-
Notifications
You must be signed in to change notification settings - Fork 17.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
x/net/http2: support h2c for http2 #14141
Comments
gRPC reimplemented http2 because they had to. We started with the same http2.Framer code and they wrote their gRPC-specific http2 client & server in parallel with me writing net/http's http2 client and server. I'm currently working on unifying them (see grpc/grpc-go#75). I might accept h2c as a separate package under x/net/http2, but it won't be part of the Go standard library or on by default. Some of the work was already done as part of golang/net@d62542d |
@bradfitz thanks.
My thought is that eventually, another framework like gRPC should not have to reimplement http2 because the standard (or at least To be clear, for my use-case I'm mostly interested in inter-service communication where I assume I control both client and server, so I really want HTTP/2 over TCP with "prior knowledge" (section 3.4) rather than HTTP/1.1 Upgrade. But I think it makes sense that there is at least a somewhat official library (even if it stays in x) that supports the whole spec. I looked at the old PR for upgrade support which seems a little hairy in the way it integrates with http server - I can see it's not a simple request to support.
Fair enough. Can I ask how the two interact? I'm using 1.6rc which uses http2 automatically when serving TLS via net/http (standard library) but the http2 implementation appears to still be this one in I will take a look through the commit you linked and get a better idea what the options are in terms of optional h2c upgrade and/or prior knowledge support with Thanks again for your help. |
Hmm I see your linked commit is against |
We vendor a copy of x/net/http2 into net/http so http2 can be automatic for the standard library. What I mean is that I will not enable h2c automatically for the standard library, but h2c work can happen in a subpackage of x/net, like x/net/http2/h2c. |
Thanks Brad, that's clear now.
|
CL https://golang.org/cl/19176 mentions this issue. |
Fixes golang/go#12737 Updates golang/go#14141 Change-Id: I552b603b63a7c87d7fcdb4eb09f96ab9fd0ec0aa Reviewed-on: https://go-review.googlesource.com/19176 Reviewed-by: Andrew Gerrand <[email protected]> Run-TryBot: Brad Fitzpatrick <[email protected]> TryBot-Result: Gobot Gobot <[email protected]>
CL https://golang.org/cl/21327 mentions this issue. |
The http2 spec defines a magic string which initates an http2 session: "PRI * HTTP/2.0\r\n\r\nSM\r\n\r\n" It was intentionally chosen to kinda look like an HTTP request, but just different enough to break things not ready for it. This change makes Go ready for it. Notably: Go now accepts the request header (the prefix "PRI * HTTP/2.0\r\n\r\n") as a valid request, even though it doesn't have a Host header. But we now mark it as "Connection: close" and teach the Server to never read a second request from the connection once that's seen. If the http.Handler wants to deal with the upgrade, it has to hijack the request, read out the "body", compare it against "SM\r\n\r\n", and then speak http2. One of the new tests demonstrates that hijacking. Fixes #14451 Updates #14141 (h2c) Change-Id: Ib46142f31c55be7d00c56fa2624ec8a232e00c43 Reviewed-on: https://go-review.googlesource.com/21327 Reviewed-by: Andrew Gerrand <[email protected]> Run-TryBot: Brad Fitzpatrick <[email protected]> TryBot-Result: Gobot Gobot <[email protected]>
Sorry to bother, but i found it's quite easy to make a h2c client with a little modification to package main
import (
"crypto/tls"
"fmt"
"io/ioutil"
"log"
"net"
"net/http"
"golang.org/x/net/http2"
)
func main() {
client := http.Client{
// Skip TLS dial
Transport: &http2.Transport{
DialTLS: func(netw, addr string, cfg *tls.Config) (net.Conn, error) {
return net.Dial(netw, addr)
},
},
}
resp, err := client.Get("http://localhost:8080/test")
if err != nil {
log.Fatal(err)
return
}
fmt.Printf("resp: %#v\n", resp)
body, err := ioutil.ReadAll(resp.Body)
if err != nil {
log.Fatal(err)
return
}
fmt.Println(string(body))
} then remove the restriction from
I agree that nearly all browsers will not support h2c, but there are some h2c server serving traffic as another RPC service (not gRPC), so can https restriction be loosen a little to support h2c client? |
I'd be fine with an |
@bradfitz I'd be fine wth the knob too. So, will it be shipped? |
If somebody sends me a change. |
@bradfitz It seems Here is a diff. Since http port is not 443, i add some if-else to use 80 diff --git a/http2/configure_transport.go b/http2/configure_transport.go
index daa17f5..dde578a 100644
--- a/http2/configure_transport.go
+++ b/http2/configure_transport.go
@@ -32,7 +32,7 @@ func configureTransport(t1 *http.Transport) (*Transport, error) {
t1.TLSClientConfig.NextProtos = append(t1.TLSClientConfig.NextProtos, "http/1.1")
}
upgradeFn := func(authority string, c *tls.Conn) http.RoundTripper {
- addr := authorityAddr(authority)
+ addr := authorityAddr(authority, "https")
if used, err := connPool.addConnIfNeeded(addr, t2, c); err != nil {
go c.Close()
return erringRoundTripper{err}
diff --git a/http2/transport.go b/http2/transport.go
index a7ada8f..a9f24dd 100644
--- a/http2/transport.go
+++ b/http2/transport.go
@@ -75,6 +75,9 @@ type Transport struct {
// explicitly requested gzip it is not automatically
// uncompressed.
DisableCompression bool
+
+ // Allow plain text HTTP/2
+ AllowHTTPScheme bool
// MaxHeaderListSize is the http2 SETTINGS_MAX_HEADER_LIST_SIZE to
// send in the initial settings frame. It is how many bytes
@@ -268,20 +271,27 @@ func (t *Transport) RoundTrip(req *http.Request) (*http.Response, error) {
// authorityAddr returns a given authority (a host/IP, or host:port / ip:port)
// and returns a host:port. The port 443 is added if needed.
-func authorityAddr(authority string) (addr string) {
+func authorityAddr(authority string, scheme string) (addr string) {
if _, _, err := net.SplitHostPort(authority); err == nil {
return authority
}
- return net.JoinHostPort(authority, "443")
+ var defaultPort string
+ if scheme == "http" {
+ defaultPort = "80"
+ } else {
+ defaultPort = "443"
+ }
+ return net.JoinHostPort(authority, defaultPort)
}
// RoundTripOpt is like RoundTrip, but takes options.
func (t *Transport) RoundTripOpt(req *http.Request, opt RoundTripOpt) (*http.Response, error) {
- if req.URL.Scheme != "https" {
+ if !(req.URL.Scheme == "https" || (req.URL.Scheme == "http" && t.AllowHTTPScheme)) {
return nil, errors.New("http2: unsupported scheme")
}
- addr := authorityAddr(req.URL.Host)
+
+ addr := authorityAddr(req.URL.Host, req.URL.Scheme)
for {
cc, err := t.connPool().GetClientConn(req, addr)
if err != nil { |
@ayanamist, we don't accept patches on Github or in comments. See https://golang.org/doc/contribute.html#Code_review That is a wall of text, but there's only like 3 real steps there. |
CL https://golang.org/cl/23181 mentions this issue. |
Since CL https://golang.org/cl/23181 was merged, what is the correct way to configure the server side for this? Given this server: package main
import (
"fmt"
"io"
"log"
"net/http"
"golang.org/x/net/http2"
)
func handler(w http.ResponseWriter, r *http.Request) {
fmt.Println(r)
io.WriteString(w, "hello\n")
}
func main() {
s := &http.Server{
Addr: ":8543",
Handler: http.HandlerFunc(handler),
}
http2.ConfigureServer(s, &http2.Server{})
log.Fatal(s.ListenAndServe())
} And this client: package main
import (
"crypto/tls"
"fmt"
"log"
"net"
"net/http"
"golang.org/x/net/http2"
)
func main() {
client := http.Client{
// Skip TLS dial
Transport: &http2.Transport{
AllowHTTP: true,
DialTLS: func(network, addr string, cfg *tls.Config) (net.Conn, error) {
return net.Dial(network, addr)
},
},
}
resp, err := client.Get("http://localhost:8543")
if err != nil {
log.Fatal(fmt.Errorf("error making request: %v", err))
}
fmt.Println(resp.StatusCode)
fmt.Println(resp.Proto)
} When the client makes a request, the output is: $ go run client/main.go
2016/08/24 22:31:07 error making request: Get http://localhost:8543: unexpected EOF
exit status 1 And the server output from the handler is: $ go run server/main.go
&{PRI * HTTP/2.0 2 0 map[] 0x3bd4e0 -1 [] true map[] map[] <nil> map[] [::1]:63792 * <nil> <nil> <nil> 0xc4200744c0} If I switch the server to use http2.Server.ServeConn directly, like in the tests in the above CL, it works as expected. |
I believe this is due to how http2 is implemented in the standard library: https://github.com/golang/go/blob/release-branch.go1.7/src/net/http/server.go#L1500:L1520 HTTP2 dispatch is implemented in It would take some additional non-ALPN logic to make h2c work the way you describe, and to my understanding @bradfitz has indicated that such additional logic will not be accepted. |
I've already accepted a ton of logic for h2c. My policy has been and remains: it should be possible for people to do h2c via a separate package, but it won't be supported by default. All the pieces should be there now for you to glue it together yourself. For server-side support you'll need to hijack in the Somebody should make a package to make that easier. I have other priorities, but can review CLs. |
Awesome, thanks @bradfitz. Would the Go team accept a CL to add h2c support, if it was opt-in (e.g. an I'd imagine a lot of people writing backend services are pretty excited about the performance advantages of http/2, but would rather terminate TLS at a load balancer. Thanks for all your hard work on http/2 thus far! |
I'd prefer it to be in its own package, hidden away and very much opt-in. |
|
@ejholmes it's been a few months, any update on getting h2c to work with http.Server? I think it'd be very useful to have plain http2 as an option for testing and backend services. Thanks |
To save others some time, @hkwi wrote an implementation here: https://github.com/hkwi/h2c |
Fixes golang/go#12737 Updates golang/go#14141 Change-Id: I552b603b63a7c87d7fcdb4eb09f96ab9fd0ec0aa Reviewed-on: https://go-review.googlesource.com/19176 Reviewed-by: Andrew Gerrand <[email protected]> Run-TryBot: Brad Fitzpatrick <[email protected]> TryBot-Result: Gobot Gobot <[email protected]>
I'm pretty far in getting this done. I expect I will be able submit a PR this weekend. |
Change https://golang.org/cl/112997 mentions this issue: |
Change https://golang.org/cl/112999 mentions this issue: |
I too wanted h2c for traffic between load balancers that were doing SSL Termination and my golang servers. I adapted the h2c implementation in Traefik into a standalone http.Handler: https://github.com/veqryn/h2c package main
import (
"fmt"
"net/http"
"github.com/veqryn/h2c"
"golang.org/x/net/http2"
)
func main() {
// Router/Mux (can use any http.Handler)
router := http.NewServeMux()
// Handlers...
router.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
fmt.Fprint(w, "Hello World")
})
// Wrap the Router
h2cWrapper := &h2c.HandlerH2C{
Handler: router,
H2Server: &http2.Server{},
}
// Server
srv := http.Server{
Addr: ":8080",
Handler: h2cWrapper,
}
srv.ListenAndServe()
} |
I'm happy to see that there is now an |
I would like to say THANK YOU @misterwilliam (and @bradfitz for reviewing), fantastic job. |
Implements h2c by leveraging the existing http2.Server by implementing the 2 ways to start an h2c connection as described in RFC 7540, which are: (1) create a connection starting with HTTP/1 and then upgrading to h2c [Section 3.2] and (2) starting a connection directly speaking h2c (aka starting with prior knowledge) [Section 3.4]. For both of the above connection methods the implementation strategy is to hijack a HTTP/1 connection, perform the h2c connection on the hijacked net.Conn, and create a suitably configured net.Conn to pass into http2.Server.ServeConn. For h2c with prior knowledge this is relatively simple. For that we just have to verify the HTTP/2 client preface has been written to the net.Conn, and then reforward the client preface to the hijacked connection. For h2c upgraded from HTTP/1, this is a bit more involved. First we validate the HTTP/1 Upgrade request, and respond to the client with 101 Switching Protocols. Then we write a HTTP/2 client preface on behalf of the client, and a settings frame and a headers frame which correspond to what was in the upgrade request. Then since http2.Server is going respond with a settings ACK, we swallow it so that it is not forwarded to the client since for h2c upgrade from HTTP/1 the 101 Switching Protocols response replaces the settins ACK. Fixes golang/go#14141 Change-Id: I435f40216c883809c0dcb75426c9c59e2ea31182 Reviewed-on: https://go-review.googlesource.com/112999 Reviewed-by: Brad Fitzpatrick <[email protected]>
Previous discussion: bradfitz/http2#59
h2c (http2 without TLS) is somewhat controversial part of the http2 spec. As discussed in the linked ticket above, major browser vendors are refusing to implement it.
I still think it is important to have first class support in Go standard library though.
The primary reason is that https is an ideal transport for inter-service RPC - c.f. grpc.io which in grpc/grpc-go has to re-implement large parts of http2 client and server transport to work around this issue.
But wanting to use http2 between internal services is not limited to gRPC.
My primary motivation is that I generally want my load balancer to terminate SSL. This is not just to reduce crypto overhead on individual application servers, but can actually be more secure.
To enable perfect forward secrecy, you have to keep tight controls over session ticket keys and rotate them frequently. This is much easier to achieve if only two (or a few) load balancer machines must terminate SSL rather than possibly hundreds or thousands of application servers, not to mention fewer places for an intruder to get hold of those keys.
Of course it's possible to terminate public SSL with perfect forward secrecy at LB and then have backend connections use internally signed certificates and forgo the same stringency, but in this setup TLS is pure overhead, and additional complexity to administer.
So I think it's pretty reasonable that Go, a language whose core competency is building network services, should at least provide a means to serve and consume h2c via the standard library http2 package. I'd even argue that it should be automatic, but if opinion is strong, I'd settle for just being possible without re-implementing large chunks of the http2 transports.
If there is agreement from core (@bradfitz has most insight I'd guess) on how this could be introduced, I'd be really happy to contribute the code. How about if it was a separate package
x/net/http2/h2c
and for now it's required to manually choose h2c transport in client/server?If not, this could be built as a (standalone) third party library that anyone who needs this can use for now.
Thanks for you help with this.
The text was updated successfully, but these errors were encountered: