ChildProcess design flaw: impossible to concurrently read a piped stream and kill the child

[matklad]

ChildProcess.kill not only kills the process, but also helpfully waits for it and closes the streams:

https://github.com/ziglang/zig/blob/197d9a9eb3557b1feae7580a82ab60defc9e12a1/lib/std/child_process.zig#L234-L236 https://github.com/ziglang/zig/blob/197d9a9eb3557b1feae7580a82ab60defc9e12a1/lib/std/child_process.zig#L386 https://github.com/ziglang/zig/blob/197d9a9eb3557b1feae7580a82ab60defc9e12a1/lib/std/child_process.zig#L420-L423

Unfortunately, this DWIM behavior makes some useful programs impossible (as far as I am aware, but maybe I am missing something).

Consider a program which spawns a child with stderr=.Pipe, then spawns a thread to read out stderr, and then, some time later, decides to kill the child on the main thread.

When one thread kills the child, the other is using the stderr file descriptor. As killing also waits and closes the stderr file descriptor, the reading thread might end up operating on an invalid FD!. As far as I am aware, a correct solution here would be an intricate dance of:

• main thread sigkills the child
• stderr reader thread gets unblocked via receiving EOF for the stderr stream.
• main thread joins the stderr reader thread to guarantee that the file descriptor is no longer in use
• main thread waits for the child process and closes file descriptros

I think this can be worked-around by duping the relevant FD, but that's an extra FD!

[matklad]

Luckily, this seems easy to work-arond, by killing the child manually:

pub fn deinit(tb: *TmpTigerBeetle, gpa: std.mem.Allocator) void {
    // Signal to the `stderr_reader_thread` that it can exit
    // TODO(Zig) https://github.com/ziglang/zig/issues/16820
    if (builtin.os.tag == .windows) {
        std.os.windows.TerminateProcess(tb.process.handle, 1) catch {};
    } else {
        std.os.kill(tb.process.pid, std.os.SIG.TERM) catch {};
    }

    tb.stderr_reader_thread.join();
    tb.stderr_reader.destroy(gpa);
    _ = tb.process.wait() catch unreachable;
    tb.tmp_dir.cleanup();
}