1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* net/sched/sch_teql.c "True" (or "trivial") link equalizer.
3 *
4 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
5 */
6
7 #include <linux/module.h>
8 #include <linux/types.h>
9 #include <linux/kernel.h>
10 #include <linux/slab.h>
11 #include <linux/string.h>
12 #include <linux/errno.h>
13 #include <linux/if_arp.h>
14 #include <linux/netdevice.h>
15 #include <linux/init.h>
16 #include <linux/skbuff.h>
17 #include <linux/moduleparam.h>
18 #include <net/dst.h>
19 #include <net/neighbour.h>
20 #include <net/pkt_sched.h>
21
22 /*
23 How to setup it.
24 ----------------
25
26 After loading this module you will find a new device teqlN
27 and new qdisc with the same name. To join a slave to the equalizer
28 you should just set this qdisc on a device f.e.
29
30 # tc qdisc add dev eth0 root teql0
31 # tc qdisc add dev eth1 root teql0
32
33 That's all. Full PnP 8)
34
35 Applicability.
36 --------------
37
38 1. Slave devices MUST be active devices, i.e., they must raise the tbusy
39 signal and generate EOI events. If you want to equalize virtual devices
40 like tunnels, use a normal eql device.
41 2. This device puts no limitations on physical slave characteristics
42 f.e. it will equalize 9600baud line and 100Mb ethernet perfectly :-)
43 Certainly, large difference in link speeds will make the resulting
44 eqalized link unusable, because of huge packet reordering.
45 I estimate an upper useful difference as ~10 times.
46 3. If the slave requires address resolution, only protocols using
47 neighbour cache (IPv4/IPv6) will work over the equalized link.
48 Other protocols are still allowed to use the slave device directly,
49 which will not break load balancing, though native slave
50 traffic will have the highest priority. */
51
52 struct teql_master {
53 struct Qdisc_ops qops;
54 struct net_device *dev;
55 struct Qdisc *slaves;
56 struct list_head master_list;
57 unsigned long tx_bytes;
58 unsigned long tx_packets;
59 unsigned long tx_errors;
60 unsigned long tx_dropped;
61 };
62
63 struct teql_sched_data {
64 struct Qdisc *next;
65 struct teql_master *m;
66 struct sk_buff_head q;
67 };
68
69 #define NEXT_SLAVE(q) (((struct teql_sched_data *)qdisc_priv(q))->next)
70
71 #define FMASK (IFF_BROADCAST | IFF_POINTOPOINT)
72
73 /* "teql*" qdisc routines */
74
75 static int
teql_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)76 teql_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
77 {
78 struct net_device *dev = qdisc_dev(sch);
79 struct teql_sched_data *q = qdisc_priv(sch);
80
81 if (q->q.qlen < dev->tx_queue_len) {
82 __skb_queue_tail(&q->q, skb);
83 return NET_XMIT_SUCCESS;
84 }
85
86 return qdisc_drop(skb, sch, to_free);
87 }
88
89 static struct sk_buff *
teql_dequeue(struct Qdisc * sch)90 teql_dequeue(struct Qdisc *sch)
91 {
92 struct teql_sched_data *dat = qdisc_priv(sch);
93 struct netdev_queue *dat_queue;
94 struct sk_buff *skb;
95 struct Qdisc *q;
96
97 skb = __skb_dequeue(&dat->q);
98 dat_queue = netdev_get_tx_queue(dat->m->dev, 0);
99 q = rcu_dereference_bh(dat_queue->qdisc);
100
101 if (skb == NULL) {
102 struct net_device *m = qdisc_dev(q);
103 if (m) {
104 dat->m->slaves = sch;
105 netif_wake_queue(m);
106 }
107 } else {
108 qdisc_bstats_update(sch, skb);
109 }
110 sch->q.qlen = dat->q.qlen + q->q.qlen;
111 return skb;
112 }
113
114 static struct sk_buff *
teql_peek(struct Qdisc * sch)115 teql_peek(struct Qdisc *sch)
116 {
117 /* teql is meant to be used as root qdisc */
118 return NULL;
119 }
120
121 static void
teql_reset(struct Qdisc * sch)122 teql_reset(struct Qdisc *sch)
123 {
124 struct teql_sched_data *dat = qdisc_priv(sch);
125
126 skb_queue_purge(&dat->q);
127 sch->q.qlen = 0;
128 }
129
130 static void
teql_destroy(struct Qdisc * sch)131 teql_destroy(struct Qdisc *sch)
132 {
133 struct Qdisc *q, *prev;
134 struct teql_sched_data *dat = qdisc_priv(sch);
135 struct teql_master *master = dat->m;
136
137 if (!master)
138 return;
139
140 prev = master->slaves;
141 if (prev) {
142 do {
143 q = NEXT_SLAVE(prev);
144 if (q == sch) {
145 NEXT_SLAVE(prev) = NEXT_SLAVE(q);
146 if (q == master->slaves) {
147 master->slaves = NEXT_SLAVE(q);
148 if (q == master->slaves) {
149 struct netdev_queue *txq;
150 spinlock_t *root_lock;
151
152 txq = netdev_get_tx_queue(master->dev, 0);
153 master->slaves = NULL;
154
155 root_lock = qdisc_root_sleeping_lock(rtnl_dereference(txq->qdisc));
156 spin_lock_bh(root_lock);
157 qdisc_reset(rtnl_dereference(txq->qdisc));
158 spin_unlock_bh(root_lock);
159 }
160 }
161 skb_queue_purge(&dat->q);
162 break;
163 }
164
165 } while ((prev = q) != master->slaves);
166 }
167 }
168
teql_qdisc_init(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)169 static int teql_qdisc_init(struct Qdisc *sch, struct nlattr *opt,
170 struct netlink_ext_ack *extack)
171 {
172 struct net_device *dev = qdisc_dev(sch);
173 struct teql_master *m = (struct teql_master *)sch->ops;
174 struct teql_sched_data *q = qdisc_priv(sch);
175
176 if (dev->hard_header_len > m->dev->hard_header_len)
177 return -EINVAL;
178
179 if (m->dev == dev)
180 return -ELOOP;
181
182 q->m = m;
183
184 skb_queue_head_init(&q->q);
185
186 if (m->slaves) {
187 if (m->dev->flags & IFF_UP) {
188 if ((m->dev->flags & IFF_POINTOPOINT &&
189 !(dev->flags & IFF_POINTOPOINT)) ||
190 (m->dev->flags & IFF_BROADCAST &&
191 !(dev->flags & IFF_BROADCAST)) ||
192 (m->dev->flags & IFF_MULTICAST &&
193 !(dev->flags & IFF_MULTICAST)) ||
194 dev->mtu < m->dev->mtu)
195 return -EINVAL;
196 } else {
197 if (!(dev->flags&IFF_POINTOPOINT))
198 m->dev->flags &= ~IFF_POINTOPOINT;
199 if (!(dev->flags&IFF_BROADCAST))
200 m->dev->flags &= ~IFF_BROADCAST;
201 if (!(dev->flags&IFF_MULTICAST))
202 m->dev->flags &= ~IFF_MULTICAST;
203 if (dev->mtu < m->dev->mtu)
204 m->dev->mtu = dev->mtu;
205 }
206 q->next = NEXT_SLAVE(m->slaves);
207 NEXT_SLAVE(m->slaves) = sch;
208 } else {
209 q->next = sch;
210 m->slaves = sch;
211 m->dev->mtu = dev->mtu;
212 m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
213 }
214 return 0;
215 }
216
217
218 static int
__teql_resolve(struct sk_buff * skb,struct sk_buff * skb_res,struct net_device * dev,struct netdev_queue * txq,struct dst_entry * dst)219 __teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res,
220 struct net_device *dev, struct netdev_queue *txq,
221 struct dst_entry *dst)
222 {
223 struct neighbour *n;
224 int err = 0;
225
226 n = dst_neigh_lookup_skb(dst, skb);
227 if (!n)
228 return -ENOENT;
229
230 if (dst->dev != dev) {
231 struct neighbour *mn;
232
233 mn = __neigh_lookup_errno(n->tbl, n->primary_key, dev);
234 neigh_release(n);
235 if (IS_ERR(mn))
236 return PTR_ERR(mn);
237 n = mn;
238 }
239
240 if (neigh_event_send(n, skb_res) == 0) {
241 int err;
242 char haddr[MAX_ADDR_LEN];
243
244 neigh_ha_snapshot(haddr, n, dev);
245 err = dev_hard_header(skb, dev, ntohs(skb_protocol(skb, false)),
246 haddr, NULL, skb->len);
247
248 if (err < 0)
249 err = -EINVAL;
250 } else {
251 err = (skb_res == NULL) ? -EAGAIN : 1;
252 }
253 neigh_release(n);
254 return err;
255 }
256
teql_resolve(struct sk_buff * skb,struct sk_buff * skb_res,struct net_device * dev,struct netdev_queue * txq)257 static inline int teql_resolve(struct sk_buff *skb,
258 struct sk_buff *skb_res,
259 struct net_device *dev,
260 struct netdev_queue *txq)
261 {
262 struct dst_entry *dst = skb_dst(skb);
263 int res;
264
265 if (rcu_access_pointer(txq->qdisc) == &noop_qdisc)
266 return -ENODEV;
267
268 if (!dev->header_ops || !dst)
269 return 0;
270
271 rcu_read_lock();
272 res = __teql_resolve(skb, skb_res, dev, txq, dst);
273 rcu_read_unlock();
274
275 return res;
276 }
277
teql_master_xmit(struct sk_buff * skb,struct net_device * dev)278 static netdev_tx_t teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
279 {
280 struct teql_master *master = netdev_priv(dev);
281 struct Qdisc *start, *q;
282 int busy;
283 int nores;
284 int subq = skb_get_queue_mapping(skb);
285 struct sk_buff *skb_res = NULL;
286
287 start = master->slaves;
288
289 restart:
290 nores = 0;
291 busy = 0;
292
293 q = start;
294 if (!q)
295 goto drop;
296
297 do {
298 struct net_device *slave = qdisc_dev(q);
299 struct netdev_queue *slave_txq = netdev_get_tx_queue(slave, 0);
300
301 if (slave_txq->qdisc_sleeping != q)
302 continue;
303 if (netif_xmit_stopped(netdev_get_tx_queue(slave, subq)) ||
304 !netif_running(slave)) {
305 busy = 1;
306 continue;
307 }
308
309 switch (teql_resolve(skb, skb_res, slave, slave_txq)) {
310 case 0:
311 if (__netif_tx_trylock(slave_txq)) {
312 unsigned int length = qdisc_pkt_len(skb);
313
314 if (!netif_xmit_frozen_or_stopped(slave_txq) &&
315 netdev_start_xmit(skb, slave, slave_txq, false) ==
316 NETDEV_TX_OK) {
317 __netif_tx_unlock(slave_txq);
318 master->slaves = NEXT_SLAVE(q);
319 netif_wake_queue(dev);
320 master->tx_packets++;
321 master->tx_bytes += length;
322 return NETDEV_TX_OK;
323 }
324 __netif_tx_unlock(slave_txq);
325 }
326 if (netif_xmit_stopped(netdev_get_tx_queue(dev, 0)))
327 busy = 1;
328 break;
329 case 1:
330 master->slaves = NEXT_SLAVE(q);
331 return NETDEV_TX_OK;
332 default:
333 nores = 1;
334 break;
335 }
336 __skb_pull(skb, skb_network_offset(skb));
337 } while ((q = NEXT_SLAVE(q)) != start);
338
339 if (nores && skb_res == NULL) {
340 skb_res = skb;
341 goto restart;
342 }
343
344 if (busy) {
345 netif_stop_queue(dev);
346 return NETDEV_TX_BUSY;
347 }
348 master->tx_errors++;
349
350 drop:
351 master->tx_dropped++;
352 dev_kfree_skb(skb);
353 return NETDEV_TX_OK;
354 }
355
teql_master_open(struct net_device * dev)356 static int teql_master_open(struct net_device *dev)
357 {
358 struct Qdisc *q;
359 struct teql_master *m = netdev_priv(dev);
360 int mtu = 0xFFFE;
361 unsigned int flags = IFF_NOARP | IFF_MULTICAST;
362
363 if (m->slaves == NULL)
364 return -EUNATCH;
365
366 flags = FMASK;
367
368 q = m->slaves;
369 do {
370 struct net_device *slave = qdisc_dev(q);
371
372 if (slave == NULL)
373 return -EUNATCH;
374
375 if (slave->mtu < mtu)
376 mtu = slave->mtu;
377 if (slave->hard_header_len > LL_MAX_HEADER)
378 return -EINVAL;
379
380 /* If all the slaves are BROADCAST, master is BROADCAST
381 If all the slaves are PtP, master is PtP
382 Otherwise, master is NBMA.
383 */
384 if (!(slave->flags&IFF_POINTOPOINT))
385 flags &= ~IFF_POINTOPOINT;
386 if (!(slave->flags&IFF_BROADCAST))
387 flags &= ~IFF_BROADCAST;
388 if (!(slave->flags&IFF_MULTICAST))
389 flags &= ~IFF_MULTICAST;
390 } while ((q = NEXT_SLAVE(q)) != m->slaves);
391
392 m->dev->mtu = mtu;
393 m->dev->flags = (m->dev->flags&~FMASK) | flags;
394 netif_start_queue(m->dev);
395 return 0;
396 }
397
teql_master_close(struct net_device * dev)398 static int teql_master_close(struct net_device *dev)
399 {
400 netif_stop_queue(dev);
401 return 0;
402 }
403
teql_master_stats64(struct net_device * dev,struct rtnl_link_stats64 * stats)404 static void teql_master_stats64(struct net_device *dev,
405 struct rtnl_link_stats64 *stats)
406 {
407 struct teql_master *m = netdev_priv(dev);
408
409 stats->tx_packets = m->tx_packets;
410 stats->tx_bytes = m->tx_bytes;
411 stats->tx_errors = m->tx_errors;
412 stats->tx_dropped = m->tx_dropped;
413 }
414
teql_master_mtu(struct net_device * dev,int new_mtu)415 static int teql_master_mtu(struct net_device *dev, int new_mtu)
416 {
417 struct teql_master *m = netdev_priv(dev);
418 struct Qdisc *q;
419
420 q = m->slaves;
421 if (q) {
422 do {
423 if (new_mtu > qdisc_dev(q)->mtu)
424 return -EINVAL;
425 } while ((q = NEXT_SLAVE(q)) != m->slaves);
426 }
427
428 dev->mtu = new_mtu;
429 return 0;
430 }
431
432 static const struct net_device_ops teql_netdev_ops = {
433 .ndo_open = teql_master_open,
434 .ndo_stop = teql_master_close,
435 .ndo_start_xmit = teql_master_xmit,
436 .ndo_get_stats64 = teql_master_stats64,
437 .ndo_change_mtu = teql_master_mtu,
438 };
439
teql_master_setup(struct net_device * dev)440 static __init void teql_master_setup(struct net_device *dev)
441 {
442 struct teql_master *master = netdev_priv(dev);
443 struct Qdisc_ops *ops = &master->qops;
444
445 master->dev = dev;
446 ops->priv_size = sizeof(struct teql_sched_data);
447
448 ops->enqueue = teql_enqueue;
449 ops->dequeue = teql_dequeue;
450 ops->peek = teql_peek;
451 ops->init = teql_qdisc_init;
452 ops->reset = teql_reset;
453 ops->destroy = teql_destroy;
454 ops->owner = THIS_MODULE;
455
456 dev->netdev_ops = &teql_netdev_ops;
457 dev->type = ARPHRD_VOID;
458 dev->mtu = 1500;
459 dev->min_mtu = 68;
460 dev->max_mtu = 65535;
461 dev->tx_queue_len = 100;
462 dev->flags = IFF_NOARP;
463 dev->hard_header_len = LL_MAX_HEADER;
464 netif_keep_dst(dev);
465 }
466
467 static LIST_HEAD(master_dev_list);
468 static int max_equalizers = 1;
469 module_param(max_equalizers, int, 0);
470 MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
471
teql_init(void)472 static int __init teql_init(void)
473 {
474 int i;
475 int err = -ENODEV;
476
477 for (i = 0; i < max_equalizers; i++) {
478 struct net_device *dev;
479 struct teql_master *master;
480
481 dev = alloc_netdev(sizeof(struct teql_master), "teql%d",
482 NET_NAME_UNKNOWN, teql_master_setup);
483 if (!dev) {
484 err = -ENOMEM;
485 break;
486 }
487
488 if ((err = register_netdev(dev))) {
489 free_netdev(dev);
490 break;
491 }
492
493 master = netdev_priv(dev);
494
495 strlcpy(master->qops.id, dev->name, IFNAMSIZ);
496 err = register_qdisc(&master->qops);
497
498 if (err) {
499 unregister_netdev(dev);
500 free_netdev(dev);
501 break;
502 }
503
504 list_add_tail(&master->master_list, &master_dev_list);
505 }
506 return i ? 0 : err;
507 }
508
teql_exit(void)509 static void __exit teql_exit(void)
510 {
511 struct teql_master *master, *nxt;
512
513 list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) {
514
515 list_del(&master->master_list);
516
517 unregister_qdisc(&master->qops);
518 unregister_netdev(master->dev);
519 free_netdev(master->dev);
520 }
521 }
522
523 module_init(teql_init);
524 module_exit(teql_exit);
525
526 MODULE_LICENSE("GPL");
527