fix(unix): retry EINTR in poll/select/read

[jh-block]

Fixes #212.

read_single_key treats any Err(Interrupted) from read_single_key_impl as Ctrl-C and forwards it via libc::raise(SIGINT). That's correct only for the synthesized Err(Interrupted) read_bytes returns when it sees a \x03 byte. EINTR returned by poll/select/read because of an unrelated signal (SIGCHLD, SIGWINCH, etc.) takes the same path and incorrectly kills the program with SIGINT.

Retry EINTR inside the syscall wrappers so Err(Interrupted) reaching read_single_key unambiguously means a \x03 byte was read.

[djc]

Thanks for the clippy fix -- can you split it out in a separate commit please?

For the first hunk's loop, I think we could do break instead of return at the end? Seems slightly nicer.

[djc]

This does not seem like a clean split. Also, the first commit looks like it makes some unrelated/unnecessary changes. If you are using an LLM to help with this, please self-review before asking for my review.

[jh-block]

Can you clarify what you mean? The only perhaps-unrelated change in the first commit is reducing the scope of the unsafe block in select_fd to the actual unsafe code, rather than covering the whole method with it, which seemed appropriate since otherwise the entire loop (which is just normal safe logic) would be within the unsafe block. Do you want me to revert that, or split it into a separate commit, or are you referring to something else?

[jh-block]

@djc is there any chance you could take another look at the two commits? I am not sure how to interpret your earlier comment; the split of the clippy fix and the logic fix into two commits is clean, and the first commit (the EINTR logic fix) contains only the changes I intended to make. If you would like something changed can you indicate what it is?

[djc]

This hunk from the first commit seems like an unnecessary increase in scope of the unsafe block?

[djc]

Nit: suggest writing this as

match err.kind() {
    io::ErrorKind::Interrupted => continue,
    _ => break Err(err),
}

[djc]

Same here.