]> bbs.cooldavid.org Git - net-next-2.6.git/blame - net/sched/sch_teql.c
net: Use queue aware tests throughout.
[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>
1da177e4
LT
12#include <linux/types.h>
13#include <linux/kernel.h>
1da177e4 14#include <linux/string.h>
1da177e4 15#include <linux/errno.h>
14c85021 16#include <linux/if_arp.h>
1da177e4 17#include <linux/netdevice.h>
1da177e4 18#include <linux/init.h>
1da177e4
LT
19#include <linux/skbuff.h>
20#include <linux/moduleparam.h>
0ba48053
PM
21#include <net/dst.h>
22#include <net/neighbour.h>
1da177e4
LT
23#include <net/pkt_sched.h>
24
25/*
26 How to setup it.
27 ----------------
28
29 After loading this module you will find a new device teqlN
30 and new qdisc with the same name. To join a slave to the equalizer
31 you should just set this qdisc on a device f.e.
32
33 # tc qdisc add dev eth0 root teql0
34 # tc qdisc add dev eth1 root teql0
35
36 That's all. Full PnP 8)
37
38 Applicability.
39 --------------
40
41 1. Slave devices MUST be active devices, i.e., they must raise the tbusy
42 signal and generate EOI events. If you want to equalize virtual devices
43 like tunnels, use a normal eql device.
44 2. This device puts no limitations on physical slave characteristics
45 f.e. it will equalize 9600baud line and 100Mb ethernet perfectly :-)
46 Certainly, large difference in link speeds will make the resulting
47 eqalized link unusable, because of huge packet reordering.
48 I estimate an upper useful difference as ~10 times.
49 3. If the slave requires address resolution, only protocols using
50 neighbour cache (IPv4/IPv6) will work over the equalized link.
51 Other protocols are still allowed to use the slave device directly,
52 which will not break load balancing, though native slave
53 traffic will have the highest priority. */
54
55struct teql_master
56{
57 struct Qdisc_ops qops;
58 struct net_device *dev;
59 struct Qdisc *slaves;
60 struct list_head master_list;
61 struct net_device_stats stats;
62};
63
64struct teql_sched_data
65{
66 struct Qdisc *next;
67 struct teql_master *m;
68 struct neighbour *ncache;
69 struct sk_buff_head q;
70};
71
72#define NEXT_SLAVE(q) (((struct teql_sched_data*)qdisc_priv(q))->next)
73
cc8fd14d 74#define FMASK (IFF_BROADCAST|IFF_POINTOPOINT)
1da177e4
LT
75
76/* "teql*" qdisc routines */
77
78static int
79teql_enqueue(struct sk_buff *skb, struct Qdisc* sch)
80{
5ce2d488 81 struct net_device *dev = qdisc_dev(sch);
1da177e4
LT
82 struct teql_sched_data *q = qdisc_priv(sch);
83
4cd8c9e8
KK
84 if (q->q.qlen < dev->tx_queue_len) {
85 __skb_queue_tail(&q->q, skb);
1da177e4
LT
86 sch->bstats.bytes += skb->len;
87 sch->bstats.packets++;
88 return 0;
89 }
90
1da177e4
LT
91 kfree_skb(skb);
92 sch->qstats.drops++;
93 return NET_XMIT_DROP;
94}
95
96static int
97teql_requeue(struct sk_buff *skb, struct Qdisc* sch)
98{
99 struct teql_sched_data *q = qdisc_priv(sch);
100
101 __skb_queue_head(&q->q, skb);
102 sch->qstats.requeues++;
103 return 0;
104}
105
106static struct sk_buff *
107teql_dequeue(struct Qdisc* sch)
108{
109 struct teql_sched_data *dat = qdisc_priv(sch);
b0e1e646 110 struct netdev_queue *dat_queue;
1da177e4
LT
111 struct sk_buff *skb;
112
113 skb = __skb_dequeue(&dat->q);
e8a0464c 114 dat_queue = netdev_get_tx_queue(dat->m->dev, 0);
1da177e4 115 if (skb == NULL) {
b0e1e646 116 struct net_device *m = qdisc_dev(dat_queue->qdisc);
1da177e4
LT
117 if (m) {
118 dat->m->slaves = sch;
119 netif_wake_queue(m);
120 }
121 }
b0e1e646 122 sch->q.qlen = dat->q.qlen + dat_queue->qdisc->q.qlen;
1da177e4
LT
123 return skb;
124}
125
126static __inline__ void
127teql_neigh_release(struct neighbour *n)
128{
129 if (n)
130 neigh_release(n);
131}
132
133static void
134teql_reset(struct Qdisc* sch)
135{
136 struct teql_sched_data *dat = qdisc_priv(sch);
137
138 skb_queue_purge(&dat->q);
139 sch->q.qlen = 0;
140 teql_neigh_release(xchg(&dat->ncache, NULL));
141}
142
143static void
144teql_destroy(struct Qdisc* sch)
145{
146 struct Qdisc *q, *prev;
147 struct teql_sched_data *dat = qdisc_priv(sch);
148 struct teql_master *master = dat->m;
149
150 if ((prev = master->slaves) != NULL) {
151 do {
152 q = NEXT_SLAVE(prev);
153 if (q == sch) {
154 NEXT_SLAVE(prev) = NEXT_SLAVE(q);
155 if (q == master->slaves) {
156 master->slaves = NEXT_SLAVE(q);
157 if (q == master->slaves) {
e8a0464c
DM
158 struct netdev_queue *txq;
159
160 txq = netdev_get_tx_queue(master->dev, 0);
1da177e4 161 master->slaves = NULL;
e8a0464c
DM
162 spin_lock_bh(&txq->lock);
163 qdisc_reset(txq->qdisc);
164 spin_unlock_bh(&txq->lock);
1da177e4
LT
165 }
166 }
167 skb_queue_purge(&dat->q);
168 teql_neigh_release(xchg(&dat->ncache, NULL));
169 break;
170 }
10297b99 171
1da177e4
LT
172 } while ((prev = q) != master->slaves);
173 }
174}
175
1e90474c 176static int teql_qdisc_init(struct Qdisc *sch, struct nlattr *opt)
1da177e4 177{
5ce2d488 178 struct net_device *dev = qdisc_dev(sch);
1da177e4
LT
179 struct teql_master *m = (struct teql_master*)sch->ops;
180 struct teql_sched_data *q = qdisc_priv(sch);
181
182 if (dev->hard_header_len > m->dev->hard_header_len)
183 return -EINVAL;
184
185 if (m->dev == dev)
186 return -ELOOP;
187
188 q->m = m;
189
190 skb_queue_head_init(&q->q);
191
192 if (m->slaves) {
193 if (m->dev->flags & IFF_UP) {
194 if ((m->dev->flags&IFF_POINTOPOINT && !(dev->flags&IFF_POINTOPOINT))
195 || (m->dev->flags&IFF_BROADCAST && !(dev->flags&IFF_BROADCAST))
196 || (m->dev->flags&IFF_MULTICAST && !(dev->flags&IFF_MULTICAST))
197 || dev->mtu < m->dev->mtu)
198 return -EINVAL;
199 } else {
200 if (!(dev->flags&IFF_POINTOPOINT))
201 m->dev->flags &= ~IFF_POINTOPOINT;
202 if (!(dev->flags&IFF_BROADCAST))
203 m->dev->flags &= ~IFF_BROADCAST;
204 if (!(dev->flags&IFF_MULTICAST))
205 m->dev->flags &= ~IFF_MULTICAST;
206 if (dev->mtu < m->dev->mtu)
207 m->dev->mtu = dev->mtu;
208 }
209 q->next = NEXT_SLAVE(m->slaves);
210 NEXT_SLAVE(m->slaves) = sch;
211 } else {
212 q->next = sch;
213 m->slaves = sch;
214 m->dev->mtu = dev->mtu;
215 m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
216 }
217 return 0;
218}
219
1da177e4
LT
220
221static int
222__teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, struct net_device *dev)
223{
e8a0464c
DM
224 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, 0);
225 struct teql_sched_data *q = qdisc_priv(dev_queue->qdisc);
1da177e4
LT
226 struct neighbour *mn = skb->dst->neighbour;
227 struct neighbour *n = q->ncache;
228
229 if (mn->tbl == NULL)
230 return -EINVAL;
231 if (n && n->tbl == mn->tbl &&
232 memcmp(n->primary_key, mn->primary_key, mn->tbl->key_len) == 0) {
233 atomic_inc(&n->refcnt);
234 } else {
235 n = __neigh_lookup_errno(mn->tbl, mn->primary_key, dev);
236 if (IS_ERR(n))
237 return PTR_ERR(n);
238 }
239 if (neigh_event_send(n, skb_res) == 0) {
240 int err;
0c4e8581 241
1da177e4 242 read_lock(&n->lock);
0c4e8581
SH
243 err = dev_hard_header(skb, dev, ntohs(skb->protocol),
244 n->ha, NULL, skb->len);
1da177e4 245 read_unlock(&n->lock);
0c4e8581 246
1da177e4
LT
247 if (err < 0) {
248 neigh_release(n);
249 return -EINVAL;
250 }
251 teql_neigh_release(xchg(&q->ncache, n));
252 return 0;
253 }
254 neigh_release(n);
255 return (skb_res == NULL) ? -EAGAIN : 1;
256}
257
3b04ddde
SH
258static inline int teql_resolve(struct sk_buff *skb,
259 struct sk_buff *skb_res, struct net_device *dev)
1da177e4 260{
e8a0464c
DM
261 struct netdev_queue *txq = netdev_get_tx_queue(dev, 0);
262 if (txq->qdisc == &noop_qdisc)
4f9f8311
EP
263 return -ENODEV;
264
3b04ddde 265 if (dev->header_ops == NULL ||
1da177e4
LT
266 skb->dst == NULL ||
267 skb->dst->neighbour == NULL)
268 return 0;
269 return __teql_resolve(skb, skb_res, dev);
270}
271
272static int teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
273{
2941a486 274 struct teql_master *master = netdev_priv(dev);
1da177e4
LT
275 struct Qdisc *start, *q;
276 int busy;
277 int nores;
278 int len = skb->len;
4e3ab47a 279 int subq = skb_get_queue_mapping(skb);
1da177e4
LT
280 struct sk_buff *skb_res = NULL;
281
282 start = master->slaves;
283
284restart:
285 nores = 0;
286 busy = 0;
287
288 if ((q = start) == NULL)
289 goto drop;
290
291 do {
5ce2d488 292 struct net_device *slave = qdisc_dev(q);
e8a0464c 293 struct netdev_queue *slave_txq;
10297b99 294
e8a0464c
DM
295 slave_txq = netdev_get_tx_queue(slave, 0);
296 if (slave_txq->qdisc_sleeping != q)
1da177e4 297 continue;
fd2ea0a7 298 if (__netif_subqueue_stopped(slave, subq) ||
f25f4e44 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)) {
fd2ea0a7 307 if (!__netif_subqueue_stopped(slave, subq) &&
1da177e4 308 slave->hard_start_xmit(skb, slave) == 0) {
932ff279 309 netif_tx_unlock(slave);
1da177e4
LT
310 master->slaves = NEXT_SLAVE(q);
311 netif_wake_queue(dev);
312 master->stats.tx_packets++;
313 master->stats.tx_bytes += len;
314 return 0;
315 }
932ff279 316 netif_tx_unlock(slave);
1da177e4
LT
317 }
318 if (netif_queue_stopped(dev))
319 busy = 1;
320 break;
321 case 1:
322 master->slaves = NEXT_SLAVE(q);
323 return 0;
324 default:
325 nores = 1;
326 break;
327 }
bbe735e4 328 __skb_pull(skb, skb_network_offset(skb));
1da177e4
LT
329 } while ((q = NEXT_SLAVE(q)) != start);
330
331 if (nores && skb_res == NULL) {
332 skb_res = skb;
333 goto restart;
334 }
335
336 if (busy) {
337 netif_stop_queue(dev);
338 return 1;
339 }
340 master->stats.tx_errors++;
341
342drop:
343 master->stats.tx_dropped++;
344 dev_kfree_skb(skb);
345 return 0;
346}
347
348static int teql_master_open(struct net_device *dev)
349{
350 struct Qdisc * q;
2941a486 351 struct teql_master *m = netdev_priv(dev);
1da177e4
LT
352 int mtu = 0xFFFE;
353 unsigned flags = IFF_NOARP|IFF_MULTICAST;
354
355 if (m->slaves == NULL)
356 return -EUNATCH;
357
358 flags = FMASK;
359
360 q = m->slaves;
361 do {
5ce2d488 362 struct net_device *slave = qdisc_dev(q);
1da177e4
LT
363
364 if (slave == NULL)
365 return -EUNATCH;
366
367 if (slave->mtu < mtu)
368 mtu = slave->mtu;
369 if (slave->hard_header_len > LL_MAX_HEADER)
370 return -EINVAL;
371
372 /* If all the slaves are BROADCAST, master is BROADCAST
373 If all the slaves are PtP, master is PtP
374 Otherwise, master is NBMA.
375 */
376 if (!(slave->flags&IFF_POINTOPOINT))
377 flags &= ~IFF_POINTOPOINT;
378 if (!(slave->flags&IFF_BROADCAST))
379 flags &= ~IFF_BROADCAST;
380 if (!(slave->flags&IFF_MULTICAST))
381 flags &= ~IFF_MULTICAST;
382 } while ((q = NEXT_SLAVE(q)) != m->slaves);
383
384 m->dev->mtu = mtu;
385 m->dev->flags = (m->dev->flags&~FMASK) | flags;
386 netif_start_queue(m->dev);
387 return 0;
388}
389
390static int teql_master_close(struct net_device *dev)
391{
392 netif_stop_queue(dev);
393 return 0;
394}
395
396static struct net_device_stats *teql_master_stats(struct net_device *dev)
397{
2941a486 398 struct teql_master *m = netdev_priv(dev);
1da177e4
LT
399 return &m->stats;
400}
401
402static int teql_master_mtu(struct net_device *dev, int new_mtu)
403{
2941a486 404 struct teql_master *m = netdev_priv(dev);
1da177e4
LT
405 struct Qdisc *q;
406
407 if (new_mtu < 68)
408 return -EINVAL;
409
410 q = m->slaves;
411 if (q) {
412 do {
5ce2d488 413 if (new_mtu > qdisc_dev(q)->mtu)
1da177e4
LT
414 return -EINVAL;
415 } while ((q=NEXT_SLAVE(q)) != m->slaves);
416 }
417
418 dev->mtu = new_mtu;
419 return 0;
420}
421
422static __init void teql_master_setup(struct net_device *dev)
423{
2941a486 424 struct teql_master *master = netdev_priv(dev);
1da177e4
LT
425 struct Qdisc_ops *ops = &master->qops;
426
427 master->dev = dev;
428 ops->priv_size = sizeof(struct teql_sched_data);
10297b99 429
1da177e4
LT
430 ops->enqueue = teql_enqueue;
431 ops->dequeue = teql_dequeue;
432 ops->requeue = teql_requeue;
433 ops->init = teql_qdisc_init;
434 ops->reset = teql_reset;
435 ops->destroy = teql_destroy;
436 ops->owner = THIS_MODULE;
437
438 dev->open = teql_master_open;
439 dev->hard_start_xmit = teql_master_xmit;
440 dev->stop = teql_master_close;
441 dev->get_stats = teql_master_stats;
442 dev->change_mtu = teql_master_mtu;
443 dev->type = ARPHRD_VOID;
444 dev->mtu = 1500;
445 dev->tx_queue_len = 100;
446 dev->flags = IFF_NOARP;
447 dev->hard_header_len = LL_MAX_HEADER;
1da177e4
LT
448}
449
450static LIST_HEAD(master_dev_list);
451static int max_equalizers = 1;
452module_param(max_equalizers, int, 0);
453MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
454
455static int __init teql_init(void)
456{
457 int i;
458 int err = -ENODEV;
459
460 for (i = 0; i < max_equalizers; i++) {
461 struct net_device *dev;
462 struct teql_master *master;
463
464 dev = alloc_netdev(sizeof(struct teql_master),
465 "teql%d", teql_master_setup);
466 if (!dev) {
467 err = -ENOMEM;
468 break;
469 }
470
471 if ((err = register_netdev(dev))) {
472 free_netdev(dev);
473 break;
474 }
475
2941a486 476 master = netdev_priv(dev);
1da177e4
LT
477
478 strlcpy(master->qops.id, dev->name, IFNAMSIZ);
479 err = register_qdisc(&master->qops);
480
481 if (err) {
482 unregister_netdev(dev);
483 free_netdev(dev);
484 break;
485 }
486
487 list_add_tail(&master->master_list, &master_dev_list);
488 }
489 return i ? 0 : err;
490}
491
10297b99 492static void __exit teql_exit(void)
1da177e4
LT
493{
494 struct teql_master *master, *nxt;
495
496 list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) {
497
498 list_del(&master->master_list);
499
500 unregister_qdisc(&master->qops);
501 unregister_netdev(master->dev);
502 free_netdev(master->dev);
503 }
504}
505
506module_init(teql_init);
507module_exit(teql_exit);
508
509MODULE_LICENSE("GPL");