On Wed, Mar 22, 2023 at 04:53:42PM +0100, Laszlo Ersek wrote: ...
What is supposed to happen is this:
(1) bash shall find test-execvp in the current directory per PATH, (2) execvp() shall find "hello.sh" in the current directory per PATH, (3) execvp() shall hit an internal failure -1/ENOEXEC, (4) execvp() shall then invoke the shell (under an unspecified pathname), (5) the shell shall get "foobar" for its argv[0], and "hello.sh" for its argv[1] (6) we shall see "hello" on the standard output.
That's exactly what happens on Linux/glibc. (Note: this result has absolutely nothing to do with my execvpe() implementation, or libnbd in the first place.)
Now, according to my above description of the busybox bug, we're tempted to believe that step (6) fails on Alpine Linux (using musl + busybox). We expect the busybox binary to be launched, via the /bin/sh symlink, in step (4), and we expect it to fail after step (5), due to it not recognizing "foobar" as an "applet name".
It turns out however that step (4) does not happen. musl does not handle ENOEXEC:
It's getting crazier by the hour.
No joke. Locally, I wrote: $ cat foo echo "hello from $0:$@" $ chmod +x foo $ bash ./foo foobar arg1 hello from ./foo:foobar arg1 $ bash -c ./foo foobar arg1 hello from ./foo: $ bash -c 'exec -a "$0" ./foo "$@"' foobar arg1 hello from foobar:arg1 $ bash -c '. ./foo' foobar arg1 hello from foobar:arg1 The latter two match the behavior we want to see. Alas, POSIX has not yet standardized 'exec -a', but the ability to use 'sh -c' to set argv0, plus . to source a file as shell commands, gets us to the same place.
I thought to create a reproducer for busybox, in spite of musl breaking down at an earlier point (see above). For that, I *statically linked* "test-execvp" on RHEL-9.1 (using glibc), and then executed the binary in the Alpine Linux container. This should eliminate musl from the picture, and exercise the ENOEXEC fallback (from glibc), invoking /bin/sh (aka busybox) under argv[0] "foobar", and trigger the "unknown applet" bug in busybox.
However, this does not happen. Instead, I get "hello" printed. How is that possible?
The solution is that glibc *too* has a bug, and that bug hides the busybox bug. Namely, in glibc, going back to historical commit
commit 6a032d81581978187f562e5533a32e0a6a3d352b (tag: cvs/libc-960210) Author: Roland McGrath <roland@gnu.org> Date: Sat Feb 10 10:00:27 1996 +0000
Sat Feb 10 04:18:48 1996 Roland McGrath <roland@churchy.gnu.ai.mit.edu>
* posix/execvp.c: If execv fails with ENOEXEC, run the shell on the file.
Fri Feb 9 11:46:45 1996 Roland McGrath <roland@churchy.gnu.ai.mit.edu>
* time/Makefile (CFLAGS-zdump.c, CFLAGS-zic.c, CFLAGS-ialloc.c, CFLAGS-scheck.c): Use -DNOID instead of -Wno-unused.
* hurd/Makefile (user-interfaces): Added hurd/tioctl.
(note the date: 1996!), the POSIX-mandated fallback
execv(<shell path>, { argv[0], file, argv[1], ..., NULL })
is not being done. Instead, the following is done:
execv(<shell path>, { <shell path>, file, argv[1], ..., NULL })
In other words, the original argv[0] is not preserved, but is replaced by <shell path>. (Look for _PATH_BSHELL in said historical glibc commit, and also in today's glibc file "posix/execvpe.c".)
Wow. This is so historically old that it's probably worth asking on the Austin Group if argv[0] must be preserved or can be changed to the shell name, and/or whether the standard should add support for 'exec -a argv0 cmd args' as a way to set argv[0] despite this long-standing glibc behavior. I'll reply back when I have some URLs to list archives of my pending questions.
This can be demonstrated with:
$ PATH=.:$PATH strace -etrace=execve test-execvp
execve("./test-execvp", ["test-execvp"], 0x7ffc0d1e5248 /* 85 vars */) = 0 execve("./hello.sh", ["foobar"], 0x7ffc528e14a8 /* 85 vars */) = -1 ENOEXEC (Exec format error) execve("/bin/sh", ["/bin/sh", "./hello.sh"], 0x7ffc528e14a8 /* 85 vars */) = 0 hello +++ exited with 0 +++
The third execve() call should be:
execve("/bin/sh", ["foobar", "./hello.sh"], 0x7ffc528e14a8 /* 85 vars */) = 0 ^^^^^^^^
Laszlo
-- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3266 Virtualization: qemu.org | libvirt.org