On (03/30/16 13:15), Eric Dumazet wrote: > You might try netperf -t TCP_SENDFILE -- -m 150 > > to let netperf use sendfile() on small frags.
that still did not reproduce it but let me try beating on that approach with more permutations. BTW, another data-point that may help debug this: even with i40e, if you use the "-o" option to the rds-stress invocation, there are no problems: the "-o" option enforces uni-directional data transfer, so one side is pure-Tx, other side is pure-Rx. It is only when both sides simultaneously do both Tx and Rx on the tcp socket that you see the issue. I dont know if that provides any clues. --Sowmini