Respond with an HTTP 4xx than only a `RST_STREAM` to malformed requests
franfastly opened this issue · comments
Issue
h2
returns RST_STREAM
frames with the PROTOCOL_ERROR
bit set as a response to many types of errors in client requests. Many of those cases, when handled by an HTTP/1 server such as the one used in hyper
, would result in an HTTP 400 Bad Request response returned to the client rather than a TCP reset (the HTTP/1 equivalent of a RST_STREAM
). As a consequence, a client will observe differences in behaviour between HTTP/1 and HTTP/2: a malformed request will result in a 400 response when HTTP/1 is used whereas the same request will result in a reset stream with h2
.
Proposal
There was a conversation in the mailing list of the IETF-HTTP-WG about whether a server can send a response to malformed requests instead of treating them as a stream error. From Steffan Eising's last response to that thread:
- Sending a HTTP response is preferable, since this gives much better error reporting on the client side.
- Violations of HTTP/2 framing layer should error on the h2 layer. Basically, the server no longer trusts the client.
- Sending a RESPONSE with subsequent RST is a good pattern
Although there was no consensus about specifying a single correct behaviour, I think it was clear that a 400 response was acceptable.
I could create a PR to make h2
reply a HEADERS
+400
+END_STREAM
frame followed by a RST_STREAM
+PROTOCOL_ERROR
frame to all the malformed!()
macro invocations in
Line 1506 in 0077d3d
I remember that conversation. I think that's probably generally fine.
cc @acfoltzer @nox @Noah-Kennedy, thoughts?
@seanmonstar I've been working with @franfastly on this investigation and an internal patch we've been cooking up to address the issue. I think overall I would prefer that h2
just return HEADERS+END_STREAM
in these cases where the framing of h2 itself is not at risk, but at the moment that behavior would break h2spec tests. Pending movement on that issue, Fran and I believe this is the best solution.