On Tue, Nov 7, 2017 at 3:36 PM, Laura Abbott <labbott@redhat.com> wrote:
>
> I'd probably put /proc/kallsyms and /proc/modules on the omit list
> since those are designed to leak addresses to userspace.

Well, they are indeed designed to leak addresses, but not a lot of
people should care.

So I think we could tighten them up.

For example, maybe /proc/kallsyms could just default to not showing
values to non-root users.

We *did* originally try to use "kptr_restrict" with a default value of
1, it's just that it was never fixable on a case-by-case basis as
people started saying "that breaks my flow, because xyz".

But if we do it for one file at a time, we probably *can* try to fix complaints.

Something like the attached TOTALLY UNTESTED patch. It's meant more as
an RFC, not for application, but it's also meant to show how we can
tailor the behavior for specific workflow issues.

So take that "kallsyms_for_perf()" thing as an example of how we can
say "hey, if you already have access to kernel profiling anyway,
there's no point in hiding kallsyms".

And there may be other similar things we can do.

The situation with /proc/modules should be similar. Using
kptr_restrict was a big hammer and might have broken something
unrelated, but did anybody actually care about the particular case of
/proc/modules not showing the module address to normal users? probably
not. "lsmod" certainly doesn't care, and that's what people really
want.

Both /proc/kallsyms and /proc/modules _used_ to be really important
for oops reporting, but that was long ago when the kernel didn't
report symbol information of its own. So we have historical reasons
for people to be able to read those files, but those are mainly things
that aren't relevant (or even possible) on modern kernels anyway.

So I don'r think we should omit /proc/kallsyms and /proc/modules - we
should just fix them.

The attached patch may not be good enough as is, but maybe something
_like_ it will work well enough that people won't care?

(And do note the "TOTALLY UNTESTED". It seems to compile. But maybe I
got some test exactly the wrong way around and it doesn't actually
_work_. Caveat testor).

                     Linus