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 __rcu *slaves;
56 spinlock_t slaves_lock; /* serializes writes to ->slaves */
57 struct list_head master_list;
58 unsigned long tx_bytes;
59 unsigned long tx_packets;
60 unsigned long tx_errors;
61 unsigned long tx_dropped;
62 };
63
64 struct teql_sched_data {
65 struct Qdisc __rcu *next;
66 struct teql_master *m;
67 struct sk_buff_head q;
68 };
69
70 #define NEXT_SLAVE(q) (((struct teql_sched_data *)qdisc_priv(q))->next)
71
72 #define FMASK (IFF_BROADCAST | IFF_POINTOPOINT)
73
74 /* "teql*" qdisc routines */
75
76 static int
teql_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)77 teql_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
78 {
79 struct net_device *dev = qdisc_dev(sch);
80 struct teql_sched_data *q = qdisc_priv(sch);
81
82 if (q->q.qlen < READ_ONCE(dev->tx_queue_len)) {
83 __skb_queue_tail(&q->q, skb);
84 return NET_XMIT_SUCCESS;
85 }
86
87 return qdisc_drop(skb, sch, to_free);
88 }
89
90 static struct sk_buff *
teql_dequeue(struct Qdisc * sch)91 teql_dequeue(struct Qdisc *sch)
92 {
93 struct teql_sched_data *dat = qdisc_priv(sch);
94 struct netdev_queue *dat_queue;
95 struct sk_buff *skb;
96 struct Qdisc *q;
97
98 skb = __skb_dequeue(&dat->q);
99 dat_queue = netdev_get_tx_queue(dat->m->dev, 0);
100 q = rcu_dereference_bh(dat_queue->qdisc);
101
102 if (skb == NULL) {
103 struct net_device *m = qdisc_dev(q);
104 if (m) {
105 spin_lock_bh(&dat->m->slaves_lock);
106 rcu_assign_pointer(dat->m->slaves, sch);
107 spin_unlock_bh(&dat->m->slaves_lock);
108 netif_wake_queue(m);
109 }
110 } else {
111 qdisc_bstats_update(sch, skb);
112 }
113 WRITE_ONCE(sch->q.qlen, dat->q.qlen + READ_ONCE(q->q.qlen));
114 return skb;
115 }
116
117 static struct sk_buff *
teql_peek(struct Qdisc * sch)118 teql_peek(struct Qdisc *sch)
119 {
120 /* teql is meant to be used as root qdisc */
121 return NULL;
122 }
123
124 static void
teql_reset(struct Qdisc * sch)125 teql_reset(struct Qdisc *sch)
126 {
127 struct teql_sched_data *dat = qdisc_priv(sch);
128
129 skb_queue_purge(&dat->q);
130 }
131
132 static void
teql_destroy(struct Qdisc * sch)133 teql_destroy(struct Qdisc *sch)
134 {
135 struct Qdisc *q, *prev;
136 struct teql_sched_data *dat = qdisc_priv(sch);
137 struct teql_master *master = dat->m;
138 struct netdev_queue *txq = NULL;
139 bool reset_master_queue = false;
140
141 if (!master)
142 return;
143
144 spin_lock_bh(&master->slaves_lock);
145 prev = rcu_dereference_protected(master->slaves,
146 lockdep_is_held(&master->slaves_lock));
147 if (prev) {
148 do {
149 struct Qdisc *head, *next;
150
151 q = rcu_dereference_protected(NEXT_SLAVE(prev),
152 lockdep_is_held(&master->slaves_lock));
153 if (q != sch) {
154 prev = q;
155 continue;
156 }
157
158 next = rcu_dereference_protected(NEXT_SLAVE(q),
159 lockdep_is_held(&master->slaves_lock));
160 rcu_assign_pointer(NEXT_SLAVE(prev), next);
161
162 head = rcu_dereference_protected(master->slaves,
163 lockdep_is_held(&master->slaves_lock));
164 if (q == head) {
165 rcu_assign_pointer(master->slaves, next);
166 if (q == next) {
167 txq = netdev_get_tx_queue(master->dev, 0);
168 rcu_assign_pointer(master->slaves, NULL);
169 reset_master_queue = true;
170 }
171 }
172 skb_queue_purge(&dat->q);
173 break;
174 } while (prev != rcu_dereference_protected(master->slaves,
175 lockdep_is_held(&master->slaves_lock)));
176 }
177 spin_unlock_bh(&master->slaves_lock);
178
179 if (reset_master_queue)
180 dev_reset_queue(master->dev, txq, NULL);
181 }
182
teql_qdisc_init(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)183 static int teql_qdisc_init(struct Qdisc *sch, struct nlattr *opt,
184 struct netlink_ext_ack *extack)
185 {
186 struct net_device *dev = qdisc_dev(sch);
187 struct teql_master *m = (struct teql_master *)sch->ops;
188 struct teql_sched_data *q = qdisc_priv(sch);
189 struct Qdisc *first;
190
191 if (dev->hard_header_len > m->dev->hard_header_len)
192 return -EINVAL;
193
194 if (m->dev == dev)
195 return -ELOOP;
196
197 if (sch->parent != TC_H_ROOT) {
198 NL_SET_ERR_MSG_MOD(extack, "teql can only be used as root");
199 return -EOPNOTSUPP;
200 }
201
202 q->m = m;
203
204 skb_queue_head_init(&q->q);
205
206 spin_lock_bh(&m->slaves_lock);
207 first = rcu_dereference_protected(m->slaves, lockdep_is_held(&m->slaves_lock));
208 if (first) {
209 if (m->dev->flags & IFF_UP) {
210 if ((m->dev->flags & IFF_POINTOPOINT &&
211 !(dev->flags & IFF_POINTOPOINT)) ||
212 (m->dev->flags & IFF_BROADCAST &&
213 !(dev->flags & IFF_BROADCAST)) ||
214 (m->dev->flags & IFF_MULTICAST &&
215 !(dev->flags & IFF_MULTICAST)) ||
216 dev->mtu < m->dev->mtu) {
217 spin_unlock_bh(&m->slaves_lock);
218 return -EINVAL;
219 }
220 } else {
221 if (!(dev->flags&IFF_POINTOPOINT))
222 m->dev->flags &= ~IFF_POINTOPOINT;
223 if (!(dev->flags&IFF_BROADCAST))
224 m->dev->flags &= ~IFF_BROADCAST;
225 if (!(dev->flags&IFF_MULTICAST))
226 m->dev->flags &= ~IFF_MULTICAST;
227 if (dev->mtu < m->dev->mtu)
228 m->dev->mtu = dev->mtu;
229 }
230 rcu_assign_pointer(q->next,
231 rcu_dereference_protected(NEXT_SLAVE(first),
232 lockdep_is_held(&m->slaves_lock)));
233 rcu_assign_pointer(NEXT_SLAVE(first), sch);
234 } else {
235 rcu_assign_pointer(q->next, sch);
236 rcu_assign_pointer(m->slaves, sch);
237 m->dev->mtu = dev->mtu;
238 m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
239 }
240 spin_unlock_bh(&m->slaves_lock);
241 return 0;
242 }
243
244
245 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)246 __teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res,
247 struct net_device *dev, struct netdev_queue *txq,
248 struct dst_entry *dst)
249 {
250 struct neighbour *n;
251 int err = 0;
252
253 n = dst_neigh_lookup_skb(dst, skb);
254 if (!n)
255 return -ENOENT;
256
257 if (dst->dev != dev) {
258 struct neighbour *mn;
259
260 mn = __neigh_lookup_errno(n->tbl, n->primary_key, dev);
261 neigh_release(n);
262 if (IS_ERR(mn))
263 return PTR_ERR(mn);
264 n = mn;
265 }
266
267 if (neigh_event_send(n, skb_res) == 0) {
268 int err;
269 char haddr[MAX_ADDR_LEN];
270
271 neigh_ha_snapshot(haddr, n, dev);
272 err = dev_hard_header(skb, dev, ntohs(skb_protocol(skb, false)),
273 haddr, NULL, skb->len);
274
275 if (err < 0)
276 err = -EINVAL;
277 } else {
278 err = (skb_res == NULL) ? -EAGAIN : 1;
279 }
280 neigh_release(n);
281 return err;
282 }
283
teql_resolve(struct sk_buff * skb,struct sk_buff * skb_res,struct net_device * dev,struct netdev_queue * txq)284 static inline int teql_resolve(struct sk_buff *skb,
285 struct sk_buff *skb_res,
286 struct net_device *dev,
287 struct netdev_queue *txq)
288 {
289 struct dst_entry *dst = skb_dst(skb);
290 int res;
291
292 if (rcu_access_pointer(txq->qdisc) == &noop_qdisc)
293 return -ENODEV;
294
295 if (!dev->header_ops || !dst)
296 return 0;
297
298 rcu_read_lock();
299 res = __teql_resolve(skb, skb_res, dev, txq, dst);
300 rcu_read_unlock();
301
302 return res;
303 }
304
teql_master_xmit(struct sk_buff * skb,struct net_device * dev)305 static netdev_tx_t teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
306 {
307 struct teql_master *master = netdev_priv(dev);
308 struct Qdisc *start, *q;
309 int busy;
310 int nores;
311 int subq = skb_get_queue_mapping(skb);
312 struct sk_buff *skb_res = NULL;
313
314 rcu_read_lock_bh();
315
316 start = rcu_dereference_bh(master->slaves);
317
318 restart:
319 nores = 0;
320 busy = 0;
321
322 q = start;
323 if (!q)
324 goto drop;
325
326 do {
327 struct net_device *slave = qdisc_dev(q);
328 struct netdev_queue *slave_txq = netdev_get_tx_queue(slave, 0);
329
330 if (rcu_access_pointer(slave_txq->qdisc_sleeping) != q)
331 continue;
332 if (netif_xmit_stopped(netdev_get_tx_queue(slave, subq)) ||
333 !netif_running(slave)) {
334 busy = 1;
335 continue;
336 }
337
338 switch (teql_resolve(skb, skb_res, slave, slave_txq)) {
339 case 0:
340 if (__netif_tx_trylock(slave_txq)) {
341 unsigned int length = qdisc_pkt_len(skb);
342
343 skb->dev = slave;
344 if (!netif_xmit_frozen_or_stopped(slave_txq) &&
345 netdev_start_xmit(skb, slave, slave_txq, false) ==
346 NETDEV_TX_OK) {
347 __netif_tx_unlock(slave_txq);
348 spin_lock_bh(&master->slaves_lock);
349 if (rcu_dereference_protected(master->slaves,
350 lockdep_is_held(&master->slaves_lock)) == q)
351 rcu_assign_pointer(master->slaves,
352 rcu_dereference_protected(NEXT_SLAVE(q),
353 lockdep_is_held(&master->slaves_lock)));
354 spin_unlock_bh(&master->slaves_lock);
355 netif_wake_queue(dev);
356 master->tx_packets++;
357 master->tx_bytes += length;
358 rcu_read_unlock_bh();
359 return NETDEV_TX_OK;
360 }
361 __netif_tx_unlock(slave_txq);
362 }
363 if (netif_xmit_stopped(netdev_get_tx_queue(dev, 0)))
364 busy = 1;
365 break;
366 case 1:
367 spin_lock_bh(&master->slaves_lock);
368 if (rcu_dereference_protected(master->slaves,
369 lockdep_is_held(&master->slaves_lock)) == q)
370 rcu_assign_pointer(master->slaves,
371 rcu_dereference_protected(NEXT_SLAVE(q),
372 lockdep_is_held(&master->slaves_lock)));
373 spin_unlock_bh(&master->slaves_lock);
374 rcu_read_unlock_bh();
375 return NETDEV_TX_OK;
376 default:
377 nores = 1;
378 break;
379 }
380 __skb_pull(skb, skb_network_offset(skb));
381 } while ((q = rcu_dereference_bh(NEXT_SLAVE(q))) != start);
382
383 if (nores && skb_res == NULL) {
384 skb_res = skb;
385 goto restart;
386 }
387
388 if (busy) {
389 netif_stop_queue(dev);
390 rcu_read_unlock_bh();
391 return NETDEV_TX_BUSY;
392 }
393 master->tx_errors++;
394
395 drop:
396 master->tx_dropped++;
397 rcu_read_unlock_bh();
398 dev_kfree_skb(skb);
399 return NETDEV_TX_OK;
400 }
401
teql_master_open(struct net_device * dev)402 static int teql_master_open(struct net_device *dev)
403 {
404 struct Qdisc *q, *first;
405 struct teql_master *m = netdev_priv(dev);
406 int mtu = 0xFFFE;
407 unsigned int flags = IFF_NOARP | IFF_MULTICAST;
408
409 first = rtnl_dereference(m->slaves);
410 if (!first)
411 return -EUNATCH;
412
413 flags = FMASK;
414
415 q = first;
416 do {
417 struct net_device *slave = qdisc_dev(q);
418
419 if (slave == NULL)
420 return -EUNATCH;
421
422 if (slave->mtu < mtu)
423 mtu = slave->mtu;
424 if (slave->hard_header_len > LL_MAX_HEADER)
425 return -EINVAL;
426
427 /* If all the slaves are BROADCAST, master is BROADCAST
428 If all the slaves are PtP, master is PtP
429 Otherwise, master is NBMA.
430 */
431 if (!(slave->flags&IFF_POINTOPOINT))
432 flags &= ~IFF_POINTOPOINT;
433 if (!(slave->flags&IFF_BROADCAST))
434 flags &= ~IFF_BROADCAST;
435 if (!(slave->flags&IFF_MULTICAST))
436 flags &= ~IFF_MULTICAST;
437 } while ((q = rtnl_dereference(NEXT_SLAVE(q))) != first);
438
439 m->dev->mtu = mtu;
440 m->dev->flags = (m->dev->flags&~FMASK) | flags;
441 netif_start_queue(m->dev);
442 return 0;
443 }
444
teql_master_close(struct net_device * dev)445 static int teql_master_close(struct net_device *dev)
446 {
447 netif_stop_queue(dev);
448 return 0;
449 }
450
teql_master_stats64(struct net_device * dev,struct rtnl_link_stats64 * stats)451 static void teql_master_stats64(struct net_device *dev,
452 struct rtnl_link_stats64 *stats)
453 {
454 struct teql_master *m = netdev_priv(dev);
455
456 stats->tx_packets = m->tx_packets;
457 stats->tx_bytes = m->tx_bytes;
458 stats->tx_errors = m->tx_errors;
459 stats->tx_dropped = m->tx_dropped;
460 }
461
teql_master_mtu(struct net_device * dev,int new_mtu)462 static int teql_master_mtu(struct net_device *dev, int new_mtu)
463 {
464 struct teql_master *m = netdev_priv(dev);
465 struct Qdisc *q, *first;
466
467 first = rtnl_dereference(m->slaves);
468 q = first;
469 if (q) {
470 do {
471 if (new_mtu > qdisc_dev(q)->mtu)
472 return -EINVAL;
473 } while ((q = rtnl_dereference(NEXT_SLAVE(q))) != first);
474 }
475
476 WRITE_ONCE(dev->mtu, new_mtu);
477 return 0;
478 }
479
480 static const struct net_device_ops teql_netdev_ops = {
481 .ndo_open = teql_master_open,
482 .ndo_stop = teql_master_close,
483 .ndo_start_xmit = teql_master_xmit,
484 .ndo_get_stats64 = teql_master_stats64,
485 .ndo_change_mtu = teql_master_mtu,
486 };
487
teql_master_setup(struct net_device * dev)488 static __init void teql_master_setup(struct net_device *dev)
489 {
490 struct teql_master *master = netdev_priv(dev);
491 struct Qdisc_ops *ops = &master->qops;
492
493 spin_lock_init(&master->slaves_lock);
494 master->dev = dev;
495 ops->priv_size = sizeof(struct teql_sched_data);
496
497 ops->enqueue = teql_enqueue;
498 ops->dequeue = teql_dequeue;
499 ops->peek = teql_peek;
500 ops->init = teql_qdisc_init;
501 ops->reset = teql_reset;
502 ops->destroy = teql_destroy;
503 ops->owner = THIS_MODULE;
504
505 dev->netdev_ops = &teql_netdev_ops;
506 dev->type = ARPHRD_VOID;
507 dev->mtu = 1500;
508 dev->min_mtu = 68;
509 dev->max_mtu = 65535;
510 dev->tx_queue_len = 100;
511 dev->flags = IFF_NOARP;
512 dev->hard_header_len = LL_MAX_HEADER;
513 netif_keep_dst(dev);
514 }
515
516 static LIST_HEAD(master_dev_list);
517 static int max_equalizers = 1;
518 module_param(max_equalizers, int, 0);
519 MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
520
teql_init(void)521 static int __init teql_init(void)
522 {
523 int i;
524 int err = -ENODEV;
525
526 for (i = 0; i < max_equalizers; i++) {
527 struct net_device *dev;
528 struct teql_master *master;
529
530 dev = alloc_netdev(sizeof(struct teql_master), "teql%d",
531 NET_NAME_UNKNOWN, teql_master_setup);
532 if (!dev) {
533 err = -ENOMEM;
534 break;
535 }
536
537 if ((err = register_netdev(dev))) {
538 free_netdev(dev);
539 break;
540 }
541
542 master = netdev_priv(dev);
543
544 strscpy(master->qops.id, dev->name, IFNAMSIZ);
545 err = register_qdisc(&master->qops);
546
547 if (err) {
548 unregister_netdev(dev);
549 free_netdev(dev);
550 break;
551 }
552
553 list_add_tail(&master->master_list, &master_dev_list);
554 }
555 return i ? 0 : err;
556 }
557
teql_exit(void)558 static void __exit teql_exit(void)
559 {
560 struct teql_master *master, *nxt;
561
562 list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) {
563
564 list_del(&master->master_list);
565
566 unregister_qdisc(&master->qops);
567 unregister_netdev(master->dev);
568 free_netdev(master->dev);
569 }
570 }
571
572 module_init(teql_init);
573 module_exit(teql_exit);
574
575 MODULE_LICENSE("GPL");
576 MODULE_DESCRIPTION("True (or trivial) link equalizer qdisc");
577