rt: fix *_closed false positives

[satakuma]

Motivation

Readiness futures inconsistently return current readiness of a I/O resource, if it is immediately available, or all readiness relevant for the given Interest, if a future needs to wait. In particular, it always returns read_closed for Interest::READABLE and write_closed for Interest::WRITABLE, what often is not true. Tokio should not tolerate false positives for *_closed events because they are considered final states and are not cleared internally.

In case of a io_resource.ready(Interest::READABLE | Interest::WRITABLE) call, this behavior may also lead to false positives of other events.

Solution

If a Readiness future is woken up by a new event, store the relevant readiness and return it later.
It pumps the Waiter and enclosing structs a bit (8 bytes on x86_64). If it's an issue, readiness can be packed into something smaller.

Closes: #5098

[Darksonn]

Looks reasonable to me, but I would also like a review from Carl.

[carllerche]

Thanks for doing this. You are correct the current implementation is buggy. Most of this looks good. I left a comment inline.

[carllerche]

If a readiness bit is not included in curr_ready then something else already cleared it. Including w.ready here would result in additional false positives & unnecessary I/O ops.

Based on discord, it sounds like you want to avoid empty readiness events, but that seems better to me than false positives. The caller would loop and wait again.

[satakuma]

@carllerche I applied your suggestion to return the current readiness.

I also updated the docs of every ready method to mention false-positives and empty events. My motivation is that empty Ready events were not really possible before this change. As far as I can tell from the docs, there is no explicit guarantee about no empty return values, but I can still imagine a program logic which would assume that at least one of the requested state is set. For example, code which would deduce that if !io.ready(Interest::READABLE).await?.is_readable(), then the read must be closed. For that reason I think an explicit mention in the docs will be helpful.

[carllerche]

LGTM thanks 👍