Skip to contentSkip to conversation

ifb: remove function declarations

PATCH In review Changli Gao 4 participants 21 messages Dec 13, 2010 – Dec 15, 2010

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/5 ifb: remove function declarations

Restructure to remove function declarations.

---
 drivers/net/ifb.c |   51 +++++++++++++++++++++++----------------------------
 1 file changed, 23 insertions(+), 28 deletions(-)
drivers/net/ifb.c +23 -28
index bfa03db..0667a61 100644
@@ -46,11 +46,6 @@ struct ifb_private {
 
 static int numifbs = 2;
 
-static void ri_tasklet(unsigned long dev);
-static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev);
-static int ifb_open(struct net_device *dev);
-static int ifb_close(struct net_device *dev);
-
 static void ri_tasklet(unsigned long dev)
 {
 
@@ -119,29 +114,6 @@ resched:
 
 }
 
-static const struct net_device_ops ifb_netdev_ops = {
-	.ndo_open	= ifb_open,
-	.ndo_stop	= ifb_close,
-	.ndo_start_xmit	= ifb_xmit,
-	.ndo_validate_addr = eth_validate_addr,
-};
-
-static void ifb_setup(struct net_device *dev)
-{
-	/* Initialize the device structure. */
-	dev->destructor = free_netdev;
-	dev->netdev_ops = &ifb_netdev_ops;
-
-	/* Fill in device structure with ethernet-generic values. */
-	ether_setup(dev);
-	dev->tx_queue_len = TX_Q_LIMIT;
-
-	dev->flags |= IFF_NOARP;
-	dev->flags &= ~IFF_MULTICAST;
-	dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
-	random_ether_addr(dev->dev_addr);
-}
-
 static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct ifb_private *dp = netdev_priv(dev);
@@ -193,6 +165,29 @@ static int ifb_open(struct net_device *dev)
 	return 0;
 }
 
+static const struct net_device_ops ifb_netdev_ops = {
+	.ndo_open		= ifb_open,
+	.ndo_stop		= ifb_close,
+	.ndo_start_xmit		= ifb_xmit,
+	.ndo_validate_addr	= eth_validate_addr,
+};
+
+static void ifb_setup(struct net_device *dev)
+{
+	/* Initialize the device structure. */
+	dev->destructor = free_netdev;
+	dev->netdev_ops = &ifb_netdev_ops;
+
+	/* Fill in device structure with ethernet-generic values. */
+	ether_setup(dev);
+	dev->tx_queue_len = TX_Q_LIMIT;
+
+	dev->flags |= IFF_NOARP;
+	dev->flags &= ~IFF_MULTICAST;
+	dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
+	random_ether_addr(dev->dev_addr);
+}
+
 static int ifb_validate(struct nlattr *tb[], struct nlattr *data[])
 {
 	if (tb[IFLA_ADDRESS]) {
#1 lore raw

PATCH 2/5 ifb: code cleanup

Clean up code according to scripts/checkpatch.pl

---
 drivers/net/ifb.c |   22 ++++++++++++----------
 1 file changed, 12 insertions(+), 10 deletions(-)
drivers/net/ifb.c +12 -10
index 0667a61..1628d01 100644
@@ -15,14 +15,14 @@
 	by Patrick McHardy and then maintained by Andre Correa.
 
 	You need the tc action  mirror or redirect to feed this device
-       	packets.
+	packets.
 
 	This program is free software; you can redistribute it and/or
 	modify it under the terms of the GNU General Public License
 	as published by the Free Software Foundation; either version
 	2 of the License, or (at your option) any later version.
 
-  	Authors:	Jamal Hadi Salim (2005)
+	Authors:	Jamal Hadi Salim (2005)
 
 */
 
@@ -56,7 +56,8 @@ static void ri_tasklet(unsigned long dev)
 	struct sk_buff *skb;
 
 	txq = netdev_get_tx_queue(_dev, 0);
-	if ((skb = skb_peek(&dp->tq)) == NULL) {
+	skb = skb_peek(&dp->tq);
+	if (skb == NULL) {
 		if (__netif_tx_trylock(txq)) {
 			skb_queue_splice_tail_init(&dp->rq, &dp->tq);
 			__netif_tx_unlock(txq);
@@ -72,7 +73,7 @@ static void ri_tasklet(unsigned long dev)
 		skb->tc_verd = 0;
 		skb->tc_verd = SET_TC_NCLS(skb->tc_verd);
 		stats->tx_packets++;
-		stats->tx_bytes +=skb->len;
+		stats->tx_bytes += skb->len;
 
 		rcu_read_lock();
 		skb->dev = dev_get_by_index_rcu(&init_net, skb->skb_iif);
@@ -97,7 +98,8 @@ static void ri_tasklet(unsigned long dev)
 	}
 
 	if (__netif_tx_trylock(txq)) {
-		if ((skb = skb_peek(&dp->rq)) == NULL) {
+		skb = skb_peek(&dp->rq);
+		if (skb == NULL) {
 			dp->tasklet_pending = 0;
 			if (netif_queue_stopped(_dev))
 				netif_wake_queue(_dev);
@@ -121,7 +123,7 @@ static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
 	u32 from = G_TC_FROM(skb->tc_verd);
 
 	stats->rx_packets++;
-	stats->rx_bytes+=skb->len;
+	stats->rx_bytes += skb->len;
 
 	if (!(from & (AT_INGRESS|AT_EGRESS)) || !skb->skb_iif) {
 		dev_kfree_skb(skb);
@@ -129,9 +131,8 @@ static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
 		return NETDEV_TX_OK;
 	}
 
-	if (skb_queue_len(&dp->rq) >= dev->tx_queue_len) {
+	if (skb_queue_len(&dp->rq) >= dev->tx_queue_len)
 		netif_stop_queue(dev);
-	}
 
 	__skb_queue_tail(&dp->rq, skb);
 	if (!dp->tasklet_pending) {
@@ -150,6 +151,7 @@ static int ifb_close(struct net_device *dev)
 	netif_stop_queue(dev);
 	__skb_queue_purge(&dp->rq);
 	__skb_queue_purge(&dp->tq);
+
 	return 0;
 }
 
@@ -196,6 +198,7 @@ static int ifb_validate(struct nlattr *tb[], struct nlattr *data[])
 		if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
 			return -EADDRNOTAVAIL;
 	}
+
 	return 0;
 }
 
@@ -215,8 +218,7 @@ static int __init ifb_init_one(int index)
 	struct net_device *dev_ifb;
 	int err;
 
-	dev_ifb = alloc_netdev(sizeof(struct ifb_private),
-				 "ifb%d", ifb_setup);
+	dev_ifb = alloc_netdev(sizeof(struct ifb_private), "ifb%d", ifb_setup);
 
 	if (!dev_ifb)
 		return -ENOMEM;
#3 lore raw

PATCH 3/5 ifb: fix tx_queue_len overlimit

In reply to Changli Gao

We should check the length of rq after enqueuing, otherwise the length of rq will be over tx_queue_len.

---
 drivers/net/ifb.c |    6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/net/ifb.c +3 -3
index 1628d01..57c5cfb 100644
@@ -131,15 +131,15 @@ static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
 		return NETDEV_TX_OK;
 	}
 
-	if (skb_queue_len(&dp->rq) >= dev->tx_queue_len)
-		netif_stop_queue(dev);
-
 	__skb_queue_tail(&dp->rq, skb);
 	if (!dp->tasklet_pending) {
 		dp->tasklet_pending = 1;
 		tasklet_schedule(&dp->ifb_tasklet);
 	}
 
+	if (skb_queue_len(&dp->rq) >= dev->tx_queue_len)
+		netif_stop_queue(dev);
+
 	return NETDEV_TX_OK;
 }
 
#4 lore raw

Re: [PATCH 3/5] ifb: fix tx_queue_len overlimit

Quoted 4 lines · Changli Gao

We should check the length of rq after enqueuing, otherwise the length of rq will be over tx_queue_len.

Not sure about this one. The check is also for ==, so we are going to stop before we go over tx_queue_len. What am i missing?

cheers,
jamal

#5 lore raw

Re: [PATCH 3/5] ifb: fix tx_queue_len overlimit

Quoted 9 lines · jamal
Quoted 4 lines · Changli Gao

We should check the length of rq after enqueuing, otherwise the length of rq will be over tx_queue_len.

Not sure about this one. The check is also for ==, so we are going to stop before we go over tx_queue_len. What am i missing?

But then, we put this skb into rq in any way, and the length of rq becomes tx_queue_len + 1.

Signature · 2 lines
Regards,
Changli Gao([email protected])
#6 lore raw

Re: [PATCH 3/5] ifb: fix tx_queue_len overlimit

Quoted 3 lines · Changli Gao

But then, we put this skb into rq in any way, and the length of rq becomes tx_queue_len + 1.

;->

cheers,
jamal

#7 lore raw

PATCH 4/5 ifb: add multiqueue support

In reply to Changli Gao

Each ifb NIC has nr_cpu_ids rx queues and nr_cpu_ids queues. Packets transmitted to ifb are enqueued to the corresponding per cpu tx queues, and processed in the corresponding per cpu tasklet latter.

The stats are converted to the u64 ones.

tq is a stack variable now. It makes ifb_q_private smaller and tx queue locked only once in ri_tasklet.

---
 drivers/net/ifb.c |  211 ++++++++++++++++++++++++++++++++++++------------------
 1 file changed, 141 insertions(+), 70 deletions(-)
drivers/net/ifb.c +141 -70
index 57c5cfb..16c767b 100644
@@ -37,56 +37,63 @@
 #include <net/net_namespace.h>
 
 #define TX_Q_LIMIT    32
+struct ifb_q_private {
+	struct tasklet_struct	ifb_tasklet;
+	struct sk_buff_head	rq;
+	struct u64_stats_sync	syncp;
+	u64			rx_packets;
+	u64			rx_bytes;
+	u64			rx_dropped;
+};
+
 struct ifb_private {
-	struct tasklet_struct   ifb_tasklet;
-	int     tasklet_pending;
-	struct sk_buff_head     rq;
-	struct sk_buff_head     tq;
+	struct ifb_q_private __percpu	*q;
 };
 
 static int numifbs = 2;
 
-static void ri_tasklet(unsigned long dev)
+static void ri_tasklet(unsigned long _dev)
 {
-
-	struct net_device *_dev = (struct net_device *)dev;
-	struct ifb_private *dp = netdev_priv(_dev);
-	struct net_device_stats *stats = &_dev->stats;
+	struct net_device *dev = (struct net_device *)_dev;
+	struct ifb_private *p = netdev_priv(dev);
+	struct ifb_q_private *qp;
 	struct netdev_queue *txq;
 	struct sk_buff *skb;
-
-	txq = netdev_get_tx_queue(_dev, 0);
-	skb = skb_peek(&dp->tq);
-	if (skb == NULL) {
-		if (__netif_tx_trylock(txq)) {
-			skb_queue_splice_tail_init(&dp->rq, &dp->tq);
-			__netif_tx_unlock(txq);
-		} else {
-			/* reschedule */
-			goto resched;
-		}
+	struct sk_buff_head tq;
+
+	__skb_queue_head_init(&tq);
+	txq = netdev_get_tx_queue(dev, raw_smp_processor_id());
+	qp = per_cpu_ptr(p->q, raw_smp_processor_id());
+	if (!__netif_tx_trylock(txq)) {
+		tasklet_schedule(&qp->ifb_tasklet);
+		return;
 	}
+	skb_queue_splice_tail_init(&qp->rq, &tq);
+	if (netif_tx_queue_stopped(txq))
+		netif_tx_wake_queue(txq);
+	__netif_tx_unlock(txq);
 
-	while ((skb = skb_dequeue(&dp->tq)) != NULL) {
+	while ((skb = __skb_dequeue(&tq)) != NULL) {
 		u32 from = G_TC_FROM(skb->tc_verd);
 
 		skb->tc_verd = 0;
 		skb->tc_verd = SET_TC_NCLS(skb->tc_verd);
-		stats->tx_packets++;
-		stats->tx_bytes += skb->len;
+		u64_stats_update_begin(&qp->syncp);
+		txq->tx_packets++;
+		txq->tx_bytes += skb->len;
 
 		rcu_read_lock();
 		skb->dev = dev_get_by_index_rcu(&init_net, skb->skb_iif);
 		if (!skb->dev) {
 			rcu_read_unlock();
+			txq->tx_dropped++;
+			u64_stats_update_end(&qp->syncp);
 			dev_kfree_skb(skb);
-			stats->tx_dropped++;
-			if (skb_queue_len(&dp->tq) != 0)
-				goto resched;
-			break;
+			continue;
 		}
 		rcu_read_unlock();
-		skb->skb_iif = _dev->ifindex;
+		u64_stats_update_end(&qp->syncp);
+		skb->skb_iif = dev->ifindex;
 
 		if (from & AT_EGRESS) {
 			dev_queue_xmit(skb);
@@ -96,48 +103,32 @@ static void ri_tasklet(unsigned long dev)
 		} else
 			BUG();
 	}
-
-	if (__netif_tx_trylock(txq)) {
-		skb = skb_peek(&dp->rq);
-		if (skb == NULL) {
-			dp->tasklet_pending = 0;
-			if (netif_queue_stopped(_dev))
-				netif_wake_queue(_dev);
-		} else {
-			__netif_tx_unlock(txq);
-			goto resched;
-		}
-		__netif_tx_unlock(txq);
-	} else {
-resched:
-		dp->tasklet_pending = 1;
-		tasklet_schedule(&dp->ifb_tasklet);
-	}
-
 }
 
 static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
 {
-	struct ifb_private *dp = netdev_priv(dev);
-	struct net_device_stats *stats = &dev->stats;
+	struct ifb_private *p = netdev_priv(dev);
+	struct ifb_q_private *qp = per_cpu_ptr(p->q,
+					       skb_get_queue_mapping(skb));
 	u32 from = G_TC_FROM(skb->tc_verd);
 
-	stats->rx_packets++;
-	stats->rx_bytes += skb->len;
+	u64_stats_update_begin(&qp->syncp);
+	qp->rx_packets++;
+	qp->rx_bytes += skb->len;
 
 	if (!(from & (AT_INGRESS|AT_EGRESS)) || !skb->skb_iif) {
+		qp->rx_dropped++;
+		u64_stats_update_end(&qp->syncp);
 		dev_kfree_skb(skb);
-		stats->rx_dropped++;
 		return NETDEV_TX_OK;
 	}
+	u64_stats_update_end(&qp->syncp);
 
-	__skb_queue_tail(&dp->rq, skb);
-	if (!dp->tasklet_pending) {
-		dp->tasklet_pending = 1;
-		tasklet_schedule(&dp->ifb_tasklet);
-	}
+	__skb_queue_tail(&qp->rq, skb);
+	if (skb_queue_len(&qp->rq) == 1)
+		tasklet_schedule(&qp->ifb_tasklet);
 
-	if (skb_queue_len(&dp->rq) >= dev->tx_queue_len)
+	if (skb_queue_len(&qp->rq) >= dev->tx_queue_len)
 		netif_stop_queue(dev);
 
 	return NETDEV_TX_OK;
@@ -145,33 +136,103 @@ static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
 
 static int ifb_close(struct net_device *dev)
 {
-	struct ifb_private *dp = netdev_priv(dev);
-
-	tasklet_kill(&dp->ifb_tasklet);
-	netif_stop_queue(dev);
-	__skb_queue_purge(&dp->rq);
-	__skb_queue_purge(&dp->tq);
+	struct ifb_private *p = netdev_priv(dev);
+	struct ifb_q_private *qp;
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		qp = per_cpu_ptr(p->q, cpu);
+		tasklet_kill(&qp->ifb_tasklet);
+		netif_tx_stop_queue(netdev_get_tx_queue(dev, cpu));
+		__skb_queue_purge(&qp->rq);
+	}
 
 	return 0;
 }
 
 static int ifb_open(struct net_device *dev)
 {
-	struct ifb_private *dp = netdev_priv(dev);
+	int cpu;
+
+	for_each_possible_cpu(cpu)
+		netif_tx_start_queue(netdev_get_tx_queue(dev, cpu));
+
+	return 0;
+}
+
+static int ifb_init(struct net_device *dev)
+{
+	struct ifb_private *p = netdev_priv(dev);
+	struct ifb_q_private *q;
+	int cpu;
 
-	tasklet_init(&dp->ifb_tasklet, ri_tasklet, (unsigned long)dev);
-	__skb_queue_head_init(&dp->rq);
-	__skb_queue_head_init(&dp->tq);
-	netif_start_queue(dev);
+	p->q = alloc_percpu(struct ifb_q_private);
+	if (!p->q)
+		return -ENOMEM;
+	for_each_possible_cpu(cpu) {
+		q = per_cpu_ptr(p->q, cpu);
+		tasklet_init(&q->ifb_tasklet, ri_tasklet, (unsigned long)dev);
+		__skb_queue_head_init(&q->rq);
+	}
 
 	return 0;
 }
 
+static void ifb_uninit(struct net_device *dev)
+{
+	struct ifb_private *p = netdev_priv(dev);
+
+	free_percpu(p->q);
+}
+
+static u16 ifb_select_queue(struct net_device *dev, struct sk_buff *skb)
+{
+	return smp_processor_id();
+}
+
+static struct rtnl_link_stats64 *ifb_get_stats64(struct net_device *dev,
+		struct rtnl_link_stats64 *stats)
+{
+	struct ifb_private *p = netdev_priv(dev);
+	struct ifb_q_private *q;
+	struct netdev_queue *txq;
+	int cpu;
+	u64 rx_packets, rx_bytes, rx_dropped;
+	u64 tx_packets, tx_bytes, tx_dropped;
+	unsigned int start;
+
+	for_each_possible_cpu(cpu) {
+		q = per_cpu_ptr(p->q, cpu);
+		txq = netdev_get_tx_queue(dev, cpu);
+		do {
+			start = u64_stats_fetch_begin_bh(&q->syncp);
+			rx_packets = q->rx_packets;
+			rx_bytes = q->rx_bytes;
+			rx_dropped = q->rx_dropped;
+			tx_packets = txq->tx_packets;
+			tx_bytes = txq->tx_bytes;
+			tx_dropped = txq->tx_dropped;
+		} while (u64_stats_fetch_retry_bh(&q->syncp, start));
+		stats->rx_packets += rx_packets;
+		stats->rx_bytes += rx_bytes;
+		stats->rx_dropped += rx_dropped;
+		stats->tx_packets += tx_packets;
+		stats->tx_bytes += tx_bytes;
+		stats->tx_dropped += tx_dropped;
+	}
+
+	return stats;
+}
+
 static const struct net_device_ops ifb_netdev_ops = {
+	.ndo_init		= ifb_init,
+	.ndo_uninit		= ifb_uninit,
 	.ndo_open		= ifb_open,
 	.ndo_stop		= ifb_close,
 	.ndo_start_xmit		= ifb_xmit,
 	.ndo_validate_addr	= eth_validate_addr,
+	.ndo_select_queue	= ifb_select_queue,
+	.ndo_get_stats64	= ifb_get_stats64,
 };
 
 static void ifb_setup(struct net_device *dev)
@@ -202,11 +263,21 @@ static int ifb_validate(struct nlattr *tb[], struct nlattr *data[])
 	return 0;
 }
 
+static int ifb_get_tx_queues(struct net *net, struct nlattr *tb[],
+			     unsigned int *tx_queues,
+			     unsigned int *real_tx_queues)
+{
+	*real_tx_queues = *tx_queues = nr_cpu_ids;
+
+	return 0;
+}
+
 static struct rtnl_link_ops ifb_link_ops __read_mostly = {
 	.kind		= "ifb",
 	.priv_size	= sizeof(struct ifb_private),
 	.setup		= ifb_setup,
 	.validate	= ifb_validate,
+	.get_tx_queues	= ifb_get_tx_queues,
 };
 
 /* Number of ifb devices to be set up by this module. */
@@ -218,8 +289,8 @@ static int __init ifb_init_one(int index)
 	struct net_device *dev_ifb;
 	int err;
 
-	dev_ifb = alloc_netdev(sizeof(struct ifb_private), "ifb%d", ifb_setup);
-
+	dev_ifb = alloc_netdev_mq(sizeof(struct ifb_private), "ifb%d",
+				  ifb_setup, nr_cpu_ids);
 	if (!dev_ifb)
 		return -ENOMEM;
 
#8 lore raw

Re: [PATCH 4/5] ifb: add multiqueue support

Quoted 9 lines · Changli Gao

Each ifb NIC has nr_cpu_ids rx queues and nr_cpu_ids queues. Packets transmitted to ifb are enqueued to the corresponding per cpu tx queues, and processed in the corresponding per cpu tasklet latter.

The stats are converted to the u64 ones.

tq is a stack variable now. It makes ifb_q_private smaller and tx queue locked only once in ri_tasklet.

Might be good to say the tx_queue_len is multiplied by number of online cpus ;)

Quoted 17 lines
---
 drivers/net/ifb.c |  211 ++++++++++++++++++++++++++++++++++++------------------
 1 file changed, 141 insertions(+), 70 deletions(-)
drivers/net/ifb.c +5 -0
index 57c5cfb..16c767b 100644
@@ -37,56 +37,63 @@
 #include <net/net_namespace.h>
 
 #define TX_Q_LIMIT    32
+struct ifb_q_private {
+	struct tasklet_struct	ifb_tasklet;
+	struct sk_buff_head	rq;
+	struct u64_stats_sync	syncp;
+	u64			rx_packets;

drivers/net/ifb.c line 50

Quoted 10 lines
drivers/net/ifb.c +5 -4
+	u64			rx_bytes;
+	u64			rx_dropped;
+};
+
 struct ifb_private {
-	struct tasklet_struct   ifb_tasklet;
-	int     tasklet_pending;
-	struct sk_buff_head     rq;
-	struct sk_buff_head     tq;
+	struct ifb_q_private __percpu	*q;

You probably could use dev->ml_priv (lstats/tstats/dstats) so that ifb_private just disapears (we save a full cache line)

Quoted 22 lines
drivers/net/ifb.c +4 -11
 };
 
 static int numifbs = 2;
 
-static void ri_tasklet(unsigned long dev)
+static void ri_tasklet(unsigned long _dev)
 {
-
-	struct net_device *_dev = (struct net_device *)dev;
-	struct ifb_private *dp = netdev_priv(_dev);
-	struct net_device_stats *stats = &_dev->stats;
+	struct net_device *dev = (struct net_device *)_dev;
+	struct ifb_private *p = netdev_priv(dev);
+	struct ifb_q_private *qp;
 	struct netdev_queue *txq;
 	struct sk_buff *skb;
-
-	txq = netdev_get_tx_queue(_dev, 0);
-	skb = skb_peek(&dp->tq);
-	if (skb == NULL) {
-		if (__netif_tx_trylock(txq)) {
-			skb_queue_splice_tail_init(&dp->rq, &dp->tq);

drivers/net/ifb.c line 66

Quoted 10 lines
drivers/net/ifb.c +5 -5
-			__netif_tx_unlock(txq);
-		} else {
-			/* reschedule */
-			goto resched;
-		}
+	struct sk_buff_head tq;
+
+	__skb_queue_head_init(&tq);
+	txq = netdev_get_tx_queue(dev, raw_smp_processor_id());
+	qp = per_cpu_ptr(p->q, raw_smp_processor_id());
	qp = this_cpu_ptr(dev->ifb_qp); is faster
Quoted 28 lines
drivers/net/ifb.c +13 -3
+	if (!__netif_tx_trylock(txq)) {
+		tasklet_schedule(&qp->ifb_tasklet);
+		return;
 	}
+	skb_queue_splice_tail_init(&qp->rq, &tq);
+	if (netif_tx_queue_stopped(txq))
+		netif_tx_wake_queue(txq);
+	__netif_tx_unlock(txq);
 
-	while ((skb = skb_dequeue(&dp->tq)) != NULL) {
+	while ((skb = __skb_dequeue(&tq)) != NULL) {
 		u32 from = G_TC_FROM(skb->tc_verd);
 
 		skb->tc_verd = 0;
 		skb->tc_verd = SET_TC_NCLS(skb->tc_verd);
-		stats->tx_packets++;
-		stats->tx_bytes += skb->len;
+		u64_stats_update_begin(&qp->syncp);
+		txq->tx_packets++;
+		txq->tx_bytes += skb->len;
 
 		rcu_read_lock();
 		skb->dev = dev_get_by_index_rcu(&init_net, skb->skb_iif);
 		if (!skb->dev) {
 			rcu_read_unlock();
+			txq->tx_dropped++;
+			u64_stats_update_end(&qp->syncp);
 			dev_kfree_skb(skb);

drivers/net/ifb.c line 96

Quoted 10 lines
drivers/net/ifb.c +3 -5
-			stats->tx_dropped++;
-			if (skb_queue_len(&dp->tq) != 0)
-				goto resched;
-			break;
+			continue;
 		}
 		rcu_read_unlock();
-		skb->skb_iif = _dev->ifindex;
+		u64_stats_update_end(&qp->syncp);
+		skb->skb_iif = dev->ifindex;

Why is this necessary ? shouldnt skb->skb_iif already be dev->ifindex ?

Quoted 24 lines
drivers/net/ifb.c +0 -17
 		if (from & AT_EGRESS) {
 			dev_queue_xmit(skb);
@@ -96,48 +103,32 @@ static void ri_tasklet(unsigned long dev)
 		} else
 			BUG();
 	}
-
-	if (__netif_tx_trylock(txq)) {
-		skb = skb_peek(&dp->rq);
-		if (skb == NULL) {
-			dp->tasklet_pending = 0;
-			if (netif_queue_stopped(_dev))
-				netif_wake_queue(_dev);
-		} else {
-			__netif_tx_unlock(txq);
-			goto resched;
-		}
-		__netif_tx_unlock(txq);
-	} else {
-resched:
-		dp->tasklet_pending = 1;
-		tasklet_schedule(&dp->ifb_tasklet);
-	}

drivers/net/ifb.c static void ri_tasklet(unsigned long dev) · line 112

Quoted 10 lines
drivers/net/ifb.c +3 -3
-
 }
 
 static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
 {
-	struct ifb_private *dp = netdev_priv(dev);
-	struct net_device_stats *stats = &dev->stats;
+	struct ifb_private *p = netdev_priv(dev);
+	struct ifb_q_private *qp = per_cpu_ptr(p->q,
+					       skb_get_queue_mapping(skb));

Would be good to add a

WARN_ON(skb_get_queue_mapping(skb) != smp_processor_id());

Quoted 18 lines
drivers/net/ifb.c +6 -4
 	u32 from = G_TC_FROM(skb->tc_verd);
 
-	stats->rx_packets++;
-	stats->rx_bytes += skb->len;
+	u64_stats_update_begin(&qp->syncp);
+	qp->rx_packets++;
+	qp->rx_bytes += skb->len;
 
 	if (!(from & (AT_INGRESS|AT_EGRESS)) || !skb->skb_iif) {
+		qp->rx_dropped++;
+		u64_stats_update_end(&qp->syncp);
 		dev_kfree_skb(skb);
-		stats->rx_dropped++;
 		return NETDEV_TX_OK;
 	}
+	u64_stats_update_end(&qp->syncp);
 
-	__skb_queue_tail(&dp->rq, skb);

drivers/net/ifb.c static void ri_tasklet(unsigned long dev) · line 131

Quoted 10 lines
drivers/net/ifb.c +4 -5
-	if (!dp->tasklet_pending) {
-		dp->tasklet_pending = 1;
-		tasklet_schedule(&dp->ifb_tasklet);
-	}
+	__skb_queue_tail(&qp->rq, skb);
+	if (skb_queue_len(&qp->rq) == 1)
+		tasklet_schedule(&qp->ifb_tasklet);
 
-	if (skb_queue_len(&dp->rq) >= dev->tx_queue_len)
+	if (skb_queue_len(&qp->rq) >= dev->tx_queue_len)

This seems wrong...
You need to change to netif_tx_stop_queue(txq)

Quoted 44 lines
drivers/net/ifb.c +21 -7
 		netif_stop_queue(dev);
 
 	return NETDEV_TX_OK;
@@ -145,33 +136,103 @@ static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
 
 static int ifb_close(struct net_device *dev)
 {
-	struct ifb_private *dp = netdev_priv(dev);
-
-	tasklet_kill(&dp->ifb_tasklet);
-	netif_stop_queue(dev);
-	__skb_queue_purge(&dp->rq);
-	__skb_queue_purge(&dp->tq);
+	struct ifb_private *p = netdev_priv(dev);
+	struct ifb_q_private *qp;
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		qp = per_cpu_ptr(p->q, cpu);
+		tasklet_kill(&qp->ifb_tasklet);
+		netif_tx_stop_queue(netdev_get_tx_queue(dev, cpu));
+		__skb_queue_purge(&qp->rq);
+	}
 
 	return 0;
 }
 
 static int ifb_open(struct net_device *dev)
 {
-	struct ifb_private *dp = netdev_priv(dev);
+	int cpu;
+
 
 
 
+	for_each_possible_cpu(cpu)
+		netif_tx_start_queue(netdev_get_tx_queue(dev, cpu));
+
+	return 0;
+}
+
+static int ifb_init(struct net_device *dev)
+{
+	struct ifb_private *p = netdev_priv(dev);

drivers/net/ifb.c static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev) · line 174

Quoted 10 lines
drivers/net/ifb.c +5 -4
+	struct ifb_q_private *q;
+	int cpu;
 
-	tasklet_init(&dp->ifb_tasklet, ri_tasklet, (unsigned long)dev);
-	__skb_queue_head_init(&dp->rq);
-	__skb_queue_head_init(&dp->tq);
-	netif_start_queue(dev);
+	p->q = alloc_percpu(struct ifb_q_private);
+	if (!p->q)
+		return -ENOMEM;

Hmm, maybe use netdev_queue_numa_node_write() somewhere, so that qdisc_alloc() can use NUMA affinities.

Quoted 100 lines
drivers/net/ifb.c +55 -0
+	for_each_possible_cpu(cpu) {
+		q = per_cpu_ptr(p->q, cpu);
+		tasklet_init(&q->ifb_tasklet, ri_tasklet, (unsigned long)dev);
+		__skb_queue_head_init(&q->rq);
+	}
 
 	return 0;
 }
 
+static void ifb_uninit(struct net_device *dev)
+{
+	struct ifb_private *p = netdev_priv(dev);
+
+	free_percpu(p->q);
+}
+
+static u16 ifb_select_queue(struct net_device *dev, struct sk_buff *skb)
+{
+	return smp_processor_id();
+}
+
+static struct rtnl_link_stats64 *ifb_get_stats64(struct net_device *dev,
+		struct rtnl_link_stats64 *stats)
+{
+	struct ifb_private *p = netdev_priv(dev);
+	struct ifb_q_private *q;
+	struct netdev_queue *txq;
+	int cpu;
+	u64 rx_packets, rx_bytes, rx_dropped;
+	u64 tx_packets, tx_bytes, tx_dropped;
+	unsigned int start;
+
+	for_each_possible_cpu(cpu) {
+		q = per_cpu_ptr(p->q, cpu);
+		txq = netdev_get_tx_queue(dev, cpu);
+		do {
+			start = u64_stats_fetch_begin_bh(&q->syncp);
+			rx_packets = q->rx_packets;
+			rx_bytes = q->rx_bytes;
+			rx_dropped = q->rx_dropped;
+			tx_packets = txq->tx_packets;
+			tx_bytes = txq->tx_bytes;
+			tx_dropped = txq->tx_dropped;
+		} while (u64_stats_fetch_retry_bh(&q->syncp, start));
+		stats->rx_packets += rx_packets;
+		stats->rx_bytes += rx_bytes;
+		stats->rx_dropped += rx_dropped;
+		stats->tx_packets += tx_packets;
+		stats->tx_bytes += tx_bytes;
+		stats->tx_dropped += tx_dropped;
+	}
+
+	return stats;
+}
+
 static const struct net_device_ops ifb_netdev_ops = {
+	.ndo_init		= ifb_init,
+	.ndo_uninit		= ifb_uninit,
 	.ndo_open		= ifb_open,
 	.ndo_stop		= ifb_close,
 	.ndo_start_xmit		= ifb_xmit,
 	.ndo_validate_addr	= eth_validate_addr,
+	.ndo_select_queue	= ifb_select_queue,
+	.ndo_get_stats64	= ifb_get_stats64,

};

static void ifb_setup(struct net_device *dev)

drivers/net/ifb.c +12 -2
@@ -202,11 +263,21 @@ static int ifb_validate(struct nlattr *tb[], struct nlattr *data[])
 	return 0;
 }
 
+static int ifb_get_tx_queues(struct net *net, struct nlattr *tb[],
+			     unsigned int *tx_queues,
+			     unsigned int *real_tx_queues)
+{
+	*real_tx_queues = *tx_queues = nr_cpu_ids;
+
+	return 0;
+}
+
 static struct rtnl_link_ops ifb_link_ops __read_mostly = {
 	.kind		= "ifb",
 	.priv_size	= sizeof(struct ifb_private),
 	.setup		= ifb_setup,
 	.validate	= ifb_validate,
+	.get_tx_queues	= ifb_get_tx_queues,
 };
 
 /* Number of ifb devices to be set up by this module. */
@@ -218,8 +289,8 @@ static int __init ifb_init_one(int index)
 	struct net_device *dev_ifb;
 	int err;
 
-	dev_ifb = alloc_netdev(sizeof(struct ifb_private), "ifb%d", ifb_setup);
-
+	dev_ifb = alloc_netdev_mq(sizeof(struct ifb_private), "ifb%d",
+				  ifb_setup, nr_cpu_ids);
 	if (!dev_ifb)
 		return -ENOMEM;
 
#9 lore raw

Re: [PATCH 4/5] ifb: add multiqueue support

Quoted 5 lines · Eric Dumazet
Quoted 4 lines · Changli Gao

Each ifb NIC has nr_cpu_ids rx queues and nr_cpu_ids queues. Packets transmitted to ifb are enqueued to the corresponding per cpu tx queues, and processed in the corresponding per cpu tasklet latter.

Quoted 8 lines
Quoted 5 lines

The stats are converted to the u64 ones.

tq is a stack variable now. It makes ifb_q_private smaller and tx queue locked only once in ri_tasklet.

Might be good to say the tx_queue_len is multiplied by number of online cpus ;)

Thanks for completion.

Quoted 23 lines
Quoted 23 lines

---
 drivers/net/ifb.c |  211 ++++++++++++++++++++++++++++++++++++------------------

 1 file changed, 141 insertions(+), 70 deletions(-)
drivers/net/ifb.c +9 -0
index 57c5cfb..16c767b 100644
@@ -37,56 +37,63 @@
  #include <net/net_namespace.h>
 
  #define TX_Q_LIMIT    32
+struct ifb_q_private {
+     struct tasklet_struct   ifb_tasklet;
+     struct sk_buff_head     rq;
+     struct u64_stats_sync   syncp;
+     u64                     rx_packets;
+     u64                     rx_bytes;
+     u64                     rx_dropped;
+};
+
  struct ifb_private {
Quoted 8 lines

drivers/net/ifb.c line 50

Quoted 5 lines
drivers/net/ifb.c +1 -4
-     struct tasklet_struct   ifb_tasklet;
-     int     tasklet_pending;
-     struct sk_buff_head     rq;
-     struct sk_buff_head     tq;
+     struct ifb_q_private __percpu   *q;

You probably could use dev->ml_priv (lstats/tstats/dstats) so that ifb_private just disapears (we save a full cache line)

Good idea. Thanks.

Quoted 27 lines
Quoted 27 lines
drivers/net/ifb.c +4 -15
  };
 
  static int numifbs = 2;
 
-static void ri_tasklet(unsigned long dev)
+static void ri_tasklet(unsigned long _dev)
  {
-
-     struct net_device *_dev = (struct net_device *)dev;
-     struct ifb_private *dp = netdev_priv(_dev);
-     struct net_device_stats *stats = &_dev->stats;
+     struct net_device *dev = (struct net_device *)_dev;
+     struct ifb_private *p = netdev_priv(dev);
+     struct ifb_q_private *qp;
       struct netdev_queue *txq;
       struct sk_buff *skb;
-
-     txq = netdev_get_tx_queue(_dev, 0);
-     skb = skb_peek(&dp->tq);
-     if (skb == NULL) {
-             if (__netif_tx_trylock(txq)) {
-                     skb_queue_splice_tail_init(&dp->rq, &dp->tq);
-                     __netif_tx_unlock(txq);
-             } else {
-                     /* reschedule */
-                     goto resched;
Quoted 8 lines

drivers/net/ifb.c line 66

Quoted 6 lines
drivers/net/ifb.c +5 -1
-             }
+     struct sk_buff_head tq;
+
+     __skb_queue_head_init(&tq);
+     txq = netdev_get_tx_queue(dev, raw_smp_processor_id());
+     qp = per_cpu_ptr(p->q, raw_smp_processor_id());

       qp = this_cpu_ptr(dev->ifb_qp); is faster

and less, thanks. :)

Quoted 40 lines
Quoted 29 lines
drivers/net/ifb.c +13 -3
+     if (!__netif_tx_trylock(txq)) {
+             tasklet_schedule(&qp->ifb_tasklet);
+             return;
       }
+     skb_queue_splice_tail_init(&qp->rq, &tq);
+     if (netif_tx_queue_stopped(txq))
+             netif_tx_wake_queue(txq);
+     __netif_tx_unlock(txq);
 
-     while ((skb = skb_dequeue(&dp->tq)) != NULL) {
+     while ((skb = __skb_dequeue(&tq)) != NULL) {
               u32 from = G_TC_FROM(skb->tc_verd);
 
               skb->tc_verd = 0;
               skb->tc_verd = SET_TC_NCLS(skb->tc_verd);
-             stats->tx_packets++;
-             stats->tx_bytes += skb->len;
+             u64_stats_update_begin(&qp->syncp);
+             txq->tx_packets++;
+             txq->tx_bytes += skb->len;
 
               rcu_read_lock();
               skb->dev = dev_get_by_index_rcu(&init_net, skb->skb_iif);
               if (!skb->dev) {
                       rcu_read_unlock();
+                     txq->tx_dropped++;
+                     u64_stats_update_end(&qp->syncp);
                       dev_kfree_skb(skb);

drivers/net/ifb.c line 96

Quoted 10 lines
drivers/net/ifb.c +3 -5
-                     stats->tx_dropped++;
-                     if (skb_queue_len(&dp->tq) != 0)
-                             goto resched;
-                     break;
+                     continue;
               }
               rcu_read_unlock();
-             skb->skb_iif = _dev->ifindex;
+             u64_stats_update_end(&qp->syncp);
+             skb->skb_iif = dev->ifindex;
Quoted 1 line

Why is this necessary ? shouldnt skb->skb_iif already be dev->ifindex ?

No. In fact, act_mirred use skb_iff to save the original dev.

        skb2->skb_iif = skb->dev->ifindex;
        skb2->dev = dev;
Quoted 38 lines
Quoted 25 lines
drivers/net/ifb.c +0 -17
               if (from & AT_EGRESS) {
                       dev_queue_xmit(skb);
@@ -96,48 +103,32 @@ static void ri_tasklet(unsigned long dev)
               } else
                       BUG();
       }
-
-     if (__netif_tx_trylock(txq)) {
-             skb = skb_peek(&dp->rq);
-             if (skb == NULL) {
-                     dp->tasklet_pending = 0;
-                     if (netif_queue_stopped(_dev))
-                             netif_wake_queue(_dev);
-             } else {
-                     __netif_tx_unlock(txq);
-                     goto resched;
-             }
-             __netif_tx_unlock(txq);
-     } else {
-resched:
-             dp->tasklet_pending = 1;
-             tasklet_schedule(&dp->ifb_tasklet);
-     }

drivers/net/ifb.c static void ri_tasklet(unsigned long dev) · line 112

Quoted 10 lines
drivers/net/ifb.c +3 -3
-
  }
 
  static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
  {
-     struct ifb_private *dp = netdev_priv(dev);
-     struct net_device_stats *stats = &dev->stats;
+     struct ifb_private *p = netdev_priv(dev);
+     struct ifb_q_private *qp = per_cpu_ptr(p->q,
+                                            skb_get_queue_mapping(skb));

Would be good to add a

Quoted 1 line

WARN_ON(skb_get_queue_mapping(skb) != smp_processor_id());

I think it maybe what Jamal wants too. Thanks.

Quoted 28 lines
Quoted 28 lines
drivers/net/ifb.c +9 -8
       u32 from = G_TC_FROM(skb->tc_verd);
 
-     stats->rx_packets++;
-     stats->rx_bytes += skb->len;
+     u64_stats_update_begin(&qp->syncp);
+     qp->rx_packets++;
+     qp->rx_bytes += skb->len;
 
       if (!(from & (AT_INGRESS|AT_EGRESS)) || !skb->skb_iif) {
+             qp->rx_dropped++;
+             u64_stats_update_end(&qp->syncp);
               dev_kfree_skb(skb);
-             stats->rx_dropped++;
               return NETDEV_TX_OK;
       }
+     u64_stats_update_end(&qp->syncp);
 
-     __skb_queue_tail(&dp->rq, skb);
-     if (!dp->tasklet_pending) {
-             dp->tasklet_pending = 1;
-             tasklet_schedule(&dp->ifb_tasklet);
-     }
+     __skb_queue_tail(&qp->rq, skb);
+     if (skb_queue_len(&qp->rq) == 1)
+             tasklet_schedule(&qp->ifb_tasklet);
 
Quoted 5 lines

drivers/net/ifb.c static void ri_tasklet(unsigned long dev) · line 131

Quoted 2 lines
drivers/net/ifb.c +1 -1
-     if (skb_queue_len(&dp->rq) >= dev->tx_queue_len)
+     if (skb_queue_len(&qp->rq) >= dev->tx_queue_len)

This seems wrong...
You need to change to netif_tx_stop_queue(txq)

Thanks for catching this. I missed it.

Quoted 48 lines
Quoted 48 lines
drivers/net/ifb.c +23 -7
               netif_stop_queue(dev);
 
       return NETDEV_TX_OK;
@@ -145,33 +136,103 @@ static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
 
  static int ifb_close(struct net_device *dev)
  {
-     struct ifb_private *dp = netdev_priv(dev);
-
-     tasklet_kill(&dp->ifb_tasklet);
-     netif_stop_queue(dev);
-     __skb_queue_purge(&dp->rq);
-     __skb_queue_purge(&dp->tq);
+     struct ifb_private *p = netdev_priv(dev);
+     struct ifb_q_private *qp;
+     int cpu;
+
+     for_each_possible_cpu(cpu) {
+             qp = per_cpu_ptr(p->q, cpu);
+             tasklet_kill(&qp->ifb_tasklet);
+             netif_tx_stop_queue(netdev_get_tx_queue(dev, cpu));
+             __skb_queue_purge(&qp->rq);
+     }
 
       return 0;
  }
 
  static int ifb_open(struct net_device *dev)
  {
-     struct ifb_private *dp = netdev_priv(dev);
+     int cpu;
+
 
 
 
+     for_each_possible_cpu(cpu)
+             netif_tx_start_queue(netdev_get_tx_queue(dev, cpu));
+
+     return 0;
+}
+
+static int ifb_init(struct net_device *dev)
+{
+     struct ifb_private *p = netdev_priv(dev);
+     struct ifb_q_private *q;
+     int cpu;
 
Quoted 10 lines

drivers/net/ifb.c static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev) · line 171

Quoted 7 lines
drivers/net/ifb.c +3 -4
-     tasklet_init(&dp->ifb_tasklet, ri_tasklet, (unsigned long)dev);
-     __skb_queue_head_init(&dp->rq);
-     __skb_queue_head_init(&dp->tq);
-     netif_start_queue(dev);
+     p->q = alloc_percpu(struct ifb_q_private);
+     if (!p->q)
+             return -ENOMEM;

Hmm, maybe  use netdev_queue_numa_node_write() somewhere, so that qdisc_alloc() can use NUMA affinities.

I'll add it, thanks.

Signature · 2 lines
Regards,
Changli Gao([email protected])
#10 lore raw

Re: [PATCH 4/5] ifb: add multiqueue support

In reply to Changli Gao
Quoted 58 lines · Changli Gao

Each ifb NIC has nr_cpu_ids rx queues and nr_cpu_ids queues. Packets transmitted to ifb are enqueued to the corresponding per cpu tx queues, and processed in the corresponding per cpu tasklet latter.

The stats are converted to the u64 ones.

tq is a stack variable now. It makes ifb_q_private smaller and tx queue locked only once in ri_tasklet.

---
 drivers/net/ifb.c |  211 ++++++++++++++++++++++++++++++++++++------------------
 1 file changed, 141 insertions(+), 70 deletions(-)
drivers/net/ifb.c +14 -15
index 57c5cfb..16c767b 100644
@@ -37,56 +37,63 @@
 #include <net/net_namespace.h>
 
 #define TX_Q_LIMIT    32
+struct ifb_q_private {
+	struct tasklet_struct	ifb_tasklet;
+	struct sk_buff_head	rq;
+	struct u64_stats_sync	syncp;
+	u64			rx_packets;
+	u64			rx_bytes;
+	u64			rx_dropped;
+};
+
 struct ifb_private {
-	struct tasklet_struct   ifb_tasklet;
-	int     tasklet_pending;
-	struct sk_buff_head     rq;
-	struct sk_buff_head     tq;
+	struct ifb_q_private __percpu	*q;
 };
 
 static int numifbs = 2;
 
-static void ri_tasklet(unsigned long dev)
+static void ri_tasklet(unsigned long _dev)
 {
-
-	struct net_device *_dev = (struct net_device *)dev;
-	struct ifb_private *dp = netdev_priv(_dev);
-	struct net_device_stats *stats = &_dev->stats;
+	struct net_device *dev = (struct net_device *)_dev;
+	struct ifb_private *p = netdev_priv(dev);
+	struct ifb_q_private *qp;
 	struct netdev_queue *txq;
 	struct sk_buff *skb;
-
-	txq = netdev_get_tx_queue(_dev, 0);
-	skb = skb_peek(&dp->tq);
-	if (skb == NULL) {
-		if (__netif_tx_trylock(txq)) {
-			skb_queue_splice_tail_init(&dp->rq, &dp->tq);

drivers/net/ifb.c line 66

Quoted 10 lines
drivers/net/ifb.c +5 -5
-			__netif_tx_unlock(txq);
-		} else {
-			/* reschedule */
-			goto resched;
-		}
+	struct sk_buff_head tq;
+
+	__skb_queue_head_init(&tq);
+	txq = netdev_get_tx_queue(dev, raw_smp_processor_id());
+	qp = per_cpu_ptr(p->q, raw_smp_processor_id());

Hmm, this wont work with CPU HOTPLUG.

When we put a cpu offline, we can transfert tasklets from this cpu to another 'online cpu'

To solve this, you need that ri_tasklet() not use a "device pointer" parameter but a pointer to 'cpu' private data, since it can be different than the data of the current cpu.

static void ri_tasklet(unsigned long arg)
{
struct ifb_q_private *qp = (struct ifb_q_private *)arg;

	...
}
#11 lore raw

Re: [PATCH 4/5] ifb: add multiqueue support

Quoted 11 lines · Eric Dumazet

Hmm, this wont work with CPU HOTPLUG.

When we put a cpu offline, we can transfert tasklets from this cpu to another 'online cpu'

To solve this, you need that ri_tasklet() not use a "device pointer" parameter but a pointer to 'cpu' private data, since it can be different than the data of the current cpu.

Quoted 7 lines

static void ri_tasklet(unsigned long arg)
{
struct ifb_q_private *qp = (struct ifb_q_private *)arg;

       ...
}

You are right, and it was my original way. Thanks.

Signature · 2 lines
Regards,
Changli Gao([email protected])
#12 lore raw

Re: [PATCH 4/5] ifb: add multiqueue support

In reply to Changli Gao
Quoted 10 lines · Changli Gao

Each ifb NIC has nr_cpu_ids rx queues and nr_cpu_ids queues. Packets transmitted to ifb are enqueued to the corresponding per cpu tx queues, and processed in the corresponding per cpu tasklet latter.

The stats are converted to the u64 ones.

tq is a stack variable now. It makes ifb_q_private smaller and tx queue locked only once in ri_tasklet.

I will look closely at this one later and catchup with the discussion you had with Eric.

cheers,
jamal

#13 lore raw

PATCH 5/5 net: add skb.old_queue_mapping

In reply to Changli Gao

For the skbs returned from ifb, we should use the queue_mapping saved before ifb.

We save old queue_mapping in old_queue_mapping just before calling dev_queue_xmit, and restore the old_queue_mapping to queue_mapping just before reinjecting the skb.

dev_pick_tx() use the current queue_mapping for the skbs reinjected by ifb.

---
 drivers/net/ifb.c      |    1 +
 include/linux/skbuff.h |    3 +++
 net/core/dev.c         |    5 +++++
 net/sched/act_mirred.c |    1 +
 4 files changed, 10 insertions(+)
drivers/net/ifb.c +1 -0
index 16c767b..481e4b1 100644
@@ -95,6 +95,7 @@ static void ri_tasklet(unsigned long _dev)
 		u64_stats_update_end(&qp->syncp);
 		skb->skb_iif = dev->ifindex;
 
+		skb->queue_mapping = skb->old_queue_mapping;
 		if (from & AT_EGRESS) {
 			dev_queue_xmit(skb);
 		} else if (from & AT_INGRESS) {
include/linux/skbuff.h +3 -0
index 19f37a6..2ce2a96 100644
@@ -403,6 +403,9 @@ struct sk_buff {
 	};
 
 	__u16			vlan_tci;
+#ifdef CONFIG_NET_CLS_ACT
+	__u16			old_queue_mapping;
+#endif
 
 	sk_buff_data_t		transport_header;
 	sk_buff_data_t		network_header;
net/core/dev.c +5 -0
index d28b3a0..8e97cfd 100644
@@ -2190,6 +2190,11 @@ static struct netdev_queue *dev_pick_tx(struct net_device *dev,
 	int queue_index;
 	const struct net_device_ops *ops = dev->netdev_ops;
 
+#ifdef CONFIG_NET_CLS_ACT
+	if (skb->tc_verd & TC_NCLS)
+		queue_index = skb_get_queue_mapping(skb);
+	else
+#endif
 	if (dev->real_num_tx_queues == 1)
 		queue_index = 0;
 	else if (ops->ndo_select_queue) {
net/sched/act_mirred.c +1 -0
index 0c311be..853eb30 100644
@@ -197,6 +197,7 @@ static int tcf_mirred(struct sk_buff *skb, struct tc_action *a,
 
 	skb2->skb_iif = skb->dev->ifindex;
 	skb2->dev = dev;
+	skb2->old_queue_mapping = skb->queue_mapping;
 	dev_queue_xmit(skb2);
 	err = 0;
 
#14 lore raw

Re: [PATCH 5/5] net: add skb.old_queue_mapping

Quoted 34 lines · Changli Gao

For the skbs returned from ifb, we should use the queue_mapping saved before ifb.

We save old queue_mapping in old_queue_mapping just before calling dev_queue_xmit, and restore the old_queue_mapping to queue_mapping just before reinjecting the skb.

dev_pick_tx() use the current queue_mapping for the skbs reinjected by ifb.

---
 drivers/net/ifb.c      |    1 +
 include/linux/skbuff.h |    3 +++
 net/core/dev.c         |    5 +++++
 net/sched/act_mirred.c |    1 +
 4 files changed, 10 insertions(+)
drivers/net/ifb.c +1 -0
index 16c767b..481e4b1 100644
@@ -95,6 +95,7 @@ static void ri_tasklet(unsigned long _dev)
 		u64_stats_update_end(&qp->syncp);
 		skb->skb_iif = dev->ifindex;
 
+		skb->queue_mapping = skb->old_queue_mapping;
 		if (from & AT_EGRESS) {
 			dev_queue_xmit(skb);
 		} else if (from & AT_INGRESS) {
include/linux/skbuff.h +0 -0
index 19f37a6..2ce2a96 100644

include/linux/skbuff.h struct sk_buff { · line 409

Quoted 7 lines
include/linux/skbuff.h +3 -0
@@ -403,6 +403,9 @@ struct sk_buff {
 	};
 
 	__u16			vlan_tci;
+#ifdef CONFIG_NET_CLS_ACT
+	__u16			old_queue_mapping;
+#endif
 

Are you sure we need this field here ? Why not using cb[] for example ?

Quoted 30 lines
include/linux/skbuff.h +0 -0
 	sk_buff_data_t		transport_header;
 	sk_buff_data_t		network_header;
net/core/dev.c +5 -0
index d28b3a0..8e97cfd 100644
@@ -2190,6 +2190,11 @@ static struct netdev_queue *dev_pick_tx(struct net_device *dev,
 	int queue_index;
 	const struct net_device_ops *ops = dev->netdev_ops;
 
+#ifdef CONFIG_NET_CLS_ACT
+	if (skb->tc_verd & TC_NCLS)
+		queue_index = skb_get_queue_mapping(skb);
+	else
+#endif
 	if (dev->real_num_tx_queues == 1)
 		queue_index = 0;
 	else if (ops->ndo_select_queue) {
net/sched/act_mirred.c +1 -0
index 0c311be..853eb30 100644
@@ -197,6 +197,7 @@ static int tcf_mirred(struct sk_buff *skb, struct tc_action *a,
 
 	skb2->skb_iif = skb->dev->ifindex;
 	skb2->dev = dev;
+	skb2->old_queue_mapping = skb->queue_mapping;
 	dev_queue_xmit(skb2);
 	err = 0;
 
maintainer
#15 lore raw

Re: [PATCH 5/5] net: add skb.old_queue_mapping

From: Eric Dumazet <[email protected]>
Date: Mon, 13 Dec 2010 17:56:52 +0100

Quoted 8 lines · Changli Gao
Quoted 7 lines
include/linux/skbuff.h +0 -0
index 19f37a6..2ce2a96 100644
@@ -403,6 +403,9 @@ struct sk_buff {
 	};
 
Quoted 7 lines

include/linux/skbuff.h struct sk_buff { · line 409

Quoted 5 lines
include/linux/skbuff.h +3 -0
 	__u16			vlan_tci;
+#ifdef CONFIG_NET_CLS_ACT
+	__u16			old_queue_mapping;
+#endif
 

Are you sure we need this field here ? Why not using cb[] for example ?

Agreed, we should be removing sk_buff members not adding new ones.

We should especially not be adding new members to this critical structure for more obscure facilities like ifb.

#16 lore raw

Re: [PATCH 5/5] net: add skb.old_queue_mapping

In reply to Eric Dumazet
Quoted 37 lines · Eric Dumazet
Quoted 36 lines · Changli Gao

For the skbs returned from ifb, we should use the queue_mapping saved before ifb.

We save old queue_mapping in old_queue_mapping just before calling dev_queue_xmit, and restore the old_queue_mapping to queue_mapping just before reinjecting the skb.

dev_pick_tx() use the current queue_mapping for the skbs reinjected by ifb.

---
 drivers/net/ifb.c      |    1 +
 include/linux/skbuff.h |    3 +++
 net/core/dev.c         |    5 +++++
 net/sched/act_mirred.c |    1 +

 4 files changed, 10 insertions(+)
drivers/net/ifb.c +1 -0
index 16c767b..481e4b1 100644
@@ -95,6 +95,7 @@ static void ri_tasklet(unsigned long _dev)
               u64_stats_update_end(&qp->syncp);
               skb->skb_iif = dev->ifindex;
 
+             skb->queue_mapping = skb->old_queue_mapping;
               if (from & AT_EGRESS) {
                       dev_queue_xmit(skb);
               } else if (from & AT_INGRESS) {
include/linux/skbuff.h +0 -0
index 19f37a6..2ce2a96 100644
@@ -403,6 +403,9 @@ struct sk_buff {
       };
 
Quoted 8 lines

include/linux/skbuff.h struct sk_buff { · line 409

Quoted 5 lines
include/linux/skbuff.h +3 -0
       __u16                   vlan_tci;
+#ifdef CONFIG_NET_CLS_ACT
+     __u16                   old_queue_mapping;
+#endif
 

Are you sure we need this field here ? Why not using cb[] for example ?

skb->cb can only be used in the same layer, so we can't use it to save a value crossing layers. A quick grep finds sch_netem just uses skb->cb. Maybe we can put all the skb->cb in the qdisc layer and old_queue_mapping in a structure as a new skb->cb. Is it OK? Thanks.

Signature · 2 lines
Regards,
Changli Gao([email protected])
#17 lore raw

Re: [PATCH 5/5] net: add skb.old_queue_mapping

Quoted 6 lines · Changli Gao

skb->cb can only be used in the same layer, so we can't use it to save a value crossing layers. A quick grep finds sch_netem just uses skb->cb. Maybe we can put all the skb->cb in the qdisc layer and old_queue_mapping in a structure as a new skb->cb. Is it OK? Thanks.

I can't do that. Below the qdisc layer, there is another user of skb->cb GSO. Though we can set the GSO features of ifb to skip gso_segments(), it is too tricky and weak.

Signature · 2 lines
Regards,
Changli Gao([email protected])
#18 lore raw

Re: [PATCH 5/5] net: add skb.old_queue_mapping

In reply to Changli Gao
Quoted 9 lines · Changli Gao

For the skbs returned from ifb, we should use the queue_mapping saved before ifb.

We save old queue_mapping in old_queue_mapping just before calling dev_queue_xmit, and restore the old_queue_mapping to queue_mapping just before reinjecting the skb.

dev_pick_tx() use the current queue_mapping for the skbs reinjected by ifb.

You are hard-coding policy here, no? ifb can do a lot of funky things which change the nature of the flow. I can shape, i can edit packets etc. Why is the queue mapping any different?

cheers,
jamal

#19 lore raw

Re: [PATCH 5/5] net: add skb.old_queue_mapping

Quoted 9 lines · jamal
Quoted 3 lines · Changli Gao

dev_pick_tx() use the current queue_mapping for the skbs reinjected by ifb.

You are hard-coding policy here, no? ifb can do a lot of funky things which change the nature of the flow. I can shape, i can edit packets etc. Why is the queue mapping any different?

It seems reasonable. Thanks. I'll remove this hard-coding policy.

Signature · 2 lines
Regards,
Changli Gao([email protected])