From: Simon Guinot <simon.guinot@xxxxxxxxxxxx> Date: Wed, 8 Nov 2017 17:58:35 +0100 > @@ -2413,8 +2416,7 @@ static int mvneta_tx(struct sk_buff *skb, struct net_device *dev) > if (txq->count >= txq->tx_stop_threshold) > netif_tx_stop_queue(nq); > > - if (!skb->xmit_more || netif_xmit_stopped(nq) || > - txq->pending + frags > MVNETA_TXQ_DEC_SENT_MASK) > + if (!skb->xmit_more || netif_xmit_stopped(nq)) > mvneta_txq_pend_desc_add(pp, txq, frags); > else > txq->pending += frags; As David Laight said, you should not allow unlimited amounts of ->xmit_more frames to be queued without a TX doorbell update. Therefore, please keep some kind of limit here otherwise latency will spike in some circumstances.