Hi everyone While investigating a bug report on CAKE[0], I've run into the following behaviour:
When running CAKE as an ingress shaper on an IFB interface, if the GSO splitting feature is turned on, TCP throughput will drop dramatically on 6in4 (sit) tunnels running over the interface in question. Looking at a traffic dump, I'm seeing ~15% packet loss on the encapsulated TCP stream. IPv4 traffic is fine on the same interface, as is native IPv6 traffic. And turning off GSO splitting in CAKE makes the packet loss go away. The issue only seems to appear on IFB interfaces. So I'm wondering if there is some interaction that corrupts packets when they are being split in this configuration? Steps to reproduce (assuming the box you are running on has IP 10.0.0.2 on eth0, and has a peer at 10.0.0.1 with a suitably configured sit tunnel): # modprobe ifb # ip link set dev ifb0 up # tc qdisc add dev eth0 handle ffff: ingress # tc filter add dev eth0 parent ffff: protocol all prio 10 matchall action mirred egress redirect dev ifb0 # tc qdisc replace dev ifb0 root cake # ip link add type sit local 10.0.0.2 remote 10.0.0.1 # ip link set dev sit1 up # netperf -H fe80::a00:1%sit1 -t TCP_MAERTS Whereas, in the same setup, this will work fine: # netperf -H 10.0.0.1 -t TCP_MAERTS As will this: # tc qdisc replace dev ifb0 root cake no-split-gso # netperf -H fe80::a00:1%sit1 -t TCP_MAERTS Does anyone have any ideas? :) -Toke [0] https://github.com/tohojo/sqm-scripts/issues/72 _______________________________________________ Cake mailing list [email protected] https://lists.bufferbloat.net/listinfo/cake
