xref: /openbmc/linux/net/sched/sch_teql.c (revision e2c75e76)
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, struct sk_buff **to_free)
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, to_free);
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 			   struct netlink_ext_ack *extack)
172 {
173 	struct net_device *dev = qdisc_dev(sch);
174 	struct teql_master *m = (struct teql_master *)sch->ops;
175 	struct teql_sched_data *q = qdisc_priv(sch);
176 
177 	if (dev->hard_header_len > m->dev->hard_header_len)
178 		return -EINVAL;
179 
180 	if (m->dev == dev)
181 		return -ELOOP;
182 
183 	q->m = m;
184 
185 	skb_queue_head_init(&q->q);
186 
187 	if (m->slaves) {
188 		if (m->dev->flags & IFF_UP) {
189 			if ((m->dev->flags & IFF_POINTOPOINT &&
190 			     !(dev->flags & IFF_POINTOPOINT)) ||
191 			    (m->dev->flags & IFF_BROADCAST &&
192 			     !(dev->flags & IFF_BROADCAST)) ||
193 			    (m->dev->flags & IFF_MULTICAST &&
194 			     !(dev->flags & IFF_MULTICAST)) ||
195 			    dev->mtu < m->dev->mtu)
196 				return -EINVAL;
197 		} else {
198 			if (!(dev->flags&IFF_POINTOPOINT))
199 				m->dev->flags &= ~IFF_POINTOPOINT;
200 			if (!(dev->flags&IFF_BROADCAST))
201 				m->dev->flags &= ~IFF_BROADCAST;
202 			if (!(dev->flags&IFF_MULTICAST))
203 				m->dev->flags &= ~IFF_MULTICAST;
204 			if (dev->mtu < m->dev->mtu)
205 				m->dev->mtu = dev->mtu;
206 		}
207 		q->next = NEXT_SLAVE(m->slaves);
208 		NEXT_SLAVE(m->slaves) = sch;
209 	} else {
210 		q->next = sch;
211 		m->slaves = sch;
212 		m->dev->mtu = dev->mtu;
213 		m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK);
214 	}
215 	return 0;
216 }
217 
218 
219 static int
220 __teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res,
221 	       struct net_device *dev, struct netdev_queue *txq,
222 	       struct dst_entry *dst)
223 {
224 	struct neighbour *n;
225 	int err = 0;
226 
227 	n = dst_neigh_lookup_skb(dst, skb);
228 	if (!n)
229 		return -ENOENT;
230 
231 	if (dst->dev != dev) {
232 		struct neighbour *mn;
233 
234 		mn = __neigh_lookup_errno(n->tbl, n->primary_key, dev);
235 		neigh_release(n);
236 		if (IS_ERR(mn))
237 			return PTR_ERR(mn);
238 		n = mn;
239 	}
240 
241 	if (neigh_event_send(n, skb_res) == 0) {
242 		int err;
243 		char haddr[MAX_ADDR_LEN];
244 
245 		neigh_ha_snapshot(haddr, n, dev);
246 		err = dev_hard_header(skb, dev, ntohs(tc_skb_protocol(skb)),
247 				      haddr, NULL, skb->len);
248 
249 		if (err < 0)
250 			err = -EINVAL;
251 	} else {
252 		err = (skb_res == NULL) ? -EAGAIN : 1;
253 	}
254 	neigh_release(n);
255 	return err;
256 }
257 
258 static inline int teql_resolve(struct sk_buff *skb,
259 			       struct sk_buff *skb_res,
260 			       struct net_device *dev,
261 			       struct netdev_queue *txq)
262 {
263 	struct dst_entry *dst = skb_dst(skb);
264 	int res;
265 
266 	if (rcu_access_pointer(txq->qdisc) == &noop_qdisc)
267 		return -ENODEV;
268 
269 	if (!dev->header_ops || !dst)
270 		return 0;
271 
272 	rcu_read_lock();
273 	res = __teql_resolve(skb, skb_res, dev, txq, dst);
274 	rcu_read_unlock();
275 
276 	return res;
277 }
278 
279 static netdev_tx_t teql_master_xmit(struct sk_buff *skb, struct net_device *dev)
280 {
281 	struct teql_master *master = netdev_priv(dev);
282 	struct Qdisc *start, *q;
283 	int busy;
284 	int nores;
285 	int subq = skb_get_queue_mapping(skb);
286 	struct sk_buff *skb_res = NULL;
287 
288 	start = master->slaves;
289 
290 restart:
291 	nores = 0;
292 	busy = 0;
293 
294 	q = start;
295 	if (!q)
296 		goto drop;
297 
298 	do {
299 		struct net_device *slave = qdisc_dev(q);
300 		struct netdev_queue *slave_txq = netdev_get_tx_queue(slave, 0);
301 
302 		if (slave_txq->qdisc_sleeping != q)
303 			continue;
304 		if (netif_xmit_stopped(netdev_get_tx_queue(slave, subq)) ||
305 		    !netif_running(slave)) {
306 			busy = 1;
307 			continue;
308 		}
309 
310 		switch (teql_resolve(skb, skb_res, slave, slave_txq)) {
311 		case 0:
312 			if (__netif_tx_trylock(slave_txq)) {
313 				unsigned int length = qdisc_pkt_len(skb);
314 
315 				if (!netif_xmit_frozen_or_stopped(slave_txq) &&
316 				    netdev_start_xmit(skb, slave, slave_txq, false) ==
317 				    NETDEV_TX_OK) {
318 					__netif_tx_unlock(slave_txq);
319 					master->slaves = NEXT_SLAVE(q);
320 					netif_wake_queue(dev);
321 					master->tx_packets++;
322 					master->tx_bytes += length;
323 					return NETDEV_TX_OK;
324 				}
325 				__netif_tx_unlock(slave_txq);
326 			}
327 			if (netif_xmit_stopped(netdev_get_tx_queue(dev, 0)))
328 				busy = 1;
329 			break;
330 		case 1:
331 			master->slaves = NEXT_SLAVE(q);
332 			return NETDEV_TX_OK;
333 		default:
334 			nores = 1;
335 			break;
336 		}
337 		__skb_pull(skb, skb_network_offset(skb));
338 	} while ((q = NEXT_SLAVE(q)) != start);
339 
340 	if (nores && skb_res == NULL) {
341 		skb_res = skb;
342 		goto restart;
343 	}
344 
345 	if (busy) {
346 		netif_stop_queue(dev);
347 		return NETDEV_TX_BUSY;
348 	}
349 	master->tx_errors++;
350 
351 drop:
352 	master->tx_dropped++;
353 	dev_kfree_skb(skb);
354 	return NETDEV_TX_OK;
355 }
356 
357 static int teql_master_open(struct net_device *dev)
358 {
359 	struct Qdisc *q;
360 	struct teql_master *m = netdev_priv(dev);
361 	int mtu = 0xFFFE;
362 	unsigned int flags = IFF_NOARP | IFF_MULTICAST;
363 
364 	if (m->slaves == NULL)
365 		return -EUNATCH;
366 
367 	flags = FMASK;
368 
369 	q = m->slaves;
370 	do {
371 		struct net_device *slave = qdisc_dev(q);
372 
373 		if (slave == NULL)
374 			return -EUNATCH;
375 
376 		if (slave->mtu < mtu)
377 			mtu = slave->mtu;
378 		if (slave->hard_header_len > LL_MAX_HEADER)
379 			return -EINVAL;
380 
381 		/* If all the slaves are BROADCAST, master is BROADCAST
382 		   If all the slaves are PtP, master is PtP
383 		   Otherwise, master is NBMA.
384 		 */
385 		if (!(slave->flags&IFF_POINTOPOINT))
386 			flags &= ~IFF_POINTOPOINT;
387 		if (!(slave->flags&IFF_BROADCAST))
388 			flags &= ~IFF_BROADCAST;
389 		if (!(slave->flags&IFF_MULTICAST))
390 			flags &= ~IFF_MULTICAST;
391 	} while ((q = NEXT_SLAVE(q)) != m->slaves);
392 
393 	m->dev->mtu = mtu;
394 	m->dev->flags = (m->dev->flags&~FMASK) | flags;
395 	netif_start_queue(m->dev);
396 	return 0;
397 }
398 
399 static int teql_master_close(struct net_device *dev)
400 {
401 	netif_stop_queue(dev);
402 	return 0;
403 }
404 
405 static void teql_master_stats64(struct net_device *dev,
406 				struct rtnl_link_stats64 *stats)
407 {
408 	struct teql_master *m = netdev_priv(dev);
409 
410 	stats->tx_packets	= m->tx_packets;
411 	stats->tx_bytes		= m->tx_bytes;
412 	stats->tx_errors	= m->tx_errors;
413 	stats->tx_dropped	= m->tx_dropped;
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 	q = m->slaves;
422 	if (q) {
423 		do {
424 			if (new_mtu > qdisc_dev(q)->mtu)
425 				return -EINVAL;
426 		} while ((q = NEXT_SLAVE(q)) != m->slaves);
427 	}
428 
429 	dev->mtu = new_mtu;
430 	return 0;
431 }
432 
433 static const struct net_device_ops teql_netdev_ops = {
434 	.ndo_open	= teql_master_open,
435 	.ndo_stop	= teql_master_close,
436 	.ndo_start_xmit	= teql_master_xmit,
437 	.ndo_get_stats64 = teql_master_stats64,
438 	.ndo_change_mtu	= teql_master_mtu,
439 };
440 
441 static __init void teql_master_setup(struct net_device *dev)
442 {
443 	struct teql_master *master = netdev_priv(dev);
444 	struct Qdisc_ops *ops = &master->qops;
445 
446 	master->dev	= dev;
447 	ops->priv_size  = sizeof(struct teql_sched_data);
448 
449 	ops->enqueue	=	teql_enqueue;
450 	ops->dequeue	=	teql_dequeue;
451 	ops->peek	=	teql_peek;
452 	ops->init	=	teql_qdisc_init;
453 	ops->reset	=	teql_reset;
454 	ops->destroy	=	teql_destroy;
455 	ops->owner	=	THIS_MODULE;
456 
457 	dev->netdev_ops =       &teql_netdev_ops;
458 	dev->type		= ARPHRD_VOID;
459 	dev->mtu		= 1500;
460 	dev->min_mtu		= 68;
461 	dev->max_mtu		= 65535;
462 	dev->tx_queue_len	= 100;
463 	dev->flags		= IFF_NOARP;
464 	dev->hard_header_len	= LL_MAX_HEADER;
465 	netif_keep_dst(dev);
466 }
467 
468 static LIST_HEAD(master_dev_list);
469 static int max_equalizers = 1;
470 module_param(max_equalizers, int, 0);
471 MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers");
472 
473 static int __init teql_init(void)
474 {
475 	int i;
476 	int err = -ENODEV;
477 
478 	for (i = 0; i < max_equalizers; i++) {
479 		struct net_device *dev;
480 		struct teql_master *master;
481 
482 		dev = alloc_netdev(sizeof(struct teql_master), "teql%d",
483 				   NET_NAME_UNKNOWN, teql_master_setup);
484 		if (!dev) {
485 			err = -ENOMEM;
486 			break;
487 		}
488 
489 		if ((err = register_netdev(dev))) {
490 			free_netdev(dev);
491 			break;
492 		}
493 
494 		master = netdev_priv(dev);
495 
496 		strlcpy(master->qops.id, dev->name, IFNAMSIZ);
497 		err = register_qdisc(&master->qops);
498 
499 		if (err) {
500 			unregister_netdev(dev);
501 			free_netdev(dev);
502 			break;
503 		}
504 
505 		list_add_tail(&master->master_list, &master_dev_list);
506 	}
507 	return i ? 0 : err;
508 }
509 
510 static void __exit teql_exit(void)
511 {
512 	struct teql_master *master, *nxt;
513 
514 	list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) {
515 
516 		list_del(&master->master_list);
517 
518 		unregister_qdisc(&master->qops);
519 		unregister_netdev(master->dev);
520 		free_netdev(master->dev);
521 	}
522 }
523 
524 module_init(teql_init);
525 module_exit(teql_exit);
526 
527 MODULE_LICENSE("GPL");
528