Re: CONFIG_DEBUG_INFO_BTF and CONFIG_GCC_PLUGIN_RANDSTRUCT

From: Andrii Nakryiko <andrii.nakryiko@gmail.com>
To: Kees Cook <keescook@chromium.org>
Cc: Daniel Borkmann <daniel@iogearbox.net>,
	Jann Horn <jannh@google.com>,
	Alexei Starovoitov <alexei.starovoitov@gmail.com>,
	bpf <bpf@vger.kernel.org>,
	Kernel Hardening <kernel-hardening@lists.openwall.com>
Subject: Re: CONFIG_DEBUG_INFO_BTF and CONFIG_GCC_PLUGIN_RANDSTRUCT
Date: Tue, 31 Mar 2020 12:50:07 -0700	[thread overview]
Message-ID: <CAEf4BzYZsiuQGYVozwB=7nNhVYzCr=fQq6PLgHF3M5AXbhZyig@mail.gmail.com> (raw)
In-Reply-To: <202003311110.2B08091E@keescook>

On Tue, Mar 31, 2020 at 11:12 AM Kees Cook <keescook@chromium.org> wrote:
>
> On Tue, Mar 31, 2020 at 12:41:04AM +0200, Daniel Borkmann wrote:
> > On 3/30/20 7:20 PM, Kees Cook wrote:
> > > On Mon, Mar 30, 2020 at 06:17:32PM +0200, Jann Horn wrote:
> > > > On Mon, Mar 30, 2020 at 5:59 PM Alexei Starovoitov
> > > > <alexei.starovoitov@gmail.com> wrote:
> > > > > On Mon, Mar 30, 2020 at 8:14 AM Jann Horn <jannh@google.com> wrote:
> > > > > >
> > > > > > I noticed that CONFIG_DEBUG_INFO_BTF seems to partly defeat the point
> > > > > > of CONFIG_GCC_PLUGIN_RANDSTRUCT.
> > > > >
> > > > > Is it a theoretical stmt or you have data?
> > > > > I think it's the other way around.
> > > > > gcc-plugin breaks dwarf and breaks btf.
> > > > > But I only looked at gcc patches without applying them.
> > > >
> > > > Ah, interesting - I haven't actually tested it, I just assumed
> > > > (perhaps incorrectly) that the GCC plugin would deal with DWARF info
> > > > properly.
> > >
> > > Yeah, GCC appears to create DWARF before the plugin does the
> > > randomization[1], so it's not an exposure, but yes, struct randomization
> > > is pretty completely incompatible with a bunch of things in the kernel
> > > (by design). I'm happy to add negative "depends" in the Kconfig if it
> > > helps clarify anything.
> >
> > Is this expected to get fixed at some point wrt DWARF? Perhaps would make
>
> No, gcc closed the issue as "won't fix".
>
> > sense then to add a negative "depends" for both DWARF and BTF if the option
> > GCC_PLUGIN_RANDSTRUCT is set given both would be incompatible/broken.
>
> I hadn't just to keep wider randconfig build test coverage. That said, I
> could make it be: depends COMPILE_TEST || !DWARF ...
>
> I can certainly do that.

I've asked Slava in [0] to disable all three known configs that break
DWARF and subsequently BTF, I hope it's ok to just do it in one patch.
Currently all these appear to result in invalid BTF due to various
DWARF modifications:

  - DEBUG_INFO_REDUCED (see [1])
  - DEBUG_INFO_SPLIT (see [0]
  - GCC_PLUGIN_RANDSTRUCT (this discussion).

  [0] https://lore.kernel.org/bpf/CAEf4BzadnfAwfa1D0jZb=01Ou783GpK_U7PAYeEJca-L9kdnVA@mail.gmail.com/
  [1] https://lore.kernel.org/bpf/CAEf4BzZri8KpwLcoPgjiVx_=QmJ2W9UzBkDqSO2rUWMzWogkKg@mail.gmail.com/

>
> -Kees
>
> [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84052
>
> --
> Kees Cook