http: Make it possible to pass in a pre-existing connection

[laanwj]

This patch adds functionality to pass a pre-existing connection as a bufferevent to evhttp_connection_base_bufferevent_new. As far as I know, this was not possible before.

When the bufferevent has an existing fd, the evcon starts in state EVCON_IDLE so that requests can be made immediately.

I am using this in Bitcoin Core for doing HTTP requests over a UNIX socket (bitcoin/bitcoin#9919).

It could also be useful in sandboxed environments such as CloudABI that cannot directly make connections but can pass around fds.

[coveralls]

Coverage decreased (-0.03%) to 80.48% when pulling 403b793 on laanwj:2017_03_evhttp_existing_fd into cc0e04d on libevent:master.

[coveralls]

Coverage decreased (-0.03%) to 80.48% when pulling 403b793 on laanwj:2017_03_evhttp_existing_fd into cc0e04d on libevent:master.

[azat]

@laanwj sorry for the huge delay, patch LGTM, but with it evhttp_connection_set_retries will be no-op, even though this will work for you it will make things tricky for others.

That's why I hadn't merged it already. But I had an idea of how we can make it works for everybody. What do you think?

[ghazel]

You'll also want to set evcon->flags |= EVHTTP_CON_OUTGOING; so that the connection can read more than one request.

[laanwj]

but with it evhttp_connection_set_retries will be no-op, even though this will work for you it will make things tricky for others.

I figured automatic retries aren't really a possibility if the application provides the connection. This uses libevent only for the protocol handling. If the application needs retries, it has to implement them itself.

evcon->flags |= EVHTTP_CON_OUTGOING; so that the connection can read more than one request.

Good idea.

[laanwj]

I added EVHTTP_CON_OUTGOING. Making the retry work with a callback (at either http or bufferevent level) as suggested by @azat sounds like a good idea to me, but that exceeds my expertise with the libevent code.

[coveralls]

Coverage decreased (-0.1%) to 80.418% when pulling 08e0114 on laanwj:2017_03_evhttp_existing_fd into cc0e04d on libevent:master.

[coveralls]

Coverage decreased (-0.1%) to 80.418% when pulling 08e0114 on laanwj:2017_03_evhttp_existing_fd into cc0e04d on libevent:master.

[laanwj]

I'm not entirely sure what is blocking this, but will be looking at a different solution for this bypassing libevent.

[azat]

I'm not entirely sure what is blocking this, but will be looking at a different solution for this bypassing libevent.

@laanwj sorry for my ignorance, last time I looked at this I did not like that this makes no-op some of functionality (evhttp_connection_set_retries), and also that fact that http layer is not in a good shape.

I'm OK to merge this change if:

• it will be done under some connection flag (evhttp_connection_set_flags)
• it will have a test (in test/regress_http.c)

[laanwj]

Thanks for the elaboration!

I can't guarantee to get around to that but will keep it in mind.

(it's not that I really insist on this solution, I just need something lightweight to connect to a HTTP server on a UNIX oscket 🙂 )

[azat]

it's not that I really insist on this solution

It may be useful to inject something into http layer indeed (and not only for unix sockets)