Merge branch 'upstream-linus' of master.kernel.org:/pub/scm/linux/kernel/git/jgarzik...
[linux-2.6] / net / sched / sch_teql.c
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 <linux/types.h>
13 #include <linux/kernel.h>
14 #include <linux/string.h>
15 #include <linux/errno.h>
16 #include <linux/if_arp.h>
17 #include <linux/netdevice.h>
18 #include <linux/init.h>
19 #include <linux/skbuff.h>
20 #include <linux/moduleparam.h>
21 #include <net/dst.h>
22 #include <net/neighbour.h>
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
55 struct 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
64 struct 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
74 #define FMASK (IFF_BROADCAST|IFF_POINTOPOINT|IFF_BROADCAST)
75
76 /* "teql*" qdisc routines */
77
78 static int
79 teql_enqueue(struct sk_buff *skb, struct Qdisc* sch)
80 {
81         struct net_device *dev = sch->dev;
82         struct teql_sched_data *q = qdisc_priv(sch);
83
84         if (q->q.qlen < dev->tx_queue_len) {
85                 __skb_queue_tail(&q->q, skb);
86                 sch->bstats.bytes += skb->len;
87                 sch->bstats.packets++;
88                 return 0;
89         }
90
91         kfree_skb(skb);
92         sch->qstats.drops++;
93         return NET_XMIT_DROP;
94 }
95
96 static int
97 teql_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
106 static struct sk_buff *
107 teql_dequeue(struct Qdisc* sch)
108 {
109         struct teql_sched_data *dat = qdisc_priv(sch);
110         struct sk_buff *skb;
111
112         skb = __skb_dequeue(&dat->q);
113         if (skb == NULL) {
114                 struct net_device *m = dat->m->dev->qdisc->dev;
115                 if (m) {
116                         dat->m->slaves = sch;
117                         netif_wake_queue(m);
118                 }
119         }
120         sch->q.qlen = dat->q.qlen + dat->m->dev->qdisc->q.qlen;
121         return skb;
122 }
123
124 static __inline__ void
125 teql_neigh_release(struct neighbour *n)
126 {
127         if (n)
128                 neigh_release(n);
129 }
130
131 static void
132 teql_reset(struct Qdisc* sch)
133 {
134         struct teql_sched_data *dat = qdisc_priv(sch);
135
136         skb_queue_purge(&dat->q);
137         sch->q.qlen = 0;
138         teql_neigh_release(xchg(&dat->ncache, NULL));
139 }
140
141 static void
142 teql_destroy(struct Qdisc* sch)
143 {
144         struct Qdisc *q, *prev;
145         struct teql_sched_data *dat = qdisc_priv(sch);
146         struct teql_master *master = dat->m;
147
148         if ((prev = master->slaves) != NULL) {
149                 do {
150                         q = NEXT_SLAVE(prev);
151                         if (q == sch) {
152                                 NEXT_SLAVE(prev) = NEXT_SLAVE(q);
153                                 if (q == master->slaves) {
154                                         master->slaves = NEXT_SLAVE(q);
155                                         if (q == master->slaves) {
156                                                 master->slaves = NULL;
157                                                 spin_lock_bh(&master->dev->queue_lock);
158                                                 qdisc_reset(master->dev->qdisc);
159                                                 spin_unlock_bh(&master->dev->queue_lock);
160                                         }
161                                 }
162                                 skb_queue_purge(&dat->q);
163                                 teql_neigh_release(xchg(&dat->ncache, NULL));
164                                 break;
165                         }
166
167                 } while ((prev = q) != master->slaves);
168         }
169 }
170
171 static int teql_qdisc_init(struct Qdisc *sch, struct rtattr *opt)
172 {
173         struct net_device *dev = sch->dev;
174         struct teql_master *m = (struct teql_master*)sch->ops;
175         struct teql_sched_data *q = qdisc_priv(sch);
176
177         if (dev->hard_header_len > m->dev->hard_header_len)
178                 return -EINVAL;
179
180         if (m->dev == dev)
181                 return -ELOOP;
182
183         q->m = m;
184
185         skb_queue_head_init(&q->q);
186
187         if (m->slaves) {
188                 if (m->dev->flags & IFF_UP) {
189                         if ((m->dev->flags&IFF_POINTOPOINT && !(dev->flags&IFF_POINTOPOINT))
190                             || (m->dev->flags&IFF_BROADCAST && !(dev->flags&IFF_BROADCAST))
191                             || (m->dev->flags&IFF_MULTICAST && !(dev->flags&IFF_MULTICAST))
192                             || dev->mtu < m->dev->mtu)
193                                 return -EINVAL;
194                 } else {
195                         if (!(dev->flags&IFF_POINTOPOINT))
196                                 m->dev->flags &= ~IFF_POINTOPOINT;
197                         if (!(dev->flags&IFF_BROADCAST))
198                                 m->dev->flags &= ~IFF_BROADCAST;
199                         if (!(dev->flags&IFF_MULTICAST))
200                                 m->dev->flags &= ~IFF_MULTICAST;
201                         if (dev->mtu < m->dev->mtu)
202                                 m->dev->mtu = dev->mtu;
203                 }
204                 q->next = NEXT_SLAVE(m->slaves);
205                 NEXT_SLAVE(m->slaves) = sch;
206         } else {
207                 q->next = sch;
208                 m->slaves = sch;
209                 m->dev->mtu = dev->mtu;
210                 m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
211         }
212         return 0;
213 }
214
215
216 static int
217 __teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, struct net_device *dev)
218 {
219         struct teql_sched_data *q = qdisc_priv(dev->qdisc);
220         struct neighbour *mn = skb->dst->neighbour;
221         struct neighbour *n = q->ncache;
222
223         if (mn->tbl == NULL)
224                 return -EINVAL;
225         if (n && n->tbl == mn->tbl &&
226             memcmp(n->primary_key, mn->primary_key, mn->tbl->key_len) == 0) {
227                 atomic_inc(&n->refcnt);
228         } else {
229                 n = __neigh_lookup_errno(mn->tbl, mn->primary_key, dev);
230                 if (IS_ERR(n))
231                         return PTR_ERR(n);
232         }
233         if (neigh_event_send(n, skb_res) == 0) {
234                 int err;
235                 read_lock(&n->lock);
236                 err = dev->hard_header(skb, dev, ntohs(skb->protocol), n->ha, NULL, skb->len);
237                 read_unlock(&n->lock);
238                 if (err < 0) {
239                         neigh_release(n);
240                         return -EINVAL;
241                 }
242                 teql_neigh_release(xchg(&q->ncache, n));
243                 return 0;
244         }
245         neigh_release(n);
246         return (skb_res == NULL) ? -EAGAIN : 1;
247 }
248
249 static __inline__ int
250 teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, struct net_device *dev)
251 {
252         if (dev->hard_header == NULL ||
253             skb->dst == NULL ||
254             skb->dst->neighbour == NULL)
255                 return 0;
256         return __teql_resolve(skb, skb_res, dev);
257 }
258
259 static int teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
260 {
261         struct teql_master *master = netdev_priv(dev);
262         struct Qdisc *start, *q;
263         int busy;
264         int nores;
265         int len = skb->len;
266         int subq = skb->queue_mapping;
267         struct sk_buff *skb_res = NULL;
268
269         start = master->slaves;
270
271 restart:
272         nores = 0;
273         busy = 0;
274
275         if ((q = start) == NULL)
276                 goto drop;
277
278         do {
279                 struct net_device *slave = q->dev;
280
281                 if (slave->qdisc_sleeping != q)
282                         continue;
283                 if (netif_queue_stopped(slave) ||
284                     netif_subqueue_stopped(slave, subq) ||
285                     !netif_running(slave)) {
286                         busy = 1;
287                         continue;
288                 }
289
290                 switch (teql_resolve(skb, skb_res, slave)) {
291                 case 0:
292                         if (netif_tx_trylock(slave)) {
293                                 if (!netif_queue_stopped(slave) &&
294                                     !netif_subqueue_stopped(slave, subq) &&
295                                     slave->hard_start_xmit(skb, slave) == 0) {
296                                         netif_tx_unlock(slave);
297                                         master->slaves = NEXT_SLAVE(q);
298                                         netif_wake_queue(dev);
299                                         master->stats.tx_packets++;
300                                         master->stats.tx_bytes += len;
301                                         return 0;
302                                 }
303                                 netif_tx_unlock(slave);
304                         }
305                         if (netif_queue_stopped(dev))
306                                 busy = 1;
307                         break;
308                 case 1:
309                         master->slaves = NEXT_SLAVE(q);
310                         return 0;
311                 default:
312                         nores = 1;
313                         break;
314                 }
315                 __skb_pull(skb, skb_network_offset(skb));
316         } while ((q = NEXT_SLAVE(q)) != start);
317
318         if (nores && skb_res == NULL) {
319                 skb_res = skb;
320                 goto restart;
321         }
322
323         if (busy) {
324                 netif_stop_queue(dev);
325                 return 1;
326         }
327         master->stats.tx_errors++;
328
329 drop:
330         master->stats.tx_dropped++;
331         dev_kfree_skb(skb);
332         return 0;
333 }
334
335 static int teql_master_open(struct net_device *dev)
336 {
337         struct Qdisc * q;
338         struct teql_master *m = netdev_priv(dev);
339         int mtu = 0xFFFE;
340         unsigned flags = IFF_NOARP|IFF_MULTICAST;
341
342         if (m->slaves == NULL)
343                 return -EUNATCH;
344
345         flags = FMASK;
346
347         q = m->slaves;
348         do {
349                 struct net_device *slave = q->dev;
350
351                 if (slave == NULL)
352                         return -EUNATCH;
353
354                 if (slave->mtu < mtu)
355                         mtu = slave->mtu;
356                 if (slave->hard_header_len > LL_MAX_HEADER)
357                         return -EINVAL;
358
359                 /* If all the slaves are BROADCAST, master is BROADCAST
360                    If all the slaves are PtP, master is PtP
361                    Otherwise, master is NBMA.
362                  */
363                 if (!(slave->flags&IFF_POINTOPOINT))
364                         flags &= ~IFF_POINTOPOINT;
365                 if (!(slave->flags&IFF_BROADCAST))
366                         flags &= ~IFF_BROADCAST;
367                 if (!(slave->flags&IFF_MULTICAST))
368                         flags &= ~IFF_MULTICAST;
369         } while ((q = NEXT_SLAVE(q)) != m->slaves);
370
371         m->dev->mtu = mtu;
372         m->dev->flags = (m->dev->flags&~FMASK) | flags;
373         netif_start_queue(m->dev);
374         return 0;
375 }
376
377 static int teql_master_close(struct net_device *dev)
378 {
379         netif_stop_queue(dev);
380         return 0;
381 }
382
383 static struct net_device_stats *teql_master_stats(struct net_device *dev)
384 {
385         struct teql_master *m = netdev_priv(dev);
386         return &m->stats;
387 }
388
389 static int teql_master_mtu(struct net_device *dev, int new_mtu)
390 {
391         struct teql_master *m = netdev_priv(dev);
392         struct Qdisc *q;
393
394         if (new_mtu < 68)
395                 return -EINVAL;
396
397         q = m->slaves;
398         if (q) {
399                 do {
400                         if (new_mtu > q->dev->mtu)
401                                 return -EINVAL;
402                 } while ((q=NEXT_SLAVE(q)) != m->slaves);
403         }
404
405         dev->mtu = new_mtu;
406         return 0;
407 }
408
409 static __init void teql_master_setup(struct net_device *dev)
410 {
411         struct teql_master *master = netdev_priv(dev);
412         struct Qdisc_ops *ops = &master->qops;
413
414         master->dev     = dev;
415         ops->priv_size  = sizeof(struct teql_sched_data);
416
417         ops->enqueue    =       teql_enqueue;
418         ops->dequeue    =       teql_dequeue;
419         ops->requeue    =       teql_requeue;
420         ops->init       =       teql_qdisc_init;
421         ops->reset      =       teql_reset;
422         ops->destroy    =       teql_destroy;
423         ops->owner      =       THIS_MODULE;
424
425         dev->open               = teql_master_open;
426         dev->hard_start_xmit    = teql_master_xmit;
427         dev->stop               = teql_master_close;
428         dev->get_stats          = teql_master_stats;
429         dev->change_mtu         = teql_master_mtu;
430         dev->type               = ARPHRD_VOID;
431         dev->mtu                = 1500;
432         dev->tx_queue_len       = 100;
433         dev->flags              = IFF_NOARP;
434         dev->hard_header_len    = LL_MAX_HEADER;
435         SET_MODULE_OWNER(dev);
436 }
437
438 static LIST_HEAD(master_dev_list);
439 static int max_equalizers = 1;
440 module_param(max_equalizers, int, 0);
441 MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
442
443 static int __init teql_init(void)
444 {
445         int i;
446         int err = -ENODEV;
447
448         for (i = 0; i < max_equalizers; i++) {
449                 struct net_device *dev;
450                 struct teql_master *master;
451
452                 dev = alloc_netdev(sizeof(struct teql_master),
453                                   "teql%d", teql_master_setup);
454                 if (!dev) {
455                         err = -ENOMEM;
456                         break;
457                 }
458
459                 if ((err = register_netdev(dev))) {
460                         free_netdev(dev);
461                         break;
462                 }
463
464                 master = netdev_priv(dev);
465
466                 strlcpy(master->qops.id, dev->name, IFNAMSIZ);
467                 err = register_qdisc(&master->qops);
468
469                 if (err) {
470                         unregister_netdev(dev);
471                         free_netdev(dev);
472                         break;
473                 }
474
475                 list_add_tail(&master->master_list, &master_dev_list);
476         }
477         return i ? 0 : err;
478 }
479
480 static void __exit teql_exit(void)
481 {
482         struct teql_master *master, *nxt;
483
484         list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) {
485
486                 list_del(&master->master_list);
487
488                 unregister_qdisc(&master->qops);
489                 unregister_netdev(master->dev);
490                 free_netdev(master->dev);
491         }
492 }
493
494 module_init(teql_init);
495 module_exit(teql_exit);
496
497 MODULE_LICENSE("GPL");