Commit | Line | Data |
---|---|---|
e314dbdc PE |
1 | /* |
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 | ||
e314dbdc | 11 | #include <linux/netdevice.h> |
5a0e3ad6 | 12 | #include <linux/slab.h> |
e314dbdc PE |
13 | #include <linux/ethtool.h> |
14 | #include <linux/etherdevice.h> | |
cf05c700 | 15 | #include <linux/u64_stats_sync.h> |
e314dbdc PE |
16 | |
17 | #include <net/dst.h> | |
18 | #include <net/xfrm.h> | |
ecef969e | 19 | #include <linux/veth.h> |
9d9779e7 | 20 | #include <linux/module.h> |
e314dbdc PE |
21 | |
22 | #define DRV_NAME "veth" | |
23 | #define DRV_VERSION "1.0" | |
24 | ||
38d40815 EB |
25 | #define MIN_MTU 68 /* Min L3 MTU */ |
26 | #define MAX_MTU 65535 /* Max L3 MTU (arbitrary) */ | |
38d40815 | 27 | |
2681128f ED |
28 | struct pcpu_vstats { |
29 | u64 packets; | |
30 | u64 bytes; | |
cf05c700 | 31 | struct u64_stats_sync syncp; |
e314dbdc PE |
32 | }; |
33 | ||
34 | struct veth_priv { | |
d0e2c55e | 35 | struct net_device __rcu *peer; |
2681128f | 36 | atomic64_t dropped; |
e314dbdc PE |
37 | }; |
38 | ||
e314dbdc PE |
39 | /* |
40 | * ethtool interface | |
41 | */ | |
42 | ||
43 | static struct { | |
44 | const char string[ETH_GSTRING_LEN]; | |
45 | } ethtool_stats_keys[] = { | |
46 | { "peer_ifindex" }, | |
47 | }; | |
48 | ||
49 | static int veth_get_settings(struct net_device *dev, struct ethtool_cmd *cmd) | |
50 | { | |
51 | cmd->supported = 0; | |
52 | cmd->advertising = 0; | |
70739497 | 53 | ethtool_cmd_speed_set(cmd, SPEED_10000); |
e314dbdc PE |
54 | cmd->duplex = DUPLEX_FULL; |
55 | cmd->port = PORT_TP; | |
56 | cmd->phy_address = 0; | |
57 | cmd->transceiver = XCVR_INTERNAL; | |
58 | cmd->autoneg = AUTONEG_DISABLE; | |
59 | cmd->maxtxpkt = 0; | |
60 | cmd->maxrxpkt = 0; | |
61 | return 0; | |
62 | } | |
63 | ||
64 | static void veth_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info) | |
65 | { | |
33a5ba14 RJ |
66 | strlcpy(info->driver, DRV_NAME, sizeof(info->driver)); |
67 | strlcpy(info->version, DRV_VERSION, sizeof(info->version)); | |
e314dbdc PE |
68 | } |
69 | ||
70 | static void veth_get_strings(struct net_device *dev, u32 stringset, u8 *buf) | |
71 | { | |
72 | switch(stringset) { | |
73 | case ETH_SS_STATS: | |
74 | memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys)); | |
75 | break; | |
76 | } | |
77 | } | |
78 | ||
b9f2c044 | 79 | static int veth_get_sset_count(struct net_device *dev, int sset) |
e314dbdc | 80 | { |
b9f2c044 JG |
81 | switch (sset) { |
82 | case ETH_SS_STATS: | |
83 | return ARRAY_SIZE(ethtool_stats_keys); | |
84 | default: | |
85 | return -EOPNOTSUPP; | |
86 | } | |
e314dbdc PE |
87 | } |
88 | ||
89 | static void veth_get_ethtool_stats(struct net_device *dev, | |
90 | struct ethtool_stats *stats, u64 *data) | |
91 | { | |
d0e2c55e ED |
92 | struct veth_priv *priv = netdev_priv(dev); |
93 | struct net_device *peer = rtnl_dereference(priv->peer); | |
e314dbdc | 94 | |
d0e2c55e | 95 | data[0] = peer ? peer->ifindex : 0; |
e314dbdc PE |
96 | } |
97 | ||
0fc0b732 | 98 | static const struct ethtool_ops veth_ethtool_ops = { |
e314dbdc PE |
99 | .get_settings = veth_get_settings, |
100 | .get_drvinfo = veth_get_drvinfo, | |
101 | .get_link = ethtool_op_get_link, | |
e314dbdc | 102 | .get_strings = veth_get_strings, |
b9f2c044 | 103 | .get_sset_count = veth_get_sset_count, |
e314dbdc PE |
104 | .get_ethtool_stats = veth_get_ethtool_stats, |
105 | }; | |
106 | ||
424efe9c | 107 | static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev) |
e314dbdc | 108 | { |
2681128f | 109 | struct veth_priv *priv = netdev_priv(dev); |
d0e2c55e | 110 | struct net_device *rcv; |
2681128f | 111 | int length = skb->len; |
e314dbdc | 112 | |
d0e2c55e ED |
113 | rcu_read_lock(); |
114 | rcv = rcu_dereference(priv->peer); | |
115 | if (unlikely(!rcv)) { | |
116 | kfree_skb(skb); | |
117 | goto drop; | |
118 | } | |
0b796750 | 119 | /* don't change ip_summed == CHECKSUM_PARTIAL, as that |
2681128f ED |
120 | * will cause bad checksum on forwarded packets |
121 | */ | |
a2c725fa MM |
122 | if (skb->ip_summed == CHECKSUM_NONE && |
123 | rcv->features & NETIF_F_RXCSUM) | |
124 | skb->ip_summed = CHECKSUM_UNNECESSARY; | |
e314dbdc | 125 | |
2681128f ED |
126 | if (likely(dev_forward_skb(rcv, skb) == NET_RX_SUCCESS)) { |
127 | struct pcpu_vstats *stats = this_cpu_ptr(dev->vstats); | |
e314dbdc | 128 | |
2681128f ED |
129 | u64_stats_update_begin(&stats->syncp); |
130 | stats->bytes += length; | |
131 | stats->packets++; | |
132 | u64_stats_update_end(&stats->syncp); | |
133 | } else { | |
d0e2c55e | 134 | drop: |
2681128f ED |
135 | atomic64_inc(&priv->dropped); |
136 | } | |
d0e2c55e | 137 | rcu_read_unlock(); |
6ed10654 | 138 | return NETDEV_TX_OK; |
e314dbdc PE |
139 | } |
140 | ||
141 | /* | |
142 | * general routines | |
143 | */ | |
144 | ||
2681128f | 145 | static u64 veth_stats_one(struct pcpu_vstats *result, struct net_device *dev) |
e314dbdc | 146 | { |
cf05c700 | 147 | struct veth_priv *priv = netdev_priv(dev); |
11687a10 | 148 | int cpu; |
e314dbdc | 149 | |
2681128f ED |
150 | result->packets = 0; |
151 | result->bytes = 0; | |
2b1c8b0f | 152 | for_each_possible_cpu(cpu) { |
2681128f ED |
153 | struct pcpu_vstats *stats = per_cpu_ptr(dev->vstats, cpu); |
154 | u64 packets, bytes; | |
cf05c700 ED |
155 | unsigned int start; |
156 | ||
157 | do { | |
158 | start = u64_stats_fetch_begin_bh(&stats->syncp); | |
2681128f ED |
159 | packets = stats->packets; |
160 | bytes = stats->bytes; | |
cf05c700 | 161 | } while (u64_stats_fetch_retry_bh(&stats->syncp, start)); |
2681128f ED |
162 | result->packets += packets; |
163 | result->bytes += bytes; | |
11687a10 | 164 | } |
2681128f ED |
165 | return atomic64_read(&priv->dropped); |
166 | } | |
167 | ||
168 | static struct rtnl_link_stats64 *veth_get_stats64(struct net_device *dev, | |
169 | struct rtnl_link_stats64 *tot) | |
170 | { | |
171 | struct veth_priv *priv = netdev_priv(dev); | |
d0e2c55e | 172 | struct net_device *peer; |
2681128f ED |
173 | struct pcpu_vstats one; |
174 | ||
175 | tot->tx_dropped = veth_stats_one(&one, dev); | |
176 | tot->tx_bytes = one.bytes; | |
177 | tot->tx_packets = one.packets; | |
178 | ||
d0e2c55e ED |
179 | rcu_read_lock(); |
180 | peer = rcu_dereference(priv->peer); | |
181 | if (peer) { | |
182 | tot->rx_dropped = veth_stats_one(&one, peer); | |
183 | tot->rx_bytes = one.bytes; | |
184 | tot->rx_packets = one.packets; | |
185 | } | |
186 | rcu_read_unlock(); | |
6311cc44 | 187 | |
188 | return tot; | |
e314dbdc PE |
189 | } |
190 | ||
191 | static int veth_open(struct net_device *dev) | |
192 | { | |
d0e2c55e ED |
193 | struct veth_priv *priv = netdev_priv(dev); |
194 | struct net_device *peer = rtnl_dereference(priv->peer); | |
e314dbdc | 195 | |
d0e2c55e | 196 | if (!peer) |
e314dbdc PE |
197 | return -ENOTCONN; |
198 | ||
d0e2c55e | 199 | if (peer->flags & IFF_UP) { |
e314dbdc | 200 | netif_carrier_on(dev); |
d0e2c55e | 201 | netif_carrier_on(peer); |
e314dbdc PE |
202 | } |
203 | return 0; | |
204 | } | |
205 | ||
2cf48a10 EB |
206 | static int veth_close(struct net_device *dev) |
207 | { | |
208 | struct veth_priv *priv = netdev_priv(dev); | |
2efd32ee | 209 | struct net_device *peer = rtnl_dereference(priv->peer); |
2cf48a10 EB |
210 | |
211 | netif_carrier_off(dev); | |
2efd32ee ED |
212 | if (peer) |
213 | netif_carrier_off(peer); | |
2cf48a10 EB |
214 | |
215 | return 0; | |
216 | } | |
217 | ||
38d40815 EB |
218 | static int is_valid_veth_mtu(int new_mtu) |
219 | { | |
807540ba | 220 | return new_mtu >= MIN_MTU && new_mtu <= MAX_MTU; |
38d40815 EB |
221 | } |
222 | ||
223 | static int veth_change_mtu(struct net_device *dev, int new_mtu) | |
224 | { | |
225 | if (!is_valid_veth_mtu(new_mtu)) | |
226 | return -EINVAL; | |
227 | dev->mtu = new_mtu; | |
228 | return 0; | |
229 | } | |
230 | ||
e314dbdc PE |
231 | static int veth_dev_init(struct net_device *dev) |
232 | { | |
2681128f ED |
233 | dev->vstats = alloc_percpu(struct pcpu_vstats); |
234 | if (!dev->vstats) | |
e314dbdc PE |
235 | return -ENOMEM; |
236 | ||
e314dbdc PE |
237 | return 0; |
238 | } | |
239 | ||
11687a10 DM |
240 | static void veth_dev_free(struct net_device *dev) |
241 | { | |
2681128f | 242 | free_percpu(dev->vstats); |
11687a10 DM |
243 | free_netdev(dev); |
244 | } | |
245 | ||
4456e7bd | 246 | static const struct net_device_ops veth_netdev_ops = { |
ee923623 DL |
247 | .ndo_init = veth_dev_init, |
248 | .ndo_open = veth_open, | |
2cf48a10 | 249 | .ndo_stop = veth_close, |
ee923623 | 250 | .ndo_start_xmit = veth_xmit, |
38d40815 | 251 | .ndo_change_mtu = veth_change_mtu, |
6311cc44 | 252 | .ndo_get_stats64 = veth_get_stats64, |
ee923623 | 253 | .ndo_set_mac_address = eth_mac_addr, |
4456e7bd SH |
254 | }; |
255 | ||
8093315a ED |
256 | #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_ALL_TSO | \ |
257 | NETIF_F_HW_CSUM | NETIF_F_RXCSUM | NETIF_F_HIGHDMA | \ | |
258 | NETIF_F_HW_VLAN_TX | NETIF_F_HW_VLAN_RX) | |
259 | ||
e314dbdc PE |
260 | static void veth_setup(struct net_device *dev) |
261 | { | |
262 | ether_setup(dev); | |
263 | ||
550fd08c | 264 | dev->priv_flags &= ~IFF_TX_SKB_SHARING; |
23ea5a96 | 265 | dev->priv_flags |= IFF_LIVE_ADDR_CHANGE; |
550fd08c | 266 | |
4456e7bd | 267 | dev->netdev_ops = &veth_netdev_ops; |
e314dbdc PE |
268 | dev->ethtool_ops = &veth_ethtool_ops; |
269 | dev->features |= NETIF_F_LLTX; | |
8093315a | 270 | dev->features |= VETH_FEATURES; |
11687a10 | 271 | dev->destructor = veth_dev_free; |
a2c725fa | 272 | |
8093315a | 273 | dev->hw_features = VETH_FEATURES; |
e314dbdc PE |
274 | } |
275 | ||
276 | /* | |
277 | * netlink interface | |
278 | */ | |
279 | ||
280 | static int veth_validate(struct nlattr *tb[], struct nlattr *data[]) | |
281 | { | |
282 | if (tb[IFLA_ADDRESS]) { | |
283 | if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) | |
284 | return -EINVAL; | |
285 | if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) | |
286 | return -EADDRNOTAVAIL; | |
287 | } | |
38d40815 EB |
288 | if (tb[IFLA_MTU]) { |
289 | if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU]))) | |
290 | return -EINVAL; | |
291 | } | |
e314dbdc PE |
292 | return 0; |
293 | } | |
294 | ||
295 | static struct rtnl_link_ops veth_link_ops; | |
296 | ||
81adee47 | 297 | static int veth_newlink(struct net *src_net, struct net_device *dev, |
e314dbdc PE |
298 | struct nlattr *tb[], struct nlattr *data[]) |
299 | { | |
300 | int err; | |
301 | struct net_device *peer; | |
302 | struct veth_priv *priv; | |
303 | char ifname[IFNAMSIZ]; | |
304 | struct nlattr *peer_tb[IFLA_MAX + 1], **tbp; | |
3729d502 | 305 | struct ifinfomsg *ifmp; |
81adee47 | 306 | struct net *net; |
e314dbdc PE |
307 | |
308 | /* | |
309 | * create and register peer first | |
e314dbdc | 310 | */ |
e314dbdc PE |
311 | if (data != NULL && data[VETH_INFO_PEER] != NULL) { |
312 | struct nlattr *nla_peer; | |
313 | ||
314 | nla_peer = data[VETH_INFO_PEER]; | |
3729d502 | 315 | ifmp = nla_data(nla_peer); |
e314dbdc PE |
316 | err = nla_parse(peer_tb, IFLA_MAX, |
317 | nla_data(nla_peer) + sizeof(struct ifinfomsg), | |
318 | nla_len(nla_peer) - sizeof(struct ifinfomsg), | |
319 | ifla_policy); | |
320 | if (err < 0) | |
321 | return err; | |
322 | ||
323 | err = veth_validate(peer_tb, NULL); | |
324 | if (err < 0) | |
325 | return err; | |
326 | ||
327 | tbp = peer_tb; | |
3729d502 PM |
328 | } else { |
329 | ifmp = NULL; | |
e314dbdc | 330 | tbp = tb; |
3729d502 | 331 | } |
e314dbdc PE |
332 | |
333 | if (tbp[IFLA_IFNAME]) | |
334 | nla_strlcpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ); | |
335 | else | |
336 | snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d"); | |
337 | ||
81adee47 EB |
338 | net = rtnl_link_get_net(src_net, tbp); |
339 | if (IS_ERR(net)) | |
340 | return PTR_ERR(net); | |
341 | ||
c0713563 | 342 | peer = rtnl_create_link(net, ifname, &veth_link_ops, tbp); |
81adee47 EB |
343 | if (IS_ERR(peer)) { |
344 | put_net(net); | |
e314dbdc | 345 | return PTR_ERR(peer); |
81adee47 | 346 | } |
e314dbdc PE |
347 | |
348 | if (tbp[IFLA_ADDRESS] == NULL) | |
f2cedb63 | 349 | eth_hw_addr_random(peer); |
e6f8f1a7 PE |
350 | |
351 | if (ifmp && (dev->ifindex != 0)) | |
352 | peer->ifindex = ifmp->ifi_index; | |
e314dbdc PE |
353 | |
354 | err = register_netdevice(peer); | |
81adee47 EB |
355 | put_net(net); |
356 | net = NULL; | |
e314dbdc PE |
357 | if (err < 0) |
358 | goto err_register_peer; | |
359 | ||
360 | netif_carrier_off(peer); | |
361 | ||
3729d502 PM |
362 | err = rtnl_configure_link(peer, ifmp); |
363 | if (err < 0) | |
364 | goto err_configure_peer; | |
365 | ||
e314dbdc PE |
366 | /* |
367 | * register dev last | |
368 | * | |
369 | * note, that since we've registered new device the dev's name | |
370 | * should be re-allocated | |
371 | */ | |
372 | ||
373 | if (tb[IFLA_ADDRESS] == NULL) | |
f2cedb63 | 374 | eth_hw_addr_random(dev); |
e314dbdc | 375 | |
6c8c4446 JP |
376 | if (tb[IFLA_IFNAME]) |
377 | nla_strlcpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ); | |
378 | else | |
379 | snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d"); | |
380 | ||
381 | if (strchr(dev->name, '%')) { | |
382 | err = dev_alloc_name(dev, dev->name); | |
383 | if (err < 0) | |
384 | goto err_alloc_name; | |
385 | } | |
386 | ||
e314dbdc PE |
387 | err = register_netdevice(dev); |
388 | if (err < 0) | |
389 | goto err_register_dev; | |
390 | ||
391 | netif_carrier_off(dev); | |
392 | ||
393 | /* | |
394 | * tie the deviced together | |
395 | */ | |
396 | ||
397 | priv = netdev_priv(dev); | |
d0e2c55e | 398 | rcu_assign_pointer(priv->peer, peer); |
e314dbdc PE |
399 | |
400 | priv = netdev_priv(peer); | |
d0e2c55e | 401 | rcu_assign_pointer(priv->peer, dev); |
e314dbdc PE |
402 | return 0; |
403 | ||
404 | err_register_dev: | |
405 | /* nothing to do */ | |
6c8c4446 | 406 | err_alloc_name: |
3729d502 | 407 | err_configure_peer: |
e314dbdc PE |
408 | unregister_netdevice(peer); |
409 | return err; | |
410 | ||
411 | err_register_peer: | |
412 | free_netdev(peer); | |
413 | return err; | |
414 | } | |
415 | ||
23289a37 | 416 | static void veth_dellink(struct net_device *dev, struct list_head *head) |
e314dbdc PE |
417 | { |
418 | struct veth_priv *priv; | |
419 | struct net_device *peer; | |
420 | ||
421 | priv = netdev_priv(dev); | |
d0e2c55e ED |
422 | peer = rtnl_dereference(priv->peer); |
423 | ||
424 | /* Note : dellink() is called from default_device_exit_batch(), | |
425 | * before a rcu_synchronize() point. The devices are guaranteed | |
426 | * not being freed before one RCU grace period. | |
427 | */ | |
428 | RCU_INIT_POINTER(priv->peer, NULL); | |
24540535 | 429 | unregister_netdevice_queue(dev, head); |
f45a5c26 ED |
430 | |
431 | if (peer) { | |
432 | priv = netdev_priv(peer); | |
433 | RCU_INIT_POINTER(priv->peer, NULL); | |
434 | unregister_netdevice_queue(peer, head); | |
435 | } | |
e314dbdc PE |
436 | } |
437 | ||
23711438 TG |
438 | static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = { |
439 | [VETH_INFO_PEER] = { .len = sizeof(struct ifinfomsg) }, | |
440 | }; | |
e314dbdc PE |
441 | |
442 | static struct rtnl_link_ops veth_link_ops = { | |
443 | .kind = DRV_NAME, | |
444 | .priv_size = sizeof(struct veth_priv), | |
445 | .setup = veth_setup, | |
446 | .validate = veth_validate, | |
447 | .newlink = veth_newlink, | |
448 | .dellink = veth_dellink, | |
449 | .policy = veth_policy, | |
450 | .maxtype = VETH_INFO_MAX, | |
451 | }; | |
452 | ||
453 | /* | |
454 | * init/fini | |
455 | */ | |
456 | ||
457 | static __init int veth_init(void) | |
458 | { | |
459 | return rtnl_link_register(&veth_link_ops); | |
460 | } | |
461 | ||
462 | static __exit void veth_exit(void) | |
463 | { | |
68365458 | 464 | rtnl_link_unregister(&veth_link_ops); |
e314dbdc PE |
465 | } |
466 | ||
467 | module_init(veth_init); | |
468 | module_exit(veth_exit); | |
469 | ||
470 | MODULE_DESCRIPTION("Virtual Ethernet Tunnel"); | |
471 | MODULE_LICENSE("GPL v2"); | |
472 | MODULE_ALIAS_RTNL_LINK(DRV_NAME); |