1 /* net/sched/sch_teql.c "True" (or "trivial") link equalizer.
3 * This program is free software; you can redistribute it and/or
4 * modify it under the terms of the GNU General Public License
5 * as published by the Free Software Foundation; either version
6 * 2 of the License, or (at your option) any later version.
8 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
11 #include <linux/module.h>
12 #include <asm/uaccess.h>
13 #include <asm/system.h>
14 #include <linux/bitops.h>
15 #include <linux/types.h>
16 #include <linux/kernel.h>
17 #include <linux/sched.h>
18 #include <linux/string.h>
20 #include <linux/socket.h>
21 #include <linux/sockios.h>
23 #include <linux/errno.h>
24 #include <linux/interrupt.h>
25 #include <linux/if_arp.h>
26 #include <linux/if_ether.h>
27 #include <linux/inet.h>
28 #include <linux/netdevice.h>
29 #include <linux/etherdevice.h>
30 #include <linux/notifier.h>
31 #include <linux/init.h>
33 #include <net/route.h>
34 #include <linux/skbuff.h>
35 #include <linux/moduleparam.h>
37 #include <net/pkt_sched.h>
43 After loading this module you will find a new device teqlN
44 and new qdisc with the same name. To join a slave to the equalizer
45 you should just set this qdisc on a device f.e.
47 # tc qdisc add dev eth0 root teql0
48 # tc qdisc add dev eth1 root teql0
50 That's all. Full PnP 8)
55 1. Slave devices MUST be active devices, i.e., they must raise the tbusy
56 signal and generate EOI events. If you want to equalize virtual devices
57 like tunnels, use a normal eql device.
58 2. This device puts no limitations on physical slave characteristics
59 f.e. it will equalize 9600baud line and 100Mb ethernet perfectly :-)
60 Certainly, large difference in link speeds will make the resulting
61 eqalized link unusable, because of huge packet reordering.
62 I estimate an upper useful difference as ~10 times.
63 3. If the slave requires address resolution, only protocols using
64 neighbour cache (IPv4/IPv6) will work over the equalized link.
65 Other protocols are still allowed to use the slave device directly,
66 which will not break load balancing, though native slave
67 traffic will have the highest priority. */
71 struct Qdisc_ops qops;
72 struct net_device *dev;
74 struct list_head master_list;
75 struct net_device_stats stats;
78 struct teql_sched_data
81 struct teql_master *m;
82 struct neighbour *ncache;
83 struct sk_buff_head q;
86 #define NEXT_SLAVE(q) (((struct teql_sched_data*)qdisc_priv(q))->next)
88 #define FMASK (IFF_BROADCAST|IFF_POINTOPOINT|IFF_BROADCAST)
90 /* "teql*" qdisc routines */
93 teql_enqueue(struct sk_buff *skb, struct Qdisc* sch)
95 struct net_device *dev = sch->dev;
96 struct teql_sched_data *q = qdisc_priv(sch);
98 __skb_queue_tail(&q->q, skb);
99 if (q->q.qlen <= dev->tx_queue_len) {
100 sch->bstats.bytes += skb->len;
101 sch->bstats.packets++;
105 __skb_unlink(skb, &q->q);
108 return NET_XMIT_DROP;
112 teql_requeue(struct sk_buff *skb, struct Qdisc* sch)
114 struct teql_sched_data *q = qdisc_priv(sch);
116 __skb_queue_head(&q->q, skb);
117 sch->qstats.requeues++;
121 static struct sk_buff *
122 teql_dequeue(struct Qdisc* sch)
124 struct teql_sched_data *dat = qdisc_priv(sch);
127 skb = __skb_dequeue(&dat->q);
129 struct net_device *m = dat->m->dev->qdisc->dev;
131 dat->m->slaves = sch;
135 sch->q.qlen = dat->q.qlen + dat->m->dev->qdisc->q.qlen;
139 static __inline__ void
140 teql_neigh_release(struct neighbour *n)
147 teql_reset(struct Qdisc* sch)
149 struct teql_sched_data *dat = qdisc_priv(sch);
151 skb_queue_purge(&dat->q);
153 teql_neigh_release(xchg(&dat->ncache, NULL));
157 teql_destroy(struct Qdisc* sch)
159 struct Qdisc *q, *prev;
160 struct teql_sched_data *dat = qdisc_priv(sch);
161 struct teql_master *master = dat->m;
163 if ((prev = master->slaves) != NULL) {
165 q = NEXT_SLAVE(prev);
167 NEXT_SLAVE(prev) = NEXT_SLAVE(q);
168 if (q == master->slaves) {
169 master->slaves = NEXT_SLAVE(q);
170 if (q == master->slaves) {
171 master->slaves = NULL;
172 spin_lock_bh(&master->dev->queue_lock);
173 qdisc_reset(master->dev->qdisc);
174 spin_unlock_bh(&master->dev->queue_lock);
177 skb_queue_purge(&dat->q);
178 teql_neigh_release(xchg(&dat->ncache, NULL));
182 } while ((prev = q) != master->slaves);
186 static int teql_qdisc_init(struct Qdisc *sch, struct rtattr *opt)
188 struct net_device *dev = sch->dev;
189 struct teql_master *m = (struct teql_master*)sch->ops;
190 struct teql_sched_data *q = qdisc_priv(sch);
192 if (dev->hard_header_len > m->dev->hard_header_len)
200 skb_queue_head_init(&q->q);
203 if (m->dev->flags & IFF_UP) {
204 if ((m->dev->flags&IFF_POINTOPOINT && !(dev->flags&IFF_POINTOPOINT))
205 || (m->dev->flags&IFF_BROADCAST && !(dev->flags&IFF_BROADCAST))
206 || (m->dev->flags&IFF_MULTICAST && !(dev->flags&IFF_MULTICAST))
207 || dev->mtu < m->dev->mtu)
210 if (!(dev->flags&IFF_POINTOPOINT))
211 m->dev->flags &= ~IFF_POINTOPOINT;
212 if (!(dev->flags&IFF_BROADCAST))
213 m->dev->flags &= ~IFF_BROADCAST;
214 if (!(dev->flags&IFF_MULTICAST))
215 m->dev->flags &= ~IFF_MULTICAST;
216 if (dev->mtu < m->dev->mtu)
217 m->dev->mtu = dev->mtu;
219 q->next = NEXT_SLAVE(m->slaves);
220 NEXT_SLAVE(m->slaves) = sch;
224 m->dev->mtu = dev->mtu;
225 m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
230 /* "teql*" netdevice routines */
233 __teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, struct net_device *dev)
235 struct teql_sched_data *q = qdisc_priv(dev->qdisc);
236 struct neighbour *mn = skb->dst->neighbour;
237 struct neighbour *n = q->ncache;
241 if (n && n->tbl == mn->tbl &&
242 memcmp(n->primary_key, mn->primary_key, mn->tbl->key_len) == 0) {
243 atomic_inc(&n->refcnt);
245 n = __neigh_lookup_errno(mn->tbl, mn->primary_key, dev);
249 if (neigh_event_send(n, skb_res) == 0) {
252 err = dev->hard_header(skb, dev, ntohs(skb->protocol), n->ha, NULL, skb->len);
253 read_unlock(&n->lock);
258 teql_neigh_release(xchg(&q->ncache, n));
262 return (skb_res == NULL) ? -EAGAIN : 1;
265 static __inline__ int
266 teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, struct net_device *dev)
268 if (dev->hard_header == NULL ||
270 skb->dst->neighbour == NULL)
272 return __teql_resolve(skb, skb_res, dev);
275 static int teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
277 struct teql_master *master = netdev_priv(dev);
278 struct Qdisc *start, *q;
282 struct sk_buff *skb_res = NULL;
284 start = master->slaves;
290 if ((q = start) == NULL)
294 struct net_device *slave = q->dev;
296 if (slave->qdisc_sleeping != q)
298 if (netif_queue_stopped(slave) || ! netif_running(slave)) {
303 switch (teql_resolve(skb, skb_res, slave)) {
305 if (netif_tx_trylock(slave)) {
306 if (!netif_queue_stopped(slave) &&
307 slave->hard_start_xmit(skb, slave) == 0) {
308 netif_tx_unlock(slave);
309 master->slaves = NEXT_SLAVE(q);
310 netif_wake_queue(dev);
311 master->stats.tx_packets++;
312 master->stats.tx_bytes += len;
315 netif_tx_unlock(slave);
317 if (netif_queue_stopped(dev))
321 master->slaves = NEXT_SLAVE(q);
327 __skb_pull(skb, skb->nh.raw - skb->data);
328 } while ((q = NEXT_SLAVE(q)) != start);
330 if (nores && skb_res == NULL) {
336 netif_stop_queue(dev);
339 master->stats.tx_errors++;
342 master->stats.tx_dropped++;
347 static int teql_master_open(struct net_device *dev)
350 struct teql_master *m = netdev_priv(dev);
352 unsigned flags = IFF_NOARP|IFF_MULTICAST;
354 if (m->slaves == NULL)
361 struct net_device *slave = q->dev;
366 if (slave->mtu < mtu)
368 if (slave->hard_header_len > LL_MAX_HEADER)
371 /* If all the slaves are BROADCAST, master is BROADCAST
372 If all the slaves are PtP, master is PtP
373 Otherwise, master is NBMA.
375 if (!(slave->flags&IFF_POINTOPOINT))
376 flags &= ~IFF_POINTOPOINT;
377 if (!(slave->flags&IFF_BROADCAST))
378 flags &= ~IFF_BROADCAST;
379 if (!(slave->flags&IFF_MULTICAST))
380 flags &= ~IFF_MULTICAST;
381 } while ((q = NEXT_SLAVE(q)) != m->slaves);
384 m->dev->flags = (m->dev->flags&~FMASK) | flags;
385 netif_start_queue(m->dev);
389 static int teql_master_close(struct net_device *dev)
391 netif_stop_queue(dev);
395 static struct net_device_stats *teql_master_stats(struct net_device *dev)
397 struct teql_master *m = netdev_priv(dev);
401 static int teql_master_mtu(struct net_device *dev, int new_mtu)
403 struct teql_master *m = netdev_priv(dev);
412 if (new_mtu > q->dev->mtu)
414 } while ((q=NEXT_SLAVE(q)) != m->slaves);
421 static __init void teql_master_setup(struct net_device *dev)
423 struct teql_master *master = netdev_priv(dev);
424 struct Qdisc_ops *ops = &master->qops;
427 ops->priv_size = sizeof(struct teql_sched_data);
429 ops->enqueue = teql_enqueue;
430 ops->dequeue = teql_dequeue;
431 ops->requeue = teql_requeue;
432 ops->init = teql_qdisc_init;
433 ops->reset = teql_reset;
434 ops->destroy = teql_destroy;
435 ops->owner = THIS_MODULE;
437 dev->open = teql_master_open;
438 dev->hard_start_xmit = teql_master_xmit;
439 dev->stop = teql_master_close;
440 dev->get_stats = teql_master_stats;
441 dev->change_mtu = teql_master_mtu;
442 dev->type = ARPHRD_VOID;
444 dev->tx_queue_len = 100;
445 dev->flags = IFF_NOARP;
446 dev->hard_header_len = LL_MAX_HEADER;
447 SET_MODULE_OWNER(dev);
450 static LIST_HEAD(master_dev_list);
451 static int max_equalizers = 1;
452 module_param(max_equalizers, int, 0);
453 MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
455 static int __init teql_init(void)
460 for (i = 0; i < max_equalizers; i++) {
461 struct net_device *dev;
462 struct teql_master *master;
464 dev = alloc_netdev(sizeof(struct teql_master),
465 "teql%d", teql_master_setup);
471 if ((err = register_netdev(dev))) {
476 master = netdev_priv(dev);
478 strlcpy(master->qops.id, dev->name, IFNAMSIZ);
479 err = register_qdisc(&master->qops);
482 unregister_netdev(dev);
487 list_add_tail(&master->master_list, &master_dev_list);
492 static void __exit teql_exit(void)
494 struct teql_master *master, *nxt;
496 list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) {
498 list_del(&master->master_list);
500 unregister_qdisc(&master->qops);
501 unregister_netdev(master->dev);
502 free_netdev(master->dev);
506 module_init(teql_init);
507 module_exit(teql_exit);
509 MODULE_LICENSE("GPL");