3 The purpose of this driver is to provide a device that allows
4 for sharing of resources:
6 1) qdiscs/policies that are per device as opposed to system wide.
7 ifb allows for a device which can be redirected to thus providing
8 an impression of sharing.
10 2) Allows for queueing incoming traffic for shaping instead of
13 The original concept is based on what is known as the IMQ
14 driver initially written by Martin Devera, later rewritten
15 by Patrick McHardy and then maintained by Andre Correa.
17 You need the tc action mirror or redirect to feed this device
20 This program is free software; you can redistribute it and/or
21 modify it under the terms of the GNU General Public License
22 as published by the Free Software Foundation; either version
23 2 of the License, or (at your option) any later version.
25 Authors: Jamal Hadi Salim (2005)
30 #include <linux/module.h>
31 #include <linux/kernel.h>
32 #include <linux/netdevice.h>
33 #include <linux/etherdevice.h>
34 #include <linux/init.h>
35 #include <linux/interrupt.h>
36 #include <linux/moduleparam.h>
37 #include <net/pkt_sched.h>
38 #include <net/net_namespace.h>
41 struct ifb_q_private {
42 struct net_device *dev;
43 struct tasklet_struct ifb_tasklet;
46 struct sk_buff_head rq;
49 struct u64_stats_sync rsync;
51 struct u64_stats_sync tsync;
54 struct sk_buff_head tq;
55 } ____cacheline_aligned_in_smp;
57 struct ifb_dev_private {
58 struct ifb_q_private *tx_private;
61 static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev);
62 static int ifb_open(struct net_device *dev);
63 static int ifb_close(struct net_device *dev);
65 static void ifb_ri_tasklet(unsigned long _txp)
67 struct ifb_q_private *txp = (struct ifb_q_private *)_txp;
68 struct netdev_queue *txq;
71 txq = netdev_get_tx_queue(txp->dev, txp->txqnum);
72 skb = skb_peek(&txp->tq);
74 if (!__netif_tx_trylock(txq))
76 skb_queue_splice_tail_init(&txp->rq, &txp->tq);
77 __netif_tx_unlock(txq);
80 while ((skb = __skb_dequeue(&txp->tq)) != NULL) {
81 u32 from = G_TC_FROM(skb->tc_verd);
84 skb->tc_verd = SET_TC_NCLS(skb->tc_verd);
86 u64_stats_update_begin(&txp->tsync);
88 txp->tx_bytes += skb->len;
89 u64_stats_update_end(&txp->tsync);
92 skb->dev = dev_get_by_index_rcu(dev_net(txp->dev), skb->skb_iif);
96 txp->dev->stats.tx_dropped++;
97 if (skb_queue_len(&txp->tq) != 0)
102 skb->skb_iif = txp->dev->ifindex;
104 if (from & AT_EGRESS) {
106 } else if (from & AT_INGRESS) {
107 skb_pull(skb, skb->mac_len);
108 netif_receive_skb(skb);
113 if (__netif_tx_trylock(txq)) {
114 skb = skb_peek(&txp->rq);
116 txp->tasklet_pending = 0;
117 if (netif_tx_queue_stopped(txq))
118 netif_tx_wake_queue(txq);
120 __netif_tx_unlock(txq);
123 __netif_tx_unlock(txq);
126 txp->tasklet_pending = 1;
127 tasklet_schedule(&txp->ifb_tasklet);
132 static struct rtnl_link_stats64 *ifb_stats64(struct net_device *dev,
133 struct rtnl_link_stats64 *stats)
135 struct ifb_dev_private *dp = netdev_priv(dev);
136 struct ifb_q_private *txp = dp->tx_private;
141 for (i = 0; i < dev->num_tx_queues; i++,txp++) {
143 start = u64_stats_fetch_begin_irq(&txp->rsync);
144 packets = txp->rx_packets;
145 bytes = txp->rx_bytes;
146 } while (u64_stats_fetch_retry_irq(&txp->rsync, start));
147 stats->rx_packets += packets;
148 stats->rx_bytes += bytes;
151 start = u64_stats_fetch_begin_irq(&txp->tsync);
152 packets = txp->tx_packets;
153 bytes = txp->tx_bytes;
154 } while (u64_stats_fetch_retry_irq(&txp->tsync, start));
155 stats->tx_packets += packets;
156 stats->tx_bytes += bytes;
158 stats->rx_dropped = dev->stats.rx_dropped;
159 stats->tx_dropped = dev->stats.tx_dropped;
164 static int ifb_dev_init(struct net_device *dev)
166 struct ifb_dev_private *dp = netdev_priv(dev);
167 struct ifb_q_private *txp;
170 txp = kcalloc(dev->num_tx_queues, sizeof(*txp), GFP_KERNEL);
173 dp->tx_private = txp;
174 for (i = 0; i < dev->num_tx_queues; i++,txp++) {
177 __skb_queue_head_init(&txp->rq);
178 __skb_queue_head_init(&txp->tq);
179 u64_stats_init(&txp->rsync);
180 u64_stats_init(&txp->tsync);
181 tasklet_init(&txp->ifb_tasklet, ifb_ri_tasklet,
183 netif_tx_start_queue(netdev_get_tx_queue(dev, i));
188 static const struct net_device_ops ifb_netdev_ops = {
189 .ndo_open = ifb_open,
190 .ndo_stop = ifb_close,
191 .ndo_get_stats64 = ifb_stats64,
192 .ndo_start_xmit = ifb_xmit,
193 .ndo_validate_addr = eth_validate_addr,
194 .ndo_init = ifb_dev_init,
197 #define IFB_FEATURES (NETIF_F_HW_CSUM | NETIF_F_SG | NETIF_F_FRAGLIST | \
198 NETIF_F_TSO_ECN | NETIF_F_TSO | NETIF_F_TSO6 | \
199 NETIF_F_HIGHDMA | NETIF_F_HW_VLAN_CTAG_TX | \
200 NETIF_F_HW_VLAN_STAG_TX)
202 static void ifb_dev_free(struct net_device *dev)
204 struct ifb_dev_private *dp = netdev_priv(dev);
205 struct ifb_q_private *txp = dp->tx_private;
208 for (i = 0; i < dev->num_tx_queues; i++,txp++) {
209 tasklet_kill(&txp->ifb_tasklet);
210 __skb_queue_purge(&txp->rq);
211 __skb_queue_purge(&txp->tq);
213 kfree(dp->tx_private);
217 static void ifb_setup(struct net_device *dev)
219 /* Initialize the device structure. */
220 dev->netdev_ops = &ifb_netdev_ops;
222 /* Fill in device structure with ethernet-generic values. */
224 dev->tx_queue_len = TX_Q_LIMIT;
226 dev->features |= IFB_FEATURES;
227 dev->vlan_features |= IFB_FEATURES & ~(NETIF_F_HW_VLAN_CTAG_TX |
228 NETIF_F_HW_VLAN_STAG_TX);
230 dev->flags |= IFF_NOARP;
231 dev->flags &= ~IFF_MULTICAST;
232 dev->priv_flags &= ~IFF_TX_SKB_SHARING;
234 eth_hw_addr_random(dev);
235 dev->destructor = ifb_dev_free;
238 static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
240 struct ifb_dev_private *dp = netdev_priv(dev);
241 u32 from = G_TC_FROM(skb->tc_verd);
242 struct ifb_q_private *txp = dp->tx_private + skb_get_queue_mapping(skb);
244 u64_stats_update_begin(&txp->rsync);
246 txp->rx_bytes += skb->len;
247 u64_stats_update_end(&txp->rsync);
249 if (!(from & (AT_INGRESS|AT_EGRESS)) || !skb->skb_iif) {
251 dev->stats.rx_dropped++;
255 if (skb_queue_len(&txp->rq) >= dev->tx_queue_len)
256 netif_tx_stop_queue(netdev_get_tx_queue(dev, txp->txqnum));
258 __skb_queue_tail(&txp->rq, skb);
259 if (!txp->tasklet_pending) {
260 txp->tasklet_pending = 1;
261 tasklet_schedule(&txp->ifb_tasklet);
267 static int ifb_close(struct net_device *dev)
269 netif_tx_stop_all_queues(dev);
273 static int ifb_open(struct net_device *dev)
275 netif_tx_start_all_queues(dev);
279 static int ifb_validate(struct nlattr *tb[], struct nlattr *data[])
281 if (tb[IFLA_ADDRESS]) {
282 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
284 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
285 return -EADDRNOTAVAIL;
290 static struct rtnl_link_ops ifb_link_ops __read_mostly = {
292 .priv_size = sizeof(struct ifb_dev_private),
294 .validate = ifb_validate,
297 /* Number of ifb devices to be set up by this module.
298 * Note that these legacy devices have one queue.
299 * Prefer something like : ip link add ifb10 numtxqueues 8 type ifb
301 static int numifbs = 2;
302 module_param(numifbs, int, 0);
303 MODULE_PARM_DESC(numifbs, "Number of ifb devices");
305 static int __init ifb_init_one(int index)
307 struct net_device *dev_ifb;
310 dev_ifb = alloc_netdev(sizeof(struct ifb_dev_private), "ifb%d",
311 NET_NAME_UNKNOWN, ifb_setup);
316 dev_ifb->rtnl_link_ops = &ifb_link_ops;
317 err = register_netdevice(dev_ifb);
324 free_netdev(dev_ifb);
328 static int __init ifb_init_module(void)
333 err = __rtnl_link_register(&ifb_link_ops);
337 for (i = 0; i < numifbs && !err; i++) {
338 err = ifb_init_one(i);
342 __rtnl_link_unregister(&ifb_link_ops);
350 static void __exit ifb_cleanup_module(void)
352 rtnl_link_unregister(&ifb_link_ops);
355 module_init(ifb_init_module);
356 module_exit(ifb_cleanup_module);
357 MODULE_LICENSE("GPL");
358 MODULE_AUTHOR("Jamal Hadi Salim");
359 MODULE_ALIAS_RTNL_LINK("ifb");