From: Eric Dumazet <eric.duma...@gmail.com> Date: Fri, 7 Jun 2019 12:18:21 -0700
> Note that this suggests the driver is not performing TX completion > fast enough. > > Looking at the driver, I do not see anything requesting interrupt > mitigation, so this might also be caused by a race in the driver > (some skbs being not TX completed until another unrelated xmit is > requested) I think the INTR bit needs to be set in the individual TX descriptors. This is a classic case of a chip that only interrupts when the entire TX queue is emptied and the TX engine enters the idle state. I don't like this skb_orphan() solution as it kills UDP flow control. I'm not applying this.
