On Tue, Mar 05, 2024 at 07:34:34PM +0000, Adrian Ratiu wrote:
> On Tuesday, March 05, 2024 20:37 EET, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>
> > On Tue, Mar 05, 2024 at 11:32:04AM +0100, Christian Brauner wrote:
> > > On Tue, Mar 05, 2024 at 02:12:26AM -0800, Kees Cook wrote:
> > > > On Tue, Mar 05, 2024 at 10:58:25AM +0100, Christian Brauner wrote:
> > > > > Since the write handler for /proc/<pid>/mem does raise FOLL_FORCE
> > > > > unconditionally it likely would implicitly. But I'm not familiar enough
> > > > > with FOLL_FORCE to say for sure.
> > > >
> > > > I should phrase the question better. :) Is the supervisor writing into
> > > > read-only regions of the child process?
> > >
> > > Hm... I suspect we don't. Let's take two concrete examples so you can
> > > tell me.
> > >
> > > Incus intercepts the sysinfo() syscall. It prepares a struct sysinfo
> > > with cgroup aware values for the supervised process and then does:
> > >
> > > unix.Pwrite(siov.memFd, &sysinfo, sizeof(struct sysinfo), seccomp_data.args[0]))
> > >
> > > It also intercepts some bpf system calls attaching bpf programs for the
> > > caller. If that fails we update the log buffer for the supervised
> > > process:
> > >
> > > union bpf_attr attr = {}, new_attr = {};
> > >
> > > // read struct bpf_attr from mem_fd
> > > ret = pread(mem_fd, &attr, attr_len, req->data.args[1]);
> > > if (ret < 0)
> > > return -errno;
> > >
> > > // Do stuff with attr. Stuff fails. Update log buffer for supervised process:
> > > if ((new_attr.log_size) > 0 && (pwrite(mem_fd, new_attr.log_buf, new_attr.log_size, attr.log_buf) != new_attr.log_size))
> >
> > This is almost certainly in writable memory (either stack or .data).
>
> Mostly yes, but we can't be certain where it comes from, because
> SECCOMP_IOCTL_NOTIF_RECV passes any addresses set by the
> caller to the supervisor process.
>
> It is a kind of "implementation defined" behavior, just like we
> can't predict what the supervisor will do with the caller mem :)
>
> >
> > > But I'm not sure if there are other use-cases that would require this.
> >
> > Maybe this option needs to be per-process (like no_new_privs), and with
> > a few access levels:
> >
> > - as things are now
> > - no FOLL_FORCE unless by ptracer
> > - no writes unless by ptracer
> > - no FOLL_FORCE ever
> > - no writes ever
> > - no reads unless by ptracer
> > - no reads ever
> >
> > Which feels more like 3 toggles: read, write, FOLL_FORCE. Each set to
> > "DAC", "ptracer", and "none"?
>
> I really like this approach because it provides a mechanism
> with maximum flexibility without imposing a specific policy.
>
> What does DAC mean in this context? My mind jumps to
> Digital to Analog Converter :)
Ah yes, sorry, this is Discretionary Access Control (which is my
short-hand for saying "basic file permissions"). But I guess that's kind
of not really true since the open() access checks are doing a
"ptrace-able" check in addition to the file perms check.
> Shall I give it a try in v3?
Yeah, though maybe see if Mike or Jann chime in over the next few days?
-Kees
--
Kees Cook
