Re: [PATCH] fs/binfmt_elf: Add padding NULL when argc == 0

[Matthew Wilcox <willy@infradead.org>]

On Wed, Jan 26, 2022 at 11:58:39AM -0800, Kees Cook wrote:
> On Wed, Jan 26, 2022 at 08:50:39PM +0100, Jann Horn wrote:
> > On Wed, Jan 26, 2022 at 7:42 PM Ariadne Conill <ariadne@dereferenced.org> wrote:
> > > On Wed, 26 Jan 2022, Jann Horn wrote:
> > > > On Wed, Jan 26, 2022 at 6:58 PM Kees Cook <keescook@chromium.org> wrote:
> > > >> Quoting Ariadne Conill:
> > > >>
> > > >> "In several other operating systems, it is a hard requirement that the
> > > >> first argument to execve(2) be the name of a program, thus prohibiting
> > > >> a scenario where argc < 1. POSIX 2017 also recommends this behaviour,
> > > >> but it is not an explicit requirement[1]:
> > > >>
> > > >>     The argument arg0 should point to a filename string that is
> > > >>     associated with the process being started by one of the exec
> > > >>     functions.
> > > >> ...
> > > >> Interestingly, Michael Kerrisk opened an issue about this in 2008[2],
> > > >> but there was no consensus to support fixing this issue then.
> > > >> Hopefully now that CVE-2021-4034 shows practical exploitative use[3]
> > > >> of this bug in a shellcode, we can reconsider."
> > > >>
> > > >> An examination of existing[4] users of execve(..., NULL, NULL) shows
> > > >> mostly test code, or example rootkit code. While rejecting a NULL argv
> > > >> would be preferred, it looks like the main cause of userspace confusion
> > > >> is an assumption that argc >= 1, and buggy programs may skip argv[0]
> > > >> when iterating. To protect against userspace bugs of this nature, insert
> > > >> an extra NULL pointer in argv when argc == 0, so that argv[1] != envp[0].
> > > >>
> > > >> Note that this is only done in the argc == 0 case because some userspace
> > > >> programs expect to find envp at exactly argv[argc]. The overlap of these
> > > >> two misguided assumptions is believed to be zero.
> > > >
> > > > Will this result in the executed program being told that argc==0 but
> > > > having an extra NULL pointer on the stack?
> > > > If so, I believe this breaks the x86-64 ABI documented at
> > > > https://refspecs.linuxbase.org/elf/x86_64-abi-0.99.pdf - page 29,
> > > > figure 3.9 describes the layout of the initial process stack.
> > >
> > > I'm presently compiling a kernel with the patch to see if it works or not.
> > >
> > > > Actually, does this even work? Can a program still properly access its
> > > > environment variables when invoked with argc==0 with this patch
> > > > applied? AFAIU the way userspace locates envv on x86-64 is by
> > > > calculating 8*(argc+1)?
> > >
> > > In the other thread, it was suggested that perhaps we should set up an
> > > argv of {"", NULL}.  In that case, it seems like it would be safe to claim
> > > argc == 1.
> > >
> > > What do you think?
> > 
> > Sounds good to me, since that's something that could also happen
> > normally if userspace calls execve(..., {"", NULL}, ...).
> > 
> > (I'd like it even better if we could just bail out with an error code,
> > but I guess the risk of breakage might be too high with that
> > approach?)
> 
> We can't mutate argc; it'll turn at least some userspace into an
> infinite loop:
> https://sources.debian.org/src/valgrind/1:3.18.1-1/none/tests/execve.c/?hl=22#L22

How does that become an infinite loop?  We obviously wouldn't mutate
argc in the caller, just the callee.

Also, there's a version of this where we only mutate argc if we're
executing a setuid program, which would remove the privilege
escalation part of things.