netif_rx_schedule(dev, &priv->napi);
 }
 
+static void drain_tx_cq(struct net_device *dev)
+{
+       struct ipoib_dev_priv *priv = netdev_priv(dev);
+       unsigned long flags;
+
+       spin_lock_irqsave(&priv->tx_lock, flags);
+       while (poll_tx(priv))
+               ; /* nothing */
+
+       if (netif_queue_stopped(dev))
+               mod_timer(&priv->poll_timer, jiffies + 1);
+
+       spin_unlock_irqrestore(&priv->tx_lock, flags);
+}
+
+void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
+{
+       drain_tx_cq((struct net_device *)dev_ptr);
+}
+
 static inline int post_send(struct ipoib_dev_priv *priv,
                            unsigned int wr_id,
                            struct ib_ah *address, u32 qpn,
        else
                priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
 
+       if (++priv->tx_outstanding == ipoib_sendq_size) {
+               ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
+               if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
+                       ipoib_warn(priv, "request notify on send CQ failed\n");
+               netif_stop_queue(dev);
+       }
+
        if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
                               address->ah, qpn, tx_req, phead, hlen))) {
                ipoib_warn(priv, "post_send failed\n");
                ++dev->stats.tx_errors;
+               --priv->tx_outstanding;
                ipoib_dma_unmap_tx(priv->ca, tx_req);
                dev_kfree_skb_any(skb);
+               if (netif_queue_stopped(dev))
+                       netif_wake_queue(dev);
        } else {
                dev->trans_start = jiffies;
 
                ++priv->tx_head;
                skb_orphan(skb);
 
-               if (++priv->tx_outstanding == ipoib_sendq_size) {
-                       ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
-                       netif_stop_queue(dev);
-               }
        }
 
        if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
-               poll_tx(priv);
+               while (poll_tx(priv))
+                       ; /* nothing */
 }
 
 static void __ipoib_reap_ah(struct net_device *dev)
                                   round_jiffies_relative(HZ));
 }
 
+static void ipoib_ib_tx_timer_func(unsigned long ctx)
+{
+       drain_tx_cq((struct net_device *)ctx);
+}
+
 int ipoib_ib_dev_open(struct net_device *dev)
 {
        struct ipoib_dev_priv *priv = netdev_priv(dev);
        queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
                           round_jiffies_relative(HZ));
 
+       init_timer(&priv->poll_timer);
+       priv->poll_timer.function = ipoib_ib_tx_timer_func;
+       priv->poll_timer.data = (unsigned long)dev;
+
        set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
 
        return 0;
        ipoib_dbg(priv, "All sends and receives done.\n");
 
 timeout:
+       del_timer_sync(&priv->poll_timer);
        qp_attr.qp_state = IB_QPS_RESET;
        if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
                ipoib_warn(priv, "Failed to modify QP to RESET state\n");