xref: /openbmc/linux/net/sched/sch_teql.c (revision d8b9605d)
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 sk_buff_head q;
71 };
72 
73 #define NEXT_SLAVE(q) (((struct teql_sched_data *)qdisc_priv(q))->next)
74 
75 #define FMASK (IFF_BROADCAST | IFF_POINTOPOINT)
76 
77 /* "teql*" qdisc routines */
78 
79 static int
80 teql_enqueue(struct sk_buff *skb, struct Qdisc *sch)
81 {
82 	struct net_device *dev = qdisc_dev(sch);
83 	struct teql_sched_data *q = qdisc_priv(sch);
84 
85 	if (q->q.qlen < dev->tx_queue_len) {
86 		__skb_queue_tail(&q->q, skb);
87 		return NET_XMIT_SUCCESS;
88 	}
89 
90 	return qdisc_drop(skb, sch);
91 }
92 
93 static struct sk_buff *
94 teql_dequeue(struct Qdisc *sch)
95 {
96 	struct teql_sched_data *dat = qdisc_priv(sch);
97 	struct netdev_queue *dat_queue;
98 	struct sk_buff *skb;
99 	struct Qdisc *q;
100 
101 	skb = __skb_dequeue(&dat->q);
102 	dat_queue = netdev_get_tx_queue(dat->m->dev, 0);
103 	q = rcu_dereference_bh(dat_queue->qdisc);
104 
105 	if (skb == NULL) {
106 		struct net_device *m = qdisc_dev(q);
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 + q->q.qlen;
115 	return skb;
116 }
117 
118 static struct sk_buff *
119 teql_peek(struct Qdisc *sch)
120 {
121 	/* teql is meant to be used as root qdisc */
122 	return NULL;
123 }
124 
125 static void
126 teql_reset(struct Qdisc *sch)
127 {
128 	struct teql_sched_data *dat = qdisc_priv(sch);
129 
130 	skb_queue_purge(&dat->q);
131 	sch->q.qlen = 0;
132 }
133 
134 static void
135 teql_destroy(struct Qdisc *sch)
136 {
137 	struct Qdisc *q, *prev;
138 	struct teql_sched_data *dat = qdisc_priv(sch);
139 	struct teql_master *master = dat->m;
140 
141 	prev = master->slaves;
142 	if (prev) {
143 		do {
144 			q = NEXT_SLAVE(prev);
145 			if (q == sch) {
146 				NEXT_SLAVE(prev) = NEXT_SLAVE(q);
147 				if (q == master->slaves) {
148 					master->slaves = NEXT_SLAVE(q);
149 					if (q == master->slaves) {
150 						struct netdev_queue *txq;
151 						spinlock_t *root_lock;
152 
153 						txq = netdev_get_tx_queue(master->dev, 0);
154 						master->slaves = NULL;
155 
156 						root_lock = qdisc_root_sleeping_lock(rtnl_dereference(txq->qdisc));
157 						spin_lock_bh(root_lock);
158 						qdisc_reset(rtnl_dereference(txq->qdisc));
159 						spin_unlock_bh(root_lock);
160 					}
161 				}
162 				skb_queue_purge(&dat->q);
163 				break;
164 			}
165 
166 		} while ((prev = q) != master->slaves);
167 	}
168 }
169 
170 static int teql_qdisc_init(struct Qdisc *sch, struct nlattr *opt)
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
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(tc_skb_protocol(skb)),
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 
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 
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 
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 
398 static int teql_master_close(struct net_device *dev)
399 {
400 	netif_stop_queue(dev);
401 	return 0;
402 }
403 
404 static struct rtnl_link_stats64 *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 	return stats;
414 }
415 
416 static int teql_master_mtu(struct net_device *dev, int new_mtu)
417 {
418 	struct teql_master *m = netdev_priv(dev);
419 	struct Qdisc *q;
420 
421 	if (new_mtu < 68)
422 		return -EINVAL;
423 
424 	q = m->slaves;
425 	if (q) {
426 		do {
427 			if (new_mtu > qdisc_dev(q)->mtu)
428 				return -EINVAL;
429 		} while ((q = NEXT_SLAVE(q)) != m->slaves);
430 	}
431 
432 	dev->mtu = new_mtu;
433 	return 0;
434 }
435 
436 static const struct net_device_ops teql_netdev_ops = {
437 	.ndo_open	= teql_master_open,
438 	.ndo_stop	= teql_master_close,
439 	.ndo_start_xmit	= teql_master_xmit,
440 	.ndo_get_stats64 = teql_master_stats64,
441 	.ndo_change_mtu	= teql_master_mtu,
442 };
443 
444 static __init void teql_master_setup(struct net_device *dev)
445 {
446 	struct teql_master *master = netdev_priv(dev);
447 	struct Qdisc_ops *ops = &master->qops;
448 
449 	master->dev	= dev;
450 	ops->priv_size  = sizeof(struct teql_sched_data);
451 
452 	ops->enqueue	=	teql_enqueue;
453 	ops->dequeue	=	teql_dequeue;
454 	ops->peek	=	teql_peek;
455 	ops->init	=	teql_qdisc_init;
456 	ops->reset	=	teql_reset;
457 	ops->destroy	=	teql_destroy;
458 	ops->owner	=	THIS_MODULE;
459 
460 	dev->netdev_ops =       &teql_netdev_ops;
461 	dev->type		= ARPHRD_VOID;
462 	dev->mtu		= 1500;
463 	dev->tx_queue_len	= 100;
464 	dev->flags		= IFF_NOARP;
465 	dev->hard_header_len	= LL_MAX_HEADER;
466 	netif_keep_dst(dev);
467 }
468 
469 static LIST_HEAD(master_dev_list);
470 static int max_equalizers = 1;
471 module_param(max_equalizers, int, 0);
472 MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
473 
474 static int __init teql_init(void)
475 {
476 	int i;
477 	int err = -ENODEV;
478 
479 	for (i = 0; i < max_equalizers; i++) {
480 		struct net_device *dev;
481 		struct teql_master *master;
482 
483 		dev = alloc_netdev(sizeof(struct teql_master), "teql%d",
484 				   NET_NAME_UNKNOWN, teql_master_setup);
485 		if (!dev) {
486 			err = -ENOMEM;
487 			break;
488 		}
489 
490 		if ((err = register_netdev(dev))) {
491 			free_netdev(dev);
492 			break;
493 		}
494 
495 		master = netdev_priv(dev);
496 
497 		strlcpy(master->qops.id, dev->name, IFNAMSIZ);
498 		err = register_qdisc(&master->qops);
499 
500 		if (err) {
501 			unregister_netdev(dev);
502 			free_netdev(dev);
503 			break;
504 		}
505 
506 		list_add_tail(&master->master_list, &master_dev_list);
507 	}
508 	return i ? 0 : err;
509 }
510 
511 static void __exit teql_exit(void)
512 {
513 	struct teql_master *master, *nxt;
514 
515 	list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) {
516 
517 		list_del(&master->master_list);
518 
519 		unregister_qdisc(&master->qops);
520 		unregister_netdev(master->dev);
521 		free_netdev(master->dev);
522 	}
523 }
524 
525 module_init(teql_init);
526 module_exit(teql_exit);
527 
528 MODULE_LICENSE("GPL");
529