Envoy HTTP2 Connection sends RST_STREAM before initial HEADERS which causes upstream server to close connection with PROTOCOL_ERROR

Description

Preface: we’re using envoy as an L7 proxy with a HTTP2 upstream. We’ve noticed that sometimes Envoy sometimes sends a RST_STREAM before actually sending the HEADERS frame to change the stream state from idle to open.

The following is a pcap of a single TCP connection between envoy and the upstream server. The upstream server is listening on 8080 while the ephemeral port that envoy cx is using is 63692.

Pcap file

On #243, we can see that envoy starts sending RST_STREAM for streams that haven’t received HEADERS or PUSH-PROMISE frames. Additionally, some of these stream resets are repeated (e.g. stream id 27, 23 from above). Because of this the server closes the connection with a PROTOCOL_ERROR as per HTTP2 spec.

Excerpt:

RST_STREAM frames MUST NOT be sent for a stream in the "idle" state. If a RST_STREAM frame identifying an idle stream is received, the recipient MUST treat this as a connection error (Section 5.4.1) of type PROTOCOL_ERROR.

This leads to envoy dropping in-flight requests on the same HTTP2 connection with a 503 UC. We believe this is because of the way we do deferred reset after reading this code snippet. In other words, we believe that from envoy’s prospective, this stream exists, but since the first frame hasn’t made it to the server yet, the server believes that the stream is idle and does an active close.

Repro Steps

A very contrived example can be done using a single upstream go service coupled with a load tester and envoy.

Version details:

• Go 1.14.2
• envoyproxy/envoy:v1.14.1 docker image
• Hey - 0.1.3

Configuration:

• HTTP2 Server - https://gist.github.com/tan-stripe/4ea1e8f3dfbb3af6210f655dc2883647
• envoy.yaml - https://gist.github.com/tan-stripe/8965f551d7ffe68cdb4c4d738d143a63
• Dockerfile - https://gist.github.com/tan-stripe/d6bfb04c5ad8c044442b0aefa07d898b

Steps:

1. Start up a simple HTTP2 server with some delay in the response
2. Start envoy with given configuration
3. Run a hey loadtest - hey -n 50 -c 50 -h2 -t 3 http://localhost:10000
   • You may have to tune the parameters a little bit, but we were able to reliably cause an error message on the go program side. Error: http2: server connection error from 127.0.0.1:65100: connection error: PROTOCOL_ERROR
   • This error comes from receiving a RST_STREAM on an idle stream
   • Any connection on the same TCP connection gets dropped due to a TCP-FIN initiated by the go server