net/ipv4/geneve.c

   1 /*
   2  * Geneve: Generic Network Virtualization Encapsulation
   3  *
   4  * Copyright (c) 2014 Nicira, Inc.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11
  12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  13
  14 #include <linux/kernel.h>
  15 #include <linux/types.h>
  16 #include <linux/module.h>
  17 #include <linux/errno.h>
  18 #include <linux/slab.h>
  19 #include <linux/skbuff.h>
  20 #include <linux/rculist.h>
  21 #include <linux/netdevice.h>
  22 #include <linux/in.h>
  23 #include <linux/ip.h>
  24 #include <linux/udp.h>
  25 #include <linux/igmp.h>
  26 #include <linux/etherdevice.h>
  27 #include <linux/if_ether.h>
  28 #include <linux/if_vlan.h>
  29 #include <linux/hash.h>
  30 #include <linux/ethtool.h>
  31 #include <net/arp.h>
  32 #include <net/ndisc.h>
  33 #include <net/ip.h>
  34 #include <net/ip_tunnels.h>
  35 #include <net/icmp.h>
  36 #include <net/udp.h>
  37 #include <net/rtnetlink.h>
  38 #include <net/route.h>
  39 #include <net/dsfield.h>
  40 #include <net/inet_ecn.h>
  41 #include <net/net_namespace.h>
  42 #include <net/netns/generic.h>
  43 #include <net/geneve.h>
  44 #include <net/protocol.h>
  45 #include <net/udp_tunnel.h>
  46 #if IS_ENABLED(CONFIG_IPV6)
  47 #include <net/ipv6.h>
  48 #include <net/addrconf.h>
  49 #include <net/ip6_tunnel.h>
  50 #include <net/ip6_checksum.h>
  51 #endif
  52
  53 #define PORT_HASH_BITS 8
  54 #define PORT_HASH_SIZE (1<<PORT_HASH_BITS)
  55
  56 /* per-network namespace private data for this module */
  57 struct geneve_net {
  58         struct hlist_head       sock_list[PORT_HASH_SIZE];
  59         spinlock_t              sock_lock;   /* Protects sock_list */
  60 };
  61
  62 static int geneve_net_id;
  63
  64 static struct workqueue_struct *geneve_wq;
  65
  66 static inline struct genevehdr *geneve_hdr(const struct sk_buff *skb)
  67 {
  68         return (struct genevehdr *)(udp_hdr(skb) + 1);
  69 }
  70
  71 static struct hlist_head *gs_head(struct net *net, __be16 port)
  72 {
  73         struct geneve_net *gn = net_generic(net, geneve_net_id);
  74
  75         return &gn->sock_list[hash_32(ntohs(port), PORT_HASH_BITS)];
  76 }
  77
  78 /* Find geneve socket based on network namespace and UDP port */
  79 static struct geneve_sock *geneve_find_sock(struct net *net, __be16 port)
  80 {
  81         struct geneve_sock *gs;
  82
  83         hlist_for_each_entry_rcu(gs, gs_head(net, port), hlist) {
  84                 if (inet_sk(gs->sock->sk)->inet_sport == port)
  85                         return gs;
  86         }
  87
  88         return NULL;
  89 }
  90
  91 static void geneve_build_header(struct genevehdr *geneveh,
  92                                 __be16 tun_flags, u8 vni[3],
  93                                 u8 options_len, u8 *options)
  94 {
  95         geneveh->ver = GENEVE_VER;
  96         geneveh->opt_len = options_len / 4;
  97         geneveh->oam = !!(tun_flags & TUNNEL_OAM);
  98         geneveh->critical = !!(tun_flags & TUNNEL_CRIT_OPT);
  99         geneveh->rsvd1 = 0;
 100         memcpy(geneveh->vni, vni, 3);
 101         geneveh->proto_type = htons(ETH_P_TEB);
 102         geneveh->rsvd2 = 0;
 103
 104         memcpy(geneveh->options, options, options_len);
 105 }
 106
 107 /* Transmit a fully formatted Geneve frame.
 108  *
 109  * When calling this function. The skb->data should point
 110  * to the geneve header which is fully formed.
 111  *
 112  * This function will add other UDP tunnel headers.
 113  */
 114 int geneve_xmit_skb(struct geneve_sock *gs, struct rtable *rt,
 115                     struct sk_buff *skb, __be32 src, __be32 dst, __u8 tos,
 116                     __u8 ttl, __be16 df, __be16 src_port, __be16 dst_port,
 117                     __be16 tun_flags, u8 vni[3], u8 opt_len, u8 *opt,
 118                     bool xnet)
 119 {
 120         struct genevehdr *gnvh;
 121         int min_headroom;
 122         int err;
 123
 124         skb = udp_tunnel_handle_offloads(skb, !gs->sock->sk->sk_no_check_tx);
 125         if (IS_ERR(skb))
 126                 return PTR_ERR(skb);
 127
 128         min_headroom = LL_RESERVED_SPACE(rt->dst.dev) + rt->dst.header_len
 129                         + GENEVE_BASE_HLEN + opt_len + sizeof(struct iphdr)
 130                         + (vlan_tx_tag_present(skb) ? VLAN_HLEN : 0);
 131
 132         err = skb_cow_head(skb, min_headroom);
 133         if (unlikely(err)) {
 134                 kfree_skb(skb);
 135                 return err;
 136         }
 137
 138         skb = vlan_hwaccel_push_inside(skb);
 139         if (unlikely(!skb))
 140                 return -ENOMEM;
 141
 142         gnvh = (struct genevehdr *)__skb_push(skb, sizeof(*gnvh) + opt_len);
 143         geneve_build_header(gnvh, tun_flags, vni, opt_len, opt);
 144
 145         skb_set_inner_protocol(skb, htons(ETH_P_TEB));
 146
 147         return udp_tunnel_xmit_skb(gs->sock, rt, skb, src, dst,
 148                                    tos, ttl, df, src_port, dst_port, xnet);
 149 }
 150 EXPORT_SYMBOL_GPL(geneve_xmit_skb);
 151
 152 static void geneve_notify_add_rx_port(struct geneve_sock *gs)
 153 {
 154         struct sock *sk = gs->sock->sk;
 155         sa_family_t sa_family = sk->sk_family;
 156         int err;
 157
 158         if (sa_family == AF_INET) {
 159                 err = udp_add_offload(&gs->udp_offloads);
 160                 if (err)
 161                         pr_warn("geneve: udp_add_offload failed with status %d\n",
 162                                 err);
 163         }
 164 }
 165
 166 static void geneve_notify_del_rx_port(struct geneve_sock *gs)
 167 {
 168         struct sock *sk = gs->sock->sk;
 169         sa_family_t sa_family = sk->sk_family;
 170
 171         if (sa_family == AF_INET)
 172                 udp_del_offload(&gs->udp_offloads);
 173 }
 174
 175 /* Callback from net/ipv4/udp.c to receive packets */
 176 static int geneve_udp_encap_recv(struct sock *sk, struct sk_buff *skb)
 177 {
 178         struct genevehdr *geneveh;
 179         struct geneve_sock *gs;
 180         int opts_len;
 181
 182         /* Need Geneve and inner Ethernet header to be present */
 183         if (unlikely(!pskb_may_pull(skb, GENEVE_BASE_HLEN)))
 184                 goto error;
 185
 186         /* Return packets with reserved bits set */
 187         geneveh = geneve_hdr(skb);
 188
 189         if (unlikely(geneveh->ver != GENEVE_VER))
 190                 goto error;
 191
 192         if (unlikely(geneveh->proto_type != htons(ETH_P_TEB)))
 193                 goto error;
 194
 195         opts_len = geneveh->opt_len * 4;
 196         if (iptunnel_pull_header(skb, GENEVE_BASE_HLEN + opts_len,
 197                                  htons(ETH_P_TEB)))
 198                 goto drop;
 199
 200         gs = rcu_dereference_sk_user_data(sk);
 201         if (!gs)
 202                 goto drop;
 203
 204         gs->rcv(gs, skb);
 205         return 0;
 206
 207 drop:
 208         /* Consume bad packet */
 209         kfree_skb(skb);
 210         return 0;
 211
 212 error:
 213         /* Let the UDP layer deal with the skb */
 214         return 1;
 215 }
 216
 217 static void geneve_del_work(struct work_struct *work)
 218 {
 219         struct geneve_sock *gs = container_of(work, struct geneve_sock,
 220                                               del_work);
 221
 222         udp_tunnel_sock_release(gs->sock);
 223         kfree_rcu(gs, rcu);
 224 }
 225
 226 static struct socket *geneve_create_sock(struct net *net, bool ipv6,
 227                                          __be16 port)
 228 {
 229         struct socket *sock;
 230         struct udp_port_cfg udp_conf;
 231         int err;
 232
 233         memset(&udp_conf, 0, sizeof(udp_conf));
 234
 235         if (ipv6) {
 236                 udp_conf.family = AF_INET6;
 237         } else {
 238                 udp_conf.family = AF_INET;
 239                 udp_conf.local_ip.s_addr = htonl(INADDR_ANY);
 240         }
 241
 242         udp_conf.local_udp_port = port;
 243
 244         /* Open UDP socket */
 245         err = udp_sock_create(net, &udp_conf, &sock);
 246         if (err < 0)
 247                 return ERR_PTR(err);
 248
 249         return sock;
 250 }
 251
 252 /* Create new listen socket if needed */
 253 static struct geneve_sock *geneve_socket_create(struct net *net, __be16 port,
 254                                                 geneve_rcv_t *rcv, void *data,
 255                                                 bool ipv6)
 256 {
 257         struct geneve_net *gn = net_generic(net, geneve_net_id);
 258         struct geneve_sock *gs;
 259         struct socket *sock;
 260         struct udp_tunnel_sock_cfg tunnel_cfg;
 261
 262         gs = kzalloc(sizeof(*gs), GFP_KERNEL);
 263         if (!gs)
 264                 return ERR_PTR(-ENOMEM);
 265
 266         INIT_WORK(&gs->del_work, geneve_del_work);
 267
 268         sock = geneve_create_sock(net, ipv6, port);
 269         if (IS_ERR(sock)) {
 270                 kfree(gs);
 271                 return ERR_CAST(sock);
 272         }
 273
 274         gs->sock = sock;
 275         atomic_set(&gs->refcnt, 1);
 276         gs->rcv = rcv;
 277         gs->rcv_data = data;
 278
 279         /* Initialize the geneve udp offloads structure */
 280         gs->udp_offloads.port = port;
 281         gs->udp_offloads.callbacks.gro_receive = NULL;
 282         gs->udp_offloads.callbacks.gro_complete = NULL;
 283
 284         spin_lock(&gn->sock_lock);
 285         hlist_add_head_rcu(&gs->hlist, gs_head(net, port));
 286         geneve_notify_add_rx_port(gs);
 287         spin_unlock(&gn->sock_lock);
 288
 289         /* Mark socket as an encapsulation socket */
 290         tunnel_cfg.sk_user_data = gs;
 291         tunnel_cfg.encap_type = 1;
 292         tunnel_cfg.encap_rcv = geneve_udp_encap_recv;
 293         tunnel_cfg.encap_destroy = NULL;
 294         setup_udp_tunnel_sock(net, sock, &tunnel_cfg);
 295
 296         return gs;
 297 }
 298
 299 struct geneve_sock *geneve_sock_add(struct net *net, __be16 port,
 300                                     geneve_rcv_t *rcv, void *data,
 301                                     bool no_share, bool ipv6)
 302 {
 303         struct geneve_net *gn = net_generic(net, geneve_net_id);
 304         struct geneve_sock *gs;
 305
 306         gs = geneve_socket_create(net, port, rcv, data, ipv6);
 307         if (!IS_ERR(gs))
 308                 return gs;
 309
 310         if (no_share)   /* Return error if sharing is not allowed. */
 311                 return ERR_PTR(-EINVAL);
 312
 313         spin_lock(&gn->sock_lock);
 314         gs = geneve_find_sock(net, port);
 315         if (gs && ((gs->rcv != rcv) ||
 316                    !atomic_add_unless(&gs->refcnt, 1, 0)))
 317                         gs = ERR_PTR(-EBUSY);
 318         spin_unlock(&gn->sock_lock);
 319
 320         if (!gs)
 321                 gs = ERR_PTR(-EINVAL);
 322
 323         return gs;
 324 }
 325 EXPORT_SYMBOL_GPL(geneve_sock_add);
 326
 327 void geneve_sock_release(struct geneve_sock *gs)
 328 {
 329         struct net *net = sock_net(gs->sock->sk);
 330         struct geneve_net *gn = net_generic(net, geneve_net_id);
 331
 332         if (!atomic_dec_and_test(&gs->refcnt))
 333                 return;
 334
 335         spin_lock(&gn->sock_lock);
 336         hlist_del_rcu(&gs->hlist);
 337         geneve_notify_del_rx_port(gs);
 338         spin_unlock(&gn->sock_lock);
 339
 340         queue_work(geneve_wq, &gs->del_work);
 341 }
 342 EXPORT_SYMBOL_GPL(geneve_sock_release);
 343
 344 static __net_init int geneve_init_net(struct net *net)
 345 {
 346         struct geneve_net *gn = net_generic(net, geneve_net_id);
 347         unsigned int h;
 348
 349         spin_lock_init(&gn->sock_lock);
 350
 351         for (h = 0; h < PORT_HASH_SIZE; ++h)
 352                 INIT_HLIST_HEAD(&gn->sock_list[h]);
 353
 354         return 0;
 355 }
 356
 357 static struct pernet_operations geneve_net_ops = {
 358         .init = geneve_init_net,
 359         .exit = NULL,
 360         .id   = &geneve_net_id,
 361         .size = sizeof(struct geneve_net),
 362 };
 363
 364 static int __init geneve_init_module(void)
 365 {
 366         int rc;
 367
 368         geneve_wq = alloc_workqueue("geneve", 0, 0);
 369         if (!geneve_wq)
 370                 return -ENOMEM;
 371
 372         rc = register_pernet_subsys(&geneve_net_ops);
 373         if (rc)
 374                 return rc;
 375
 376         pr_info("Geneve driver\n");
 377
 378         return 0;
 379 }
 380 late_initcall(geneve_init_module);
 381
 382 static void __exit geneve_cleanup_module(void)
 383 {
 384         destroy_workqueue(geneve_wq);
 385         unregister_pernet_subsys(&geneve_net_ops);
 386 }
 387 module_exit(geneve_cleanup_module);
 388
 389 MODULE_LICENSE("GPL");
 390 MODULE_AUTHOR("Jesse Gross <jesse@nicira.com>");
 391 MODULE_DESCRIPTION("Driver for GENEVE encapsulated traffic");
 392 MODULE_ALIAS_RTNL_LINK("geneve");