stat: Use fstat when called with `-`/stdin as parameter (was MacOS CI: test_stat::test_stdin_pipe_fifo1/fifo2 are flaky)

[drinkcat]

Seeing a bunch of MacOS test failures recently, in tests that call stat -.

Sometimes in test_stdin_pipe_fifo1 (in that test, stdin is set to std::process::Stdio::piped()):

thread 'test_stat::test_stdin_pipe_fifo1' panicked at tests/by-util/test_stat.rs:349:10:
Command was expected to succeed. code: 1
stdout = 
 stderr = stat: cannot stat '-': Bad file descriptor (os error 9)

Sometimes in test_stdin_pipe_fifo2 (in that test stdin in set to std::process::Stdio::null()), where - shows up as a weird file, instead of a character special file

---- test_stat::test_stdin_pipe_fifo2 stdout ----
run: /Users/runner/work/coreutils/coreutils/target/aarch64-apple-darwin/debug/coreutils stat -

thread 'test_stat::test_stdin_pipe_fifo2' panicked at tests/by-util/test_stat.rs:366:10:
'  File: -
  Size: 0         	Blocks: 0          IO Block: 65536  weird file
Device: ffffffffbe5e09cfh/18446744072608418255d	Inode: 334         Links: 1
Access: (0666/?rw-rw-rw-)  Uid: (    0/    root)   Gid: (    0/   wheel)
Access: 2025-03-26 13:31:00.961845000 +0000
Modify: 2025-03-26 13:31:00.831524000 +0000
Change: 2025-03-26 13:31:00.831524000 +0000
 Birth: 1970-01-01 00:00:00.000000000 +0000
' does not contain 'character special file'

I'm a bit surprised that those tests are flaky (and not just always failing or passing). At least for the second one, we could try to understand what is that weird file.

[drinkcat]

In uutils, when - is passed as a parameter, we call fs::metadata on /dev/stdin. That seems to work well on Linux, but I suspect it's not portable.

Looked a bit at what Apple and OpenBSD stat command do for reference (https://github.com/apple-oss-distributions/file_cmds/blob/main/stat/stat.c#L352, https://github.com/openbsd/src/blob/master/usr.bin/stat/stat.c#L254):

• First, the calling syntax is a bit different: they do not accept - as parameter, but will look at stdin if no argument is provided.
• But then, they just call fstat on fd 0 (stdin), that's already open. I think GNU coreutils does the same (looking at strace, I see a statx system call on Linux).

So, I think the "correct" solution here would be to bypass the fs::metadata call when - is passed as a parameter, and make a direct call to libc::fstat with the stdin file description. It's a bit unfortunate to have to bypass core Rust libraries, but I don't see another option.

---

So, I still don't know why these tests have been starting to fail regularly on Mac, but I think it's probably just easier to disable them for now, and I'll keep this issue open to reimplement - handling.

[drinkcat]

There is some code in #7619 that we could probably use a starting point to fix this (so we can wait for that other PR first).

[oech3]

Is #9485 dupe of this issue?