[RFC] Add BPF_PROG_TYPE_CGROUP_IOCTL

[Alexei Starovoitov]

On Mon, Nov 02, 2020 at 02:23:02PM -0500, Kenny Ho wrote:
> Adding a few more emails from get_maintainer.pl and bumping this
> thread since there hasn't been any comments so far.  Is this too
> crazy?  Am I missing something fundamental?

sorry for delay. Missed it earlier. Feel free to ping the mailing list
sooner next time.

> On Wed, Oct 7, 2020 at 11:24 AM Kenny Ho <Kenny.Ho at amd.com> wrote:
> >
> > This is a skeleton implementation to invite comments and generate
> > discussion around the idea of introducing a bpf-cgroup program type to
> > control ioctl access.  This is modelled after
> > BPF_PROG_TYPE_CGROUP_DEVICE.  The premise is to allow system admins to
> > write bpf programs to block some ioctl access, potentially in conjunction
> > with data collected by other bpf programs stored in some bpf maps and
> > with bpf_spin_lock.
> >
> > For example, a bpf program has been accumulating resource usaging
> > statistic and a second bpf program of BPF_PROG_TYPE_CGROUP_IOCTL would
> > block access to previously mentioned resource via ioctl when the stats
> > stored in a bpf map reaches certain threshold.
> >
> > Like BPF_PROG_TYPE_CGROUP_DEVICE, the default is permissive (i.e.,
> > ioctls are not blocked if no bpf program is present for the cgroup.) to
> > maintain current interface behaviour when this functionality is unused.
> >
> > Performance impact to ioctl calls is minimal as bpf's in-kernel verifier
> > ensure attached bpf programs cannot crash and always terminate quickly.
> >
> > TODOs:
> > - correct usage of the verifier
> > - toolings
> > - samples
> > - device driver may provide helper functions that take
> > bpf_cgroup_ioctl_ctx and return something more useful for specific
> > device
> >
> > Signed-off-by: Kenny Ho <Kenny.Ho at amd.com>
...
> > @@ -45,6 +46,10 @@ long vfs_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
> >         if (!filp->f_op->unlocked_ioctl)
> >                 goto out;
> >
> > +       error = BPF_CGROUP_RUN_PROG_IOCTL(filp, cmd, arg);
> > +       if (error)
> > +               goto out;
> > +

That's a bit problematic, since we have bpf_lsm now.
Could you use security_file_ioctl hook and do the same filtering there?
It's not cgroup based though. Is it a concern?
If cgroup scoping is really necessary then it's probably better
to add it to bpf_lsm. Then all hooks will become cgroup aware.