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/slab.h>
15 #include <linux/string.h>
16 #include <linux/errno.h>
17 #include <linux/if_arp.h>
18 #include <linux/netdevice.h>
19 #include <linux/init.h>
20 #include <linux/skbuff.h>
21 #include <linux/moduleparam.h>
22 #include <net/dst.h>
23 #include <net/neighbour.h>
24 #include <net/pkt_sched.h>
25 
26 /*
27    How to setup it.
28    ----------------
29 
30    After loading this module you will find a new device teqlN
31    and new qdisc with the same name. To join a slave to the equalizer
32    you should just set this qdisc on a device f.e.
33 
34    # tc qdisc add dev eth0 root teql0
35    # tc qdisc add dev eth1 root teql0
36 
37    That's all. Full PnP 8)
38 
39    Applicability.
40    --------------
41 
42    1. Slave devices MUST be active devices, i.e., they must raise the tbusy
43       signal and generate EOI events. If you want to equalize virtual devices
44       like tunnels, use a normal eql device.
45    2. This device puts no limitations on physical slave characteristics
46       f.e. it will equalize 9600baud line and 100Mb ethernet perfectly :-)
47       Certainly, large difference in link speeds will make the resulting
48       eqalized link unusable, because of huge packet reordering.
49       I estimate an upper useful difference as ~10 times.
50    3. If the slave requires address resolution, only protocols using
51       neighbour cache (IPv4/IPv6) will work over the equalized link.
52       Other protocols are still allowed to use the slave device directly,
53       which will not break load balancing, though native slave
54       traffic will have the highest priority.  */
55 
56 struct teql_master {
57 	struct Qdisc_ops qops;
58 	struct net_device *dev;
59 	struct Qdisc *slaves;
60 	struct list_head master_list;
61 	unsigned long	tx_bytes;
62 	unsigned long	tx_packets;
63 	unsigned long	tx_errors;
64 	unsigned long	tx_dropped;
65 };
66 
67 struct teql_sched_data {
68 	struct Qdisc *next;
69 	struct teql_master *m;
70 	struct neighbour *ncache;
71 	struct sk_buff_head q;
72 };
73 
74 #define NEXT_SLAVE(q) (((struct teql_sched_data *)qdisc_priv(q))->next)
75 
76 #define FMASK (IFF_BROADCAST | IFF_POINTOPOINT)
77 
78 /* "teql*" qdisc routines */
79 
80 static int
teql_enqueue(struct sk_buff * skb,struct Qdisc * sch)81 teql_enqueue(struct sk_buff *skb, struct Qdisc *sch)
82 {
83 	struct net_device *dev = qdisc_dev(sch);
84 	struct teql_sched_data *q = qdisc_priv(sch);
85 
86 	if (q->q.qlen < dev->tx_queue_len) {
87 		__skb_queue_tail(&q->q, skb);
88 		return NET_XMIT_SUCCESS;
89 	}
90 
91 	kfree_skb(skb);
92 	sch->qstats.drops++;
93 	return NET_XMIT_DROP;
94 }
95 
96 static struct sk_buff *
teql_dequeue(struct Qdisc * sch)97 teql_dequeue(struct Qdisc *sch)
98 {
99 	struct teql_sched_data *dat = qdisc_priv(sch);
100 	struct netdev_queue *dat_queue;
101 	struct sk_buff *skb;
102 
103 	skb = __skb_dequeue(&dat->q);
104 	dat_queue = netdev_get_tx_queue(dat->m->dev, 0);
105 	if (skb == NULL) {
106 		struct net_device *m = qdisc_dev(dat_queue->qdisc);
107 		if (m) {
108 			dat->m->slaves = sch;
109 			netif_wake_queue(m);
110 		}
111 	} else {
112 		qdisc_bstats_update(sch, skb);
113 	}
114 	sch->q.qlen = dat->q.qlen + dat_queue->qdisc->q.qlen;
115 	return skb;
116 }
117 
118 static struct sk_buff *
teql_peek(struct Qdisc * sch)119 teql_peek(struct Qdisc *sch)
120 {
121 	/* teql is meant to be used as root qdisc */
122 	return NULL;
123 }
124 
125 static inline void
teql_neigh_release(struct neighbour * n)126 teql_neigh_release(struct neighbour *n)
127 {
128 	if (n)
129 		neigh_release(n);
130 }
131 
132 static void
teql_reset(struct Qdisc * sch)133 teql_reset(struct Qdisc *sch)
134 {
135 	struct teql_sched_data *dat = qdisc_priv(sch);
136 
137 	skb_queue_purge(&dat->q);
138 	sch->q.qlen = 0;
139 	teql_neigh_release(xchg(&dat->ncache, NULL));
140 }
141 
142 static void
teql_destroy(struct Qdisc * sch)143 teql_destroy(struct Qdisc *sch)
144 {
145 	struct Qdisc *q, *prev;
146 	struct teql_sched_data *dat = qdisc_priv(sch);
147 	struct teql_master *master = dat->m;
148 
149 	prev = master->slaves;
150 	if (prev) {
151 		do {
152 			q = NEXT_SLAVE(prev);
153 			if (q == sch) {
154 				NEXT_SLAVE(prev) = NEXT_SLAVE(q);
155 				if (q == master->slaves) {
156 					master->slaves = NEXT_SLAVE(q);
157 					if (q == master->slaves) {
158 						struct netdev_queue *txq;
159 						spinlock_t *root_lock;
160 
161 						txq = netdev_get_tx_queue(master->dev, 0);
162 						master->slaves = NULL;
163 
164 						root_lock = qdisc_root_sleeping_lock(txq->qdisc);
165 						spin_lock_bh(root_lock);
166 						qdisc_reset(txq->qdisc);
167 						spin_unlock_bh(root_lock);
168 					}
169 				}
170 				skb_queue_purge(&dat->q);
171 				teql_neigh_release(xchg(&dat->ncache, NULL));
172 				break;
173 			}
174 
175 		} while ((prev = q) != master->slaves);
176 	}
177 }
178 
teql_qdisc_init(struct Qdisc * sch,struct nlattr * opt)179 static int teql_qdisc_init(struct Qdisc *sch, struct nlattr *opt)
180 {
181 	struct net_device *dev = qdisc_dev(sch);
182 	struct teql_master *m = (struct teql_master *)sch->ops;
183 	struct teql_sched_data *q = qdisc_priv(sch);
184 
185 	if (dev->hard_header_len > m->dev->hard_header_len)
186 		return -EINVAL;
187 
188 	if (m->dev == dev)
189 		return -ELOOP;
190 
191 	q->m = m;
192 
193 	skb_queue_head_init(&q->q);
194 
195 	if (m->slaves) {
196 		if (m->dev->flags & IFF_UP) {
197 			if ((m->dev->flags & IFF_POINTOPOINT &&
198 			     !(dev->flags & IFF_POINTOPOINT)) ||
199 			    (m->dev->flags & IFF_BROADCAST &&
200 			     !(dev->flags & IFF_BROADCAST)) ||
201 			    (m->dev->flags & IFF_MULTICAST &&
202 			     !(dev->flags & IFF_MULTICAST)) ||
203 			    dev->mtu < m->dev->mtu)
204 				return -EINVAL;
205 		} else {
206 			if (!(dev->flags&IFF_POINTOPOINT))
207 				m->dev->flags &= ~IFF_POINTOPOINT;
208 			if (!(dev->flags&IFF_BROADCAST))
209 				m->dev->flags &= ~IFF_BROADCAST;
210 			if (!(dev->flags&IFF_MULTICAST))
211 				m->dev->flags &= ~IFF_MULTICAST;
212 			if (dev->mtu < m->dev->mtu)
213 				m->dev->mtu = dev->mtu;
214 		}
215 		q->next = NEXT_SLAVE(m->slaves);
216 		NEXT_SLAVE(m->slaves) = sch;
217 	} else {
218 		q->next = sch;
219 		m->slaves = sch;
220 		m->dev->mtu = dev->mtu;
221 		m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
222 	}
223 	return 0;
224 }
225 
226 
227 static int
__teql_resolve(struct sk_buff * skb,struct sk_buff * skb_res,struct net_device * dev,struct netdev_queue * txq,struct neighbour * mn)228 __teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res,
229 	       struct net_device *dev, struct netdev_queue *txq,
230 	       struct neighbour *mn)
231 {
232 	struct teql_sched_data *q = qdisc_priv(txq->qdisc);
233 	struct neighbour *n = q->ncache;
234 
235 	if (mn->tbl == NULL)
236 		return -EINVAL;
237 	if (n && n->tbl == mn->tbl &&
238 	    memcmp(n->primary_key, mn->primary_key, mn->tbl->key_len) == 0) {
239 		atomic_inc(&n->refcnt);
240 	} else {
241 		n = __neigh_lookup_errno(mn->tbl, mn->primary_key, dev);
242 		if (IS_ERR(n))
243 			return PTR_ERR(n);
244 	}
245 	if (neigh_event_send(n, skb_res) == 0) {
246 		int err;
247 		char haddr[MAX_ADDR_LEN];
248 
249 		neigh_ha_snapshot(haddr, n, dev);
250 		err = dev_hard_header(skb, dev, ntohs(skb->protocol), haddr,
251 				      NULL, skb->len);
252 
253 		if (err < 0) {
254 			neigh_release(n);
255 			return -EINVAL;
256 		}
257 		teql_neigh_release(xchg(&q->ncache, n));
258 		return 0;
259 	}
260 	neigh_release(n);
261 	return (skb_res == NULL) ? -EAGAIN : 1;
262 }
263 
teql_resolve(struct sk_buff * skb,struct sk_buff * skb_res,struct net_device * dev,struct netdev_queue * txq)264 static inline int teql_resolve(struct sk_buff *skb,
265 			       struct sk_buff *skb_res,
266 			       struct net_device *dev,
267 			       struct netdev_queue *txq)
268 {
269 	struct dst_entry *dst = skb_dst(skb);
270 	struct neighbour *mn;
271 	int res;
272 
273 	if (txq->qdisc == &noop_qdisc)
274 		return -ENODEV;
275 
276 	if (!dev->header_ops || !dst)
277 		return 0;
278 
279 	rcu_read_lock();
280 	mn = dst_get_neighbour_noref(dst);
281 	res = mn ? __teql_resolve(skb, skb_res, dev, txq, mn) : 0;
282 	rcu_read_unlock();
283 
284 	return res;
285 }
286 
teql_master_xmit(struct sk_buff * skb,struct net_device * dev)287 static netdev_tx_t teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
288 {
289 	struct teql_master *master = netdev_priv(dev);
290 	struct Qdisc *start, *q;
291 	int busy;
292 	int nores;
293 	int subq = skb_get_queue_mapping(skb);
294 	struct sk_buff *skb_res = NULL;
295 
296 	start = master->slaves;
297 
298 restart:
299 	nores = 0;
300 	busy = 0;
301 
302 	q = start;
303 	if (!q)
304 		goto drop;
305 
306 	do {
307 		struct net_device *slave = qdisc_dev(q);
308 		struct netdev_queue *slave_txq = netdev_get_tx_queue(slave, 0);
309 		const struct net_device_ops *slave_ops = slave->netdev_ops;
310 
311 		if (slave_txq->qdisc_sleeping != q)
312 			continue;
313 		if (netif_xmit_stopped(netdev_get_tx_queue(slave, subq)) ||
314 		    !netif_running(slave)) {
315 			busy = 1;
316 			continue;
317 		}
318 
319 		switch (teql_resolve(skb, skb_res, slave, slave_txq)) {
320 		case 0:
321 			if (__netif_tx_trylock(slave_txq)) {
322 				unsigned int length = qdisc_pkt_len(skb);
323 
324 				if (!netif_xmit_frozen_or_stopped(slave_txq) &&
325 				    slave_ops->ndo_start_xmit(skb, slave) == NETDEV_TX_OK) {
326 					txq_trans_update(slave_txq);
327 					__netif_tx_unlock(slave_txq);
328 					master->slaves = NEXT_SLAVE(q);
329 					netif_wake_queue(dev);
330 					master->tx_packets++;
331 					master->tx_bytes += length;
332 					return NETDEV_TX_OK;
333 				}
334 				__netif_tx_unlock(slave_txq);
335 			}
336 			if (netif_xmit_stopped(netdev_get_tx_queue(dev, 0)))
337 				busy = 1;
338 			break;
339 		case 1:
340 			master->slaves = NEXT_SLAVE(q);
341 			return NETDEV_TX_OK;
342 		default:
343 			nores = 1;
344 			break;
345 		}
346 		__skb_pull(skb, skb_network_offset(skb));
347 	} while ((q = NEXT_SLAVE(q)) != start);
348 
349 	if (nores && skb_res == NULL) {
350 		skb_res = skb;
351 		goto restart;
352 	}
353 
354 	if (busy) {
355 		netif_stop_queue(dev);
356 		return NETDEV_TX_BUSY;
357 	}
358 	master->tx_errors++;
359 
360 drop:
361 	master->tx_dropped++;
362 	dev_kfree_skb(skb);
363 	return NETDEV_TX_OK;
364 }
365 
teql_master_open(struct net_device * dev)366 static int teql_master_open(struct net_device *dev)
367 {
368 	struct Qdisc *q;
369 	struct teql_master *m = netdev_priv(dev);
370 	int mtu = 0xFFFE;
371 	unsigned int flags = IFF_NOARP | IFF_MULTICAST;
372 
373 	if (m->slaves == NULL)
374 		return -EUNATCH;
375 
376 	flags = FMASK;
377 
378 	q = m->slaves;
379 	do {
380 		struct net_device *slave = qdisc_dev(q);
381 
382 		if (slave == NULL)
383 			return -EUNATCH;
384 
385 		if (slave->mtu < mtu)
386 			mtu = slave->mtu;
387 		if (slave->hard_header_len > LL_MAX_HEADER)
388 			return -EINVAL;
389 
390 		/* If all the slaves are BROADCAST, master is BROADCAST
391 		   If all the slaves are PtP, master is PtP
392 		   Otherwise, master is NBMA.
393 		 */
394 		if (!(slave->flags&IFF_POINTOPOINT))
395 			flags &= ~IFF_POINTOPOINT;
396 		if (!(slave->flags&IFF_BROADCAST))
397 			flags &= ~IFF_BROADCAST;
398 		if (!(slave->flags&IFF_MULTICAST))
399 			flags &= ~IFF_MULTICAST;
400 	} while ((q = NEXT_SLAVE(q)) != m->slaves);
401 
402 	m->dev->mtu = mtu;
403 	m->dev->flags = (m->dev->flags&~FMASK) | flags;
404 	netif_start_queue(m->dev);
405 	return 0;
406 }
407 
teql_master_close(struct net_device * dev)408 static int teql_master_close(struct net_device *dev)
409 {
410 	netif_stop_queue(dev);
411 	return 0;
412 }
413 
teql_master_stats64(struct net_device * dev,struct rtnl_link_stats64 * stats)414 static struct rtnl_link_stats64 *teql_master_stats64(struct net_device *dev,
415 						     struct rtnl_link_stats64 *stats)
416 {
417 	struct teql_master *m = netdev_priv(dev);
418 
419 	stats->tx_packets	= m->tx_packets;
420 	stats->tx_bytes		= m->tx_bytes;
421 	stats->tx_errors	= m->tx_errors;
422 	stats->tx_dropped	= m->tx_dropped;
423 	return stats;
424 }
425 
teql_master_mtu(struct net_device * dev,int new_mtu)426 static int teql_master_mtu(struct net_device *dev, int new_mtu)
427 {
428 	struct teql_master *m = netdev_priv(dev);
429 	struct Qdisc *q;
430 
431 	if (new_mtu < 68)
432 		return -EINVAL;
433 
434 	q = m->slaves;
435 	if (q) {
436 		do {
437 			if (new_mtu > qdisc_dev(q)->mtu)
438 				return -EINVAL;
439 		} while ((q = NEXT_SLAVE(q)) != m->slaves);
440 	}
441 
442 	dev->mtu = new_mtu;
443 	return 0;
444 }
445 
446 static const struct net_device_ops teql_netdev_ops = {
447 	.ndo_open	= teql_master_open,
448 	.ndo_stop	= teql_master_close,
449 	.ndo_start_xmit	= teql_master_xmit,
450 	.ndo_get_stats64 = teql_master_stats64,
451 	.ndo_change_mtu	= teql_master_mtu,
452 };
453 
teql_master_setup(struct net_device * dev)454 static __init void teql_master_setup(struct net_device *dev)
455 {
456 	struct teql_master *master = netdev_priv(dev);
457 	struct Qdisc_ops *ops = &master->qops;
458 
459 	master->dev	= dev;
460 	ops->priv_size  = sizeof(struct teql_sched_data);
461 
462 	ops->enqueue	=	teql_enqueue;
463 	ops->dequeue	=	teql_dequeue;
464 	ops->peek	=	teql_peek;
465 	ops->init	=	teql_qdisc_init;
466 	ops->reset	=	teql_reset;
467 	ops->destroy	=	teql_destroy;
468 	ops->owner	=	THIS_MODULE;
469 
470 	dev->netdev_ops =       &teql_netdev_ops;
471 	dev->type		= ARPHRD_VOID;
472 	dev->mtu		= 1500;
473 	dev->tx_queue_len	= 100;
474 	dev->flags		= IFF_NOARP;
475 	dev->hard_header_len	= LL_MAX_HEADER;
476 	dev->priv_flags		&= ~IFF_XMIT_DST_RELEASE;
477 }
478 
479 static LIST_HEAD(master_dev_list);
480 static int max_equalizers = 1;
481 module_param(max_equalizers, int, 0);
482 MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
483 
teql_init(void)484 static int __init teql_init(void)
485 {
486 	int i;
487 	int err = -ENODEV;
488 
489 	for (i = 0; i < max_equalizers; i++) {
490 		struct net_device *dev;
491 		struct teql_master *master;
492 
493 		dev = alloc_netdev(sizeof(struct teql_master),
494 				  "teql%d", teql_master_setup);
495 		if (!dev) {
496 			err = -ENOMEM;
497 			break;
498 		}
499 
500 		if ((err = register_netdev(dev))) {
501 			free_netdev(dev);
502 			break;
503 		}
504 
505 		master = netdev_priv(dev);
506 
507 		strlcpy(master->qops.id, dev->name, IFNAMSIZ);
508 		err = register_qdisc(&master->qops);
509 
510 		if (err) {
511 			unregister_netdev(dev);
512 			free_netdev(dev);
513 			break;
514 		}
515 
516 		list_add_tail(&master->master_list, &master_dev_list);
517 	}
518 	return i ? 0 : err;
519 }
520 
teql_exit(void)521 static void __exit teql_exit(void)
522 {
523 	struct teql_master *master, *nxt;
524 
525 	list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) {
526 
527 		list_del(&master->master_list);
528 
529 		unregister_qdisc(&master->qops);
530 		unregister_netdev(master->dev);
531 		free_netdev(master->dev);
532 	}
533 }
534 
535 module_init(teql_init);
536 module_exit(teql_exit);
537 
538 MODULE_LICENSE("GPL");
539