fix: peek reregister after would block

[discord9]

CC #1755 tokio-rs/tokio#3789

as title, use a fix that on windows peek should reregister after either success or would block, so that it's behavior would be the same as on linux(see peek_readinesspeek_ok and peek_would_block).

This should fix windows peek problem, @Darksonn might want to take a look.

The problem it seems is that in src\sys\windows\selector.rs SockState::feed_event mio have to simulate Edge-triggered behavior and clean readable interest after send a readable interest, but since peek is also a kind of special read, it also cancelled interest to readable event, hence cause the blocking peek bug on windows because interest is lost so blocking is forever

edited: found out that under linux with cfg(mio_unsupported_force_poll_poll) peek would also have similiar problem so just write do_io on all platform instead

Alternatives

only do do_io on #[cfg(any(windows, mio_unsupported_force_poll_poll))] which also seem to pass all tests on all platform/cfg combination

[Darksonn]

Reregister is an expensive operation and I don't think that's the solution we want.

[discord9]

Reregister is an expensive operation and I don't think that's the solution we want.

How about handle peek event specially in feed_event on windows somehow(maybe add a is_peeking flag somewhere)? Maybe even a special Interest::Peek for it only on windows?

Also the CI test failed strangely, I don't understand why, the TestPoll failed with not found expected event)
but the Test (ubuntu-latest) successed, I don't understand why, also I don't think my patch affect behavior on linux platform so even more confusedOh different rustcfg flag I see

[discord9]

the falling CI seems related to #1883 and not caused by this patch

[Thomasdezeeuw]

I think this comment is outdated and can be removed.

[Thomasdezeeuw]

Thanks @discord9

[Sytten]

For what it is worth, this is the exact fix that I proposed 2 years ago.

@Thomasdezeeuw did something fundamentally changed in core for your opinion to change on that?

See #1756

[Darksonn]

I think we just misunderstood the problem space last time we looked at it. I don't think there is anything more to it. Sorry about that.