IPoIB TX NAPI

index 7cc2b755413d..19c3ba2368d2 100644
@@ -381,7 +381,6 @@ struct ipoib_dev_priv {
 	unsigned	     tx_tail;
 	struct ib_sge	     tx_sge[MAX_SKB_FRAGS + 1];
 	struct ib_ud_wr      tx_wr;
-	unsigned	     tx_outstanding;
 	struct ib_wc	     send_wc[MAX_SEND_CQE];
 
 	struct ib_recv_wr    rx_wr;

index 7500c28eac6b..6e0fc592791e 100644
@@ -769,8 +769,8 @@ void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_
 	} else {
 		netif_trans_update(dev);
 		++tx->tx_head;
-
-		if (++priv->tx_outstanding == ipoib_sendq_size) {
+		++priv->tx_head;
+		if ((priv->tx_head - priv->tx_tail) == ipoib_sendq_size) {
 			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 				  tx->qp->qp_num);
 			netif_stop_queue(dev);

@@ -814,7 +814,8 @@ void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 	netif_tx_lock(dev);
 
 	++tx->tx_tail;
-	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
+	++priv->tx_tail;
+	if (unlikely((priv->tx_head - priv->tx_tail) == ipoib_sendq_size >> 1) &&
 	    netif_queue_stopped(dev) &&
 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 		netif_wake_queue(dev);

@@ -1220,8 +1221,9 @@ static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
 		ipoib_dma_unmap_tx(priv, tx_req);
 		dev_kfree_skb_any(tx_req->skb);
 		++p->tx_tail;
+		++priv->tx_tail;
 		netif_tx_lock_bh(p->dev);
-		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
+		if (unlikely(priv->tx_head - priv->tx_tail == ipoib_sendq_size >> 1) &&
 		    netif_queue_stopped(p->dev) &&
 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 			netif_wake_queue(p->dev);

index 89d82e27a445..c978f8ffd2bb 100644
@@ -406,7 +406,7 @@ static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 	dev_kfree_skb_any(tx_req->skb);
 
 	++priv->tx_tail;
-	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
+	if (unlikely((priv->tx_head - priv->tx_tail) == ipoib_sendq_size >> 1) &&
 	    netif_queue_stopped(dev) &&
 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 		netif_wake_queue(dev);

@@ -611,8 +611,8 @@ int ipoib_send(struct net_device *dev, struct sk_buff *skb,
 		priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
 	else
 		priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
-
-	if (++priv->tx_outstanding == ipoib_sendq_size) {
+	/* increase the tx_head after send success, but use it for queue state */
+	if (priv->tx_head - priv->tx_tail == ipoib_sendq_size - 1) {
 		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");

@@ -627,7 +627,6 @@ int ipoib_send(struct net_device *dev, struct sk_buff *skb,
 	if (unlikely(rc)) {
 		ipoib_warn(priv, "post_send failed, error %d\n", rc);
 		++dev->stats.tx_errors;
-		--priv->tx_outstanding;
 		ipoib_dma_unmap_tx(priv, tx_req);
 		dev_kfree_skb_any(skb);
 		if (netif_queue_stopped(dev))

@@ -640,7 +639,7 @@ int ipoib_send(struct net_device *dev, struct sk_buff *skb,
 		++priv->tx_head;
 	}
 
-	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
+	if (unlikely(priv->tx_head - priv->tx_tail > MAX_SEND_CQE))
 		while (poll_tx(priv))
 			; /* nothing */
 

@@ -773,7 +772,6 @@ int ipoib_ib_dev_stop_default(struct net_device *dev)
 				ipoib_dma_unmap_tx(priv, tx_req);
 				dev_kfree_skb_any(tx_req->skb);
 				++priv->tx_tail;
-				--priv->tx_outstanding;
 			}
 
 			for (i = 0; i < ipoib_recvq_size; ++i) {

index 19c3ba2368d2..b534d9412a1e 100644
@@ -331,7 +331,8 @@ struct ipoib_dev_priv {
 
 	struct net_device *dev;
 
-	struct napi_struct napi;
+	struct napi_struct send_napi;
+	struct napi_struct recv_napi;
 
 	unsigned long flags;
 

@@ -408,7 +409,6 @@ struct ipoib_dev_priv {
 #endif
 	u64	hca_caps;
 	struct ipoib_ethtool_st ethtool;
-	struct timer_list poll_timer;
 	unsigned max_send_sge;
 	bool sm_fullmember_sendonly_support;
 	const struct net_device_ops	*rn_ops;

@@ -475,9 +475,10 @@ extern struct workqueue_struct *ipoib_workqueue;
 
 /* functions */
 
-int ipoib_poll(struct napi_struct *napi, int budget);
-void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr);
-void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr);
+int ipoib_rx_poll(struct napi_struct *napi, int budget);
+int ipoib_tx_poll(struct napi_struct *napi, int budget);
+void ipoib_ib_rx_completion(struct ib_cq *cq, void *ctx_ptr);
+void ipoib_ib_tx_completion(struct ib_cq *cq, void *ctx_ptr);
 
 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
 				 struct ib_pd *pd, struct rdma_ah_attr *attr);

index 6e0fc592791e..87f4bd99cdf7 100644
@@ -757,30 +757,35 @@ void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_
 		return;
 	}
 
+	if ((priv->tx_head - priv->tx_tail) == ipoib_sendq_size - 1) {
+		ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
+			  tx->qp->qp_num);
+		netif_stop_queue(dev);
+	}
+
 	skb_orphan(skb);
 	skb_dst_drop(skb);
 
+	if (netif_queue_stopped(dev))
+		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
+				     IB_CQ_REPORT_MISSED_EVENTS)) {
+			ipoib_warn(priv, "IPoIB/CM:request notify on send CQ failed\n");
+			napi_schedule(&priv->send_napi);
+		}
+
 	rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1), tx_req);
 	if (unlikely(rc)) {
-		ipoib_warn(priv, "post_send failed, error %d\n", rc);
+		ipoib_warn(priv, "IPoIB/CM:post_send failed, error %d\n", rc);
 		++dev->stats.tx_errors;
 		ipoib_dma_unmap_tx(priv, tx_req);
 		dev_kfree_skb_any(skb);
+
+		if (netif_queue_stopped(dev))
+			netif_wake_queue(dev);
 	} else {
 		netif_trans_update(dev);
 		++tx->tx_head;
 		++priv->tx_head;
-		if ((priv->tx_head - priv->tx_tail) == ipoib_sendq_size) {
-			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
-				  tx->qp->qp_num);
-			netif_stop_queue(dev);
-			rc = ib_req_notify_cq(priv->send_cq,
-				IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
-			if (rc < 0)
-				ipoib_warn(priv, "request notify on send CQ failed\n");
-			else if (rc)
-				ipoib_send_comp_handler(priv->send_cq, dev);
-		}
 	}
 }
 

@@ -815,9 +820,10 @@ void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 
 	++tx->tx_tail;
 	++priv->tx_tail;
-	if (unlikely((priv->tx_head - priv->tx_tail) == ipoib_sendq_size >> 1) &&
-	    netif_queue_stopped(dev) &&
-	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
+
+	if (unlikely(netif_queue_stopped(dev) &&
+		     (priv->tx_head - priv->tx_tail) <= ipoib_sendq_size >> 1 &&
+		     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
 		netif_wake_queue(dev);
 
 	if (wc->status != IB_WC_SUCCESS &&

@@ -1046,7 +1052,7 @@ static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_
 {
 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 	struct ib_qp_init_attr attr = {
-		.send_cq		= priv->recv_cq,
+		.send_cq		= priv->send_cq,
 		.recv_cq		= priv->recv_cq,
 		.srq			= priv->cm.srq,
 		.cap.max_send_wr	= ipoib_sendq_size,

@@ -1220,9 +1226,9 @@ static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
 		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
 		ipoib_dma_unmap_tx(priv, tx_req);
 		dev_kfree_skb_any(tx_req->skb);
+		netif_tx_lock_bh(p->dev);
 		++p->tx_tail;
 		++priv->tx_tail;
-		netif_tx_lock_bh(p->dev);
 		if (unlikely(priv->tx_head - priv->tx_tail == ipoib_sendq_size >> 1) &&
 		    netif_queue_stopped(p->dev) &&
 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))

223 lines not on this page · raw message

index 0d6b4053cdf7..12b7f911f0e5 100644
@@ -1616,13 +1616,29 @@ static void ipoib_neigh_hash_uninit(struct net_device *dev)
 	wait_for_completion(&priv->ntbl.deleted);
 }
 
+static void ipoib_napi_add(struct net_device *dev)
+{
+	struct ipoib_dev_priv *priv = ipoib_priv(dev);
+
+	netif_napi_add(dev, &priv->recv_napi, ipoib_rx_poll, IPOIB_NUM_WC);
+	netif_napi_add(dev, &priv->send_napi, ipoib_tx_poll, MAX_SEND_CQE);
+}
+
+static void ipoib_napi_del(struct net_device *dev)
+{
+	struct ipoib_dev_priv *priv = ipoib_priv(dev);
+
+	netif_napi_del(&priv->recv_napi);
+	netif_napi_del(&priv->send_napi);
+}
+
 static void ipoib_dev_uninit_default(struct net_device *dev)
 {
 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 
 	ipoib_transport_dev_cleanup(dev);
 
-	netif_napi_del(&priv->napi);
+	ipoib_napi_del(dev);
 
 	ipoib_cm_dev_cleanup(dev);
 

@@ -1637,7 +1653,7 @@ static int ipoib_dev_init_default(struct net_device *dev)
 {
 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 
-	netif_napi_add(dev, &priv->napi, ipoib_poll, NAPI_POLL_WEIGHT);
+	ipoib_napi_add(dev);
 
 	/* Allocate RX/TX "rings" to hold queued skbs */
 	priv->rx_ring =	kzalloc(ipoib_recvq_size * sizeof *priv->rx_ring,

@@ -1665,8 +1681,6 @@ static int ipoib_dev_init_default(struct net_device *dev)
 	priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
 	priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff;
 
-	timer_setup(&priv->poll_timer, ipoib_ib_tx_timer_func, 0);
-
 	return 0;
 
 out_tx_ring_cleanup:

@@ -1676,7 +1690,7 @@ static int ipoib_dev_init_default(struct net_device *dev)
 	kfree(priv->rx_ring);
 
 out:
-	netif_napi_del(&priv->napi);
+	ipoib_napi_del(dev);
 	return -ENOMEM;
 }
 

index bb64baf25309..a1ed25422b72 100644
@@ -156,7 +156,7 @@ int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
 	};
 	struct ib_cq_init_attr cq_attr = {};
 
-	int ret, size;
+	int ret, size, req_vec;
 	int i;
 
 	size = ipoib_recvq_size + 1;

@@ -171,17 +171,21 @@ int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
 		if (ret != -ENOSYS)
 			return -ENODEV;
 
+	req_vec = (priv->port - 1) * 2;
+
 	cq_attr.cqe = size;
-	priv->recv_cq = ib_create_cq(priv->ca, ipoib_ib_completion, NULL,
-				     dev, &cq_attr);
+	cq_attr.comp_vector = req_vec % priv->ca->num_comp_vectors;
+	priv->recv_cq = ib_create_cq(priv->ca, ipoib_ib_rx_completion, NULL,
+				     priv, &cq_attr);
 	if (IS_ERR(priv->recv_cq)) {
 		printk(KERN_WARNING "%s: failed to create receive CQ\n", ca->name);
 		goto out_cm_dev_cleanup;
 	}
 
 	cq_attr.cqe = ipoib_sendq_size;
-	priv->send_cq = ib_create_cq(priv->ca, ipoib_send_comp_handler, NULL,
-				     dev, &cq_attr);
+	cq_attr.comp_vector = (req_vec + 1) % priv->ca->num_comp_vectors;
+	priv->send_cq = ib_create_cq(priv->ca, ipoib_ib_tx_completion, NULL,
+				     priv, &cq_attr);
 	if (IS_ERR(priv->send_cq)) {
 		printk(KERN_WARNING "%s: failed to create send CQ\n", ca->name);
 		goto out_free_recv_cq;

@@ -208,6 +212,9 @@ int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
 		goto out_free_send_cq;
 	}
 
+	if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
+		goto out_free_send_cq;
+
 	for (i = 0; i < MAX_SKB_FRAGS + 1; ++i)
 		priv->tx_sge[i].lkey = priv->pd->local_dma_lkey;
 

index b534d9412a1e..8033a006277f 100644
@@ -111,7 +111,7 @@ enum {
 	IPOIB_MCAST_FLAG_BUSY	  = 2,
 	IPOIB_MCAST_FLAG_ATTACHED = 3,
 
-	MAX_SEND_CQE		  = 16,
+	MAX_SEND_CQE		  = 64,
 	IPOIB_CM_COPYBREAK	  = 256,
 
 	IPOIB_NON_CHILD		  = 0,