The UDP proxy is setting a deadline of 90 seconds when reading from the
backend. If no data is received within this interval, it reclaims the
connection.
This means, the backend would see a different connection every 90
seconds if the backend never sends back any reply to a client.
This change prevents the proxy from eagerly GC'ing such connections by
taking into account the last time a datagram was proxyed to the backend.
Signed-off-by: Albin Kerouanton <albinker@gmail.com>
The UDP proxy used by cmd/docker-proxy is executing Write and Close in
two separate goroutines, such that a Close could interrupt an in-flight
Write.
Introduce a `connTrackEntry` that wraps a `net.Conn` and a `sync.Mutex`
to ensure that Write and Close are serialized.
Signed-off-by: Albin Kerouanton <albinker@gmail.com>
When a UDP server is running on a multihomed server, as is the case with
pretty much _all_ Docker hosts (eg. eth0 + docker0), the kernel has to
choose which source address is used when replying to a UDP client. But
that process is based on heuristics and is fallible.
If the address picked doesn't match the original destination address
used by the client, it'll drop the datagram and return an ICMP Port
Unreachable.
To prevent that, we need to:
- `setsockopt(IP_PKTINFO)` on proxy's sockets.
- Extract the original destination address from an ancillary message
every time a new 'UDP connection' is 'established' (ie. every time we
insert a new entry into the UDP conntrack table).
- And finally, pass a control message containing the desired source
address to the kernel, every time we send a response back to the
client.
Also, update the inline comment on read errors in `(*UDPProxy).Run()`.
This comment was misleadingly referencing ECONNREFUSED - Linux's UDP
implementation never returns this error (see [1]). Instead, state why
`net.ErrClosed` is perfectly fine and doesn't need to be logged
(although, docker-proxy currently logs to nowhere).
[1]: https://github.com/search?q=repo%3Atorvalds%2Flinux+ECONNREFUSED+path%3A%2F%5Enet%5C%2F%28ipv4%7Cipv6%29%5C%2F%28udp%7Ctcp%29%2F&type=code
Signed-off-by: Albin Kerouanton <albinker@gmail.com>
In preparation for the daemon passing a listen fd, add command line
option -use-listen-fd to indicate that the fd is present (as fd 4).
If the new flag isn't given, open the listener as normal.
Refactor the TCP and UDP proxies to be constructed with an existing
TCPListener or UDPConn, respectively. Lift the responsibilty of opening
the listener to the entrypoint. Per the Single Responsibility Principle,
this structure affords changing how the listener is created without
having to touch the proxy implementations.
Co-authored-by: Cory Snider <csnider@mirantis.com>
Signed-off-by: Rob Murray <rob.murray@docker.com>
