Lwt.pause deadlock on spt

[Firobe]

open Lwt.Syntax

module Hello (T : Mirage_time.S) = struct
    let f1c = ref 0
    let f2c = ref 0

    let rec f1 () = 
        Logs.info (fun f -> f "f1 (%d)" !f1c);
        incr f1c;
        let* () = Lwt.pause () in
        f1 ()

    let rec f2 () = 
        Logs.info (fun f -> f "f2 (%d)" !f2c);
        incr f2c;
        let* () = Lwt.pause () in
        f2 ()

    let start _ =
        Logs.info (fun f -> f "Start");
        let* _ = Lwt.all [f1 (); f2 ()] in Lwt.return_unit
end

The above unikernel, when compiled and executed for the hvt target, produces the expected output:

2024-09-19T09:27:21-00:00: [INFO] [application] Start
2024-09-19T09:27:21-00:00: [INFO] [application] f2 (0)
2024-09-19T09:27:21-00:00: [INFO] [application] f1 (0)
2024-09-19T09:27:21-00:00: [INFO] [application] f2 (1)
2024-09-19T09:27:21-00:00: [INFO] [application] f1 (1)
2024-09-19T09:27:21-00:00: [INFO] [application] f2 (2)
2024-09-19T09:27:21-00:00: [INFO] [application] f1 (2)
2024-09-19T09:27:21-00:00: [INFO] [application] f2 (3)
2024-09-19T09:27:21-00:00: [INFO] [application] f1 (3)
2024-09-19T09:27:21-00:00: [INFO] [application] f2 (4)
2024-09-19T09:27:21-00:00: [INFO] [application] f1 (4)
2024-09-19T09:27:21-00:00: [INFO] [application] f2 (5)
2024-09-19T09:27:21-00:00: [INFO] [application] f1 (5)
2024-09-19T09:27:21-00:00: [INFO] [application] f2 (6)
2024-09-19T09:27:21-00:00: [INFO] [application] f1 (6)
2024-09-19T09:27:21-00:00: [INFO] [application] f2 (7)
2024-09-19T09:27:21-00:00: [INFO] [application] f1 (7)
2024-09-19T09:27:21-00:00: [INFO] [application] f2 (8)
2024-09-19T09:27:21-00:00: [INFO] [application] f1 (8)
[...]

However when compiled and executed for the spt target, fibers that have have "paused" are never woken up, and the application deadlocks:

2024-09-19T09:32:32-00:00: [INFO] [application] Start
2024-09-19T09:32:32-00:00: [INFO] [application] f2 (0)
2024-09-19T09:32:32-00:00: [INFO] [application] f1 (0)
[no further logs]

[reynir]

It seems to me in this case we end up here with a timeout of 0L and call solo5_yield:

mirage-solo5/lib/main.ml

Lines 51 to 58 in 0d9b87b

	let timeout =
	if Lwt.paused_count () > 0 then 0L
	else
	match Time.select_next () with
	| None -> Int64.add (Time.time ()) (Duration.of_day 1)
	| Some tm -> tm
	in
	let ready_set = solo5_yield timeout in

Then here I lose track of what happens (it calls some assembly)
https://github.com/Solo5/solo5/blob/5478c255911e1dae6c9ab1f448daddf0fb46ea37/bindings/spt/net.c#L91-L114

[reynir]

Ok, it seems if you call timerfd_settime() with all zeroes it disarms the timer. I'll look into a fix in solo5 spt bindings...

[hannesm]

What does hvt do in contrast? Shouldn't that be similar code?

[reynir]

So hvt had a similar issue:
Solo5/solo5@97a66f6

I tried applying the same fix in reynir/solo5@683529b but in my test it doesn't seem to fix it - but then again when I run strace it seems to call with a zero it_value so I'm doubting my test...

[reynir]

It was dune cache that was causing trouble...