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
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 < READ_ONCE(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 *
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 *
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
122 teql_reset(struct Qdisc *sch)
123 {
124 struct teql_sched_data *dat = qdisc_priv(sch);
125
126 skb_queue_purge(&dat->q);
127 }
128
129 static void
130 teql_destroy(struct Qdisc *sch)
131 {
132 struct Qdisc *q, *prev;
133 struct teql_sched_data *dat = qdisc_priv(sch);
134 struct teql_master *master = dat->m;
135
136 if (!master)
137 return;
138
139 prev = master->slaves;
140 if (prev) {
141 do {
142 q = NEXT_SLAVE(prev);
143 if (q == sch) {
144 NEXT_SLAVE(prev) = NEXT_SLAVE(q);
145 if (q == master->slaves) {
146 master->slaves = NEXT_SLAVE(q);
147 if (q == master->slaves) {
148 struct netdev_queue *txq;
149 spinlock_t *root_lock;
150
151 txq = netdev_get_tx_queue(master->dev, 0);
152 master->slaves = NULL;
153
154 root_lock = qdisc_root_sleeping_lock(rtnl_dereference(txq->qdisc));
155 spin_lock_bh(root_lock);
156 qdisc_reset(rtnl_dereference(txq->qdisc));
157 spin_unlock_bh(root_lock);
158 }
159 }
160 skb_queue_purge(&dat->q);
161 break;
162 }
163
164 } while ((prev = q) != master->slaves);
165 }
166 }
167
168 static int teql_qdisc_init(struct Qdisc *sch, struct nlattr *opt,
169 struct netlink_ext_ack *extack)
170 {
171 struct net_device *dev = qdisc_dev(sch);
172 struct teql_master *m = (struct teql_master *)sch->ops;
173 struct teql_sched_data *q = qdisc_priv(sch);
174
175 if (dev->hard_header_len > m->dev->hard_header_len)
176 return -EINVAL;
177
178 if (m->dev == dev)
179 return -ELOOP;
180
181 q->m = m;
182
183 skb_queue_head_init(&q->q);
184
185 if (m->slaves) {
186 if (m->dev->flags & IFF_UP) {
187 if ((m->dev->flags & IFF_POINTOPOINT &&
188 !(dev->flags & IFF_POINTOPOINT)) ||
189 (m->dev->flags & IFF_BROADCAST &&
190 !(dev->flags & IFF_BROADCAST)) ||
191 (m->dev->flags & IFF_MULTICAST &&
192 !(dev->flags & IFF_MULTICAST)) ||
193 dev->mtu < m->dev->mtu)
194 return -EINVAL;
195 } else {
196 if (!(dev->flags&IFF_POINTOPOINT))
197 m->dev->flags &= ~IFF_POINTOPOINT;
198 if (!(dev->flags&IFF_BROADCAST))
199 m->dev->flags &= ~IFF_BROADCAST;
200 if (!(dev->flags&IFF_MULTICAST))
201 m->dev->flags &= ~IFF_MULTICAST;
202 if (dev->mtu < m->dev->mtu)
203 m->dev->mtu = dev->mtu;
204 }
205 q->next = NEXT_SLAVE(m->slaves);
206 NEXT_SLAVE(m->slaves) = sch;
207 } else {
208 q->next = sch;
209 m->slaves = sch;
210 m->dev->mtu = dev->mtu;
211 m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
212 }
213 return 0;
214 }
215
216
217 static int
218 __teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res,
219 struct net_device *dev, struct netdev_queue *txq,
220 struct dst_entry *dst)
221 {
222 struct neighbour *n;
223 int err = 0;
224
225 n = dst_neigh_lookup_skb(dst, skb);
226 if (!n)
227 return -ENOENT;
228
229 if (dst->dev != dev) {
230 struct neighbour *mn;
231
232 mn = __neigh_lookup_errno(n->tbl, n->primary_key, dev);
233 neigh_release(n);
234 if (IS_ERR(mn))
235 return PTR_ERR(mn);
236 n = mn;
237 }
238
239 if (neigh_event_send(n, skb_res) == 0) {
240 int err;
241 char haddr[MAX_ADDR_LEN];
242
243 neigh_ha_snapshot(haddr, n, dev);
244 err = dev_hard_header(skb, dev, ntohs(skb_protocol(skb, false)),
245 haddr, NULL, skb->len);
246
247 if (err < 0)
248 err = -EINVAL;
249 } else {
250 err = (skb_res == NULL) ? -EAGAIN : 1;
251 }
252 neigh_release(n);
253 return err;
254 }
255
256 static inline int teql_resolve(struct sk_buff *skb,
257 struct sk_buff *skb_res,
258 struct net_device *dev,
259 struct netdev_queue *txq)
260 {
261 struct dst_entry *dst = skb_dst(skb);
262 int res;
263
264 if (rcu_access_pointer(txq->qdisc) == &noop_qdisc)
265 return -ENODEV;
266
267 if (!dev->header_ops || !dst)
268 return 0;
269
270 rcu_read_lock();
271 res = __teql_resolve(skb, skb_res, dev, txq, dst);
272 rcu_read_unlock();
273
274 return res;
275 }
276
277 static netdev_tx_t teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
278 {
279 struct teql_master *master = netdev_priv(dev);
280 struct Qdisc *start, *q;
281 int busy;
282 int nores;
283 int subq = skb_get_queue_mapping(skb);
284 struct sk_buff *skb_res = NULL;
285
286 start = master->slaves;
287
288 restart:
289 nores = 0;
290 busy = 0;
291
292 q = start;
293 if (!q)
294 goto drop;
295
296 do {
297 struct net_device *slave = qdisc_dev(q);
298 struct netdev_queue *slave_txq = netdev_get_tx_queue(slave, 0);
299
300 if (rcu_access_pointer(slave_txq->qdisc_sleeping) != q)
301 continue;
302 if (netif_xmit_stopped(netdev_get_tx_queue(slave, subq)) ||
303 !netif_running(slave)) {
304 busy = 1;
305 continue;
306 }
307
308 switch (teql_resolve(skb, skb_res, slave, slave_txq)) {
309 case 0:
310 if (__netif_tx_trylock(slave_txq)) {
311 unsigned int length = qdisc_pkt_len(skb);
312
313 if (!netif_xmit_frozen_or_stopped(slave_txq) &&
314 netdev_start_xmit(skb, slave, slave_txq, false) ==
315 NETDEV_TX_OK) {
316 __netif_tx_unlock(slave_txq);
317 master->slaves = NEXT_SLAVE(q);
318 netif_wake_queue(dev);
319 master->tx_packets++;
320 master->tx_bytes += length;
321 return NETDEV_TX_OK;
322 }
323 __netif_tx_unlock(slave_txq);
324 }
325 if (netif_xmit_stopped(netdev_get_tx_queue(dev, 0)))
326 busy = 1;
327 break;
328 case 1:
329 master->slaves = NEXT_SLAVE(q);
330 return NETDEV_TX_OK;
331 default:
332 nores = 1;
333 break;
334 }
335 __skb_pull(skb, skb_network_offset(skb));
336 } while ((q = NEXT_SLAVE(q)) != start);
337
338 if (nores && skb_res == NULL) {
339 skb_res = skb;
340 goto restart;
341 }
342
343 if (busy) {
344 netif_stop_queue(dev);
345 return NETDEV_TX_BUSY;
346 }
347 master->tx_errors++;
348
349 drop:
350 master->tx_dropped++;
351 dev_kfree_skb(skb);
352 return NETDEV_TX_OK;
353 }
354
355 static int teql_master_open(struct net_device *dev)
356 {
357 struct Qdisc *q;
358 struct teql_master *m = netdev_priv(dev);
359 int mtu = 0xFFFE;
360 unsigned int flags = IFF_NOARP | IFF_MULTICAST;
361
362 if (m->slaves == NULL)
363 return -EUNATCH;
364
365 flags = FMASK;
366
367 q = m->slaves;
368 do {
369 struct net_device *slave = qdisc_dev(q);
370
371 if (slave == NULL)
372 return -EUNATCH;
373
374 if (slave->mtu < mtu)
375 mtu = slave->mtu;
376 if (slave->hard_header_len > LL_MAX_HEADER)
377 return -EINVAL;
378
379 /* If all the slaves are BROADCAST, master is BROADCAST
380 If all the slaves are PtP, master is PtP
381 Otherwise, master is NBMA.
382 */
383 if (!(slave->flags&IFF_POINTOPOINT))
384 flags &= ~IFF_POINTOPOINT;
385 if (!(slave->flags&IFF_BROADCAST))
386 flags &= ~IFF_BROADCAST;
387 if (!(slave->flags&IFF_MULTICAST))
388 flags &= ~IFF_MULTICAST;
389 } while ((q = NEXT_SLAVE(q)) != m->slaves);
390
391 m->dev->mtu = mtu;
392 m->dev->flags = (m->dev->flags&~FMASK) | flags;
393 netif_start_queue(m->dev);
394 return 0;
395 }
396
397 static int teql_master_close(struct net_device *dev)
398 {
399 netif_stop_queue(dev);
400 return 0;
401 }
402
403 static void teql_master_stats64(struct net_device *dev,
404 struct rtnl_link_stats64 *stats)
405 {
406 struct teql_master *m = netdev_priv(dev);
407
408 stats->tx_packets = m->tx_packets;
409 stats->tx_bytes = m->tx_bytes;
410 stats->tx_errors = m->tx_errors;
411 stats->tx_dropped = m->tx_dropped;
412 }
413
414 static int teql_master_mtu(struct net_device *dev, int new_mtu)
415 {
416 struct teql_master *m = netdev_priv(dev);
417 struct Qdisc *q;
418
419 q = m->slaves;
420 if (q) {
421 do {
422 if (new_mtu > qdisc_dev(q)->mtu)
423 return -EINVAL;
424 } while ((q = NEXT_SLAVE(q)) != m->slaves);
425 }
426
427 WRITE_ONCE(dev->mtu, new_mtu);
428 return 0;
429 }
430
431 static const struct net_device_ops teql_netdev_ops = {
432 .ndo_open = teql_master_open,
433 .ndo_stop = teql_master_close,
434 .ndo_start_xmit = teql_master_xmit,
435 .ndo_get_stats64 = teql_master_stats64,
436 .ndo_change_mtu = teql_master_mtu,
437 };
438
439 static __init void teql_master_setup(struct net_device *dev)
440 {
441 struct teql_master *master = netdev_priv(dev);
442 struct Qdisc_ops *ops = &master->qops;
443
444 master->dev = dev;
445 ops->priv_size = sizeof(struct teql_sched_data);
446
447 ops->enqueue = teql_enqueue;
448 ops->dequeue = teql_dequeue;
449 ops->peek = teql_peek;
450 ops->init = teql_qdisc_init;
451 ops->reset = teql_reset;
452 ops->destroy = teql_destroy;
453 ops->owner = THIS_MODULE;
454
455 dev->netdev_ops = &teql_netdev_ops;
456 dev->type = ARPHRD_VOID;
457 dev->mtu = 1500;
458 dev->min_mtu = 68;
459 dev->max_mtu = 65535;
460 dev->tx_queue_len = 100;
461 dev->flags = IFF_NOARP;
462 dev->hard_header_len = LL_MAX_HEADER;
463 netif_keep_dst(dev);
464 }
465
466 static LIST_HEAD(master_dev_list);
467 static int max_equalizers = 1;
468 module_param(max_equalizers, int, 0);
469 MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
470
471 static int __init teql_init(void)
472 {
473 int i;
474 int err = -ENODEV;
475
476 for (i = 0; i < max_equalizers; i++) {
477 struct net_device *dev;
478 struct teql_master *master;
479
480 dev = alloc_netdev(sizeof(struct teql_master), "teql%d",
481 NET_NAME_UNKNOWN, teql_master_setup);
482 if (!dev) {
483 err = -ENOMEM;
484 break;
485 }
486
487 if ((err = register_netdev(dev))) {
488 free_netdev(dev);
489 break;
490 }
491
492 master = netdev_priv(dev);
493
494 strscpy(master->qops.id, dev->name, IFNAMSIZ);
495 err = register_qdisc(&master->qops);
496
497 if (err) {
498 unregister_netdev(dev);
499 free_netdev(dev);
500 break;
501 }
502
503 list_add_tail(&master->master_list, &master_dev_list);
504 }
505 return i ? 0 : err;
506 }
507
508 static void __exit teql_exit(void)
509 {
510 struct teql_master *master, *nxt;
511
512 list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) {
513
514 list_del(&master->master_list);
515
516 unregister_qdisc(&master->qops);
517 unregister_netdev(master->dev);
518 free_netdev(master->dev);
519 }
520 }
521
522 module_init(teql_init);
523 module_exit(teql_exit);
524
525 MODULE_LICENSE("GPL");
526 MODULE_DESCRIPTION("True (or trivial) link equalizer qdisc");
527
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.