Channel refactor

[cramertj]

Refactors channels to use "waker" terminology rather than "parking"

Splits unbounded channels into a separate, simplified underlying type

Calls poll_ready in poll_flush impl for bounded channels in order to
provide backpressure on ::send.

Replacement for #984.

Fix #800
Fix rustasync/team#11

cc @danburkert, @seanmonstar, @carllerche

[cramertj]

This is necessary in order to prevent this test from flaking due to the race condition when writing into a closed channel.

[carllerche]

I assume that this is the race that you mentioned in gitter and you are punting on a fix? If so, could you add a TODO referencing the issue?

[carllerche]

Skimming over it looks like the unbounded variant also has the same issue?

[cramertj]

Yes, they both have this issue.

[carllerche]

Could you explain the reasoning behind splitting up the unbounded & bounded implementations?

[carllerche]

It's a pretty large diff and hard for me to reason about due to how terminology changes are mixed with code and behavior changes.

[cramertj]

@carllerche

Could you explain the reasoning behind splitting up the unbounded & bounded implementations?

Sure! I thought it was clearer and easier to read with them split up this way, and seemed like it might be more performant since there are fewer queues, atomics, and other checks needed to supply simple receiver wakeups than the comparatively complex sender wakeup semantics.

[cramertj]

Before this PR:

test bounded_100_tx        ... bench:     290,574 ns/iter (+/- 2,578)
test bounded_1_tx          ... bench:     278,866 ns/iter (+/- 11,897)
test unbounded_100_tx      ... bench:     150,284 ns/iter (+/- 7,048)
test unbounded_1_tx        ... bench:     134,440 ns/iter (+/- 3,558)
test unbounded_uncontended ... bench:      66,844 ns/iter (+/- 5,163)

After this PR:

test bounded_100_tx        ... bench:     281,201 ns/iter (+/- 1,249)
test bounded_1_tx          ... bench:     278,706 ns/iter (+/- 36,851)
test unbounded_100_tx      ... bench:     105,909 ns/iter (+/- 22,730)
test unbounded_1_tx        ... bench:      98,740 ns/iter (+/- 2,073)
test unbounded_uncontended ... bench:      59,876 ns/iter (+/- 5,405)

[carllerche]

The main argument against duplicating code is that it introduces more locations for bugs. For example, the close race bug is now duplicated in two code locations.

I would suggest that, if performance is the driver, at the very least benchmarks illustrating the performance difference should be present and probably at least a basic tuning pass of the bounded mpsc should be done (given that none have been done to date).

[carllerche]

At the end of the day, you are the ones who have to maintain it, so feel free to do whatever.

[aturon]

LGTM modulo nit.

[aturon]

LGTM modulo nit.

[aturon]

Seems worth factoring this out into its own method, since the code is repeated verbatim below.

[cramertj]

Done