[WIP] fix(benches): re-enable the recently disabled benches

[oddgrd]

Re-enable two of the benches disabled in #2929, using hyper::server::conn rather than the higher-level Server struct which is getting removed.

• benches/pipeline.rs
• benches/server.rs

I tested the updated benches by running them on this branch and also on the commit before #2929, and they have the same performance. I tried to implement graceful shutdown, but I'm not sure how to do it without Server. I see there is an issue (#2862) to add a helper for this to hyper-util, but if it can be implemented without it I am happy to keep trying.

PS. I don't have a firm grasp on async yet, so please let me know if I did something weird 😄

Partially closes#2930

[oddgrd]

I've managed to get all but two of the end_to_end benches to work, but a few of them have quite degraded performance compared to the previous implementation. I assume this is in part due to the fact that the previously used Client has a connection pool built in? Of course there is also something wrong with my implementation, as two of the benches just hang and do nothing. I've pushed a commit with my solution so far, with TODOs on the benches that have reduced performance and the two broken commented out.

I've also attached the logs from one of the broken benches (they both get stuck like this).

Logs from failing benches

TRACE mio::poll > registering event source with poller: token=Token(0), interests=READABLE | WRITABLE
TRACE mio::poll > registering event source with poller: token=Token(1), interests=READABLE | WRITABLE
TRACE hyper::client::conn > client handshake Http1
TRACE mio::poll > registering event source with poller: token=Token(2), interests=READABLE | WRITABLE
TRACE hyper::proto::h1::conn > Conn::read_head
TRACE hyper::proto::h1::conn > flushed({role=server}): State { reading: Init, writing: Init, keep_alive: Busy }
TRACE want > signal: Want
TRACE want > signal found waiting giver, notifying
TRACE hyper::proto::h1::conn > flushed({role=client}): State { reading: Init, writing: Init, keep_alive: Busy }
TRACE want > poll_want: taker wants!
TRACE tracing::span > encode_headers;
TRACE tracing::span::active > -> encode_headers;
TRACE hyper::proto::h1::role > Client::encode method=GET, body=None
TRACE tracing::span::active > <- encode_headers;
TRACE tracing::span > -- encode_headers;
DEBUG hyper::proto::h1::io > flushed 45 bytes
TRACE hyper::proto::h1::conn > flushed({role=client}): State { reading: Init, writing: KeepAlive, keep_alive: Busy }
TRACE hyper::proto::h1::conn > Conn::read_head
TRACE hyper::proto::h1::io > received 45 bytes
TRACE tracing::span > parse_headers;
TRACE tracing::span::active > -> parse_headers;
TRACE hyper::proto::h1::role > Request.parse bytes=45
TRACE hyper::proto::h1::role > Request.parse Complete(45)
TRACE tracing::span::active > <- parse_headers;
TRACE tracing::span > -- parse_headers;
DEBUG hyper::proto::h1::io > parsed 0 headers
DEBUG hyper::proto::h1::conn > incoming body is empty
TRACE tracing::span > encode_headers;
TRACE tracing::span::active > -> encode_headers;
TRACE hyper::proto::h1::role > Server::encode status=200, body=Some(Known(10485760)), req_method=Some(GET)
TRACE tracing::span::active > <- encode_headers;
TRACE tracing::span > -- encode_headers;
TRACE hyper::proto::h1::io > buffer.queue self.len=82 buf.len=10485760
DEBUG hyper::proto::h1::io > flushed 2588672 bytes
TRACE hyper::proto::h1::conn > Conn::read_head
TRACE hyper::proto::h1::io > received 8192 bytes
TRACE tracing::span > parse_headers;
TRACE tracing::span::active > -> parse_headers;
TRACE hyper::proto::h1::role > Response.parse bytes=8192
TRACE hyper::proto::h1::role > Response.parse Complete(82)
TRACE tracing::span::active > <- parse_headers;
TRACE tracing::span > -- parse_headers;
DEBUG hyper::proto::h1::io > parsed 2 headers
DEBUG hyper::proto::h1::conn > incoming body is content-length (10485760 bytes)
TRACE hyper::proto::h1::decode > decode; state=Length(10485760)
TRACE hyper::proto::h1::conn > flushed({role=client}): State { reading: Body(Length(10477650)), writing: KeepAlive, keep_alive: Busy }

[seanmonstar]

Yes, the end_to_end benches were using a connection pool. It'd be important to adjust the benchmarks to reuse the connection. I assumed we'd need to change that file to spawn N connections, and save the SendRequest handles and keep sending on them. But we can merge the first two first.

[oddgrd]

Hi, thanks for the tip on the e2e bench! I stashed the e2e commit for now and fixed up the others so they are ready to merge first, as you suggested. They perform as before. I'll have a look at the e2e again, but if anyone else wants to take a shot that is fine by me of course.