit user space running on a 64-bit kernel, each subsystem or driver
> >> > that implements an ioctl callback handler must also implement the
> >> > corresponding compat_ioctl handler. The compat_ptr_ioctl() helper can
> >> > be used in place of a custom compat_ioctl file operation for drivers
> >> > that only take arguments that are pointers to compatible data
> >> > structures.
> >> >
> >> > In case of NS_* ioctls only NS_GET_OWNER_UID accepts an argument, and
> >> > this argument is a pointer to uid_t type, which is universally defined
> >> > to __kernel_uid32_t.
> >>
> >> This is potentially dangerous to rely on, as there are two parts that
> >> are mismatched:
> >>
> >> - user space does not see the kernel's uid_t definition, but has its own,
> >> which may be either the 16-bit or the 32-bit type. 32-bit uid_t was
> >> introduced with linux-2.3.39 in back in 2000. glibc was already
> >> using 32-bit uid_t at the time in user space, but uclibc only changed
> >> in 2003, and others may have been even later.
> >>
> >> - the ioctl command number is defined (incorrectly) as if there was no
> >> argument, so if there is any user space that happens to be built with
> >> a 16-bit uid_t, this does not get caught.
> >
> > Note that NS_GET_OWNER_UID is provided on 32-bit architectures, too, so
> > this 16-bit vs 32-bit uid_t issue was exposed to userspace long time ago
> > when NS_GET_OWNER_UID was introduced, and making NS_GET_OWNER_UID
> > available for compat processes won't make any difference, as the mismatch
> > is not between native and compat types, but rather between 16-bit and
> > 32-bit uid_t types.
> >
> > I agree it would be correct to define NS_GET_OWNER_UID as
> > _IOR(NSIO, 0x4, uid_t) instead of _IO(NSIO, 0x4), but nobody Cc'ed me
> > on this topic when NS_GET_OWNER_UID was discussed, and that ship has long
> > sailed.
> >
> >> > This change fixes compat strace --pidns-translation.
> >> >
> >> > Note: when backporting this patch to stable kernels, commit
> >> > "compat_ioctl: add compat_ptr_ioctl()" is needed as well.
> >> >
> >> > Reported-by: Ákos Uzonyi <uzonyi.akos@xxxxxxxxx>
> >> > Fixes: 6786741dbf99 ("nsfs: add ioctl to get an owning user namespace for ns file descriptor")
> >> > Cc: stable@xxxxxxxxxxxxxxx # v4.9+
> >> > Signed-off-by: Dmitry V. Levin <ldv@xxxxxxxxxxxx>
> >> > ---
> >> > fs/nsfs.c | 1 +
> >> > 1 file changed, 1 insertion(+)
> >> >
> >> > diff --git a/fs/nsfs.c b/fs/nsfs.c
> >> > index 800c1d0eb0d0..a00236bffa2c 100644
> >> > --- a/fs/nsfs.c
> >> > +++ b/fs/nsfs.c
> >> > @@ -21,6 +21,7 @@ static long ns_ioctl(struct file *filp, unsigned int ioctl,
> >> > static const struct file_operations ns_file_operations = {
> >> > .llseek = no_llseek,
> >> > .unlocked_ioctl = ns_ioctl,
> >> > + .compat_ioctl = compat_ptr_ioctl,
> >> > };
> >> >
> >> > static char *ns_dname(struct dentry *dentry, char *buffer, int buflen)
>
> Thank you,
> Eric
--
ldv
