[Toybox] faccessat and AT_SYM_NOFOLLOW

[Rich Felker]

This message is a follow-up to a discussion Rob Landley and I had on
#musl regarding musl's returning an error of EINVAL when the
AT_SYM_NOFOLLOW flag is passed to faccessat (a nonstandard usage),
which is affecting toybox. Rob's idea for using it came from the Linux
man pages, which document this flag as supported and do not make it
clear that it's glibc, not Linux, providing the support.

Issue 1: Is the inclusion of AT_SYM_NOFOLLOW in the man page a total
documentation error (not actually supported by glibc at all) or just a
failure to mark it as a glibc extension?

Here's the relevant glibc file:

https://sourceware.org/git/?p=glibc.git;a=blob;f=sysdeps/unix/sysv/linux/faccessat.c;h=4a6048ec7930c8fc249ee629b1d2618cd81084b0;hb=HEAD

As I read it:

1. It sets EINVAL for flags that are invalid, but considers
   AT_SYM_NOFOLLOW valid. So far so good...

2. It flag is 0 or the only flag set is AT_EACCESS and the binary is
   not suid, it just makes the syscall directly. OOPS, big bug -- it
   does not honor changes to the effective uid made by programs
   initially started as root!

3. Otherwise, it uses fstatat, possibly with the AT_SYM_NOFOLLOW flag,
   to get the file ownership and mode and performs its own access
   permissions check in userspace. This is imprecise and does not
   respect ACLs or any other advanced permission models provided by
   LSMs etc.

I was not even aware that this imprecise emulation was used in the
AT_SYM_NOFOLLOW case; I figured they would do it in an exact way,
which is possible with some /proc/self/fd tricks.

So my conclusion? There are some moderate-level documentation errors.
glibc implements the flag, but not correctly. The changes I would
recommend to the documentation:

1. Document that AT_SYM_NOFOLLOW is not standard for this function,
   and is a glibc extension. (uclibc is just a copy of glibc code)

2. Document that AT_SYM_NOFOLLOW and AT_EACCESS are emulated and
   unreliable on glibc.

3. Document that the man page is covering the POSIX/glibc function
   details, and the kernel syscall does not support flags at all.
   (This might aid in getting the kernel folks to add a new faccessat4
   syscall that would do flags at the kernel level.)

Do these sound reasonable?

Issue 2: Should musl support or ignore the AT_SYM_NOFOLLOW with
faccessat?

If anyone would like to see support, I can consider it for a future
agenda item (and of course: patches welcome), but I'm not going to add
an implementation that reports incorrect information. Just ignoring
the flag would give the wrong results (possibly dangerous) and would
leave the application no way to know that the flag was ignored
(whereas failing with EINVAL makes it clear, and is explicitly
documented as an optional error condition for invalid flags).

However:

Issue 3: Does the AT_SYM_NOFOLLOW flag even make sense with faccessat?

The whole point of AT_SYM_NOFOLLOW is to avoid TOCTOU races checking
for symlinks. However, any use of faccessat is fundamentally a TOCTOU
race -- the information it obtained is not necessarily correct by the
time it returns. Reducing two TOCTOU issues in succession to one
TOCTOU issue does not seem useful for most purposes.

Issue 4: Does faccessat, with or without AT_SYM_NOFOLLOW, make sense
for implementing rm -rf?

No. At best it's a wasted syscall that slows you down, and at worst it
gives you wrong information. The efficient and correct implementation
is to simply _try_ the operation which might fail (openat?) and only
change file permissions and retry if it failed with EACCESS. This
requires zero extra syscalls in the success case (versus one extra
with faccessat).

As for AT_SYM_NOFOLLOW, I'm not clear why it was being used. Even if
the inaccessible file is a symlink target, you'd need to use
AT_SYM_NOFOLLOW or similar when doing the fchmodat, and that one's
what actually protects you from races used to trick rm into changing
permissions on files it shouldn't. Use of the result from faccessat to
make the decision is a TOCTOU race.

Rich