Deprecate the Lwt_unix.async_method machinery and eventually remove it

[aantron]

Here is a description of it:

lwt/src/unix/lwt_unix.cppo.mli

Lines 68 to 87 in 0aa5063

	(** For system calls that cannot be made asynchronously, Lwt uses one
	of the following method: *)
	type async_method =
	| Async_none
	(** System calls are made synchronously, and may block the
	entire program. *)
	| Async_detach
	(** System calls are made in another system thread, thus without
	blocking other Lwt threads. The drawback is that it may
	degrade performance in some cases.
	This is the default. *)
	| Async_switch
	(** System calls are made in the main thread, and if one blocks
	the execution continue in another system thread. This method
	is the most efficient, also you will get better performance
	if you force all threads to run on the same cpu. On linux
	this can be done by using the command [taskset].
	Note that this method is still experimental. *)

Of the three methods, Async_switch, AFAICT, was only ever working on Linux, but might be broken now (#184). Since it is non-portable, marked as experimental, probably broken, and unsupported, I think we should deprecate it and eventually remove it.

Of the remaining two methods, Async_none and Async_detach, Async_detach is the reason why someone would want to use Lwt at all, over Unix. In case we bind a system call that is not available in blocking form in Unix, we can expose both synchronous and asynchronous bindings from Lwt. So, we should also remove Async_none, and thus the whole async_method type.

IIRC, setting the async method explicitly is basically not used in opam. I would have to look again to see what the exact impact is.

[chambart]

I'm all for removing Async_switch. It was a neat trick, but is a too brittle to seriously consider keeping. But Async_none has its uses. There are not that many system call that are completely blocking for an unbounded time. Most of the things that doesn't have a usable non-blocking equivalent are disk IO, which are often fast enough not to care. In particular, if the disk cache has a good hit-rate, Async_none is faster than Async_detach. It matters when you are CPU bound rather than IO bound.
But the main reason for me to use it is that it is often far easier to debug with GDB with Async_none.

By the way the only way I use it is through the environment variable.

[aantron]

Thanks. I'll see about keeping Async_none somehow, maybe through a simpler API, or just the environment variable.

[chambart]

Thanks, I would be happy with only the environment variable

[aantron]

I looked in more detail, and I am now leaning towards making Async_switch be effectively a synonym for Async_detach in 5.0.0, and leaving Async_none working as now. I'll experiment with Async_none some more, and perhaps un-deprecate it. Once the code implementing Async_switch is gone, it should also be much clearer what to do next (probably nothing :)).

[adrianbrink]

What is the current status of Async_none?

[aantron]

It's unchanged from before this issue was opened.

[adrianbrink]

Does that mean that I can still use Lwt_unix.async_method Async_none even in 5.0.0?