blob: 2de9b90c5f8f079859b695979acb4871c5a0ea08 [file] [log] [blame]
Pavel Emelyanove314dbd2007-09-25 16:14:46 -07001/*
2 * drivers/net/veth.c
3 *
4 * Copyright (C) 2007 OpenVZ http://openvz.org, SWsoft Inc
5 *
6 * Author: Pavel Emelianov <xemul@openvz.org>
7 * Ethtool interface from: Eric W. Biederman <ebiederm@xmission.com>
8 *
9 */
10
Pavel Emelyanove314dbd2007-09-25 16:14:46 -070011#include <linux/netdevice.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090012#include <linux/slab.h>
Pavel Emelyanove314dbd2007-09-25 16:14:46 -070013#include <linux/ethtool.h>
14#include <linux/etherdevice.h>
15
16#include <net/dst.h>
17#include <net/xfrm.h>
Stephen Hemmingerecef9692007-12-25 17:23:59 -080018#include <linux/veth.h>
Pavel Emelyanove314dbd2007-09-25 16:14:46 -070019
20#define DRV_NAME "veth"
21#define DRV_VERSION "1.0"
22
Eric Biederman38d40812009-03-03 23:36:04 -080023#define MIN_MTU 68 /* Min L3 MTU */
24#define MAX_MTU 65535 /* Max L3 MTU (arbitrary) */
25#define MTU_PAD (ETH_HLEN + 4) /* Max difference between L2 and L3 size MTU */
26
Pavel Emelyanove314dbd2007-09-25 16:14:46 -070027struct veth_net_stats {
28 unsigned long rx_packets;
29 unsigned long tx_packets;
30 unsigned long rx_bytes;
31 unsigned long tx_bytes;
32 unsigned long tx_dropped;
Eric Biederman38d40812009-03-03 23:36:04 -080033 unsigned long rx_dropped;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -070034};
35
36struct veth_priv {
37 struct net_device *peer;
Tejun Heo47d74272010-02-16 15:21:08 +000038 struct veth_net_stats __percpu *stats;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -070039 unsigned ip_summed;
40};
41
Pavel Emelyanove314dbd2007-09-25 16:14:46 -070042/*
43 * ethtool interface
44 */
45
46static struct {
47 const char string[ETH_GSTRING_LEN];
48} ethtool_stats_keys[] = {
49 { "peer_ifindex" },
50};
51
52static int veth_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
53{
54 cmd->supported = 0;
55 cmd->advertising = 0;
56 cmd->speed = SPEED_10000;
57 cmd->duplex = DUPLEX_FULL;
58 cmd->port = PORT_TP;
59 cmd->phy_address = 0;
60 cmd->transceiver = XCVR_INTERNAL;
61 cmd->autoneg = AUTONEG_DISABLE;
62 cmd->maxtxpkt = 0;
63 cmd->maxrxpkt = 0;
64 return 0;
65}
66
67static void veth_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info)
68{
69 strcpy(info->driver, DRV_NAME);
70 strcpy(info->version, DRV_VERSION);
71 strcpy(info->fw_version, "N/A");
72}
73
74static void veth_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
75{
76 switch(stringset) {
77 case ETH_SS_STATS:
78 memcpy(buf, &ethtool_stats_keys, sizeof(ethtool_stats_keys));
79 break;
80 }
81}
82
Jeff Garzikb9f2c042007-10-03 18:07:32 -070083static int veth_get_sset_count(struct net_device *dev, int sset)
Pavel Emelyanove314dbd2007-09-25 16:14:46 -070084{
Jeff Garzikb9f2c042007-10-03 18:07:32 -070085 switch (sset) {
86 case ETH_SS_STATS:
87 return ARRAY_SIZE(ethtool_stats_keys);
88 default:
89 return -EOPNOTSUPP;
90 }
Pavel Emelyanove314dbd2007-09-25 16:14:46 -070091}
92
93static void veth_get_ethtool_stats(struct net_device *dev,
94 struct ethtool_stats *stats, u64 *data)
95{
96 struct veth_priv *priv;
97
98 priv = netdev_priv(dev);
99 data[0] = priv->peer->ifindex;
100}
101
102static u32 veth_get_rx_csum(struct net_device *dev)
103{
104 struct veth_priv *priv;
105
106 priv = netdev_priv(dev);
107 return priv->ip_summed == CHECKSUM_UNNECESSARY;
108}
109
110static int veth_set_rx_csum(struct net_device *dev, u32 data)
111{
112 struct veth_priv *priv;
113
114 priv = netdev_priv(dev);
115 priv->ip_summed = data ? CHECKSUM_UNNECESSARY : CHECKSUM_NONE;
116 return 0;
117}
118
119static u32 veth_get_tx_csum(struct net_device *dev)
120{
121 return (dev->features & NETIF_F_NO_CSUM) != 0;
122}
123
124static int veth_set_tx_csum(struct net_device *dev, u32 data)
125{
126 if (data)
127 dev->features |= NETIF_F_NO_CSUM;
128 else
129 dev->features &= ~NETIF_F_NO_CSUM;
130 return 0;
131}
132
Stephen Hemminger0fc0b732009-09-02 01:03:33 -0700133static const struct ethtool_ops veth_ethtool_ops = {
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700134 .get_settings = veth_get_settings,
135 .get_drvinfo = veth_get_drvinfo,
136 .get_link = ethtool_op_get_link,
137 .get_rx_csum = veth_get_rx_csum,
138 .set_rx_csum = veth_set_rx_csum,
139 .get_tx_csum = veth_get_tx_csum,
140 .set_tx_csum = veth_set_tx_csum,
141 .get_sg = ethtool_op_get_sg,
142 .set_sg = ethtool_op_set_sg,
143 .get_strings = veth_get_strings,
Jeff Garzikb9f2c042007-10-03 18:07:32 -0700144 .get_sset_count = veth_get_sset_count,
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700145 .get_ethtool_stats = veth_get_ethtool_stats,
146};
147
148/*
149 * xmit
150 */
151
Stephen Hemminger424efe92009-08-31 19:50:51 +0000152static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700153{
154 struct net_device *rcv = NULL;
155 struct veth_priv *priv, *rcv_priv;
Eric Biederman38d40812009-03-03 23:36:04 -0800156 struct veth_net_stats *stats, *rcv_stats;
Christoph Lametere7dcaa42009-10-03 19:48:23 +0900157 int length;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700158
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700159 priv = netdev_priv(dev);
160 rcv = priv->peer;
161 rcv_priv = netdev_priv(rcv);
162
Christoph Lametere7dcaa42009-10-03 19:48:23 +0900163 stats = this_cpu_ptr(priv->stats);
164 rcv_stats = this_cpu_ptr(rcv_priv->stats);
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700165
166 if (!(rcv->flags & IFF_UP))
Eric Biederman38d40812009-03-03 23:36:04 -0800167 goto tx_drop;
168
Michał Mirosław0b796752010-12-14 12:35:13 +0000169 /* don't change ip_summed == CHECKSUM_PARTIAL, as that
170 will cause bad checksum on forwarded packets */
171 if (skb->ip_summed == CHECKSUM_NONE)
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700172 skb->ip_summed = rcv_priv->ip_summed;
173
Eric W. Biederman675071a2011-03-21 18:24:53 -0700174 length = skb->len;
Arnd Bergmann44540962009-11-26 06:07:08 +0000175 if (dev_forward_skb(rcv, skb) != NET_RX_SUCCESS)
176 goto rx_drop;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700177
178 stats->tx_bytes += length;
179 stats->tx_packets++;
180
Eric Biederman38d40812009-03-03 23:36:04 -0800181 rcv_stats->rx_bytes += length;
182 rcv_stats->rx_packets++;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700183
Patrick McHardy6ed10652009-06-23 06:03:08 +0000184 return NETDEV_TX_OK;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700185
Eric Biederman38d40812009-03-03 23:36:04 -0800186tx_drop:
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700187 kfree_skb(skb);
188 stats->tx_dropped++;
Patrick McHardy6ed10652009-06-23 06:03:08 +0000189 return NETDEV_TX_OK;
Eric Biederman38d40812009-03-03 23:36:04 -0800190
191rx_drop:
Eric Biederman38d40812009-03-03 23:36:04 -0800192 rcv_stats->rx_dropped++;
Patrick McHardy6ed10652009-06-23 06:03:08 +0000193 return NETDEV_TX_OK;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700194}
195
196/*
197 * general routines
198 */
199
200static struct net_device_stats *veth_get_stats(struct net_device *dev)
201{
David S. Miller11687a12009-06-25 02:45:42 -0700202 struct veth_priv *priv;
David S. Miller11687a12009-06-25 02:45:42 -0700203 int cpu;
Eric Dumazet2b1c8b02009-11-18 07:09:39 +0000204 struct veth_net_stats *stats, total = {0};
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700205
David S. Miller11687a12009-06-25 02:45:42 -0700206 priv = netdev_priv(dev);
David S. Miller11687a12009-06-25 02:45:42 -0700207
Eric Dumazet2b1c8b02009-11-18 07:09:39 +0000208 for_each_possible_cpu(cpu) {
David S. Miller11687a12009-06-25 02:45:42 -0700209 stats = per_cpu_ptr(priv->stats, cpu);
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700210
Eric Dumazet2b1c8b02009-11-18 07:09:39 +0000211 total.rx_packets += stats->rx_packets;
212 total.tx_packets += stats->tx_packets;
213 total.rx_bytes += stats->rx_bytes;
214 total.tx_bytes += stats->tx_bytes;
215 total.tx_dropped += stats->tx_dropped;
216 total.rx_dropped += stats->rx_dropped;
David S. Miller11687a12009-06-25 02:45:42 -0700217 }
Eric Dumazet2b1c8b02009-11-18 07:09:39 +0000218 dev->stats.rx_packets = total.rx_packets;
219 dev->stats.tx_packets = total.tx_packets;
220 dev->stats.rx_bytes = total.rx_bytes;
221 dev->stats.tx_bytes = total.tx_bytes;
222 dev->stats.tx_dropped = total.tx_dropped;
223 dev->stats.rx_dropped = total.rx_dropped;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700224
Eric Dumazet2b1c8b02009-11-18 07:09:39 +0000225 return &dev->stats;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700226}
227
228static int veth_open(struct net_device *dev)
229{
230 struct veth_priv *priv;
231
232 priv = netdev_priv(dev);
233 if (priv->peer == NULL)
234 return -ENOTCONN;
235
236 if (priv->peer->flags & IFF_UP) {
237 netif_carrier_on(dev);
238 netif_carrier_on(priv->peer);
239 }
240 return 0;
241}
242
Eric W. Biederman2cf48a12009-02-25 19:47:29 +0000243static int veth_close(struct net_device *dev)
244{
245 struct veth_priv *priv = netdev_priv(dev);
246
247 netif_carrier_off(dev);
248 netif_carrier_off(priv->peer);
249
250 return 0;
251}
252
Eric Biederman38d40812009-03-03 23:36:04 -0800253static int is_valid_veth_mtu(int new_mtu)
254{
Eric Dumazet807540b2010-09-23 05:40:09 +0000255 return new_mtu >= MIN_MTU && new_mtu <= MAX_MTU;
Eric Biederman38d40812009-03-03 23:36:04 -0800256}
257
258static int veth_change_mtu(struct net_device *dev, int new_mtu)
259{
260 if (!is_valid_veth_mtu(new_mtu))
261 return -EINVAL;
262 dev->mtu = new_mtu;
263 return 0;
264}
265
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700266static int veth_dev_init(struct net_device *dev)
267{
Tejun Heo47d74272010-02-16 15:21:08 +0000268 struct veth_net_stats __percpu *stats;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700269 struct veth_priv *priv;
270
271 stats = alloc_percpu(struct veth_net_stats);
272 if (stats == NULL)
273 return -ENOMEM;
274
275 priv = netdev_priv(dev);
276 priv->stats = stats;
277 return 0;
278}
279
David S. Miller11687a12009-06-25 02:45:42 -0700280static void veth_dev_free(struct net_device *dev)
281{
282 struct veth_priv *priv;
283
284 priv = netdev_priv(dev);
285 free_percpu(priv->stats);
286 free_netdev(dev);
287}
288
Stephen Hemminger4456e7b2008-11-19 21:50:10 -0800289static const struct net_device_ops veth_netdev_ops = {
Daniel Lezcanoee923622009-02-22 00:04:45 -0800290 .ndo_init = veth_dev_init,
291 .ndo_open = veth_open,
Eric W. Biederman2cf48a12009-02-25 19:47:29 +0000292 .ndo_stop = veth_close,
Daniel Lezcanoee923622009-02-22 00:04:45 -0800293 .ndo_start_xmit = veth_xmit,
Eric Biederman38d40812009-03-03 23:36:04 -0800294 .ndo_change_mtu = veth_change_mtu,
Daniel Lezcanoee923622009-02-22 00:04:45 -0800295 .ndo_get_stats = veth_get_stats,
296 .ndo_set_mac_address = eth_mac_addr,
Stephen Hemminger4456e7b2008-11-19 21:50:10 -0800297};
298
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700299static void veth_setup(struct net_device *dev)
300{
301 ether_setup(dev);
302
Stephen Hemminger4456e7b2008-11-19 21:50:10 -0800303 dev->netdev_ops = &veth_netdev_ops;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700304 dev->ethtool_ops = &veth_ethtool_ops;
305 dev->features |= NETIF_F_LLTX;
David S. Miller11687a12009-06-25 02:45:42 -0700306 dev->destructor = veth_dev_free;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700307}
308
309/*
310 * netlink interface
311 */
312
313static int veth_validate(struct nlattr *tb[], struct nlattr *data[])
314{
315 if (tb[IFLA_ADDRESS]) {
316 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
317 return -EINVAL;
318 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
319 return -EADDRNOTAVAIL;
320 }
Eric Biederman38d40812009-03-03 23:36:04 -0800321 if (tb[IFLA_MTU]) {
322 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
323 return -EINVAL;
324 }
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700325 return 0;
326}
327
328static struct rtnl_link_ops veth_link_ops;
329
Eric W. Biederman81adee42009-11-08 00:53:51 -0800330static int veth_newlink(struct net *src_net, struct net_device *dev,
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700331 struct nlattr *tb[], struct nlattr *data[])
332{
333 int err;
334 struct net_device *peer;
335 struct veth_priv *priv;
336 char ifname[IFNAMSIZ];
337 struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
Patrick McHardy3729d502010-02-26 06:34:54 +0000338 struct ifinfomsg *ifmp;
Eric W. Biederman81adee42009-11-08 00:53:51 -0800339 struct net *net;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700340
341 /*
342 * create and register peer first
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700343 */
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700344 if (data != NULL && data[VETH_INFO_PEER] != NULL) {
345 struct nlattr *nla_peer;
346
347 nla_peer = data[VETH_INFO_PEER];
Patrick McHardy3729d502010-02-26 06:34:54 +0000348 ifmp = nla_data(nla_peer);
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700349 err = nla_parse(peer_tb, IFLA_MAX,
350 nla_data(nla_peer) + sizeof(struct ifinfomsg),
351 nla_len(nla_peer) - sizeof(struct ifinfomsg),
352 ifla_policy);
353 if (err < 0)
354 return err;
355
356 err = veth_validate(peer_tb, NULL);
357 if (err < 0)
358 return err;
359
360 tbp = peer_tb;
Patrick McHardy3729d502010-02-26 06:34:54 +0000361 } else {
362 ifmp = NULL;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700363 tbp = tb;
Patrick McHardy3729d502010-02-26 06:34:54 +0000364 }
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700365
366 if (tbp[IFLA_IFNAME])
367 nla_strlcpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
368 else
369 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
370
Eric W. Biederman81adee42009-11-08 00:53:51 -0800371 net = rtnl_link_get_net(src_net, tbp);
372 if (IS_ERR(net))
373 return PTR_ERR(net);
374
375 peer = rtnl_create_link(src_net, net, ifname, &veth_link_ops, tbp);
376 if (IS_ERR(peer)) {
377 put_net(net);
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700378 return PTR_ERR(peer);
Eric W. Biederman81adee42009-11-08 00:53:51 -0800379 }
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700380
381 if (tbp[IFLA_ADDRESS] == NULL)
382 random_ether_addr(peer->dev_addr);
383
384 err = register_netdevice(peer);
Eric W. Biederman81adee42009-11-08 00:53:51 -0800385 put_net(net);
386 net = NULL;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700387 if (err < 0)
388 goto err_register_peer;
389
390 netif_carrier_off(peer);
391
Patrick McHardy3729d502010-02-26 06:34:54 +0000392 err = rtnl_configure_link(peer, ifmp);
393 if (err < 0)
394 goto err_configure_peer;
395
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700396 /*
397 * register dev last
398 *
399 * note, that since we've registered new device the dev's name
400 * should be re-allocated
401 */
402
403 if (tb[IFLA_ADDRESS] == NULL)
404 random_ether_addr(dev->dev_addr);
405
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700406 err = register_netdevice(dev);
407 if (err < 0)
408 goto err_register_dev;
409
410 netif_carrier_off(dev);
411
412 /*
413 * tie the deviced together
414 */
415
416 priv = netdev_priv(dev);
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700417 priv->peer = peer;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700418
419 priv = netdev_priv(peer);
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700420 priv->peer = dev;
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700421 return 0;
422
423err_register_dev:
424 /* nothing to do */
Patrick McHardy3729d502010-02-26 06:34:54 +0000425err_configure_peer:
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700426 unregister_netdevice(peer);
427 return err;
428
429err_register_peer:
430 free_netdev(peer);
431 return err;
432}
433
Eric Dumazet23289a32009-10-27 07:06:36 +0000434static void veth_dellink(struct net_device *dev, struct list_head *head)
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700435{
436 struct veth_priv *priv;
437 struct net_device *peer;
438
439 priv = netdev_priv(dev);
440 peer = priv->peer;
441
Eric Dumazet24540532009-10-30 01:00:27 -0700442 unregister_netdevice_queue(dev, head);
443 unregister_netdevice_queue(peer, head);
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700444}
445
446static const struct nla_policy veth_policy[VETH_INFO_MAX + 1];
447
448static struct rtnl_link_ops veth_link_ops = {
449 .kind = DRV_NAME,
450 .priv_size = sizeof(struct veth_priv),
451 .setup = veth_setup,
452 .validate = veth_validate,
453 .newlink = veth_newlink,
454 .dellink = veth_dellink,
455 .policy = veth_policy,
456 .maxtype = VETH_INFO_MAX,
457};
458
459/*
460 * init/fini
461 */
462
463static __init int veth_init(void)
464{
465 return rtnl_link_register(&veth_link_ops);
466}
467
468static __exit void veth_exit(void)
469{
Patrick McHardy68365452008-01-20 17:25:14 -0800470 rtnl_link_unregister(&veth_link_ops);
Pavel Emelyanove314dbd2007-09-25 16:14:46 -0700471}
472
473module_init(veth_init);
474module_exit(veth_exit);
475
476MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
477MODULE_LICENSE("GPL v2");
478MODULE_ALIAS_RTNL_LINK(DRV_NAME);