net: txq_trans_update() helper
[linux-2.6.git] / 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 };
62
63 struct teql_sched_data
64 {
65         struct Qdisc *next;
66         struct teql_master *m;
67         struct neighbour *ncache;
68         struct sk_buff_head q;
69 };
70
71 #define NEXT_SLAVE(q) (((struct teql_sched_data*)qdisc_priv(q))->next)
72
73 #define FMASK (IFF_BROADCAST|IFF_POINTOPOINT)
74
75 /* "teql*" qdisc routines */
76
77 static int
78 teql_enqueue(struct sk_buff *skb, struct Qdisc* sch)
79 {
80         struct net_device *dev = qdisc_dev(sch);
81         struct teql_sched_data *q = qdisc_priv(sch);
82
83         if (q->q.qlen < dev->tx_queue_len) {
84                 __skb_queue_tail(&q->q, skb);
85                 sch->bstats.bytes += qdisc_pkt_len(skb);
86                 sch->bstats.packets++;
87                 return 0;
88         }
89
90         kfree_skb(skb);
91         sch->qstats.drops++;
92         return NET_XMIT_DROP;
93 }
94
95 static struct sk_buff *
96 teql_dequeue(struct Qdisc* sch)
97 {
98         struct teql_sched_data *dat = qdisc_priv(sch);
99         struct netdev_queue *dat_queue;
100         struct sk_buff *skb;
101
102         skb = __skb_dequeue(&dat->q);
103         dat_queue = netdev_get_tx_queue(dat->m->dev, 0);
104         if (skb == NULL) {
105                 struct net_device *m = qdisc_dev(dat_queue->qdisc);
106                 if (m) {
107                         dat->m->slaves = sch;
108                         netif_wake_queue(m);
109                 }
110         }
111         sch->q.qlen = dat->q.qlen + dat_queue->qdisc->q.qlen;
112         return skb;
113 }
114
115 static struct sk_buff *
116 teql_peek(struct Qdisc* sch)
117 {
118         /* teql is meant to be used as root qdisc */
119         return NULL;
120 }
121
122 static __inline__ void
123 teql_neigh_release(struct neighbour *n)
124 {
125         if (n)
126                 neigh_release(n);
127 }
128
129 static void
130 teql_reset(struct Qdisc* sch)
131 {
132         struct teql_sched_data *dat = qdisc_priv(sch);
133
134         skb_queue_purge(&dat->q);
135         sch->q.qlen = 0;
136         teql_neigh_release(xchg(&dat->ncache, NULL));
137 }
138
139 static void
140 teql_destroy(struct Qdisc* sch)
141 {
142         struct Qdisc *q, *prev;
143         struct teql_sched_data *dat = qdisc_priv(sch);
144         struct teql_master *master = dat->m;
145
146         if ((prev = master->slaves) != NULL) {
147                 do {
148                         q = NEXT_SLAVE(prev);
149                         if (q == sch) {
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;
156
157                                                 txq = netdev_get_tx_queue(master->dev, 0);
158                                                 master->slaves = NULL;
159
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);
164                                         }
165                                 }
166                                 skb_queue_purge(&dat->q);
167                                 teql_neigh_release(xchg(&dat->ncache, NULL));
168                                 break;
169                         }
170
171                 } while ((prev = q) != master->slaves);
172         }
173 }
174
175 static int teql_qdisc_init(struct Qdisc *sch, struct nlattr *opt)
176 {
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);
180
181         if (dev->hard_header_len > m->dev->hard_header_len)
182                 return -EINVAL;
183
184         if (m->dev == dev)
185                 return -ELOOP;
186
187         q->m = m;
188
189         skb_queue_head_init(&q->q);
190
191         if (m->slaves) {
192                 if (m->dev->flags & IFF_UP) {
193                         if ((m->dev->flags&IFF_POINTOPOINT && !(dev->flags&IFF_POINTOPOINT))
194                             || (m->dev->flags&IFF_BROADCAST && !(dev->flags&IFF_BROADCAST))
195                             || (m->dev->flags&IFF_MULTICAST && !(dev->flags&IFF_MULTICAST))
196                             || dev->mtu < m->dev->mtu)
197                                 return -EINVAL;
198                 } else {
199                         if (!(dev->flags&IFF_POINTOPOINT))
200                                 m->dev->flags &= ~IFF_POINTOPOINT;
201                         if (!(dev->flags&IFF_BROADCAST))
202                                 m->dev->flags &= ~IFF_BROADCAST;
203                         if (!(dev->flags&IFF_MULTICAST))
204                                 m->dev->flags &= ~IFF_MULTICAST;
205                         if (dev->mtu < m->dev->mtu)
206                                 m->dev->mtu = dev->mtu;
207                 }
208                 q->next = NEXT_SLAVE(m->slaves);
209                 NEXT_SLAVE(m->slaves) = sch;
210         } else {
211                 q->next = sch;
212                 m->slaves = sch;
213                 m->dev->mtu = dev->mtu;
214                 m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
215         }
216         return 0;
217 }
218
219
220 static int
221 __teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, struct net_device *dev)
222 {
223         struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, 0);
224         struct teql_sched_data *q = qdisc_priv(dev_queue->qdisc);
225         struct neighbour *mn = skb->dst->neighbour;
226         struct neighbour *n = q->ncache;
227
228         if (mn->tbl == NULL)
229                 return -EINVAL;
230         if (n && n->tbl == mn->tbl &&
231             memcmp(n->primary_key, mn->primary_key, mn->tbl->key_len) == 0) {
232                 atomic_inc(&n->refcnt);
233         } else {
234                 n = __neigh_lookup_errno(mn->tbl, mn->primary_key, dev);
235                 if (IS_ERR(n))
236                         return PTR_ERR(n);
237         }
238         if (neigh_event_send(n, skb_res) == 0) {
239                 int err;
240
241                 read_lock(&n->lock);
242                 err = dev_hard_header(skb, dev, ntohs(skb->protocol),
243                                       n->ha, NULL, skb->len);
244                 read_unlock(&n->lock);
245
246                 if (err < 0) {
247                         neigh_release(n);
248                         return -EINVAL;
249                 }
250                 teql_neigh_release(xchg(&q->ncache, n));
251                 return 0;
252         }
253         neigh_release(n);
254         return (skb_res == NULL) ? -EAGAIN : 1;
255 }
256
257 static inline int teql_resolve(struct sk_buff *skb,
258                                struct sk_buff *skb_res, struct net_device *dev)
259 {
260         struct netdev_queue *txq = netdev_get_tx_queue(dev, 0);
261         if (txq->qdisc == &noop_qdisc)
262                 return -ENODEV;
263
264         if (dev->header_ops == NULL ||
265             skb->dst == NULL ||
266             skb->dst->neighbour == NULL)
267                 return 0;
268         return __teql_resolve(skb, skb_res, dev);
269 }
270
271 static int teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
272 {
273         struct teql_master *master = netdev_priv(dev);
274         struct netdev_queue *txq = netdev_get_tx_queue(dev, 0);
275         struct Qdisc *start, *q;
276         int busy;
277         int nores;
278         int subq = skb_get_queue_mapping(skb);
279         struct sk_buff *skb_res = NULL;
280
281         start = master->slaves;
282
283 restart:
284         nores = 0;
285         busy = 0;
286
287         if ((q = start) == NULL)
288                 goto drop;
289
290         do {
291                 struct net_device *slave = qdisc_dev(q);
292                 struct netdev_queue *slave_txq = netdev_get_tx_queue(slave, 0);
293                 const struct net_device_ops *slave_ops = slave->netdev_ops;
294
295                 if (slave_txq->qdisc_sleeping != q)
296                         continue;
297                 if (__netif_subqueue_stopped(slave, subq) ||
298                     !netif_running(slave)) {
299                         busy = 1;
300                         continue;
301                 }
302
303                 switch (teql_resolve(skb, skb_res, slave)) {
304                 case 0:
305                         if (__netif_tx_trylock(slave_txq)) {
306                                 unsigned int length = qdisc_pkt_len(skb);
307
308                                 if (!netif_tx_queue_stopped(slave_txq) &&
309                                     !netif_tx_queue_frozen(slave_txq) &&
310                                     slave_ops->ndo_start_xmit(skb, slave) == 0) {
311                                         txq_trans_update(slave_txq);
312                                         __netif_tx_unlock(slave_txq);
313                                         master->slaves = NEXT_SLAVE(q);
314                                         netif_wake_queue(dev);
315                                         txq->tx_packets++;
316                                         txq->tx_bytes += length;
317                                         return 0;
318                                 }
319                                 __netif_tx_unlock(slave_txq);
320                         }
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;
330                 }
331                 __skb_pull(skb, skb_network_offset(skb));
332         } while ((q = NEXT_SLAVE(q)) != start);
333
334         if (nores && skb_res == NULL) {
335                 skb_res = skb;
336                 goto restart;
337         }
338
339         if (busy) {
340                 netif_stop_queue(dev);
341                 return 1;
342         }
343         dev->stats.tx_errors++;
344
345 drop:
346         txq->tx_dropped++;
347         dev_kfree_skb(skb);
348         return 0;
349 }
350
351 static int teql_master_open(struct net_device *dev)
352 {
353         struct Qdisc * q;
354         struct teql_master *m = netdev_priv(dev);
355         int mtu = 0xFFFE;
356         unsigned flags = IFF_NOARP|IFF_MULTICAST;
357
358         if (m->slaves == NULL)
359                 return -EUNATCH;
360
361         flags = FMASK;
362
363         q = m->slaves;
364         do {
365                 struct net_device *slave = qdisc_dev(q);
366
367                 if (slave == NULL)
368                         return -EUNATCH;
369
370                 if (slave->mtu < mtu)
371                         mtu = slave->mtu;
372                 if (slave->hard_header_len > LL_MAX_HEADER)
373                         return -EINVAL;
374
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.
378                  */
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);
386
387         m->dev->mtu = mtu;
388         m->dev->flags = (m->dev->flags&~FMASK) | flags;
389         netif_start_queue(m->dev);
390         return 0;
391 }
392
393 static int teql_master_close(struct net_device *dev)
394 {
395         netif_stop_queue(dev);
396         return 0;
397 }
398
399 static int teql_master_mtu(struct net_device *dev, int new_mtu)
400 {
401         struct teql_master *m = netdev_priv(dev);
402         struct Qdisc *q;
403
404         if (new_mtu < 68)
405                 return -EINVAL;
406
407         q = m->slaves;
408         if (q) {
409                 do {
410                         if (new_mtu > qdisc_dev(q)->mtu)
411                                 return -EINVAL;
412                 } while ((q=NEXT_SLAVE(q)) != m->slaves);
413         }
414
415         dev->mtu = new_mtu;
416         return 0;
417 }
418
419 static const struct net_device_ops teql_netdev_ops = {
420         .ndo_open       = teql_master_open,
421         .ndo_stop       = teql_master_close,
422         .ndo_start_xmit = teql_master_xmit,
423         .ndo_change_mtu = teql_master_mtu,
424 };
425
426 static __init void teql_master_setup(struct net_device *dev)
427 {
428         struct teql_master *master = netdev_priv(dev);
429         struct Qdisc_ops *ops = &master->qops;
430
431         master->dev     = dev;
432         ops->priv_size  = sizeof(struct teql_sched_data);
433
434         ops->enqueue    =       teql_enqueue;
435         ops->dequeue    =       teql_dequeue;
436         ops->peek       =       teql_peek;
437         ops->init       =       teql_qdisc_init;
438         ops->reset      =       teql_reset;
439         ops->destroy    =       teql_destroy;
440         ops->owner      =       THIS_MODULE;
441
442         dev->netdev_ops =       &teql_netdev_ops;
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;
448 }
449
450 static LIST_HEAD(master_dev_list);
451 static int max_equalizers = 1;
452 module_param(max_equalizers, int, 0);
453 MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
454
455 static 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
476                 master = netdev_priv(dev);
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
492 static void __exit teql_exit(void)
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
506 module_init(teql_init);
507 module_exit(teql_exit);
508
509 MODULE_LICENSE("GPL");