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, <[email protected]>
   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 /* Pseudo network device */
  42 struct bareudp_dev {
  43         struct net         *net;        /* netns for packet i/o */
  44         struct net_device  *dev;        /* netdev for bareudp tunnel */
  45         __be16             ethertype;
  46         __be16             port;
  47         u16                sport_min;
  48         bool               multi_proto_mode;
  49         struct socket      __rcu *sock;
  50         struct list_head   next;        /* bareudp node  on namespace list */
  51         struct gro_cells   gro_cells;
  52 };
  53
  54 static int bareudp_udp_encap_recv(struct sock *sk, struct sk_buff *skb)
  55 {
  56         struct metadata_dst *tun_dst = NULL;
  57         struct bareudp_dev *bareudp;
  58         unsigned short family;
  59         unsigned int len;
  60         __be16 proto;
  61         void *oiph;
  62         int err;
  63
  64         bareudp = rcu_dereference_sk_user_data(sk);
  65         if (!bareudp)
  66                 goto drop;
  67
  68         if (skb->protocol ==  htons(ETH_P_IP))
  69                 family = AF_INET;
  70         else
  71                 family = AF_INET6;
  72
  73         if (bareudp->ethertype == htons(ETH_P_IP)) {
  74                 struct iphdr *iphdr;
  75
  76                 iphdr = (struct iphdr *)(skb->data + BAREUDP_BASE_HLEN);
  77                 if (iphdr->version == 4) {
  78                         proto = bareudp->ethertype;
  79                 } else if (bareudp->multi_proto_mode && (iphdr->version == 6)) {
  80                         proto = htons(ETH_P_IPV6);
  81                 } else {
  82                         bareudp->dev->stats.rx_dropped++;
  83                         goto drop;
  84                 }
  85         } else if (bareudp->ethertype == htons(ETH_P_MPLS_UC)) {
  86                 struct iphdr *tunnel_hdr;
  87
  88                 tunnel_hdr = (struct iphdr *)skb_network_header(skb);
  89                 if (tunnel_hdr->version == 4) {
  90                         if (!ipv4_is_multicast(tunnel_hdr->daddr)) {
  91                                 proto = bareudp->ethertype;
  92                         } else if (bareudp->multi_proto_mode &&
  93                                    ipv4_is_multicast(tunnel_hdr->daddr)) {
  94                                 proto = htons(ETH_P_MPLS_MC);
  95                         } else {
  96                                 bareudp->dev->stats.rx_dropped++;
  97                                 goto drop;
  98                         }
  99                 } else {
 100                         int addr_type;
 101                         struct ipv6hdr *tunnel_hdr_v6;
 102
 103                         tunnel_hdr_v6 = (struct ipv6hdr *)skb_network_header(skb);
 104                         addr_type =
 105                         ipv6_addr_type((struct in6_addr *)&tunnel_hdr_v6->daddr);
 106                         if (!(addr_type & IPV6_ADDR_MULTICAST)) {
 107                                 proto = bareudp->ethertype;
 108                         } else if (bareudp->multi_proto_mode &&
 109                                    (addr_type & IPV6_ADDR_MULTICAST)) {
 110                                 proto = htons(ETH_P_MPLS_MC);
 111                         } else {
 112                                 bareudp->dev->stats.rx_dropped++;
 113                                 goto drop;
 114                         }
 115                 }
 116         } else {
 117                 proto = bareudp->ethertype;
 118         }
 119
 120         if (iptunnel_pull_header(skb, BAREUDP_BASE_HLEN,
 121                                  proto,
 122                                  !net_eq(bareudp->net,
 123                                  dev_net(bareudp->dev)))) {
 124                 bareudp->dev->stats.rx_dropped++;
 125                 goto drop;
 126         }
 127         tun_dst = udp_tun_rx_dst(skb, family, TUNNEL_KEY, 0, 0);
 128         if (!tun_dst) {
 129                 bareudp->dev->stats.rx_dropped++;
 130                 goto drop;
 131         }
 132         skb_dst_set(skb, &tun_dst->dst);
 133         skb->dev = bareudp->dev;
 134         oiph = skb_network_header(skb);
 135         skb_reset_network_header(skb);
 136
 137         if (!IS_ENABLED(CONFIG_IPV6) || family == AF_INET)
 138                 err = IP_ECN_decapsulate(oiph, skb);
 139         else
 140                 err = IP6_ECN_decapsulate(oiph, skb);
 141
 142         if (unlikely(err)) {
 143                 if (log_ecn_error) {
 144                         if  (!IS_ENABLED(CONFIG_IPV6) || family == AF_INET)
 145                                 net_info_ratelimited("non-ECT from %pI4 "
 146                                                      "with TOS=%#x\n",
 147                                                      &((struct iphdr *)oiph)->saddr,
 148                                                      ((struct iphdr *)oiph)->tos);
 149                         else
 150                                 net_info_ratelimited("non-ECT from %pI6\n",
 151                                                      &((struct ipv6hdr *)oiph)->saddr);
 152                 }
 153                 if (err > 1) {
 154                         ++bareudp->dev->stats.rx_frame_errors;
 155                         ++bareudp->dev->stats.rx_errors;
 156                         goto drop;
 157                 }
 158         }
 159
 160         len = skb->len;
 161         err = gro_cells_receive(&bareudp->gro_cells, skb);
 162         if (likely(err == NET_RX_SUCCESS))
 163                 dev_sw_netstats_rx_add(bareudp->dev, len);
 164
 165         return 0;
 166 drop:
 167         /* Consume bad packet */
 168         kfree_skb(skb);
 169
 170         return 0;
 171 }
 172
 173 static int bareudp_err_lookup(struct sock *sk, struct sk_buff *skb)
 174 {
 175         return 0;
 176 }
 177
 178 static int bareudp_init(struct net_device *dev)
 179 {
 180         struct bareudp_dev *bareudp = netdev_priv(dev);
 181         int err;
 182
 183         dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
 184         if (!dev->tstats)
 185                 return -ENOMEM;
 186
 187         err = gro_cells_init(&bareudp->gro_cells, dev);
 188         if (err) {
 189                 free_percpu(dev->tstats);
 190                 return err;
 191         }
 192         return 0;
 193 }
 194
 195 static void bareudp_uninit(struct net_device *dev)
 196 {
 197         struct bareudp_dev *bareudp = netdev_priv(dev);
 198
 199         gro_cells_destroy(&bareudp->gro_cells);
 200         free_percpu(dev->tstats);
 201 }
 202
 203 static struct socket *bareudp_create_sock(struct net *net, __be16 port)
 204 {
 205         struct udp_port_cfg udp_conf;
 206         struct socket *sock;
 207         int err;
 208
 209         memset(&udp_conf, 0, sizeof(udp_conf));
 210 #if IS_ENABLED(CONFIG_IPV6)
 211         udp_conf.family = AF_INET6;
 212 #else
 213         udp_conf.family = AF_INET;
 214 #endif
 215         udp_conf.local_udp_port = port;
 216         /* Open UDP socket */
 217         err = udp_sock_create(net, &udp_conf, &sock);
 218         if (err < 0)
 219                 return ERR_PTR(err);
 220
 221         return sock;
 222 }
 223
 224 /* Create new listen socket if needed */
 225 static int bareudp_socket_create(struct bareudp_dev *bareudp, __be16 port)
 226 {
 227         struct udp_tunnel_sock_cfg tunnel_cfg;
 228         struct socket *sock;
 229
 230         sock = bareudp_create_sock(bareudp->net, port);
 231         if (IS_ERR(sock))
 232                 return PTR_ERR(sock);
 233
 234         /* Mark socket as an encapsulation socket */
 235         memset(&tunnel_cfg, 0, sizeof(tunnel_cfg));
 236         tunnel_cfg.sk_user_data = bareudp;
 237         tunnel_cfg.encap_type = 1;
 238         tunnel_cfg.encap_rcv = bareudp_udp_encap_recv;
 239         tunnel_cfg.encap_err_lookup = bareudp_err_lookup;
 240         tunnel_cfg.encap_destroy = NULL;
 241         setup_udp_tunnel_sock(bareudp->net, sock, &tunnel_cfg);
 242
 243         rcu_assign_pointer(bareudp->sock, sock);
 244         return 0;
 245 }
 246
 247 static int bareudp_open(struct net_device *dev)
 248 {
 249         struct bareudp_dev *bareudp = netdev_priv(dev);
 250         int ret = 0;
 251
 252         ret =  bareudp_socket_create(bareudp, bareudp->port);
 253         return ret;
 254 }
 255
 256 static void bareudp_sock_release(struct bareudp_dev *bareudp)
 257 {
 258         struct socket *sock;
 259
 260         sock = bareudp->sock;
 261         rcu_assign_pointer(bareudp->sock, NULL);
 262         synchronize_net();
 263         udp_tunnel_sock_release(sock);
 264 }
 265
 266 static int bareudp_stop(struct net_device *dev)
 267 {
 268         struct bareudp_dev *bareudp = netdev_priv(dev);
 269
 270         bareudp_sock_release(bareudp);
 271         return 0;
 272 }
 273
 274 static int bareudp_xmit_skb(struct sk_buff *skb, struct net_device *dev,
 275                             struct bareudp_dev *bareudp,
 276                             const struct ip_tunnel_info *info)
 277 {
 278         bool xnet = !net_eq(bareudp->net, dev_net(bareudp->dev));
 279         bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
 280         struct socket *sock = rcu_dereference(bareudp->sock);
 281         bool udp_sum = !!(info->key.tun_flags & TUNNEL_CSUM);
 282         const struct ip_tunnel_key *key = &info->key;
 283         struct rtable *rt;
 284         __be16 sport, df;
 285         int min_headroom;
 286         __u8 tos, ttl;
 287         __be32 saddr;
 288         int err;
 289
 290         if (!sock)
 291                 return -ESHUTDOWN;
 292
 293         rt = ip_route_output_tunnel(skb, dev, bareudp->net, &saddr, info,
 294                                     IPPROTO_UDP, use_cache);
 295
 296         if (IS_ERR(rt))
 297                 return PTR_ERR(rt);
 298
 299         skb_tunnel_check_pmtu(skb, &rt->dst,
 300                               BAREUDP_IPV4_HLEN + info->options_len, false);
 301
 302         sport = udp_flow_src_port(bareudp->net, skb,
 303                                   bareudp->sport_min, USHRT_MAX,
 304                                   true);
 305         tos = ip_tunnel_ecn_encap(key->tos, ip_hdr(skb), skb);
 306         ttl = key->ttl;
 307         df = key->tun_flags & TUNNEL_DONT_FRAGMENT ? htons(IP_DF) : 0;
 308         skb_scrub_packet(skb, xnet);
 309
 310         err = -ENOSPC;
 311         if (!skb_pull(skb, skb_network_offset(skb)))
 312                 goto free_dst;
 313
 314         min_headroom = LL_RESERVED_SPACE(rt->dst.dev) + rt->dst.header_len +
 315                 BAREUDP_BASE_HLEN + info->options_len + sizeof(struct iphdr);
 316
 317         err = skb_cow_head(skb, min_headroom);
 318         if (unlikely(err))
 319                 goto free_dst;
 320
 321         err = udp_tunnel_handle_offloads(skb, udp_sum);
 322         if (err)
 323                 goto free_dst;
 324
 325         skb_set_inner_protocol(skb, bareudp->ethertype);
 326         udp_tunnel_xmit_skb(rt, sock->sk, skb, saddr, info->key.u.ipv4.dst,
 327                             tos, ttl, df, sport, bareudp->port,
 328                             !net_eq(bareudp->net, dev_net(bareudp->dev)),
 329                             !(info->key.tun_flags & TUNNEL_CSUM));
 330         return 0;
 331
 332 free_dst:
 333         dst_release(&rt->dst);
 334         return err;
 335 }
 336
 337 static int bareudp6_xmit_skb(struct sk_buff *skb, struct net_device *dev,
 338                              struct bareudp_dev *bareudp,
 339                              const struct ip_tunnel_info *info)
 340 {
 341         bool xnet = !net_eq(bareudp->net, dev_net(bareudp->dev));
 342         bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
 343         struct socket *sock  = rcu_dereference(bareudp->sock);
 344         bool udp_sum = !!(info->key.tun_flags & TUNNEL_CSUM);
 345         const struct ip_tunnel_key *key = &info->key;
 346         struct dst_entry *dst = NULL;
 347         struct in6_addr saddr, daddr;
 348         int min_headroom;
 349         __u8 prio, ttl;
 350         __be16 sport;
 351         int err;
 352
 353         if (!sock)
 354                 return -ESHUTDOWN;
 355
 356         dst = ip6_dst_lookup_tunnel(skb, dev, bareudp->net, sock, &saddr, info,
 357                                     IPPROTO_UDP, use_cache);
 358         if (IS_ERR(dst))
 359                 return PTR_ERR(dst);
 360
 361         skb_tunnel_check_pmtu(skb, dst, BAREUDP_IPV6_HLEN + info->options_len,
 362                               false);
 363
 364         sport = udp_flow_src_port(bareudp->net, skb,
 365                                   bareudp->sport_min, USHRT_MAX,
 366                                   true);
 367         prio = ip_tunnel_ecn_encap(key->tos, ip_hdr(skb), skb);
 368         ttl = key->ttl;
 369
 370         skb_scrub_packet(skb, xnet);
 371
 372         err = -ENOSPC;
 373         if (!skb_pull(skb, skb_network_offset(skb)))
 374                 goto free_dst;
 375
 376         min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len +
 377                 BAREUDP_BASE_HLEN + info->options_len + sizeof(struct ipv6hdr);
 378
 379         err = skb_cow_head(skb, min_headroom);
 380         if (unlikely(err))
 381                 goto free_dst;
 382
 383         err = udp_tunnel_handle_offloads(skb, udp_sum);
 384         if (err)
 385                 goto free_dst;
 386
 387         daddr = info->key.u.ipv6.dst;
 388         udp_tunnel6_xmit_skb(dst, sock->sk, skb, dev,
 389                              &saddr, &daddr, prio, ttl,
 390                              info->key.label, sport, bareudp->port,
 391                              !(info->key.tun_flags & TUNNEL_CSUM));
 392         return 0;
 393
 394 free_dst:
 395         dst_release(dst);
 396         return err;
 397 }
 398
 399 static bool bareudp_proto_valid(struct bareudp_dev *bareudp, __be16 proto)
 400 {
 401         if (bareudp->ethertype == proto)
 402                 return true;
 403
 404         if (!bareudp->multi_proto_mode)
 405                 return false;
 406
 407         if (bareudp->ethertype == htons(ETH_P_MPLS_UC) &&
 408             proto == htons(ETH_P_MPLS_MC))
 409                 return true;
 410
 411         if (bareudp->ethertype == htons(ETH_P_IP) &&
 412             proto == htons(ETH_P_IPV6))
 413                 return true;
 414
 415         return false;
 416 }
 417
 418 static netdev_tx_t bareudp_xmit(struct sk_buff *skb, struct net_device *dev)
 419 {
 420         struct bareudp_dev *bareudp = netdev_priv(dev);
 421         struct ip_tunnel_info *info = NULL;
 422         int err;
 423
 424         if (!bareudp_proto_valid(bareudp, skb->protocol)) {
 425                 err = -EINVAL;
 426                 goto tx_error;
 427         }
 428
 429         info = skb_tunnel_info(skb);
 430         if (unlikely(!info || !(info->mode & IP_TUNNEL_INFO_TX))) {
 431                 err = -EINVAL;
 432                 goto tx_error;
 433         }
 434
 435         rcu_read_lock();
 436         if (IS_ENABLED(CONFIG_IPV6) && info->mode & IP_TUNNEL_INFO_IPV6)
 437                 err = bareudp6_xmit_skb(skb, dev, bareudp, info);
 438         else
 439                 err = bareudp_xmit_skb(skb, dev, bareudp, info);
 440
 441         rcu_read_unlock();
 442
 443         if (likely(!err))
 444                 return NETDEV_TX_OK;
 445 tx_error:
 446         dev_kfree_skb(skb);
 447
 448         if (err == -ELOOP)
 449                 dev->stats.collisions++;
 450         else if (err == -ENETUNREACH)
 451                 dev->stats.tx_carrier_errors++;
 452
 453         dev->stats.tx_errors++;
 454         return NETDEV_TX_OK;
 455 }
 456
 457 static int bareudp_fill_metadata_dst(struct net_device *dev,
 458                                      struct sk_buff *skb)
 459 {
 460         struct ip_tunnel_info *info = skb_tunnel_info(skb);
 461         struct bareudp_dev *bareudp = netdev_priv(dev);
 462         bool use_cache;
 463
 464         use_cache = ip_tunnel_dst_cache_usable(skb, info);
 465
 466         if (!IS_ENABLED(CONFIG_IPV6) || ip_tunnel_info_af(info) == AF_INET) {
 467                 struct rtable *rt;
 468                 __be32 saddr;
 469
 470                 rt = ip_route_output_tunnel(skb, dev, bareudp->net, &saddr,
 471                                             info, IPPROTO_UDP, use_cache);
 472                 if (IS_ERR(rt))
 473                         return PTR_ERR(rt);
 474
 475                 ip_rt_put(rt);
 476                 info->key.u.ipv4.src = saddr;
 477         } else if (ip_tunnel_info_af(info) == AF_INET6) {
 478                 struct dst_entry *dst;
 479                 struct in6_addr saddr;
 480                 struct socket *sock = rcu_dereference(bareudp->sock);
 481
 482                 dst = ip6_dst_lookup_tunnel(skb, dev, bareudp->net, sock,
 483                                             &saddr, info, IPPROTO_UDP,
 484                                             use_cache);
 485                 if (IS_ERR(dst))
 486                         return PTR_ERR(dst);
 487
 488                 dst_release(dst);
 489                 info->key.u.ipv6.src = saddr;
 490         } else {
 491                 return -EINVAL;
 492         }
 493
 494         info->key.tp_src = udp_flow_src_port(bareudp->net, skb,
 495                                              bareudp->sport_min,
 496                         USHRT_MAX, true);
 497         info->key.tp_dst = bareudp->port;
 498         return 0;
 499 }
 500
 501 static const struct net_device_ops bareudp_netdev_ops = {
 502         .ndo_init               = bareudp_init,
 503         .ndo_uninit             = bareudp_uninit,
 504         .ndo_open               = bareudp_open,
 505         .ndo_stop               = bareudp_stop,
 506         .ndo_start_xmit         = bareudp_xmit,
 507         .ndo_get_stats64        = dev_get_tstats64,
 508         .ndo_fill_metadata_dst  = bareudp_fill_metadata_dst,
 509 };
 510
 511 static const struct nla_policy bareudp_policy[IFLA_BAREUDP_MAX + 1] = {
 512         [IFLA_BAREUDP_PORT]                = { .type = NLA_U16 },
 513         [IFLA_BAREUDP_ETHERTYPE]           = { .type = NLA_U16 },
 514         [IFLA_BAREUDP_SRCPORT_MIN]         = { .type = NLA_U16 },
 515         [IFLA_BAREUDP_MULTIPROTO_MODE]     = { .type = NLA_FLAG },
 516 };
 517
 518 /* Info for udev, that this is a virtual tunnel endpoint */
 519 static const struct device_type bareudp_type = {
 520         .name = "bareudp",
 521 };
 522
 523 /* Initialize the device structure. */
 524 static void bareudp_setup(struct net_device *dev)
 525 {
 526         dev->netdev_ops = &bareudp_netdev_ops;
 527         dev->needs_free_netdev = true;
 528         SET_NETDEV_DEVTYPE(dev, &bareudp_type);
 529         dev->features    |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_FRAGLIST;
 530         dev->features    |= NETIF_F_RXCSUM;
 531         dev->features    |= NETIF_F_LLTX;
 532         dev->features    |= NETIF_F_GSO_SOFTWARE;
 533         dev->hw_features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_FRAGLIST;
 534         dev->hw_features |= NETIF_F_RXCSUM;
 535         dev->hw_features |= NETIF_F_GSO_SOFTWARE;
 536         dev->hard_header_len = 0;
 537         dev->addr_len = 0;
 538         dev->mtu = ETH_DATA_LEN;
 539         dev->min_mtu = IPV4_MIN_MTU;
 540         dev->max_mtu = IP_MAX_MTU - BAREUDP_BASE_HLEN;
 541         dev->type = ARPHRD_NONE;
 542         netif_keep_dst(dev);
 543         dev->priv_flags |= IFF_NO_QUEUE;
 544         dev->flags = IFF_POINTOPOINT | IFF_NOARP | IFF_MULTICAST;
 545 }
 546
 547 static int bareudp_validate(struct nlattr *tb[], struct nlattr *data[],
 548                             struct netlink_ext_ack *extack)
 549 {
 550         if (!data) {
 551                 NL_SET_ERR_MSG(extack,
 552                                "Not enough attributes provided to perform the operation");
 553                 return -EINVAL;
 554         }
 555         return 0;
 556 }
 557
 558 static int bareudp2info(struct nlattr *data[], struct bareudp_conf *conf,
 559                         struct netlink_ext_ack *extack)
 560 {
 561         memset(conf, 0, sizeof(*conf));
 562
 563         if (!data[IFLA_BAREUDP_PORT]) {
 564                 NL_SET_ERR_MSG(extack, "port not specified");
 565                 return -EINVAL;
 566         }
 567         if (!data[IFLA_BAREUDP_ETHERTYPE]) {
 568                 NL_SET_ERR_MSG(extack, "ethertype not specified");
 569                 return -EINVAL;
 570         }
 571
 572         if (data[IFLA_BAREUDP_PORT])
 573                 conf->port =  nla_get_u16(data[IFLA_BAREUDP_PORT]);
 574
 575         if (data[IFLA_BAREUDP_ETHERTYPE])
 576                 conf->ethertype =  nla_get_u16(data[IFLA_BAREUDP_ETHERTYPE]);
 577
 578         if (data[IFLA_BAREUDP_SRCPORT_MIN])
 579                 conf->sport_min =  nla_get_u16(data[IFLA_BAREUDP_SRCPORT_MIN]);
 580
 581         if (data[IFLA_BAREUDP_MULTIPROTO_MODE])
 582                 conf->multi_proto_mode = true;
 583
 584         return 0;
 585 }
 586
 587 static struct bareudp_dev *bareudp_find_dev(struct bareudp_net *bn,
 588                                             const struct bareudp_conf *conf)
 589 {
 590         struct bareudp_dev *bareudp, *t = NULL;
 591
 592         list_for_each_entry(bareudp, &bn->bareudp_list, next) {
 593                 if (conf->port == bareudp->port)
 594                         t = bareudp;
 595         }
 596         return t;
 597 }
 598
 599 static int bareudp_configure(struct net *net, struct net_device *dev,
 600                              struct bareudp_conf *conf)
 601 {
 602         struct bareudp_net *bn = net_generic(net, bareudp_net_id);
 603         struct bareudp_dev *t, *bareudp = netdev_priv(dev);
 604         int err;
 605
 606         bareudp->net = net;
 607         bareudp->dev = dev;
 608         t = bareudp_find_dev(bn, conf);
 609         if (t)
 610                 return -EBUSY;
 611
 612         if (conf->multi_proto_mode &&
 613             (conf->ethertype != htons(ETH_P_MPLS_UC) &&
 614              conf->ethertype != htons(ETH_P_IP)))
 615                 return -EINVAL;
 616
 617         bareudp->port = conf->port;
 618         bareudp->ethertype = conf->ethertype;
 619         bareudp->sport_min = conf->sport_min;
 620         bareudp->multi_proto_mode = conf->multi_proto_mode;
 621
 622         err = register_netdevice(dev);
 623         if (err)
 624                 return err;
 625
 626         list_add(&bareudp->next, &bn->bareudp_list);
 627         return 0;
 628 }
 629
 630 static int bareudp_link_config(struct net_device *dev,
 631                                struct nlattr *tb[])
 632 {
 633         int err;
 634
 635         if (tb[IFLA_MTU]) {
 636                 err = dev_set_mtu(dev, nla_get_u32(tb[IFLA_MTU]));
 637                 if (err)
 638                         return err;
 639         }
 640         return 0;
 641 }
 642
 643 static void bareudp_dellink(struct net_device *dev, struct list_head *head)
 644 {
 645         struct bareudp_dev *bareudp = netdev_priv(dev);
 646
 647         list_del(&bareudp->next);
 648         unregister_netdevice_queue(dev, head);
 649 }
 650
 651 static int bareudp_newlink(struct net *net, struct net_device *dev,
 652                            struct nlattr *tb[], struct nlattr *data[],
 653                            struct netlink_ext_ack *extack)
 654 {
 655         struct bareudp_conf conf;
 656         int err;
 657
 658         err = bareudp2info(data, &conf, extack);
 659         if (err)
 660                 return err;
 661
 662         err = bareudp_configure(net, dev, &conf);
 663         if (err)
 664                 return err;
 665
 666         err = bareudp_link_config(dev, tb);
 667         if (err)
 668                 goto err_unconfig;
 669
 670         return 0;
 671
 672 err_unconfig:
 673         bareudp_dellink(dev, NULL);
 674         return err;
 675 }
 676
 677 static size_t bareudp_get_size(const struct net_device *dev)
 678 {
 679         return  nla_total_size(sizeof(__be16)) +  /* IFLA_BAREUDP_PORT */
 680                 nla_total_size(sizeof(__be16)) +  /* IFLA_BAREUDP_ETHERTYPE */
 681                 nla_total_size(sizeof(__u16))  +  /* IFLA_BAREUDP_SRCPORT_MIN */
 682                 nla_total_size(0)              +  /* IFLA_BAREUDP_MULTIPROTO_MODE */
 683                 0;
 684 }
 685
 686 static int bareudp_fill_info(struct sk_buff *skb, const struct net_device *dev)
 687 {
 688         struct bareudp_dev *bareudp = netdev_priv(dev);
 689
 690         if (nla_put_be16(skb, IFLA_BAREUDP_PORT, bareudp->port))
 691                 goto nla_put_failure;
 692         if (nla_put_be16(skb, IFLA_BAREUDP_ETHERTYPE, bareudp->ethertype))
 693                 goto nla_put_failure;
 694         if (nla_put_u16(skb, IFLA_BAREUDP_SRCPORT_MIN, bareudp->sport_min))
 695                 goto nla_put_failure;
 696         if (bareudp->multi_proto_mode &&
 697             nla_put_flag(skb, IFLA_BAREUDP_MULTIPROTO_MODE))
 698                 goto nla_put_failure;
 699
 700         return 0;
 701
 702 nla_put_failure:
 703         return -EMSGSIZE;
 704 }
 705
 706 static struct rtnl_link_ops bareudp_link_ops __read_mostly = {
 707         .kind           = "bareudp",
 708         .maxtype        = IFLA_BAREUDP_MAX,
 709         .policy         = bareudp_policy,
 710         .priv_size      = sizeof(struct bareudp_dev),
 711         .setup          = bareudp_setup,
 712         .validate       = bareudp_validate,
 713         .newlink        = bareudp_newlink,
 714         .dellink        = bareudp_dellink,
 715         .get_size       = bareudp_get_size,
 716         .fill_info      = bareudp_fill_info,
 717 };
 718
 719 struct net_device *bareudp_dev_create(struct net *net, const char *name,
 720                                       u8 name_assign_type,
 721                                       struct bareudp_conf *conf)
 722 {
 723         struct nlattr *tb[IFLA_MAX + 1];
 724         struct net_device *dev;
 725         int err;
 726
 727         memset(tb, 0, sizeof(tb));
 728         dev = rtnl_create_link(net, name, name_assign_type,
 729                                &bareudp_link_ops, tb, NULL);
 730         if (IS_ERR(dev))
 731                 return dev;
 732
 733         err = bareudp_configure(net, dev, conf);
 734         if (err) {
 735                 free_netdev(dev);
 736                 return ERR_PTR(err);
 737         }
 738         err = dev_set_mtu(dev, IP_MAX_MTU - BAREUDP_BASE_HLEN);
 739         if (err)
 740                 goto err;
 741
 742         err = rtnl_configure_link(dev, NULL);
 743         if (err < 0)
 744                 goto err;
 745
 746         return dev;
 747 err:
 748         bareudp_dellink(dev, NULL);
 749         return ERR_PTR(err);
 750 }
 751 EXPORT_SYMBOL_GPL(bareudp_dev_create);
 752
 753 static __net_init int bareudp_init_net(struct net *net)
 754 {
 755         struct bareudp_net *bn = net_generic(net, bareudp_net_id);
 756
 757         INIT_LIST_HEAD(&bn->bareudp_list);
 758         return 0;
 759 }
 760
 761 static void bareudp_destroy_tunnels(struct net *net, struct list_head *head)
 762 {
 763         struct bareudp_net *bn = net_generic(net, bareudp_net_id);
 764         struct bareudp_dev *bareudp, *next;
 765
 766         list_for_each_entry_safe(bareudp, next, &bn->bareudp_list, next)
 767                 unregister_netdevice_queue(bareudp->dev, head);
 768 }
 769
 770 static void __net_exit bareudp_exit_batch_net(struct list_head *net_list)
 771 {
 772         struct net *net;
 773         LIST_HEAD(list);
 774
 775         rtnl_lock();
 776         list_for_each_entry(net, net_list, exit_list)
 777                 bareudp_destroy_tunnels(net, &list);
 778
 779         /* unregister the devices gathered above */
 780         unregister_netdevice_many(&list);
 781         rtnl_unlock();
 782 }
 783
 784 static struct pernet_operations bareudp_net_ops = {
 785         .init = bareudp_init_net,
 786         .exit_batch = bareudp_exit_batch_net,
 787         .id   = &bareudp_net_id,
 788         .size = sizeof(struct bareudp_net),
 789 };
 790
 791 static int __init bareudp_init_module(void)
 792 {
 793         int rc;
 794
 795         rc = register_pernet_subsys(&bareudp_net_ops);
 796         if (rc)
 797                 goto out1;
 798
 799         rc = rtnl_link_register(&bareudp_link_ops);
 800         if (rc)
 801                 goto out2;
 802
 803         return 0;
 804 out2:
 805         unregister_pernet_subsys(&bareudp_net_ops);
 806 out1:
 807         return rc;
 808 }
 809 late_initcall(bareudp_init_module);
 810
 811 static void __exit bareudp_cleanup_module(void)
 812 {
 813         rtnl_link_unregister(&bareudp_link_ops);
 814         unregister_pernet_subsys(&bareudp_net_ops);
 815 }
 816 module_exit(bareudp_cleanup_module);
 817
 818 MODULE_ALIAS_RTNL_LINK("bareudp");
 819 MODULE_LICENSE("GPL");
 820 MODULE_AUTHOR("Martin Varghese <[email protected]>");
 821 MODULE_DESCRIPTION("Interface driver for UDP encapsulated traffic");