Re: [musl] Re: Tweaking the program name for <err.h> functions

[Rich Felker <dalias@libc.org>]

On Tue, Mar 12, 2024 at 12:18:24AM +0000, Gabriel Ravier wrote:
> On 3/11/24 19:47, Rich Felker wrote:
> >On Mon, Mar 11, 2024 at 11:30:04AM -0400, Skyler Ferrante (RIT Student) wrote:
> >>Hmm, maybe I'm missing something, but it seems you can close(fd) for
> >>the standard fds and then call execve, and the new process image will
> >>have no fd 0,1,2. I've tried this on a default Ubuntu 22.04 system.
> >>This seems to affect shadow-utils and other setuid/setgid binaries.
> >>
> >>Here is a repo I built for testing,
> >>https://github.com/skyler-ferrante/fd_omission/. What is the correct
> >>glibc behavior? Am I misunderstanding something?
> >As Florian noted, you're missing that strace cannot invoke it suid.
> >POSIX explicitly permits the implementation to open these fds if they
> >started closed in suid execs, and IIRC indicates as a future direction
> >that it might be permitted for all execs. We do the same in musl in
> >the suid case. So really the only way that "writing attacker
> >controlled prefix strings to fd 2" becomes an issue is if the
> >application erroneously closes fd 2 and lets something else get opened
> >on it.
> >
> >(Aside: making _FORTIFY_SOURCE>1 trap close(n) with n<3 would be an
> >interesting idea... :)
> 
> Doing this would break many programs, such as:
> - most of coreutils, e.g. programs like ls, cat or head, since they
> always `close` their input and output descriptors (when they've
> written or read something) to make sure to diagnose all errors
> - grep
> - xargs
> - find

This makes it so they can malfunction during exit when it
flushes/closes the corresponding stdio FILEs. If nothing else has been
opened in the mean time, under typical implementations it should be
safe, but I think per 2.5.1 Interaction of File Descriptors and
Standard I/O Streams:

https://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_05_01

it's undefined.

The safe way to do what they want is to dup the fd they want to
close-and-check-for-errors, open /dev/null, dup2 that over the
original fd, then close the first dup.

Or, don't exit()/return-from-main, but instead _exit, so there's no
subsequent access to the FILE.

> - strace, which (using the half-closed self-pipe trick mentioned
> earlier in this thread to avoid reusing them later btw) closes the
> standard descriptors, to avoid changing the behavior of programs
> calling it if e.g. its input is a pipe (where if it left the fds
> open that'd mean the writer would get SIGPIPE later than if the
> program was ran without strace)
> - tcsh, which deliberately does `close(n)` with `n < 3` to make it
> so all the standard FDs point to `/dev/null`
> - troff and groff (and thus man)
> - git
> - many more... I have found these by simply stracing random programs
> as found on my system with `ls /bin/ | shuf -n1`

Yes, I'm quite aware it's commonplace, but it would be something nice
to get cleaned up...

Rich