Actually not anymore. The current CFS scheduler is no longer O(1) but O(logN):
Have one goroutine loop forever:
* wait for SIGCHLD
* take the childlock
* do nonblocking Wait() until no unwaited child remains
* release the childlock
The childlock would need to be taken for reading by os.Process.Kill and when a syscall that takes a PID of a child is called (after taking it we'd need to verify that the process we intend to touch isn't already dead).However, you don't need to use goroutines (or threads). You do it as you would in C (and how all real PID 1 systems are written) -- with a single thread that starts processes, receives signals, and reaps children in a non-blocking fashion.
This style of program -- a program that needs to simultaneously wait for child processes/signals and fd events -- is quite awkward in Unix, but it definitely works when you get the idea.
To wait on a fd and a signal in a single threaded program, you would use the "self pipe trick" in classic Unix. In Linux, you can ask for a signal to be delivered over a file descriptor with fdsignal(). But AFAICT there is no real reason, and portability across Unix IS a good thing IMO (but not portability to completely different OS's like Windows; in that case I would write a completely separate program using their native APIs).
node.js actually does a great job making this API easy and efficient. It is probably the only runtime (Python/Ruby/JVM/etc.), that doesn't suffer from this problem doing "async processes" (i.e. a complement to async networking).
In any case, you are not calling Wait4(<specific PID>), which is what implies the thread per process.
Re-read what I wrote. If that doesn't convince you, then download and run the code. Run "pstree" on it and observe how many child processes and threads there are. You'll learn something useful about the relationship of the Go runtime to the OS.
You can do it non-portably in Go by using os.ForkExec and Wait4(-1). The portable exec package assumes you will call Wait(pid), and not Wait(-1), which basically implies using a thread per process. Go's runtime isn't magic -- if you call libc/syscall wait(), an entire thread will be blocked, and the runtime can't use it for anything else. In this case this is the lifetime of an entire process, which is forever for server processes.
I'm pretty sure nobody would use a real PID 1 that burned a thread per process (systemd, upstart, etc.). But yes, for most use cases, in the grand scheme of things, it's probably not a big deal. I suppose Linux has an O(1) scheduler, although I'm not quite sure how this affects scheduling (interested in any comments).
But this goes to show that portable APIs are awkward and obscure for low level code. Better to use raw Unix APIs for something like an init server. Python and Java have similar problems.
IMO all interesting code nowadays is POSIX-like, so we should drop the pretension of portability and simplify our lives. Unix works.