drivers/net/bareudp.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Bareudp: UDP  tunnel encasulation for different Payload types like
   3  * MPLS, NSH, IP, etc.
   4  * Copyright (c) 2019 Nokia, Inc.
   5  * Authors:  Martin Varghese, <martin.varghese@nokia.com>
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/kernel.h>
  11 #include <linux/module.h>
  12 #include <linux/etherdevice.h>
  13 #include <linux/hash.h>
  14 #include <net/dst_metadata.h>
  15 #include <net/gro_cells.h>
  16 #include <net/rtnetlink.h>
  17 #include <net/protocol.h>
  18 #include <net/ip6_tunnel.h>
  19 #include <net/ip_tunnels.h>
  20 #include <net/udp_tunnel.h>
  21 #include <net/bareudp.h>
  22
  23 #define BAREUDP_BASE_HLEN sizeof(struct udphdr)
  24 #define BAREUDP_IPV4_HLEN (sizeof(struct iphdr) + \
  25                            sizeof(struct udphdr))
  26 #define BAREUDP_IPV6_HLEN (sizeof(struct ipv6hdr) + \
  27                            sizeof(struct udphdr))
  28
  29 static bool log_ecn_error = true;
  30 module_param(log_ecn_error, bool, 0644);
  31 MODULE_PARM_DESC(log_ecn_error, "Log packets received with corrupted ECN");
  32
  33 /* per-network namespace private data for this module */
  34
  35 static unsigned int bareudp_net_id;
  36
  37 struct bareudp_net {
  38         struct list_head        bareudp_list;
  39 };
  40
  41 struct bareudp_conf {
  42         __be16 ethertype;
  43         __be16 port;
  44         u16 sport_min;
  45         bool multi_proto_mode;
  46 };
  47
  48 /* Pseudo network device */
  49 struct bareudp_dev {
  50         struct net         *net;        /* netns for packet i/o */
  51         struct net_device  *dev;        /* netdev for bareudp tunnel */
  52         __be16             ethertype;
  53         __be16             port;
  54         u16                sport_min;
  55         bool               multi_proto_mode;
  56         struct socket      __rcu *sock;
  57         struct list_head   next;        /* bareudp node  on namespace list */
  58         struct gro_cells   gro_cells;
  59 };
  60
  61 static int bareudp_udp_encap_recv(struct sock *sk, struct sk_buff *skb)
  62 {
  63         struct metadata_dst *tun_dst = NULL;
  64         struct bareudp_dev *bareudp;
  65         unsigned short family;
  66         unsigned int len;
  67         __be16 proto;
  68         void *oiph;
  69         int err;
  70
  71         bareudp = rcu_dereference_sk_user_data(sk);
  72         if (!bareudp)
  73                 goto drop;
  74
  75         if (skb->protocol ==  htons(ETH_P_IP))
  76                 family = AF_INET;
  77         else
  78                 family = AF_INET6;
  79
  80         if (bareudp->ethertype == htons(ETH_P_IP)) {
  81                 __u8 ipversion;
  82
  83                 if (skb_copy_bits(skb, BAREUDP_BASE_HLEN, &ipversion,
  84                                   sizeof(ipversion))) {
  85                         bareudp->dev->stats.rx_dropped++;
  86                         goto drop;
  87                 }
  88                 ipversion >>= 4;
  89
  90                 if (ipversion == 4) {
  91                         proto = htons(ETH_P_IP);
  92                 } else if (ipversion == 6 && bareudp->multi_proto_mode) {
  93                         proto = htons(ETH_P_IPV6);
  94                 } else {
  95                         bareudp->dev->stats.rx_dropped++;
  96                         goto drop;
  97                 }
  98         } else if (bareudp->ethertype == htons(ETH_P_MPLS_UC)) {
  99                 struct iphdr *tunnel_hdr;
 100
 101                 tunnel_hdr = (struct iphdr *)skb_network_header(skb);
 102                 if (tunnel_hdr->version == 4) {
 103                         if (!ipv4_is_multicast(tunnel_hdr->daddr)) {
 104                                 proto = bareudp->ethertype;
 105                         } else if (bareudp->multi_proto_mode &&
 106                                    ipv4_is_multicast(tunnel_hdr->daddr)) {
 107                                 proto = htons(ETH_P_MPLS_MC);
 108                         } else {
 109                                 bareudp->dev->stats.rx_dropped++;
 110                                 goto drop;
 111                         }
 112                 } else {
 113                         int addr_type;
 114                         struct ipv6hdr *tunnel_hdr_v6;
 115
 116                         tunnel_hdr_v6 = (struct ipv6hdr *)skb_network_header(skb);
 117                         addr_type =
 118                         ipv6_addr_type((struct in6_addr *)&tunnel_hdr_v6->daddr);
 119                         if (!(addr_type & IPV6_ADDR_MULTICAST)) {
 120                                 proto = bareudp->ethertype;
 121                         } else if (bareudp->multi_proto_mode &&
 122                                    (addr_type & IPV6_ADDR_MULTICAST)) {
 123                                 proto = htons(ETH_P_MPLS_MC);
 124                         } else {
 125                                 bareudp->dev->stats.rx_dropped++;
 126                                 goto drop;
 127                         }
 128                 }
 129         } else {
 130                 proto = bareudp->ethertype;
 131         }
 132
 133         if (iptunnel_pull_header(skb, BAREUDP_BASE_HLEN,
 134                                  proto,
 135                                  !net_eq(bareudp->net,
 136                                  dev_net(bareudp->dev)))) {
 137                 bareudp->dev->stats.rx_dropped++;
 138                 goto drop;
 139         }
 140         tun_dst = udp_tun_rx_dst(skb, family, TUNNEL_KEY, 0, 0);
 141         if (!tun_dst) {
 142                 bareudp->dev->stats.rx_dropped++;
 143                 goto drop;
 144         }
 145         skb_dst_set(skb, &tun_dst->dst);
 146         skb->dev = bareudp->dev;
 147         oiph = skb_network_header(skb);
 148         skb_reset_network_header(skb);
 149         skb_reset_mac_header(skb);
 150
 151         if (!IS_ENABLED(CONFIG_IPV6) || family == AF_INET)
 152                 err = IP_ECN_decapsulate(oiph, skb);
 153         else
 154                 err = IP6_ECN_decapsulate(oiph, skb);
 155
 156         if (unlikely(err)) {
 157                 if (log_ecn_error) {
 158                         if  (!IS_ENABLED(CONFIG_IPV6) || family == AF_INET)
 159                                 net_info_ratelimited("non-ECT from %pI4 "
 160                                                      "with TOS=%#x\n",
 161                                                      &((struct iphdr *)oiph)->saddr,
 162                                                      ((struct iphdr *)oiph)->tos);
 163                         else
 164                                 net_info_ratelimited("non-ECT from %pI6\n",
 165                                                      &((struct ipv6hdr *)oiph)->saddr);
 166                 }
 167                 if (err > 1) {
 168                         ++bareudp->dev->stats.rx_frame_errors;
 169                         ++bareudp->dev->stats.rx_errors;
 170                         goto drop;
 171                 }
 172         }
 173
 174         len = skb->len;
 175         err = gro_cells_receive(&bareudp->gro_cells, skb);
 176         if (likely(err == NET_RX_SUCCESS))
 177                 dev_sw_netstats_rx_add(bareudp->dev, len);
 178
 179         return 0;
 180 drop:
 181         /* Consume bad packet */
 182         kfree_skb(skb);
 183
 184         return 0;
 185 }
 186
 187 static int bareudp_err_lookup(struct sock *sk, struct sk_buff *skb)
 188 {
 189         return 0;
 190 }
 191
 192 static int bareudp_init(struct net_device *dev)
 193 {
 194         struct bareudp_dev *bareudp = netdev_priv(dev);
 195         int err;
 196
 197         dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
 198         if (!dev->tstats)
 199                 return -ENOMEM;
 200
 201         err = gro_cells_init(&bareudp->gro_cells, dev);
 202         if (err) {
 203                 free_percpu(dev->tstats);
 204                 return err;
 205         }
 206         return 0;
 207 }
 208
 209 static void bareudp_uninit(struct net_device *dev)
 210 {
 211         struct bareudp_dev *bareudp = netdev_priv(dev);
 212
 213         gro_cells_destroy(&bareudp->gro_cells);
 214         free_percpu(dev->tstats);
 215 }
 216
 217 static struct socket *bareudp_create_sock(struct net *net, __be16 port)
 218 {
 219         struct udp_port_cfg udp_conf;
 220         struct socket *sock;
 221         int err;
 222
 223         memset(&udp_conf, 0, sizeof(udp_conf));
 224 #if IS_ENABLED(CONFIG_IPV6)
 225         udp_conf.family = AF_INET6;
 226 #else
 227         udp_conf.family = AF_INET;
 228 #endif
 229         udp_conf.local_udp_port = port;
 230         /* Open UDP socket */
 231         err = udp_sock_create(net, &udp_conf, &sock);
 232         if (err < 0)
 233                 return ERR_PTR(err);
 234
 235         udp_allow_gso(sock->sk);
 236         return sock;
 237 }
 238
 239 /* Create new listen socket if needed */
 240 static int bareudp_socket_create(struct bareudp_dev *bareudp, __be16 port)
 241 {
 242         struct udp_tunnel_sock_cfg tunnel_cfg;
 243         struct socket *sock;
 244
 245         sock = bareudp_create_sock(bareudp->net, port);
 246         if (IS_ERR(sock))
 247                 return PTR_ERR(sock);
 248
 249         /* Mark socket as an encapsulation socket */
 250         memset(&tunnel_cfg, 0, sizeof(tunnel_cfg));
 251         tunnel_cfg.sk_user_data = bareudp;
 252         tunnel_cfg.encap_type = 1;
 253         tunnel_cfg.encap_rcv = bareudp_udp_encap_recv;
 254         tunnel_cfg.encap_err_lookup = bareudp_err_lookup;
 255         tunnel_cfg.encap_destroy = NULL;
 256         setup_udp_tunnel_sock(bareudp->net, sock, &tunnel_cfg);
 257
 258         rcu_assign_pointer(bareudp->sock, sock);
 259         return 0;
 260 }
 261
 262 static int bareudp_open(struct net_device *dev)
 263 {
 264         struct bareudp_dev *bareudp = netdev_priv(dev);
 265         int ret = 0;
 266
 267         ret =  bareudp_socket_create(bareudp, bareudp->port);
 268         return ret;
 269 }
 270
 271 static void bareudp_sock_release(struct bareudp_dev *bareudp)
 272 {
 273         struct socket *sock;
 274
 275         sock = bareudp->sock;
 276         rcu_assign_pointer(bareudp->sock, NULL);
 277         synchronize_net();
 278         udp_tunnel_sock_release(sock);
 279 }
 280
 281 static int bareudp_stop(struct net_device *dev)
 282 {
 283         struct bareudp_dev *bareudp = netdev_priv(dev);
 284
 285         bareudp_sock_release(bareudp);
 286         return 0;
 287 }
 288
 289 static int bareudp_xmit_skb(struct sk_buff *skb, struct net_device *dev,
 290                             struct bareudp_dev *bareudp,
 291                             const struct ip_tunnel_info *info)
 292 {
 293         bool xnet = !net_eq(bareudp->net, dev_net(bareudp->dev));
 294         bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
 295         struct socket *sock = rcu_dereference(bareudp->sock);
 296         bool udp_sum = !!(info->key.tun_flags & TUNNEL_CSUM);
 297         const struct ip_tunnel_key *key = &info->key;
 298         struct rtable *rt;
 299         __be16 sport, df;
 300         int min_headroom;
 301         __u8 tos, ttl;
 302         __be32 saddr;
 303         int err;
 304
 305         if (!sock)
 306                 return -ESHUTDOWN;
 307
 308         rt = ip_route_output_tunnel(skb, dev, bareudp->net, &saddr, info,
 309                                     IPPROTO_UDP, use_cache);
 310
 311         if (IS_ERR(rt))
 312                 return PTR_ERR(rt);
 313
 314         skb_tunnel_check_pmtu(skb, &rt->dst,
 315                               BAREUDP_IPV4_HLEN + info->options_len, false);
 316
 317         sport = udp_flow_src_port(bareudp->net, skb,
 318                                   bareudp->sport_min, USHRT_MAX,
 319                                   true);
 320         tos = ip_tunnel_ecn_encap(key->tos, ip_hdr(skb), skb);
 321         ttl = key->ttl;
 322         df = key->tun_flags & TUNNEL_DONT_FRAGMENT ? htons(IP_DF) : 0;
 323         skb_scrub_packet(skb, xnet);
 324
 325         err = -ENOSPC;
 326         if (!skb_pull(skb, skb_network_offset(skb)))
 327                 goto free_dst;
 328
 329         min_headroom = LL_RESERVED_SPACE(rt->dst.dev) + rt->dst.header_len +
 330                 BAREUDP_BASE_HLEN + info->options_len + sizeof(struct iphdr);
 331
 332         err = skb_cow_head(skb, min_headroom);
 333         if (unlikely(err))
 334                 goto free_dst;
 335
 336         err = udp_tunnel_handle_offloads(skb, udp_sum);
 337         if (err)
 338                 goto free_dst;
 339
 340         skb_set_inner_protocol(skb, bareudp->ethertype);
 341         udp_tunnel_xmit_skb(rt, sock->sk, skb, saddr, info->key.u.ipv4.dst,
 342                             tos, ttl, df, sport, bareudp->port,
 343                             !net_eq(bareudp->net, dev_net(bareudp->dev)),
 344                             !(info->key.tun_flags & TUNNEL_CSUM));
 345         return 0;
 346
 347 free_dst:
 348         dst_release(&rt->dst);
 349         return err;
 350 }
 351
 352 static int bareudp6_xmit_skb(struct sk_buff *skb, struct net_device *dev,
 353                              struct bareudp_dev *bareudp,
 354                              const struct ip_tunnel_info *info)
 355 {
 356         bool xnet = !net_eq(bareudp->net, dev_net(bareudp->dev));
 357         bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
 358         struct socket *sock  = rcu_dereference(bareudp->sock);
 359         bool udp_sum = !!(info->key.tun_flags & TUNNEL_CSUM);
 360         const struct ip_tunnel_key *key = &info->key;
 361         struct dst_entry *dst = NULL;
 362         struct in6_addr saddr, daddr;
 363         int min_headroom;
 364         __u8 prio, ttl;
 365         __be16 sport;
 366         int err;
 367
 368         if (!sock)
 369                 return -ESHUTDOWN;
 370
 371         dst = ip6_dst_lookup_tunnel(skb, dev, bareudp->net, sock, &saddr, info,
 372                                     IPPROTO_UDP, use_cache);
 373         if (IS_ERR(dst))
 374                 return PTR_ERR(dst);
 375
 376         skb_tunnel_check_pmtu(skb, dst, BAREUDP_IPV6_HLEN + info->options_len,
 377                               false);
 378
 379         sport = udp_flow_src_port(bareudp->net, skb,
 380                                   bareudp->sport_min, USHRT_MAX,
 381                                   true);
 382         prio = ip_tunnel_ecn_encap(key->tos, ip_hdr(skb), skb);
 383         ttl = key->ttl;
 384
 385         skb_scrub_packet(skb, xnet);
 386
 387         err = -ENOSPC;
 388         if (!skb_pull(skb, skb_network_offset(skb)))
 389                 goto free_dst;
 390
 391         min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len +
 392                 BAREUDP_BASE_HLEN + info->options_len + sizeof(struct ipv6hdr);
 393
 394         err = skb_cow_head(skb, min_headroom);
 395         if (unlikely(err))
 396                 goto free_dst;
 397
 398         err = udp_tunnel_handle_offloads(skb, udp_sum);
 399         if (err)
 400                 goto free_dst;
 401
 402         daddr = info->key.u.ipv6.dst;
 403         udp_tunnel6_xmit_skb(dst, sock->sk, skb, dev,
 404                              &saddr, &daddr, prio, ttl,
 405                              info->key.label, sport, bareudp->port,
 406                              !(info->key.tun_flags & TUNNEL_CSUM));
 407         return 0;
 408
 409 free_dst:
 410         dst_release(dst);
 411         return err;
 412 }
 413
 414 static bool bareudp_proto_valid(struct bareudp_dev *bareudp, __be16 proto)
 415 {
 416         if (bareudp->ethertype == proto)
 417                 return true;
 418
 419         if (!bareudp->multi_proto_mode)
 420                 return false;
 421
 422         if (bareudp->ethertype == htons(ETH_P_MPLS_UC) &&
 423             proto == htons(ETH_P_MPLS_MC))
 424                 return true;
 425
 426         if (bareudp->ethertype == htons(ETH_P_IP) &&
 427             proto == htons(ETH_P_IPV6))
 428                 return true;
 429
 430         return false;
 431 }
 432
 433 static netdev_tx_t bareudp_xmit(struct sk_buff *skb, struct net_device *dev)
 434 {
 435         struct bareudp_dev *bareudp = netdev_priv(dev);
 436         struct ip_tunnel_info *info = NULL;
 437         int err;
 438
 439         if (!bareudp_proto_valid(bareudp, skb->protocol)) {
 440                 err = -EINVAL;
 441                 goto tx_error;
 442         }
 443
 444         info = skb_tunnel_info(skb);
 445         if (unlikely(!info || !(info->mode & IP_TUNNEL_INFO_TX))) {
 446                 err = -EINVAL;
 447                 goto tx_error;
 448         }
 449
 450         rcu_read_lock();
 451         if (IS_ENABLED(CONFIG_IPV6) && info->mode & IP_TUNNEL_INFO_IPV6)
 452                 err = bareudp6_xmit_skb(skb, dev, bareudp, info);
 453         else
 454                 err = bareudp_xmit_skb(skb, dev, bareudp, info);
 455
 456         rcu_read_unlock();
 457
 458         if (likely(!err))
 459                 return NETDEV_TX_OK;
 460 tx_error:
 461         dev_kfree_skb(skb);
 462
 463         if (err == -ELOOP)
 464                 dev->stats.collisions++;
 465         else if (err == -ENETUNREACH)
 466                 dev->stats.tx_carrier_errors++;
 467
 468         dev->stats.tx_errors++;
 469         return NETDEV_TX_OK;
 470 }
 471
 472 static int bareudp_fill_metadata_dst(struct net_device *dev,
 473                                      struct sk_buff *skb)
 474 {
 475         struct ip_tunnel_info *info = skb_tunnel_info(skb);
 476         struct bareudp_dev *bareudp = netdev_priv(dev);
 477         bool use_cache;
 478
 479         use_cache = ip_tunnel_dst_cache_usable(skb, info);
 480
 481         if (!IS_ENABLED(CONFIG_IPV6) || ip_tunnel_info_af(info) == AF_INET) {
 482                 struct rtable *rt;
 483                 __be32 saddr;
 484
 485                 rt = ip_route_output_tunnel(skb, dev, bareudp->net, &saddr,
 486                                             info, IPPROTO_UDP, use_cache);
 487                 if (IS_ERR(rt))
 488                         return PTR_ERR(rt);
 489
 490                 ip_rt_put(rt);
 491                 info->key.u.ipv4.src = saddr;
 492         } else if (ip_tunnel_info_af(info) == AF_INET6) {
 493                 struct dst_entry *dst;
 494                 struct in6_addr saddr;
 495                 struct socket *sock = rcu_dereference(bareudp->sock);
 496
 497                 dst = ip6_dst_lookup_tunnel(skb, dev, bareudp->net, sock,
 498                                             &saddr, info, IPPROTO_UDP,
 499                                             use_cache);
 500                 if (IS_ERR(dst))
 501                         return PTR_ERR(dst);
 502
 503                 dst_release(dst);
 504                 info->key.u.ipv6.src = saddr;
 505         } else {
 506                 return -EINVAL;
 507         }
 508
 509         info->key.tp_src = udp_flow_src_port(bareudp->net, skb,
 510                                              bareudp->sport_min,
 511                         USHRT_MAX, true);
 512         info->key.tp_dst = bareudp->port;
 513         return 0;
 514 }
 515
 516 static const struct net_device_ops bareudp_netdev_ops = {
 517         .ndo_init               = bareudp_init,
 518         .ndo_uninit             = bareudp_uninit,
 519         .ndo_open               = bareudp_open,
 520         .ndo_stop               = bareudp_stop,
 521         .ndo_start_xmit         = bareudp_xmit,
 522         .ndo_get_stats64        = dev_get_tstats64,
 523         .ndo_fill_metadata_dst  = bareudp_fill_metadata_dst,
 524 };
 525
 526 static const struct nla_policy bareudp_policy[IFLA_BAREUDP_MAX + 1] = {
 527         [IFLA_BAREUDP_PORT]                = { .type = NLA_U16 },
 528         [IFLA_BAREUDP_ETHERTYPE]           = { .type = NLA_U16 },
 529         [IFLA_BAREUDP_SRCPORT_MIN]         = { .type = NLA_U16 },
 530         [IFLA_BAREUDP_MULTIPROTO_MODE]     = { .type = NLA_FLAG },
 531 };
 532
 533 /* Info for udev, that this is a virtual tunnel endpoint */
 534 static const struct device_type bareudp_type = {
 535         .name = "bareudp",
 536 };
 537
 538 /* Initialize the device structure. */
 539 static void bareudp_setup(struct net_device *dev)
 540 {
 541         dev->netdev_ops = &bareudp_netdev_ops;
 542         dev->needs_free_netdev = true;
 543         SET_NETDEV_DEVTYPE(dev, &bareudp_type);
 544         dev->features    |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_FRAGLIST;
 545         dev->features    |= NETIF_F_RXCSUM;
 546         dev->features    |= NETIF_F_LLTX;
 547         dev->features    |= NETIF_F_GSO_SOFTWARE;
 548         dev->hw_features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_FRAGLIST;
 549         dev->hw_features |= NETIF_F_RXCSUM;
 550         dev->hw_features |= NETIF_F_GSO_SOFTWARE;
 551         dev->hard_header_len = 0;
 552         dev->addr_len = 0;
 553         dev->mtu = ETH_DATA_LEN;
 554         dev->min_mtu = IPV4_MIN_MTU;
 555         dev->max_mtu = IP_MAX_MTU - BAREUDP_BASE_HLEN;
 556         dev->type = ARPHRD_NONE;
 557         netif_keep_dst(dev);
 558         dev->priv_flags |= IFF_NO_QUEUE;
 559         dev->flags = IFF_POINTOPOINT | IFF_NOARP | IFF_MULTICAST;
 560 }
 561
 562 static int bareudp_validate(struct nlattr *tb[], struct nlattr *data[],
 563                             struct netlink_ext_ack *extack)
 564 {
 565         if (!data) {
 566                 NL_SET_ERR_MSG(extack,
 567                                "Not enough attributes provided to perform the operation");
 568                 return -EINVAL;
 569         }
 570         return 0;
 571 }
 572
 573 static int bareudp2info(struct nlattr *data[], struct bareudp_conf *conf,
 574                         struct netlink_ext_ack *extack)
 575 {
 576         memset(conf, 0, sizeof(*conf));
 577
 578         if (!data[IFLA_BAREUDP_PORT]) {
 579                 NL_SET_ERR_MSG(extack, "port not specified");
 580                 return -EINVAL;
 581         }
 582         if (!data[IFLA_BAREUDP_ETHERTYPE]) {
 583                 NL_SET_ERR_MSG(extack, "ethertype not specified");
 584                 return -EINVAL;
 585         }
 586
 587         conf->port = nla_get_u16(data[IFLA_BAREUDP_PORT]);
 588         conf->ethertype = nla_get_u16(data[IFLA_BAREUDP_ETHERTYPE]);
 589
 590         if (data[IFLA_BAREUDP_SRCPORT_MIN])
 591                 conf->sport_min =  nla_get_u16(data[IFLA_BAREUDP_SRCPORT_MIN]);
 592
 593         if (data[IFLA_BAREUDP_MULTIPROTO_MODE])
 594                 conf->multi_proto_mode = true;
 595
 596         return 0;
 597 }
 598
 599 static struct bareudp_dev *bareudp_find_dev(struct bareudp_net *bn,
 600                                             const struct bareudp_conf *conf)
 601 {
 602         struct bareudp_dev *bareudp, *t = NULL;
 603
 604         list_for_each_entry(bareudp, &bn->bareudp_list, next) {
 605                 if (conf->port == bareudp->port)
 606                         t = bareudp;
 607         }
 608         return t;
 609 }
 610
 611 static int bareudp_configure(struct net *net, struct net_device *dev,
 612                              struct bareudp_conf *conf,
 613                              struct netlink_ext_ack *extack)
 614 {
 615         struct bareudp_net *bn = net_generic(net, bareudp_net_id);
 616         struct bareudp_dev *t, *bareudp = netdev_priv(dev);
 617         int err;
 618
 619         bareudp->net = net;
 620         bareudp->dev = dev;
 621         t = bareudp_find_dev(bn, conf);
 622         if (t) {
 623                 NL_SET_ERR_MSG(extack, "Another bareudp device using the same port already exists");
 624                 return -EBUSY;
 625         }
 626
 627         if (conf->multi_proto_mode &&
 628             (conf->ethertype != htons(ETH_P_MPLS_UC) &&
 629              conf->ethertype != htons(ETH_P_IP))) {
 630                 NL_SET_ERR_MSG(extack, "Cannot set multiproto mode for this ethertype (only IPv4 and unicast MPLS are supported)");
 631                 return -EINVAL;
 632         }
 633
 634         bareudp->port = conf->port;
 635         bareudp->ethertype = conf->ethertype;
 636         bareudp->sport_min = conf->sport_min;
 637         bareudp->multi_proto_mode = conf->multi_proto_mode;
 638
 639         err = register_netdevice(dev);
 640         if (err)
 641                 return err;
 642
 643         list_add(&bareudp->next, &bn->bareudp_list);
 644         return 0;
 645 }
 646
 647 static int bareudp_link_config(struct net_device *dev,
 648                                struct nlattr *tb[])
 649 {
 650         int err;
 651
 652         if (tb[IFLA_MTU]) {
 653                 err = dev_set_mtu(dev, nla_get_u32(tb[IFLA_MTU]));
 654                 if (err)
 655                         return err;
 656         }
 657         return 0;
 658 }
 659
 660 static void bareudp_dellink(struct net_device *dev, struct list_head *head)
 661 {
 662         struct bareudp_dev *bareudp = netdev_priv(dev);
 663
 664         list_del(&bareudp->next);
 665         unregister_netdevice_queue(dev, head);
 666 }
 667
 668 static int bareudp_newlink(struct net *net, struct net_device *dev,
 669                            struct nlattr *tb[], struct nlattr *data[],
 670                            struct netlink_ext_ack *extack)
 671 {
 672         struct bareudp_conf conf;
 673         int err;
 674
 675         err = bareudp2info(data, &conf, extack);
 676         if (err)
 677                 return err;
 678
 679         err = bareudp_configure(net, dev, &conf, extack);
 680         if (err)
 681                 return err;
 682
 683         err = bareudp_link_config(dev, tb);
 684         if (err)
 685                 goto err_unconfig;
 686
 687         return 0;
 688
 689 err_unconfig:
 690         bareudp_dellink(dev, NULL);
 691         return err;
 692 }
 693
 694 static size_t bareudp_get_size(const struct net_device *dev)
 695 {
 696         return  nla_total_size(sizeof(__be16)) +  /* IFLA_BAREUDP_PORT */
 697                 nla_total_size(sizeof(__be16)) +  /* IFLA_BAREUDP_ETHERTYPE */
 698                 nla_total_size(sizeof(__u16))  +  /* IFLA_BAREUDP_SRCPORT_MIN */
 699                 nla_total_size(0)              +  /* IFLA_BAREUDP_MULTIPROTO_MODE */
 700                 0;
 701 }
 702
 703 static int bareudp_fill_info(struct sk_buff *skb, const struct net_device *dev)
 704 {
 705         struct bareudp_dev *bareudp = netdev_priv(dev);
 706
 707         if (nla_put_be16(skb, IFLA_BAREUDP_PORT, bareudp->port))
 708                 goto nla_put_failure;
 709         if (nla_put_be16(skb, IFLA_BAREUDP_ETHERTYPE, bareudp->ethertype))
 710                 goto nla_put_failure;
 711         if (nla_put_u16(skb, IFLA_BAREUDP_SRCPORT_MIN, bareudp->sport_min))
 712                 goto nla_put_failure;
 713         if (bareudp->multi_proto_mode &&
 714             nla_put_flag(skb, IFLA_BAREUDP_MULTIPROTO_MODE))
 715                 goto nla_put_failure;
 716
 717         return 0;
 718
 719 nla_put_failure:
 720         return -EMSGSIZE;
 721 }
 722
 723 static struct rtnl_link_ops bareudp_link_ops __read_mostly = {
 724         .kind           = "bareudp",
 725         .maxtype        = IFLA_BAREUDP_MAX,
 726         .policy         = bareudp_policy,
 727         .priv_size      = sizeof(struct bareudp_dev),
 728         .setup          = bareudp_setup,
 729         .validate       = bareudp_validate,
 730         .newlink        = bareudp_newlink,
 731         .dellink        = bareudp_dellink,
 732         .get_size       = bareudp_get_size,
 733         .fill_info      = bareudp_fill_info,
 734 };
 735
 736 static __net_init int bareudp_init_net(struct net *net)
 737 {
 738         struct bareudp_net *bn = net_generic(net, bareudp_net_id);
 739
 740         INIT_LIST_HEAD(&bn->bareudp_list);
 741         return 0;
 742 }
 743
 744 static void bareudp_destroy_tunnels(struct net *net, struct list_head *head)
 745 {
 746         struct bareudp_net *bn = net_generic(net, bareudp_net_id);
 747         struct bareudp_dev *bareudp, *next;
 748
 749         list_for_each_entry_safe(bareudp, next, &bn->bareudp_list, next)
 750                 unregister_netdevice_queue(bareudp->dev, head);
 751 }
 752
 753 static void __net_exit bareudp_exit_batch_net(struct list_head *net_list)
 754 {
 755         struct net *net;
 756         LIST_HEAD(list);
 757
 758         rtnl_lock();
 759         list_for_each_entry(net, net_list, exit_list)
 760                 bareudp_destroy_tunnels(net, &list);
 761
 762         /* unregister the devices gathered above */
 763         unregister_netdevice_many(&list);
 764         rtnl_unlock();
 765 }
 766
 767 static struct pernet_operations bareudp_net_ops = {
 768         .init = bareudp_init_net,
 769         .exit_batch = bareudp_exit_batch_net,
 770         .id   = &bareudp_net_id,
 771         .size = sizeof(struct bareudp_net),
 772 };
 773
 774 static int __init bareudp_init_module(void)
 775 {
 776         int rc;
 777
 778         rc = register_pernet_subsys(&bareudp_net_ops);
 779         if (rc)
 780                 goto out1;
 781
 782         rc = rtnl_link_register(&bareudp_link_ops);
 783         if (rc)
 784                 goto out2;
 785
 786         return 0;
 787 out2:
 788         unregister_pernet_subsys(&bareudp_net_ops);
 789 out1:
 790         return rc;
 791 }
 792 late_initcall(bareudp_init_module);
 793
 794 static void __exit bareudp_cleanup_module(void)
 795 {
 796         rtnl_link_unregister(&bareudp_link_ops);
 797         unregister_pernet_subsys(&bareudp_net_ops);
 798 }
 799 module_exit(bareudp_cleanup_module);
 800
 801 MODULE_ALIAS_RTNL_LINK("bareudp");
 802 MODULE_LICENSE("GPL");
 803 MODULE_AUTHOR("Martin Varghese <martin.varghese@nokia.com>");
 804 MODULE_DESCRIPTION("Interface driver for UDP encapsulated traffic");