Add std::os::unix::{register, unregister, Interest}

[leo60228]

Closes #293, based on #293 (comment)

[yoshuawuyts]

@leo60228 thanks so much for starting this work! I think this is an interesting step. I've got a few design notes though:

• We don't want to expose mio::Evented as part of the public interface.
• Any design here should provide people access to the epoll instance that we're running; do you have an idea how you would go about this?
• Why did you choose to expose Entry instead of Watcher?

I'm thinking exposing std::os::unix::Watcher might make more sense than exposing std::os::Entry for most cases. In many ways Entry is purely an implementation detail that we don't want to expose to end-users. I really like the idea of having an async_std::os::unix version of our watcher abstraction, and another for async_std::os::windows. That way we can differentiate between RawFd and RawSocket in constructor interfaces.

edit: to answer my own question: the design in #293 (comment) doesn't mention Watcher but instead only talks about register/unregister/Interest. I guess that's what this is based on. Though I do feel that having a Watcher would likely be closer to what we would want to expose.

[yoshuawuyts]

Evented shouldn't be part of our pub interface, and fd should be RawFd.

[leo60228]

This module is pub(crate), and RawFd is Unix-only. I didn't want to copy/paste the method and struct, though perhaps that's preferable for memory usage? It'd be heavily preferable to use Option<!> instead of having to put every call in a cfg, though.

[yoshuawuyts]

Evented shouldn't be part of our pub interface.

[leo60228]

This module is pub(crate).

[sdroege]

I really like the idea of having an async_std::os::unix version of our watcher abstraction, and another for async_std::os::windows. That way we can differentiate between RawFd and RawSocket in constructor interfaces.

Note that you will also want to support RawHandle on Windows, e.g. for event objects.

[leo60228]

How does that work? Is there an EventedHandle in mio like there is EventedFd?

[leo60228]

mio doesn't seem to have a way to register a RawHandle.

[leo60228]

FWIW, I switched from impl AsRawFd to RawFd because RawFd doesn't implement AsRawFd.

[Demi-Marie]

Would it be possible to also expose methods on UdpSocket et al? I really don’t want to write more code than I must for Windows.

[Keruspe]

It would be great if we could have access to what was the actual event.
With the current implementation, the waker gets called, but no way of knowing if the event was for read or write (and how are treated the error, read_hup and write_hup events?)

[dignifiedquire]

This will need a different approach, now that we are no longer using mio, and use smol as the underlying runtime.

[leo60228]

Yeah, this design doesn't make sense anymore, especially since you can just use smol::Async directly. It might make sense to re-export that somewhere, but that's about it.