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 <linux/types.h>
13 #include <linux/kernel.h>
14 #include <linux/slab.h>
15 #include <linux/string.h>
16 #include <linux/errno.h>
17 #include <linux/if_arp.h>
18 #include <linux/netdevice.h>
19 #include <linux/init.h>
20 #include <linux/skbuff.h>
21 #include <linux/moduleparam.h>
23 #include <net/neighbour.h>
24 #include <net/pkt_sched.h>
30 After loading this module you will find a new device teqlN
31 and new qdisc with the same name. To join a slave to the equalizer
32 you should just set this qdisc on a device f.e.
34 # tc qdisc add dev eth0 root teql0
35 # tc qdisc add dev eth1 root teql0
37 That's all. Full PnP 8)
42 1. Slave devices MUST be active devices, i.e., they must raise the tbusy
43 signal and generate EOI events. If you want to equalize virtual devices
44 like tunnels, use a normal eql device.
45 2. This device puts no limitations on physical slave characteristics
46 f.e. it will equalize 9600baud line and 100Mb ethernet perfectly :-)
47 Certainly, large difference in link speeds will make the resulting
48 eqalized link unusable, because of huge packet reordering.
49 I estimate an upper useful difference as ~10 times.
50 3. If the slave requires address resolution, only protocols using
51 neighbour cache (IPv4/IPv6) will work over the equalized link.
52 Other protocols are still allowed to use the slave device directly,
53 which will not break load balancing, though native slave
54 traffic will have the highest priority. */
58 struct Qdisc_ops qops;
59 struct net_device *dev;
61 struct list_head master_list;
64 struct teql_sched_data
67 struct teql_master *m;
68 struct neighbour *ncache;
69 struct sk_buff_head q;
72 #define NEXT_SLAVE(q) (((struct teql_sched_data*)qdisc_priv(q))->next)
74 #define FMASK (IFF_BROADCAST|IFF_POINTOPOINT)
76 /* "teql*" qdisc routines */
79 teql_enqueue(struct sk_buff *skb, struct Qdisc* sch)
81 struct net_device *dev = qdisc_dev(sch);
82 struct teql_sched_data *q = qdisc_priv(sch);
84 if (q->q.qlen < dev->tx_queue_len) {
85 __skb_queue_tail(&q->q, skb);
86 qdisc_bstats_update(sch, skb);
87 return NET_XMIT_SUCCESS;
95 static struct sk_buff *
96 teql_dequeue(struct Qdisc* sch)
98 struct teql_sched_data *dat = qdisc_priv(sch);
99 struct netdev_queue *dat_queue;
102 skb = __skb_dequeue(&dat->q);
103 dat_queue = netdev_get_tx_queue(dat->m->dev, 0);
105 struct net_device *m = qdisc_dev(dat_queue->qdisc);
107 dat->m->slaves = sch;
111 sch->q.qlen = dat->q.qlen + dat_queue->qdisc->q.qlen;
115 static struct sk_buff *
116 teql_peek(struct Qdisc* sch)
118 /* teql is meant to be used as root qdisc */
122 static __inline__ void
123 teql_neigh_release(struct neighbour *n)
130 teql_reset(struct Qdisc* sch)
132 struct teql_sched_data *dat = qdisc_priv(sch);
134 skb_queue_purge(&dat->q);
136 teql_neigh_release(xchg(&dat->ncache, NULL));
140 teql_destroy(struct Qdisc* sch)
142 struct Qdisc *q, *prev;
143 struct teql_sched_data *dat = qdisc_priv(sch);
144 struct teql_master *master = dat->m;
146 if ((prev = master->slaves) != NULL) {
148 q = NEXT_SLAVE(prev);
150 NEXT_SLAVE(prev) = NEXT_SLAVE(q);
151 if (q == master->slaves) {
152 master->slaves = NEXT_SLAVE(q);
153 if (q == master->slaves) {
154 struct netdev_queue *txq;
155 spinlock_t *root_lock;
157 txq = netdev_get_tx_queue(master->dev, 0);
158 master->slaves = NULL;
160 root_lock = qdisc_root_sleeping_lock(txq->qdisc);
161 spin_lock_bh(root_lock);
162 qdisc_reset(txq->qdisc);
163 spin_unlock_bh(root_lock);
166 skb_queue_purge(&dat->q);
167 teql_neigh_release(xchg(&dat->ncache, NULL));
171 } while ((prev = q) != master->slaves);
175 static int teql_qdisc_init(struct Qdisc *sch, struct nlattr *opt)
177 struct net_device *dev = qdisc_dev(sch);
178 struct teql_master *m = (struct teql_master*)sch->ops;
179 struct teql_sched_data *q = qdisc_priv(sch);
181 if (dev->hard_header_len > m->dev->hard_header_len)
189 skb_queue_head_init(&q->q);
192 if (m->dev->flags & IFF_UP) {
193 if ((m->dev->flags & IFF_POINTOPOINT &&
194 !(dev->flags & IFF_POINTOPOINT)) ||
195 (m->dev->flags & IFF_BROADCAST &&
196 !(dev->flags & IFF_BROADCAST)) ||
197 (m->dev->flags & IFF_MULTICAST &&
198 !(dev->flags & IFF_MULTICAST)) ||
199 dev->mtu < m->dev->mtu)
202 if (!(dev->flags&IFF_POINTOPOINT))
203 m->dev->flags &= ~IFF_POINTOPOINT;
204 if (!(dev->flags&IFF_BROADCAST))
205 m->dev->flags &= ~IFF_BROADCAST;
206 if (!(dev->flags&IFF_MULTICAST))
207 m->dev->flags &= ~IFF_MULTICAST;
208 if (dev->mtu < m->dev->mtu)
209 m->dev->mtu = dev->mtu;
211 q->next = NEXT_SLAVE(m->slaves);
212 NEXT_SLAVE(m->slaves) = sch;
216 m->dev->mtu = dev->mtu;
217 m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
224 __teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, struct net_device *dev)
226 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, 0);
227 struct teql_sched_data *q = qdisc_priv(dev_queue->qdisc);
228 struct neighbour *mn = skb_dst(skb)->neighbour;
229 struct neighbour *n = q->ncache;
233 if (n && n->tbl == mn->tbl &&
234 memcmp(n->primary_key, mn->primary_key, mn->tbl->key_len) == 0) {
235 atomic_inc(&n->refcnt);
237 n = __neigh_lookup_errno(mn->tbl, mn->primary_key, dev);
241 if (neigh_event_send(n, skb_res) == 0) {
243 char haddr[MAX_ADDR_LEN];
245 neigh_ha_snapshot(haddr, n, dev);
246 err = dev_hard_header(skb, dev, ntohs(skb->protocol), haddr,
253 teql_neigh_release(xchg(&q->ncache, n));
257 return (skb_res == NULL) ? -EAGAIN : 1;
260 static inline int teql_resolve(struct sk_buff *skb,
261 struct sk_buff *skb_res, struct net_device *dev)
263 struct netdev_queue *txq = netdev_get_tx_queue(dev, 0);
264 if (txq->qdisc == &noop_qdisc)
267 if (dev->header_ops == NULL ||
268 skb_dst(skb) == NULL ||
269 skb_dst(skb)->neighbour == NULL)
271 return __teql_resolve(skb, skb_res, dev);
274 static netdev_tx_t teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
276 struct teql_master *master = netdev_priv(dev);
277 struct netdev_queue *txq = netdev_get_tx_queue(dev, 0);
278 struct Qdisc *start, *q;
281 int subq = skb_get_queue_mapping(skb);
282 struct sk_buff *skb_res = NULL;
284 start = master->slaves;
290 if ((q = start) == NULL)
294 struct net_device *slave = qdisc_dev(q);
295 struct netdev_queue *slave_txq = netdev_get_tx_queue(slave, 0);
296 const struct net_device_ops *slave_ops = slave->netdev_ops;
298 if (slave_txq->qdisc_sleeping != q)
300 if (__netif_subqueue_stopped(slave, subq) ||
301 !netif_running(slave)) {
306 switch (teql_resolve(skb, skb_res, slave)) {
308 if (__netif_tx_trylock(slave_txq)) {
309 unsigned int length = qdisc_pkt_len(skb);
311 if (!netif_tx_queue_frozen_or_stopped(slave_txq) &&
312 slave_ops->ndo_start_xmit(skb, slave) == NETDEV_TX_OK) {
313 txq_trans_update(slave_txq);
314 __netif_tx_unlock(slave_txq);
315 master->slaves = NEXT_SLAVE(q);
316 netif_wake_queue(dev);
318 txq->tx_bytes += length;
321 __netif_tx_unlock(slave_txq);
323 if (netif_queue_stopped(dev))
327 master->slaves = NEXT_SLAVE(q);
333 __skb_pull(skb, skb_network_offset(skb));
334 } while ((q = NEXT_SLAVE(q)) != start);
336 if (nores && skb_res == NULL) {
342 netif_stop_queue(dev);
343 return NETDEV_TX_BUSY;
345 dev->stats.tx_errors++;
353 static int teql_master_open(struct net_device *dev)
356 struct teql_master *m = netdev_priv(dev);
358 unsigned flags = IFF_NOARP|IFF_MULTICAST;
360 if (m->slaves == NULL)
367 struct net_device *slave = qdisc_dev(q);
372 if (slave->mtu < mtu)
374 if (slave->hard_header_len > LL_MAX_HEADER)
377 /* If all the slaves are BROADCAST, master is BROADCAST
378 If all the slaves are PtP, master is PtP
379 Otherwise, master is NBMA.
381 if (!(slave->flags&IFF_POINTOPOINT))
382 flags &= ~IFF_POINTOPOINT;
383 if (!(slave->flags&IFF_BROADCAST))
384 flags &= ~IFF_BROADCAST;
385 if (!(slave->flags&IFF_MULTICAST))
386 flags &= ~IFF_MULTICAST;
387 } while ((q = NEXT_SLAVE(q)) != m->slaves);
390 m->dev->flags = (m->dev->flags&~FMASK) | flags;
391 netif_start_queue(m->dev);
395 static int teql_master_close(struct net_device *dev)
397 netif_stop_queue(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 > qdisc_dev(q)->mtu)
414 } while ((q=NEXT_SLAVE(q)) != m->slaves);
421 static const struct net_device_ops teql_netdev_ops = {
422 .ndo_open = teql_master_open,
423 .ndo_stop = teql_master_close,
424 .ndo_start_xmit = teql_master_xmit,
425 .ndo_change_mtu = teql_master_mtu,
428 static __init void teql_master_setup(struct net_device *dev)
430 struct teql_master *master = netdev_priv(dev);
431 struct Qdisc_ops *ops = &master->qops;
434 ops->priv_size = sizeof(struct teql_sched_data);
436 ops->enqueue = teql_enqueue;
437 ops->dequeue = teql_dequeue;
438 ops->peek = teql_peek;
439 ops->init = teql_qdisc_init;
440 ops->reset = teql_reset;
441 ops->destroy = teql_destroy;
442 ops->owner = THIS_MODULE;
444 dev->netdev_ops = &teql_netdev_ops;
445 dev->type = ARPHRD_VOID;
447 dev->tx_queue_len = 100;
448 dev->flags = IFF_NOARP;
449 dev->hard_header_len = LL_MAX_HEADER;
450 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
453 static LIST_HEAD(master_dev_list);
454 static int max_equalizers = 1;
455 module_param(max_equalizers, int, 0);
456 MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
458 static int __init teql_init(void)
463 for (i = 0; i < max_equalizers; i++) {
464 struct net_device *dev;
465 struct teql_master *master;
467 dev = alloc_netdev(sizeof(struct teql_master),
468 "teql%d", teql_master_setup);
474 if ((err = register_netdev(dev))) {
479 master = netdev_priv(dev);
481 strlcpy(master->qops.id, dev->name, IFNAMSIZ);
482 err = register_qdisc(&master->qops);
485 unregister_netdev(dev);
490 list_add_tail(&master->master_list, &master_dev_list);
495 static void __exit teql_exit(void)
497 struct teql_master *master, *nxt;
499 list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) {
501 list_del(&master->master_list);
503 unregister_qdisc(&master->qops);
504 unregister_netdev(master->dev);
505 free_netdev(master->dev);
509 module_init(teql_init);
510 module_exit(teql_exit);
512 MODULE_LICENSE("GPL");