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/in.h>
  14 #include <linux/in_route.h>
  15 #include <linux/jhash.h>
  16 #include <linux/kernel.h>
  17 #include <linux/version.h>
  18 #include <linux/workqueue.h>
  19
  20 #include <net/dsfield.h>
  21 #include <net/dst.h>
  22 #include <net/icmp.h>
  23 #include <net/inet_ecn.h>
  24 #include <net/ip.h>
  25 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
  26 #include <net/ipv6.h>
  27 #endif
  28 #include <net/route.h>
  29 #include <net/xfrm.h>
  30
  31 #include "actions.h"
  32 #include "checksum.h"
  33 #include "datapath.h"
  34 #include "table.h"
  35 #include "tunnel.h"
  36 #include "vlan.h"
  37 #include "vport.h"
  38 #include "vport-generic.h"
  39 #include "vport-internal_dev.h"
  40
  41 #ifdef NEED_CACHE_TIMEOUT
  42 /*
  43  * On kernels where we can't quickly detect changes in the rest of the system
  44  * we use an expiration time to invalidate the cache.  A shorter expiration
  45  * reduces the length of time that we may potentially blackhole packets while
  46  * a longer time increases performance by reducing the frequency that the
  47  * cache needs to be rebuilt.  A variety of factors may cause the cache to be
  48  * invalidated before the expiration time but this is the maximum.  The time
  49  * is expressed in jiffies.
  50  */
  51 #define MAX_CACHE_EXP HZ
  52 #endif
  53
  54 /*
  55  * Interval to check for and remove caches that are no longer valid.  Caches
  56  * are checked for validity before they are used for packet encapsulation and
  57  * old caches are removed at that time.  However, if no packets are sent through
  58  * the tunnel then the cache will never be destroyed.  Since it holds
  59  * references to a number of system objects, the cache will continue to use
  60  * system resources by not allowing those objects to be destroyed.  The cache
  61  * cleaner is periodically run to free invalid caches.  It does not
  62  * significantly affect system performance.  A lower interval will release
  63  * resources faster but will itself consume resources by requiring more frequent
  64  * checks.  A longer interval may result in messages being printed to the kernel
  65  * message buffer about unreleased resources.  The interval is expressed in
  66  * jiffies.
  67  */
  68 #define CACHE_CLEANER_INTERVAL (5 * HZ)
  69
  70 #define CACHE_DATA_ALIGN 16
  71
  72 static struct tbl __rcu *port_table __read_mostly;
  73
  74 static void cache_cleaner(struct work_struct *work);
  75 static DECLARE_DELAYED_WORK(cache_cleaner_wq, cache_cleaner);
  76
  77 /*
  78  * These are just used as an optimization: they don't require any kind of
  79  * synchronization because we could have just as easily read the value before
  80  * the port change happened.
  81  */
  82 static unsigned int key_local_remote_ports __read_mostly;
  83 static unsigned int key_remote_ports __read_mostly;
  84 static unsigned int local_remote_ports __read_mostly;
  85 static unsigned int remote_ports __read_mostly;
  86
  87 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,36)
  88 #define rt_dst(rt) (rt->dst)
  89 #else
  90 #define rt_dst(rt) (rt->u.dst)
  91 #endif
  92
  93 static inline struct vport *tnl_vport_to_vport(const struct tnl_vport *tnl_vport)
  94 {
  95         return vport_from_priv(tnl_vport);
  96 }
  97
  98 static inline struct tnl_vport *tnl_vport_table_cast(const struct tbl_node *node)
  99 {
 100         return container_of(node, struct tnl_vport, tbl_node);
 101 }
 102
 103 /* This is analogous to rtnl_dereference for the tunnel cache.  It checks that
 104  * cache_lock is held, so it is only for update side code.
 105  */
 106 static inline struct tnl_cache *cache_dereference(struct tnl_vport *tnl_vport)
 107 {
 108         return rcu_dereference_protected(tnl_vport->cache,
 109                                          lockdep_is_held(&tnl_vport->cache_lock));
 110 }
 111
 112 static inline void schedule_cache_cleaner(void)
 113 {
 114         schedule_delayed_work(&cache_cleaner_wq, CACHE_CLEANER_INTERVAL);
 115 }
 116
 117 static void free_cache(struct tnl_cache *cache)
 118 {
 119         if (!cache)
 120                 return;
 121
 122         flow_put(cache->flow);
 123         ip_rt_put(cache->rt);
 124         kfree(cache);
 125 }
 126
 127 static void free_config_rcu(struct rcu_head *rcu)
 128 {
 129         struct tnl_mutable_config *c = container_of(rcu, struct tnl_mutable_config, rcu);
 130         kfree(c);
 131 }
 132
 133 static void free_cache_rcu(struct rcu_head *rcu)
 134 {
 135         struct tnl_cache *c = container_of(rcu, struct tnl_cache, rcu);
 136         free_cache(c);
 137 }
 138
 139 static void assign_config_rcu(struct vport *vport,
 140                               struct tnl_mutable_config *new_config)
 141 {
 142         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 143         struct tnl_mutable_config *old_config;
 144
 145         old_config = rtnl_dereference(tnl_vport->mutable);
 146         rcu_assign_pointer(tnl_vport->mutable, new_config);
 147         call_rcu(&old_config->rcu, free_config_rcu);
 148 }
 149
 150 static void assign_cache_rcu(struct vport *vport, struct tnl_cache *new_cache)
 151 {
 152         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 153         struct tnl_cache *old_cache;
 154
 155         old_cache = cache_dereference(tnl_vport);
 156         rcu_assign_pointer(tnl_vport->cache, new_cache);
 157
 158         if (old_cache)
 159                 call_rcu(&old_cache->rcu, free_cache_rcu);
 160 }
 161
 162 static unsigned int *find_port_pool(const struct tnl_mutable_config *mutable)
 163 {
 164         if (mutable->flags & TNL_F_IN_KEY_MATCH) {
 165                 if (mutable->saddr)
 166                         return &local_remote_ports;
 167                 else
 168                         return &remote_ports;
 169         } else {
 170                 if (mutable->saddr)
 171                         return &key_local_remote_ports;
 172                 else
 173                         return &key_remote_ports;
 174         }
 175 }
 176
 177 struct port_lookup_key {
 178         const struct tnl_mutable_config *mutable;
 179         __be64 key;
 180         u32 tunnel_type;
 181         __be32 saddr;
 182         __be32 daddr;
 183 };
 184
 185 /*
 186  * Modifies 'target' to store the rcu_dereferenced pointer that was used to do
 187  * the comparision.
 188  */
 189 static int port_cmp(const struct tbl_node *node, void *target, int unused)
 190 {
 191         const struct tnl_vport *tnl_vport = tnl_vport_table_cast(node);
 192         struct port_lookup_key *lookup = target;
 193
 194         lookup->mutable = rcu_dereference_rtnl(tnl_vport->mutable);
 195
 196         return (lookup->mutable->tunnel_type == lookup->tunnel_type &&
 197                 lookup->mutable->daddr == lookup->daddr &&
 198                 lookup->mutable->in_key == lookup->key &&
 199                 lookup->mutable->saddr == lookup->saddr);
 200 }
 201
 202 static u32 port_hash(struct port_lookup_key *k)
 203 {
 204         u32 x = jhash_3words((__force u32)k->saddr, (__force u32)k->daddr,
 205                              k->tunnel_type, 0);
 206         return jhash_2words((__force u64)k->key >> 32, (__force u32)k->key, x);
 207 }
 208
 209 static u32 mutable_hash(const struct tnl_mutable_config *mutable)
 210 {
 211         struct port_lookup_key lookup;
 212
 213         lookup.saddr = mutable->saddr;
 214         lookup.daddr = mutable->daddr;
 215         lookup.key = mutable->in_key;
 216         lookup.tunnel_type = mutable->tunnel_type;
 217
 218         return port_hash(&lookup);
 219 }
 220
 221 static void check_table_empty(void)
 222 {
 223         struct tbl *old_table = rtnl_dereference(port_table);
 224
 225         if (tbl_count(old_table) == 0) {
 226                 cancel_delayed_work_sync(&cache_cleaner_wq);
 227                 rcu_assign_pointer(port_table, NULL);
 228                 tbl_deferred_destroy(old_table, NULL);
 229         }
 230 }
 231
 232 static int add_port(struct vport *vport)
 233 {
 234         struct tbl *cur_table = rtnl_dereference(port_table);
 235         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 236         int err;
 237
 238         if (!port_table) {
 239                 struct tbl *new_table;
 240
 241                 new_table = tbl_create(TBL_MIN_BUCKETS);
 242                 if (!new_table)
 243                         return -ENOMEM;
 244
 245                 rcu_assign_pointer(port_table, new_table);
 246                 schedule_cache_cleaner();
 247
 248         } else if (tbl_count(cur_table) > tbl_n_buckets(cur_table)) {
 249                 struct tbl *new_table;
 250
 251                 new_table = tbl_expand(cur_table);
 252                 if (IS_ERR(new_table))
 253                         return PTR_ERR(new_table);
 254
 255                 rcu_assign_pointer(port_table, new_table);
 256                 tbl_deferred_destroy(cur_table, NULL);
 257         }
 258
 259         err = tbl_insert(rtnl_dereference(port_table), &tnl_vport->tbl_node,
 260                          mutable_hash(rtnl_dereference(tnl_vport->mutable)));
 261         if (err) {
 262                 check_table_empty();
 263                 return err;
 264         }
 265
 266         (*find_port_pool(rtnl_dereference(tnl_vport->mutable)))++;
 267
 268         return 0;
 269 }
 270
 271 static int move_port(struct vport *vport, struct tnl_mutable_config *new_mutable)
 272 {
 273         int err;
 274         struct tbl *cur_table = rtnl_dereference(port_table);
 275         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 276         u32 hash;
 277
 278         hash = mutable_hash(new_mutable);
 279         if (hash == tnl_vport->tbl_node.hash)
 280                 goto table_updated;
 281
 282         /*
 283          * Ideally we should make this move atomic to avoid having gaps in
 284          * finding tunnels or the possibility of failure.  However, if we do
 285          * find a tunnel it will always be consistent.
 286          */
 287         err = tbl_remove(cur_table, &tnl_vport->tbl_node);
 288         if (err)
 289                 return err;
 290
 291         err = tbl_insert(cur_table, &tnl_vport->tbl_node, hash);
 292         if (err) {
 293                 (*find_port_pool(rtnl_dereference(tnl_vport->mutable)))--;
 294                 check_table_empty();
 295                 return err;
 296         }
 297
 298 table_updated:
 299         (*find_port_pool(rtnl_dereference(tnl_vport->mutable)))--;
 300         assign_config_rcu(vport, new_mutable);
 301         (*find_port_pool(rtnl_dereference(tnl_vport->mutable)))++;
 302
 303         return 0;
 304 }
 305
 306 static int del_port(struct vport *vport)
 307 {
 308         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 309         int err;
 310
 311         err = tbl_remove(rtnl_dereference(port_table), &tnl_vport->tbl_node);
 312         if (err)
 313                 return err;
 314
 315         check_table_empty();
 316         (*find_port_pool(rtnl_dereference(tnl_vport->mutable)))--;
 317
 318         return 0;
 319 }
 320
 321 struct vport *tnl_find_port(__be32 saddr, __be32 daddr, __be64 key,
 322                             int tunnel_type,
 323                             const struct tnl_mutable_config **mutable)
 324 {
 325         struct port_lookup_key lookup;
 326         struct tbl *table = rcu_dereference_rtnl(port_table);
 327         struct tbl_node *tbl_node;
 328
 329         if (unlikely(!table))
 330                 return NULL;
 331
 332         lookup.saddr = saddr;
 333         lookup.daddr = daddr;
 334
 335         if (tunnel_type & TNL_T_KEY_EXACT) {
 336                 lookup.key = key;
 337                 lookup.tunnel_type = tunnel_type & ~TNL_T_KEY_MATCH;
 338
 339                 if (key_local_remote_ports) {
 340                         tbl_node = tbl_lookup(table, &lookup, sizeof(lookup),
 341                                               port_hash(&lookup), port_cmp);
 342                         if (tbl_node)
 343                                 goto found;
 344                 }
 345
 346                 if (key_remote_ports) {
 347                         lookup.saddr = 0;
 348
 349                         tbl_node = tbl_lookup(table, &lookup, sizeof(lookup),
 350                                               port_hash(&lookup), port_cmp);
 351                         if (tbl_node)
 352                                 goto found;
 353
 354                         lookup.saddr = saddr;
 355                 }
 356         }
 357
 358         if (tunnel_type & TNL_T_KEY_MATCH) {
 359                 lookup.key = 0;
 360                 lookup.tunnel_type = tunnel_type & ~TNL_T_KEY_EXACT;
 361
 362                 if (local_remote_ports) {
 363                         tbl_node = tbl_lookup(table, &lookup, sizeof(lookup),
 364                                               port_hash(&lookup), port_cmp);
 365                         if (tbl_node)
 366                                 goto found;
 367                 }
 368
 369                 if (remote_ports) {
 370                         lookup.saddr = 0;
 371
 372                         tbl_node = tbl_lookup(table, &lookup, sizeof(lookup),
 373                                               port_hash(&lookup), port_cmp);
 374                         if (tbl_node)
 375                                 goto found;
 376                 }
 377         }
 378
 379         return NULL;
 380
 381 found:
 382         *mutable = lookup.mutable;
 383         return tnl_vport_to_vport(tnl_vport_table_cast(tbl_node));
 384 }
 385
 386 static void ecn_decapsulate(struct sk_buff *skb, u8 tos)
 387 {
 388         if (unlikely(INET_ECN_is_ce(tos))) {
 389                 __be16 protocol = skb->protocol;
 390
 391                 skb_set_network_header(skb, ETH_HLEN);
 392
 393                 if (protocol == htons(ETH_P_8021Q)) {
 394                         if (unlikely(!pskb_may_pull(skb, VLAN_ETH_HLEN)))
 395                                 return;
 396
 397                         protocol = vlan_eth_hdr(skb)->h_vlan_encapsulated_proto;
 398                         skb_set_network_header(skb, VLAN_ETH_HLEN);
 399                 }
 400
 401                 if (protocol == htons(ETH_P_IP)) {
 402                         if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
 403                             + sizeof(struct iphdr))))
 404                                 return;
 405
 406                         IP_ECN_set_ce(ip_hdr(skb));
 407                 }
 408 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 409                 else if (protocol == htons(ETH_P_IPV6)) {
 410                         if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
 411                             + sizeof(struct ipv6hdr))))
 412                                 return;
 413
 414                         IP6_ECN_set_ce(ipv6_hdr(skb));
 415                 }
 416 #endif
 417         }
 418 }
 419
 420 /**
 421  *      tnl_rcv - ingress point for generic tunnel code
 422  *
 423  * @vport: port this packet was received on
 424  * @skb: received packet
 425  * @tos: ToS from encapsulating IP packet, used to copy ECN bits
 426  *
 427  * Must be called with rcu_read_lock.
 428  *
 429  * Packets received by this function are in the following state:
 430  * - skb->data points to the inner Ethernet header.
 431  * - The inner Ethernet header is in the linear data area.
 432  * - skb->csum does not include the inner Ethernet header.
 433  * - The layer pointers are undefined.
 434  */
 435 void tnl_rcv(struct vport *vport, struct sk_buff *skb, u8 tos)
 436 {
 437         struct ethhdr *eh;
 438
 439         skb_reset_mac_header(skb);
 440         eh = eth_hdr(skb);
 441
 442         if (likely(ntohs(eh->h_proto) >= 1536))
 443                 skb->protocol = eh->h_proto;
 444         else
 445                 skb->protocol = htons(ETH_P_802_2);
 446
 447         skb_dst_drop(skb);
 448         nf_reset(skb);
 449         skb_clear_rxhash(skb);
 450         secpath_reset(skb);
 451
 452         ecn_decapsulate(skb, tos);
 453         compute_ip_summed(skb, false);
 454         vlan_set_tci(skb, 0);
 455
 456         vport_receive(vport, skb);
 457 }
 458
 459 static bool check_ipv4_address(__be32 addr)
 460 {
 461         if (ipv4_is_multicast(addr) || ipv4_is_lbcast(addr)
 462             || ipv4_is_loopback(addr) || ipv4_is_zeronet(addr))
 463                 return false;
 464
 465         return true;
 466 }
 467
 468 static bool ipv4_should_icmp(struct sk_buff *skb)
 469 {
 470         struct iphdr *old_iph = ip_hdr(skb);
 471
 472         /* Don't respond to L2 broadcast. */
 473         if (is_multicast_ether_addr(eth_hdr(skb)->h_dest))
 474                 return false;
 475
 476         /* Don't respond to L3 broadcast or invalid addresses. */
 477         if (!check_ipv4_address(old_iph->daddr) ||
 478             !check_ipv4_address(old_iph->saddr))
 479                 return false;
 480
 481         /* Only respond to the first fragment. */
 482         if (old_iph->frag_off & htons(IP_OFFSET))
 483                 return false;
 484
 485         /* Don't respond to ICMP error messages. */
 486         if (old_iph->protocol == IPPROTO_ICMP) {
 487                 u8 icmp_type, *icmp_typep;
 488
 489                 icmp_typep = skb_header_pointer(skb, (u8 *)old_iph +
 490                                                 (old_iph->ihl << 2) +
 491                                                 offsetof(struct icmphdr, type) -
 492                                                 skb->data, sizeof(icmp_type),
 493                                                 &icmp_type);
 494
 495                 if (!icmp_typep)
 496                         return false;
 497
 498                 if (*icmp_typep > NR_ICMP_TYPES
 499                         || (*icmp_typep <= ICMP_PARAMETERPROB
 500                                 && *icmp_typep != ICMP_ECHOREPLY
 501                                 && *icmp_typep != ICMP_ECHO))
 502                         return false;
 503         }
 504
 505         return true;
 506 }
 507
 508 static void ipv4_build_icmp(struct sk_buff *skb, struct sk_buff *nskb,
 509                             unsigned int mtu, unsigned int payload_length)
 510 {
 511         struct iphdr *iph, *old_iph = ip_hdr(skb);
 512         struct icmphdr *icmph;
 513         u8 *payload;
 514
 515         iph = (struct iphdr *)skb_put(nskb, sizeof(struct iphdr));
 516         icmph = (struct icmphdr *)skb_put(nskb, sizeof(struct icmphdr));
 517         payload = skb_put(nskb, payload_length);
 518
 519         /* IP */
 520         iph->version            =       4;
 521         iph->ihl                =       sizeof(struct iphdr) >> 2;
 522         iph->tos                =       (old_iph->tos & IPTOS_TOS_MASK) |
 523                                         IPTOS_PREC_INTERNETCONTROL;
 524         iph->tot_len            =       htons(sizeof(struct iphdr)
 525                                               + sizeof(struct icmphdr)
 526                                               + payload_length);
 527         get_random_bytes(&iph->id, sizeof(iph->id));
 528         iph->frag_off           =       0;
 529         iph->ttl                =       IPDEFTTL;
 530         iph->protocol           =       IPPROTO_ICMP;
 531         iph->daddr              =       old_iph->saddr;
 532         iph->saddr              =       old_iph->daddr;
 533
 534         ip_send_check(iph);
 535
 536         /* ICMP */
 537         icmph->type             =       ICMP_DEST_UNREACH;
 538         icmph->code             =       ICMP_FRAG_NEEDED;
 539         icmph->un.gateway       =       htonl(mtu);
 540         icmph->checksum         =       0;
 541
 542         nskb->csum = csum_partial((u8 *)icmph, sizeof(struct icmphdr), 0);
 543         nskb->csum = skb_copy_and_csum_bits(skb, (u8 *)old_iph - skb->data,
 544                                             payload, payload_length,
 545                                             nskb->csum);
 546         icmph->checksum = csum_fold(nskb->csum);
 547 }
 548
 549 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 550 static bool ipv6_should_icmp(struct sk_buff *skb)
 551 {
 552         struct ipv6hdr *old_ipv6h = ipv6_hdr(skb);
 553         int addr_type;
 554         int payload_off = (u8 *)(old_ipv6h + 1) - skb->data;
 555         u8 nexthdr = ipv6_hdr(skb)->nexthdr;
 556
 557         /* Check source address is valid. */
 558         addr_type = ipv6_addr_type(&old_ipv6h->saddr);
 559         if (addr_type & IPV6_ADDR_MULTICAST || addr_type == IPV6_ADDR_ANY)
 560                 return false;
 561
 562         /* Don't reply to unspecified addresses. */
 563         if (ipv6_addr_type(&old_ipv6h->daddr) == IPV6_ADDR_ANY)
 564                 return false;
 565
 566         /* Don't respond to ICMP error messages. */
 567         payload_off = ipv6_skip_exthdr(skb, payload_off, &nexthdr);
 568         if (payload_off < 0)
 569                 return false;
 570
 571         if (nexthdr == NEXTHDR_ICMP) {
 572                 u8 icmp_type, *icmp_typep;
 573
 574                 icmp_typep = skb_header_pointer(skb, payload_off +
 575                                                 offsetof(struct icmp6hdr,
 576                                                         icmp6_type),
 577                                                 sizeof(icmp_type), &icmp_type);
 578
 579                 if (!icmp_typep || !(*icmp_typep & ICMPV6_INFOMSG_MASK))
 580                         return false;
 581         }
 582
 583         return true;
 584 }
 585
 586 static void ipv6_build_icmp(struct sk_buff *skb, struct sk_buff *nskb,
 587                             unsigned int mtu, unsigned int payload_length)
 588 {
 589         struct ipv6hdr *ipv6h, *old_ipv6h = ipv6_hdr(skb);
 590         struct icmp6hdr *icmp6h;
 591         u8 *payload;
 592
 593         ipv6h = (struct ipv6hdr *)skb_put(nskb, sizeof(struct ipv6hdr));
 594         icmp6h = (struct icmp6hdr *)skb_put(nskb, sizeof(struct icmp6hdr));
 595         payload = skb_put(nskb, payload_length);
 596
 597         /* IPv6 */
 598         ipv6h->version          =       6;
 599         ipv6h->priority         =       0;
 600         memset(&ipv6h->flow_lbl, 0, sizeof(ipv6h->flow_lbl));
 601         ipv6h->payload_len      =       htons(sizeof(struct icmp6hdr)
 602                                               + payload_length);
 603         ipv6h->nexthdr          =       NEXTHDR_ICMP;
 604         ipv6h->hop_limit        =       IPV6_DEFAULT_HOPLIMIT;
 605         ipv6_addr_copy(&ipv6h->daddr, &old_ipv6h->saddr);
 606         ipv6_addr_copy(&ipv6h->saddr, &old_ipv6h->daddr);
 607
 608         /* ICMPv6 */
 609         icmp6h->icmp6_type      =       ICMPV6_PKT_TOOBIG;
 610         icmp6h->icmp6_code      =       0;
 611         icmp6h->icmp6_cksum     =       0;
 612         icmp6h->icmp6_mtu       =       htonl(mtu);
 613
 614         nskb->csum = csum_partial((u8 *)icmp6h, sizeof(struct icmp6hdr), 0);
 615         nskb->csum = skb_copy_and_csum_bits(skb, (u8 *)old_ipv6h - skb->data,
 616                                             payload, payload_length,
 617                                             nskb->csum);
 618         icmp6h->icmp6_cksum = csum_ipv6_magic(&ipv6h->saddr, &ipv6h->daddr,
 619                                                 sizeof(struct icmp6hdr)
 620                                                 + payload_length,
 621                                                 ipv6h->nexthdr, nskb->csum);
 622 }
 623 #endif /* IPv6 */
 624
 625 bool tnl_frag_needed(struct vport *vport, const struct tnl_mutable_config *mutable,
 626                      struct sk_buff *skb, unsigned int mtu, __be64 flow_key)
 627 {
 628         unsigned int eth_hdr_len = ETH_HLEN;
 629         unsigned int total_length = 0, header_length = 0, payload_length;
 630         struct ethhdr *eh, *old_eh = eth_hdr(skb);
 631         struct sk_buff *nskb;
 632
 633         /* Sanity check */
 634         if (skb->protocol == htons(ETH_P_IP)) {
 635                 if (mtu < IP_MIN_MTU)
 636                         return false;
 637
 638                 if (!ipv4_should_icmp(skb))
 639                         return true;
 640         }
 641 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 642         else if (skb->protocol == htons(ETH_P_IPV6)) {
 643                 if (mtu < IPV6_MIN_MTU)
 644                         return false;
 645
 646                 /*
 647                  * In theory we should do PMTUD on IPv6 multicast messages but
 648                  * we don't have an address to send from so just fragment.
 649                  */
 650                 if (ipv6_addr_type(&ipv6_hdr(skb)->daddr) & IPV6_ADDR_MULTICAST)
 651                         return false;
 652
 653                 if (!ipv6_should_icmp(skb))
 654                         return true;
 655         }
 656 #endif
 657         else
 658                 return false;
 659
 660         /* Allocate */
 661         if (old_eh->h_proto == htons(ETH_P_8021Q))
 662                 eth_hdr_len = VLAN_ETH_HLEN;
 663
 664         payload_length = skb->len - eth_hdr_len;
 665         if (skb->protocol == htons(ETH_P_IP)) {
 666                 header_length = sizeof(struct iphdr) + sizeof(struct icmphdr);
 667                 total_length = min_t(unsigned int, header_length +
 668                                                    payload_length, 576);
 669         }
 670 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 671         else {
 672                 header_length = sizeof(struct ipv6hdr) +
 673                                 sizeof(struct icmp6hdr);
 674                 total_length = min_t(unsigned int, header_length +
 675                                                   payload_length, IPV6_MIN_MTU);
 676         }
 677 #endif
 678
 679         payload_length = total_length - header_length;
 680
 681         nskb = dev_alloc_skb(NET_IP_ALIGN + eth_hdr_len + header_length +
 682                              payload_length);
 683         if (!nskb)
 684                 return false;
 685
 686         skb_reserve(nskb, NET_IP_ALIGN);
 687
 688         /* Ethernet / VLAN */
 689         eh = (struct ethhdr *)skb_put(nskb, eth_hdr_len);
 690         memcpy(eh->h_dest, old_eh->h_source, ETH_ALEN);
 691         memcpy(eh->h_source, mutable->eth_addr, ETH_ALEN);
 692         nskb->protocol = eh->h_proto = old_eh->h_proto;
 693         if (old_eh->h_proto == htons(ETH_P_8021Q)) {
 694                 struct vlan_ethhdr *vh = (struct vlan_ethhdr *)eh;
 695
 696                 vh->h_vlan_TCI = vlan_eth_hdr(skb)->h_vlan_TCI;
 697                 vh->h_vlan_encapsulated_proto = skb->protocol;
 698         } else
 699                 vlan_set_tci(nskb, vlan_get_tci(skb));
 700         skb_reset_mac_header(nskb);
 701
 702         /* Protocol */
 703         if (skb->protocol == htons(ETH_P_IP))
 704                 ipv4_build_icmp(skb, nskb, mtu, payload_length);
 705 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 706         else
 707                 ipv6_build_icmp(skb, nskb, mtu, payload_length);
 708 #endif
 709
 710         /*
 711          * Assume that flow based keys are symmetric with respect to input
 712          * and output and use the key that we were going to put on the
 713          * outgoing packet for the fake received packet.  If the keys are
 714          * not symmetric then PMTUD needs to be disabled since we won't have
 715          * any way of synthesizing packets.
 716          */
 717         if ((mutable->flags & (TNL_F_IN_KEY_MATCH | TNL_F_OUT_KEY_ACTION)) ==
 718             (TNL_F_IN_KEY_MATCH | TNL_F_OUT_KEY_ACTION))
 719                 OVS_CB(nskb)->tun_id = flow_key;
 720
 721         compute_ip_summed(nskb, false);
 722         vport_receive(vport, nskb);
 723
 724         return true;
 725 }
 726
 727 static bool check_mtu(struct sk_buff *skb,
 728                       struct vport *vport,
 729                       const struct tnl_mutable_config *mutable,
 730                       const struct rtable *rt, __be16 *frag_offp)
 731 {
 732         bool df_inherit = mutable->flags & TNL_F_DF_INHERIT;
 733         bool pmtud = mutable->flags & TNL_F_PMTUD;
 734         __be16 frag_off = mutable->flags & TNL_F_DF_DEFAULT ? htons(IP_DF) : 0;
 735         int mtu = 0;
 736         unsigned int packet_length = skb->len - ETH_HLEN;
 737
 738         /* Allow for one level of tagging in the packet length. */
 739         if (!vlan_tx_tag_present(skb) &&
 740             eth_hdr(skb)->h_proto == htons(ETH_P_8021Q))
 741                 packet_length -= VLAN_HLEN;
 742
 743         if (pmtud) {
 744                 int vlan_header = 0;
 745
 746                 /* The tag needs to go in packet regardless of where it
 747                  * currently is, so subtract it from the MTU.
 748                  */
 749                 if (vlan_tx_tag_present(skb) ||
 750                     eth_hdr(skb)->h_proto == htons(ETH_P_8021Q))
 751                         vlan_header = VLAN_HLEN;
 752
 753                 mtu = dst_mtu(&rt_dst(rt))
 754                         - ETH_HLEN
 755                         - mutable->tunnel_hlen
 756                         - vlan_header;
 757         }
 758
 759         if (skb->protocol == htons(ETH_P_IP)) {
 760                 struct iphdr *iph = ip_hdr(skb);
 761
 762                 if (df_inherit)
 763                         frag_off = iph->frag_off & htons(IP_DF);
 764
 765                 if (pmtud && iph->frag_off & htons(IP_DF)) {
 766                         mtu = max(mtu, IP_MIN_MTU);
 767
 768                         if (packet_length > mtu &&
 769                             tnl_frag_needed(vport, mutable, skb, mtu,
 770                                             OVS_CB(skb)->tun_id))
 771                                 return false;
 772                 }
 773         }
 774 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 775         else if (skb->protocol == htons(ETH_P_IPV6)) {
 776                 /* IPv6 requires end hosts to do fragmentation
 777                  * if the packet is above the minimum MTU.
 778                  */
 779                 if (df_inherit && packet_length > IPV6_MIN_MTU)
 780                         frag_off = htons(IP_DF);
 781
 782                 if (pmtud) {
 783                         mtu = max(mtu, IPV6_MIN_MTU);
 784
 785                         if (packet_length > mtu &&
 786                             tnl_frag_needed(vport, mutable, skb, mtu,
 787                                             OVS_CB(skb)->tun_id))
 788                                 return false;
 789                 }
 790         }
 791 #endif
 792
 793         *frag_offp = frag_off;
 794         return true;
 795 }
 796
 797 static void create_tunnel_header(const struct vport *vport,
 798                                  const struct tnl_mutable_config *mutable,
 799                                  const struct rtable *rt, void *header)
 800 {
 801         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 802         struct iphdr *iph = header;
 803
 804         iph->version    = 4;
 805         iph->ihl        = sizeof(struct iphdr) >> 2;
 806         iph->frag_off   = htons(IP_DF);
 807         iph->protocol   = tnl_vport->tnl_ops->ipproto;
 808         iph->tos        = mutable->tos;
 809         iph->daddr      = rt->rt_dst;
 810         iph->saddr      = rt->rt_src;
 811         iph->ttl        = mutable->ttl;
 812         if (!iph->ttl)
 813                 iph->ttl = ip4_dst_hoplimit(&rt_dst(rt));
 814
 815         tnl_vport->tnl_ops->build_header(vport, mutable, iph + 1);
 816 }
 817
 818 static inline void *get_cached_header(const struct tnl_cache *cache)
 819 {
 820         return (void *)cache + ALIGN(sizeof(struct tnl_cache), CACHE_DATA_ALIGN);
 821 }
 822
 823 static inline bool check_cache_valid(const struct tnl_cache *cache,
 824                                      const struct tnl_mutable_config *mutable)
 825 {
 826         return cache &&
 827 #ifdef NEED_CACHE_TIMEOUT
 828                 time_before(jiffies, cache->expiration) &&
 829 #endif
 830 #ifdef HAVE_RT_GENID
 831                 atomic_read(&init_net.ipv4.rt_genid) == cache->rt->rt_genid &&
 832 #endif
 833 #ifdef HAVE_HH_SEQ
 834                 rt_dst(cache->rt).hh->hh_lock.sequence == cache->hh_seq &&
 835 #endif
 836                 mutable->seq == cache->mutable_seq &&
 837                 (!is_internal_dev(rt_dst(cache->rt).dev) ||
 838                 (cache->flow && !cache->flow->dead));
 839 }
 840
 841 static int cache_cleaner_cb(struct tbl_node *tbl_node, void *aux)
 842 {
 843         struct tnl_vport *tnl_vport = tnl_vport_table_cast(tbl_node);
 844         const struct tnl_mutable_config *mutable = rcu_dereference(tnl_vport->mutable);
 845         const struct tnl_cache *cache = rcu_dereference(tnl_vport->cache);
 846
 847         if (cache && !check_cache_valid(cache, mutable) &&
 848             spin_trylock_bh(&tnl_vport->cache_lock)) {
 849                 assign_cache_rcu(tnl_vport_to_vport(tnl_vport), NULL);
 850                 spin_unlock_bh(&tnl_vport->cache_lock);
 851         }
 852
 853         return 0;
 854 }
 855
 856 static void cache_cleaner(struct work_struct *work)
 857 {
 858         schedule_cache_cleaner();
 859
 860         rcu_read_lock();
 861         tbl_foreach(rcu_dereference(port_table), cache_cleaner_cb, NULL);
 862         rcu_read_unlock();
 863 }
 864
 865 static inline void create_eth_hdr(struct tnl_cache *cache,
 866                                   const struct rtable *rt)
 867 {
 868         void *cache_data = get_cached_header(cache);
 869         int hh_len = rt_dst(rt).hh->hh_len;
 870         int hh_off = HH_DATA_ALIGN(rt_dst(rt).hh->hh_len) - hh_len;
 871
 872 #ifdef HAVE_HH_SEQ
 873         unsigned hh_seq;
 874
 875         do {
 876                 hh_seq = read_seqbegin(&rt_dst(rt).hh->hh_lock);
 877                 memcpy(cache_data, (void *)rt_dst(rt).hh->hh_data + hh_off, hh_len);
 878         } while (read_seqretry(&rt_dst(rt).hh->hh_lock, hh_seq));
 879
 880         cache->hh_seq = hh_seq;
 881 #else
 882         read_lock_bh(&rt_dst(rt).hh->hh_lock);
 883         memcpy(cache_data, (void *)rt_dst(rt).hh->hh_data + hh_off, hh_len);
 884         read_unlock_bh(&rt_dst(rt).hh->hh_lock);
 885 #endif
 886 }
 887
 888 static struct tnl_cache *build_cache(struct vport *vport,
 889                                      const struct tnl_mutable_config *mutable,
 890                                      struct rtable *rt)
 891 {
 892         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 893         struct tnl_cache *cache;
 894         void *cache_data;
 895         int cache_len;
 896
 897         if (!(mutable->flags & TNL_F_HDR_CACHE))
 898                 return NULL;
 899
 900         /*
 901          * If there is no entry in the ARP cache or if this device does not
 902          * support hard header caching just fall back to the IP stack.
 903          */
 904         if (!rt_dst(rt).hh)
 905                 return NULL;
 906
 907         /*
 908          * If lock is contended fall back to directly building the header.
 909          * We're not going to help performance by sitting here spinning.
 910          */
 911         if (!spin_trylock_bh(&tnl_vport->cache_lock))
 912                 return NULL;
 913
 914         cache = cache_dereference(tnl_vport);
 915         if (check_cache_valid(cache, mutable))
 916                 goto unlock;
 917         else
 918                 cache = NULL;
 919
 920         cache_len = rt_dst(rt).hh->hh_len + mutable->tunnel_hlen;
 921
 922         cache = kzalloc(ALIGN(sizeof(struct tnl_cache), CACHE_DATA_ALIGN) +
 923                         cache_len, GFP_ATOMIC);
 924         if (!cache)
 925                 goto unlock;
 926
 927         cache->len = cache_len;
 928
 929         create_eth_hdr(cache, rt);
 930         cache_data = get_cached_header(cache) + rt_dst(rt).hh->hh_len;
 931
 932         create_tunnel_header(vport, mutable, rt, cache_data);
 933
 934         cache->mutable_seq = mutable->seq;
 935         cache->rt = rt;
 936 #ifdef NEED_CACHE_TIMEOUT
 937         cache->expiration = jiffies + tnl_vport->cache_exp_interval;
 938 #endif
 939
 940         if (is_internal_dev(rt_dst(rt).dev)) {
 941                 struct sw_flow_key flow_key;
 942                 struct tbl_node *flow_node;
 943                 struct vport *dst_vport;
 944                 struct sk_buff *skb;
 945                 bool is_frag;
 946                 int err;
 947                 int flow_key_len;
 948
 949                 dst_vport = internal_dev_get_vport(rt_dst(rt).dev);
 950                 if (!dst_vport)
 951                         goto done;
 952
 953                 skb = alloc_skb(cache->len, GFP_ATOMIC);
 954                 if (!skb)
 955                         goto done;
 956
 957                 __skb_put(skb, cache->len);
 958                 memcpy(skb->data, get_cached_header(cache), cache->len);
 959
 960                 err = flow_extract(skb, dst_vport->port_no, &flow_key,
 961                                    &flow_key_len, &is_frag);
 962
 963                 kfree_skb(skb);
 964                 if (err || is_frag)
 965                         goto done;
 966
 967                 flow_node = tbl_lookup(rcu_dereference(dst_vport->dp->table),
 968                                        &flow_key, flow_key_len,
 969                                        flow_hash(&flow_key, flow_key_len),
 970                                        flow_cmp);
 971                 if (flow_node) {
 972                         struct sw_flow *flow = flow_cast(flow_node);
 973
 974                         cache->flow = flow;
 975                         flow_hold(flow);
 976                 }
 977         }
 978
 979 done:
 980         assign_cache_rcu(vport, cache);
 981
 982 unlock:
 983         spin_unlock_bh(&tnl_vport->cache_lock);
 984
 985         return cache;
 986 }
 987
 988 static struct rtable *find_route(struct vport *vport,
 989                                  const struct tnl_mutable_config *mutable,
 990                                  u8 tos, struct tnl_cache **cache)
 991 {
 992         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 993         struct tnl_cache *cur_cache = rcu_dereference(tnl_vport->cache);
 994
 995         *cache = NULL;
 996         tos = RT_TOS(tos);
 997
 998         if (likely(tos == mutable->tos && check_cache_valid(cur_cache, mutable))) {
 999                 *cache = cur_cache;
1000                 return cur_cache->rt;
1001         } else {
1002                 struct rtable *rt;
1003 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,39)
1004                 struct flowi fl = { .nl_u = { .ip4_u =
1005                                               { .daddr = mutable->daddr,
1006                                                 .saddr = mutable->saddr,
1007                                                 .tos = tos } },
1008                                     .proto = tnl_vport->tnl_ops->ipproto };
1009
1010                 if (unlikely(ip_route_output_key(&init_net, &rt, &fl)))
1011                         return NULL;
1012 #else
1013                 struct flowi4 fl = { .daddr = mutable->daddr,
1014                                      .saddr = mutable->saddr,
1015                                      .flowi4_tos = tos,
1016                                      .flowi4_proto = tnl_vport->tnl_ops->ipproto };
1017
1018                 rt = ip_route_output_key(&init_net, &fl);
1019                 if (IS_ERR(rt))
1020                         return NULL;
1021 #endif
1022
1023                 if (likely(tos == mutable->tos))
1024                         *cache = build_cache(vport, mutable, rt);
1025
1026                 return rt;
1027         }
1028 }
1029
1030 static struct sk_buff *check_headroom(struct sk_buff *skb, int headroom)
1031 {
1032         if (skb_headroom(skb) < headroom || skb_header_cloned(skb)) {
1033                 struct sk_buff *nskb = skb_realloc_headroom(skb, headroom + 16);
1034                 if (unlikely(!nskb)) {
1035                         kfree_skb(skb);
1036                         return ERR_PTR(-ENOMEM);
1037                 }
1038
1039                 set_skb_csum_bits(skb, nskb);
1040
1041                 if (skb->sk)
1042                         skb_set_owner_w(nskb, skb->sk);
1043
1044                 kfree_skb(skb);
1045                 return nskb;
1046         }
1047
1048         return skb;
1049 }
1050
1051 static inline bool need_linearize(const struct sk_buff *skb)
1052 {
1053         int i;
1054
1055         if (unlikely(skb_shinfo(skb)->frag_list))
1056                 return true;
1057
1058         /*
1059          * Generally speaking we should linearize if there are paged frags.
1060          * However, if all of the refcounts are 1 we know nobody else can
1061          * change them from underneath us and we can skip the linearization.
1062          */
1063         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
1064                 if (unlikely(page_count(skb_shinfo(skb)->frags[i].page) > 1))
1065                         return true;
1066
1067         return false;
1068 }
1069
1070 static struct sk_buff *handle_offloads(struct sk_buff *skb,
1071                                        const struct tnl_mutable_config *mutable,
1072                                        const struct rtable *rt)
1073 {
1074         int min_headroom;
1075         int err;
1076
1077         forward_ip_summed(skb);
1078
1079         err = vswitch_skb_checksum_setup(skb);
1080         if (unlikely(err))
1081                 goto error_free;
1082
1083         min_headroom = LL_RESERVED_SPACE(rt_dst(rt).dev) + rt_dst(rt).header_len
1084                         + mutable->tunnel_hlen
1085                         + (vlan_tx_tag_present(skb) ? VLAN_HLEN : 0);
1086
1087         skb = check_headroom(skb, min_headroom);
1088         if (IS_ERR(skb)) {
1089                 err = PTR_ERR(skb);
1090                 goto error;
1091         }
1092
1093         if (skb_is_gso(skb)) {
1094                 struct sk_buff *nskb;
1095
1096                 nskb = skb_gso_segment(skb, 0);
1097                 kfree_skb(skb);
1098                 if (IS_ERR(nskb)) {
1099                         err = PTR_ERR(nskb);
1100                         goto error;
1101                 }
1102
1103                 skb = nskb;
1104         } else if (skb->ip_summed == CHECKSUM_PARTIAL) {
1105                 /* Pages aren't locked and could change at any time.
1106                  * If this happens after we compute the checksum, the
1107                  * checksum will be wrong.  We linearize now to avoid
1108                  * this problem.
1109                  */
1110                 if (unlikely(need_linearize(skb))) {
1111                         err = __skb_linearize(skb);
1112                         if (unlikely(err))
1113                                 goto error_free;
1114                 }
1115
1116                 err = skb_checksum_help(skb);
1117                 if (unlikely(err))
1118                         goto error_free;
1119         } else if (skb->ip_summed == CHECKSUM_COMPLETE)
1120                 skb->ip_summed = CHECKSUM_NONE;
1121
1122         return skb;
1123
1124 error_free:
1125         kfree_skb(skb);
1126 error:
1127         return ERR_PTR(err);
1128 }
1129
1130 static int send_frags(struct sk_buff *skb,
1131                       const struct tnl_mutable_config *mutable)
1132 {
1133         int sent_len;
1134
1135         sent_len = 0;
1136         while (skb) {
1137                 struct sk_buff *next = skb->next;
1138                 int frag_len = skb->len - mutable->tunnel_hlen;
1139                 int err;
1140
1141                 skb->next = NULL;
1142                 memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
1143
1144                 err = ip_local_out(skb);
1145                 skb = next;
1146                 if (unlikely(net_xmit_eval(err)))
1147                         goto free_frags;
1148                 sent_len += frag_len;
1149         }
1150
1151         return sent_len;
1152
1153 free_frags:
1154         /*
1155          * There's no point in continuing to send fragments once one has been
1156          * dropped so just free the rest.  This may help improve the congestion
1157          * that caused the first packet to be dropped.
1158          */
1159         tnl_free_linked_skbs(skb);
1160         return sent_len;
1161 }
1162
1163 int tnl_send(struct vport *vport, struct sk_buff *skb)
1164 {
1165         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1166         const struct tnl_mutable_config *mutable = rcu_dereference(tnl_vport->mutable);
1167
1168         enum vport_err_type err = VPORT_E_TX_ERROR;
1169         struct rtable *rt;
1170         struct dst_entry *unattached_dst = NULL;
1171         struct tnl_cache *cache;
1172         int sent_len = 0;
1173         __be16 frag_off = 0;
1174         u8 ttl;
1175         u8 inner_tos;
1176         u8 tos;
1177
1178         /* Validate the protocol headers before we try to use them. */
1179         if (skb->protocol == htons(ETH_P_8021Q) &&
1180             !vlan_tx_tag_present(skb)) {
1181                 if (unlikely(!pskb_may_pull(skb, VLAN_ETH_HLEN)))
1182                         goto error_free;
1183
1184                 skb->protocol = vlan_eth_hdr(skb)->h_vlan_encapsulated_proto;
1185                 skb_set_network_header(skb, VLAN_ETH_HLEN);
1186         }
1187
1188         if (skb->protocol == htons(ETH_P_IP)) {
1189                 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
1190                     + sizeof(struct iphdr))))
1191                         skb->protocol = 0;
1192         }
1193 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1194         else if (skb->protocol == htons(ETH_P_IPV6)) {
1195                 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
1196                     + sizeof(struct ipv6hdr))))
1197                         skb->protocol = 0;
1198         }
1199 #endif
1200
1201         /* ToS */
1202         if (skb->protocol == htons(ETH_P_IP))
1203                 inner_tos = ip_hdr(skb)->tos;
1204 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1205         else if (skb->protocol == htons(ETH_P_IPV6))
1206                 inner_tos = ipv6_get_dsfield(ipv6_hdr(skb));
1207 #endif
1208         else
1209                 inner_tos = 0;
1210
1211         if (mutable->flags & TNL_F_TOS_INHERIT)
1212                 tos = inner_tos;
1213         else
1214                 tos = mutable->tos;
1215
1216         tos = INET_ECN_encapsulate(tos, inner_tos);
1217
1218         /* Route lookup */
1219         rt = find_route(vport, mutable, tos, &cache);
1220         if (unlikely(!rt))
1221                 goto error_free;
1222         if (unlikely(!cache))
1223                 unattached_dst = &rt_dst(rt);
1224
1225         /* Reset SKB */
1226         nf_reset(skb);
1227         secpath_reset(skb);
1228         skb_dst_drop(skb);
1229         skb_clear_rxhash(skb);
1230
1231         /* Offloading */
1232         skb = handle_offloads(skb, mutable, rt);
1233         if (IS_ERR(skb))
1234                 goto error;
1235
1236         /* MTU */
1237         if (unlikely(!check_mtu(skb, vport, mutable, rt, &frag_off))) {
1238                 err = VPORT_E_TX_DROPPED;
1239                 goto error_free;
1240         }
1241
1242         /*
1243          * If we are over the MTU, allow the IP stack to handle fragmentation.
1244          * Fragmentation is a slow path anyways.
1245          */
1246         if (unlikely(skb->len + mutable->tunnel_hlen > dst_mtu(&rt_dst(rt)) &&
1247                      cache)) {
1248                 unattached_dst = &rt_dst(rt);
1249                 dst_hold(unattached_dst);
1250                 cache = NULL;
1251         }
1252
1253         /* TTL */
1254         ttl = mutable->ttl;
1255         if (!ttl)
1256                 ttl = ip4_dst_hoplimit(&rt_dst(rt));
1257
1258         if (mutable->flags & TNL_F_TTL_INHERIT) {
1259                 if (skb->protocol == htons(ETH_P_IP))
1260                         ttl = ip_hdr(skb)->ttl;
1261 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1262                 else if (skb->protocol == htons(ETH_P_IPV6))
1263                         ttl = ipv6_hdr(skb)->hop_limit;
1264 #endif
1265         }
1266
1267         while (skb) {
1268                 struct iphdr *iph;
1269                 struct sk_buff *next_skb = skb->next;
1270                 skb->next = NULL;
1271
1272                 if (unlikely(vlan_deaccel_tag(skb)))
1273                         goto next;
1274
1275                 if (likely(cache)) {
1276                         skb_push(skb, cache->len);
1277                         memcpy(skb->data, get_cached_header(cache), cache->len);
1278                         skb_reset_mac_header(skb);
1279                         skb_set_network_header(skb, rt_dst(rt).hh->hh_len);
1280
1281                 } else {
1282                         skb_push(skb, mutable->tunnel_hlen);
1283                         create_tunnel_header(vport, mutable, rt, skb->data);
1284                         skb_reset_network_header(skb);
1285
1286                         if (next_skb)
1287                                 skb_dst_set(skb, dst_clone(unattached_dst));
1288                         else {
1289                                 skb_dst_set(skb, unattached_dst);
1290                                 unattached_dst = NULL;
1291                         }
1292                 }
1293                 skb_set_transport_header(skb, skb_network_offset(skb) + sizeof(struct iphdr));
1294
1295                 iph = ip_hdr(skb);
1296                 iph->tos = tos;
1297                 iph->ttl = ttl;
1298                 iph->frag_off = frag_off;
1299                 ip_select_ident(iph, &rt_dst(rt), NULL);
1300
1301                 skb = tnl_vport->tnl_ops->update_header(vport, mutable, &rt_dst(rt), skb);
1302                 if (unlikely(!skb))
1303                         goto next;
1304
1305                 if (likely(cache)) {
1306                         int orig_len = skb->len - cache->len;
1307                         struct vport *cache_vport = internal_dev_get_vport(rt_dst(rt).dev);
1308
1309                         skb->protocol = htons(ETH_P_IP);
1310                         iph = ip_hdr(skb);
1311                         iph->tot_len = htons(skb->len - skb_network_offset(skb));
1312                         ip_send_check(iph);
1313
1314                         if (cache_vport) {
1315                                 OVS_CB(skb)->flow = cache->flow;
1316                                 compute_ip_summed(skb, true);
1317                                 vport_receive(cache_vport, skb);
1318                                 sent_len += orig_len;
1319                         } else {
1320                                 int xmit_err;
1321
1322                                 skb->dev = rt_dst(rt).dev;
1323                                 xmit_err = dev_queue_xmit(skb);
1324
1325                                 if (likely(net_xmit_eval(xmit_err) == 0))
1326                                         sent_len += orig_len;
1327                         }
1328                 } else
1329                         sent_len += send_frags(skb, mutable);
1330
1331 next:
1332                 skb = next_skb;
1333         }
1334
1335         if (unlikely(sent_len == 0))
1336                 vport_record_error(vport, VPORT_E_TX_DROPPED);
1337
1338         goto out;
1339
1340 error_free:
1341         tnl_free_linked_skbs(skb);
1342 error:
1343         vport_record_error(vport, err);
1344 out:
1345         dst_release(unattached_dst);
1346         return sent_len;
1347 }
1348
1349 static const struct nla_policy tnl_policy[ODP_TUNNEL_ATTR_MAX + 1] = {
1350         [ODP_TUNNEL_ATTR_FLAGS]    = { .type = NLA_U32 },
1351         [ODP_TUNNEL_ATTR_DST_IPV4] = { .type = NLA_U32 },
1352         [ODP_TUNNEL_ATTR_SRC_IPV4] = { .type = NLA_U32 },
1353         [ODP_TUNNEL_ATTR_OUT_KEY]  = { .type = NLA_U64 },
1354         [ODP_TUNNEL_ATTR_IN_KEY]   = { .type = NLA_U64 },
1355         [ODP_TUNNEL_ATTR_TOS]      = { .type = NLA_U8 },
1356         [ODP_TUNNEL_ATTR_TTL]      = { .type = NLA_U8 },
1357 };
1358
1359 /* Sets ODP_TUNNEL_ATTR_* fields in 'mutable', which must initially be zeroed. */
1360 static int tnl_set_config(struct nlattr *options, const struct tnl_ops *tnl_ops,
1361                           const struct vport *cur_vport,
1362                           struct tnl_mutable_config *mutable)
1363 {
1364         const struct vport *old_vport;
1365         const struct tnl_mutable_config *old_mutable;
1366         struct nlattr *a[ODP_TUNNEL_ATTR_MAX + 1];
1367         int err;
1368
1369         if (!options)
1370                 return -EINVAL;
1371
1372         err = nla_parse_nested(a, ODP_TUNNEL_ATTR_MAX, options, tnl_policy);
1373         if (err)
1374                 return err;
1375
1376         if (!a[ODP_TUNNEL_ATTR_FLAGS] || !a[ODP_TUNNEL_ATTR_DST_IPV4])
1377                 return -EINVAL;
1378
1379         mutable->flags = nla_get_u32(a[ODP_TUNNEL_ATTR_FLAGS]) & TNL_F_PUBLIC;
1380
1381         if (a[ODP_TUNNEL_ATTR_SRC_IPV4])
1382                 mutable->saddr = nla_get_be32(a[ODP_TUNNEL_ATTR_SRC_IPV4]);
1383         mutable->daddr = nla_get_be32(a[ODP_TUNNEL_ATTR_DST_IPV4]);
1384
1385         if (a[ODP_TUNNEL_ATTR_TOS]) {
1386                 mutable->tos = nla_get_u8(a[ODP_TUNNEL_ATTR_TOS]);
1387                 if (mutable->tos != RT_TOS(mutable->tos))
1388                         return -EINVAL;
1389         }
1390
1391         if (a[ODP_TUNNEL_ATTR_TTL])
1392                 mutable->ttl = nla_get_u8(a[ODP_TUNNEL_ATTR_TTL]);
1393
1394         mutable->tunnel_type = tnl_ops->tunnel_type;
1395         if (!a[ODP_TUNNEL_ATTR_IN_KEY]) {
1396                 mutable->tunnel_type |= TNL_T_KEY_MATCH;
1397                 mutable->flags |= TNL_F_IN_KEY_MATCH;
1398         } else {
1399                 mutable->tunnel_type |= TNL_T_KEY_EXACT;
1400                 mutable->in_key = nla_get_be64(a[ODP_TUNNEL_ATTR_IN_KEY]);
1401         }
1402
1403         if (!a[ODP_TUNNEL_ATTR_OUT_KEY])
1404                 mutable->flags |= TNL_F_OUT_KEY_ACTION;
1405         else
1406                 mutable->out_key = nla_get_be64(a[ODP_TUNNEL_ATTR_OUT_KEY]);
1407
1408         mutable->tunnel_hlen = tnl_ops->hdr_len(mutable);
1409         if (mutable->tunnel_hlen < 0)
1410                 return mutable->tunnel_hlen;
1411
1412         mutable->tunnel_hlen += sizeof(struct iphdr);
1413
1414         old_vport = tnl_find_port(mutable->saddr, mutable->daddr,
1415                                   mutable->in_key, mutable->tunnel_type,
1416                                   &old_mutable);
1417
1418         if (old_vport && old_vport != cur_vport)
1419                 return -EEXIST;
1420
1421         return 0;
1422 }
1423
1424 struct vport *tnl_create(const struct vport_parms *parms,
1425                          const struct vport_ops *vport_ops,
1426                          const struct tnl_ops *tnl_ops)
1427 {
1428         struct vport *vport;
1429         struct tnl_vport *tnl_vport;
1430         struct tnl_mutable_config *mutable;
1431         int initial_frag_id;
1432         int err;
1433
1434         vport = vport_alloc(sizeof(struct tnl_vport), vport_ops, parms);
1435         if (IS_ERR(vport)) {
1436                 err = PTR_ERR(vport);
1437                 goto error;
1438         }
1439
1440         tnl_vport = tnl_vport_priv(vport);
1441
1442         strcpy(tnl_vport->name, parms->name);
1443         tnl_vport->tnl_ops = tnl_ops;
1444
1445         mutable = kzalloc(sizeof(struct tnl_mutable_config), GFP_KERNEL);
1446         if (!mutable) {
1447                 err = -ENOMEM;
1448                 goto error_free_vport;
1449         }
1450
1451         vport_gen_rand_ether_addr(mutable->eth_addr);
1452
1453         get_random_bytes(&initial_frag_id, sizeof(int));
1454         atomic_set(&tnl_vport->frag_id, initial_frag_id);
1455
1456         err = tnl_set_config(parms->options, tnl_ops, NULL, mutable);
1457         if (err)
1458                 goto error_free_mutable;
1459
1460         spin_lock_init(&tnl_vport->cache_lock);
1461
1462 #ifdef NEED_CACHE_TIMEOUT
1463         tnl_vport->cache_exp_interval = MAX_CACHE_EXP -
1464                                        (net_random() % (MAX_CACHE_EXP / 2));
1465 #endif
1466
1467         rcu_assign_pointer(tnl_vport->mutable, mutable);
1468
1469         err = add_port(vport);
1470         if (err)
1471                 goto error_free_mutable;
1472
1473         return vport;
1474
1475 error_free_mutable:
1476         kfree(mutable);
1477 error_free_vport:
1478         vport_free(vport);
1479 error:
1480         return ERR_PTR(err);
1481 }
1482
1483 int tnl_set_options(struct vport *vport, struct nlattr *options)
1484 {
1485         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1486         const struct tnl_mutable_config *old_mutable;
1487         struct tnl_mutable_config *mutable;
1488         int err;
1489
1490         mutable = kzalloc(sizeof(struct tnl_mutable_config), GFP_KERNEL);
1491         if (!mutable) {
1492                 err = -ENOMEM;
1493                 goto error;
1494         }
1495
1496         /* Copy fields whose values should be retained. */
1497         old_mutable = rtnl_dereference(tnl_vport->mutable);
1498         mutable->seq = old_mutable->seq + 1;
1499         memcpy(mutable->eth_addr, old_mutable->eth_addr, ETH_ALEN);
1500
1501         /* Parse the others configured by userspace. */
1502         err = tnl_set_config(options, tnl_vport->tnl_ops, vport, mutable);
1503         if (err)
1504                 goto error_free;
1505
1506         err = move_port(vport, mutable);
1507         if (err)
1508                 goto error_free;
1509
1510         return 0;
1511
1512 error_free:
1513         kfree(mutable);
1514 error:
1515         return err;
1516 }
1517
1518 int tnl_get_options(const struct vport *vport, struct sk_buff *skb)
1519 {
1520         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1521         const struct tnl_mutable_config *mutable = rcu_dereference_rtnl(tnl_vport->mutable);
1522
1523         NLA_PUT_U32(skb, ODP_TUNNEL_ATTR_FLAGS, mutable->flags & TNL_F_PUBLIC);
1524         NLA_PUT_BE32(skb, ODP_TUNNEL_ATTR_DST_IPV4, mutable->daddr);
1525
1526         if (!(mutable->flags & TNL_F_IN_KEY_MATCH))
1527                 NLA_PUT_BE64(skb, ODP_TUNNEL_ATTR_IN_KEY, mutable->in_key);
1528         if (!(mutable->flags & TNL_F_OUT_KEY_ACTION))
1529                 NLA_PUT_BE64(skb, ODP_TUNNEL_ATTR_OUT_KEY, mutable->out_key);
1530         if (mutable->saddr)
1531                 NLA_PUT_BE32(skb, ODP_TUNNEL_ATTR_SRC_IPV4, mutable->saddr);
1532         if (mutable->tos)
1533                 NLA_PUT_U8(skb, ODP_TUNNEL_ATTR_TOS, mutable->tos);
1534         if (mutable->ttl)
1535                 NLA_PUT_U8(skb, ODP_TUNNEL_ATTR_TTL, mutable->ttl);
1536
1537         return 0;
1538
1539 nla_put_failure:
1540         return -EMSGSIZE;
1541 }
1542
1543 static void free_port_rcu(struct rcu_head *rcu)
1544 {
1545         struct tnl_vport *tnl_vport = container_of(rcu,
1546                                                    struct tnl_vport, rcu);
1547
1548         free_cache((struct tnl_cache __force *)tnl_vport->cache);
1549         kfree((struct tnl_mutable __force *)tnl_vport->mutable);
1550         vport_free(tnl_vport_to_vport(tnl_vport));
1551 }
1552
1553 int tnl_destroy(struct vport *vport)
1554 {
1555         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1556         const struct tnl_mutable_config *mutable, *old_mutable;
1557
1558         mutable = rtnl_dereference(tnl_vport->mutable);
1559
1560         if (vport == tnl_find_port(mutable->saddr, mutable->daddr,
1561                                    mutable->in_key, mutable->tunnel_type,
1562                                    &old_mutable))
1563                 del_port(vport);
1564
1565         call_rcu(&tnl_vport->rcu, free_port_rcu);
1566
1567         return 0;
1568 }
1569
1570 int tnl_set_addr(struct vport *vport, const unsigned char *addr)
1571 {
1572         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1573         struct tnl_mutable_config *mutable;
1574
1575         mutable = kmemdup(rtnl_dereference(tnl_vport->mutable),
1576                           sizeof(struct tnl_mutable_config), GFP_KERNEL);
1577         if (!mutable)
1578                 return -ENOMEM;
1579
1580         memcpy(mutable->eth_addr, addr, ETH_ALEN);
1581         assign_config_rcu(vport, mutable);
1582
1583         return 0;
1584 }
1585
1586 const char *tnl_get_name(const struct vport *vport)
1587 {
1588         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1589         return tnl_vport->name;
1590 }
1591
1592 const unsigned char *tnl_get_addr(const struct vport *vport)
1593 {
1594         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1595         return rcu_dereference_rtnl(tnl_vport->mutable)->eth_addr;
1596 }
1597
1598 void tnl_free_linked_skbs(struct sk_buff *skb)
1599 {
1600         while (skb) {
1601                 struct sk_buff *next = skb->next;
1602                 kfree_skb(skb);
1603                 skb = next;
1604         }
1605 }