make copy continue filling the buffer when writer stalls

[farnz]

Fixes: #3694

Motivation

When a reader is giving small chunks of data (e.g. reading over a congested TCP link), and the writer stalls because it's also overloaded, we can end up issuing smaller writes than a buffer because we do at most one read for every write.

In the overload case, demonstrated by the copy_chunk_to_slow_hdd benchmark, this results in us being slower than we could be if we aimed to fill the buffer each time we did a read.

Solution

Refactor so that we can read every time we get a Poll::Pending from the writer. This has no significant impact on times if either the reader is able to fill the copy buffer each time, or the writer never stalls, but on my Xeon E3-1245v2, I see copy_chunk_to_slow_hdd speed up:

Before:
test copy_chunk_to_slow_hdd ... bench: 161,070,661 ns/iter (+/- 4,341,313)

After:
test copy_chunk_to_slow_hdd ... bench: 141,173,126 ns/iter (+/- 348,824)

[Darksonn]

A current-thread runtime probably makes more sense for the benchmark. You don't do any spawning.

[farnz]

I copied that from fs.rs, which also doesn't do any spawning - it uses block_in_place, though, which might be the reason.

[Darksonn]

The block_in_place method involves spawning.

[Darksonn]

In general, I think this looks pretty good.

[Darksonn]

LGTM!