Until this PR, we were naively closing the quic.Stream whenever
the callstack for handling the request (HTTP or TCP) finished.
However, our proxy handler may still be reading or writing from
the quic.Stream at that point, because we return the callstack if
either side finishes, but not necessarily both.
This is a problem for quic-go library because quic.Stream#Close
cannot be called concurrently with quic.Stream#Write
Furthermore, we also noticed that quic.Stream#Close does nothing
to do receiving stream (since, underneath, quic.Stream has 2 streams,
1 for each direction), thus leaking memory, as explained in:
https://github.com/lucas-clemente/quic-go/issues/3322
This PR addresses both problems by wrapping the quic.Stream that
is passed down to the proxying logic and handle all these concerns.
This adds various bug fixes when investigating why QUIC transports were
not being unregistered when they should (and only when the graceful shutdown
started).
Most of these bug fixes are making the QUIC transport implementation closer
to its HTTP2 counterpart:
- ServeControlStream is now a blocking function (it's up to the transport to handle that)
- QUIC transport then handles the control plane as part of its Serve, thus waiting for it on shutdown
- QUIC transport now returns "non recoverable" for connections with similar semantics to HTTP2 and H2mux
- QUIC transport no longer has a loop around its Serve logic that retries connections on its own (that logic is upstream)
This does a few fixes to make sure that the QUICConnection returns from
Serve when the context is cancelled.
QUIC transport now behaves like other transports: closes as soon as there
is no traffic, or at most by grace-period. Note that we do not wait for
UDP traffic since that's connectionless by design.
Creates an abstraction over UDP Conn for origin "connection" which can
be useful for future support of complex protocols that may require
changing ports during protocol negotiation (eg. SIP, TFTP)
In addition, it removes a dependency from ingress on connection package.
- Refactors some h2mux specific logic from connection/header.go to connection/h2mux_header.go
- Do the same for the unit tests
- Add a non-h2mux "is control response header" function (we don't need one for the request flow)
- In that new function, do not consider "content-length" as a control header
- Use that function in the non-h2mux flow for response (and it will be used also in origintunneld)
The default max streams value of 100 is rather small when subject to
high load in terms of connecting QUIC with streams faster than it can
create new ones. This high value allows for more throughput.
Go's client defaults to chunked encoding after a 200ms delay if the following cases are true:
* the request body blocks
* the content length is not set (or set to -1)
* the method doesn't usually have a body (GET, HEAD, DELETE, ...)
* there is no transfer-encoding=chunked already set.
So for non websocket requests, if transfer-encoding isn't chunked and content length is 0, we dont set a request body.
ServeControlStream accidentally became non-blocking in the last quic
change causing stream to not be returned until a SIGTERM was received.
This change makes ServeControlStream be non-blocking for QUIC streams.
This maximum grace period will be honored by Cloudflare edge such that
either side will close the connection after unregistration at most
by this time (3min as of this commit):
- If the connection is unused, it is already closed as soon as possible.
- If the connection is still used, it is closed on the cloudflared configured grace-period.
Even if cloudflared does not close the connection by the grace-period time,
the edge will do so.
Sudarsan Reddy
Nuno Diegues
João Oliveirinha
cthuang
Dimitris Apostolou
Areg Harutyunyan
Rishabh Bector