projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
vxlan: break dependency with netdev drivers
[deliverable/linux.git] / drivers / net / geneve.c
CommitLineData
2d07dc79
JL		 1/*
2 * GENEVE: Generic Network Virtualization Encapsulation
3 *
4 * Copyright (c) 2015 Red Hat, Inc.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
9 */
10
11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12
13#include <linux/kernel.h>
14#include <linux/module.h>
15#include <linux/netdevice.h>
16#include <linux/etherdevice.h>
17#include <linux/hash.h>
e305ac6c 18#include <net/dst_metadata.h>
8e816df8 19#include <net/gro_cells.h>
2d07dc79
JL		 20#include <net/rtnetlink.h>
21#include <net/geneve.h>
371bd106 22#include <net/protocol.h>
2d07dc79
JL		 23
24#define GENEVE_NETDEV_VER "0.6"
25
26#define GENEVE_UDP_PORT 6081
27
28#define GENEVE_N_VID (1u << 24)
29#define GENEVE_VID_MASK (GENEVE_N_VID - 1)
30
31#define VNI_HASH_BITS 10
32#define VNI_HASH_SIZE (1<<VNI_HASH_BITS)
33
34static bool log_ecn_error = true;
35module_param(log_ecn_error, bool, 0644);
36MODULE_PARM_DESC(log_ecn_error, "Log packets received with corrupted ECN");
37
371bd106
PS		 38#define GENEVE_VER 0
39#define GENEVE_BASE_HLEN (sizeof(struct udphdr) + sizeof(struct genevehdr))
40
2d07dc79
JL		 41/* per-network namespace private data for this module */
42struct geneve_net {
371bd106 43 struct list_head geneve_list;
371bd106 44 struct list_head sock_list;
2d07dc79
JL		 45};
46
371bd106
PS		 47static int geneve_net_id;
48
8ed66f0e
JL		 49union geneve_addr {
50 struct sockaddr_in sin;
51 struct sockaddr_in6 sin6;
52 struct sockaddr sa;
53};
54
55static union geneve_addr geneve_remote_unspec = { .sa.sa_family = AF_UNSPEC, };
56
2d07dc79
JL		 57/* Pseudo network device */
58struct geneve_dev {
59 struct hlist_node hlist; /* vni hash table */
60 struct net *net; /* netns for packet i/o */
61 struct net_device *dev; /* netdev for geneve tunnel */
8ed66f0e
JL		 62 struct geneve_sock *sock4; /* IPv4 socket used for geneve tunnel */
63#if IS_ENABLED(CONFIG_IPV6)
64 struct geneve_sock *sock6; /* IPv6 socket used for geneve tunnel */
65#endif
8760ce58
JL		 66 u8 vni[3]; /* virtual network ID for tunnel */
67 u8 ttl; /* TTL override */
d8951125 68 u8 tos; /* TOS override */
8ed66f0e 69 union geneve_addr remote; /* IP address for link partner */
2d07dc79 70 struct list_head next; /* geneve's per namespace list */
8eb3b995 71 __be32 label; /* IPv6 flowlabel override */
cd7918b3 72 __be16 dst_port;
e305ac6c 73 bool collect_md;
8e816df8 74 struct gro_cells gro_cells;
abe492b4 75 u32 flags;
468dfffc 76 struct dst_cache dst_cache;
2d07dc79
JL		 77};
78
abe492b4 79/* Geneve device flags */
14f1f724 80#define GENEVE_F_UDP_ZERO_CSUM_TX BIT(0)
abe492b4
TH		 81#define GENEVE_F_UDP_ZERO_CSUM6_TX BIT(1)
82#define GENEVE_F_UDP_ZERO_CSUM6_RX BIT(2)
83
371bd106
PS		 84struct geneve_sock {
85 bool collect_md;
371bd106
PS		 86 struct list_head list;
87 struct socket *sock;
88 struct rcu_head rcu;
89 int refcnt;
66d47003 90 struct hlist_head vni_list[VNI_HASH_SIZE];
abe492b4 91 u32 flags;
371bd106 92};
2d07dc79
JL		 93
94static inline __u32 geneve_net_vni_hash(u8 vni[3])
95{
96 __u32 vnid;
97
98 vnid = (vni[0] << 16) | (vni[1] << 8) | vni[2];
99 return hash_32(vnid, VNI_HASH_BITS);
100}
101
e305ac6c
PS		 102static __be64 vni_to_tunnel_id(const __u8 *vni)
103{
104#ifdef __BIG_ENDIAN
105 return (vni[0] << 16) | (vni[1] << 8) | vni[2];
106#else
107 return (__force __be64)(((__force u64)vni[0] << 40) |
108 ((__force u64)vni[1] << 48) |
109 ((__force u64)vni[2] << 56));
110#endif
111}
112
1e9f12ec
JB		 113static sa_family_t geneve_get_sk_family(struct geneve_sock *gs)
114{
115 return gs->sock->sk->sk_family;
116}
117
66d47003 118static struct geneve_dev *geneve_lookup(struct geneve_sock *gs,
371bd106 119 __be32 addr, u8 vni[])
2d07dc79 120{
2d07dc79 121 struct hlist_head *vni_list_head;
e305ac6c 122 struct geneve_dev *geneve;
2d07dc79
JL		 123 __u32 hash;
124
2d07dc79 125 /* Find the device for this VNI */
371bd106 126 hash = geneve_net_vni_hash(vni);
66d47003 127 vni_list_head = &gs->vni_list[hash];
e305ac6c 128 hlist_for_each_entry_rcu(geneve, vni_list_head, hlist) {
371bd106 129 if (!memcmp(vni, geneve->vni, sizeof(geneve->vni)) &&
8ed66f0e
JL		 130 addr == geneve->remote.sin.sin_addr.s_addr)
131 return geneve;
132 }
133 return NULL;
134}
135
136#if IS_ENABLED(CONFIG_IPV6)
137static struct geneve_dev *geneve6_lookup(struct geneve_sock *gs,
138 struct in6_addr addr6, u8 vni[])
139{
140 struct hlist_head *vni_list_head;
141 struct geneve_dev *geneve;
142 __u32 hash;
143
144 /* Find the device for this VNI */
145 hash = geneve_net_vni_hash(vni);
146 vni_list_head = &gs->vni_list[hash];
147 hlist_for_each_entry_rcu(geneve, vni_list_head, hlist) {
148 if (!memcmp(vni, geneve->vni, sizeof(geneve->vni)) &&
149 ipv6_addr_equal(&addr6, &geneve->remote.sin6.sin6_addr))
e305ac6c 150 return geneve;
2d07dc79 151 }
e305ac6c
PS		 152 return NULL;
153}
8ed66f0e 154#endif
e305ac6c 155
371bd106
PS		 156static inline struct genevehdr *geneve_hdr(const struct sk_buff *skb)
157{
158 return (struct genevehdr *)(udp_hdr(skb) + 1);
159}
160
9fc47545
JB		 161static struct geneve_dev *geneve_lookup_skb(struct geneve_sock *gs,
162 struct sk_buff *skb)
e305ac6c 163{
9fc47545 164 u8 *vni;
371bd106 165 __be32 addr;
8ed66f0e 166 static u8 zero_vni[3];
8ed66f0e 167#if IS_ENABLED(CONFIG_IPV6)
8ed66f0e
JL		 168 static struct in6_addr zero_addr6;
169#endif
e305ac6c 170
1e9f12ec 171 if (geneve_get_sk_family(gs) == AF_INET) {
9fc47545
JB		 172 struct iphdr *iph;
173
8ed66f0e 174 iph = ip_hdr(skb); /* outer IP header... */
371bd106 175
8ed66f0e
JL		 176 if (gs->collect_md) {
177 vni = zero_vni;
178 addr = 0;
179 } else {
9fc47545 180 vni = geneve_hdr(skb)->vni;
8ed66f0e
JL		 181 addr = iph->saddr;
182 }
183
9fc47545 184 return geneve_lookup(gs, addr, vni);
8ed66f0e 185#if IS_ENABLED(CONFIG_IPV6)
1e9f12ec 186 } else if (geneve_get_sk_family(gs) == AF_INET6) {
9fc47545
JB		 187 struct ipv6hdr *ip6h;
188 struct in6_addr addr6;
189
8ed66f0e 190 ip6h = ipv6_hdr(skb); /* outer IPv6 header... */
371bd106 191
8ed66f0e
JL		 192 if (gs->collect_md) {
193 vni = zero_vni;
194 addr6 = zero_addr6;
195 } else {
9fc47545 196 vni = geneve_hdr(skb)->vni;
8ed66f0e
JL		 197 addr6 = ip6h->saddr;
198 }
199
9fc47545 200 return geneve6_lookup(gs, addr6, vni);
8ed66f0e
JL		 201#endif
202 }
9fc47545
JB		 203 return NULL;
204}
205
206/* geneve receive/decap routine */
207static void geneve_rx(struct geneve_dev *geneve, struct geneve_sock *gs,
208 struct sk_buff *skb)
209{
210 struct genevehdr *gnvh = geneve_hdr(skb);
211 struct metadata_dst *tun_dst = NULL;
212 struct pcpu_sw_netstats *stats;
213 int err = 0;
214 void *oiph;
2d07dc79 215
371bd106 216 if (ip_tunnel_collect_metadata() || gs->collect_md) {
e305ac6c 217 __be16 flags;
e305ac6c
PS		 218
219 flags = TUNNEL_KEY | TUNNEL_GENEVE_OPT |
220 (gnvh->oam ? TUNNEL_OAM : 0) |
221 (gnvh->critical ? TUNNEL_CRIT_OPT : 0);
222
1e9f12ec 223 tun_dst = udp_tun_rx_dst(skb, geneve_get_sk_family(gs), flags,
e305ac6c
PS		 224 vni_to_tunnel_id(gnvh->vni),
225 gnvh->opt_len * 4);
226 if (!tun_dst)
227 goto drop;
e305ac6c 228 /* Update tunnel dst according to Geneve options. */
4c222798
PS		 229 ip_tunnel_info_opts_set(&tun_dst->u.tun_info,
230 gnvh->options, gnvh->opt_len * 4);
e305ac6c
PS		 231 } else {
232 /* Drop packets w/ critical options,
233 * since we don't support any...
234 */
235 if (gnvh->critical)
236 goto drop;
237 }
2d07dc79
JL		 238
239 skb_reset_mac_header(skb);
2d07dc79
JL		 240 skb->protocol = eth_type_trans(skb, geneve->dev);
241 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
242
e305ac6c
PS		 243 if (tun_dst)
244 skb_dst_set(skb, &tun_dst->dst);
245
2d07dc79
JL		 246 /* Ignore packet loops (and multicast echo) */
247 if (ether_addr_equal(eth_hdr(skb)->h_source, geneve->dev->dev_addr))
248 goto drop;
249
9fc47545 250 oiph = skb_network_header(skb);
2d07dc79
JL		 251 skb_reset_network_header(skb);
252
9fc47545
JB		 253 if (geneve_get_sk_family(gs) == AF_INET)
254 err = IP_ECN_decapsulate(oiph, skb);
8ed66f0e 255#if IS_ENABLED(CONFIG_IPV6)
9fc47545
JB		 256 else
257 err = IP6_ECN_decapsulate(oiph, skb);
8ed66f0e 258#endif
2d07dc79
JL		 259
260 if (unlikely(err)) {
8ed66f0e 261 if (log_ecn_error) {
9fc47545 262 if (geneve_get_sk_family(gs) == AF_INET)
8ed66f0e
JL		 263 net_info_ratelimited("non-ECT from %pI4 "
264 "with TOS=%#x\n",
9fc47545
JB		 265 &((struct iphdr *)oiph)->saddr,
266 ((struct iphdr *)oiph)->tos);
8ed66f0e 267#if IS_ENABLED(CONFIG_IPV6)
9fc47545 268 else
8ed66f0e 269 net_info_ratelimited("non-ECT from %pI6\n",
9fc47545 270 &((struct ipv6hdr *)oiph)->saddr);
8ed66f0e
JL		 271#endif
272 }
2d07dc79
JL		 273 if (err > 1) {
274 ++geneve->dev->stats.rx_frame_errors;
275 ++geneve->dev->stats.rx_errors;
276 goto drop;
277 }
278 }
279
280 stats = this_cpu_ptr(geneve->dev->tstats);
281 u64_stats_update_begin(&stats->syncp);
282 stats->rx_packets++;
283 stats->rx_bytes += skb->len;
284 u64_stats_update_end(&stats->syncp);
285
8e816df8 286 gro_cells_receive(&geneve->gro_cells, skb);
2d07dc79
JL		 287 return;
288drop:
289 /* Consume bad packet */
290 kfree_skb(skb);
291}
292
293/* Setup stats when device is created */
294static int geneve_init(struct net_device *dev)
295{
8e816df8
JG		 296 struct geneve_dev *geneve = netdev_priv(dev);
297 int err;
298
2d07dc79
JL		 299 dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
300 if (!dev->tstats)
301 return -ENOMEM;
8e816df8
JG		 302
303 err = gro_cells_init(&geneve->gro_cells, dev);
304 if (err) {
305 free_percpu(dev->tstats);
306 return err;
307 }
308
468dfffc
PA		 309 err = dst_cache_init(&geneve->dst_cache, GFP_KERNEL);
310 if (err) {
311 free_percpu(dev->tstats);
312 gro_cells_destroy(&geneve->gro_cells);
313 return err;
314 }
315
2d07dc79
JL		 316 return 0;
317}
318
319static void geneve_uninit(struct net_device *dev)
320{
8e816df8
JG		 321 struct geneve_dev *geneve = netdev_priv(dev);
322
468dfffc 323 dst_cache_destroy(&geneve->dst_cache);
8e816df8 324 gro_cells_destroy(&geneve->gro_cells);
2d07dc79
JL		 325 free_percpu(dev->tstats);
326}
327
371bd106
PS		 328/* Callback from net/ipv4/udp.c to receive packets */
329static int geneve_udp_encap_recv(struct sock *sk, struct sk_buff *skb)
330{
331 struct genevehdr *geneveh;
9fc47545 332 struct geneve_dev *geneve;
371bd106
PS		 333 struct geneve_sock *gs;
334 int opts_len;
335
336 /* Need Geneve and inner Ethernet header to be present */
337 if (unlikely(!pskb_may_pull(skb, GENEVE_BASE_HLEN)))
338 goto error;
339
340 /* Return packets with reserved bits set */
341 geneveh = geneve_hdr(skb);
342 if (unlikely(geneveh->ver != GENEVE_VER))
343 goto error;
344
345 if (unlikely(geneveh->proto_type != htons(ETH_P_TEB)))
346 goto error;
347
9fc47545
JB		 348 gs = rcu_dereference_sk_user_data(sk);
349 if (!gs)
350 goto drop;
351
352 geneve = geneve_lookup_skb(gs, skb);
353 if (!geneve)
354 goto drop;
355
371bd106
PS		 356 opts_len = geneveh->opt_len * 4;
357 if (iptunnel_pull_header(skb, GENEVE_BASE_HLEN + opts_len,
7f290c94
JB		 358 htons(ETH_P_TEB),
359 !net_eq(geneve->net, dev_net(geneve->dev))))
371bd106
PS		 360 goto drop;
361
9fc47545 362 geneve_rx(geneve, gs, skb);
371bd106
PS		 363 return 0;
364
365drop:
366 /* Consume bad packet */
367 kfree_skb(skb);
368 return 0;
369
370error:
371 /* Let the UDP layer deal with the skb */
372 return 1;
373}
374
375static struct socket *geneve_create_sock(struct net *net, bool ipv6,
abe492b4 376 __be16 port, u32 flags)
371bd106
PS		 377{
378 struct socket *sock;
379 struct udp_port_cfg udp_conf;
380 int err;
381
382 memset(&udp_conf, 0, sizeof(udp_conf));
383
384 if (ipv6) {
385 udp_conf.family = AF_INET6;
8ed66f0e 386 udp_conf.ipv6_v6only = 1;
abe492b4
TH		 387 udp_conf.use_udp6_rx_checksums =
388 !(flags & GENEVE_F_UDP_ZERO_CSUM6_RX);
371bd106
PS		 389 } else {
390 udp_conf.family = AF_INET;
391 udp_conf.local_ip.s_addr = htonl(INADDR_ANY);
392 }
393
394 udp_conf.local_udp_port = port;
395
396 /* Open UDP socket */
397 err = udp_sock_create(net, &udp_conf, &sock);
398 if (err < 0)
399 return ERR_PTR(err);
400
401 return sock;
402}
403
404static void geneve_notify_add_rx_port(struct geneve_sock *gs)
405{
a8170d2b 406 struct net_device *dev;
371bd106 407 struct sock *sk = gs->sock->sk;
a8170d2b 408 struct net *net = sock_net(sk);
1e9f12ec 409 sa_family_t sa_family = geneve_get_sk_family(gs);
a8170d2b 410 __be16 port = inet_sk(sk)->inet_sport;
a8170d2b
SA		 411
412 rcu_read_lock();
413 for_each_netdev_rcu(net, dev) {
414 if (dev->netdev_ops->ndo_add_geneve_port)
415 dev->netdev_ops->ndo_add_geneve_port(dev, sa_family,
416 port);
417 }
418 rcu_read_unlock();
371bd106
PS		 419}
420
421static int geneve_hlen(struct genevehdr *gh)
422{
423 return sizeof(*gh) + gh->opt_len * 4;
424}
425
4a0090a9
TH		 426static struct sk_buff **geneve_gro_receive(struct sock *sk,
427 struct sk_buff **head,
428 struct sk_buff *skb)
371bd106
PS		 429{
430 struct sk_buff *p, **pp = NULL;
431 struct genevehdr *gh, *gh2;
432 unsigned int hlen, gh_len, off_gnv;
433 const struct packet_offload *ptype;
434 __be16 type;
435 int flush = 1;
436
437 off_gnv = skb_gro_offset(skb);
438 hlen = off_gnv + sizeof(*gh);
439 gh = skb_gro_header_fast(skb, off_gnv);
440 if (skb_gro_header_hard(skb, hlen)) {
441 gh = skb_gro_header_slow(skb, hlen, off_gnv);
442 if (unlikely(!gh))
443 goto out;
444 }
445
446 if (gh->ver != GENEVE_VER || gh->oam)
447 goto out;
448 gh_len = geneve_hlen(gh);
449
450 hlen = off_gnv + gh_len;
451 if (skb_gro_header_hard(skb, hlen)) {
452 gh = skb_gro_header_slow(skb, hlen, off_gnv);
453 if (unlikely(!gh))
454 goto out;
455 }
456
371bd106
PS		 457 for (p = *head; p; p = p->next) {
458 if (!NAPI_GRO_CB(p)->same_flow)
459 continue;
460
461 gh2 = (struct genevehdr *)(p->data + off_gnv);
462 if (gh->opt_len != gh2->opt_len ||
463 memcmp(gh, gh2, gh_len)) {
464 NAPI_GRO_CB(p)->same_flow = 0;
465 continue;
466 }
467 }
468
469 type = gh->proto_type;
470
471 rcu_read_lock();
472 ptype = gro_find_receive_by_type(type);
c194cf93 473 if (!ptype)
371bd106 474 goto out_unlock;
371bd106
PS		 475
476 skb_gro_pull(skb, gh_len);
477 skb_gro_postpull_rcsum(skb, gh, gh_len);
478 pp = ptype->callbacks.gro_receive(head, skb);
c194cf93 479 flush = 0;
371bd106
PS		 480
481out_unlock:
482 rcu_read_unlock();
483out:
484 NAPI_GRO_CB(skb)->flush |= flush;
485
486 return pp;
487}
488
4a0090a9
TH		 489static int geneve_gro_complete(struct sock *sk, struct sk_buff *skb,
490 int nhoff)
371bd106
PS		 491{
492 struct genevehdr *gh;
493 struct packet_offload *ptype;
494 __be16 type;
495 int gh_len;
496 int err = -ENOSYS;
497
498 udp_tunnel_gro_complete(skb, nhoff);
499
500 gh = (struct genevehdr *)(skb->data + nhoff);
501 gh_len = geneve_hlen(gh);
502 type = gh->proto_type;
503
504 rcu_read_lock();
505 ptype = gro_find_complete_by_type(type);
506 if (ptype)
507 err = ptype->callbacks.gro_complete(skb, nhoff + gh_len);
508
509 rcu_read_unlock();
510 return err;
511}
512
513/* Create new listen socket if needed */
514static struct geneve_sock *geneve_socket_create(struct net *net, __be16 port,
abe492b4 515 bool ipv6, u32 flags)
371bd106
PS		 516{
517 struct geneve_net *gn = net_generic(net, geneve_net_id);
518 struct geneve_sock *gs;
519 struct socket *sock;
520 struct udp_tunnel_sock_cfg tunnel_cfg;
66d47003 521 int h;
371bd106
PS		 522
523 gs = kzalloc(sizeof(*gs), GFP_KERNEL);
524 if (!gs)
525 return ERR_PTR(-ENOMEM);
526
abe492b4 527 sock = geneve_create_sock(net, ipv6, port, flags);
371bd106
PS		 528 if (IS_ERR(sock)) {
529 kfree(gs);
530 return ERR_CAST(sock);
531 }
532
533 gs->sock = sock;
534 gs->refcnt = 1;
66d47003
PS		 535 for (h = 0; h < VNI_HASH_SIZE; ++h)
536 INIT_HLIST_HEAD(&gs->vni_list[h]);
371bd106
PS		 537
538 /* Initialize the geneve udp offloads structure */
371bd106
PS		 539 geneve_notify_add_rx_port(gs);
540
541 /* Mark socket as an encapsulation socket */
4a0090a9 542 memset(&tunnel_cfg, 0, sizeof(tunnel_cfg));
371bd106
PS		 543 tunnel_cfg.sk_user_data = gs;
544 tunnel_cfg.encap_type = 1;
4a0090a9
TH		 545 tunnel_cfg.gro_receive = geneve_gro_receive;
546 tunnel_cfg.gro_complete = geneve_gro_complete;
371bd106
PS		 547 tunnel_cfg.encap_rcv = geneve_udp_encap_recv;
548 tunnel_cfg.encap_destroy = NULL;
549 setup_udp_tunnel_sock(net, sock, &tunnel_cfg);
371bd106
PS		 550 list_add(&gs->list, &gn->sock_list);
551 return gs;
552}
553
554static void geneve_notify_del_rx_port(struct geneve_sock *gs)
555{
a8170d2b 556 struct net_device *dev;
371bd106 557 struct sock *sk = gs->sock->sk;
a8170d2b 558 struct net *net = sock_net(sk);
1e9f12ec 559 sa_family_t sa_family = geneve_get_sk_family(gs);
a8170d2b
SA		 560 __be16 port = inet_sk(sk)->inet_sport;
561
562 rcu_read_lock();
563 for_each_netdev_rcu(net, dev) {
564 if (dev->netdev_ops->ndo_del_geneve_port)
565 dev->netdev_ops->ndo_del_geneve_port(dev, sa_family,
566 port);
567 }
568
569 rcu_read_unlock();
371bd106
PS		 570}
571
8ed66f0e 572static void __geneve_sock_release(struct geneve_sock *gs)
371bd106 573{
8ed66f0e 574 if (!gs || --gs->refcnt)
371bd106
PS		 575 return;
576
577 list_del(&gs->list);
578 geneve_notify_del_rx_port(gs);
579 udp_tunnel_sock_release(gs->sock);
580 kfree_rcu(gs, rcu);
581}
582
8ed66f0e
JL		 583static void geneve_sock_release(struct geneve_dev *geneve)
584{
585 __geneve_sock_release(geneve->sock4);
586#if IS_ENABLED(CONFIG_IPV6)
587 __geneve_sock_release(geneve->sock6);
588#endif
589}
590
371bd106 591static struct geneve_sock *geneve_find_sock(struct geneve_net *gn,
8ed66f0e 592 sa_family_t family,
371bd106
PS		 593 __be16 dst_port)
594{
595 struct geneve_sock *gs;
596
597 list_for_each_entry(gs, &gn->sock_list, list) {
598 if (inet_sk(gs->sock->sk)->inet_sport == dst_port &&
1e9f12ec 599 geneve_get_sk_family(gs) == family) {
371bd106
PS		 600 return gs;
601 }
602 }
603 return NULL;
604}
605
8ed66f0e 606static int geneve_sock_add(struct geneve_dev *geneve, bool ipv6)
2d07dc79 607{
2d07dc79 608 struct net *net = geneve->net;
371bd106 609 struct geneve_net *gn = net_generic(net, geneve_net_id);
2d07dc79 610 struct geneve_sock *gs;
66d47003 611 __u32 hash;
2d07dc79 612
8ed66f0e 613 gs = geneve_find_sock(gn, ipv6 ? AF_INET6 : AF_INET, geneve->dst_port);
371bd106
PS		 614 if (gs) {
615 gs->refcnt++;
616 goto out;
617 }
618
abe492b4 619 gs = geneve_socket_create(net, geneve->dst_port, ipv6, geneve->flags);
2d07dc79
JL		 620 if (IS_ERR(gs))
621 return PTR_ERR(gs);
622
371bd106
PS		 623out:
624 gs->collect_md = geneve->collect_md;
abe492b4 625 gs->flags = geneve->flags;
8ed66f0e
JL		 626#if IS_ENABLED(CONFIG_IPV6)
627 if (ipv6)
628 geneve->sock6 = gs;
629 else
630#endif
631 geneve->sock4 = gs;
66d47003
PS		 632
633 hash = geneve_net_vni_hash(geneve->vni);
634 hlist_add_head_rcu(&geneve->hlist, &gs->vni_list[hash]);
2d07dc79
JL		 635 return 0;
636}
637
8ed66f0e
JL		 638static int geneve_open(struct net_device *dev)
639{
640 struct geneve_dev *geneve = netdev_priv(dev);
641 bool ipv6 = geneve->remote.sa.sa_family == AF_INET6;
642 bool metadata = geneve->collect_md;
643 int ret = 0;
644
645 geneve->sock4 = NULL;
646#if IS_ENABLED(CONFIG_IPV6)
647 geneve->sock6 = NULL;
648 if (ipv6 || metadata)
649 ret = geneve_sock_add(geneve, true);
650#endif
651 if (!ret && (!ipv6 || metadata))
652 ret = geneve_sock_add(geneve, false);
653 if (ret < 0)
654 geneve_sock_release(geneve);
655
656 return ret;
657}
658
2d07dc79
JL		 659static int geneve_stop(struct net_device *dev)
660{
661 struct geneve_dev *geneve = netdev_priv(dev);
2d07dc79 662
66d47003
PS		 663 if (!hlist_unhashed(&geneve->hlist))
664 hlist_del_rcu(&geneve->hlist);
8ed66f0e 665 geneve_sock_release(geneve);
371bd106
PS		 666 return 0;
667}
668
8ed66f0e
JL		 669static void geneve_build_header(struct genevehdr *geneveh,
670 __be16 tun_flags, u8 vni[3],
671 u8 options_len, u8 *options)
672{
673 geneveh->ver = GENEVE_VER;
674 geneveh->opt_len = options_len / 4;
675 geneveh->oam = !!(tun_flags & TUNNEL_OAM);
676 geneveh->critical = !!(tun_flags & TUNNEL_CRIT_OPT);
677 geneveh->rsvd1 = 0;
678 memcpy(geneveh->vni, vni, 3);
679 geneveh->proto_type = htons(ETH_P_TEB);
680 geneveh->rsvd2 = 0;
681
682 memcpy(geneveh->options, options, options_len);
683}
684
371bd106
PS		 685static int geneve_build_skb(struct rtable *rt, struct sk_buff *skb,
686 __be16 tun_flags, u8 vni[3], u8 opt_len, u8 *opt,
abe492b4 687 u32 flags, bool xnet)
371bd106
PS		 688{
689 struct genevehdr *gnvh;
690 int min_headroom;
691 int err;
14f1f724 692 bool udp_sum = !(flags & GENEVE_F_UDP_ZERO_CSUM_TX);
2d07dc79 693
8ed66f0e
JL		 694 skb_scrub_packet(skb, xnet);
695
371bd106
PS		 696 min_headroom = LL_RESERVED_SPACE(rt->dst.dev) + rt->dst.header_len
697 + GENEVE_BASE_HLEN + opt_len + sizeof(struct iphdr);
698 err = skb_cow_head(skb, min_headroom);
aed069df 699 if (unlikely(err))
371bd106 700 goto free_rt;
371bd106 701
aed069df
AD		 702 err = udp_tunnel_handle_offloads(skb, udp_sum);
703 if (err)
371bd106 704 goto free_rt;
371bd106
PS		 705
706 gnvh = (struct genevehdr *)__skb_push(skb, sizeof(*gnvh) + opt_len);
8ed66f0e 707 geneve_build_header(gnvh, tun_flags, vni, opt_len, opt);
371bd106
PS		 708
709 skb_set_inner_protocol(skb, htons(ETH_P_TEB));
2d07dc79 710 return 0;
371bd106
PS		 711
712free_rt:
713 ip_rt_put(rt);
714 return err;
2d07dc79
JL		 715}
716
8ed66f0e
JL		 717#if IS_ENABLED(CONFIG_IPV6)
718static int geneve6_build_skb(struct dst_entry *dst, struct sk_buff *skb,
719 __be16 tun_flags, u8 vni[3], u8 opt_len, u8 *opt,
abe492b4 720 u32 flags, bool xnet)
8ed66f0e
JL		 721{
722 struct genevehdr *gnvh;
723 int min_headroom;
724 int err;
abe492b4 725 bool udp_sum = !(flags & GENEVE_F_UDP_ZERO_CSUM6_TX);
8ed66f0e
JL		 726
727 skb_scrub_packet(skb, xnet);
728
729 min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len
730 + GENEVE_BASE_HLEN + opt_len + sizeof(struct ipv6hdr);
731 err = skb_cow_head(skb, min_headroom);
aed069df 732 if (unlikely(err))
8ed66f0e 733 goto free_dst;
8ed66f0e 734
aed069df 735 err = udp_tunnel_handle_offloads(skb, udp_sum);
1ba64fac 736 if (err)
8ed66f0e 737 goto free_dst;
8ed66f0e
JL		 738
739 gnvh = (struct genevehdr *)__skb_push(skb, sizeof(*gnvh) + opt_len);
740 geneve_build_header(gnvh, tun_flags, vni, opt_len, opt);
741
742 skb_set_inner_protocol(skb, htons(ETH_P_TEB));
743 return 0;
744
745free_dst:
746 dst_release(dst);
747 return err;
748}
749#endif
750
751static struct rtable *geneve_get_v4_rt(struct sk_buff *skb,
752 struct net_device *dev,
753 struct flowi4 *fl4,
754 struct ip_tunnel_info *info)
e305ac6c 755{
db3c6139 756 bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
e305ac6c 757 struct geneve_dev *geneve = netdev_priv(dev);
468dfffc 758 struct dst_cache *dst_cache;
e305ac6c
PS		 759 struct rtable *rt = NULL;
760 __u8 tos;
761
762 memset(fl4, 0, sizeof(*fl4));
763 fl4->flowi4_mark = skb->mark;
764 fl4->flowi4_proto = IPPROTO_UDP;
765
766 if (info) {
767 fl4->daddr = info->key.u.ipv4.dst;
768 fl4->saddr = info->key.u.ipv4.src;
769 fl4->flowi4_tos = RT_TOS(info->key.tos);
468dfffc 770 dst_cache = &info->dst_cache;
e305ac6c
PS		 771 } else {
772 tos = geneve->tos;
773 if (tos == 1) {
774 const struct iphdr *iip = ip_hdr(skb);
775
776 tos = ip_tunnel_get_dsfield(iip, skb);
468dfffc 777 use_cache = false;
e305ac6c
PS		 778 }
779
780 fl4->flowi4_tos = RT_TOS(tos);
8ed66f0e 781 fl4->daddr = geneve->remote.sin.sin_addr.s_addr;
468dfffc
PA		 782 dst_cache = &geneve->dst_cache;
783 }
784
468dfffc
PA		 785 if (use_cache) {
786 rt = dst_cache_get_ip4(dst_cache, &fl4->saddr);
787 if (rt)
788 return rt;
e305ac6c
PS		 789 }
790
791 rt = ip_route_output_key(geneve->net, fl4);
792 if (IS_ERR(rt)) {
793 netdev_dbg(dev, "no route to %pI4\n", &fl4->daddr);
fc4099f1 794 return ERR_PTR(-ENETUNREACH);
e305ac6c
PS		 795 }
796 if (rt->dst.dev == dev) { /* is this necessary? */
797 netdev_dbg(dev, "circular route to %pI4\n", &fl4->daddr);
e305ac6c 798 ip_rt_put(rt);
fc4099f1 799 return ERR_PTR(-ELOOP);
e305ac6c 800 }
468dfffc
PA		 801 if (use_cache)
802 dst_cache_set_ip4(dst_cache, &rt->dst, fl4->saddr);
e305ac6c
PS		 803 return rt;
804}
805
8ed66f0e
JL		 806#if IS_ENABLED(CONFIG_IPV6)
807static struct dst_entry *geneve_get_v6_dst(struct sk_buff *skb,
808 struct net_device *dev,
809 struct flowi6 *fl6,
810 struct ip_tunnel_info *info)
811{
db3c6139 812 bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
8ed66f0e
JL		 813 struct geneve_dev *geneve = netdev_priv(dev);
814 struct geneve_sock *gs6 = geneve->sock6;
815 struct dst_entry *dst = NULL;
468dfffc 816 struct dst_cache *dst_cache;
3a56f86f 817 __u8 prio;
8ed66f0e
JL		 818
819 memset(fl6, 0, sizeof(*fl6));
820 fl6->flowi6_mark = skb->mark;
821 fl6->flowi6_proto = IPPROTO_UDP;
822
823 if (info) {
824 fl6->daddr = info->key.u.ipv6.dst;
825 fl6->saddr = info->key.u.ipv6.src;
95caf6f7
DB		 826 fl6->flowlabel = ip6_make_flowinfo(RT_TOS(info->key.tos),
827 info->key.label);
468dfffc 828 dst_cache = &info->dst_cache;
8ed66f0e 829 } else {
3a56f86f
JL		 830 prio = geneve->tos;
831 if (prio == 1) {
832 const struct iphdr *iip = ip_hdr(skb);
833
834 prio = ip_tunnel_get_dsfield(iip, skb);
468dfffc 835 use_cache = false;
3a56f86f
JL		 836 }
837
95caf6f7
DB		 838 fl6->flowlabel = ip6_make_flowinfo(RT_TOS(prio),
839 geneve->label);
8ed66f0e 840 fl6->daddr = geneve->remote.sin6.sin6_addr;
468dfffc
PA		 841 dst_cache = &geneve->dst_cache;
842 }
843
468dfffc
PA		 844 if (use_cache) {
845 dst = dst_cache_get_ip6(dst_cache, &fl6->saddr);
846 if (dst)
847 return dst;
8ed66f0e
JL		 848 }
849
850 if (ipv6_stub->ipv6_dst_lookup(geneve->net, gs6->sock->sk, &dst, fl6)) {
851 netdev_dbg(dev, "no route to %pI6\n", &fl6->daddr);
852 return ERR_PTR(-ENETUNREACH);
853 }
854 if (dst->dev == dev) { /* is this necessary? */
855 netdev_dbg(dev, "circular route to %pI6\n", &fl6->daddr);
856 dst_release(dst);
857 return ERR_PTR(-ELOOP);
858 }
859
468dfffc
PA		 860 if (use_cache)
861 dst_cache_set_ip6(dst_cache, dst, &fl6->saddr);
8ed66f0e
JL		 862 return dst;
863}
864#endif
865
e305ac6c
PS		 866/* Convert 64 bit tunnel ID to 24 bit VNI. */
867static void tunnel_id_to_vni(__be64 tun_id, __u8 *vni)
868{
869#ifdef __BIG_ENDIAN
870 vni[0] = (__force __u8)(tun_id >> 16);
871 vni[1] = (__force __u8)(tun_id >> 8);
872 vni[2] = (__force __u8)tun_id;
873#else
874 vni[0] = (__force __u8)((__force u64)tun_id >> 40);
875 vni[1] = (__force __u8)((__force u64)tun_id >> 48);
876 vni[2] = (__force __u8)((__force u64)tun_id >> 56);
877#endif
878}
879
8ed66f0e
JL		 880static netdev_tx_t geneve_xmit_skb(struct sk_buff *skb, struct net_device *dev,
881 struct ip_tunnel_info *info)
2d07dc79
JL		 882{
883 struct geneve_dev *geneve = netdev_priv(dev);
8ed66f0e 884 struct geneve_sock *gs4 = geneve->sock4;
2d07dc79 885 struct rtable *rt = NULL;
08399efc 886 const struct iphdr *iip; /* interior IP header */
fc4099f1 887 int err = -EINVAL;
2d07dc79 888 struct flowi4 fl4;
8760ce58 889 __u8 tos, ttl;
e305ac6c 890 __be16 sport;
371bd106 891 __be16 df;
8ed66f0e 892 bool xnet = !net_eq(geneve->net, dev_net(geneve->dev));
abe492b4 893 u32 flags = geneve->flags;
2d07dc79 894
e305ac6c 895 if (geneve->collect_md) {
8ed66f0e 896 if (unlikely(!info || !(info->mode & IP_TUNNEL_INFO_TX))) {
e305ac6c
PS		 897 netdev_dbg(dev, "no tunnel metadata\n");
898 goto tx_error;
899 }
7f9562a1
JB		 900 if (info && ip_tunnel_info_af(info) != AF_INET)
901 goto tx_error;
e305ac6c 902 }
980c394c 903
8ed66f0e 904 rt = geneve_get_v4_rt(skb, dev, &fl4, info);
2d07dc79 905 if (IS_ERR(rt)) {
fc4099f1 906 err = PTR_ERR(rt);
2d07dc79
JL		 907 goto tx_error;
908 }
371bd106
PS		 909
910 sport = udp_flow_src_port(geneve->net, skb, 1, USHRT_MAX, true);
e305ac6c 911 skb_reset_mac_header(skb);
e305ac6c 912
08399efc
JL		 913 iip = ip_hdr(skb);
914
e305ac6c
PS		 915 if (info) {
916 const struct ip_tunnel_key *key = &info->key;
e305ac6c
PS		 917 u8 *opts = NULL;
918 u8 vni[3];
e305ac6c
PS		 919
920 tunnel_id_to_vni(key->tun_id, vni);
14ca0751 921 if (info->options_len)
4c222798 922 opts = ip_tunnel_info_opts(info);
e305ac6c 923
abe492b4 924 if (key->tun_flags & TUNNEL_CSUM)
14f1f724 925 flags &= ~GENEVE_F_UDP_ZERO_CSUM_TX;
abe492b4 926 else
14f1f724 927 flags |= GENEVE_F_UDP_ZERO_CSUM_TX;
abe492b4 928
371bd106 929 err = geneve_build_skb(rt, skb, key->tun_flags, vni,
abe492b4 930 info->options_len, opts, flags, xnet);
371bd106 931 if (unlikely(err))
aed069df 932 goto tx_error;
371bd106 933
08399efc 934 tos = ip_tunnel_ecn_encap(key->tos, iip, skb);
371bd106
PS		 935 ttl = key->ttl;
936 df = key->tun_flags & TUNNEL_DONT_FRAGMENT ? htons(IP_DF) : 0;
e305ac6c 937 } else {
371bd106 938 err = geneve_build_skb(rt, skb, 0, geneve->vni,
abe492b4 939 0, NULL, flags, xnet);
371bd106 940 if (unlikely(err))
aed069df 941 goto tx_error;
371bd106 942
e305ac6c 943 tos = ip_tunnel_ecn_encap(fl4.flowi4_tos, iip, skb);
e305ac6c
PS		 944 ttl = geneve->ttl;
945 if (!ttl && IN_MULTICAST(ntohl(fl4.daddr)))
946 ttl = 1;
e305ac6c 947 ttl = ttl ? : ip4_dst_hoplimit(&rt->dst);
371bd106 948 df = 0;
2d07dc79 949 }
039f5062
PS		 950 udp_tunnel_xmit_skb(rt, gs4->sock->sk, skb, fl4.saddr, fl4.daddr,
951 tos, ttl, df, sport, geneve->dst_port,
952 !net_eq(geneve->net, dev_net(geneve->dev)),
14f1f724 953 !!(flags & GENEVE_F_UDP_ZERO_CSUM_TX));
2d07dc79 954
2d07dc79
JL		 955 return NETDEV_TX_OK;
956
2d07dc79 957tx_error:
2d07dc79 958 dev_kfree_skb(skb);
aed069df 959
fc4099f1
PS		 960 if (err == -ELOOP)
961 dev->stats.collisions++;
962 else if (err == -ENETUNREACH)
963 dev->stats.tx_carrier_errors++;
964 else
965 dev->stats.tx_errors++;
2d07dc79
JL		 966 return NETDEV_TX_OK;
967}
968
8ed66f0e
JL		 969#if IS_ENABLED(CONFIG_IPV6)
970static netdev_tx_t geneve6_xmit_skb(struct sk_buff *skb, struct net_device *dev,
971 struct ip_tunnel_info *info)
972{
973 struct geneve_dev *geneve = netdev_priv(dev);
974 struct geneve_sock *gs6 = geneve->sock6;
975 struct dst_entry *dst = NULL;
3a56f86f 976 const struct iphdr *iip; /* interior IP header */
8ed66f0e
JL		 977 int err = -EINVAL;
978 struct flowi6 fl6;
3a56f86f 979 __u8 prio, ttl;
8ed66f0e 980 __be16 sport;
8eb3b995 981 __be32 label;
8ed66f0e 982 bool xnet = !net_eq(geneve->net, dev_net(geneve->dev));
abe492b4 983 u32 flags = geneve->flags;
8ed66f0e
JL		 984
985 if (geneve->collect_md) {
986 if (unlikely(!info || !(info->mode & IP_TUNNEL_INFO_TX))) {
987 netdev_dbg(dev, "no tunnel metadata\n");
988 goto tx_error;
989 }
990 }
991
992 dst = geneve_get_v6_dst(skb, dev, &fl6, info);
993 if (IS_ERR(dst)) {
994 err = PTR_ERR(dst);
995 goto tx_error;
996 }
997
998 sport = udp_flow_src_port(geneve->net, skb, 1, USHRT_MAX, true);
999 skb_reset_mac_header(skb);
1000
3a56f86f
JL		 1001 iip = ip_hdr(skb);
1002
8ed66f0e
JL		 1003 if (info) {
1004 const struct ip_tunnel_key *key = &info->key;
1005 u8 *opts = NULL;
1006 u8 vni[3];
1007
1008 tunnel_id_to_vni(key->tun_id, vni);
14ca0751 1009 if (info->options_len)
8ed66f0e
JL		 1010 opts = ip_tunnel_info_opts(info);
1011
abe492b4 1012 if (key->tun_flags & TUNNEL_CSUM)
35e2d115 1013 flags &= ~GENEVE_F_UDP_ZERO_CSUM6_TX;
abe492b4 1014 else
35e2d115 1015 flags |= GENEVE_F_UDP_ZERO_CSUM6_TX;
abe492b4 1016
8ed66f0e
JL		 1017 err = geneve6_build_skb(dst, skb, key->tun_flags, vni,
1018 info->options_len, opts,
abe492b4 1019 flags, xnet);
8ed66f0e 1020 if (unlikely(err))
aed069df 1021 goto tx_error;
8ed66f0e 1022
3a56f86f 1023 prio = ip_tunnel_ecn_encap(key->tos, iip, skb);
8ed66f0e 1024 ttl = key->ttl;
8eb3b995 1025 label = info->key.label;
8ed66f0e 1026 } else {
8ed66f0e 1027 err = geneve6_build_skb(dst, skb, 0, geneve->vni,
abe492b4 1028 0, NULL, flags, xnet);
8ed66f0e 1029 if (unlikely(err))
aed069df 1030 goto tx_error;
8ed66f0e 1031
95caf6f7
DB		 1032 prio = ip_tunnel_ecn_encap(ip6_tclass(fl6.flowlabel),
1033 iip, skb);
8ed66f0e
JL		 1034 ttl = geneve->ttl;
1035 if (!ttl && ipv6_addr_is_multicast(&fl6.daddr))
1036 ttl = 1;
1037 ttl = ttl ? : ip6_dst_hoplimit(dst);
8eb3b995 1038 label = geneve->label;
8ed66f0e 1039 }
8eb3b995 1040
039f5062 1041 udp_tunnel6_xmit_skb(dst, gs6->sock->sk, skb, dev,
8eb3b995 1042 &fl6.saddr, &fl6.daddr, prio, ttl, label,
039f5062
PS		 1043 sport, geneve->dst_port,
1044 !!(flags & GENEVE_F_UDP_ZERO_CSUM6_TX));
8ed66f0e
JL		 1045 return NETDEV_TX_OK;
1046
1047tx_error:
1048 dev_kfree_skb(skb);
aed069df 1049
8ed66f0e
JL		 1050 if (err == -ELOOP)
1051 dev->stats.collisions++;
1052 else if (err == -ENETUNREACH)
1053 dev->stats.tx_carrier_errors++;
1054 else
1055 dev->stats.tx_errors++;
1056 return NETDEV_TX_OK;
1057}
1058#endif
1059
1060static netdev_tx_t geneve_xmit(struct sk_buff *skb, struct net_device *dev)
1061{
1062 struct geneve_dev *geneve = netdev_priv(dev);
1063 struct ip_tunnel_info *info = NULL;
1064
1065 if (geneve->collect_md)
1066 info = skb_tunnel_info(skb);
1067
1068#if IS_ENABLED(CONFIG_IPV6)
1069 if ((info && ip_tunnel_info_af(info) == AF_INET6) ||
1070 (!info && geneve->remote.sa.sa_family == AF_INET6))
1071 return geneve6_xmit_skb(skb, dev, info);
1072#endif
1073 return geneve_xmit_skb(skb, dev, info);
1074}
1075
aeee0e66 1076static int __geneve_change_mtu(struct net_device *dev, int new_mtu, bool strict)
55e5bfb5 1077{
aeee0e66
DW		 1078 /* The max_mtu calculation does not take account of GENEVE
1079 * options, to avoid excluding potentially valid
1080 * configurations.
55e5bfb5 1081 */
aeee0e66
DW		 1082 int max_mtu = IP_MAX_MTU - GENEVE_BASE_HLEN - sizeof(struct iphdr)
1083 - dev->hard_header_len;
1084
1085 if (new_mtu < 68)
55e5bfb5 1086 return -EINVAL;
aeee0e66
DW		 1087
1088 if (new_mtu > max_mtu) {
1089 if (strict)
1090 return -EINVAL;
1091
1092 new_mtu = max_mtu;
1093 }
1094
55e5bfb5
DW		 1095 dev->mtu = new_mtu;
1096 return 0;
1097}
1098
aeee0e66
DW		 1099static int geneve_change_mtu(struct net_device *dev, int new_mtu)
1100{
1101 return __geneve_change_mtu(dev, new_mtu, true);
1102}
1103
fc4099f1
PS		 1104static int geneve_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
1105{
1106 struct ip_tunnel_info *info = skb_tunnel_info(skb);
1107 struct geneve_dev *geneve = netdev_priv(dev);
1108 struct rtable *rt;
1109 struct flowi4 fl4;
b8812fa8
JL		 1110#if IS_ENABLED(CONFIG_IPV6)
1111 struct dst_entry *dst;
1112 struct flowi6 fl6;
1113#endif
fc4099f1 1114
b8812fa8
JL		 1115 if (ip_tunnel_info_af(info) == AF_INET) {
1116 rt = geneve_get_v4_rt(skb, dev, &fl4, info);
1117 if (IS_ERR(rt))
1118 return PTR_ERR(rt);
fc4099f1 1119
b8812fa8
JL		 1120 ip_rt_put(rt);
1121 info->key.u.ipv4.src = fl4.saddr;
1122#if IS_ENABLED(CONFIG_IPV6)
1123 } else if (ip_tunnel_info_af(info) == AF_INET6) {
1124 dst = geneve_get_v6_dst(skb, dev, &fl6, info);
1125 if (IS_ERR(dst))
1126 return PTR_ERR(dst);
1127
1128 dst_release(dst);
1129 info->key.u.ipv6.src = fl6.saddr;
1130#endif
1131 } else {
1132 return -EINVAL;
1133 }
fc4099f1 1134
fc4099f1
PS		 1135 info->key.tp_src = udp_flow_src_port(geneve->net, skb,
1136 1, USHRT_MAX, true);
1137 info->key.tp_dst = geneve->dst_port;
1138 return 0;
1139}
1140
2d07dc79
JL		 1141static const struct net_device_ops geneve_netdev_ops = {
1142 .ndo_init = geneve_init,
1143 .ndo_uninit = geneve_uninit,
1144 .ndo_open = geneve_open,
1145 .ndo_stop = geneve_stop,
1146 .ndo_start_xmit = geneve_xmit,
1147 .ndo_get_stats64 = ip_tunnel_get_stats64,
55e5bfb5 1148 .ndo_change_mtu = geneve_change_mtu,
2d07dc79
JL		 1149 .ndo_validate_addr = eth_validate_addr,
1150 .ndo_set_mac_address = eth_mac_addr,
fc4099f1 1151 .ndo_fill_metadata_dst = geneve_fill_metadata_dst,
2d07dc79
JL		 1152};
1153
1154static void geneve_get_drvinfo(struct net_device *dev,
1155 struct ethtool_drvinfo *drvinfo)
1156{
1157 strlcpy(drvinfo->version, GENEVE_NETDEV_VER, sizeof(drvinfo->version));
1158 strlcpy(drvinfo->driver, "geneve", sizeof(drvinfo->driver));
1159}
1160
1161static const struct ethtool_ops geneve_ethtool_ops = {
1162 .get_drvinfo = geneve_get_drvinfo,
1163 .get_link = ethtool_op_get_link,
1164};
1165
1166/* Info for udev, that this is a virtual tunnel endpoint */
1167static struct device_type geneve_type = {
1168 .name = "geneve",
1169};
1170
05ca4029
SA		 1171/* Calls the ndo_add_geneve_port of the caller in order to
1172 * supply the listening GENEVE udp ports. Callers are expected
1173 * to implement the ndo_add_geneve_port.
1174 */
1175void geneve_get_rx_port(struct net_device *dev)
1176{
1177 struct net *net = dev_net(dev);
1178 struct geneve_net *gn = net_generic(net, geneve_net_id);
1179 struct geneve_sock *gs;
1180 sa_family_t sa_family;
1181 struct sock *sk;
1182 __be16 port;
1183
1184 rcu_read_lock();
1185 list_for_each_entry_rcu(gs, &gn->sock_list, list) {
1186 sk = gs->sock->sk;
1187 sa_family = sk->sk_family;
1188 port = inet_sk(sk)->inet_sport;
1189 dev->netdev_ops->ndo_add_geneve_port(dev, sa_family, port);
1190 }
1191 rcu_read_unlock();
1192}
1193EXPORT_SYMBOL_GPL(geneve_get_rx_port);
1194
2d07dc79
JL		 1195/* Initialize the device structure. */
1196static void geneve_setup(struct net_device *dev)
1197{
1198 ether_setup(dev);
1199
1200 dev->netdev_ops = &geneve_netdev_ops;
1201 dev->ethtool_ops = &geneve_ethtool_ops;
1202 dev->destructor = free_netdev;
1203
1204 SET_NETDEV_DEVTYPE(dev, &geneve_type);
1205
2d07dc79
JL		 1206 dev->features |= NETIF_F_LLTX;
1207 dev->features |= NETIF_F_SG | NETIF_F_HW_CSUM;
1208 dev->features |= NETIF_F_RXCSUM;
1209 dev->features |= NETIF_F_GSO_SOFTWARE;
1210
2d07dc79
JL		 1211 dev->hw_features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_RXCSUM;
1212 dev->hw_features |= NETIF_F_GSO_SOFTWARE;
2d07dc79
JL		 1213
1214 netif_keep_dst(dev);
fc41cdb3 1215 dev->priv_flags &= ~IFF_TX_SKB_SHARING;
ed961ac2 1216 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
87cd3dca 1217 eth_hw_addr_random(dev);
2d07dc79
JL		 1218}
1219
1220static const struct nla_policy geneve_policy[IFLA_GENEVE_MAX + 1] = {
1221 [IFLA_GENEVE_ID] = { .type = NLA_U32 },
1222 [IFLA_GENEVE_REMOTE] = { .len = FIELD_SIZEOF(struct iphdr, daddr) },
8ed66f0e 1223 [IFLA_GENEVE_REMOTE6] = { .len = sizeof(struct in6_addr) },
8760ce58 1224 [IFLA_GENEVE_TTL] = { .type = NLA_U8 },
d8951125 1225 [IFLA_GENEVE_TOS] = { .type = NLA_U8 },
8eb3b995 1226 [IFLA_GENEVE_LABEL] = { .type = NLA_U32 },
cd7918b3 1227 [IFLA_GENEVE_PORT] = { .type = NLA_U16 },
e305ac6c 1228 [IFLA_GENEVE_COLLECT_METADATA] = { .type = NLA_FLAG },
abe492b4
TH		 1229 [IFLA_GENEVE_UDP_CSUM] = { .type = NLA_U8 },
1230 [IFLA_GENEVE_UDP_ZERO_CSUM6_TX] = { .type = NLA_U8 },
1231 [IFLA_GENEVE_UDP_ZERO_CSUM6_RX] = { .type = NLA_U8 },
2d07dc79
JL		 1232};
1233
1234static int geneve_validate(struct nlattr *tb[], struct nlattr *data[])
1235{
1236 if (tb[IFLA_ADDRESS]) {
1237 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1238 return -EINVAL;
1239
1240 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1241 return -EADDRNOTAVAIL;
1242 }
1243
1244 if (!data)
1245 return -EINVAL;
1246
1247 if (data[IFLA_GENEVE_ID]) {
1248 __u32 vni = nla_get_u32(data[IFLA_GENEVE_ID]);
1249
1250 if (vni >= GENEVE_VID_MASK)
1251 return -ERANGE;
1252 }
1253
1254 return 0;
1255}
1256
371bd106
PS		 1257static struct geneve_dev *geneve_find_dev(struct geneve_net *gn,
1258 __be16 dst_port,
8ed66f0e 1259 union geneve_addr *remote,
371bd106
PS		 1260 u8 vni[],
1261 bool *tun_on_same_port,
1262 bool *tun_collect_md)
1263{
1264 struct geneve_dev *geneve, *t;
1265
1266 *tun_on_same_port = false;
1267 *tun_collect_md = false;
1268 t = NULL;
1269 list_for_each_entry(geneve, &gn->geneve_list, next) {
1270 if (geneve->dst_port == dst_port) {
1271 *tun_collect_md = geneve->collect_md;
1272 *tun_on_same_port = true;
1273 }
1274 if (!memcmp(vni, geneve->vni, sizeof(geneve->vni)) &&
8ed66f0e 1275 !memcmp(remote, &geneve->remote, sizeof(geneve->remote)) &&
371bd106
PS		 1276 dst_port == geneve->dst_port)
1277 t = geneve;
1278 }
1279 return t;
1280}
1281
e305ac6c 1282static int geneve_configure(struct net *net, struct net_device *dev,
8ed66f0e 1283 union geneve_addr *remote,
8eb3b995
DB		 1284 __u32 vni, __u8 ttl, __u8 tos, __be32 label,
1285 __be16 dst_port, bool metadata, u32 flags)
2d07dc79
JL		 1286{
1287 struct geneve_net *gn = net_generic(net, geneve_net_id);
371bd106
PS		 1288 struct geneve_dev *t, *geneve = netdev_priv(dev);
1289 bool tun_collect_md, tun_on_same_port;
184fc8b5 1290 int err, encap_len;
2d07dc79 1291
8ed66f0e
JL		 1292 if (!remote)
1293 return -EINVAL;
1294 if (metadata &&
8eb3b995 1295 (remote->sa.sa_family != AF_UNSPEC || vni || tos || ttl || label))
8ed66f0e 1296 return -EINVAL;
2d07dc79
JL		 1297
1298 geneve->net = net;
1299 geneve->dev = dev;
1300
2d07dc79
JL		 1301 geneve->vni[0] = (vni & 0x00ff0000) >> 16;
1302 geneve->vni[1] = (vni & 0x0000ff00) >> 8;
1303 geneve->vni[2] = vni & 0x000000ff;
1304
8ed66f0e
JL		 1305 if ((remote->sa.sa_family == AF_INET &&
1306 IN_MULTICAST(ntohl(remote->sin.sin_addr.s_addr))) ||
1307 (remote->sa.sa_family == AF_INET6 &&
1308 ipv6_addr_is_multicast(&remote->sin6.sin6_addr)))
2d07dc79 1309 return -EINVAL;
8eb3b995
DB		 1310 if (label && remote->sa.sa_family != AF_INET6)
1311 return -EINVAL;
1312
8ed66f0e 1313 geneve->remote = *remote;
2d07dc79 1314
e305ac6c
PS		 1315 geneve->ttl = ttl;
1316 geneve->tos = tos;
8eb3b995 1317 geneve->label = label;
7bbe33ff 1318 geneve->dst_port = dst_port;
e305ac6c 1319 geneve->collect_md = metadata;
abe492b4 1320 geneve->flags = flags;
e305ac6c 1321
8ed66f0e 1322 t = geneve_find_dev(gn, dst_port, remote, geneve->vni,
371bd106
PS		 1323 &tun_on_same_port, &tun_collect_md);
1324 if (t)
1325 return -EBUSY;
1326
184fc8b5
PA		 1327 /* make enough headroom for basic scenario */
1328 encap_len = GENEVE_BASE_HLEN + ETH_HLEN;
1329 if (remote->sa.sa_family == AF_INET)
1330 encap_len += sizeof(struct iphdr);
1331 else
1332 encap_len += sizeof(struct ipv6hdr);
1333 dev->needed_headroom = encap_len + ETH_HLEN;
1334
371bd106
PS		 1335 if (metadata) {
1336 if (tun_on_same_port)
1337 return -EPERM;
1338 } else {
1339 if (tun_collect_md)
1340 return -EPERM;
1341 }
1342
468dfffc
PA		 1343 dst_cache_reset(&geneve->dst_cache);
1344
2d07dc79
JL		 1345 err = register_netdevice(dev);
1346 if (err)
1347 return err;
1348
e305ac6c 1349 list_add(&geneve->next, &gn->geneve_list);
e305ac6c
PS		 1350 return 0;
1351}
1352
1353static int geneve_newlink(struct net *net, struct net_device *dev,
1354 struct nlattr *tb[], struct nlattr *data[])
1355{
7bbe33ff 1356 __be16 dst_port = htons(GENEVE_UDP_PORT);
e305ac6c
PS		 1357 __u8 ttl = 0, tos = 0;
1358 bool metadata = false;
8ed66f0e 1359 union geneve_addr remote = geneve_remote_unspec;
8eb3b995 1360 __be32 label = 0;
e277de5f 1361 __u32 vni = 0;
abe492b4 1362 u32 flags = 0;
e305ac6c 1363
8ed66f0e
JL		 1364 if (data[IFLA_GENEVE_REMOTE] && data[IFLA_GENEVE_REMOTE6])
1365 return -EINVAL;
1366
1367 if (data[IFLA_GENEVE_REMOTE]) {
1368 remote.sa.sa_family = AF_INET;
1369 remote.sin.sin_addr.s_addr =
1370 nla_get_in_addr(data[IFLA_GENEVE_REMOTE]);
1371 }
1372
1373 if (data[IFLA_GENEVE_REMOTE6]) {
1374 if (!IS_ENABLED(CONFIG_IPV6))
1375 return -EPFNOSUPPORT;
1376
1377 remote.sa.sa_family = AF_INET6;
1378 remote.sin6.sin6_addr =
1379 nla_get_in6_addr(data[IFLA_GENEVE_REMOTE6]);
1380
1381 if (ipv6_addr_type(&remote.sin6.sin6_addr) &
1382 IPV6_ADDR_LINKLOCAL) {
1383 netdev_dbg(dev, "link-local remote is unsupported\n");
1384 return -EINVAL;
1385 }
1386 }
1387
e277de5f
JG		 1388 if (data[IFLA_GENEVE_ID])
1389 vni = nla_get_u32(data[IFLA_GENEVE_ID]);
e305ac6c 1390
8760ce58 1391 if (data[IFLA_GENEVE_TTL])
e305ac6c 1392 ttl = nla_get_u8(data[IFLA_GENEVE_TTL]);
8760ce58 1393
d8951125 1394 if (data[IFLA_GENEVE_TOS])
e305ac6c 1395 tos = nla_get_u8(data[IFLA_GENEVE_TOS]);
d8951125 1396
8eb3b995
DB		 1397 if (data[IFLA_GENEVE_LABEL])
1398 label = nla_get_be32(data[IFLA_GENEVE_LABEL]) &
1399 IPV6_FLOWLABEL_MASK;
1400
e305ac6c 1401 if (data[IFLA_GENEVE_PORT])
7bbe33ff 1402 dst_port = nla_get_be16(data[IFLA_GENEVE_PORT]);
2d07dc79 1403
e305ac6c
PS		 1404 if (data[IFLA_GENEVE_COLLECT_METADATA])
1405 metadata = true;
2d07dc79 1406
abe492b4 1407 if (data[IFLA_GENEVE_UDP_CSUM] &&
14f1f724
AD		 1408 !nla_get_u8(data[IFLA_GENEVE_UDP_CSUM]))
1409 flags |= GENEVE_F_UDP_ZERO_CSUM_TX;
abe492b4
TH		 1410
1411 if (data[IFLA_GENEVE_UDP_ZERO_CSUM6_TX] &&
1412 nla_get_u8(data[IFLA_GENEVE_UDP_ZERO_CSUM6_TX]))
1413 flags |= GENEVE_F_UDP_ZERO_CSUM6_TX;
1414
1415 if (data[IFLA_GENEVE_UDP_ZERO_CSUM6_RX] &&
1416 nla_get_u8(data[IFLA_GENEVE_UDP_ZERO_CSUM6_RX]))
1417 flags |= GENEVE_F_UDP_ZERO_CSUM6_RX;
1418
8eb3b995
DB		 1419 return geneve_configure(net, dev, &remote, vni, ttl, tos, label,
1420 dst_port, metadata, flags);
2d07dc79
JL		 1421}
1422
1423static void geneve_dellink(struct net_device *dev, struct list_head *head)
1424{
1425 struct geneve_dev *geneve = netdev_priv(dev);
1426
2d07dc79
JL		 1427 list_del(&geneve->next);
1428 unregister_netdevice_queue(dev, head);
1429}
1430
1431static size_t geneve_get_size(const struct net_device *dev)
1432{
1433 return nla_total_size(sizeof(__u32)) + /* IFLA_GENEVE_ID */
8ed66f0e 1434 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_GENEVE_REMOTE{6} */
8760ce58 1435 nla_total_size(sizeof(__u8)) + /* IFLA_GENEVE_TTL */
d8951125 1436 nla_total_size(sizeof(__u8)) + /* IFLA_GENEVE_TOS */
8eb3b995 1437 nla_total_size(sizeof(__be32)) + /* IFLA_GENEVE_LABEL */
7bbe33ff 1438 nla_total_size(sizeof(__be16)) + /* IFLA_GENEVE_PORT */
e305ac6c 1439 nla_total_size(0) + /* IFLA_GENEVE_COLLECT_METADATA */
abe492b4
TH		 1440 nla_total_size(sizeof(__u8)) + /* IFLA_GENEVE_UDP_CSUM */
1441 nla_total_size(sizeof(__u8)) + /* IFLA_GENEVE_UDP_ZERO_CSUM6_TX */
1442 nla_total_size(sizeof(__u8)) + /* IFLA_GENEVE_UDP_ZERO_CSUM6_RX */
2d07dc79
JL		 1443 0;
1444}
1445
1446static int geneve_fill_info(struct sk_buff *skb, const struct net_device *dev)
1447{
1448 struct geneve_dev *geneve = netdev_priv(dev);
1449 __u32 vni;
1450
1451 vni = (geneve->vni[0] << 16) | (geneve->vni[1] << 8) | geneve->vni[2];
1452 if (nla_put_u32(skb, IFLA_GENEVE_ID, vni))
1453 goto nla_put_failure;
1454
8ed66f0e
JL		 1455 if (geneve->remote.sa.sa_family == AF_INET) {
1456 if (nla_put_in_addr(skb, IFLA_GENEVE_REMOTE,
1457 geneve->remote.sin.sin_addr.s_addr))
1458 goto nla_put_failure;
1459#if IS_ENABLED(CONFIG_IPV6)
1460 } else {
1461 if (nla_put_in6_addr(skb, IFLA_GENEVE_REMOTE6,
1462 &geneve->remote.sin6.sin6_addr))
1463 goto nla_put_failure;
1464#endif
1465 }
2d07dc79 1466
d8951125 1467 if (nla_put_u8(skb, IFLA_GENEVE_TTL, geneve->ttl) ||
8eb3b995
DB		 1468 nla_put_u8(skb, IFLA_GENEVE_TOS, geneve->tos) ||
1469 nla_put_be32(skb, IFLA_GENEVE_LABEL, geneve->label))
8760ce58
JL		 1470 goto nla_put_failure;
1471
7bbe33ff 1472 if (nla_put_be16(skb, IFLA_GENEVE_PORT, geneve->dst_port))
cd7918b3
PS		 1473 goto nla_put_failure;
1474
e305ac6c
PS		 1475 if (geneve->collect_md) {
1476 if (nla_put_flag(skb, IFLA_GENEVE_COLLECT_METADATA))
1477 goto nla_put_failure;
1478 }
1479
abe492b4 1480 if (nla_put_u8(skb, IFLA_GENEVE_UDP_CSUM,
14f1f724 1481 !(geneve->flags & GENEVE_F_UDP_ZERO_CSUM_TX)) ||
abe492b4
TH		 1482 nla_put_u8(skb, IFLA_GENEVE_UDP_ZERO_CSUM6_TX,
1483 !!(geneve->flags & GENEVE_F_UDP_ZERO_CSUM6_TX)) ||
1484 nla_put_u8(skb, IFLA_GENEVE_UDP_ZERO_CSUM6_RX,
1485 !!(geneve->flags & GENEVE_F_UDP_ZERO_CSUM6_RX)))
1486 goto nla_put_failure;
1487
2d07dc79
JL		 1488 return 0;
1489
1490nla_put_failure:
1491 return -EMSGSIZE;
1492}
1493
1494static struct rtnl_link_ops geneve_link_ops __read_mostly = {
1495 .kind = "geneve",
1496 .maxtype = IFLA_GENEVE_MAX,
1497 .policy = geneve_policy,
1498 .priv_size = sizeof(struct geneve_dev),
1499 .setup = geneve_setup,
1500 .validate = geneve_validate,
1501 .newlink = geneve_newlink,
1502 .dellink = geneve_dellink,
1503 .get_size = geneve_get_size,
1504 .fill_info = geneve_fill_info,
1505};
1506
e305ac6c
PS		 1507struct net_device *geneve_dev_create_fb(struct net *net, const char *name,
1508 u8 name_assign_type, u16 dst_port)
1509{
1510 struct nlattr *tb[IFLA_MAX + 1];
1511 struct net_device *dev;
1512 int err;
1513
1514 memset(tb, 0, sizeof(tb));
1515 dev = rtnl_create_link(net, name, name_assign_type,
1516 &geneve_link_ops, tb);
1517 if (IS_ERR(dev))
1518 return dev;
1519
8ed66f0e 1520 err = geneve_configure(net, dev, &geneve_remote_unspec,
8eb3b995 1521 0, 0, 0, 0, htons(dst_port), true,
c868ee70 1522 GENEVE_F_UDP_ZERO_CSUM6_RX);
7e059158
DW		 1523 if (err)
1524 goto err;
1525
1526 /* openvswitch users expect packet sizes to be unrestricted,
1527 * so set the largest MTU we can.
1528 */
aeee0e66 1529 err = __geneve_change_mtu(dev, IP_MAX_MTU, false);
7e059158
DW		 1530 if (err)
1531 goto err;
1532
e305ac6c 1533 return dev;
7e059158
DW		 1534
1535 err:
1536 free_netdev(dev);
1537 return ERR_PTR(err);
e305ac6c
PS		 1538}
1539EXPORT_SYMBOL_GPL(geneve_dev_create_fb);
1540
2d07dc79
JL		 1541static __net_init int geneve_init_net(struct net *net)
1542{
1543 struct geneve_net *gn = net_generic(net, geneve_net_id);
2d07dc79
JL		 1544
1545 INIT_LIST_HEAD(&gn->geneve_list);
371bd106 1546 INIT_LIST_HEAD(&gn->sock_list);
2d07dc79
JL		 1547 return 0;
1548}
1549
1550static void __net_exit geneve_exit_net(struct net *net)
1551{
1552 struct geneve_net *gn = net_generic(net, geneve_net_id);
1553 struct geneve_dev *geneve, *next;
1554 struct net_device *dev, *aux;
1555 LIST_HEAD(list);
1556
1557 rtnl_lock();
1558
1559 /* gather any geneve devices that were moved into this ns */
1560 for_each_netdev_safe(net, dev, aux)
1561 if (dev->rtnl_link_ops == &geneve_link_ops)
1562 unregister_netdevice_queue(dev, &list);
1563
1564 /* now gather any other geneve devices that were created in this ns */
1565 list_for_each_entry_safe(geneve, next, &gn->geneve_list, next) {
1566 /* If geneve->dev is in the same netns, it was already added
1567 * to the list by the previous loop.
1568 */
1569 if (!net_eq(dev_net(geneve->dev), net))
1570 unregister_netdevice_queue(geneve->dev, &list);
1571 }
1572
1573 /* unregister the devices gathered above */
1574 unregister_netdevice_many(&list);
1575 rtnl_unlock();
1576}
1577
1578static struct pernet_operations geneve_net_ops = {
1579 .init = geneve_init_net,
1580 .exit = geneve_exit_net,
1581 .id = &geneve_net_id,
1582 .size = sizeof(struct geneve_net),
1583};
1584
1585static int __init geneve_init_module(void)
1586{
1587 int rc;
1588
1589 rc = register_pernet_subsys(&geneve_net_ops);
1590 if (rc)
1591 goto out1;
1592
1593 rc = rtnl_link_register(&geneve_link_ops);
1594 if (rc)
1595 goto out2;
1596
1597 return 0;
1598out2:
1599 unregister_pernet_subsys(&geneve_net_ops);
1600out1:
1601 return rc;
1602}
1603late_initcall(geneve_init_module);
1604
1605static void __exit geneve_cleanup_module(void)
1606{
1607 rtnl_link_unregister(&geneve_link_ops);
1608 unregister_pernet_subsys(&geneve_net_ops);
1609}
1610module_exit(geneve_cleanup_module);
1611
1612MODULE_LICENSE("GPL");
1613MODULE_VERSION(GENEVE_NETDEV_VER);
1614MODULE_AUTHOR("John W. Linville <linville@tuxdriver.com>");
1615MODULE_DESCRIPTION("Interface driver for GENEVE encapsulated traffic");
1616MODULE_ALIAS_RTNL_LINK("geneve");
Linux kernel - Deliverables
This page took 0.159353 seconds and 5 git commands to generate.