Re: [PATCH] fs/binfmt_elf.c: allocate initialized memory in fill_thread_core_info()

[Jann Horn]

On Tue, Apr 21, 2020 at 6:05 PM Yu-cheng Yu <yu-cheng.yu@xxxxxxxxx> wrote:
> On Tue, 2020-04-21 at 17:09 +0200, Jann Horn wrote:
> > +x86 folks
> >
> > (rest of thread is on lore
> > <https://lore.kernel.org/lkml/20200419100848.63472-1-glider@xxxxxxxxxx/>;,
> > with original bug report on github
> > <https://github.com/google/kmsan/issues/76>;)
> >
> > On Tue, Apr 21, 2020 at 2:54 PM Alexander Potapenko <glider@xxxxxxxxxx> wrote:
> > > On Tue, Apr 21, 2020 at 5:42 AM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> > > > On Mon, Apr 20, 2020 at 03:41:40PM -0700, Kees Cook wrote:
> > > > > On Mon, Apr 20, 2020 at 03:33:52PM -0700, Andrew Morton wrote:
> > > > > > On Sun, 19 Apr 2020 12:08:48 +0200 glider@xxxxxxxxxx wrote:
> > > > > >
> > > > > > > KMSAN reported uninitialized data being written to disk when dumping
> > > > > > > core. As a result, several kilobytes of kmalloc memory may be written to
> > > > > > > the core file and then read by a non-privileged user.
> > > > >
> > > > > Ewww. That's been there for 12 years. Did something change in
> > > > > regset_size() or regset->get()? Do you know what leaves the hole?
> > > >
> > > > Not lately and I would also like to hear the details; which regset it is?
> > > > Should be reasonably easy to find - just memset() the damn thing to something
> > > > recognizable, do whatever triggers that KMSAN report and look at that
> > > > resulting coredump.
> > > >
> > >
> > > Seems to be REGSET_XSTATE filled by xstateregs_get().
> > > Is there a ptrace interface also using that function?
> >
> > It looks to me like the problem KMSAN found is that
> > copy_xstate_to_kernel() will not fill out memory for unused xstates? I
> > think this may have been introduced by commit 91c3dba7dbc1
> > ("x86/fpu/xstate: Fix PTRACE frames for XSAVES", introduced in v4.8).
> >
> > There seem to be no other functions that reach that path other than
> > coredumping; I think the correct fix would be to change
> > copy_xstate_to_kernel() to always fully initialize the output buffer.
>
> Yes, that makes sense.  On the other hand, the kzalloc() fix prevents potential
> similar problems for other regsets.

I don't really have anything against using kzalloc() there; but in my
opinion that's not a fix, that's hardening. The real problem, in my
opinion, is that regset->get() claims to have filled out a buffer
without actually having done so; and if someone happens to add another
caller to that thing in the future, I don't want them to run into
exactly the same problem again.