net/ipv6/ip6_gre.c

   1 /*
   2  *      GRE over IPv6 protocol decoder.
   3  *
   4  *      Authors: Dmitry Kozlov (xeb@mail.ru)
   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
  13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  14
  15 #include <linux/capability.h>
  16 #include <linux/module.h>
  17 #include <linux/types.h>
  18 #include <linux/kernel.h>
  19 #include <linux/slab.h>
  20 #include <linux/uaccess.h>
  21 #include <linux/skbuff.h>
  22 #include <linux/netdevice.h>
  23 #include <linux/in.h>
  24 #include <linux/tcp.h>
  25 #include <linux/udp.h>
  26 #include <linux/if_arp.h>
  27 #include <linux/init.h>
  28 #include <linux/in6.h>
  29 #include <linux/inetdevice.h>
  30 #include <linux/igmp.h>
  31 #include <linux/netfilter_ipv4.h>
  32 #include <linux/etherdevice.h>
  33 #include <linux/if_ether.h>
  34 #include <linux/hash.h>
  35 #include <linux/if_tunnel.h>
  36 #include <linux/ip6_tunnel.h>
  37
  38 #include <net/sock.h>
  39 #include <net/ip.h>
  40 #include <net/ip_tunnels.h>
  41 #include <net/icmp.h>
  42 #include <net/protocol.h>
  43 #include <net/addrconf.h>
  44 #include <net/arp.h>
  45 #include <net/checksum.h>
  46 #include <net/dsfield.h>
  47 #include <net/inet_ecn.h>
  48 #include <net/xfrm.h>
  49 #include <net/net_namespace.h>
  50 #include <net/netns/generic.h>
  51 #include <net/rtnetlink.h>
  52
  53 #include <net/ipv6.h>
  54 #include <net/ip6_fib.h>
  55 #include <net/ip6_route.h>
  56 #include <net/ip6_tunnel.h>
  57 #include <net/gre.h>
  58
  59
  60 static bool log_ecn_error = true;
  61 module_param(log_ecn_error, bool, 0644);
  62 MODULE_PARM_DESC(log_ecn_error, "Log packets received with corrupted ECN");
  63
  64 #define HASH_SIZE_SHIFT  5
  65 #define HASH_SIZE (1 << HASH_SIZE_SHIFT)
  66
  67 static int ip6gre_net_id __read_mostly;
  68 struct ip6gre_net {
  69         struct ip6_tnl __rcu *tunnels[4][HASH_SIZE];
  70
  71         struct net_device *fb_tunnel_dev;
  72 };
  73
  74 static struct rtnl_link_ops ip6gre_link_ops __read_mostly;
  75 static struct rtnl_link_ops ip6gre_tap_ops __read_mostly;
  76 static int ip6gre_tunnel_init(struct net_device *dev);
  77 static void ip6gre_tunnel_setup(struct net_device *dev);
  78 static void ip6gre_tunnel_link(struct ip6gre_net *ign, struct ip6_tnl *t);
  79 static void ip6gre_tnl_link_config(struct ip6_tnl *t, int set_mtu);
  80
  81 /* Tunnel hash table */
  82
  83 /*
  84    4 hash tables:
  85
  86    3: (remote,local)
  87    2: (remote,*)
  88    1: (*,local)
  89    0: (*,*)
  90
  91    We require exact key match i.e. if a key is present in packet
  92    it will match only tunnel with the same key; if it is not present,
  93    it will match only keyless tunnel.
  94
  95    All keysless packets, if not matched configured keyless tunnels
  96    will match fallback tunnel.
  97  */
  98
  99 #define HASH_KEY(key) (((__force u32)key^((__force u32)key>>4))&(HASH_SIZE - 1))
 100 static u32 HASH_ADDR(const struct in6_addr *addr)
 101 {
 102         u32 hash = ipv6_addr_hash(addr);
 103
 104         return hash_32(hash, HASH_SIZE_SHIFT);
 105 }
 106
 107 #define tunnels_r_l     tunnels[3]
 108 #define tunnels_r       tunnels[2]
 109 #define tunnels_l       tunnels[1]
 110 #define tunnels_wc      tunnels[0]
 111
 112 /* Given src, dst and key, find appropriate for input tunnel. */
 113
 114 static struct ip6_tnl *ip6gre_tunnel_lookup(struct net_device *dev,
 115                 const struct in6_addr *remote, const struct in6_addr *local,
 116                 __be32 key, __be16 gre_proto)
 117 {
 118         struct net *net = dev_net(dev);
 119         int link = dev->ifindex;
 120         unsigned int h0 = HASH_ADDR(remote);
 121         unsigned int h1 = HASH_KEY(key);
 122         struct ip6_tnl *t, *cand = NULL;
 123         struct ip6gre_net *ign = net_generic(net, ip6gre_net_id);
 124         int dev_type = (gre_proto == htons(ETH_P_TEB)) ?
 125                        ARPHRD_ETHER : ARPHRD_IP6GRE;
 126         int score, cand_score = 4;
 127
 128         for_each_ip_tunnel_rcu(t, ign->tunnels_r_l[h0 ^ h1]) {
 129                 if (!ipv6_addr_equal(local, &t->parms.laddr) ||
 130                     !ipv6_addr_equal(remote, &t->parms.raddr) ||
 131                     key != t->parms.i_key ||
 132                     !(t->dev->flags & IFF_UP))
 133                         continue;
 134
 135                 if (t->dev->type != ARPHRD_IP6GRE &&
 136                     t->dev->type != dev_type)
 137                         continue;
 138
 139                 score = 0;
 140                 if (t->parms.link != link)
 141                         score |= 1;
 142                 if (t->dev->type != dev_type)
 143                         score |= 2;
 144                 if (score == 0)
 145                         return t;
 146
 147                 if (score < cand_score) {
 148                         cand = t;
 149                         cand_score = score;
 150                 }
 151         }
 152
 153         for_each_ip_tunnel_rcu(t, ign->tunnels_r[h0 ^ h1]) {
 154                 if (!ipv6_addr_equal(remote, &t->parms.raddr) ||
 155                     key != t->parms.i_key ||
 156                     !(t->dev->flags & IFF_UP))
 157                         continue;
 158
 159                 if (t->dev->type != ARPHRD_IP6GRE &&
 160                     t->dev->type != dev_type)
 161                         continue;
 162
 163                 score = 0;
 164                 if (t->parms.link != link)
 165                         score |= 1;
 166                 if (t->dev->type != dev_type)
 167                         score |= 2;
 168                 if (score == 0)
 169                         return t;
 170
 171                 if (score < cand_score) {
 172                         cand = t;
 173                         cand_score = score;
 174                 }
 175         }
 176
 177         for_each_ip_tunnel_rcu(t, ign->tunnels_l[h1]) {
 178                 if ((!ipv6_addr_equal(local, &t->parms.laddr) &&
 179                           (!ipv6_addr_equal(local, &t->parms.raddr) ||
 180                                  !ipv6_addr_is_multicast(local))) ||
 181                     key != t->parms.i_key ||
 182                     !(t->dev->flags & IFF_UP))
 183                         continue;
 184
 185                 if (t->dev->type != ARPHRD_IP6GRE &&
 186                     t->dev->type != dev_type)
 187                         continue;
 188
 189                 score = 0;
 190                 if (t->parms.link != link)
 191                         score |= 1;
 192                 if (t->dev->type != dev_type)
 193                         score |= 2;
 194                 if (score == 0)
 195                         return t;
 196
 197                 if (score < cand_score) {
 198                         cand = t;
 199                         cand_score = score;
 200                 }
 201         }
 202
 203         for_each_ip_tunnel_rcu(t, ign->tunnels_wc[h1]) {
 204                 if (t->parms.i_key != key ||
 205                     !(t->dev->flags & IFF_UP))
 206                         continue;
 207
 208                 if (t->dev->type != ARPHRD_IP6GRE &&
 209                     t->dev->type != dev_type)
 210                         continue;
 211
 212                 score = 0;
 213                 if (t->parms.link != link)
 214                         score |= 1;
 215                 if (t->dev->type != dev_type)
 216                         score |= 2;
 217                 if (score == 0)
 218                         return t;
 219
 220                 if (score < cand_score) {
 221                         cand = t;
 222                         cand_score = score;
 223                 }
 224         }
 225
 226         if (cand)
 227                 return cand;
 228
 229         dev = ign->fb_tunnel_dev;
 230         if (dev->flags & IFF_UP)
 231                 return netdev_priv(dev);
 232
 233         return NULL;
 234 }
 235
 236 static struct ip6_tnl __rcu **__ip6gre_bucket(struct ip6gre_net *ign,
 237                 const struct __ip6_tnl_parm *p)
 238 {
 239         const struct in6_addr *remote = &p->raddr;
 240         const struct in6_addr *local = &p->laddr;
 241         unsigned int h = HASH_KEY(p->i_key);
 242         int prio = 0;
 243
 244         if (!ipv6_addr_any(local))
 245                 prio |= 1;
 246         if (!ipv6_addr_any(remote) && !ipv6_addr_is_multicast(remote)) {
 247                 prio |= 2;
 248                 h ^= HASH_ADDR(remote);
 249         }
 250
 251         return &ign->tunnels[prio][h];
 252 }
 253
 254 static inline struct ip6_tnl __rcu **ip6gre_bucket(struct ip6gre_net *ign,
 255                 const struct ip6_tnl *t)
 256 {
 257         return __ip6gre_bucket(ign, &t->parms);
 258 }
 259
 260 static void ip6gre_tunnel_link(struct ip6gre_net *ign, struct ip6_tnl *t)
 261 {
 262         struct ip6_tnl __rcu **tp = ip6gre_bucket(ign, t);
 263
 264         rcu_assign_pointer(t->next, rtnl_dereference(*tp));
 265         rcu_assign_pointer(*tp, t);
 266 }
 267
 268 static void ip6gre_tunnel_unlink(struct ip6gre_net *ign, struct ip6_tnl *t)
 269 {
 270         struct ip6_tnl __rcu **tp;
 271         struct ip6_tnl *iter;
 272
 273         for (tp = ip6gre_bucket(ign, t);
 274              (iter = rtnl_dereference(*tp)) != NULL;
 275              tp = &iter->next) {
 276                 if (t == iter) {
 277                         rcu_assign_pointer(*tp, t->next);
 278                         break;
 279                 }
 280         }
 281 }
 282
 283 static struct ip6_tnl *ip6gre_tunnel_find(struct net *net,
 284                                            const struct __ip6_tnl_parm *parms,
 285                                            int type)
 286 {
 287         const struct in6_addr *remote = &parms->raddr;
 288         const struct in6_addr *local = &parms->laddr;
 289         __be32 key = parms->i_key;
 290         int link = parms->link;
 291         struct ip6_tnl *t;
 292         struct ip6_tnl __rcu **tp;
 293         struct ip6gre_net *ign = net_generic(net, ip6gre_net_id);
 294
 295         for (tp = __ip6gre_bucket(ign, parms);
 296              (t = rtnl_dereference(*tp)) != NULL;
 297              tp = &t->next)
 298                 if (ipv6_addr_equal(local, &t->parms.laddr) &&
 299                     ipv6_addr_equal(remote, &t->parms.raddr) &&
 300                     key == t->parms.i_key &&
 301                     link == t->parms.link &&
 302                     type == t->dev->type)
 303                         break;
 304
 305         return t;
 306 }
 307
 308 static struct ip6_tnl *ip6gre_tunnel_locate(struct net *net,
 309                 const struct __ip6_tnl_parm *parms, int create)
 310 {
 311         struct ip6_tnl *t, *nt;
 312         struct net_device *dev;
 313         char name[IFNAMSIZ];
 314         struct ip6gre_net *ign = net_generic(net, ip6gre_net_id);
 315
 316         t = ip6gre_tunnel_find(net, parms, ARPHRD_IP6GRE);
 317         if (t && create)
 318                 return NULL;
 319         if (t || !create)
 320                 return t;
 321
 322         if (parms->name[0])
 323                 strlcpy(name, parms->name, IFNAMSIZ);
 324         else
 325                 strcpy(name, "ip6gre%d");
 326
 327         dev = alloc_netdev(sizeof(*t), name, NET_NAME_UNKNOWN,
 328                            ip6gre_tunnel_setup);
 329         if (!dev)
 330                 return NULL;
 331
 332         dev_net_set(dev, net);
 333
 334         nt = netdev_priv(dev);
 335         nt->parms = *parms;
 336         dev->rtnl_link_ops = &ip6gre_link_ops;
 337
 338         nt->dev = dev;
 339         nt->net = dev_net(dev);
 340         ip6gre_tnl_link_config(nt, 1);
 341
 342         if (register_netdevice(dev) < 0)
 343                 goto failed_free;
 344
 345         /* Can use a lockless transmit, unless we generate output sequences */
 346         if (!(nt->parms.o_flags & TUNNEL_SEQ))
 347                 dev->features |= NETIF_F_LLTX;
 348
 349         dev_hold(dev);
 350         ip6gre_tunnel_link(ign, nt);
 351         return nt;
 352
 353 failed_free:
 354         free_netdev(dev);
 355         return NULL;
 356 }
 357
 358 static void ip6gre_tunnel_uninit(struct net_device *dev)
 359 {
 360         struct ip6_tnl *t = netdev_priv(dev);
 361         struct ip6gre_net *ign = net_generic(t->net, ip6gre_net_id);
 362
 363         ip6gre_tunnel_unlink(ign, t);
 364         dst_cache_reset(&t->dst_cache);
 365         dev_put(dev);
 366 }
 367
 368
 369 static void ip6gre_err(struct sk_buff *skb, struct inet6_skb_parm *opt,
 370                 u8 type, u8 code, int offset, __be32 info)
 371 {
 372         const struct ipv6hdr *ipv6h = (const struct ipv6hdr *)skb->data;
 373         __be16 *p = (__be16 *)(skb->data + offset);
 374         int grehlen = offset + 4;
 375         struct ip6_tnl *t;
 376         __be16 flags;
 377
 378         flags = p[0];
 379         if (flags&(GRE_CSUM|GRE_KEY|GRE_SEQ|GRE_ROUTING|GRE_VERSION)) {
 380                 if (flags&(GRE_VERSION|GRE_ROUTING))
 381                         return;
 382                 if (flags&GRE_KEY) {
 383                         grehlen += 4;
 384                         if (flags&GRE_CSUM)
 385                                 grehlen += 4;
 386                 }
 387         }
 388
 389         /* If only 8 bytes returned, keyed message will be dropped here */
 390         if (!pskb_may_pull(skb, grehlen))
 391                 return;
 392         ipv6h = (const struct ipv6hdr *)skb->data;
 393         p = (__be16 *)(skb->data + offset);
 394
 395         t = ip6gre_tunnel_lookup(skb->dev, &ipv6h->daddr, &ipv6h->saddr,
 396                                 flags & GRE_KEY ?
 397                                 *(((__be32 *)p) + (grehlen / 4) - 1) : 0,
 398                                 p[1]);
 399         if (!t)
 400                 return;
 401
 402         switch (type) {
 403                 __u32 teli;
 404                 struct ipv6_tlv_tnl_enc_lim *tel;
 405                 __u32 mtu;
 406         case ICMPV6_DEST_UNREACH:
 407                 net_dbg_ratelimited("%s: Path to destination invalid or inactive!\n",
 408                                     t->parms.name);
 409                 break;
 410         case ICMPV6_TIME_EXCEED:
 411                 if (code == ICMPV6_EXC_HOPLIMIT) {
 412                         net_dbg_ratelimited("%s: Too small hop limit or routing loop in tunnel!\n",
 413                                             t->parms.name);
 414                 }
 415                 break;
 416         case ICMPV6_PARAMPROB:
 417                 teli = 0;
 418                 if (code == ICMPV6_HDR_FIELD)
 419                         teli = ip6_tnl_parse_tlv_enc_lim(skb, skb->data);
 420
 421                 if (teli && teli == be32_to_cpu(info) - 2) {
 422                         tel = (struct ipv6_tlv_tnl_enc_lim *) &skb->data[teli];
 423                         if (tel->encap_limit == 0) {
 424                                 net_dbg_ratelimited("%s: Too small encapsulation limit or routing loop in tunnel!\n",
 425                                                     t->parms.name);
 426                         }
 427                 } else {
 428                         net_dbg_ratelimited("%s: Recipient unable to parse tunneled packet!\n",
 429                                             t->parms.name);
 430                 }
 431                 break;
 432         case ICMPV6_PKT_TOOBIG:
 433                 mtu = be32_to_cpu(info) - offset;
 434                 if (mtu < IPV6_MIN_MTU)
 435                         mtu = IPV6_MIN_MTU;
 436                 t->dev->mtu = mtu;
 437                 break;
 438         }
 439
 440         if (time_before(jiffies, t->err_time + IP6TUNNEL_ERR_TIMEO))
 441                 t->err_count++;
 442         else
 443                 t->err_count = 1;
 444         t->err_time = jiffies;
 445 }
 446
 447 static int ip6gre_rcv(struct sk_buff *skb, const struct tnl_ptk_info *tpi)
 448 {
 449         const struct ipv6hdr *ipv6h;
 450         struct ip6_tnl *tunnel;
 451
 452         ipv6h = ipv6_hdr(skb);
 453         tunnel = ip6gre_tunnel_lookup(skb->dev,
 454                                       &ipv6h->saddr, &ipv6h->daddr, tpi->key,
 455                                       tpi->proto);
 456         if (tunnel) {
 457                 ip6_tnl_rcv(tunnel, skb, tpi, NULL, false);
 458
 459                 return PACKET_RCVD;
 460         }
 461
 462         return PACKET_REJECT;
 463 }
 464
 465 static int gre_rcv(struct sk_buff *skb)
 466 {
 467         struct tnl_ptk_info tpi;
 468         bool csum_err = false;
 469         int hdr_len;
 470
 471         hdr_len = gre_parse_header(skb, &tpi, &csum_err, htons(ETH_P_IPV6));
 472         if (hdr_len < 0)
 473                 goto drop;
 474
 475         if (iptunnel_pull_header(skb, hdr_len, tpi.proto, false))
 476                 goto drop;
 477
 478         if (ip6gre_rcv(skb, &tpi) == PACKET_RCVD)
 479                 return 0;
 480
 481         icmpv6_send(skb, ICMPV6_DEST_UNREACH, ICMPV6_PORT_UNREACH, 0);
 482 drop:
 483         kfree_skb(skb);
 484         return 0;
 485 }
 486
 487 struct ipv6_tel_txoption {
 488         struct ipv6_txoptions ops;
 489         __u8 dst_opt[8];
 490 };
 491
 492 static int gre_handle_offloads(struct sk_buff *skb, bool csum)
 493 {
 494         return iptunnel_handle_offloads(skb,
 495                                         csum ? SKB_GSO_GRE_CSUM : SKB_GSO_GRE);
 496 }
 497
 498 static netdev_tx_t __gre6_xmit(struct sk_buff *skb,
 499                                struct net_device *dev, __u8 dsfield,
 500                                struct flowi6 *fl6, int encap_limit,
 501                                __u32 *pmtu, __be16 proto)
 502 {
 503         struct ip6_tnl *tunnel = netdev_priv(dev);
 504         __be16 protocol = (dev->type == ARPHRD_ETHER) ?
 505                           htons(ETH_P_TEB) : proto;
 506
 507         if (dev->type == ARPHRD_ETHER)
 508                 IPCB(skb)->flags = 0;
 509
 510         if (dev->header_ops && dev->type == ARPHRD_IP6GRE)
 511                 fl6->daddr = ((struct ipv6hdr *)skb->data)->daddr;
 512         else
 513                 fl6->daddr = tunnel->parms.raddr;
 514
 515         if (tunnel->parms.o_flags & TUNNEL_SEQ)
 516                 tunnel->o_seqno++;
 517
 518         /* Push GRE header. */
 519         gre_build_header(skb, tunnel->tun_hlen, tunnel->parms.o_flags,
 520                          protocol, tunnel->parms.o_key, htonl(tunnel->o_seqno));
 521
 522         skb_set_inner_protocol(skb, protocol);
 523
 524         return ip6_tnl_xmit(skb, dev, dsfield, fl6, encap_limit, pmtu,
 525                             NEXTHDR_GRE);
 526 }
 527
 528 static inline int ip6gre_xmit_ipv4(struct sk_buff *skb, struct net_device *dev)
 529 {
 530         struct ip6_tnl *t = netdev_priv(dev);
 531         const struct iphdr  *iph = ip_hdr(skb);
 532         int encap_limit = -1;
 533         struct flowi6 fl6;
 534         __u8 dsfield;
 535         __u32 mtu;
 536         int err;
 537
 538         memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
 539
 540         if (!(t->parms.flags & IP6_TNL_F_IGN_ENCAP_LIMIT))
 541                 encap_limit = t->parms.encap_limit;
 542
 543         memcpy(&fl6, &t->fl.u.ip6, sizeof(fl6));
 544
 545         dsfield = ipv4_get_dsfield(iph);
 546
 547         if (t->parms.flags & IP6_TNL_F_USE_ORIG_TCLASS)
 548                 fl6.flowlabel |= htonl((__u32)iph->tos << IPV6_TCLASS_SHIFT)
 549                                           & IPV6_TCLASS_MASK;
 550         if (t->parms.flags & IP6_TNL_F_USE_ORIG_FWMARK)
 551                 fl6.flowi6_mark = skb->mark;
 552
 553         err = gre_handle_offloads(skb, !!(t->parms.o_flags & TUNNEL_CSUM));
 554         if (err)
 555                 return -1;
 556
 557         err = __gre6_xmit(skb, dev, dsfield, &fl6, encap_limit, &mtu,
 558                           skb->protocol);
 559         if (err != 0) {
 560                 /* XXX: send ICMP error even if DF is not set. */
 561                 if (err == -EMSGSIZE)
 562                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 563                                   htonl(mtu));
 564                 return -1;
 565         }
 566
 567         return 0;
 568 }
 569
 570 static inline int ip6gre_xmit_ipv6(struct sk_buff *skb, struct net_device *dev)
 571 {
 572         struct ip6_tnl *t = netdev_priv(dev);
 573         struct ipv6hdr *ipv6h = ipv6_hdr(skb);
 574         int encap_limit = -1;
 575         __u16 offset;
 576         struct flowi6 fl6;
 577         __u8 dsfield;
 578         __u32 mtu;
 579         int err;
 580
 581         if (ipv6_addr_equal(&t->parms.raddr, &ipv6h->saddr))
 582                 return -1;
 583
 584         offset = ip6_tnl_parse_tlv_enc_lim(skb, skb_network_header(skb));
 585         if (offset > 0) {
 586                 struct ipv6_tlv_tnl_enc_lim *tel;
 587                 tel = (struct ipv6_tlv_tnl_enc_lim *)&skb_network_header(skb)[offset];
 588                 if (tel->encap_limit == 0) {
 589                         icmpv6_send(skb, ICMPV6_PARAMPROB,
 590                                     ICMPV6_HDR_FIELD, offset + 2);
 591                         return -1;
 592                 }
 593                 encap_limit = tel->encap_limit - 1;
 594         } else if (!(t->parms.flags & IP6_TNL_F_IGN_ENCAP_LIMIT))
 595                 encap_limit = t->parms.encap_limit;
 596
 597         memcpy(&fl6, &t->fl.u.ip6, sizeof(fl6));
 598
 599         dsfield = ipv6_get_dsfield(ipv6h);
 600         if (t->parms.flags & IP6_TNL_F_USE_ORIG_TCLASS)
 601                 fl6.flowlabel |= (*(__be32 *) ipv6h & IPV6_TCLASS_MASK);
 602         if (t->parms.flags & IP6_TNL_F_USE_ORIG_FLOWLABEL)
 603                 fl6.flowlabel |= ip6_flowlabel(ipv6h);
 604         if (t->parms.flags & IP6_TNL_F_USE_ORIG_FWMARK)
 605                 fl6.flowi6_mark = skb->mark;
 606
 607         if (gre_handle_offloads(skb, !!(t->parms.o_flags & TUNNEL_CSUM)))
 608                 return -1;
 609
 610         err = __gre6_xmit(skb, dev, dsfield, &fl6, encap_limit,
 611                           &mtu, skb->protocol);
 612         if (err != 0) {
 613                 if (err == -EMSGSIZE)
 614                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
 615                 return -1;
 616         }
 617
 618         return 0;
 619 }
 620
 621 /**
 622  * ip6_tnl_addr_conflict - compare packet addresses to tunnel's own
 623  *   @t: the outgoing tunnel device
 624  *   @hdr: IPv6 header from the incoming packet
 625  *
 626  * Description:
 627  *   Avoid trivial tunneling loop by checking that tunnel exit-point
 628  *   doesn't match source of incoming packet.
 629  *
 630  * Return:
 631  *   1 if conflict,
 632  *   0 else
 633  **/
 634
 635 static inline bool ip6gre_tnl_addr_conflict(const struct ip6_tnl *t,
 636         const struct ipv6hdr *hdr)
 637 {
 638         return ipv6_addr_equal(&t->parms.raddr, &hdr->saddr);
 639 }
 640
 641 static int ip6gre_xmit_other(struct sk_buff *skb, struct net_device *dev)
 642 {
 643         struct ip6_tnl *t = netdev_priv(dev);
 644         int encap_limit = -1;
 645         struct flowi6 fl6;
 646         __u32 mtu;
 647         int err;
 648
 649         if (!(t->parms.flags & IP6_TNL_F_IGN_ENCAP_LIMIT))
 650                 encap_limit = t->parms.encap_limit;
 651
 652         memcpy(&fl6, &t->fl.u.ip6, sizeof(fl6));
 653         fl6.flowi6_proto = skb->protocol;
 654
 655         err = gre_handle_offloads(skb, !!(t->parms.o_flags & TUNNEL_CSUM));
 656         if (err)
 657                 return err;
 658
 659         err = __gre6_xmit(skb, dev, 0, &fl6, encap_limit, &mtu, skb->protocol);
 660
 661         return err;
 662 }
 663
 664 static netdev_tx_t ip6gre_tunnel_xmit(struct sk_buff *skb,
 665         struct net_device *dev)
 666 {
 667         struct ip6_tnl *t = netdev_priv(dev);
 668         struct net_device_stats *stats = &t->dev->stats;
 669         int ret;
 670
 671         if (!ip6_tnl_xmit_ctl(t, &t->parms.laddr, &t->parms.raddr))
 672                 goto tx_err;
 673
 674         switch (skb->protocol) {
 675         case htons(ETH_P_IP):
 676                 ret = ip6gre_xmit_ipv4(skb, dev);
 677                 break;
 678         case htons(ETH_P_IPV6):
 679                 ret = ip6gre_xmit_ipv6(skb, dev);
 680                 break;
 681         default:
 682                 ret = ip6gre_xmit_other(skb, dev);
 683                 break;
 684         }
 685
 686         if (ret < 0)
 687                 goto tx_err;
 688
 689         return NETDEV_TX_OK;
 690
 691 tx_err:
 692         stats->tx_errors++;
 693         stats->tx_dropped++;
 694         kfree_skb(skb);
 695         return NETDEV_TX_OK;
 696 }
 697
 698 static void ip6gre_tnl_link_config(struct ip6_tnl *t, int set_mtu)
 699 {
 700         struct net_device *dev = t->dev;
 701         struct __ip6_tnl_parm *p = &t->parms;
 702         struct flowi6 *fl6 = &t->fl.u.ip6;
 703         int t_hlen;
 704
 705         if (dev->type != ARPHRD_ETHER) {
 706                 memcpy(dev->dev_addr, &p->laddr, sizeof(struct in6_addr));
 707                 memcpy(dev->broadcast, &p->raddr, sizeof(struct in6_addr));
 708         }
 709
 710         /* Set up flowi template */
 711         fl6->saddr = p->laddr;
 712         fl6->daddr = p->raddr;
 713         fl6->flowi6_oif = p->link;
 714         fl6->flowlabel = 0;
 715         fl6->flowi6_proto = IPPROTO_GRE;
 716
 717         if (!(p->flags&IP6_TNL_F_USE_ORIG_TCLASS))
 718                 fl6->flowlabel |= IPV6_TCLASS_MASK & p->flowinfo;
 719         if (!(p->flags&IP6_TNL_F_USE_ORIG_FLOWLABEL))
 720                 fl6->flowlabel |= IPV6_FLOWLABEL_MASK & p->flowinfo;
 721
 722         p->flags &= ~(IP6_TNL_F_CAP_XMIT|IP6_TNL_F_CAP_RCV|IP6_TNL_F_CAP_PER_PACKET);
 723         p->flags |= ip6_tnl_get_cap(t, &p->laddr, &p->raddr);
 724
 725         if (p->flags&IP6_TNL_F_CAP_XMIT &&
 726                         p->flags&IP6_TNL_F_CAP_RCV && dev->type != ARPHRD_ETHER)
 727                 dev->flags |= IFF_POINTOPOINT;
 728         else
 729                 dev->flags &= ~IFF_POINTOPOINT;
 730
 731         t->tun_hlen = gre_calc_hlen(t->parms.o_flags);
 732
 733         t->hlen = t->encap_hlen + t->tun_hlen;
 734
 735         t_hlen = t->hlen + sizeof(struct ipv6hdr);
 736
 737         if (p->flags & IP6_TNL_F_CAP_XMIT) {
 738                 int strict = (ipv6_addr_type(&p->raddr) &
 739                               (IPV6_ADDR_MULTICAST|IPV6_ADDR_LINKLOCAL));
 740
 741                 struct rt6_info *rt = rt6_lookup(t->net,
 742                                                  &p->raddr, &p->laddr,
 743                                                  p->link, strict);
 744
 745                 if (!rt)
 746                         return;
 747
 748                 if (rt->dst.dev) {
 749                         dev->hard_header_len = rt->dst.dev->hard_header_len +
 750                                                t_hlen;
 751
 752                         if (set_mtu) {
 753                                 dev->mtu = rt->dst.dev->mtu - t_hlen;
 754                                 if (!(t->parms.flags & IP6_TNL_F_IGN_ENCAP_LIMIT))
 755                                         dev->mtu -= 8;
 756                                 if (dev->type == ARPHRD_ETHER)
 757                                         dev->mtu -= ETH_HLEN;
 758
 759                                 if (dev->mtu < IPV6_MIN_MTU)
 760                                         dev->mtu = IPV6_MIN_MTU;
 761                         }
 762                 }
 763                 ip6_rt_put(rt);
 764         }
 765 }
 766
 767 static int ip6gre_tnl_change(struct ip6_tnl *t,
 768         const struct __ip6_tnl_parm *p, int set_mtu)
 769 {
 770         t->parms.laddr = p->laddr;
 771         t->parms.raddr = p->raddr;
 772         t->parms.flags = p->flags;
 773         t->parms.hop_limit = p->hop_limit;
 774         t->parms.encap_limit = p->encap_limit;
 775         t->parms.flowinfo = p->flowinfo;
 776         t->parms.link = p->link;
 777         t->parms.proto = p->proto;
 778         t->parms.i_key = p->i_key;
 779         t->parms.o_key = p->o_key;
 780         t->parms.i_flags = p->i_flags;
 781         t->parms.o_flags = p->o_flags;
 782         dst_cache_reset(&t->dst_cache);
 783         ip6gre_tnl_link_config(t, set_mtu);
 784         return 0;
 785 }
 786
 787 static void ip6gre_tnl_parm_from_user(struct __ip6_tnl_parm *p,
 788         const struct ip6_tnl_parm2 *u)
 789 {
 790         p->laddr = u->laddr;
 791         p->raddr = u->raddr;
 792         p->flags = u->flags;
 793         p->hop_limit = u->hop_limit;
 794         p->encap_limit = u->encap_limit;
 795         p->flowinfo = u->flowinfo;
 796         p->link = u->link;
 797         p->i_key = u->i_key;
 798         p->o_key = u->o_key;
 799         p->i_flags = gre_flags_to_tnl_flags(u->i_flags);
 800         p->o_flags = gre_flags_to_tnl_flags(u->o_flags);
 801         memcpy(p->name, u->name, sizeof(u->name));
 802 }
 803
 804 static void ip6gre_tnl_parm_to_user(struct ip6_tnl_parm2 *u,
 805         const struct __ip6_tnl_parm *p)
 806 {
 807         u->proto = IPPROTO_GRE;
 808         u->laddr = p->laddr;
 809         u->raddr = p->raddr;
 810         u->flags = p->flags;
 811         u->hop_limit = p->hop_limit;
 812         u->encap_limit = p->encap_limit;
 813         u->flowinfo = p->flowinfo;
 814         u->link = p->link;
 815         u->i_key = p->i_key;
 816         u->o_key = p->o_key;
 817         u->i_flags = gre_tnl_flags_to_gre_flags(p->i_flags);
 818         u->o_flags = gre_tnl_flags_to_gre_flags(p->o_flags);
 819         memcpy(u->name, p->name, sizeof(u->name));
 820 }
 821
 822 static int ip6gre_tunnel_ioctl(struct net_device *dev,
 823         struct ifreq *ifr, int cmd)
 824 {
 825         int err = 0;
 826         struct ip6_tnl_parm2 p;
 827         struct __ip6_tnl_parm p1;
 828         struct ip6_tnl *t = netdev_priv(dev);
 829         struct net *net = t->net;
 830         struct ip6gre_net *ign = net_generic(net, ip6gre_net_id);
 831
 832         memset(&p1, 0, sizeof(p1));
 833
 834         switch (cmd) {
 835         case SIOCGETTUNNEL:
 836                 if (dev == ign->fb_tunnel_dev) {
 837                         if (copy_from_user(&p, ifr->ifr_ifru.ifru_data, sizeof(p))) {
 838                                 err = -EFAULT;
 839                                 break;
 840                         }
 841                         ip6gre_tnl_parm_from_user(&p1, &p);
 842                         t = ip6gre_tunnel_locate(net, &p1, 0);
 843                         if (!t)
 844                                 t = netdev_priv(dev);
 845                 }
 846                 memset(&p, 0, sizeof(p));
 847                 ip6gre_tnl_parm_to_user(&p, &t->parms);
 848                 if (copy_to_user(ifr->ifr_ifru.ifru_data, &p, sizeof(p)))
 849                         err = -EFAULT;
 850                 break;
 851
 852         case SIOCADDTUNNEL:
 853         case SIOCCHGTUNNEL:
 854                 err = -EPERM;
 855                 if (!ns_capable(net->user_ns, CAP_NET_ADMIN))
 856                         goto done;
 857
 858                 err = -EFAULT;
 859                 if (copy_from_user(&p, ifr->ifr_ifru.ifru_data, sizeof(p)))
 860                         goto done;
 861
 862                 err = -EINVAL;
 863                 if ((p.i_flags|p.o_flags)&(GRE_VERSION|GRE_ROUTING))
 864                         goto done;
 865
 866                 if (!(p.i_flags&GRE_KEY))
 867                         p.i_key = 0;
 868                 if (!(p.o_flags&GRE_KEY))
 869                         p.o_key = 0;
 870
 871                 ip6gre_tnl_parm_from_user(&p1, &p);
 872                 t = ip6gre_tunnel_locate(net, &p1, cmd == SIOCADDTUNNEL);
 873
 874                 if (dev != ign->fb_tunnel_dev && cmd == SIOCCHGTUNNEL) {
 875                         if (t) {
 876                                 if (t->dev != dev) {
 877                                         err = -EEXIST;
 878                                         break;
 879                                 }
 880                         } else {
 881                                 t = netdev_priv(dev);
 882
 883                                 ip6gre_tunnel_unlink(ign, t);
 884                                 synchronize_net();
 885                                 ip6gre_tnl_change(t, &p1, 1);
 886                                 ip6gre_tunnel_link(ign, t);
 887                                 netdev_state_change(dev);
 888                         }
 889                 }
 890
 891                 if (t) {
 892                         err = 0;
 893
 894                         memset(&p, 0, sizeof(p));
 895                         ip6gre_tnl_parm_to_user(&p, &t->parms);
 896                         if (copy_to_user(ifr->ifr_ifru.ifru_data, &p, sizeof(p)))
 897                                 err = -EFAULT;
 898                 } else
 899                         err = (cmd == SIOCADDTUNNEL ? -ENOBUFS : -ENOENT);
 900                 break;
 901
 902         case SIOCDELTUNNEL:
 903                 err = -EPERM;
 904                 if (!ns_capable(net->user_ns, CAP_NET_ADMIN))
 905                         goto done;
 906
 907                 if (dev == ign->fb_tunnel_dev) {
 908                         err = -EFAULT;
 909                         if (copy_from_user(&p, ifr->ifr_ifru.ifru_data, sizeof(p)))
 910                                 goto done;
 911                         err = -ENOENT;
 912                         ip6gre_tnl_parm_from_user(&p1, &p);
 913                         t = ip6gre_tunnel_locate(net, &p1, 0);
 914                         if (!t)
 915                                 goto done;
 916                         err = -EPERM;
 917                         if (t == netdev_priv(ign->fb_tunnel_dev))
 918                                 goto done;
 919                         dev = t->dev;
 920                 }
 921                 unregister_netdevice(dev);
 922                 err = 0;
 923                 break;
 924
 925         default:
 926                 err = -EINVAL;
 927         }
 928
 929 done:
 930         return err;
 931 }
 932
 933 static int ip6gre_header(struct sk_buff *skb, struct net_device *dev,
 934                         unsigned short type,
 935                         const void *daddr, const void *saddr, unsigned int len)
 936 {
 937         struct ip6_tnl *t = netdev_priv(dev);
 938         struct ipv6hdr *ipv6h = (struct ipv6hdr *)skb_push(skb, t->hlen);
 939         __be16 *p = (__be16 *)(ipv6h+1);
 940
 941         ip6_flow_hdr(ipv6h, 0,
 942                      ip6_make_flowlabel(dev_net(dev), skb,
 943                                         t->fl.u.ip6.flowlabel, true,
 944                                         &t->fl.u.ip6));
 945         ipv6h->hop_limit = t->parms.hop_limit;
 946         ipv6h->nexthdr = NEXTHDR_GRE;
 947         ipv6h->saddr = t->parms.laddr;
 948         ipv6h->daddr = t->parms.raddr;
 949
 950         p[0]            = t->parms.o_flags;
 951         p[1]            = htons(type);
 952
 953         /*
 954          *      Set the source hardware address.
 955          */
 956
 957         if (saddr)
 958                 memcpy(&ipv6h->saddr, saddr, sizeof(struct in6_addr));
 959         if (daddr)
 960                 memcpy(&ipv6h->daddr, daddr, sizeof(struct in6_addr));
 961         if (!ipv6_addr_any(&ipv6h->daddr))
 962                 return t->hlen;
 963
 964         return -t->hlen;
 965 }
 966
 967 static const struct header_ops ip6gre_header_ops = {
 968         .create = ip6gre_header,
 969 };
 970
 971 static const struct net_device_ops ip6gre_netdev_ops = {
 972         .ndo_init               = ip6gre_tunnel_init,
 973         .ndo_uninit             = ip6gre_tunnel_uninit,
 974         .ndo_start_xmit         = ip6gre_tunnel_xmit,
 975         .ndo_do_ioctl           = ip6gre_tunnel_ioctl,
 976         .ndo_change_mtu         = ip6_tnl_change_mtu,
 977         .ndo_get_stats64        = ip_tunnel_get_stats64,
 978         .ndo_get_iflink         = ip6_tnl_get_iflink,
 979 };
 980
 981 static void ip6gre_dev_free(struct net_device *dev)
 982 {
 983         struct ip6_tnl *t = netdev_priv(dev);
 984
 985         dst_cache_destroy(&t->dst_cache);
 986         free_percpu(dev->tstats);
 987         free_netdev(dev);
 988 }
 989
 990 static void ip6gre_tunnel_setup(struct net_device *dev)
 991 {
 992         dev->netdev_ops = &ip6gre_netdev_ops;
 993         dev->destructor = ip6gre_dev_free;
 994
 995         dev->type = ARPHRD_IP6GRE;
 996
 997         dev->flags |= IFF_NOARP;
 998         dev->addr_len = sizeof(struct in6_addr);
 999         netif_keep_dst(dev);
1000 }
1001
1002 static int ip6gre_tunnel_init_common(struct net_device *dev)
1003 {
1004         struct ip6_tnl *tunnel;
1005         int ret;
1006         int t_hlen;
1007
1008         tunnel = netdev_priv(dev);
1009
1010         tunnel->dev = dev;
1011         tunnel->net = dev_net(dev);
1012         strcpy(tunnel->parms.name, dev->name);
1013
1014         dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
1015         if (!dev->tstats)
1016                 return -ENOMEM;
1017
1018         ret = dst_cache_init(&tunnel->dst_cache, GFP_KERNEL);
1019         if (ret) {
1020                 free_percpu(dev->tstats);
1021                 dev->tstats = NULL;
1022                 return ret;
1023         }
1024
1025         tunnel->tun_hlen = gre_calc_hlen(tunnel->parms.o_flags);
1026         tunnel->hlen = tunnel->tun_hlen + tunnel->encap_hlen;
1027         t_hlen = tunnel->hlen + sizeof(struct ipv6hdr);
1028
1029         dev->hard_header_len = LL_MAX_HEADER + t_hlen;
1030         dev->mtu = ETH_DATA_LEN - t_hlen;
1031         if (dev->type == ARPHRD_ETHER)
1032                 dev->mtu -= ETH_HLEN;
1033         if (!(tunnel->parms.flags & IP6_TNL_F_IGN_ENCAP_LIMIT))
1034                 dev->mtu -= 8;
1035
1036         return 0;
1037 }
1038
1039 static int ip6gre_tunnel_init(struct net_device *dev)
1040 {
1041         struct ip6_tnl *tunnel;
1042         int ret;
1043
1044         ret = ip6gre_tunnel_init_common(dev);
1045         if (ret)
1046                 return ret;
1047
1048         tunnel = netdev_priv(dev);
1049
1050         memcpy(dev->dev_addr, &tunnel->parms.laddr, sizeof(struct in6_addr));
1051         memcpy(dev->broadcast, &tunnel->parms.raddr, sizeof(struct in6_addr));
1052
1053         if (ipv6_addr_any(&tunnel->parms.raddr))
1054                 dev->header_ops = &ip6gre_header_ops;
1055
1056         return 0;
1057 }
1058
1059 static void ip6gre_fb_tunnel_init(struct net_device *dev)
1060 {
1061         struct ip6_tnl *tunnel = netdev_priv(dev);
1062
1063         tunnel->dev = dev;
1064         tunnel->net = dev_net(dev);
1065         strcpy(tunnel->parms.name, dev->name);
1066
1067         tunnel->hlen            = sizeof(struct ipv6hdr) + 4;
1068
1069         dev_hold(dev);
1070 }
1071
1072
1073 static struct inet6_protocol ip6gre_protocol __read_mostly = {
1074         .handler     = gre_rcv,
1075         .err_handler = ip6gre_err,
1076         .flags       = INET6_PROTO_NOPOLICY|INET6_PROTO_FINAL,
1077 };
1078
1079 static void ip6gre_destroy_tunnels(struct net *net, struct list_head *head)
1080 {
1081         struct ip6gre_net *ign = net_generic(net, ip6gre_net_id);
1082         struct net_device *dev, *aux;
1083         int prio;
1084
1085         for_each_netdev_safe(net, dev, aux)
1086                 if (dev->rtnl_link_ops == &ip6gre_link_ops ||
1087                     dev->rtnl_link_ops == &ip6gre_tap_ops)
1088                         unregister_netdevice_queue(dev, head);
1089
1090         for (prio = 0; prio < 4; prio++) {
1091                 int h;
1092                 for (h = 0; h < HASH_SIZE; h++) {
1093                         struct ip6_tnl *t;
1094
1095                         t = rtnl_dereference(ign->tunnels[prio][h]);
1096
1097                         while (t) {
1098                                 /* If dev is in the same netns, it has already
1099                                  * been added to the list by the previous loop.
1100                                  */
1101                                 if (!net_eq(dev_net(t->dev), net))
1102                                         unregister_netdevice_queue(t->dev,
1103                                                                    head);
1104                                 t = rtnl_dereference(t->next);
1105                         }
1106                 }
1107         }
1108 }
1109
1110 static int __net_init ip6gre_init_net(struct net *net)
1111 {
1112         struct ip6gre_net *ign = net_generic(net, ip6gre_net_id);
1113         int err;
1114
1115         ign->fb_tunnel_dev = alloc_netdev(sizeof(struct ip6_tnl), "ip6gre0",
1116                                           NET_NAME_UNKNOWN,
1117                                           ip6gre_tunnel_setup);
1118         if (!ign->fb_tunnel_dev) {
1119                 err = -ENOMEM;
1120                 goto err_alloc_dev;
1121         }
1122         dev_net_set(ign->fb_tunnel_dev, net);
1123         /* FB netdevice is special: we have one, and only one per netns.
1124          * Allowing to move it to another netns is clearly unsafe.
1125          */
1126         ign->fb_tunnel_dev->features |= NETIF_F_NETNS_LOCAL;
1127
1128
1129         ip6gre_fb_tunnel_init(ign->fb_tunnel_dev);
1130         ign->fb_tunnel_dev->rtnl_link_ops = &ip6gre_link_ops;
1131
1132         err = register_netdev(ign->fb_tunnel_dev);
1133         if (err)
1134                 goto err_reg_dev;
1135
1136         rcu_assign_pointer(ign->tunnels_wc[0],
1137                            netdev_priv(ign->fb_tunnel_dev));
1138         return 0;
1139
1140 err_reg_dev:
1141         ip6gre_dev_free(ign->fb_tunnel_dev);
1142 err_alloc_dev:
1143         return err;
1144 }
1145
1146 static void __net_exit ip6gre_exit_net(struct net *net)
1147 {
1148         LIST_HEAD(list);
1149
1150         rtnl_lock();
1151         ip6gre_destroy_tunnels(net, &list);
1152         unregister_netdevice_many(&list);
1153         rtnl_unlock();
1154 }
1155
1156 static struct pernet_operations ip6gre_net_ops = {
1157         .init = ip6gre_init_net,
1158         .exit = ip6gre_exit_net,
1159         .id   = &ip6gre_net_id,
1160         .size = sizeof(struct ip6gre_net),
1161 };
1162
1163 static int ip6gre_tunnel_validate(struct nlattr *tb[], struct nlattr *data[])
1164 {
1165         __be16 flags;
1166
1167         if (!data)
1168                 return 0;
1169
1170         flags = 0;
1171         if (data[IFLA_GRE_IFLAGS])
1172                 flags |= nla_get_be16(data[IFLA_GRE_IFLAGS]);
1173         if (data[IFLA_GRE_OFLAGS])
1174                 flags |= nla_get_be16(data[IFLA_GRE_OFLAGS]);
1175         if (flags & (GRE_VERSION|GRE_ROUTING))
1176                 return -EINVAL;
1177
1178         return 0;
1179 }
1180
1181 static int ip6gre_tap_validate(struct nlattr *tb[], struct nlattr *data[])
1182 {
1183         struct in6_addr daddr;
1184
1185         if (tb[IFLA_ADDRESS]) {
1186                 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1187                         return -EINVAL;
1188                 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1189                         return -EADDRNOTAVAIL;
1190         }
1191
1192         if (!data)
1193                 goto out;
1194
1195         if (data[IFLA_GRE_REMOTE]) {
1196                 daddr = nla_get_in6_addr(data[IFLA_GRE_REMOTE]);
1197                 if (ipv6_addr_any(&daddr))
1198                         return -EINVAL;
1199         }
1200
1201 out:
1202         return ip6gre_tunnel_validate(tb, data);
1203 }
1204
1205
1206 static void ip6gre_netlink_parms(struct nlattr *data[],
1207                                 struct __ip6_tnl_parm *parms)
1208 {
1209         memset(parms, 0, sizeof(*parms));
1210
1211         if (!data)
1212                 return;
1213
1214         if (data[IFLA_GRE_LINK])
1215                 parms->link = nla_get_u32(data[IFLA_GRE_LINK]);
1216
1217         if (data[IFLA_GRE_IFLAGS])
1218                 parms->i_flags = gre_flags_to_tnl_flags(
1219                                 nla_get_be16(data[IFLA_GRE_IFLAGS]));
1220
1221         if (data[IFLA_GRE_OFLAGS])
1222                 parms->o_flags = gre_flags_to_tnl_flags(
1223                                 nla_get_be16(data[IFLA_GRE_OFLAGS]));
1224
1225         if (data[IFLA_GRE_IKEY])
1226                 parms->i_key = nla_get_be32(data[IFLA_GRE_IKEY]);
1227
1228         if (data[IFLA_GRE_OKEY])
1229                 parms->o_key = nla_get_be32(data[IFLA_GRE_OKEY]);
1230
1231         if (data[IFLA_GRE_LOCAL])
1232                 parms->laddr = nla_get_in6_addr(data[IFLA_GRE_LOCAL]);
1233
1234         if (data[IFLA_GRE_REMOTE])
1235                 parms->raddr = nla_get_in6_addr(data[IFLA_GRE_REMOTE]);
1236
1237         if (data[IFLA_GRE_TTL])
1238                 parms->hop_limit = nla_get_u8(data[IFLA_GRE_TTL]);
1239
1240         if (data[IFLA_GRE_ENCAP_LIMIT])
1241                 parms->encap_limit = nla_get_u8(data[IFLA_GRE_ENCAP_LIMIT]);
1242
1243         if (data[IFLA_GRE_FLOWINFO])
1244                 parms->flowinfo = nla_get_u32(data[IFLA_GRE_FLOWINFO]);
1245
1246         if (data[IFLA_GRE_FLAGS])
1247                 parms->flags = nla_get_u32(data[IFLA_GRE_FLAGS]);
1248 }
1249
1250 static int ip6gre_tap_init(struct net_device *dev)
1251 {
1252         struct ip6_tnl *tunnel;
1253         int ret;
1254
1255         ret = ip6gre_tunnel_init_common(dev);
1256         if (ret)
1257                 return ret;
1258
1259         dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1260
1261         tunnel = netdev_priv(dev);
1262
1263         ip6gre_tnl_link_config(tunnel, 1);
1264
1265         return 0;
1266 }
1267
1268 static const struct net_device_ops ip6gre_tap_netdev_ops = {
1269         .ndo_init = ip6gre_tap_init,
1270         .ndo_uninit = ip6gre_tunnel_uninit,
1271         .ndo_start_xmit = ip6gre_tunnel_xmit,
1272         .ndo_set_mac_address = eth_mac_addr,
1273         .ndo_validate_addr = eth_validate_addr,
1274         .ndo_change_mtu = ip6_tnl_change_mtu,
1275         .ndo_get_stats64 = ip_tunnel_get_stats64,
1276         .ndo_get_iflink = ip6_tnl_get_iflink,
1277 };
1278
1279 #define GRE6_FEATURES (NETIF_F_SG |             \
1280                        NETIF_F_FRAGLIST |       \
1281                        NETIF_F_HIGHDMA |                \
1282                        NETIF_F_HW_CSUM)
1283
1284 static void ip6gre_tap_setup(struct net_device *dev)
1285 {
1286
1287         ether_setup(dev);
1288
1289         dev->netdev_ops = &ip6gre_tap_netdev_ops;
1290         dev->destructor = ip6gre_dev_free;
1291
1292         dev->features |= NETIF_F_NETNS_LOCAL;
1293         dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1294         dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1295 }
1296
1297 static bool ip6gre_netlink_encap_parms(struct nlattr *data[],
1298                                        struct ip_tunnel_encap *ipencap)
1299 {
1300         bool ret = false;
1301
1302         memset(ipencap, 0, sizeof(*ipencap));
1303
1304         if (!data)
1305                 return ret;
1306
1307         if (data[IFLA_GRE_ENCAP_TYPE]) {
1308                 ret = true;
1309                 ipencap->type = nla_get_u16(data[IFLA_GRE_ENCAP_TYPE]);
1310         }
1311
1312         if (data[IFLA_GRE_ENCAP_FLAGS]) {
1313                 ret = true;
1314                 ipencap->flags = nla_get_u16(data[IFLA_GRE_ENCAP_FLAGS]);
1315         }
1316
1317         if (data[IFLA_GRE_ENCAP_SPORT]) {
1318                 ret = true;
1319                 ipencap->sport = nla_get_be16(data[IFLA_GRE_ENCAP_SPORT]);
1320         }
1321
1322         if (data[IFLA_GRE_ENCAP_DPORT]) {
1323                 ret = true;
1324                 ipencap->dport = nla_get_be16(data[IFLA_GRE_ENCAP_DPORT]);
1325         }
1326
1327         return ret;
1328 }
1329
1330 static int ip6gre_newlink(struct net *src_net, struct net_device *dev,
1331         struct nlattr *tb[], struct nlattr *data[])
1332 {
1333         struct ip6_tnl *nt;
1334         struct net *net = dev_net(dev);
1335         struct ip6gre_net *ign = net_generic(net, ip6gre_net_id);
1336         struct ip_tunnel_encap ipencap;
1337         int err;
1338
1339         nt = netdev_priv(dev);
1340
1341         if (ip6gre_netlink_encap_parms(data, &ipencap)) {
1342                 int err = ip6_tnl_encap_setup(nt, &ipencap);
1343
1344                 if (err < 0)
1345                         return err;
1346         }
1347
1348         ip6gre_netlink_parms(data, &nt->parms);
1349
1350         if (ip6gre_tunnel_find(net, &nt->parms, dev->type))
1351                 return -EEXIST;
1352
1353         if (dev->type == ARPHRD_ETHER && !tb[IFLA_ADDRESS])
1354                 eth_hw_addr_random(dev);
1355
1356         nt->dev = dev;
1357         nt->net = dev_net(dev);
1358         ip6gre_tnl_link_config(nt, !tb[IFLA_MTU]);
1359
1360         dev->features           |= GRE6_FEATURES;
1361         dev->hw_features        |= GRE6_FEATURES;
1362
1363         if (!(nt->parms.o_flags & TUNNEL_SEQ)) {
1364                 /* TCP offload with GRE SEQ is not supported, nor
1365                  * can we support 2 levels of outer headers requiring
1366                  * an update.
1367                  */
1368                 if (!(nt->parms.o_flags & TUNNEL_CSUM) ||
1369                     (nt->encap.type == TUNNEL_ENCAP_NONE)) {
1370                         dev->features    |= NETIF_F_GSO_SOFTWARE;
1371                         dev->hw_features |= NETIF_F_GSO_SOFTWARE;
1372                 }
1373
1374                 /* Can use a lockless transmit, unless we generate
1375                  * output sequences
1376                  */
1377                 dev->features |= NETIF_F_LLTX;
1378         }
1379
1380         err = register_netdevice(dev);
1381         if (err)
1382                 goto out;
1383
1384         dev_hold(dev);
1385         ip6gre_tunnel_link(ign, nt);
1386
1387 out:
1388         return err;
1389 }
1390
1391 static int ip6gre_changelink(struct net_device *dev, struct nlattr *tb[],
1392                             struct nlattr *data[])
1393 {
1394         struct ip6_tnl *t, *nt = netdev_priv(dev);
1395         struct net *net = nt->net;
1396         struct ip6gre_net *ign = net_generic(net, ip6gre_net_id);
1397         struct __ip6_tnl_parm p;
1398         struct ip_tunnel_encap ipencap;
1399
1400         if (dev == ign->fb_tunnel_dev)
1401                 return -EINVAL;
1402
1403         if (ip6gre_netlink_encap_parms(data, &ipencap)) {
1404                 int err = ip6_tnl_encap_setup(nt, &ipencap);
1405
1406                 if (err < 0)
1407                         return err;
1408         }
1409
1410         ip6gre_netlink_parms(data, &p);
1411
1412         t = ip6gre_tunnel_locate(net, &p, 0);
1413
1414         if (t) {
1415                 if (t->dev != dev)
1416                         return -EEXIST;
1417         } else {
1418                 t = nt;
1419         }
1420
1421         ip6gre_tunnel_unlink(ign, t);
1422         ip6gre_tnl_change(t, &p, !tb[IFLA_MTU]);
1423         ip6gre_tunnel_link(ign, t);
1424         return 0;
1425 }
1426
1427 static void ip6gre_dellink(struct net_device *dev, struct list_head *head)
1428 {
1429         struct net *net = dev_net(dev);
1430         struct ip6gre_net *ign = net_generic(net, ip6gre_net_id);
1431
1432         if (dev != ign->fb_tunnel_dev)
1433                 unregister_netdevice_queue(dev, head);
1434 }
1435
1436 static size_t ip6gre_get_size(const struct net_device *dev)
1437 {
1438         return
1439                 /* IFLA_GRE_LINK */
1440                 nla_total_size(4) +
1441                 /* IFLA_GRE_IFLAGS */
1442                 nla_total_size(2) +
1443                 /* IFLA_GRE_OFLAGS */
1444                 nla_total_size(2) +
1445                 /* IFLA_GRE_IKEY */
1446                 nla_total_size(4) +
1447                 /* IFLA_GRE_OKEY */
1448                 nla_total_size(4) +
1449                 /* IFLA_GRE_LOCAL */
1450                 nla_total_size(sizeof(struct in6_addr)) +
1451                 /* IFLA_GRE_REMOTE */
1452                 nla_total_size(sizeof(struct in6_addr)) +
1453                 /* IFLA_GRE_TTL */
1454                 nla_total_size(1) +
1455                 /* IFLA_GRE_ENCAP_LIMIT */
1456                 nla_total_size(1) +
1457                 /* IFLA_GRE_FLOWINFO */
1458                 nla_total_size(4) +
1459                 /* IFLA_GRE_FLAGS */
1460                 nla_total_size(4) +
1461                 /* IFLA_GRE_ENCAP_TYPE */
1462                 nla_total_size(2) +
1463                 /* IFLA_GRE_ENCAP_FLAGS */
1464                 nla_total_size(2) +
1465                 /* IFLA_GRE_ENCAP_SPORT */
1466                 nla_total_size(2) +
1467                 /* IFLA_GRE_ENCAP_DPORT */
1468                 nla_total_size(2) +
1469                 0;
1470 }
1471
1472 static int ip6gre_fill_info(struct sk_buff *skb, const struct net_device *dev)
1473 {
1474         struct ip6_tnl *t = netdev_priv(dev);
1475         struct __ip6_tnl_parm *p = &t->parms;
1476
1477         if (nla_put_u32(skb, IFLA_GRE_LINK, p->link) ||
1478             nla_put_be16(skb, IFLA_GRE_IFLAGS,
1479                          gre_tnl_flags_to_gre_flags(p->i_flags)) ||
1480             nla_put_be16(skb, IFLA_GRE_OFLAGS,
1481                          gre_tnl_flags_to_gre_flags(p->o_flags)) ||
1482             nla_put_be32(skb, IFLA_GRE_IKEY, p->i_key) ||
1483             nla_put_be32(skb, IFLA_GRE_OKEY, p->o_key) ||
1484             nla_put_in6_addr(skb, IFLA_GRE_LOCAL, &p->laddr) ||
1485             nla_put_in6_addr(skb, IFLA_GRE_REMOTE, &p->raddr) ||
1486             nla_put_u8(skb, IFLA_GRE_TTL, p->hop_limit) ||
1487             nla_put_u8(skb, IFLA_GRE_ENCAP_LIMIT, p->encap_limit) ||
1488             nla_put_be32(skb, IFLA_GRE_FLOWINFO, p->flowinfo) ||
1489             nla_put_u32(skb, IFLA_GRE_FLAGS, p->flags))
1490                 goto nla_put_failure;
1491
1492         if (nla_put_u16(skb, IFLA_GRE_ENCAP_TYPE,
1493                         t->encap.type) ||
1494             nla_put_be16(skb, IFLA_GRE_ENCAP_SPORT,
1495                          t->encap.sport) ||
1496             nla_put_be16(skb, IFLA_GRE_ENCAP_DPORT,
1497                          t->encap.dport) ||
1498             nla_put_u16(skb, IFLA_GRE_ENCAP_FLAGS,
1499                         t->encap.flags))
1500                 goto nla_put_failure;
1501
1502         return 0;
1503
1504 nla_put_failure:
1505         return -EMSGSIZE;
1506 }
1507
1508 static const struct nla_policy ip6gre_policy[IFLA_GRE_MAX + 1] = {
1509         [IFLA_GRE_LINK]        = { .type = NLA_U32 },
1510         [IFLA_GRE_IFLAGS]      = { .type = NLA_U16 },
1511         [IFLA_GRE_OFLAGS]      = { .type = NLA_U16 },
1512         [IFLA_GRE_IKEY]        = { .type = NLA_U32 },
1513         [IFLA_GRE_OKEY]        = { .type = NLA_U32 },
1514         [IFLA_GRE_LOCAL]       = { .len = FIELD_SIZEOF(struct ipv6hdr, saddr) },
1515         [IFLA_GRE_REMOTE]      = { .len = FIELD_SIZEOF(struct ipv6hdr, daddr) },
1516         [IFLA_GRE_TTL]         = { .type = NLA_U8 },
1517         [IFLA_GRE_ENCAP_LIMIT] = { .type = NLA_U8 },
1518         [IFLA_GRE_FLOWINFO]    = { .type = NLA_U32 },
1519         [IFLA_GRE_FLAGS]       = { .type = NLA_U32 },
1520         [IFLA_GRE_ENCAP_TYPE]   = { .type = NLA_U16 },
1521         [IFLA_GRE_ENCAP_FLAGS]  = { .type = NLA_U16 },
1522         [IFLA_GRE_ENCAP_SPORT]  = { .type = NLA_U16 },
1523         [IFLA_GRE_ENCAP_DPORT]  = { .type = NLA_U16 },
1524 };
1525
1526 static struct rtnl_link_ops ip6gre_link_ops __read_mostly = {
1527         .kind           = "ip6gre",
1528         .maxtype        = IFLA_GRE_MAX,
1529         .policy         = ip6gre_policy,
1530         .priv_size      = sizeof(struct ip6_tnl),
1531         .setup          = ip6gre_tunnel_setup,
1532         .validate       = ip6gre_tunnel_validate,
1533         .newlink        = ip6gre_newlink,
1534         .changelink     = ip6gre_changelink,
1535         .dellink        = ip6gre_dellink,
1536         .get_size       = ip6gre_get_size,
1537         .fill_info      = ip6gre_fill_info,
1538         .get_link_net   = ip6_tnl_get_link_net,
1539 };
1540
1541 static struct rtnl_link_ops ip6gre_tap_ops __read_mostly = {
1542         .kind           = "ip6gretap",
1543         .maxtype        = IFLA_GRE_MAX,
1544         .policy         = ip6gre_policy,
1545         .priv_size      = sizeof(struct ip6_tnl),
1546         .setup          = ip6gre_tap_setup,
1547         .validate       = ip6gre_tap_validate,
1548         .newlink        = ip6gre_newlink,
1549         .changelink     = ip6gre_changelink,
1550         .get_size       = ip6gre_get_size,
1551         .fill_info      = ip6gre_fill_info,
1552         .get_link_net   = ip6_tnl_get_link_net,
1553 };
1554
1555 /*
1556  *      And now the modules code and kernel interface.
1557  */
1558
1559 static int __init ip6gre_init(void)
1560 {
1561         int err;
1562
1563         pr_info("GRE over IPv6 tunneling driver\n");
1564
1565         err = register_pernet_device(&ip6gre_net_ops);
1566         if (err < 0)
1567                 return err;
1568
1569         err = inet6_add_protocol(&ip6gre_protocol, IPPROTO_GRE);
1570         if (err < 0) {
1571                 pr_info("%s: can't add protocol\n", __func__);
1572                 goto add_proto_failed;
1573         }
1574
1575         err = rtnl_link_register(&ip6gre_link_ops);
1576         if (err < 0)
1577                 goto rtnl_link_failed;
1578
1579         err = rtnl_link_register(&ip6gre_tap_ops);
1580         if (err < 0)
1581                 goto tap_ops_failed;
1582
1583 out:
1584         return err;
1585
1586 tap_ops_failed:
1587         rtnl_link_unregister(&ip6gre_link_ops);
1588 rtnl_link_failed:
1589         inet6_del_protocol(&ip6gre_protocol, IPPROTO_GRE);
1590 add_proto_failed:
1591         unregister_pernet_device(&ip6gre_net_ops);
1592         goto out;
1593 }
1594
1595 static void __exit ip6gre_fini(void)
1596 {
1597         rtnl_link_unregister(&ip6gre_tap_ops);
1598         rtnl_link_unregister(&ip6gre_link_ops);
1599         inet6_del_protocol(&ip6gre_protocol, IPPROTO_GRE);
1600         unregister_pernet_device(&ip6gre_net_ops);
1601 }
1602
1603 module_init(ip6gre_init);
1604 module_exit(ip6gre_fini);
1605 MODULE_LICENSE("GPL");
1606 MODULE_AUTHOR("D. Kozlov (xeb@mail.ru)");
1607 MODULE_DESCRIPTION("GRE over IPv6 tunneling device");
1608 MODULE_ALIAS_RTNL_LINK("ip6gre");
1609 MODULE_ALIAS_RTNL_LINK("ip6gretap");
1610 MODULE_ALIAS_NETDEV("ip6gre0");