datapath/tunnel.c

   1 /*
   2  * Copyright (c) 2010, 2011 Nicira Networks.
   3  * Distributed under the terms of the GNU GPL version 2.
   4  *
   5  * Significant portions of this file may be copied from parts of the Linux
   6  * kernel, by Linus Torvalds and others.
   7  */
   8
   9 #include <linux/if_arp.h>
  10 #include <linux/if_ether.h>
  11 #include <linux/ip.h>
  12 #include <linux/if_vlan.h>
  13 #include <linux/igmp.h>
  14 #include <linux/in.h>
  15 #include <linux/in_route.h>
  16 #include <linux/inetdevice.h>
  17 #include <linux/jhash.h>
  18 #include <linux/list.h>
  19 #include <linux/kernel.h>
  20 #include <linux/version.h>
  21 #include <linux/workqueue.h>
  22 #include <linux/rculist.h>
  23
  24 #include <net/dsfield.h>
  25 #include <net/dst.h>
  26 #include <net/icmp.h>
  27 #include <net/inet_ecn.h>
  28 #include <net/ip.h>
  29 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
  30 #include <net/ipv6.h>
  31 #endif
  32 #include <net/route.h>
  33 #include <net/xfrm.h>
  34
  35 #include "actions.h"
  36 #include "checksum.h"
  37 #include "datapath.h"
  38 #include "tunnel.h"
  39 #include "vlan.h"
  40 #include "vport.h"
  41 #include "vport-generic.h"
  42 #include "vport-internal_dev.h"
  43
  44 #ifdef NEED_CACHE_TIMEOUT
  45 /*
  46  * On kernels where we can't quickly detect changes in the rest of the system
  47  * we use an expiration time to invalidate the cache.  A shorter expiration
  48  * reduces the length of time that we may potentially blackhole packets while
  49  * a longer time increases performance by reducing the frequency that the
  50  * cache needs to be rebuilt.  A variety of factors may cause the cache to be
  51  * invalidated before the expiration time but this is the maximum.  The time
  52  * is expressed in jiffies.
  53  */
  54 #define MAX_CACHE_EXP HZ
  55 #endif
  56
  57 /*
  58  * Interval to check for and remove caches that are no longer valid.  Caches
  59  * are checked for validity before they are used for packet encapsulation and
  60  * old caches are removed at that time.  However, if no packets are sent through
  61  * the tunnel then the cache will never be destroyed.  Since it holds
  62  * references to a number of system objects, the cache will continue to use
  63  * system resources by not allowing those objects to be destroyed.  The cache
  64  * cleaner is periodically run to free invalid caches.  It does not
  65  * significantly affect system performance.  A lower interval will release
  66  * resources faster but will itself consume resources by requiring more frequent
  67  * checks.  A longer interval may result in messages being printed to the kernel
  68  * message buffer about unreleased resources.  The interval is expressed in
  69  * jiffies.
  70  */
  71 #define CACHE_CLEANER_INTERVAL (5 * HZ)
  72
  73 #define CACHE_DATA_ALIGN 16
  74 #define PORT_TABLE_SIZE  1024
  75
  76 static struct hlist_head *port_table __read_mostly;
  77 static int port_table_count;
  78
  79 static void cache_cleaner(struct work_struct *work);
  80 static DECLARE_DELAYED_WORK(cache_cleaner_wq, cache_cleaner);
  81
  82 /*
  83  * These are just used as an optimization: they don't require any kind of
  84  * synchronization because we could have just as easily read the value before
  85  * the port change happened.
  86  */
  87 static unsigned int key_local_remote_ports __read_mostly;
  88 static unsigned int key_remote_ports __read_mostly;
  89 static unsigned int local_remote_ports __read_mostly;
  90 static unsigned int remote_ports __read_mostly;
  91
  92 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,36)
  93 #define rt_dst(rt) (rt->dst)
  94 #else
  95 #define rt_dst(rt) (rt->u.dst)
  96 #endif
  97
  98 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,1,0)
  99 static struct hh_cache *rt_hh(struct rtable *rt)
 100 {
 101         struct neighbour *neigh = dst_get_neighbour(&rt->dst);
 102         if (!neigh || !(neigh->nud_state & NUD_CONNECTED) ||
 103                         !neigh->hh.hh_len)
 104                 return NULL;
 105         return &neigh->hh;
 106 }
 107 #else
 108 #define rt_hh(rt) (rt_dst(rt).hh)
 109 #endif
 110
 111 static struct vport *tnl_vport_to_vport(const struct tnl_vport *tnl_vport)
 112 {
 113         return vport_from_priv(tnl_vport);
 114 }
 115
 116 /* This is analogous to rtnl_dereference for the tunnel cache.  It checks that
 117  * cache_lock is held, so it is only for update side code.
 118  */
 119 static struct tnl_cache *cache_dereference(struct tnl_vport *tnl_vport)
 120 {
 121         return rcu_dereference_protected(tnl_vport->cache,
 122                                  lockdep_is_held(&tnl_vport->cache_lock));
 123 }
 124
 125 static void schedule_cache_cleaner(void)
 126 {
 127         schedule_delayed_work(&cache_cleaner_wq, CACHE_CLEANER_INTERVAL);
 128 }
 129
 130 static void free_cache(struct tnl_cache *cache)
 131 {
 132         if (!cache)
 133                 return;
 134
 135         flow_put(cache->flow);
 136         ip_rt_put(cache->rt);
 137         kfree(cache);
 138 }
 139
 140 static void free_config_rcu(struct rcu_head *rcu)
 141 {
 142         struct tnl_mutable_config *c = container_of(rcu, struct tnl_mutable_config, rcu);
 143         kfree(c);
 144 }
 145
 146 static void free_cache_rcu(struct rcu_head *rcu)
 147 {
 148         struct tnl_cache *c = container_of(rcu, struct tnl_cache, rcu);
 149         free_cache(c);
 150 }
 151
 152 /* Frees the portion of 'mutable' that requires RTNL and thus can't happen
 153  * within an RCU callback.  Fortunately this part doesn't require waiting for
 154  * an RCU grace period.
 155  */
 156 static void free_mutable_rtnl(struct tnl_mutable_config *mutable)
 157 {
 158         ASSERT_RTNL();
 159         if (ipv4_is_multicast(mutable->key.daddr) && mutable->mlink) {
 160                 struct in_device *in_dev;
 161                 in_dev = inetdev_by_index(&init_net, mutable->mlink);
 162                 if (in_dev)
 163                         ip_mc_dec_group(in_dev, mutable->key.daddr);
 164         }
 165 }
 166
 167 static void assign_config_rcu(struct vport *vport,
 168                               struct tnl_mutable_config *new_config)
 169 {
 170         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 171         struct tnl_mutable_config *old_config;
 172
 173         old_config = rtnl_dereference(tnl_vport->mutable);
 174         rcu_assign_pointer(tnl_vport->mutable, new_config);
 175
 176         free_mutable_rtnl(old_config);
 177         call_rcu(&old_config->rcu, free_config_rcu);
 178 }
 179
 180 static void assign_cache_rcu(struct vport *vport, struct tnl_cache *new_cache)
 181 {
 182         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 183         struct tnl_cache *old_cache;
 184
 185         old_cache = cache_dereference(tnl_vport);
 186         rcu_assign_pointer(tnl_vport->cache, new_cache);
 187
 188         if (old_cache)
 189                 call_rcu(&old_cache->rcu, free_cache_rcu);
 190 }
 191
 192 static unsigned int *find_port_pool(const struct tnl_mutable_config *mutable)
 193 {
 194         if (mutable->flags & TNL_F_IN_KEY_MATCH) {
 195                 if (mutable->key.saddr)
 196                         return &local_remote_ports;
 197                 else
 198                         return &remote_ports;
 199         } else {
 200                 if (mutable->key.saddr)
 201                         return &key_local_remote_ports;
 202                 else
 203                         return &key_remote_ports;
 204         }
 205 }
 206
 207 static u32 port_hash(const struct port_lookup_key *key)
 208 {
 209         return jhash2((u32 *)key, (PORT_KEY_LEN / sizeof(u32)), 0);
 210 }
 211
 212 static struct hlist_head *find_bucket(u32 hash)
 213 {
 214         return &port_table[(hash & (PORT_TABLE_SIZE - 1))];
 215 }
 216
 217 static void port_table_add_port(struct vport *vport)
 218 {
 219         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 220         const struct tnl_mutable_config *mutable;
 221         u32 hash;
 222
 223         if (port_table_count == 0)
 224                 schedule_cache_cleaner();
 225
 226         mutable = rtnl_dereference(tnl_vport->mutable);
 227         hash = port_hash(&mutable->key);
 228         hlist_add_head_rcu(&tnl_vport->hash_node, find_bucket(hash));
 229         port_table_count++;
 230
 231         (*find_port_pool(rtnl_dereference(tnl_vport->mutable)))++;
 232 }
 233
 234 static void port_table_move_port(struct vport *vport,
 235                       struct tnl_mutable_config *new_mutable)
 236 {
 237         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 238         u32 hash;
 239
 240         hash = port_hash(&new_mutable->key);
 241         hlist_del_init_rcu(&tnl_vport->hash_node);
 242         hlist_add_head_rcu(&tnl_vport->hash_node, find_bucket(hash));
 243
 244         (*find_port_pool(rtnl_dereference(tnl_vport->mutable)))--;
 245         assign_config_rcu(vport, new_mutable);
 246         (*find_port_pool(rtnl_dereference(tnl_vport->mutable)))++;
 247 }
 248
 249 static void port_table_remove_port(struct vport *vport)
 250 {
 251         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 252
 253         hlist_del_init_rcu(&tnl_vport->hash_node);
 254
 255         port_table_count--;
 256         if (port_table_count == 0)
 257                 cancel_delayed_work_sync(&cache_cleaner_wq);
 258
 259         (*find_port_pool(rtnl_dereference(tnl_vport->mutable)))--;
 260 }
 261
 262 static struct vport *port_table_lookup(struct port_lookup_key *key,
 263                                        const struct tnl_mutable_config **pmutable)
 264 {
 265         struct hlist_node *n;
 266         struct hlist_head *bucket;
 267         u32 hash = port_hash(key);
 268         struct tnl_vport *tnl_vport;
 269
 270         bucket = find_bucket(hash);
 271
 272         hlist_for_each_entry_rcu(tnl_vport, n, bucket, hash_node) {
 273                 struct tnl_mutable_config *mutable;
 274
 275                 mutable = rcu_dereference_rtnl(tnl_vport->mutable);
 276                 if (!memcmp(&mutable->key, key, PORT_KEY_LEN)) {
 277                         *pmutable = mutable;
 278                         return tnl_vport_to_vport(tnl_vport);
 279                 }
 280         }
 281
 282         return NULL;
 283 }
 284
 285 struct vport *tnl_find_port(__be32 saddr, __be32 daddr, __be64 key,
 286                             int tunnel_type,
 287                             const struct tnl_mutable_config **mutable)
 288 {
 289         struct port_lookup_key lookup;
 290         struct vport *vport;
 291
 292         if (ipv4_is_multicast(saddr)) {
 293                 lookup.saddr = 0;
 294                 lookup.daddr = saddr;
 295                 if (key_remote_ports) {
 296                         lookup.tunnel_type = tunnel_type | TNL_T_KEY_EXACT;
 297                         lookup.in_key = key;
 298                         vport = port_table_lookup(&lookup, mutable);
 299                         if (vport)
 300                                 return vport;
 301                 }
 302                 if (remote_ports) {
 303                         lookup.tunnel_type = tunnel_type | TNL_T_KEY_MATCH;
 304                         lookup.in_key = 0;
 305                         vport = port_table_lookup(&lookup, mutable);
 306                         if (vport)
 307                                 return vport;
 308                 }
 309                 return NULL;
 310         }
 311
 312         lookup.saddr = saddr;
 313         lookup.daddr = daddr;
 314
 315         /* First try for exact match on in_key. */
 316         lookup.in_key = key;
 317         lookup.tunnel_type = tunnel_type | TNL_T_KEY_EXACT;
 318         if (key_local_remote_ports) {
 319                 vport = port_table_lookup(&lookup, mutable);
 320                 if (vport)
 321                         return vport;
 322         }
 323         if (key_remote_ports) {
 324                 lookup.saddr = 0;
 325                 vport = port_table_lookup(&lookup, mutable);
 326                 if (vport)
 327                         return vport;
 328
 329                 lookup.saddr = saddr;
 330         }
 331
 332         /* Then try matches that wildcard in_key. */
 333         lookup.in_key = 0;
 334         lookup.tunnel_type = tunnel_type | TNL_T_KEY_MATCH;
 335         if (local_remote_ports) {
 336                 vport = port_table_lookup(&lookup, mutable);
 337                 if (vport)
 338                         return vport;
 339         }
 340         if (remote_ports) {
 341                 lookup.saddr = 0;
 342                 vport = port_table_lookup(&lookup, mutable);
 343                 if (vport)
 344                         return vport;
 345         }
 346
 347         return NULL;
 348 }
 349
 350 static void ecn_decapsulate(struct sk_buff *skb, u8 tos)
 351 {
 352         if (unlikely(INET_ECN_is_ce(tos))) {
 353                 __be16 protocol = skb->protocol;
 354
 355                 skb_set_network_header(skb, ETH_HLEN);
 356
 357                 if (protocol == htons(ETH_P_8021Q)) {
 358                         if (unlikely(!pskb_may_pull(skb, VLAN_ETH_HLEN)))
 359                                 return;
 360
 361                         protocol = vlan_eth_hdr(skb)->h_vlan_encapsulated_proto;
 362                         skb_set_network_header(skb, VLAN_ETH_HLEN);
 363                 }
 364
 365                 if (protocol == htons(ETH_P_IP)) {
 366                         if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
 367                             + sizeof(struct iphdr))))
 368                                 return;
 369
 370                         IP_ECN_set_ce(ip_hdr(skb));
 371                 }
 372 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 373                 else if (protocol == htons(ETH_P_IPV6)) {
 374                         if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
 375                             + sizeof(struct ipv6hdr))))
 376                                 return;
 377
 378                         IP6_ECN_set_ce(ipv6_hdr(skb));
 379                 }
 380 #endif
 381         }
 382 }
 383
 384 /**
 385  *      tnl_rcv - ingress point for generic tunnel code
 386  *
 387  * @vport: port this packet was received on
 388  * @skb: received packet
 389  * @tos: ToS from encapsulating IP packet, used to copy ECN bits
 390  *
 391  * Must be called with rcu_read_lock.
 392  *
 393  * Packets received by this function are in the following state:
 394  * - skb->data points to the inner Ethernet header.
 395  * - The inner Ethernet header is in the linear data area.
 396  * - skb->csum does not include the inner Ethernet header.
 397  * - The layer pointers are undefined.
 398  */
 399 void tnl_rcv(struct vport *vport, struct sk_buff *skb, u8 tos)
 400 {
 401         struct ethhdr *eh;
 402
 403         skb_reset_mac_header(skb);
 404         eh = eth_hdr(skb);
 405
 406         if (likely(ntohs(eh->h_proto) >= 1536))
 407                 skb->protocol = eh->h_proto;
 408         else
 409                 skb->protocol = htons(ETH_P_802_2);
 410
 411         skb_dst_drop(skb);
 412         nf_reset(skb);
 413         skb_clear_rxhash(skb);
 414         secpath_reset(skb);
 415
 416         ecn_decapsulate(skb, tos);
 417         vlan_set_tci(skb, 0);
 418
 419         if (unlikely(compute_ip_summed(skb, false))) {
 420                 kfree_skb(skb);
 421                 return;
 422         }
 423
 424         vport_receive(vport, skb);
 425 }
 426
 427 static bool check_ipv4_address(__be32 addr)
 428 {
 429         if (ipv4_is_multicast(addr) || ipv4_is_lbcast(addr)
 430             || ipv4_is_loopback(addr) || ipv4_is_zeronet(addr))
 431                 return false;
 432
 433         return true;
 434 }
 435
 436 static bool ipv4_should_icmp(struct sk_buff *skb)
 437 {
 438         struct iphdr *old_iph = ip_hdr(skb);
 439
 440         /* Don't respond to L2 broadcast. */
 441         if (is_multicast_ether_addr(eth_hdr(skb)->h_dest))
 442                 return false;
 443
 444         /* Don't respond to L3 broadcast or invalid addresses. */
 445         if (!check_ipv4_address(old_iph->daddr) ||
 446             !check_ipv4_address(old_iph->saddr))
 447                 return false;
 448
 449         /* Only respond to the first fragment. */
 450         if (old_iph->frag_off & htons(IP_OFFSET))
 451                 return false;
 452
 453         /* Don't respond to ICMP error messages. */
 454         if (old_iph->protocol == IPPROTO_ICMP) {
 455                 u8 icmp_type, *icmp_typep;
 456
 457                 icmp_typep = skb_header_pointer(skb, (u8 *)old_iph +
 458                                                 (old_iph->ihl << 2) +
 459                                                 offsetof(struct icmphdr, type) -
 460                                                 skb->data, sizeof(icmp_type),
 461                                                 &icmp_type);
 462
 463                 if (!icmp_typep)
 464                         return false;
 465
 466                 if (*icmp_typep > NR_ICMP_TYPES
 467                         || (*icmp_typep <= ICMP_PARAMETERPROB
 468                                 && *icmp_typep != ICMP_ECHOREPLY
 469                                 && *icmp_typep != ICMP_ECHO))
 470                         return false;
 471         }
 472
 473         return true;
 474 }
 475
 476 static void ipv4_build_icmp(struct sk_buff *skb, struct sk_buff *nskb,
 477                             unsigned int mtu, unsigned int payload_length)
 478 {
 479         struct iphdr *iph, *old_iph = ip_hdr(skb);
 480         struct icmphdr *icmph;
 481         u8 *payload;
 482
 483         iph = (struct iphdr *)skb_put(nskb, sizeof(struct iphdr));
 484         icmph = (struct icmphdr *)skb_put(nskb, sizeof(struct icmphdr));
 485         payload = skb_put(nskb, payload_length);
 486
 487         /* IP */
 488         iph->version            =       4;
 489         iph->ihl                =       sizeof(struct iphdr) >> 2;
 490         iph->tos                =       (old_iph->tos & IPTOS_TOS_MASK) |
 491                                         IPTOS_PREC_INTERNETCONTROL;
 492         iph->tot_len            =       htons(sizeof(struct iphdr)
 493                                               + sizeof(struct icmphdr)
 494                                               + payload_length);
 495         get_random_bytes(&iph->id, sizeof(iph->id));
 496         iph->frag_off           =       0;
 497         iph->ttl                =       IPDEFTTL;
 498         iph->protocol           =       IPPROTO_ICMP;
 499         iph->daddr              =       old_iph->saddr;
 500         iph->saddr              =       old_iph->daddr;
 501
 502         ip_send_check(iph);
 503
 504         /* ICMP */
 505         icmph->type             =       ICMP_DEST_UNREACH;
 506         icmph->code             =       ICMP_FRAG_NEEDED;
 507         icmph->un.gateway       =       htonl(mtu);
 508         icmph->checksum         =       0;
 509
 510         nskb->csum = csum_partial((u8 *)icmph, sizeof(struct icmphdr), 0);
 511         nskb->csum = skb_copy_and_csum_bits(skb, (u8 *)old_iph - skb->data,
 512                                             payload, payload_length,
 513                                             nskb->csum);
 514         icmph->checksum = csum_fold(nskb->csum);
 515 }
 516
 517 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 518 static bool ipv6_should_icmp(struct sk_buff *skb)
 519 {
 520         struct ipv6hdr *old_ipv6h = ipv6_hdr(skb);
 521         int addr_type;
 522         int payload_off = (u8 *)(old_ipv6h + 1) - skb->data;
 523         u8 nexthdr = ipv6_hdr(skb)->nexthdr;
 524
 525         /* Check source address is valid. */
 526         addr_type = ipv6_addr_type(&old_ipv6h->saddr);
 527         if (addr_type & IPV6_ADDR_MULTICAST || addr_type == IPV6_ADDR_ANY)
 528                 return false;
 529
 530         /* Don't reply to unspecified addresses. */
 531         if (ipv6_addr_type(&old_ipv6h->daddr) == IPV6_ADDR_ANY)
 532                 return false;
 533
 534         /* Don't respond to ICMP error messages. */
 535         payload_off = ipv6_skip_exthdr(skb, payload_off, &nexthdr);
 536         if (payload_off < 0)
 537                 return false;
 538
 539         if (nexthdr == NEXTHDR_ICMP) {
 540                 u8 icmp_type, *icmp_typep;
 541
 542                 icmp_typep = skb_header_pointer(skb, payload_off +
 543                                                 offsetof(struct icmp6hdr,
 544                                                         icmp6_type),
 545                                                 sizeof(icmp_type), &icmp_type);
 546
 547                 if (!icmp_typep || !(*icmp_typep & ICMPV6_INFOMSG_MASK))
 548                         return false;
 549         }
 550
 551         return true;
 552 }
 553
 554 static void ipv6_build_icmp(struct sk_buff *skb, struct sk_buff *nskb,
 555                             unsigned int mtu, unsigned int payload_length)
 556 {
 557         struct ipv6hdr *ipv6h, *old_ipv6h = ipv6_hdr(skb);
 558         struct icmp6hdr *icmp6h;
 559         u8 *payload;
 560
 561         ipv6h = (struct ipv6hdr *)skb_put(nskb, sizeof(struct ipv6hdr));
 562         icmp6h = (struct icmp6hdr *)skb_put(nskb, sizeof(struct icmp6hdr));
 563         payload = skb_put(nskb, payload_length);
 564
 565         /* IPv6 */
 566         ipv6h->version          =       6;
 567         ipv6h->priority         =       0;
 568         memset(&ipv6h->flow_lbl, 0, sizeof(ipv6h->flow_lbl));
 569         ipv6h->payload_len      =       htons(sizeof(struct icmp6hdr)
 570                                               + payload_length);
 571         ipv6h->nexthdr          =       NEXTHDR_ICMP;
 572         ipv6h->hop_limit        =       IPV6_DEFAULT_HOPLIMIT;
 573         ipv6_addr_copy(&ipv6h->daddr, &old_ipv6h->saddr);
 574         ipv6_addr_copy(&ipv6h->saddr, &old_ipv6h->daddr);
 575
 576         /* ICMPv6 */
 577         icmp6h->icmp6_type      =       ICMPV6_PKT_TOOBIG;
 578         icmp6h->icmp6_code      =       0;
 579         icmp6h->icmp6_cksum     =       0;
 580         icmp6h->icmp6_mtu       =       htonl(mtu);
 581
 582         nskb->csum = csum_partial((u8 *)icmp6h, sizeof(struct icmp6hdr), 0);
 583         nskb->csum = skb_copy_and_csum_bits(skb, (u8 *)old_ipv6h - skb->data,
 584                                             payload, payload_length,
 585                                             nskb->csum);
 586         icmp6h->icmp6_cksum = csum_ipv6_magic(&ipv6h->saddr, &ipv6h->daddr,
 587                                                 sizeof(struct icmp6hdr)
 588                                                 + payload_length,
 589                                                 ipv6h->nexthdr, nskb->csum);
 590 }
 591 #endif /* IPv6 */
 592
 593 bool tnl_frag_needed(struct vport *vport,
 594                      const struct tnl_mutable_config *mutable,
 595                      struct sk_buff *skb, unsigned int mtu, __be64 flow_key)
 596 {
 597         unsigned int eth_hdr_len = ETH_HLEN;
 598         unsigned int total_length = 0, header_length = 0, payload_length;
 599         struct ethhdr *eh, *old_eh = eth_hdr(skb);
 600         struct sk_buff *nskb;
 601
 602         /* Sanity check */
 603         if (skb->protocol == htons(ETH_P_IP)) {
 604                 if (mtu < IP_MIN_MTU)
 605                         return false;
 606
 607                 if (!ipv4_should_icmp(skb))
 608                         return true;
 609         }
 610 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 611         else if (skb->protocol == htons(ETH_P_IPV6)) {
 612                 if (mtu < IPV6_MIN_MTU)
 613                         return false;
 614
 615                 /*
 616                  * In theory we should do PMTUD on IPv6 multicast messages but
 617                  * we don't have an address to send from so just fragment.
 618                  */
 619                 if (ipv6_addr_type(&ipv6_hdr(skb)->daddr) & IPV6_ADDR_MULTICAST)
 620                         return false;
 621
 622                 if (!ipv6_should_icmp(skb))
 623                         return true;
 624         }
 625 #endif
 626         else
 627                 return false;
 628
 629         /* Allocate */
 630         if (old_eh->h_proto == htons(ETH_P_8021Q))
 631                 eth_hdr_len = VLAN_ETH_HLEN;
 632
 633         payload_length = skb->len - eth_hdr_len;
 634         if (skb->protocol == htons(ETH_P_IP)) {
 635                 header_length = sizeof(struct iphdr) + sizeof(struct icmphdr);
 636                 total_length = min_t(unsigned int, header_length +
 637                                                    payload_length, 576);
 638         }
 639 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 640         else {
 641                 header_length = sizeof(struct ipv6hdr) +
 642                                 sizeof(struct icmp6hdr);
 643                 total_length = min_t(unsigned int, header_length +
 644                                                   payload_length, IPV6_MIN_MTU);
 645         }
 646 #endif
 647
 648         payload_length = total_length - header_length;
 649
 650         nskb = dev_alloc_skb(NET_IP_ALIGN + eth_hdr_len + header_length +
 651                              payload_length);
 652         if (!nskb)
 653                 return false;
 654
 655         skb_reserve(nskb, NET_IP_ALIGN);
 656
 657         /* Ethernet / VLAN */
 658         eh = (struct ethhdr *)skb_put(nskb, eth_hdr_len);
 659         memcpy(eh->h_dest, old_eh->h_source, ETH_ALEN);
 660         memcpy(eh->h_source, mutable->eth_addr, ETH_ALEN);
 661         nskb->protocol = eh->h_proto = old_eh->h_proto;
 662         if (old_eh->h_proto == htons(ETH_P_8021Q)) {
 663                 struct vlan_ethhdr *vh = (struct vlan_ethhdr *)eh;
 664
 665                 vh->h_vlan_TCI = vlan_eth_hdr(skb)->h_vlan_TCI;
 666                 vh->h_vlan_encapsulated_proto = skb->protocol;
 667         } else
 668                 vlan_set_tci(nskb, vlan_get_tci(skb));
 669         skb_reset_mac_header(nskb);
 670
 671         /* Protocol */
 672         if (skb->protocol == htons(ETH_P_IP))
 673                 ipv4_build_icmp(skb, nskb, mtu, payload_length);
 674 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 675         else
 676                 ipv6_build_icmp(skb, nskb, mtu, payload_length);
 677 #endif
 678
 679         /*
 680          * Assume that flow based keys are symmetric with respect to input
 681          * and output and use the key that we were going to put on the
 682          * outgoing packet for the fake received packet.  If the keys are
 683          * not symmetric then PMTUD needs to be disabled since we won't have
 684          * any way of synthesizing packets.
 685          */
 686         if ((mutable->flags & (TNL_F_IN_KEY_MATCH | TNL_F_OUT_KEY_ACTION)) ==
 687             (TNL_F_IN_KEY_MATCH | TNL_F_OUT_KEY_ACTION))
 688                 OVS_CB(nskb)->tun_id = flow_key;
 689
 690         if (unlikely(compute_ip_summed(nskb, false))) {
 691                 kfree_skb(nskb);
 692                 return false;
 693         }
 694
 695         vport_receive(vport, nskb);
 696
 697         return true;
 698 }
 699
 700 static bool check_mtu(struct sk_buff *skb,
 701                       struct vport *vport,
 702                       const struct tnl_mutable_config *mutable,
 703                       const struct rtable *rt, __be16 *frag_offp)
 704 {
 705         bool df_inherit = mutable->flags & TNL_F_DF_INHERIT;
 706         bool pmtud = mutable->flags & TNL_F_PMTUD;
 707         __be16 frag_off = mutable->flags & TNL_F_DF_DEFAULT ? htons(IP_DF) : 0;
 708         int mtu = 0;
 709         unsigned int packet_length = skb->len - ETH_HLEN;
 710
 711         /* Allow for one level of tagging in the packet length. */
 712         if (!vlan_tx_tag_present(skb) &&
 713             eth_hdr(skb)->h_proto == htons(ETH_P_8021Q))
 714                 packet_length -= VLAN_HLEN;
 715
 716         if (pmtud) {
 717                 int vlan_header = 0;
 718
 719                 /* The tag needs to go in packet regardless of where it
 720                  * currently is, so subtract it from the MTU.
 721                  */
 722                 if (vlan_tx_tag_present(skb) ||
 723                     eth_hdr(skb)->h_proto == htons(ETH_P_8021Q))
 724                         vlan_header = VLAN_HLEN;
 725
 726                 mtu = dst_mtu(&rt_dst(rt))
 727                         - ETH_HLEN
 728                         - mutable->tunnel_hlen
 729                         - vlan_header;
 730         }
 731
 732         if (skb->protocol == htons(ETH_P_IP)) {
 733                 struct iphdr *iph = ip_hdr(skb);
 734
 735                 if (df_inherit)
 736                         frag_off = iph->frag_off & htons(IP_DF);
 737
 738                 if (pmtud && iph->frag_off & htons(IP_DF)) {
 739                         mtu = max(mtu, IP_MIN_MTU);
 740
 741                         if (packet_length > mtu &&
 742                             tnl_frag_needed(vport, mutable, skb, mtu,
 743                                             OVS_CB(skb)->tun_id))
 744                                 return false;
 745                 }
 746         }
 747 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 748         else if (skb->protocol == htons(ETH_P_IPV6)) {
 749                 /* IPv6 requires end hosts to do fragmentation
 750                  * if the packet is above the minimum MTU.
 751                  */
 752                 if (df_inherit && packet_length > IPV6_MIN_MTU)
 753                         frag_off = htons(IP_DF);
 754
 755                 if (pmtud) {
 756                         mtu = max(mtu, IPV6_MIN_MTU);
 757
 758                         if (packet_length > mtu &&
 759                             tnl_frag_needed(vport, mutable, skb, mtu,
 760                                             OVS_CB(skb)->tun_id))
 761                                 return false;
 762                 }
 763         }
 764 #endif
 765
 766         *frag_offp = frag_off;
 767         return true;
 768 }
 769
 770 static void create_tunnel_header(const struct vport *vport,
 771                                  const struct tnl_mutable_config *mutable,
 772                                  const struct rtable *rt, void *header)
 773 {
 774         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 775         struct iphdr *iph = header;
 776
 777         iph->version    = 4;
 778         iph->ihl        = sizeof(struct iphdr) >> 2;
 779         iph->frag_off   = htons(IP_DF);
 780         iph->protocol   = tnl_vport->tnl_ops->ipproto;
 781         iph->tos        = mutable->tos;
 782         iph->daddr      = rt->rt_dst;
 783         iph->saddr      = rt->rt_src;
 784         iph->ttl        = mutable->ttl;
 785         if (!iph->ttl)
 786                 iph->ttl = ip4_dst_hoplimit(&rt_dst(rt));
 787
 788         tnl_vport->tnl_ops->build_header(vport, mutable, iph + 1);
 789 }
 790
 791 static void *get_cached_header(const struct tnl_cache *cache)
 792 {
 793         return (void *)cache + ALIGN(sizeof(struct tnl_cache), CACHE_DATA_ALIGN);
 794 }
 795
 796 static bool check_cache_valid(const struct tnl_cache *cache,
 797                               const struct tnl_mutable_config *mutable)
 798 {
 799         struct hh_cache *hh;
 800
 801         if (!cache)
 802                 return false;
 803
 804         hh = rt_hh(cache->rt);
 805         return hh &&
 806 #ifdef NEED_CACHE_TIMEOUT
 807                 time_before(jiffies, cache->expiration) &&
 808 #endif
 809 #ifdef HAVE_RT_GENID
 810                 atomic_read(&init_net.ipv4.rt_genid) == cache->rt->rt_genid &&
 811 #endif
 812 #ifdef HAVE_HH_SEQ
 813                 hh->hh_lock.sequence == cache->hh_seq &&
 814 #endif
 815                 mutable->seq == cache->mutable_seq &&
 816                 (!is_internal_dev(rt_dst(cache->rt).dev) ||
 817                 (cache->flow && !cache->flow->dead));
 818 }
 819
 820 static void __cache_cleaner(struct tnl_vport *tnl_vport)
 821 {
 822         const struct tnl_mutable_config *mutable =
 823                         rcu_dereference(tnl_vport->mutable);
 824         const struct tnl_cache *cache = rcu_dereference(tnl_vport->cache);
 825
 826         if (cache && !check_cache_valid(cache, mutable) &&
 827             spin_trylock_bh(&tnl_vport->cache_lock)) {
 828                 assign_cache_rcu(tnl_vport_to_vport(tnl_vport), NULL);
 829                 spin_unlock_bh(&tnl_vport->cache_lock);
 830         }
 831 }
 832
 833 static void cache_cleaner(struct work_struct *work)
 834 {
 835         int i;
 836
 837         schedule_cache_cleaner();
 838
 839         rcu_read_lock();
 840         for (i = 0; i < PORT_TABLE_SIZE; i++) {
 841                 struct hlist_node *n;
 842                 struct hlist_head *bucket;
 843                 struct tnl_vport  *tnl_vport;
 844
 845                 bucket = &port_table[i];
 846                 hlist_for_each_entry_rcu(tnl_vport, n, bucket, hash_node)
 847                         __cache_cleaner(tnl_vport);
 848         }
 849         rcu_read_unlock();
 850 }
 851
 852 static void create_eth_hdr(struct tnl_cache *cache, struct hh_cache *hh)
 853 {
 854         void *cache_data = get_cached_header(cache);
 855         int hh_off;
 856
 857 #ifdef HAVE_HH_SEQ
 858         unsigned hh_seq;
 859
 860         do {
 861                 hh_seq = read_seqbegin(&hh->hh_lock);
 862                 hh_off = HH_DATA_ALIGN(hh->hh_len) - hh->hh_len;
 863                 memcpy(cache_data, (void *)hh->hh_data + hh_off, hh->hh_len);
 864                 cache->hh_len = hh->hh_len;
 865         } while (read_seqretry(&hh->hh_lock, hh_seq));
 866
 867         cache->hh_seq = hh_seq;
 868 #else
 869         read_lock(&hh->hh_lock);
 870         hh_off = HH_DATA_ALIGN(hh->hh_len) - hh->hh_len;
 871         memcpy(cache_data, (void *)hh->hh_data + hh_off, hh->hh_len);
 872         cache->hh_len = hh->hh_len;
 873         read_unlock(&hh->hh_lock);
 874 #endif
 875 }
 876
 877 static struct tnl_cache *build_cache(struct vport *vport,
 878                                      const struct tnl_mutable_config *mutable,
 879                                      struct rtable *rt)
 880 {
 881         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 882         struct tnl_cache *cache;
 883         void *cache_data;
 884         int cache_len;
 885         struct hh_cache *hh;
 886
 887         if (!(mutable->flags & TNL_F_HDR_CACHE))
 888                 return NULL;
 889
 890         /*
 891          * If there is no entry in the ARP cache or if this device does not
 892          * support hard header caching just fall back to the IP stack.
 893          */
 894
 895         hh = rt_hh(rt);
 896         if (!hh)
 897                 return NULL;
 898
 899         /*
 900          * If lock is contended fall back to directly building the header.
 901          * We're not going to help performance by sitting here spinning.
 902          */
 903         if (!spin_trylock(&tnl_vport->cache_lock))
 904                 return NULL;
 905
 906         cache = cache_dereference(tnl_vport);
 907         if (check_cache_valid(cache, mutable))
 908                 goto unlock;
 909         else
 910                 cache = NULL;
 911
 912         cache_len = LL_RESERVED_SPACE(rt_dst(rt).dev) + mutable->tunnel_hlen;
 913
 914         cache = kzalloc(ALIGN(sizeof(struct tnl_cache), CACHE_DATA_ALIGN) +
 915                         cache_len, GFP_ATOMIC);
 916         if (!cache)
 917                 goto unlock;
 918
 919         create_eth_hdr(cache, hh);
 920         cache_data = get_cached_header(cache) + cache->hh_len;
 921         cache->len = cache->hh_len + mutable->tunnel_hlen;
 922
 923         create_tunnel_header(vport, mutable, rt, cache_data);
 924
 925         cache->mutable_seq = mutable->seq;
 926         cache->rt = rt;
 927 #ifdef NEED_CACHE_TIMEOUT
 928         cache->expiration = jiffies + tnl_vport->cache_exp_interval;
 929 #endif
 930
 931         if (is_internal_dev(rt_dst(rt).dev)) {
 932                 struct sw_flow_key flow_key;
 933                 struct vport *dst_vport;
 934                 struct sk_buff *skb;
 935                 int err;
 936                 int flow_key_len;
 937                 struct sw_flow *flow;
 938
 939                 dst_vport = internal_dev_get_vport(rt_dst(rt).dev);
 940                 if (!dst_vport)
 941                         goto done;
 942
 943                 skb = alloc_skb(cache->len, GFP_ATOMIC);
 944                 if (!skb)
 945                         goto done;
 946
 947                 __skb_put(skb, cache->len);
 948                 memcpy(skb->data, get_cached_header(cache), cache->len);
 949
 950                 err = flow_extract(skb, dst_vport->port_no, &flow_key,
 951                                    &flow_key_len);
 952
 953                 consume_skb(skb);
 954                 if (err)
 955                         goto done;
 956
 957                 flow = flow_tbl_lookup(rcu_dereference(dst_vport->dp->table),
 958                                          &flow_key, flow_key_len);
 959                 if (flow) {
 960                         cache->flow = flow;
 961                         flow_hold(flow);
 962                 }
 963         }
 964
 965 done:
 966         assign_cache_rcu(vport, cache);
 967
 968 unlock:
 969         spin_unlock(&tnl_vport->cache_lock);
 970
 971         return cache;
 972 }
 973
 974 static struct rtable *__find_route(const struct tnl_mutable_config *mutable,
 975                                    u8 ipproto, u8 tos)
 976 {
 977 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,39)
 978         struct flowi fl = { .nl_u = { .ip4_u = {
 979                                         .daddr = mutable->key.daddr,
 980                                         .saddr = mutable->key.saddr,
 981                                         .tos = tos } },
 982                             .proto = ipproto };
 983         struct rtable *rt;
 984
 985         if (unlikely(ip_route_output_key(&init_net, &rt, &fl)))
 986                 return ERR_PTR(-EADDRNOTAVAIL);
 987
 988         return rt;
 989 #else
 990         struct flowi4 fl = { .daddr = mutable->key.daddr,
 991                              .saddr = mutable->key.saddr,
 992                              .flowi4_tos = tos,
 993                              .flowi4_proto = ipproto };
 994
 995         return ip_route_output_key(&init_net, &fl);
 996 #endif
 997 }
 998
 999 static struct rtable *find_route(struct vport *vport,
1000                                  const struct tnl_mutable_config *mutable,
1001                                  u8 tos, struct tnl_cache **cache)
1002 {
1003         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1004         struct tnl_cache *cur_cache = rcu_dereference(tnl_vport->cache);
1005
1006         *cache = NULL;
1007         tos = RT_TOS(tos);
1008
1009         if (likely(tos == mutable->tos &&
1010             check_cache_valid(cur_cache, mutable))) {
1011                 *cache = cur_cache;
1012                 return cur_cache->rt;
1013         } else {
1014                 struct rtable *rt;
1015
1016                 rt = __find_route(mutable, tnl_vport->tnl_ops->ipproto, tos);
1017                 if (IS_ERR(rt))
1018                         return NULL;
1019
1020                 if (likely(tos == mutable->tos))
1021                         *cache = build_cache(vport, mutable, rt);
1022
1023                 return rt;
1024         }
1025 }
1026
1027 static bool need_linearize(const struct sk_buff *skb)
1028 {
1029         int i;
1030
1031         if (unlikely(skb_shinfo(skb)->frag_list))
1032                 return true;
1033
1034         /*
1035          * Generally speaking we should linearize if there are paged frags.
1036          * However, if all of the refcounts are 1 we know nobody else can
1037          * change them from underneath us and we can skip the linearization.
1038          */
1039         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
1040                 if (unlikely(page_count(skb_shinfo(skb)->frags[i].page) > 1))
1041                         return true;
1042
1043         return false;
1044 }
1045
1046 static struct sk_buff *handle_offloads(struct sk_buff *skb,
1047                                        const struct tnl_mutable_config *mutable,
1048                                        const struct rtable *rt)
1049 {
1050         int min_headroom;
1051         int err;
1052
1053         min_headroom = LL_RESERVED_SPACE(rt_dst(rt).dev) + rt_dst(rt).header_len
1054                         + mutable->tunnel_hlen
1055                         + (vlan_tx_tag_present(skb) ? VLAN_HLEN : 0);
1056
1057         if (skb_headroom(skb) < min_headroom || skb_header_cloned(skb)) {
1058                 int head_delta = SKB_DATA_ALIGN(min_headroom -
1059                                                 skb_headroom(skb) +
1060                                                 16);
1061                 err = pskb_expand_head(skb, max_t(int, head_delta, 0),
1062                                         0, GFP_ATOMIC);
1063                 if (unlikely(err))
1064                         goto error_free;
1065         }
1066
1067         forward_ip_summed(skb, true);
1068
1069         if (skb_is_gso(skb)) {
1070                 struct sk_buff *nskb;
1071
1072                 nskb = skb_gso_segment(skb, 0);
1073                 if (IS_ERR(nskb)) {
1074                         kfree_skb(skb);
1075                         err = PTR_ERR(nskb);
1076                         goto error;
1077                 }
1078
1079                 consume_skb(skb);
1080                 skb = nskb;
1081         } else if (get_ip_summed(skb) == OVS_CSUM_PARTIAL) {
1082                 /* Pages aren't locked and could change at any time.
1083                  * If this happens after we compute the checksum, the
1084                  * checksum will be wrong.  We linearize now to avoid
1085                  * this problem.
1086                  */
1087                 if (unlikely(need_linearize(skb))) {
1088                         err = __skb_linearize(skb);
1089                         if (unlikely(err))
1090                                 goto error_free;
1091                 }
1092
1093                 err = skb_checksum_help(skb);
1094                 if (unlikely(err))
1095                         goto error_free;
1096         }
1097
1098         set_ip_summed(skb, OVS_CSUM_NONE);
1099
1100         return skb;
1101
1102 error_free:
1103         kfree_skb(skb);
1104 error:
1105         return ERR_PTR(err);
1106 }
1107
1108 static int send_frags(struct sk_buff *skb,
1109                       const struct tnl_mutable_config *mutable)
1110 {
1111         int sent_len;
1112
1113         sent_len = 0;
1114         while (skb) {
1115                 struct sk_buff *next = skb->next;
1116                 int frag_len = skb->len - mutable->tunnel_hlen;
1117                 int err;
1118
1119                 skb->next = NULL;
1120                 memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
1121
1122                 err = ip_local_out(skb);
1123                 skb = next;
1124                 if (unlikely(net_xmit_eval(err)))
1125                         goto free_frags;
1126                 sent_len += frag_len;
1127         }
1128
1129         return sent_len;
1130
1131 free_frags:
1132         /*
1133          * There's no point in continuing to send fragments once one has been
1134          * dropped so just free the rest.  This may help improve the congestion
1135          * that caused the first packet to be dropped.
1136          */
1137         tnl_free_linked_skbs(skb);
1138         return sent_len;
1139 }
1140
1141 int tnl_send(struct vport *vport, struct sk_buff *skb)
1142 {
1143         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1144         const struct tnl_mutable_config *mutable = rcu_dereference(tnl_vport->mutable);
1145
1146         enum vport_err_type err = VPORT_E_TX_ERROR;
1147         struct rtable *rt;
1148         struct dst_entry *unattached_dst = NULL;
1149         struct tnl_cache *cache;
1150         int sent_len = 0;
1151         __be16 frag_off = 0;
1152         u8 ttl;
1153         u8 inner_tos;
1154         u8 tos;
1155
1156         /* Validate the protocol headers before we try to use them. */
1157         if (skb->protocol == htons(ETH_P_8021Q) &&
1158             !vlan_tx_tag_present(skb)) {
1159                 if (unlikely(!pskb_may_pull(skb, VLAN_ETH_HLEN)))
1160                         goto error_free;
1161
1162                 skb->protocol = vlan_eth_hdr(skb)->h_vlan_encapsulated_proto;
1163                 skb_set_network_header(skb, VLAN_ETH_HLEN);
1164         }
1165
1166         if (skb->protocol == htons(ETH_P_IP)) {
1167                 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
1168                     + sizeof(struct iphdr))))
1169                         skb->protocol = 0;
1170         }
1171 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1172         else if (skb->protocol == htons(ETH_P_IPV6)) {
1173                 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
1174                     + sizeof(struct ipv6hdr))))
1175                         skb->protocol = 0;
1176         }
1177 #endif
1178
1179         /* ToS */
1180         if (skb->protocol == htons(ETH_P_IP))
1181                 inner_tos = ip_hdr(skb)->tos;
1182 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1183         else if (skb->protocol == htons(ETH_P_IPV6))
1184                 inner_tos = ipv6_get_dsfield(ipv6_hdr(skb));
1185 #endif
1186         else
1187                 inner_tos = 0;
1188
1189         if (mutable->flags & TNL_F_TOS_INHERIT)
1190                 tos = inner_tos;
1191         else
1192                 tos = mutable->tos;
1193
1194         tos = INET_ECN_encapsulate(tos, inner_tos);
1195
1196         /* Route lookup */
1197         rt = find_route(vport, mutable, tos, &cache);
1198         if (unlikely(!rt))
1199                 goto error_free;
1200         if (unlikely(!cache))
1201                 unattached_dst = &rt_dst(rt);
1202
1203         /* Reset SKB */
1204         nf_reset(skb);
1205         secpath_reset(skb);
1206         skb_dst_drop(skb);
1207         skb_clear_rxhash(skb);
1208
1209         /* Offloading */
1210         skb = handle_offloads(skb, mutable, rt);
1211         if (IS_ERR(skb))
1212                 goto error;
1213
1214         /* MTU */
1215         if (unlikely(!check_mtu(skb, vport, mutable, rt, &frag_off))) {
1216                 err = VPORT_E_TX_DROPPED;
1217                 goto error_free;
1218         }
1219
1220         /*
1221          * If we are over the MTU, allow the IP stack to handle fragmentation.
1222          * Fragmentation is a slow path anyways.
1223          */
1224         if (unlikely(skb->len + mutable->tunnel_hlen > dst_mtu(&rt_dst(rt)) &&
1225                      cache)) {
1226                 unattached_dst = &rt_dst(rt);
1227                 dst_hold(unattached_dst);
1228                 cache = NULL;
1229         }
1230
1231         /* TTL */
1232         ttl = mutable->ttl;
1233         if (!ttl)
1234                 ttl = ip4_dst_hoplimit(&rt_dst(rt));
1235
1236         if (mutable->flags & TNL_F_TTL_INHERIT) {
1237                 if (skb->protocol == htons(ETH_P_IP))
1238                         ttl = ip_hdr(skb)->ttl;
1239 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1240                 else if (skb->protocol == htons(ETH_P_IPV6))
1241                         ttl = ipv6_hdr(skb)->hop_limit;
1242 #endif
1243         }
1244
1245         while (skb) {
1246                 struct iphdr *iph;
1247                 struct sk_buff *next_skb = skb->next;
1248                 skb->next = NULL;
1249
1250                 if (unlikely(vlan_deaccel_tag(skb)))
1251                         goto next;
1252
1253                 if (likely(cache)) {
1254                         skb_push(skb, cache->len);
1255                         memcpy(skb->data, get_cached_header(cache), cache->len);
1256                         skb_reset_mac_header(skb);
1257                         skb_set_network_header(skb, cache->hh_len);
1258
1259                 } else {
1260                         skb_push(skb, mutable->tunnel_hlen);
1261                         create_tunnel_header(vport, mutable, rt, skb->data);
1262                         skb_reset_network_header(skb);
1263
1264                         if (next_skb)
1265                                 skb_dst_set(skb, dst_clone(unattached_dst));
1266                         else {
1267                                 skb_dst_set(skb, unattached_dst);
1268                                 unattached_dst = NULL;
1269                         }
1270                 }
1271                 skb_set_transport_header(skb, skb_network_offset(skb) + sizeof(struct iphdr));
1272
1273                 iph = ip_hdr(skb);
1274                 iph->tos = tos;
1275                 iph->ttl = ttl;
1276                 iph->frag_off = frag_off;
1277                 ip_select_ident(iph, &rt_dst(rt), NULL);
1278
1279                 skb = tnl_vport->tnl_ops->update_header(vport, mutable,
1280                                                         &rt_dst(rt), skb);
1281                 if (unlikely(!skb))
1282                         goto next;
1283
1284                 if (likely(cache)) {
1285                         int orig_len = skb->len - cache->len;
1286                         struct vport *cache_vport;
1287
1288                         cache_vport = internal_dev_get_vport(rt_dst(rt).dev);
1289                         skb->protocol = htons(ETH_P_IP);
1290                         iph = ip_hdr(skb);
1291                         iph->tot_len = htons(skb->len - skb_network_offset(skb));
1292                         ip_send_check(iph);
1293
1294                         if (cache_vport) {
1295                                 if (unlikely(compute_ip_summed(skb, true))) {
1296                                         kfree_skb(skb);
1297                                         goto next;
1298                                 }
1299
1300                                 OVS_CB(skb)->flow = cache->flow;
1301                                 vport_receive(cache_vport, skb);
1302                                 sent_len += orig_len;
1303                         } else {
1304                                 int xmit_err;
1305
1306                                 skb->dev = rt_dst(rt).dev;
1307                                 xmit_err = dev_queue_xmit(skb);
1308
1309                                 if (likely(net_xmit_eval(xmit_err) == 0))
1310                                         sent_len += orig_len;
1311                         }
1312                 } else
1313                         sent_len += send_frags(skb, mutable);
1314
1315 next:
1316                 skb = next_skb;
1317         }
1318
1319         if (unlikely(sent_len == 0))
1320                 vport_record_error(vport, VPORT_E_TX_DROPPED);
1321
1322         goto out;
1323
1324 error_free:
1325         tnl_free_linked_skbs(skb);
1326 error:
1327         vport_record_error(vport, err);
1328 out:
1329         dst_release(unattached_dst);
1330         return sent_len;
1331 }
1332
1333 static const struct nla_policy tnl_policy[OVS_TUNNEL_ATTR_MAX + 1] = {
1334         [OVS_TUNNEL_ATTR_FLAGS]    = { .type = NLA_U32 },
1335         [OVS_TUNNEL_ATTR_DST_IPV4] = { .type = NLA_U32 },
1336         [OVS_TUNNEL_ATTR_SRC_IPV4] = { .type = NLA_U32 },
1337         [OVS_TUNNEL_ATTR_OUT_KEY]  = { .type = NLA_U64 },
1338         [OVS_TUNNEL_ATTR_IN_KEY]   = { .type = NLA_U64 },
1339         [OVS_TUNNEL_ATTR_TOS]      = { .type = NLA_U8 },
1340         [OVS_TUNNEL_ATTR_TTL]      = { .type = NLA_U8 },
1341 };
1342
1343 /* Sets OVS_TUNNEL_ATTR_* fields in 'mutable', which must initially be
1344  * zeroed. */
1345 static int tnl_set_config(struct nlattr *options, const struct tnl_ops *tnl_ops,
1346                           const struct vport *cur_vport,
1347                           struct tnl_mutable_config *mutable)
1348 {
1349         const struct vport *old_vport;
1350         const struct tnl_mutable_config *old_mutable;
1351         struct nlattr *a[OVS_TUNNEL_ATTR_MAX + 1];
1352         int err;
1353
1354         if (!options)
1355                 return -EINVAL;
1356
1357         err = nla_parse_nested(a, OVS_TUNNEL_ATTR_MAX, options, tnl_policy);
1358         if (err)
1359                 return err;
1360
1361         if (!a[OVS_TUNNEL_ATTR_FLAGS] || !a[OVS_TUNNEL_ATTR_DST_IPV4])
1362                 return -EINVAL;
1363
1364         mutable->flags = nla_get_u32(a[OVS_TUNNEL_ATTR_FLAGS]) & TNL_F_PUBLIC;
1365
1366         mutable->key.daddr = nla_get_be32(a[OVS_TUNNEL_ATTR_DST_IPV4]);
1367         if (a[OVS_TUNNEL_ATTR_SRC_IPV4]) {
1368                 if (ipv4_is_multicast(mutable->key.daddr))
1369                         return -EINVAL;
1370                 mutable->key.saddr = nla_get_be32(a[OVS_TUNNEL_ATTR_SRC_IPV4]);
1371         }
1372
1373         if (a[OVS_TUNNEL_ATTR_TOS]) {
1374                 mutable->tos = nla_get_u8(a[OVS_TUNNEL_ATTR_TOS]);
1375                 if (mutable->tos != RT_TOS(mutable->tos))
1376                         return -EINVAL;
1377         }
1378
1379         if (a[OVS_TUNNEL_ATTR_TTL])
1380                 mutable->ttl = nla_get_u8(a[OVS_TUNNEL_ATTR_TTL]);
1381
1382         mutable->key.tunnel_type = tnl_ops->tunnel_type;
1383         if (!a[OVS_TUNNEL_ATTR_IN_KEY]) {
1384                 mutable->key.tunnel_type |= TNL_T_KEY_MATCH;
1385                 mutable->flags |= TNL_F_IN_KEY_MATCH;
1386         } else {
1387                 mutable->key.tunnel_type |= TNL_T_KEY_EXACT;
1388                 mutable->key.in_key = nla_get_be64(a[OVS_TUNNEL_ATTR_IN_KEY]);
1389         }
1390
1391         if (!a[OVS_TUNNEL_ATTR_OUT_KEY])
1392                 mutable->flags |= TNL_F_OUT_KEY_ACTION;
1393         else
1394                 mutable->out_key = nla_get_be64(a[OVS_TUNNEL_ATTR_OUT_KEY]);
1395
1396         mutable->tunnel_hlen = tnl_ops->hdr_len(mutable);
1397         if (mutable->tunnel_hlen < 0)
1398                 return mutable->tunnel_hlen;
1399
1400         mutable->tunnel_hlen += sizeof(struct iphdr);
1401
1402         old_vport = port_table_lookup(&mutable->key, &old_mutable);
1403         if (old_vport && old_vport != cur_vport)
1404                 return -EEXIST;
1405
1406         mutable->mlink = 0;
1407         if (ipv4_is_multicast(mutable->key.daddr)) {
1408                 struct net_device *dev;
1409                 struct rtable *rt;
1410
1411                 rt = __find_route(mutable, tnl_ops->ipproto, mutable->tos);
1412                 if (IS_ERR(rt))
1413                         return -EADDRNOTAVAIL;
1414                 dev = rt_dst(rt).dev;
1415                 ip_rt_put(rt);
1416                 if (__in_dev_get_rtnl(dev) == NULL)
1417                         return -EADDRNOTAVAIL;
1418                 mutable->mlink = dev->ifindex;
1419                 ip_mc_inc_group(__in_dev_get_rtnl(dev), mutable->key.daddr);
1420         }
1421
1422         return 0;
1423 }
1424
1425 struct vport *tnl_create(const struct vport_parms *parms,
1426                          const struct vport_ops *vport_ops,
1427                          const struct tnl_ops *tnl_ops)
1428 {
1429         struct vport *vport;
1430         struct tnl_vport *tnl_vport;
1431         struct tnl_mutable_config *mutable;
1432         int initial_frag_id;
1433         int err;
1434
1435         vport = vport_alloc(sizeof(struct tnl_vport), vport_ops, parms);
1436         if (IS_ERR(vport)) {
1437                 err = PTR_ERR(vport);
1438                 goto error;
1439         }
1440
1441         tnl_vport = tnl_vport_priv(vport);
1442
1443         strcpy(tnl_vport->name, parms->name);
1444         tnl_vport->tnl_ops = tnl_ops;
1445
1446         mutable = kzalloc(sizeof(struct tnl_mutable_config), GFP_KERNEL);
1447         if (!mutable) {
1448                 err = -ENOMEM;
1449                 goto error_free_vport;
1450         }
1451
1452         vport_gen_rand_ether_addr(mutable->eth_addr);
1453
1454         get_random_bytes(&initial_frag_id, sizeof(int));
1455         atomic_set(&tnl_vport->frag_id, initial_frag_id);
1456
1457         err = tnl_set_config(parms->options, tnl_ops, NULL, mutable);
1458         if (err)
1459                 goto error_free_mutable;
1460
1461         spin_lock_init(&tnl_vport->cache_lock);
1462
1463 #ifdef NEED_CACHE_TIMEOUT
1464         tnl_vport->cache_exp_interval = MAX_CACHE_EXP -
1465                                        (net_random() % (MAX_CACHE_EXP / 2));
1466 #endif
1467
1468         rcu_assign_pointer(tnl_vport->mutable, mutable);
1469
1470         port_table_add_port(vport);
1471         return vport;
1472
1473 error_free_mutable:
1474         free_mutable_rtnl(mutable);
1475         kfree(mutable);
1476 error_free_vport:
1477         vport_free(vport);
1478 error:
1479         return ERR_PTR(err);
1480 }
1481
1482 int tnl_set_options(struct vport *vport, struct nlattr *options)
1483 {
1484         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1485         const struct tnl_mutable_config *old_mutable;
1486         struct tnl_mutable_config *mutable;
1487         int err;
1488
1489         mutable = kzalloc(sizeof(struct tnl_mutable_config), GFP_KERNEL);
1490         if (!mutable) {
1491                 err = -ENOMEM;
1492                 goto error;
1493         }
1494
1495         /* Copy fields whose values should be retained. */
1496         old_mutable = rtnl_dereference(tnl_vport->mutable);
1497         mutable->seq = old_mutable->seq + 1;
1498         memcpy(mutable->eth_addr, old_mutable->eth_addr, ETH_ALEN);
1499
1500         /* Parse the others configured by userspace. */
1501         err = tnl_set_config(options, tnl_vport->tnl_ops, vport, mutable);
1502         if (err)
1503                 goto error_free;
1504
1505         if (port_hash(&mutable->key) != port_hash(&old_mutable->key))
1506                 port_table_move_port(vport, mutable);
1507         else
1508                 assign_config_rcu(vport, mutable);
1509
1510         return 0;
1511
1512 error_free:
1513         free_mutable_rtnl(mutable);
1514         kfree(mutable);
1515 error:
1516         return err;
1517 }
1518
1519 int tnl_get_options(const struct vport *vport, struct sk_buff *skb)
1520 {
1521         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1522         const struct tnl_mutable_config *mutable = rcu_dereference_rtnl(tnl_vport->mutable);
1523
1524         NLA_PUT_U32(skb, OVS_TUNNEL_ATTR_FLAGS, mutable->flags & TNL_F_PUBLIC);
1525         NLA_PUT_BE32(skb, OVS_TUNNEL_ATTR_DST_IPV4, mutable->key.daddr);
1526
1527         if (!(mutable->flags & TNL_F_IN_KEY_MATCH))
1528                 NLA_PUT_BE64(skb, OVS_TUNNEL_ATTR_IN_KEY, mutable->key.in_key);
1529         if (!(mutable->flags & TNL_F_OUT_KEY_ACTION))
1530                 NLA_PUT_BE64(skb, OVS_TUNNEL_ATTR_OUT_KEY, mutable->out_key);
1531         if (mutable->key.saddr)
1532                 NLA_PUT_BE32(skb, OVS_TUNNEL_ATTR_SRC_IPV4, mutable->key.saddr);
1533         if (mutable->tos)
1534                 NLA_PUT_U8(skb, OVS_TUNNEL_ATTR_TOS, mutable->tos);
1535         if (mutable->ttl)
1536                 NLA_PUT_U8(skb, OVS_TUNNEL_ATTR_TTL, mutable->ttl);
1537
1538         return 0;
1539
1540 nla_put_failure:
1541         return -EMSGSIZE;
1542 }
1543
1544 static void free_port_rcu(struct rcu_head *rcu)
1545 {
1546         struct tnl_vport *tnl_vport = container_of(rcu,
1547                                                    struct tnl_vport, rcu);
1548
1549         free_cache((struct tnl_cache __force *)tnl_vport->cache);
1550         kfree((struct tnl_mutable __force *)tnl_vport->mutable);
1551         vport_free(tnl_vport_to_vport(tnl_vport));
1552 }
1553
1554 void tnl_destroy(struct vport *vport)
1555 {
1556         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1557         struct tnl_mutable_config *mutable;
1558
1559         mutable = rtnl_dereference(tnl_vport->mutable);
1560         port_table_remove_port(vport);
1561         free_mutable_rtnl(mutable);
1562         call_rcu(&tnl_vport->rcu, free_port_rcu);
1563 }
1564
1565 int tnl_set_addr(struct vport *vport, const unsigned char *addr)
1566 {
1567         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1568         struct tnl_mutable_config *old_mutable, *mutable;
1569
1570         old_mutable = rtnl_dereference(tnl_vport->mutable);
1571         mutable = kmemdup(old_mutable, sizeof(struct tnl_mutable_config), GFP_KERNEL);
1572         if (!mutable)
1573                 return -ENOMEM;
1574
1575         old_mutable->mlink = 0;
1576
1577         memcpy(mutable->eth_addr, addr, ETH_ALEN);
1578         assign_config_rcu(vport, mutable);
1579
1580         return 0;
1581 }
1582
1583 const char *tnl_get_name(const struct vport *vport)
1584 {
1585         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1586         return tnl_vport->name;
1587 }
1588
1589 const unsigned char *tnl_get_addr(const struct vport *vport)
1590 {
1591         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1592         return rcu_dereference_rtnl(tnl_vport->mutable)->eth_addr;
1593 }
1594
1595 void tnl_free_linked_skbs(struct sk_buff *skb)
1596 {
1597         while (skb) {
1598                 struct sk_buff *next = skb->next;
1599                 kfree_skb(skb);
1600                 skb = next;
1601         }
1602 }
1603
1604 int tnl_init(void)
1605 {
1606         int i;
1607
1608         port_table = kmalloc(PORT_TABLE_SIZE * sizeof(struct hlist_head *),
1609                         GFP_KERNEL);
1610         if (!port_table)
1611                 return -ENOMEM;
1612
1613         for (i = 0; i < PORT_TABLE_SIZE; i++)
1614                 INIT_HLIST_HEAD(&port_table[i]);
1615
1616         return 0;
1617 }
1618
1619 void tnl_exit(void)
1620 {
1621         int i;
1622
1623         for (i = 0; i < PORT_TABLE_SIZE; i++) {
1624                 struct tnl_vport *tnl_vport;
1625                 struct hlist_head *hash_head;
1626                 struct hlist_node *n;
1627
1628                 hash_head = &port_table[i];
1629                 hlist_for_each_entry(tnl_vport, n, hash_head, hash_node) {
1630                         BUG();
1631                         goto out;
1632                 }
1633         }
1634 out:
1635         kfree(port_table);
1636 }