Re: [PATCH] seccomp: Check flags on seccomp_notif is unset

[Aleksa Sarai]

On 2019-12-27, Sargun Dhillon <sargun@xxxxxxxxx> wrote:
> On Fri, Dec 27, 2019 at 6:47 AM Christian Brauner
> <christian.brauner@xxxxxxxxxx> wrote:
> >
> > On Fri, Dec 27, 2019 at 01:31:31PM +1100, Aleksa Sarai wrote:
> > > On 2019-12-27, Aleksa Sarai <cyphar@xxxxxxxxxx> wrote:
> > >
> > > Scratch that -- as Tycho just mentioned, there is un-named padding in
> > > the struct so check_zeroed_user() is the wrong thing to do. But this
> >
> > Hm, I don't think so.
> > I understood Tycho's point as _if_ there ever is padding then this would
> > not be zeroed.
> > Right now, there is no padding since the struct is correctly padded:
> >
> > struct seccomp_data {
> >         int nr;
> >         __u32 arch;
> >         __u64 instruction_pointer;
> >         __u64 args[6];
> > };
> >
> > struct seccomp_notif {
> >         __u64 id;
> >         __u32 pid;
> >         __u32 flags;
> >         struct seccomp_data data;
> > };
> >
> > which would be - using pahole:
> >
> > struct seccomp_data {
> >         int                        nr;                   /*     0     4 */
> >         __u32                      arch;                 /*     4     4 */
> >         __u64                      instruction_pointer;  /*     8     8 */
> >         __u64                      args[6];              /*    16    48 */
> >
> >         /* size: 64, cachelines: 1, members: 4 */
> > };
> > struct seccomp_notif {
> >         __u64                      id;                   /*     0     8 */
> >         __u32                      pid;                  /*     8     4 */
> >         __u32                      flags;                /*    12     4 */
> >         struct seccomp_data data;                        /*    16    64 */
> >
> >         /* size: 80, cachelines: 2, members: 4 */
> >         /* last cacheline: 16 bytes */
> > };
> >
> > The only worry would be a 2byte int type but there's no architecture
> > we support which does this right now afaict.
> >
> > > also will make extensions harder to deal with because (presumably) they
> > > will also have un-named padding, making copy_struct_from_user() the
> >
> > This all will be a non-issue if we just use __u64 for extensions.
> >
> > My point about using copy_struct_from_user() was that we should verify
> > that _all_ fields are uninitialized and not just the flags argument
> > since we might introduce a flags argument that requires another already
> > existing member in seccomp_notif to be set to a value. We should do this
> > change now so we don't have to risk breaking someone in the future.
> >
> > I'm trying to get at least Mozilla/Firefox off of their crazy
> > SECCOMP_RET_TRAP way of implementing their broker onto the user notifier
> > and they will likely need some extensions. That includes the pidfd stuff
> > for seccomp that Sargun will likely be doing and the new pidfd_getfd()
> > syscall. So it's not unlikely that we might need other already existing
> > fields in that struct to be set to some value.
> >
> > I don't particulary care how we do it:
> > - We can do a simple copy_from_user() and check each field individually.
> 
> Just doing a simple copy_from_user, and for now, calling memchr_inv
> on the whole thing. We can drop the memset, and just leave a note to
> indicate that if unpadded fields are introduced in the future, this structure
> must be manually zeroed out. Although, this might be laying a trap for
> ourselves.
> 
> This leaves us in a good position for introducing a flag field in the future.
> All we have to do is change the memchr_inv from checking on an
> entire struct basis to checking on a per-field basis.

There is no need to do memchr_inv() on copy_from_user() to check for
zero-ness. That's the entire point of check_zeroed_user() -- to not need
to do it that way.

> > - Use copy_struct_from_user().
> >   That is safe to do right now since there is no padding afaict and
> >   it'll automatically verify new fields as well.
> >   If I understand the worry correctly then the argument against
> >   copy_struct_from_user() here is that there might be padding introduced
> >   and userspace will not do an explicit memset() but rather rely on an
> >   empty inializer {} and will _accidently_ pass down a struct which has
> >   __all fields cleared__ but __uninitialized padding__ and we tell them
> >   EINVAL? That can only happen if we introduce padding in the struct
> >   which I'd argue we just don't do. That'll be in line with what we
> >   require from our ABIs already anyway.


-- 
Aleksa Sarai
Senior Software Engineer (Containers)
SUSE Linux GmbH
<https://www.cyphar.com/>

Attachment: signature.asc
Description: PGP signature