Import 2.3.47pre1
[davej-history.git] / net / sched / sch_teql.c
blobf7fed6c84c86c38c7bde68325506f84564c5aabe
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>
9 */
11 #include <linux/module.h>
12 #include <asm/uaccess.h>
13 #include <asm/system.h>
14 #include <asm/bitops.h>
15 #include <linux/types.h>
16 #include <linux/kernel.h>
17 #include <linux/sched.h>
18 #include <linux/string.h>
19 #include <linux/mm.h>
20 #include <linux/socket.h>
21 #include <linux/sockios.h>
22 #include <linux/in.h>
23 #include <linux/errno.h>
24 #include <linux/interrupt.h>
25 #include <linux/if_ether.h>
26 #include <linux/inet.h>
27 #include <linux/netdevice.h>
28 #include <linux/etherdevice.h>
29 #include <linux/notifier.h>
30 #include <linux/init.h>
31 #include <net/ip.h>
32 #include <net/route.h>
33 #include <linux/skbuff.h>
34 #include <net/sock.h>
35 #include <net/pkt_sched.h>
38 How to setup it.
39 ----------------
41 After loading this module you will find a new device teqlN
42 and new qdisc with the same name. To join a slave to the equalizer
43 you should just set this qdisc on a device f.e.
45 # tc qdisc add dev eth0 root teql0
46 # tc qdisc add dev eth1 root teql0
48 That's all. Full PnP 8)
50 Applicability.
51 --------------
53 1. Slave devices MUST be active devices, i.e., they must raise the tbusy
54 signal and generate EOI events. If you want to equalize virtual devices
55 like tunnels, use a normal eql device.
56 2. This device puts no limitations on physical slave characteristics
57 f.e. it will equalize 9600baud line and 100Mb ethernet perfectly :-)
58 Certainly, large difference in link speeds will make the resulting
59 eqalized link unusable, because of huge packet reordering.
60 I estimate an upper useful difference as ~10 times.
61 3. If the slave requires address resolution, only protocols using
62 neighbour cache (IPv4/IPv6) will work over the equalized link.
63 Other protocols are still allowed to use the slave device directly,
64 which will not break load balancing, though native slave
65 traffic will have the highest priority. */
67 struct teql_master
69 struct Qdisc_ops qops;
70 struct net_device dev;
71 struct Qdisc *slaves;
72 struct net_device_stats stats;
73 char name[IFNAMSIZ];
76 struct teql_sched_data
78 struct Qdisc *next;
79 struct teql_master *m;
80 struct neighbour *ncache;
81 struct sk_buff_head q;
84 #define NEXT_SLAVE(q) (((struct teql_sched_data*)((q)->data))->next)
86 #define FMASK (IFF_BROADCAST|IFF_POINTOPOINT|IFF_BROADCAST)
88 /* "teql*" qdisc routines */
90 static int
91 teql_enqueue(struct sk_buff *skb, struct Qdisc* sch)
93 struct net_device *dev = sch->dev;
94 struct teql_sched_data *q = (struct teql_sched_data *)sch->data;
96 __skb_queue_tail(&q->q, skb);
97 if (q->q.qlen <= dev->tx_queue_len) {
98 sch->stats.bytes += skb->len;
99 sch->stats.packets++;
100 return 0;
103 __skb_unlink(skb, &q->q);
104 kfree_skb(skb);
105 sch->stats.drops++;
106 return NET_XMIT_DROP;
109 static int
110 teql_requeue(struct sk_buff *skb, struct Qdisc* sch)
112 struct teql_sched_data *q = (struct teql_sched_data *)sch->data;
114 __skb_queue_head(&q->q, skb);
115 return 0;
118 static struct sk_buff *
119 teql_dequeue(struct Qdisc* sch)
121 struct teql_sched_data *dat = (struct teql_sched_data *)sch->data;
122 struct sk_buff *skb;
124 skb = __skb_dequeue(&dat->q);
125 if (skb == NULL) {
126 struct net_device *m = dat->m->dev.qdisc->dev;
127 if (m) {
128 dat->m->slaves = sch;
129 netif_wake_queue(m);
132 sch->q.qlen = dat->q.qlen + dat->m->dev.qdisc->q.qlen;
133 return skb;
136 static __inline__ void
137 teql_neigh_release(struct neighbour *n)
139 if (n)
140 neigh_release(n);
143 static void
144 teql_reset(struct Qdisc* sch)
146 struct teql_sched_data *dat = (struct teql_sched_data *)sch->data;
148 skb_queue_purge(&dat->q);
149 sch->q.qlen = 0;
150 teql_neigh_release(xchg(&dat->ncache, NULL));
153 static void
154 teql_destroy(struct Qdisc* sch)
156 struct Qdisc *q, *prev;
157 struct teql_sched_data *dat = (struct teql_sched_data *)sch->data;
158 struct teql_master *master = dat->m;
160 if ((prev = master->slaves) != NULL) {
161 do {
162 q = NEXT_SLAVE(prev);
163 if (q == sch) {
164 NEXT_SLAVE(prev) = NEXT_SLAVE(q);
165 if (q == master->slaves) {
166 master->slaves = NEXT_SLAVE(q);
167 if (q == master->slaves) {
168 master->slaves = NULL;
169 spin_lock_bh(&master->dev.queue_lock);
170 qdisc_reset(master->dev.qdisc);
171 spin_unlock_bh(&master->dev.queue_lock);
174 skb_queue_purge(&dat->q);
175 teql_neigh_release(xchg(&dat->ncache, NULL));
176 break;
179 } while ((prev = q) != master->slaves);
182 MOD_DEC_USE_COUNT;
185 static int teql_qdisc_init(struct Qdisc *sch, struct rtattr *opt)
187 struct net_device *dev = sch->dev;
188 struct teql_master *m = (struct teql_master*)sch->ops;
189 struct teql_sched_data *q = (struct teql_sched_data *)sch->data;
191 if (dev->hard_header_len > m->dev.hard_header_len)
192 return -EINVAL;
194 if (&m->dev == dev)
195 return -ELOOP;
197 q->m = m;
199 skb_queue_head_init(&q->q);
201 if (m->slaves) {
202 if (m->dev.flags & IFF_UP) {
203 if ((m->dev.flags&IFF_POINTOPOINT && !(dev->flags&IFF_POINTOPOINT))
204 || (m->dev.flags&IFF_BROADCAST && !(dev->flags&IFF_BROADCAST))
205 || (m->dev.flags&IFF_MULTICAST && !(dev->flags&IFF_MULTICAST))
206 || dev->mtu < m->dev.mtu)
207 return -EINVAL;
208 } else {
209 if (!(dev->flags&IFF_POINTOPOINT))
210 m->dev.flags &= ~IFF_POINTOPOINT;
211 if (!(dev->flags&IFF_BROADCAST))
212 m->dev.flags &= ~IFF_BROADCAST;
213 if (!(dev->flags&IFF_MULTICAST))
214 m->dev.flags &= ~IFF_MULTICAST;
215 if (dev->mtu < m->dev.mtu)
216 m->dev.mtu = dev->mtu;
218 q->next = NEXT_SLAVE(m->slaves);
219 NEXT_SLAVE(m->slaves) = sch;
220 } else {
221 q->next = sch;
222 m->slaves = sch;
223 m->dev.mtu = dev->mtu;
224 m->dev.flags = (m->dev.flags&~FMASK)|(dev->flags&FMASK);
227 MOD_INC_USE_COUNT;
228 return 0;
231 /* "teql*" netdevice routines */
233 static int
234 __teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, struct net_device *dev)
236 struct teql_sched_data *q = (void*)dev->qdisc->data;
237 struct neighbour *mn = skb->dst->neighbour;
238 struct neighbour *n = q->ncache;
240 if (mn->tbl == NULL)
241 return -EINVAL;
242 if (n && n->tbl == mn->tbl &&
243 memcmp(n->primary_key, mn->primary_key, mn->tbl->key_len) == 0) {
244 atomic_inc(&n->refcnt);
245 } else {
246 n = __neigh_lookup(mn->tbl, mn->primary_key, dev, 1);
247 if (n == NULL)
248 return -ENOBUFS;
250 if (neigh_event_send(n, skb_res) == 0) {
251 int err;
252 read_lock(&n->lock);
253 err = dev->hard_header(skb, dev, ntohs(skb->protocol), n->ha, NULL, skb->len);
254 read_unlock(&n->lock);
255 if (err < 0) {
256 neigh_release(n);
257 return -EINVAL;
259 teql_neigh_release(xchg(&q->ncache, n));
260 return 0;
262 neigh_release(n);
263 return (skb_res == NULL) ? -EAGAIN : 1;
266 static __inline__ int
267 teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, struct net_device *dev)
269 if (dev->hard_header == NULL ||
270 skb->dst == NULL ||
271 skb->dst->neighbour == NULL)
272 return 0;
273 return __teql_resolve(skb, skb_res, dev);
276 static int teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
278 struct teql_master *master = (void*)dev->priv;
279 struct Qdisc *start, *q;
280 int busy;
281 int nores;
282 int len = skb->len;
283 struct sk_buff *skb_res = NULL;
285 start = master->slaves;
287 restart:
288 nores = 0;
289 busy = 0;
291 if ((q = start) == NULL)
292 goto drop;
294 do {
295 struct net_device *slave = q->dev;
297 if (slave->qdisc_sleeping != q)
298 continue;
299 if (netif_queue_stopped(slave) || ! netif_running(slave)) {
300 busy = 1;
301 continue;
304 switch (teql_resolve(skb, skb_res, slave)) {
305 case 0:
306 if (spin_trylock(&slave->xmit_lock)) {
307 slave->xmit_lock_owner = smp_processor_id();
308 if (!netif_queue_stopped(slave) &&
309 slave->hard_start_xmit(skb, slave) == 0) {
310 slave->xmit_lock_owner = -1;
311 spin_unlock(&slave->xmit_lock);
312 master->slaves = NEXT_SLAVE(q);
313 netif_wake_queue(dev);
314 master->stats.tx_packets++;
315 master->stats.tx_bytes += len;
316 return 0;
318 slave->xmit_lock_owner = -1;
319 spin_unlock(&slave->xmit_lock);
321 if (netif_queue_stopped(dev))
322 busy = 1;
323 break;
324 case 1:
325 master->slaves = NEXT_SLAVE(q);
326 return 0;
327 default:
328 nores = 1;
329 break;
331 __skb_pull(skb, skb->nh.raw - skb->data);
332 } while ((q = NEXT_SLAVE(q)) != start);
334 if (nores && skb_res == NULL) {
335 skb_res = skb;
336 goto restart;
339 if (busy) {
340 netif_stop_queue(dev);
341 return 1;
343 master->stats.tx_errors++;
345 drop:
346 master->stats.tx_dropped++;
347 dev_kfree_skb(skb);
348 return 0;
351 static int teql_master_open(struct net_device *dev)
353 struct Qdisc * q;
354 struct teql_master *m = (void*)dev->priv;
355 int mtu = 0xFFFE;
356 unsigned flags = IFF_NOARP|IFF_MULTICAST;
358 if (m->slaves == NULL)
359 return -EUNATCH;
361 flags = FMASK;
363 q = m->slaves;
364 do {
365 struct net_device *slave = q->dev;
367 if (slave == NULL)
368 return -EUNATCH;
370 if (slave->mtu < mtu)
371 mtu = slave->mtu;
372 if (slave->hard_header_len > LL_MAX_HEADER)
373 return -EINVAL;
375 /* If all the slaves are BROADCAST, master is BROADCAST
376 If all the slaves are PtP, master is PtP
377 Otherwise, master is NBMA.
379 if (!(slave->flags&IFF_POINTOPOINT))
380 flags &= ~IFF_POINTOPOINT;
381 if (!(slave->flags&IFF_BROADCAST))
382 flags &= ~IFF_BROADCAST;
383 if (!(slave->flags&IFF_MULTICAST))
384 flags &= ~IFF_MULTICAST;
385 } while ((q = NEXT_SLAVE(q)) != m->slaves);
387 m->dev.mtu = mtu;
388 m->dev.flags = (m->dev.flags&~FMASK) | flags;
389 netif_start_queue(&m->dev);
390 MOD_INC_USE_COUNT;
391 return 0;
394 static int teql_master_close(struct net_device *dev)
396 netif_stop_queue(dev);
397 MOD_DEC_USE_COUNT;
398 return 0;
401 static struct net_device_stats *teql_master_stats(struct net_device *dev)
403 struct teql_master *m = (void*)dev->priv;
404 return &m->stats;
407 static int teql_master_mtu(struct net_device *dev, int new_mtu)
409 struct teql_master *m = (void*)dev->priv;
410 struct Qdisc *q;
412 if (new_mtu < 68)
413 return -EINVAL;
415 q = m->slaves;
416 if (q) {
417 do {
418 if (new_mtu > q->dev->mtu)
419 return -EINVAL;
420 } while ((q=NEXT_SLAVE(q)) != m->slaves);
423 dev->mtu = new_mtu;
424 return 0;
427 static int teql_master_init(struct net_device *dev)
429 dev->open = teql_master_open;
430 dev->hard_start_xmit = teql_master_xmit;
431 dev->stop = teql_master_close;
432 dev->get_stats = teql_master_stats;
433 dev->change_mtu = teql_master_mtu;
434 dev->type = ARPHRD_VOID;
435 dev->mtu = 1500;
436 dev->tx_queue_len = 100;
437 dev->flags = IFF_NOARP;
438 dev->hard_header_len = LL_MAX_HEADER;
439 return 0;
442 static struct teql_master the_master = {
444 NULL,
445 NULL,
447 sizeof(struct teql_sched_data),
449 teql_enqueue,
450 teql_dequeue,
451 teql_requeue,
452 NULL,
454 teql_qdisc_init,
455 teql_reset,
456 teql_destroy,
457 NULL,
458 },};
461 #ifdef MODULE
462 int init_module(void)
463 #else
464 int __init teql_init(void)
465 #endif
467 int err;
469 rtnl_lock();
471 the_master.dev.priv = (void*)&the_master;
472 the_master.dev.name = (void*)&the_master.name;
473 err = dev_alloc_name(&the_master.dev, "teql%d");
474 if (err < 0)
475 return err;
476 memcpy(the_master.qops.id, the_master.name, IFNAMSIZ);
477 the_master.dev.init = teql_master_init;
479 err = register_netdevice(&the_master.dev);
480 if (err == 0) {
481 err = register_qdisc(&the_master.qops);
482 if (err)
483 unregister_netdevice(&the_master.dev);
485 rtnl_unlock();
486 return err;
489 #ifdef MODULE
490 void cleanup_module(void)
492 rtnl_lock();
493 unregister_qdisc(&the_master.qops);
494 unregister_netdevice(&the_master.dev);
495 rtnl_unlock();
497 #endif