Re: [RESEND RFC] kernel/ksysfs.c: restrict /sys/kernel/notes to root access

[Jann Horn <jannh@google.com>]

On Sun, Feb 18, 2024 at 8:47 AM Greg KH <gregkh@linuxfoundation.org> wrote:
>
> On Sun, Feb 18, 2024 at 03:35:01PM +0800, Guixiong Wei wrote:
> > From: Guixiong Wei <weiguixiong@bytedance.com>
> >
> > Restrict non-privileged user access to /sys/kernel/notes to
> > avoid security attack.
> >
> > The non-privileged users have read access to notes. The notes
> > expose the load address of startup_xen. This address could be
> > used to bypass KASLR.
>
> How can it be used to bypass it?
>
> KASLR is, for local users, pretty much not an issue, as that's not what
> it protects from, only remote ones.
>
> > For example, the startup_xen is built at 0xffffffff82465180 and
> > commit_creds is built at 0xffffffff810ad570 which could read from
> > the /boot/System.map. And the loaded address of startup_xen is
> > 0xffffffffbc265180 which read from /sys/kernel/notes. So the loaded
> > address of commit_creds is 0xffffffffbc265180 - (0xffffffff82465180
> >  - 0xffffffff810ad570) = 0xffffffffbaead570.
>
> I've cc: the hardening list on this, I'm sure the developers there have
> opinions about this.
>
> > Signed-off-by: Guixiong Wei <weiguixiong@bytedance.com>
> > ---
> >  kernel/ksysfs.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/kernel/ksysfs.c b/kernel/ksysfs.c
> > index b1292a57c2a5..09bc0730239b 100644
> > --- a/kernel/ksysfs.c
> > +++ b/kernel/ksysfs.c
> > @@ -199,7 +199,7 @@ static ssize_t notes_read(struct file *filp, struct kobject *kobj,
> >  static struct bin_attribute notes_attr __ro_after_init  = {
> >       .attr = {
> >               .name = "notes",
> > -             .mode = S_IRUGO,
> > +             .mode = S_IRUSR,
> >       },
> >       .read = &notes_read,
> >  };
>
> No objection from me, but what userspace tool requires access to this
> file today?  Will it break if permissions are changed on it?

FWIW, https://codesearch.debian.net/search?q=%2Fsys%2Fkernel%2Fnotes&literal=1
shows that (from what I can tell) this is mostly used by stuff like
perf, libdwfl and systemtap for figuring out the kernel's build-id,
probably mostly for kernel profiling?