]> bbs.cooldavid.org Git - net-next-2.6.git/blame - net/sched/sch_teql.c
[CORE] Stack changes to add multiqueue hardware support API
[net-next-2.6.git] / net / sched / sch_teql.c
CommitLineData
1da177e4
LT
1/* net/sched/sch_teql.c "True" (or "trivial") link equalizer.
2 *
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.
7 *
8 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
9 */
10
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>
1da177e4
LT
17#include <linux/string.h>
18#include <linux/mm.h>
19#include <linux/socket.h>
20#include <linux/sockios.h>
21#include <linux/in.h>
22#include <linux/errno.h>
23#include <linux/interrupt.h>
14c85021 24#include <linux/if_arp.h>
1da177e4
LT
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 <linux/moduleparam.h>
35#include <net/sock.h>
36#include <net/pkt_sched.h>
37
38/*
39 How to setup it.
40 ----------------
41
42 After loading this module you will find a new device teqlN
43 and new qdisc with the same name. To join a slave to the equalizer
44 you should just set this qdisc on a device f.e.
45
46 # tc qdisc add dev eth0 root teql0
47 # tc qdisc add dev eth1 root teql0
48
49 That's all. Full PnP 8)
50
51 Applicability.
52 --------------
53
54 1. Slave devices MUST be active devices, i.e., they must raise the tbusy
55 signal and generate EOI events. If you want to equalize virtual devices
56 like tunnels, use a normal eql device.
57 2. This device puts no limitations on physical slave characteristics
58 f.e. it will equalize 9600baud line and 100Mb ethernet perfectly :-)
59 Certainly, large difference in link speeds will make the resulting
60 eqalized link unusable, because of huge packet reordering.
61 I estimate an upper useful difference as ~10 times.
62 3. If the slave requires address resolution, only protocols using
63 neighbour cache (IPv4/IPv6) will work over the equalized link.
64 Other protocols are still allowed to use the slave device directly,
65 which will not break load balancing, though native slave
66 traffic will have the highest priority. */
67
68struct teql_master
69{
70 struct Qdisc_ops qops;
71 struct net_device *dev;
72 struct Qdisc *slaves;
73 struct list_head master_list;
74 struct net_device_stats stats;
75};
76
77struct teql_sched_data
78{
79 struct Qdisc *next;
80 struct teql_master *m;
81 struct neighbour *ncache;
82 struct sk_buff_head q;
83};
84
85#define NEXT_SLAVE(q) (((struct teql_sched_data*)qdisc_priv(q))->next)
86
87#define FMASK (IFF_BROADCAST|IFF_POINTOPOINT|IFF_BROADCAST)
88
89/* "teql*" qdisc routines */
90
91static int
92teql_enqueue(struct sk_buff *skb, struct Qdisc* sch)
93{
94 struct net_device *dev = sch->dev;
95 struct teql_sched_data *q = qdisc_priv(sch);
96
4cd8c9e8
KK
97 if (q->q.qlen < dev->tx_queue_len) {
98 __skb_queue_tail(&q->q, skb);
1da177e4
LT
99 sch->bstats.bytes += skb->len;
100 sch->bstats.packets++;
101 return 0;
102 }
103
1da177e4
LT
104 kfree_skb(skb);
105 sch->qstats.drops++;
106 return NET_XMIT_DROP;
107}
108
109static int
110teql_requeue(struct sk_buff *skb, struct Qdisc* sch)
111{
112 struct teql_sched_data *q = qdisc_priv(sch);
113
114 __skb_queue_head(&q->q, skb);
115 sch->qstats.requeues++;
116 return 0;
117}
118
119static struct sk_buff *
120teql_dequeue(struct Qdisc* sch)
121{
122 struct teql_sched_data *dat = qdisc_priv(sch);
123 struct sk_buff *skb;
124
125 skb = __skb_dequeue(&dat->q);
126 if (skb == NULL) {
127 struct net_device *m = dat->m->dev->qdisc->dev;
128 if (m) {
129 dat->m->slaves = sch;
130 netif_wake_queue(m);
131 }
132 }
133 sch->q.qlen = dat->q.qlen + dat->m->dev->qdisc->q.qlen;
134 return skb;
135}
136
137static __inline__ void
138teql_neigh_release(struct neighbour *n)
139{
140 if (n)
141 neigh_release(n);
142}
143
144static void
145teql_reset(struct Qdisc* sch)
146{
147 struct teql_sched_data *dat = qdisc_priv(sch);
148
149 skb_queue_purge(&dat->q);
150 sch->q.qlen = 0;
151 teql_neigh_release(xchg(&dat->ncache, NULL));
152}
153
154static void
155teql_destroy(struct Qdisc* sch)
156{
157 struct Qdisc *q, *prev;
158 struct teql_sched_data *dat = qdisc_priv(sch);
159 struct teql_master *master = dat->m;
160
161 if ((prev = master->slaves) != NULL) {
162 do {
163 q = NEXT_SLAVE(prev);
164 if (q == sch) {
165 NEXT_SLAVE(prev) = NEXT_SLAVE(q);
166 if (q == master->slaves) {
167 master->slaves = NEXT_SLAVE(q);
168 if (q == master->slaves) {
169 master->slaves = NULL;
170 spin_lock_bh(&master->dev->queue_lock);
171 qdisc_reset(master->dev->qdisc);
172 spin_unlock_bh(&master->dev->queue_lock);
173 }
174 }
175 skb_queue_purge(&dat->q);
176 teql_neigh_release(xchg(&dat->ncache, NULL));
177 break;
178 }
10297b99 179
1da177e4
LT
180 } while ((prev = q) != master->slaves);
181 }
182}
183
184static int teql_qdisc_init(struct Qdisc *sch, struct rtattr *opt)
185{
186 struct net_device *dev = sch->dev;
187 struct teql_master *m = (struct teql_master*)sch->ops;
188 struct teql_sched_data *q = qdisc_priv(sch);
189
190 if (dev->hard_header_len > m->dev->hard_header_len)
191 return -EINVAL;
192
193 if (m->dev == dev)
194 return -ELOOP;
195
196 q->m = m;
197
198 skb_queue_head_init(&q->q);
199
200 if (m->slaves) {
201 if (m->dev->flags & IFF_UP) {
202 if ((m->dev->flags&IFF_POINTOPOINT && !(dev->flags&IFF_POINTOPOINT))
203 || (m->dev->flags&IFF_BROADCAST && !(dev->flags&IFF_BROADCAST))
204 || (m->dev->flags&IFF_MULTICAST && !(dev->flags&IFF_MULTICAST))
205 || dev->mtu < m->dev->mtu)
206 return -EINVAL;
207 } else {
208 if (!(dev->flags&IFF_POINTOPOINT))
209 m->dev->flags &= ~IFF_POINTOPOINT;
210 if (!(dev->flags&IFF_BROADCAST))
211 m->dev->flags &= ~IFF_BROADCAST;
212 if (!(dev->flags&IFF_MULTICAST))
213 m->dev->flags &= ~IFF_MULTICAST;
214 if (dev->mtu < m->dev->mtu)
215 m->dev->mtu = dev->mtu;
216 }
217 q->next = NEXT_SLAVE(m->slaves);
218 NEXT_SLAVE(m->slaves) = sch;
219 } else {
220 q->next = sch;
221 m->slaves = sch;
222 m->dev->mtu = dev->mtu;
223 m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
224 }
225 return 0;
226}
227
228/* "teql*" netdevice routines */
229
230static int
231__teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, struct net_device *dev)
232{
233 struct teql_sched_data *q = qdisc_priv(dev->qdisc);
234 struct neighbour *mn = skb->dst->neighbour;
235 struct neighbour *n = q->ncache;
236
237 if (mn->tbl == NULL)
238 return -EINVAL;
239 if (n && n->tbl == mn->tbl &&
240 memcmp(n->primary_key, mn->primary_key, mn->tbl->key_len) == 0) {
241 atomic_inc(&n->refcnt);
242 } else {
243 n = __neigh_lookup_errno(mn->tbl, mn->primary_key, dev);
244 if (IS_ERR(n))
245 return PTR_ERR(n);
246 }
247 if (neigh_event_send(n, skb_res) == 0) {
248 int err;
249 read_lock(&n->lock);
250 err = dev->hard_header(skb, dev, ntohs(skb->protocol), n->ha, NULL, skb->len);
251 read_unlock(&n->lock);
252 if (err < 0) {
253 neigh_release(n);
254 return -EINVAL;
255 }
256 teql_neigh_release(xchg(&q->ncache, n));
257 return 0;
258 }
259 neigh_release(n);
260 return (skb_res == NULL) ? -EAGAIN : 1;
261}
262
263static __inline__ int
264teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, struct net_device *dev)
265{
266 if (dev->hard_header == NULL ||
267 skb->dst == NULL ||
268 skb->dst->neighbour == NULL)
269 return 0;
270 return __teql_resolve(skb, skb_res, dev);
271}
272
273static int teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
274{
2941a486 275 struct teql_master *master = netdev_priv(dev);
1da177e4
LT
276 struct Qdisc *start, *q;
277 int busy;
278 int nores;
279 int len = skb->len;
f25f4e44 280 int subq = skb->queue_mapping;
1da177e4
LT
281 struct sk_buff *skb_res = NULL;
282
283 start = master->slaves;
284
285restart:
286 nores = 0;
287 busy = 0;
288
289 if ((q = start) == NULL)
290 goto drop;
291
292 do {
293 struct net_device *slave = q->dev;
10297b99 294
1da177e4
LT
295 if (slave->qdisc_sleeping != q)
296 continue;
f25f4e44
PWJ
297 if (netif_queue_stopped(slave) ||
298 netif_subqueue_stopped(slave, subq) ||
299 !netif_running(slave)) {
1da177e4
LT
300 busy = 1;
301 continue;
302 }
303
304 switch (teql_resolve(skb, skb_res, slave)) {
305 case 0:
932ff279 306 if (netif_tx_trylock(slave)) {
1da177e4 307 if (!netif_queue_stopped(slave) &&
f25f4e44 308 !netif_subqueue_stopped(slave, subq) &&
1da177e4 309 slave->hard_start_xmit(skb, slave) == 0) {
932ff279 310 netif_tx_unlock(slave);
1da177e4
LT
311 master->slaves = NEXT_SLAVE(q);
312 netif_wake_queue(dev);
313 master->stats.tx_packets++;
314 master->stats.tx_bytes += len;
315 return 0;
316 }
932ff279 317 netif_tx_unlock(slave);
1da177e4
LT
318 }
319 if (netif_queue_stopped(dev))
320 busy = 1;
321 break;
322 case 1:
323 master->slaves = NEXT_SLAVE(q);
324 return 0;
325 default:
326 nores = 1;
327 break;
328 }
bbe735e4 329 __skb_pull(skb, skb_network_offset(skb));
1da177e4
LT
330 } while ((q = NEXT_SLAVE(q)) != start);
331
332 if (nores && skb_res == NULL) {
333 skb_res = skb;
334 goto restart;
335 }
336
337 if (busy) {
338 netif_stop_queue(dev);
339 return 1;
340 }
341 master->stats.tx_errors++;
342
343drop:
344 master->stats.tx_dropped++;
345 dev_kfree_skb(skb);
346 return 0;
347}
348
349static int teql_master_open(struct net_device *dev)
350{
351 struct Qdisc * q;
2941a486 352 struct teql_master *m = netdev_priv(dev);
1da177e4
LT
353 int mtu = 0xFFFE;
354 unsigned flags = IFF_NOARP|IFF_MULTICAST;
355
356 if (m->slaves == NULL)
357 return -EUNATCH;
358
359 flags = FMASK;
360
361 q = m->slaves;
362 do {
363 struct net_device *slave = q->dev;
364
365 if (slave == NULL)
366 return -EUNATCH;
367
368 if (slave->mtu < mtu)
369 mtu = slave->mtu;
370 if (slave->hard_header_len > LL_MAX_HEADER)
371 return -EINVAL;
372
373 /* If all the slaves are BROADCAST, master is BROADCAST
374 If all the slaves are PtP, master is PtP
375 Otherwise, master is NBMA.
376 */
377 if (!(slave->flags&IFF_POINTOPOINT))
378 flags &= ~IFF_POINTOPOINT;
379 if (!(slave->flags&IFF_BROADCAST))
380 flags &= ~IFF_BROADCAST;
381 if (!(slave->flags&IFF_MULTICAST))
382 flags &= ~IFF_MULTICAST;
383 } while ((q = NEXT_SLAVE(q)) != m->slaves);
384
385 m->dev->mtu = mtu;
386 m->dev->flags = (m->dev->flags&~FMASK) | flags;
387 netif_start_queue(m->dev);
388 return 0;
389}
390
391static int teql_master_close(struct net_device *dev)
392{
393 netif_stop_queue(dev);
394 return 0;
395}
396
397static struct net_device_stats *teql_master_stats(struct net_device *dev)
398{
2941a486 399 struct teql_master *m = netdev_priv(dev);
1da177e4
LT
400 return &m->stats;
401}
402
403static int teql_master_mtu(struct net_device *dev, int new_mtu)
404{
2941a486 405 struct teql_master *m = netdev_priv(dev);
1da177e4
LT
406 struct Qdisc *q;
407
408 if (new_mtu < 68)
409 return -EINVAL;
410
411 q = m->slaves;
412 if (q) {
413 do {
414 if (new_mtu > q->dev->mtu)
415 return -EINVAL;
416 } while ((q=NEXT_SLAVE(q)) != m->slaves);
417 }
418
419 dev->mtu = new_mtu;
420 return 0;
421}
422
423static __init void teql_master_setup(struct net_device *dev)
424{
2941a486 425 struct teql_master *master = netdev_priv(dev);
1da177e4
LT
426 struct Qdisc_ops *ops = &master->qops;
427
428 master->dev = dev;
429 ops->priv_size = sizeof(struct teql_sched_data);
10297b99 430
1da177e4
LT
431 ops->enqueue = teql_enqueue;
432 ops->dequeue = teql_dequeue;
433 ops->requeue = teql_requeue;
434 ops->init = teql_qdisc_init;
435 ops->reset = teql_reset;
436 ops->destroy = teql_destroy;
437 ops->owner = THIS_MODULE;
438
439 dev->open = teql_master_open;
440 dev->hard_start_xmit = teql_master_xmit;
441 dev->stop = teql_master_close;
442 dev->get_stats = teql_master_stats;
443 dev->change_mtu = teql_master_mtu;
444 dev->type = ARPHRD_VOID;
445 dev->mtu = 1500;
446 dev->tx_queue_len = 100;
447 dev->flags = IFF_NOARP;
448 dev->hard_header_len = LL_MAX_HEADER;
449 SET_MODULE_OWNER(dev);
450}
451
452static LIST_HEAD(master_dev_list);
453static int max_equalizers = 1;
454module_param(max_equalizers, int, 0);
455MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
456
457static int __init teql_init(void)
458{
459 int i;
460 int err = -ENODEV;
461
462 for (i = 0; i < max_equalizers; i++) {
463 struct net_device *dev;
464 struct teql_master *master;
465
466 dev = alloc_netdev(sizeof(struct teql_master),
467 "teql%d", teql_master_setup);
468 if (!dev) {
469 err = -ENOMEM;
470 break;
471 }
472
473 if ((err = register_netdev(dev))) {
474 free_netdev(dev);
475 break;
476 }
477
2941a486 478 master = netdev_priv(dev);
1da177e4
LT
479
480 strlcpy(master->qops.id, dev->name, IFNAMSIZ);
481 err = register_qdisc(&master->qops);
482
483 if (err) {
484 unregister_netdev(dev);
485 free_netdev(dev);
486 break;
487 }
488
489 list_add_tail(&master->master_list, &master_dev_list);
490 }
491 return i ? 0 : err;
492}
493
10297b99 494static void __exit teql_exit(void)
1da177e4
LT
495{
496 struct teql_master *master, *nxt;
497
498 list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) {
499
500 list_del(&master->master_list);
501
502 unregister_qdisc(&master->qops);
503 unregister_netdev(master->dev);
504 free_netdev(master->dev);
505 }
506}
507
508module_init(teql_init);
509module_exit(teql_exit);
510
511MODULE_LICENSE("GPL");