xref: /linux/drivers/net/macvlan.c (revision d39d0ed196aa1685bb24771e92f78633c66ac9cb)
1 /*
2  * Copyright (c) 2007 Patrick McHardy <kaber@trash.net>
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public License as
6  * published by the Free Software Foundation; either version 2 of
7  * the License, or (at your option) any later version.
8  *
9  * The code this is based on carried the following copyright notice:
10  * ---
11  * (C) Copyright 2001-2006
12  * Alex Zeffertt, Cambridge Broadband Ltd, ajz@cambridgebroadband.com
13  * Re-worked by Ben Greear <greearb@candelatech.com>
14  * ---
15  */
16 #include <linux/kernel.h>
17 #include <linux/types.h>
18 #include <linux/module.h>
19 #include <linux/init.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/string.h>
23 #include <linux/rculist.h>
24 #include <linux/notifier.h>
25 #include <linux/netdevice.h>
26 #include <linux/etherdevice.h>
27 #include <linux/ethtool.h>
28 #include <linux/if_arp.h>
29 #include <linux/if_link.h>
30 #include <linux/if_macvlan.h>
31 #include <net/rtnetlink.h>
32 #include <net/xfrm.h>
33 
34 #define MACVLAN_HASH_SIZE	(1 << BITS_PER_BYTE)
35 
36 struct macvlan_port {
37 	struct net_device	*dev;
38 	struct hlist_head	vlan_hash[MACVLAN_HASH_SIZE];
39 	struct list_head	vlans;
40 	struct rcu_head		rcu;
41 };
42 
43 #define macvlan_port_get_rcu(dev) \
44 	((struct macvlan_port *) rcu_dereference(dev->rx_handler_data))
45 #define macvlan_port_get(dev) ((struct macvlan_port *) dev->rx_handler_data)
46 #define macvlan_port_exists(dev) (dev->priv_flags & IFF_MACVLAN_PORT)
47 
48 static struct macvlan_dev *macvlan_hash_lookup(const struct macvlan_port *port,
49 					       const unsigned char *addr)
50 {
51 	struct macvlan_dev *vlan;
52 	struct hlist_node *n;
53 
54 	hlist_for_each_entry_rcu(vlan, n, &port->vlan_hash[addr[5]], hlist) {
55 		if (!compare_ether_addr_64bits(vlan->dev->dev_addr, addr))
56 			return vlan;
57 	}
58 	return NULL;
59 }
60 
61 static void macvlan_hash_add(struct macvlan_dev *vlan)
62 {
63 	struct macvlan_port *port = vlan->port;
64 	const unsigned char *addr = vlan->dev->dev_addr;
65 
66 	hlist_add_head_rcu(&vlan->hlist, &port->vlan_hash[addr[5]]);
67 }
68 
69 static void macvlan_hash_del(struct macvlan_dev *vlan)
70 {
71 	hlist_del_rcu(&vlan->hlist);
72 	synchronize_rcu();
73 }
74 
75 static void macvlan_hash_change_addr(struct macvlan_dev *vlan,
76 					const unsigned char *addr)
77 {
78 	macvlan_hash_del(vlan);
79 	/* Now that we are unhashed it is safe to change the device
80 	 * address without confusing packet delivery.
81 	 */
82 	memcpy(vlan->dev->dev_addr, addr, ETH_ALEN);
83 	macvlan_hash_add(vlan);
84 }
85 
86 static int macvlan_addr_busy(const struct macvlan_port *port,
87 				const unsigned char *addr)
88 {
89 	/* Test to see if the specified multicast address is
90 	 * currently in use by the underlying device or
91 	 * another macvlan.
92 	 */
93 	if (!compare_ether_addr_64bits(port->dev->dev_addr, addr))
94 		return 1;
95 
96 	if (macvlan_hash_lookup(port, addr))
97 		return 1;
98 
99 	return 0;
100 }
101 
102 
103 static int macvlan_broadcast_one(struct sk_buff *skb,
104 				 const struct macvlan_dev *vlan,
105 				 const struct ethhdr *eth, bool local)
106 {
107 	struct net_device *dev = vlan->dev;
108 	if (!skb)
109 		return NET_RX_DROP;
110 
111 	if (local)
112 		return vlan->forward(dev, skb);
113 
114 	skb->dev = dev;
115 	if (!compare_ether_addr_64bits(eth->h_dest,
116 				       dev->broadcast))
117 		skb->pkt_type = PACKET_BROADCAST;
118 	else
119 		skb->pkt_type = PACKET_MULTICAST;
120 
121 	return vlan->receive(skb);
122 }
123 
124 static void macvlan_broadcast(struct sk_buff *skb,
125 			      const struct macvlan_port *port,
126 			      struct net_device *src,
127 			      enum macvlan_mode mode)
128 {
129 	const struct ethhdr *eth = eth_hdr(skb);
130 	const struct macvlan_dev *vlan;
131 	struct hlist_node *n;
132 	struct sk_buff *nskb;
133 	unsigned int i;
134 	int err;
135 
136 	if (skb->protocol == htons(ETH_P_PAUSE))
137 		return;
138 
139 	for (i = 0; i < MACVLAN_HASH_SIZE; i++) {
140 		hlist_for_each_entry_rcu(vlan, n, &port->vlan_hash[i], hlist) {
141 			if (vlan->dev == src || !(vlan->mode & mode))
142 				continue;
143 
144 			nskb = skb_clone(skb, GFP_ATOMIC);
145 			err = macvlan_broadcast_one(nskb, vlan, eth,
146 					 mode == MACVLAN_MODE_BRIDGE);
147 			macvlan_count_rx(vlan, skb->len + ETH_HLEN,
148 					 err == NET_RX_SUCCESS, 1);
149 		}
150 	}
151 }
152 
153 /* called under rcu_read_lock() from netif_receive_skb */
154 static struct sk_buff *macvlan_handle_frame(struct sk_buff *skb)
155 {
156 	struct macvlan_port *port;
157 	const struct ethhdr *eth = eth_hdr(skb);
158 	const struct macvlan_dev *vlan;
159 	const struct macvlan_dev *src;
160 	struct net_device *dev;
161 	unsigned int len = 0;
162 	int ret = NET_RX_DROP;
163 
164 	port = macvlan_port_get_rcu(skb->dev);
165 	if (is_multicast_ether_addr(eth->h_dest)) {
166 		src = macvlan_hash_lookup(port, eth->h_source);
167 		if (!src)
168 			/* frame comes from an external address */
169 			macvlan_broadcast(skb, port, NULL,
170 					  MACVLAN_MODE_PRIVATE |
171 					  MACVLAN_MODE_VEPA    |
172 					  MACVLAN_MODE_BRIDGE);
173 		else if (src->mode == MACVLAN_MODE_VEPA)
174 			/* flood to everyone except source */
175 			macvlan_broadcast(skb, port, src->dev,
176 					  MACVLAN_MODE_VEPA |
177 					  MACVLAN_MODE_BRIDGE);
178 		else if (src->mode == MACVLAN_MODE_BRIDGE)
179 			/*
180 			 * flood only to VEPA ports, bridge ports
181 			 * already saw the frame on the way out.
182 			 */
183 			macvlan_broadcast(skb, port, src->dev,
184 					  MACVLAN_MODE_VEPA);
185 		return skb;
186 	}
187 
188 	vlan = macvlan_hash_lookup(port, eth->h_dest);
189 	if (vlan == NULL)
190 		return skb;
191 
192 	dev = vlan->dev;
193 	if (unlikely(!(dev->flags & IFF_UP))) {
194 		kfree_skb(skb);
195 		return NULL;
196 	}
197 	len = skb->len + ETH_HLEN;
198 	skb = skb_share_check(skb, GFP_ATOMIC);
199 	if (!skb)
200 		goto out;
201 
202 	skb->dev = dev;
203 	skb->pkt_type = PACKET_HOST;
204 
205 	ret = vlan->receive(skb);
206 
207 out:
208 	macvlan_count_rx(vlan, len, ret == NET_RX_SUCCESS, 0);
209 	return NULL;
210 }
211 
212 static int macvlan_queue_xmit(struct sk_buff *skb, struct net_device *dev)
213 {
214 	const struct macvlan_dev *vlan = netdev_priv(dev);
215 	const struct macvlan_port *port = vlan->port;
216 	const struct macvlan_dev *dest;
217 
218 	if (vlan->mode == MACVLAN_MODE_BRIDGE) {
219 		const struct ethhdr *eth = (void *)skb->data;
220 
221 		/* send to other bridge ports directly */
222 		if (is_multicast_ether_addr(eth->h_dest)) {
223 			macvlan_broadcast(skb, port, dev, MACVLAN_MODE_BRIDGE);
224 			goto xmit_world;
225 		}
226 
227 		dest = macvlan_hash_lookup(port, eth->h_dest);
228 		if (dest && dest->mode == MACVLAN_MODE_BRIDGE) {
229 			unsigned int length = skb->len + ETH_HLEN;
230 			int ret = dest->forward(dest->dev, skb);
231 			macvlan_count_rx(dest, length,
232 					 ret == NET_RX_SUCCESS, 0);
233 
234 			return NET_XMIT_SUCCESS;
235 		}
236 	}
237 
238 xmit_world:
239 	skb_set_dev(skb, vlan->lowerdev);
240 	return dev_queue_xmit(skb);
241 }
242 
243 netdev_tx_t macvlan_start_xmit(struct sk_buff *skb,
244 			       struct net_device *dev)
245 {
246 	int i = skb_get_queue_mapping(skb);
247 	struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
248 	unsigned int len = skb->len;
249 	int ret;
250 
251 	ret = macvlan_queue_xmit(skb, dev);
252 	if (likely(ret == NET_XMIT_SUCCESS || ret == NET_XMIT_CN)) {
253 		txq->tx_packets++;
254 		txq->tx_bytes += len;
255 	} else
256 		txq->tx_dropped++;
257 
258 	return ret;
259 }
260 EXPORT_SYMBOL_GPL(macvlan_start_xmit);
261 
262 static int macvlan_hard_header(struct sk_buff *skb, struct net_device *dev,
263 			       unsigned short type, const void *daddr,
264 			       const void *saddr, unsigned len)
265 {
266 	const struct macvlan_dev *vlan = netdev_priv(dev);
267 	struct net_device *lowerdev = vlan->lowerdev;
268 
269 	return dev_hard_header(skb, lowerdev, type, daddr,
270 			       saddr ? : dev->dev_addr, len);
271 }
272 
273 static const struct header_ops macvlan_hard_header_ops = {
274 	.create  	= macvlan_hard_header,
275 	.rebuild	= eth_rebuild_header,
276 	.parse		= eth_header_parse,
277 	.cache		= eth_header_cache,
278 	.cache_update	= eth_header_cache_update,
279 };
280 
281 static int macvlan_open(struct net_device *dev)
282 {
283 	struct macvlan_dev *vlan = netdev_priv(dev);
284 	struct net_device *lowerdev = vlan->lowerdev;
285 	int err;
286 
287 	err = -EBUSY;
288 	if (macvlan_addr_busy(vlan->port, dev->dev_addr))
289 		goto out;
290 
291 	err = dev_uc_add(lowerdev, dev->dev_addr);
292 	if (err < 0)
293 		goto out;
294 	if (dev->flags & IFF_ALLMULTI) {
295 		err = dev_set_allmulti(lowerdev, 1);
296 		if (err < 0)
297 			goto del_unicast;
298 	}
299 	macvlan_hash_add(vlan);
300 	return 0;
301 
302 del_unicast:
303 	dev_uc_del(lowerdev, dev->dev_addr);
304 out:
305 	return err;
306 }
307 
308 static int macvlan_stop(struct net_device *dev)
309 {
310 	struct macvlan_dev *vlan = netdev_priv(dev);
311 	struct net_device *lowerdev = vlan->lowerdev;
312 
313 	dev_mc_unsync(lowerdev, dev);
314 	if (dev->flags & IFF_ALLMULTI)
315 		dev_set_allmulti(lowerdev, -1);
316 
317 	dev_uc_del(lowerdev, dev->dev_addr);
318 
319 	macvlan_hash_del(vlan);
320 	return 0;
321 }
322 
323 static int macvlan_set_mac_address(struct net_device *dev, void *p)
324 {
325 	struct macvlan_dev *vlan = netdev_priv(dev);
326 	struct net_device *lowerdev = vlan->lowerdev;
327 	struct sockaddr *addr = p;
328 	int err;
329 
330 	if (!is_valid_ether_addr(addr->sa_data))
331 		return -EADDRNOTAVAIL;
332 
333 	if (!(dev->flags & IFF_UP)) {
334 		/* Just copy in the new address */
335 		memcpy(dev->dev_addr, addr->sa_data, ETH_ALEN);
336 	} else {
337 		/* Rehash and update the device filters */
338 		if (macvlan_addr_busy(vlan->port, addr->sa_data))
339 			return -EBUSY;
340 
341 		err = dev_uc_add(lowerdev, addr->sa_data);
342 		if (err)
343 			return err;
344 
345 		dev_uc_del(lowerdev, dev->dev_addr);
346 
347 		macvlan_hash_change_addr(vlan, addr->sa_data);
348 	}
349 	return 0;
350 }
351 
352 static void macvlan_change_rx_flags(struct net_device *dev, int change)
353 {
354 	struct macvlan_dev *vlan = netdev_priv(dev);
355 	struct net_device *lowerdev = vlan->lowerdev;
356 
357 	if (change & IFF_ALLMULTI)
358 		dev_set_allmulti(lowerdev, dev->flags & IFF_ALLMULTI ? 1 : -1);
359 }
360 
361 static void macvlan_set_multicast_list(struct net_device *dev)
362 {
363 	struct macvlan_dev *vlan = netdev_priv(dev);
364 
365 	dev_mc_sync(vlan->lowerdev, dev);
366 }
367 
368 static int macvlan_change_mtu(struct net_device *dev, int new_mtu)
369 {
370 	struct macvlan_dev *vlan = netdev_priv(dev);
371 
372 	if (new_mtu < 68 || vlan->lowerdev->mtu < new_mtu)
373 		return -EINVAL;
374 	dev->mtu = new_mtu;
375 	return 0;
376 }
377 
378 /*
379  * macvlan network devices have devices nesting below it and are a special
380  * "super class" of normal network devices; split their locks off into a
381  * separate class since they always nest.
382  */
383 static struct lock_class_key macvlan_netdev_xmit_lock_key;
384 static struct lock_class_key macvlan_netdev_addr_lock_key;
385 
386 #define MACVLAN_FEATURES \
387 	(NETIF_F_SG | NETIF_F_ALL_CSUM | NETIF_F_HIGHDMA | NETIF_F_FRAGLIST | \
388 	 NETIF_F_GSO | NETIF_F_TSO | NETIF_F_UFO | NETIF_F_GSO_ROBUST | \
389 	 NETIF_F_TSO_ECN | NETIF_F_TSO6 | NETIF_F_GRO)
390 
391 #define MACVLAN_STATE_MASK \
392 	((1<<__LINK_STATE_NOCARRIER) | (1<<__LINK_STATE_DORMANT))
393 
394 static void macvlan_set_lockdep_class_one(struct net_device *dev,
395 					  struct netdev_queue *txq,
396 					  void *_unused)
397 {
398 	lockdep_set_class(&txq->_xmit_lock,
399 			  &macvlan_netdev_xmit_lock_key);
400 }
401 
402 static void macvlan_set_lockdep_class(struct net_device *dev)
403 {
404 	lockdep_set_class(&dev->addr_list_lock,
405 			  &macvlan_netdev_addr_lock_key);
406 	netdev_for_each_tx_queue(dev, macvlan_set_lockdep_class_one, NULL);
407 }
408 
409 static int macvlan_init(struct net_device *dev)
410 {
411 	struct macvlan_dev *vlan = netdev_priv(dev);
412 	const struct net_device *lowerdev = vlan->lowerdev;
413 
414 	dev->state		= (dev->state & ~MACVLAN_STATE_MASK) |
415 				  (lowerdev->state & MACVLAN_STATE_MASK);
416 	dev->features 		= lowerdev->features & MACVLAN_FEATURES;
417 	dev->gso_max_size	= lowerdev->gso_max_size;
418 	dev->iflink		= lowerdev->ifindex;
419 	dev->hard_header_len	= lowerdev->hard_header_len;
420 
421 	macvlan_set_lockdep_class(dev);
422 
423 	vlan->rx_stats = alloc_percpu(struct macvlan_rx_stats);
424 	if (!vlan->rx_stats)
425 		return -ENOMEM;
426 
427 	return 0;
428 }
429 
430 static void macvlan_uninit(struct net_device *dev)
431 {
432 	struct macvlan_dev *vlan = netdev_priv(dev);
433 
434 	free_percpu(vlan->rx_stats);
435 }
436 
437 static struct rtnl_link_stats64 *macvlan_dev_get_stats64(struct net_device *dev,
438 							 struct rtnl_link_stats64 *stats)
439 {
440 	struct macvlan_dev *vlan = netdev_priv(dev);
441 
442 	dev_txq_stats_fold(dev, stats);
443 
444 	if (vlan->rx_stats) {
445 		struct macvlan_rx_stats *p, accum = {0};
446 		u64 rx_packets, rx_bytes, rx_multicast;
447 		unsigned int start;
448 		int i;
449 
450 		for_each_possible_cpu(i) {
451 			p = per_cpu_ptr(vlan->rx_stats, i);
452 			do {
453 				start = u64_stats_fetch_begin_bh(&p->syncp);
454 				rx_packets	= p->rx_packets;
455 				rx_bytes	= p->rx_bytes;
456 				rx_multicast	= p->rx_multicast;
457 			} while (u64_stats_fetch_retry_bh(&p->syncp, start));
458 			accum.rx_packets	+= rx_packets;
459 			accum.rx_bytes		+= rx_bytes;
460 			accum.rx_multicast	+= rx_multicast;
461 			/* rx_errors is an ulong, updated without syncp protection */
462 			accum.rx_errors		+= p->rx_errors;
463 		}
464 		stats->rx_packets = accum.rx_packets;
465 		stats->rx_bytes   = accum.rx_bytes;
466 		stats->rx_errors  = accum.rx_errors;
467 		stats->rx_dropped = accum.rx_errors;
468 		stats->multicast  = accum.rx_multicast;
469 	}
470 	return stats;
471 }
472 
473 static void macvlan_ethtool_get_drvinfo(struct net_device *dev,
474 					struct ethtool_drvinfo *drvinfo)
475 {
476 	snprintf(drvinfo->driver, 32, "macvlan");
477 	snprintf(drvinfo->version, 32, "0.1");
478 }
479 
480 static u32 macvlan_ethtool_get_rx_csum(struct net_device *dev)
481 {
482 	const struct macvlan_dev *vlan = netdev_priv(dev);
483 	return dev_ethtool_get_rx_csum(vlan->lowerdev);
484 }
485 
486 static int macvlan_ethtool_get_settings(struct net_device *dev,
487 					struct ethtool_cmd *cmd)
488 {
489 	const struct macvlan_dev *vlan = netdev_priv(dev);
490 	return dev_ethtool_get_settings(vlan->lowerdev, cmd);
491 }
492 
493 static u32 macvlan_ethtool_get_flags(struct net_device *dev)
494 {
495 	const struct macvlan_dev *vlan = netdev_priv(dev);
496 	return dev_ethtool_get_flags(vlan->lowerdev);
497 }
498 
499 static const struct ethtool_ops macvlan_ethtool_ops = {
500 	.get_link		= ethtool_op_get_link,
501 	.get_settings		= macvlan_ethtool_get_settings,
502 	.get_rx_csum		= macvlan_ethtool_get_rx_csum,
503 	.get_drvinfo		= macvlan_ethtool_get_drvinfo,
504 	.get_flags		= macvlan_ethtool_get_flags,
505 };
506 
507 static const struct net_device_ops macvlan_netdev_ops = {
508 	.ndo_init		= macvlan_init,
509 	.ndo_uninit		= macvlan_uninit,
510 	.ndo_open		= macvlan_open,
511 	.ndo_stop		= macvlan_stop,
512 	.ndo_start_xmit		= macvlan_start_xmit,
513 	.ndo_change_mtu		= macvlan_change_mtu,
514 	.ndo_change_rx_flags	= macvlan_change_rx_flags,
515 	.ndo_set_mac_address	= macvlan_set_mac_address,
516 	.ndo_set_multicast_list	= macvlan_set_multicast_list,
517 	.ndo_get_stats64	= macvlan_dev_get_stats64,
518 	.ndo_validate_addr	= eth_validate_addr,
519 };
520 
521 void macvlan_common_setup(struct net_device *dev)
522 {
523 	ether_setup(dev);
524 
525 	dev->priv_flags	       &= ~IFF_XMIT_DST_RELEASE;
526 	dev->netdev_ops		= &macvlan_netdev_ops;
527 	dev->destructor		= free_netdev;
528 	dev->header_ops		= &macvlan_hard_header_ops,
529 	dev->ethtool_ops	= &macvlan_ethtool_ops;
530 }
531 EXPORT_SYMBOL_GPL(macvlan_common_setup);
532 
533 static void macvlan_setup(struct net_device *dev)
534 {
535 	macvlan_common_setup(dev);
536 	dev->tx_queue_len	= 0;
537 }
538 
539 static int macvlan_port_create(struct net_device *dev)
540 {
541 	struct macvlan_port *port;
542 	unsigned int i;
543 	int err;
544 
545 	if (dev->type != ARPHRD_ETHER || dev->flags & IFF_LOOPBACK)
546 		return -EINVAL;
547 
548 	port = kzalloc(sizeof(*port), GFP_KERNEL);
549 	if (port == NULL)
550 		return -ENOMEM;
551 
552 	port->dev = dev;
553 	INIT_LIST_HEAD(&port->vlans);
554 	for (i = 0; i < MACVLAN_HASH_SIZE; i++)
555 		INIT_HLIST_HEAD(&port->vlan_hash[i]);
556 
557 	err = netdev_rx_handler_register(dev, macvlan_handle_frame, port);
558 	if (err)
559 		kfree(port);
560 
561 	dev->priv_flags |= IFF_MACVLAN_PORT;
562 	return err;
563 }
564 
565 static void macvlan_port_rcu_free(struct rcu_head *head)
566 {
567 	struct macvlan_port *port;
568 
569 	port = container_of(head, struct macvlan_port, rcu);
570 	kfree(port);
571 }
572 
573 static void macvlan_port_destroy(struct net_device *dev)
574 {
575 	struct macvlan_port *port = macvlan_port_get(dev);
576 
577 	dev->priv_flags &= ~IFF_MACVLAN_PORT;
578 	netdev_rx_handler_unregister(dev);
579 	call_rcu(&port->rcu, macvlan_port_rcu_free);
580 }
581 
582 static int macvlan_validate(struct nlattr *tb[], struct nlattr *data[])
583 {
584 	if (tb[IFLA_ADDRESS]) {
585 		if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
586 			return -EINVAL;
587 		if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
588 			return -EADDRNOTAVAIL;
589 	}
590 
591 	if (data && data[IFLA_MACVLAN_MODE]) {
592 		switch (nla_get_u32(data[IFLA_MACVLAN_MODE])) {
593 		case MACVLAN_MODE_PRIVATE:
594 		case MACVLAN_MODE_VEPA:
595 		case MACVLAN_MODE_BRIDGE:
596 			break;
597 		default:
598 			return -EINVAL;
599 		}
600 	}
601 	return 0;
602 }
603 
604 static int macvlan_get_tx_queues(struct net *net,
605 				 struct nlattr *tb[],
606 				 unsigned int *num_tx_queues,
607 				 unsigned int *real_num_tx_queues)
608 {
609 	struct net_device *real_dev;
610 
611 	if (!tb[IFLA_LINK])
612 		return -EINVAL;
613 
614 	real_dev = __dev_get_by_index(net, nla_get_u32(tb[IFLA_LINK]));
615 	if (!real_dev)
616 		return -ENODEV;
617 
618 	*num_tx_queues      = real_dev->num_tx_queues;
619 	*real_num_tx_queues = real_dev->real_num_tx_queues;
620 	return 0;
621 }
622 
623 int macvlan_common_newlink(struct net *src_net, struct net_device *dev,
624 			   struct nlattr *tb[], struct nlattr *data[],
625 			   int (*receive)(struct sk_buff *skb),
626 			   int (*forward)(struct net_device *dev,
627 					  struct sk_buff *skb))
628 {
629 	struct macvlan_dev *vlan = netdev_priv(dev);
630 	struct macvlan_port *port;
631 	struct net_device *lowerdev;
632 	int err;
633 
634 	if (!tb[IFLA_LINK])
635 		return -EINVAL;
636 
637 	lowerdev = __dev_get_by_index(src_net, nla_get_u32(tb[IFLA_LINK]));
638 	if (lowerdev == NULL)
639 		return -ENODEV;
640 
641 	/* When creating macvlans on top of other macvlans - use
642 	 * the real device as the lowerdev.
643 	 */
644 	if (lowerdev->rtnl_link_ops == dev->rtnl_link_ops) {
645 		struct macvlan_dev *lowervlan = netdev_priv(lowerdev);
646 		lowerdev = lowervlan->lowerdev;
647 	}
648 
649 	if (!tb[IFLA_MTU])
650 		dev->mtu = lowerdev->mtu;
651 	else if (dev->mtu > lowerdev->mtu)
652 		return -EINVAL;
653 
654 	if (!tb[IFLA_ADDRESS])
655 		random_ether_addr(dev->dev_addr);
656 
657 	if (!macvlan_port_exists(lowerdev)) {
658 		err = macvlan_port_create(lowerdev);
659 		if (err < 0)
660 			return err;
661 	}
662 	port = macvlan_port_get(lowerdev);
663 
664 	vlan->lowerdev = lowerdev;
665 	vlan->dev      = dev;
666 	vlan->port     = port;
667 	vlan->receive  = receive;
668 	vlan->forward  = forward;
669 
670 	vlan->mode     = MACVLAN_MODE_VEPA;
671 	if (data && data[IFLA_MACVLAN_MODE])
672 		vlan->mode = nla_get_u32(data[IFLA_MACVLAN_MODE]);
673 
674 	err = register_netdevice(dev);
675 	if (err < 0)
676 		goto destroy_port;
677 
678 	list_add_tail(&vlan->list, &port->vlans);
679 	netif_stacked_transfer_operstate(lowerdev, dev);
680 
681 	return 0;
682 
683 destroy_port:
684 	if (list_empty(&port->vlans))
685 		macvlan_port_destroy(lowerdev);
686 
687 	return err;
688 }
689 EXPORT_SYMBOL_GPL(macvlan_common_newlink);
690 
691 static int macvlan_newlink(struct net *src_net, struct net_device *dev,
692 			   struct nlattr *tb[], struct nlattr *data[])
693 {
694 	return macvlan_common_newlink(src_net, dev, tb, data,
695 				      netif_rx,
696 				      dev_forward_skb);
697 }
698 
699 void macvlan_dellink(struct net_device *dev, struct list_head *head)
700 {
701 	struct macvlan_dev *vlan = netdev_priv(dev);
702 	struct macvlan_port *port = vlan->port;
703 
704 	list_del(&vlan->list);
705 	unregister_netdevice_queue(dev, head);
706 
707 	if (list_empty(&port->vlans))
708 		macvlan_port_destroy(port->dev);
709 }
710 EXPORT_SYMBOL_GPL(macvlan_dellink);
711 
712 static int macvlan_changelink(struct net_device *dev,
713 		struct nlattr *tb[], struct nlattr *data[])
714 {
715 	struct macvlan_dev *vlan = netdev_priv(dev);
716 	if (data && data[IFLA_MACVLAN_MODE])
717 		vlan->mode = nla_get_u32(data[IFLA_MACVLAN_MODE]);
718 	return 0;
719 }
720 
721 static size_t macvlan_get_size(const struct net_device *dev)
722 {
723 	return nla_total_size(4);
724 }
725 
726 static int macvlan_fill_info(struct sk_buff *skb,
727 				const struct net_device *dev)
728 {
729 	struct macvlan_dev *vlan = netdev_priv(dev);
730 
731 	NLA_PUT_U32(skb, IFLA_MACVLAN_MODE, vlan->mode);
732 	return 0;
733 
734 nla_put_failure:
735 	return -EMSGSIZE;
736 }
737 
738 static const struct nla_policy macvlan_policy[IFLA_MACVLAN_MAX + 1] = {
739 	[IFLA_MACVLAN_MODE] = { .type = NLA_U32 },
740 };
741 
742 int macvlan_link_register(struct rtnl_link_ops *ops)
743 {
744 	/* common fields */
745 	ops->priv_size		= sizeof(struct macvlan_dev);
746 	ops->get_tx_queues	= macvlan_get_tx_queues;
747 	ops->validate		= macvlan_validate;
748 	ops->maxtype		= IFLA_MACVLAN_MAX;
749 	ops->policy		= macvlan_policy;
750 	ops->changelink		= macvlan_changelink;
751 	ops->get_size		= macvlan_get_size;
752 	ops->fill_info		= macvlan_fill_info;
753 
754 	return rtnl_link_register(ops);
755 };
756 EXPORT_SYMBOL_GPL(macvlan_link_register);
757 
758 static struct rtnl_link_ops macvlan_link_ops = {
759 	.kind		= "macvlan",
760 	.setup		= macvlan_setup,
761 	.newlink	= macvlan_newlink,
762 	.dellink	= macvlan_dellink,
763 };
764 
765 static int macvlan_device_event(struct notifier_block *unused,
766 				unsigned long event, void *ptr)
767 {
768 	struct net_device *dev = ptr;
769 	struct macvlan_dev *vlan, *next;
770 	struct macvlan_port *port;
771 
772 	if (!macvlan_port_exists(dev))
773 		return NOTIFY_DONE;
774 
775 	port = macvlan_port_get(dev);
776 
777 	switch (event) {
778 	case NETDEV_CHANGE:
779 		list_for_each_entry(vlan, &port->vlans, list)
780 			netif_stacked_transfer_operstate(vlan->lowerdev,
781 							 vlan->dev);
782 		break;
783 	case NETDEV_FEAT_CHANGE:
784 		list_for_each_entry(vlan, &port->vlans, list) {
785 			vlan->dev->features = dev->features & MACVLAN_FEATURES;
786 			vlan->dev->gso_max_size = dev->gso_max_size;
787 			netdev_features_change(vlan->dev);
788 		}
789 		break;
790 	case NETDEV_UNREGISTER:
791 		list_for_each_entry_safe(vlan, next, &port->vlans, list)
792 			vlan->dev->rtnl_link_ops->dellink(vlan->dev, NULL);
793 		break;
794 	case NETDEV_PRE_TYPE_CHANGE:
795 		/* Forbid underlaying device to change its type. */
796 		return NOTIFY_BAD;
797 	}
798 	return NOTIFY_DONE;
799 }
800 
801 static struct notifier_block macvlan_notifier_block __read_mostly = {
802 	.notifier_call	= macvlan_device_event,
803 };
804 
805 static int __init macvlan_init_module(void)
806 {
807 	int err;
808 
809 	register_netdevice_notifier(&macvlan_notifier_block);
810 
811 	err = macvlan_link_register(&macvlan_link_ops);
812 	if (err < 0)
813 		goto err1;
814 	return 0;
815 err1:
816 	unregister_netdevice_notifier(&macvlan_notifier_block);
817 	return err;
818 }
819 
820 static void __exit macvlan_cleanup_module(void)
821 {
822 	rtnl_link_unregister(&macvlan_link_ops);
823 	unregister_netdevice_notifier(&macvlan_notifier_block);
824 }
825 
826 module_init(macvlan_init_module);
827 module_exit(macvlan_cleanup_module);
828 
829 MODULE_LICENSE("GPL");
830 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
831 MODULE_DESCRIPTION("Driver for MAC address based VLANs");
832 MODULE_ALIAS_RTNL_LINK("macvlan");
833