Fix tunnel proxy: HTTP requests only

[yeraydiazdiaz]

Fixes#54 split from #55 excluding support for tunneling HTTPS requests. Changes:

• First establish an HTTP connection to our proxy since CONNECT is an HTTP method handled by h11.
• Once successful discard the connection but prevent closing the socket.
• Allow passing a socket to the AsyncHTTPConnection and create a new h11 connection to the target.

[florimondmanca]

First pass… :-)

[florimondmanca]

@yeraydiazdiaz I'm actually not sure I understand the original bug in #54. My hypothesis is that closing the proxy pool calls .aclose() on the connection which we didn't add using .add_to_pool(), so the semaphore was not acquired. But I'm not quite sure. What's your diagnosis?

[florimondmanca]

@yeraydiazdiaz Okay, after playing with this myself, I ended up to the same conclusion: because h11.PAUSED moves the connection to a permanently "can't be used anymore" state (and this is now also enforced by our own state machine, which we didn't have before in HTTPX), we need to setup a new connection once CONNECT has succeeded, but without closing the socket. The socket parameter seems to be the way to go.

[yeraydiazdiaz]

@yeraydiazdiaz I'm actually not sure I understand the original bug in #54. My hypothesis is that closing the proxy pool calls .aclose() on the connection which we didn't add using .add_to_pool(), so the semaphore was not acquired. But I'm not quite sure. What's your diagnosis?

So there were several problems before. The first one was an assertion in the AsyncHTTPConnection preventing sending requests to URLs different from the originally passed origin, the connection was created with the target origin but the CONNECT requests was attempted to the proxy origin.

Working around that I also found that consuming the response body closed the connection and the socket. Eventually arriving at we discussed above.

[florimondmanca]

Eventually arriving at we discussed above.

Yep. Followed the same series of thoughts. :-)

[florimondmanca]

@yeraydiazdiaz Are you able to proxy an HTTP request via Squid/mitmproxy with the current state of this PR?

Right now on Squid I get a 503 on the CONNECT request, not sure if it's because we're missing something here or if I have something else to setup on Squid…

[(b'server', b'squid/3.5.27'), (b'mime-version', b'1.0'), (b'date', b'Sun, 12 Apr 2020 17:03:08 GMT'), (b'content-type', b'text/html;charset=utf-8'), (b'content-length', b'3481'), (b'x-squid-error', b'ERR_CONNECT_FAIL 99'), (b'vary', b'Accept-Language'), (b'content-language', b'en')]Traceback (most recent call last): File "debug/client.py", line 16, in <module> asyncio.run(main()) File "/Users/florimond/.pyenv/versions/3.8.2/lib/python3.8/asyncio/runners.py", line 43, in run return loop.run_until_complete(main) File "/Users/florimond/.pyenv/versions/3.8.2/lib/python3.8/asyncio/base_events.py", line 616, in run_until_complete return future.result() File "debug/client.py", line 10, in main response = await http.request(b"GET", url) File "/Users/florimond/Developer/python-projects/httpcore/httpcore/_async/http_proxy.py", line 83, in request return await self._tunnel_request( File "/Users/florimond/Developer/python-projects/httpcore/httpcore/_async/http_proxy.py", line 175, in _tunnel_request raise ProxyError(msg)httpcore._exceptions.ProxyError: 503 Service Unavailable

Simplified client.py I'm using:

importasyncioimporthttpcoreproxy_origin= (b"http", b"localhost", 3128)
url= (b"http", b"localhost", 8000, b"/")
asyncdefmain() ->None:
asyncwithhttpcore.AsyncHTTPProxy(proxy_origin, proxy_mode="TUNNEL_ONLY") ashttp:
response=awaithttp.request(b"GET", url)
print(response)
status_code=response[1]
assertstatus_code==200, status_codeasyncio.run(main())

Uvicorn server on localhost:8000:

fromstarlette.responsesimportPlainTextResponseapp=PlainTextResponse("Hello, world!")

[yeraydiazdiaz]

Your example using Starlette does not work using Squid, but it does when using mitmproxy without any options.

import asyncio
import httpcore
proxy_origin = (b"http", b"localhost", 8080)
url = (b"http", b"localhost", 8000, b"/")
async def main() -> None:
async with httpcore.AsyncHTTPProxy(proxy_origin, proxy_mode="TUNNEL_ONLY") as http:
response = await http.request(b"GET", url, headers=[(b"host", b"localhost")])
print(response)
status_code = response[1]
assert status_code == 200, status_code
asyncio.run(main())

It also works fine when fetching example.org.

Squid doesn't really log a lot with our current configuration so it's hard to say what's wrong there.

[florimondmanca]

Second pass, bunch of nits though, I think we're on the right track :-) I was able to make it work with mitmproxy too!

About proxy Host headers: actually, examples in RFC 7321 suggest that CONNECT should also contain a Host header, so I think this is a legit constraint to put on users, and definitely not an h11 edge case we should hide (because it's not an edge case).

[florimondmanca]

@yeraydiazdiaz Sorry if this is taking longer to review than you expected. 😅

This is great! stuff, and I think we're almost there.

Yes, currently proxying is broken in HTTPX, and it's a regression, but that's alright. We haven't released anything yet, and previous versions are still mostly working, so no pressure on that side.

This is why I think we should take the time to dissolve any fuzzy areas, make sure we understand everything that's going on (I'm discovering more and more of the new subtle aspects of HTTPCore as we're going through this), and that we enforce the "raw request in, raw response out"/tight-scope philosophy of HTTPCore, i.e. keeping the boundary between it and client libraries as clearly defined as possible.

Hope you're okay with this — I know long reviews can be a bit draining, so feel free to push back when I'm being too nitpicky as I know I can sometimes be. :-) Thanks again for tackling this 💖

[yeraydiazdiaz]

Thank you for reviewing! 🌟

It's absolutely fine, I agree we should take time to get things right, we've only just started to make actual use of HTTPCore so I expected some "where should this live" types of discussions, we're bound to have more of these, so the sooner we can define these lines the better 🙂

[yeraydiazdiaz]

I removed the _ONLY suffix here which does not match the values in HTTPX but more semantic sense.

[lovelydinosaur]

We might want to leave these the same, at least within the scope of this pull request?
It'd be helpful if we're not introducing breaking API changes from httpx 0.12 -> 0.13 wherever possible. (excepting for example, the UDS support)

[yeraydiazdiaz]

So this change stems from a longer conversation with @florimondmanca. TL;DR is where do we handle the logic behind the DEFAULT mode?

I opted for removing it completely as suggested by Florimond and have httpx call into httpcore with a mode explicitly.

I think this way we reduces complexity on httpcore but I don't mind keeping the three modes and the logic behind here if you prefer it. What do you think?

[yeraydiazdiaz]

@florimondmanca could you have another look at this when you have a second? Thanks!

[florimondmanca]

Will try to do soon :)

[yeraydiazdiaz]

I reintroduced the DEFAULT mode and *_ONLY proxy modes. The default mode is now an alias for TUNNEL_ONLY for all requests, as opposed to only on HTTPS ones. But frankly I'm not really sure how to proceed here and I'm a bit confused by the feedback in this PR.

At the moment httpx will fail to make a request via a proxy due to missing/conflicting headers.

In #59 and here we discussed httpcore should have a small as possible API and httpx should be in charge of setting the parameters are required. We discussed in this PR that it doesn't make much sense to have a DEFAULT mode here or the logic to add the proxy headers as required. Of course that means changing httpx to introduce the necessary changes there but I think it's worth it because we're moving to a smaller API surface in httpcore.

What are the next steps for this? @florimondmanca@tomchristie

[lovelydinosaur]

Okay, so I think I understand the sticking point better here now. In order to keep the conversation more tightly scoped, let me start by making one inline suggestion, and see if we can move forward from there...

[lovelydinosaur]

Looks great! Thanks so much for your time on it.

🌟👍

[yeraydiazdiaz]

🎉 Thanks @tomchristie and @florimondmanca! 🚀