When "idle" isn't idle: how a Linux kernel optimization became a QUIC bug

The article discusses a bug related to the CUBIC congestion control algorithm used in Linux that affects QUIC connections. The problem arises when the congestion window (cwnd) is pinned at its minimum after a congestion collapse, preventing recovery. This defect was revealed during tests simulating heavy packet loss, which led to a high failure rate (61%) in connection attempts.

The bug is traced back to a Linux kernel change that mismanages the ‘epoch’ reference for cwnd growth when idle, leading to a cycle where the congestion controller oscillates without growth. The solution involved adjusting how idle periods are measured in the QUIC implementation ('quiche'), optimizing the timing of recovery states to prevent this self-reinforcing issue, which has been successfully resolved with a small code change.