timeout: replace the busy loop with sigtimedwait (#9099)

[andreacorbellini]

Instead of checking for the process status in a busy loop, use sigtimedwait to wait for the process to exit.

sigtimedwait is a POSIX method to suspend the current process until a specified set of signals is received. The signals we're interested in are:

• SIGCHLD: this is sent whenever the child process quits, or is suspended (via SIGSTOP) or is resumed (via SIGCONT)
• SIGTERM: so that it can be forwarded to the child process (should this also be done for SIGINT, SIGHUP, and others? If yes, it can be added in a future PR)

This PR also does the following:

• Removes all uses of cfg(unix) because currently timeout only works on Unix. For Windows support, I suggest moving the current logic in a unix.rs module so that we don't need to prefix each and every function with cfg(unix).
• Replaces integer signals with the Signal type from the nix crate wherever possible.
• Resolves all the TODO items related to exit statuses (exit codes were based on the output of GNU's timeout --help).
• Removes the SIGTERM signal handler: now everything is caught via sigtimedwait.

Sorry for the large refactor and for addressing multiple issues in the same commit!

[github-actions]

GNU testsuite comparison:

GNU test failed: tests/timeout/timeout-large-parameters. tests/timeout/timeout-large-parameters is passing on 'main'. Maybe you have to rebase?
GNU test failed: tests/timeout/timeout-parameters. tests/timeout/timeout-parameters is passing on 'main'. Maybe you have to rebase?
Skip an intermittent issue tests/misc/usage_vs_getopt (fails in this run but passes in the 'main' branch)
Skip an intermittent issue tests/tail/overlay-headers (fails in this run but passes in the 'main' branch)
Skip an intermittent issue tests/timeout/timeout (fails in this run but passes in the 'main' branch)

[github-actions]

GNU testsuite comparison:

GNU test failed: tests/timeout/timeout-parameters. tests/timeout/timeout-parameters is passing on 'main'. Maybe you have to rebase?
Skip an intermittent issue tests/misc/usage_vs_getopt (fails in this run but passes in the 'main' branch)
Skip an intermittent issue tests/tail/overlay-headers (fails in this run but passes in the 'main' branch)
Skip an intermittent issue tests/timeout/timeout (fails in this run but passes in the 'main' branch)

[github-actions]

GNU testsuite comparison:

Skip an intermittent issue tests/misc/usage_vs_getopt (fails in this run but passes in the 'main' branch)
Skip an intermittent issue tests/timeout/timeout (fails in this run but passes in the 'main' branch)

[Alonely0]

This API appears not to be available on macOS, tho afaik the one I had suggested (setitimer, iirc) does. Also, there are multiple ones on Linux, with different time resolutions (seconds/microseconds/nanoseconds iirc). We should use the one with the greatest resolution on Linux, which iirc is timer_create(2); on macOS use the other one that's available, iirc setitimer(2); and probably also add a alarm(2) as a fallback, because this is the lowest common denominator in POSIX, which takes time in seconds, so we would have to convert it to seconds and take the nearest positive integer.

[Alonely0]

We honestly should be using the nix crate where possible, I'll replace this with nix's API and we should automatically gain macOS support; if not, I'll make sure to handle platform-specific cases.

[Alonely0]

The best way is probably going to be to use self-pipe tricks with signal handlers and pass the timeout to select()... Already on it.
EDIT: It's probably even better to just use pselect with the needed signal mask, LOL. It's POSIX 2001, but everybody has implemented it because it's crucial for efficient servers.

[andreacorbellini]

We honestly should be using the nix crate where possible, I'll replace this with nix's API and we should automatically gain macOS support; if not, I'll make sure to handle platform-specific cases.

Agreed, but nix does not support sigtimedwait as of today.

The best way is probably going to be to use self-pipe tricks with signal handlers and pass the timeout to select()... Already on it.

For MacOS: maybe. I do think however that sigtimedwait is the best approach on modern systems and we should stick to that when supported. Although I can understand the burden of having multiple implementations that do the same thing, so if that's a concern I guess I'm fine with dropping sigtimedwait

[Alonely0]

Got it fixed with select and self-pipes, I'll submit a PR after I got it tested properly.

EDIT: replaced it with pselect due to the greater code complexity and slightly higher latency of self-pipes and select.