mirror of
https://github.com/tailscale/tailscale.git
synced 2025-07-29 15:23:45 +00:00
91d65e03e8
Cherry picks bug fix #16678 and flake fix #16680 onto the 1.86 release branch. When kubectl starts an interactive attach session, it sends 2 resize messages in quick succession. It seems that particularly in HTTP mode, we often receive both of these WebSocket frames from the underlying connection in a single read. However, our parser currently assumes 0-1 frames per read, and leaves the second frame in the read buffer until the next read from the underlying connection. It doesn't take long after that before we end up failing to skip a control message as we normally should, and then we parse a control message as though it will have a stream ID (part of the Kubernetes protocol) and error out. Instead, we should keep parsing frames from the read buffer for as long as we're able to parse complete frames, so this commit refactors the messages parsing logic into a loop based on the contents of the read buffer being non-empty. k/k staging/src/k8s.io/kubectl/pkg/cmd/attach/attach.go for full details of the resize messages. There are at least a couple more multiple-frame read edge cases we should handle, but this commit is very conservatively fixing a single observed issue to make it a low-risk candidate for cherry picking. Updates #13358 Change-Id: Iafb91ad1cbeed9c5231a1525d4563164fc1f002f Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>