Re: [PATCH v2] fs/exec: require argv[0] presence in do_execveat_common()

[Ariadne Conill]

Hi,

On Wed, 26 Jan 2022, Kees Cook wrote:

> On Wed, Jan 26, 2022 at 11:44:47AM +0000, Ariadne Conill wrote:
> > 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[0]:
> >
> >     The argument arg0 should point to a filename string that is
> >     associated with the process being started by one of the exec
> >     functions.
> >
> > To ensure that execve(2) with argc < 1 is not a useful gadget for
> > shellcode to use, we can validate this in do_execveat_common() and
> > fail for this scenario, effectively blocking successful exploitation
> > of CVE-2021-4034 and similar bugs which depend on this gadget.
> >
> > The use of -EFAULT for this case is similar to other systems, such
> > as FreeBSD, OpenBSD and Solaris.  QNX uses -EINVAL for this case.
> >
> > Interestingly, Michael Kerrisk opened an issue about this in 2008[1],
> > but there was no consensus to support fixing this issue then.
> > Hopefully now that CVE-2021-4034 shows practical exploitative use
> > of this bug in a shellcode, we can reconsider.
> >
> > [0]: https://pubs.opengroup.org/onlinepubs/9699919799/functions/exec.html
> > [1]: https://bugzilla.kernel.org/show_bug.cgi?id=8408
> >
> > Changes from v1:
> > - Rework commit message significantly.
> > - Make the argv[0] check explicit rather than hijacking the error-check
> >   for count().
> >
> > Signed-off-by: Ariadne Conill <ariadne@xxxxxxxxxxxxxxxx>
> > ---
> >  fs/exec.c | 4 ++++
> >  1 file changed, 4 insertions(+)
> >
> > diff --git a/fs/exec.c b/fs/exec.c
> > index 79f2c9483302..e52c41991aab 100644
> > --- a/fs/exec.c
> > +++ b/fs/exec.c
> > @@ -1899,6 +1899,10 @@ static int do_execveat_common(int fd, struct filename *filename,
> >  	retval = count(argv, MAX_ARG_STRINGS);
> >  	if (retval < 0)
> >  		goto out_free;
> > +	if (retval == 0) {
> > +		retval = -EFAULT;
> > +		goto out_free;
> > +	}
> >  	bprm->argc = retval;
> >
> >  	retval = count(envp, MAX_ARG_STRINGS);
> > --
> > 2.34.1
>
> Okay, so, the dangerous condition is userspace iterating through envp
> when it thinks it's iterating argv.
>
> Assuming it is not okay to break valgrind's test suite:
> https://sources.debian.org/src/valgrind/1:3.18.1-1/none/tests/execve.c/?hl=22#L22
> we cannot reject a NULL argv (test will fail), and we cannot mutate
> argc=0 into argc=1 (test will enter infinite loop).
>
> Perhaps we need to reject argv=NULL when envp!=NULL, and add a
> pr_warn_once() about using a NULL argv?

Sure, I can rework the patch to do it for only the envp != NULL case.

I think we should combine it with the {NULL, NULL} padding patch in this 
case though, since it appears to work, that way the execve(..., NULL, 
NULL) case gets some protection.

> I note that glibc already warns about NULL argv:
> argc0.c:7:3: warning: null argument where non-null required (argument 2)
> [-Wnonnull]
>    7 |   execve(argv[0], NULL, envp);
>      |   ^~~~~~
>
> in the future we could expand this to only looking at argv=NULL?

I don't think musl's headers generate a diagnostic for this, but main(0, 
{NULL}) is not a supported use-case at least as far as Alpine is 
concerned.  I am sure it is the same with the other musl distributions.

Will send a v3 patch with this logic change and move to EINVAL shortly.

Ariadne