lkml.org 
[lkml]   [2020]   [May]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] seccomp: Add group_leader pid to seccomp_notif
On Sun, May 17, 2020 at 12:17:14AM -0700, Kees Cook wrote:
> On Fri, May 15, 2020 at 04:40:05PM -0700, Sargun Dhillon wrote:
> > This includes the thread group leader ID in the seccomp_notif. This is
> > immediately useful for opening up a pidfd for the group leader, as
> > pidfds only work on group leaders.
> >
> > Previously, it was considered to include an actual pidfd in the
> > seccomp_notif structure[1], but it was suggested to avoid proliferating
> > mechanisms to create pidfds[2].
> >
> > [1]: https://lkml.org/lkml/2020/1/24/133
> > [2]: https://lkml.org/lkml/2020/5/15/481
>
> nit: please use lore.kernel.org/lkml/ URLs
>
> > Suggested-by: Christian Brauner <christian.brauner@ubuntu.com>
> > Signed-off-by: Sargun Dhillon <sargun@sargun.me>
> > ---
> > include/uapi/linux/seccomp.h | 2 +
> > kernel/seccomp.c | 1 +
> > tools/testing/selftests/seccomp/seccomp_bpf.c | 50 +++++++++++++++++++
> > 3 files changed, 53 insertions(+)
> >
> > diff --git a/include/uapi/linux/seccomp.h b/include/uapi/linux/seccomp.h
> > index c1735455bc53..f0c272ef0f1e 100644
> > --- a/include/uapi/linux/seccomp.h
> > +++ b/include/uapi/linux/seccomp.h
> > @@ -75,6 +75,8 @@ struct seccomp_notif {
> > __u32 pid;
> > __u32 flags;
> > struct seccomp_data data;
> > + __u32 tgid;
> > + __u8 pad0[4];
> > };
>
> I think we need to leave off padding and instead use __packed. If we
> don't then userspace can't tell when "pad0" changes its "meaning" (i.e.
> the size of seccomp_notif becomes 88 bytes with above -- either via
> explicit padding like you've got or via implicit by the compiler. If
> some other u32 gets added in the future, user space will still see "88"
> as the size.
>
> So I *think* the right change here is:
>
> -};
> + __u32 tgid;
> +} __packed;
>
> Though tgid may need to go above seccomp_data... for when it grows.
> Agh...
>
> _However_, unfortunately, I appear to have no thought this through very
> well, and there is actually no sanity-checking in the kernel for dealing
> with an old userspace when sizes change. :( For example, if a userspace
> doesn't check sizes and calls an ioctl, etc, the kernel will clobber the
> user buffer if it's too small.

I'd just argue that that's just userspace messing up.

>
> Even the SECCOMP_GET_NOTIF_SIZES command lacks a buffer size argument.
> :(
>
> So:
>
> - should we just declare such userspace as "wrong"? I don't think
> that'll work, especially since what if we ever change the size of
> seccomp_data... that predated the ..._SIZES command.

Yeah, that's nasty since the struct member in seccomp_notif would now
clobber each other.

>
> - should we add a SECCOMP_SET_SIZES command to tell the kernel what
> we're expecting? There's no "state" associated across seccomp(2)
> calls, but maybe that doesn't matter because only user_notif writes
> back to userspace. For the ioctl, the state could be part of the
> private file data? Sending seccomp_data back to userspace only
> happens here, and any changes in seccomp_data size will just be seen
> as allowing a filter to query further into it.

Sounds ok-ish in my opinion.

>
> - should GET_SIZES report "useful" size? (i.e. exclude padding?)

Or... And that's more invasive but ultimately cleaner we v2 the whole
thing so e.g. SECCOMP_IOCTL_NOTIF_RECV2, SECCOMP_IOCTL_NOTIF_SEND2, and
embedd the size argument in the structs. Userspace sets the size
argument, we use get_user() to get the size first and then
copy_struct_from_user() to handle it cleanly based on that. A similar
model as with sched (has other unrelated quirks because they messed up
something too):

static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
{
u32 size;
int ret;

/* Zero the full structure, so that a short copy will be nice: */
memset(attr, 0, sizeof(*attr));

ret = get_user(size, &uattr->size);
if (ret)
return ret;

/* ABI compatibility quirk: */
if (!size)
size = SCHED_ATTR_SIZE_VER0;
if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
goto err_size;

ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
if (ret) {
if (ret == -E2BIG)
goto err_size;
return ret;
}

We're probably the biggest user of this right now and I'd be ok with
that change. If it's a v2 than whatever. :)

Christian

\
 
 \ /
  Last update: 2020-05-17 12:48    [W:0.112 / U:0.068 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site