WebSockets over HTTP/2

Question

WebSockets over HTTP/2

nhooyr opened this issue 6 years ago · 21 comments

nhooyr commented 6 years ago

Blocked on the drafts and Go support.

nhooyr commented 5 years ago

Answer 1 · 2018-09-06T23:22:17.000Z

Definitely worth reading https://tools.ietf.org/html/rfc8441

Answer 2 · 2019-03-29T23:19:33.000Z

This does mean we cannot expose net.Conn directly as part of the API interface like gorilla/ws because there may not be a net.Conn to expose.

Answer 3 · 2019-04-05T00:47:42.000Z

So unfortunately, there is no clean way for a client to detect whether a HTTP server supports WebSockets over HTTP/2 without actually performing a handshake. I've emailed the working group regarding this.

Answer 4 · 2019-04-05T01:35:21.000Z

Hopefully shows up on https://lists.w3.org/Archives/Public/ietf-http-wg/2019AprJun/

Answer 5 · 2019-04-07T18:12:03.000Z

https://lists.w3.org/Archives/Public/ietf-http-wg/2019AprJun/0010.html

Answer 6 · 2019-05-30T04:00:41.000Z

So it looks like my suggestion won't be implemented. The only way to implement this then is to take a HTTP2 flag on the dial options or keep a global cache.

For the server, I'm sure we can do this transparently but will require some collaboration with net/http.

Answer 7 · 2019-07-15T13:49:32.000Z

Could probably implement this reasonably quickly if we forked x/net/http2.

Need to be able to advertise support for HTTP/2 WebSockets
Use http.Hijacker on HTTP/2 Connect requests with a mocked out net.Conn for the stream.
Have *http.Client return a rw body for successful HTTP/2 Connect requests
Be able to set the :protocol pseduo header.

Answer 8 · 2019-07-16T05:04:18.000Z

Hi, I have been building a similar websocket package, and would be willing to help if I can. Some things I can share:

I also looked into the problem of identifying whether an endpoint supports HTTP/2 WebSockets. The best I came up with was similar to that suggested elsewhere - use an HTTP/2 connection first, then downgrade to HTTP/1 if necessary. However, SETTINGS_ENABLE_CONNECT_PROTOCOL, and similar host-wide mechanisms are not actually sufficient to identify whether an endpoint supports RFC8441 or not - I think it can be theoretically confused by a non-RFC8441 service behind an RFC8441-compliant reverse proxy. In reality, once this gets pushed out into the world a lot of services are likely to have partial support during an upgrade period - with some endpoints using the upgraded websocket packages and others not yet supporting it. The most foolproof way I came up with is to send a request and check for a 405 Method Not Allowed status code. Any RFC6455-compliant server will return a 405 Method Not Allowed if it recieves an unrecognized method, such as CONNECT as used here. When I checked, all implementations I found did this correctly.
The rw body for successful HTTP/2 Connect requests has already been implemented. I have tested it out, and it works.
I did open an issue which you linked about being able to set the :protocol header. I have been looking into that, and so far I have not come up with any reasonable way to expose this behavior beyond adding an override hack like the one currently used to force HTTP/1 for websockets.

I see you on some of the related threads, so some of this is just to provide context to anyone else reading this issue thread.

Answer 9 · 2019-07-16T15:53:04.000Z

The rw body for successful HTTP/2 Connect requests has already been implemented. I have tested it out, and it works.

Nice.

However, SETTINGS_ENABLE_CONNECT_PROTOCOL, and similar host-wide mechanisms are not actually sufficient to identify whether an endpoint supports RFC8441 or not - I think it can be theoretically confused by a non-RFC8441 service behind an RFC8441-compliant reverse proxy.

I think it should be fine. A RFC 8441-complaint reverse proxy would translate HTTP/2 over web sockets into HTTP/1.1 for the origin server. e.g. Nginx translates all HTTP/2 requests into HTTP/1.1 for origin servers. Otherwise it wouldn't be RFC 8441-complaint if it just proxied the request directly.

I did open an issue which you linked about being able to set the :protocol header. I have been looking into that, and so far I have not come up with any reasonable way to expose this behavior beyond adding an override hack like the one currently used to force HTTP/1 for websockets.

Yea that is unfortunate. Maybe a new field on *http.Request to set custom pseudo headers forcing HTTP/2? Or translating the HTTP/1.1 Protocol header to :protocol for HTTP/2? I think that would be reasonable.

Hi, I have been building a similar websocket package, and would be willing to help if I can.

Awesome, would love your help :)

So then the remaining items here are:

HTTP/2 CONNECT support for http.Handler
Setting custom HTTP/2 pseudo headers

Answer 10 · 2019-11-05T15:11:51.000Z

See my proposal at golang/go#32763 (comment) to make net/http transparently handle http/2 upgrades server side and client side.

Would be much nicer than implementing it here as all Go WebSocket server's would work over HTTP/2 and we would get client side support for free as we use net/http's Client directly.

Answer 11 · 2021-05-13T06:39:02.000Z

Hi... is there a workaround that can be used until go supports something in that direction?

Answer 12 · 2021-08-06T09:21:45.000Z

@hons82 Unfortunately apart from forking the go net/http library nope. I don't think it'd be very difficult to implement but there just hasn't been a response from the Go team on the linked issue.

Answer 13 · 2022-04-06T02:18:46.000Z

@nhooyr Any further updates or progress on this?

Answer 14 · 2023-08-16T03:47:45.000Z

I'm full time on websocket the next few months. There hasn't been any movement on the Go issue but I think I'll just fork net/http and implement my proposal linked above there. Then people can use my fork if they want WebSockets over http/2 with my library.

As well I'll open a CL on the standard library with my proposal so hopefully my fork won't be necessary. These things don't tend to get moving on issue trackers unless someone shows the initiative by implementing and demonstrating the demand.

Answer 15 · 2023-08-16T21:02:46.000Z

I believe the link was meant to be #402

Answer 16 · 2023-10-21T19:57:37.000Z

Have you seen golang/go#49918?

Answer 17 · 2023-10-21T19:59:11.000Z

Also golang/go#53209 and golang/go#53208.

Answer 18 · 2023-10-22T01:13:16.000Z

Nope, I missed those thanks for the links!

Answer 19 · 2023-10-23T11:37:07.000Z

That would be awesome

Answer 20 · 2023-11-13T21:51:22.000Z

While the Go team decides how (or whether?) to handle this, is there a workaround ?

My execution environment:

server: go1.21.4 darwin/arm64
client: Safari 17.1 running a wasm client compiled with go1.21.4 (GOOS=js GOARCH=wasm)

When the wasm client tries to connect, the server gets this error from websocket.Accept(w,req,nil):

failed to accept WebSocket connection: WebSocket protocol violation: Connection header "keep-alive" does not contain Upgrade

When package websocket is compiled for wasm, the field websocket.DialOptions.HTTPHeader is not available, so I cannot add the required header. This appears to be a limitation of the Javascript WebSocket API, or of how the API is used.

I presume this is all related to issue 373.

But it is possible that I am on the wrong track and have no idea what I am doing wrong.

I would be grateful for any guidance.