Make client able to use non-Send executor

[Ruben2424]

Hi,
I tried to solve #3017. But I got stuck.

I copied and modified existing examples to get one which fails to compile because of the missing send bounds.

The issue suggests to create a extension trait Executor<InternalFutureType<IO, B>>.
But what exactly is the InternalFutureType<IO, B>? I had the idea of a enum InternalFutureType<IO, B> with variants for every Future that gets spawned and implementing Future for it. But most the futures that get spawned are Anonymous or have Anonymous Parts.

hyper/src/proto/h2/client.rs

Line 158 in 47f614f

exec.execute(conn_task(conn, conn_drop_rx, cancel_tx));

hyper/src/proto/h2/client.rs

Line 259 in 47f614f

self.executor.execute(pipe);

hyper/src/proto/h2/client.rs

Line 316 in 47f614f

self.executor.execute(f.cb.send_when(fut));

Maybe the solution is obvious and I just don't see it? Or am I on the wrong path?
Any hint would be helpful.

[seanmonstar]

So, doing this is a bit unfun. Those anonymous futures would need to become structs, with manual impl Future for them. They would need to be pseudo-public. The similar one to see is proto::h2::server::H2Stream.

Then, we can add in to rt::bounds a trait that is implemented for T: Executor<ConnTask>, T: Executor<Pipe>, T: SendWhen.

[Ruben2424]

I changed the anonymous futures to structs and introduced a trait. And it seams to work mostly. Some cleanup is left.
@seanmonstar is it ok to for Body to require 'static + Unpin ?

[seanmonstar]

The line just above, B::Error: Into<..> should be enough. Is it not?

[Ruben2424]

Is it better now?

[seanmonstar]

Is it required to add something here at all? Because currently, this makes it stricter in that not only is it Into, but it must also be Error. That means Box<dyn Error> no longer qualifies (since Box<dyn Error>: Error is not true for confusing reasons).

[Ruben2424]

I think this is ready for a second review now.

I am not 100% sure if I really understand everything I did, so please review carefully.

I got some questions:

1. http1 requires still Send for the io types. Is this something that needs to be addressed? And if so, in this PR?
2. One test didn't compile because the new bounds don't allow the error type Box<dyn std::error::Error + Send + Sync> is this something that needs to be fixed? And if how can I fix this?
3. I just tested the example and started the integration test. Is there something that needs to be tested manually? Or something that I need to look at when testing? Or is a testcase missing?

I hope I haven't forgotten something important here.

[seanmonstar]

Very impressive! I think there's just some dangling clean up that I found.

[seanmonstar]

Very well done!

[seanmonstar]

Oh, looks like there's merge conflicts. Would you be up to resolving those? Then we can merge. RC4 is nearly here!