event_base_loop: reentrant invocation

[ysbaddaden]

[warn] event_base_loop: reentrant invocation. Only one event_base_loop can run on each event_base at once.

It got the issue last week, but it seems it's still possible to happen:
#15 (comment)

That needs to be investigated.

[ysbaddaden]

Ideally the event loop would return the list of runnable fibers, so the schedulers can enqueue them in single operation (minus one to resume immediately), not execute a callback for each fiber that individually enqueues each fiber one by one, then have to dequeue one.

In the meantime, maybe the event loop callbacks could bypass Fiber#enqueue and call ExecutionContext::Scheduler#enqueue directly?

[ysbaddaden]

This is the problematic call:
https://github.com/ysbaddaden/execution_context/blob/main/src/single_threaded.cr#L156

Which means only ST is affected. MT prefers to switch to unblock the current fiber (to avoid delays if another thread picks it), while ST can avoid the switch.

[ysbaddaden]

Should be fixed by 647ed93