Functions to execute processes, pass file descriptors around and return results
The low-level Unix.fork(), Unix.exec*() functions and friends are not safe to call in multithreaded programs for two reasons:
-
parallel threads opening new file descriptors will have these descriptors captured by a fork(). This leads to annoying glitches like (for example) attempts to 'umount' a filesystem being rejected because a file is still open.
-
although Unix.fork() will call (via the ocaml runtime) a pthread_atfork handler which attempts to clean up the state of the threading system in the child, this relies on quite a complex glibc implementation which has been observed to fail occasionally (typical symptom is the process spins after the fork() but before the exec())
Additionally Unix.fork(), Unix.exec*() are very low-level primitives. When we call these functions what we actually want to do is run some separate process with certain file-descriptors, optionally returning results.
The interface in this module
- is higher-level than Unix.fork(), Unix.exec*()
- allows us to offload Unix.fork(), Unix.exec*() to a single-threaded separate process where the glibc+ocaml runtime codepaths are simpler and hopefully more reliable.