IPoIB: Use separate CQ for UD send completions

Use a dedicated CQ for UD send completions. Also, do not arm the UD
send CQ, which reduces the number of interrupts generated.  This patch
farther reduces overhead by not calling poll CQ for every posted send
WR -- it does polls only when there 16 or more outstanding work requests.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

drivers/infiniband/ulp/ipoib/ipoib.h

@@ -95,6 +95,8 @@ enum {
 	IPOIB_MCAST_FLAG_SENDONLY = 1,
 	IPOIB_MCAST_FLAG_BUSY	  = 2,	/* joining or already joined */
 	IPOIB_MCAST_FLAG_ATTACHED = 3,
+
+	MAX_SEND_CQE		  = 16,
 };
 
 #define	IPOIB_OP_RECV   (1ul << 31)
@@ -285,7 +287,8 @@ struct ipoib_dev_priv {
 	u16		  pkey_index;
 	struct ib_pd	 *pd;
 	struct ib_mr	 *mr;
-	struct ib_cq	 *cq;
+	struct ib_cq	 *recv_cq;
+	struct ib_cq	 *send_cq;
 	struct ib_qp	 *qp;
 	u32		  qkey;
 
@@ -305,6 +308,7 @@ struct ipoib_dev_priv {
 	struct ib_sge	     tx_sge[MAX_SKB_FRAGS + 1];
 	struct ib_send_wr    tx_wr;
 	unsigned	     tx_outstanding;
+	struct ib_wc	     send_wc[MAX_SEND_CQE];
 
 	struct ib_recv_wr    rx_wr;
 	struct ib_sge	     rx_sge[IPOIB_UD_RX_SG];
@@ -662,7 +666,6 @@ static inline int ipoib_register_debugfs(void) { return 0; }
 static inline void ipoib_unregister_debugfs(void) { }
 #endif
 
-
 #define ipoib_printk(level, priv, format, arg...)	\
 	printk(level "%s: " format, ((struct ipoib_dev_priv *) priv)->dev->name , ## arg)
 #define ipoib_warn(priv, format, arg...)		\

drivers/infiniband/ulp/ipoib/ipoib_cm.c

@@ -249,8 +249,8 @@ static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 	struct ipoib_dev_priv *priv = netdev_priv(dev);
 	struct ib_qp_init_attr attr = {
 		.event_handler = ipoib_cm_rx_event_handler,
-		.send_cq = priv->cq, /* For drain WR */
-		.recv_cq = priv->cq,
+		.send_cq = priv->recv_cq, /* For drain WR */
+		.recv_cq = priv->recv_cq,
 		.srq = priv->cm.srq,
 		.cap.max_send_wr = 1, /* For drain WR */
 		.cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
@@ -951,8 +951,8 @@ static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_
 {
 	struct ipoib_dev_priv *priv = netdev_priv(dev);
 	struct ib_qp_init_attr attr = {
-		.send_cq		= priv->cq,
-		.recv_cq		= priv->cq,
+		.send_cq		= priv->recv_cq,
+		.recv_cq		= priv->recv_cq,
 		.srq			= priv->cm.srq,
 		.cap.max_send_wr	= ipoib_sendq_size,
 		.cap.max_send_sge	= 1,

drivers/infiniband/ulp/ipoib/ipoib_ethtool.c

@@ -71,7 +71,7 @@ static int ipoib_set_coalesce(struct net_device *dev,
 	    coal->rx_max_coalesced_frames > 0xffff)
 		return -EINVAL;
 
-	ret = ib_modify_cq(priv->cq, coal->rx_max_coalesced_frames,
+	ret = ib_modify_cq(priv->recv_cq, coal->rx_max_coalesced_frames,
 			   coal->rx_coalesce_usecs);
 	if (ret && ret != -ENOSYS) {
 		ipoib_warn(priv, "failed modifying CQ (%d)\n", ret);

drivers/infiniband/ulp/ipoib/ipoib_ib.c

@@ -364,7 +364,6 @@ static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 	struct ipoib_dev_priv *priv = netdev_priv(dev);
 	unsigned int wr_id = wc->wr_id;
 	struct ipoib_tx_buf *tx_req;
-	unsigned long flags;
 
 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
 		       wr_id, wc->status);
@@ -384,13 +383,11 @@ static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 
 	dev_kfree_skb_any(tx_req->skb);
 
-	spin_lock_irqsave(&priv->tx_lock, flags);
 	++priv->tx_tail;
 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 	    netif_queue_stopped(dev) &&
 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 		netif_wake_queue(dev);
-	spin_unlock_irqrestore(&priv->tx_lock, flags);
 
 	if (wc->status != IB_WC_SUCCESS &&
 	    wc->status != IB_WC_WR_FLUSH_ERR)
@@ -399,6 +396,17 @@ static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 			   wc->status, wr_id, wc->vendor_err);
 }
 
+static int poll_tx(struct ipoib_dev_priv *priv)
+{
+	int n, i;
+
+	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
+	for (i = 0; i < n; ++i)
+		ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
+
+	return n == MAX_SEND_CQE;
+}
+
 int ipoib_poll(struct napi_struct *napi, int budget)
 {
 	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
@@ -414,7 +422,7 @@ int ipoib_poll(struct napi_struct *napi, int budget)
 		int max = (budget - done);
 
 		t = min(IPOIB_NUM_WC, max);
-		n = ib_poll_cq(priv->cq, t, priv->ibwc);
+		n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
 
 		for (i = 0; i < n; i++) {
 			struct ib_wc *wc = priv->ibwc + i;
@@ -425,12 +433,8 @@ int ipoib_poll(struct napi_struct *napi, int budget)
 					ipoib_cm_handle_rx_wc(dev, wc);
 				else
 					ipoib_ib_handle_rx_wc(dev, wc);
-			} else {
-				if (wc->wr_id & IPOIB_OP_CM)
-					ipoib_cm_handle_tx_wc(dev, wc);
-				else
-					ipoib_ib_handle_tx_wc(dev, wc);
-			}
+			} else
+				ipoib_cm_handle_tx_wc(priv->dev, wc);
 		}
 
 		if (n != t)
@@ -439,7 +443,7 @@ int ipoib_poll(struct napi_struct *napi, int budget)
 
 	if (done < budget) {
 		netif_rx_complete(dev, napi);
-		if (unlikely(ib_req_notify_cq(priv->cq,
+		if (unlikely(ib_req_notify_cq(priv->recv_cq,
 					      IB_CQ_NEXT_COMP |
 					      IB_CQ_REPORT_MISSED_EVENTS)) &&
 		    netif_rx_reschedule(dev, napi))
@@ -562,12 +566,16 @@ void ipoib_send(struct net_device *dev, struct sk_buff *skb,
 
 		address->last_send = priv->tx_head;
 		++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);
 }
 
 static void __ipoib_reap_ah(struct net_device *dev)
@@ -714,7 +722,7 @@ void ipoib_drain_cq(struct net_device *dev)
 	struct ipoib_dev_priv *priv = netdev_priv(dev);
 	int i, n;
 	do {
-		n = ib_poll_cq(priv->cq, IPOIB_NUM_WC, priv->ibwc);
+		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
 		for (i = 0; i < n; ++i) {
 			/*
 			 * Convert any successful completions to flush
@@ -729,14 +737,13 @@ void ipoib_drain_cq(struct net_device *dev)
 					ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
 				else
 					ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
-			} else {
-				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
-					ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
-				else
-					ipoib_ib_handle_tx_wc(dev, priv->ibwc + i);
-			}
+			} else
+				ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
 		}
 	} while (n == IPOIB_NUM_WC);
+
+	while (poll_tx(priv))
+		; /* nothing */
 }
 
 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
@@ -826,7 +833,7 @@ int ipoib_ib_dev_stop(struct net_device *dev, int flush)
 		msleep(1);
 	}
 
-	ib_req_notify_cq(priv->cq, IB_CQ_NEXT_COMP);
+	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
 
 	return 0;
 }

drivers/infiniband/ulp/ipoib/ipoib_main.c

@@ -1298,7 +1298,8 @@ static int __init ipoib_init_module(void)
 
 	ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
 	ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
-	ipoib_sendq_size = max(ipoib_sendq_size, IPOIB_MIN_QUEUE_SIZE);
+	ipoib_sendq_size = max(ipoib_sendq_size, max(2 * MAX_SEND_CQE,
+						     IPOIB_MIN_QUEUE_SIZE));
 #ifdef CONFIG_INFINIBAND_IPOIB_CM
 	ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
 #endif

drivers/infiniband/ulp/ipoib/ipoib_verbs.c

@@ -171,26 +171,33 @@ int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
 		goto out_free_pd;
 	}
 
-	size = ipoib_sendq_size + ipoib_recvq_size + 1;
+	size = ipoib_recvq_size + 1;
 	ret = ipoib_cm_dev_init(dev);
 	if (!ret) {
+		size += ipoib_sendq_size;
 		if (ipoib_cm_has_srq(dev))
 			size += ipoib_recvq_size + 1; /* 1 extra for rx_drain_qp */
 		else
 			size += ipoib_recvq_size * ipoib_max_conn_qp;
 	}
 
-	priv->cq = ib_create_cq(priv->ca, ipoib_ib_completion, NULL, dev, size, 0);
-	if (IS_ERR(priv->cq)) {
-		printk(KERN_WARNING "%s: failed to create CQ\n", ca->name);
+	priv->recv_cq = ib_create_cq(priv->ca, ipoib_ib_completion, NULL, dev, size, 0);
+	if (IS_ERR(priv->recv_cq)) {
+		printk(KERN_WARNING "%s: failed to create receive CQ\n", ca->name);
 		goto out_free_mr;
 	}
 
-	if (ib_req_notify_cq(priv->cq, IB_CQ_NEXT_COMP))
-		goto out_free_cq;
+	priv->send_cq = ib_create_cq(priv->ca, NULL, NULL, dev, ipoib_sendq_size, 0);
+	if (IS_ERR(priv->send_cq)) {
+		printk(KERN_WARNING "%s: failed to create send CQ\n", ca->name);
+		goto out_free_recv_cq;
+	}
+
+	if (ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP))
+		goto out_free_send_cq;
 
-	init_attr.send_cq = priv->cq;
-	init_attr.recv_cq = priv->cq;
+	init_attr.send_cq = priv->send_cq;
+	init_attr.recv_cq = priv->recv_cq;
 
 	if (priv->hca_caps & IB_DEVICE_UD_TSO)
 		init_attr.create_flags = IB_QP_CREATE_IPOIB_UD_LSO;
@@ -201,7 +208,7 @@ int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
 	priv->qp = ib_create_qp(priv->pd, &init_attr);
 	if (IS_ERR(priv->qp)) {
 		printk(KERN_WARNING "%s: failed to create QP\n", ca->name);
-		goto out_free_cq;
+		goto out_free_send_cq;
 	}
 
 	priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
@@ -230,8 +237,11 @@ int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
 
 	return 0;
 
-out_free_cq:
-	ib_destroy_cq(priv->cq);
+out_free_send_cq:
+	ib_destroy_cq(priv->send_cq);
+
+out_free_recv_cq:
+	ib_destroy_cq(priv->recv_cq);
 
 out_free_mr:
 	ib_dereg_mr(priv->mr);
@@ -254,8 +264,11 @@ void ipoib_transport_dev_cleanup(struct net_device *dev)
 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 	}
 
-	if (ib_destroy_cq(priv->cq))
-		ipoib_warn(priv, "ib_cq_destroy failed\n");
+	if (ib_destroy_cq(priv->send_cq))
+		ipoib_warn(priv, "ib_cq_destroy (send) failed\n");
+
+	if (ib_destroy_cq(priv->recv_cq))
+		ipoib_warn(priv, "ib_cq_destroy (recv) failed\n");
 
 	ipoib_cm_dev_cleanup(dev);