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  55  * Interval to check for and remove caches that are no longer valid.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         vlan_set_tci(skb, 0);
 454
 455         if (unlikely(compute_ip_summed(skb, false))) {
 456                 kfree_skb(skb);
 457                 return;
 458         }
 459
 460         vport_receive(vport, skb);
 461 }
 462
 463 static bool check_ipv4_address(__be32 addr)
 464 {
 465         if (ipv4_is_multicast(addr) || ipv4_is_lbcast(addr)
 466             || ipv4_is_loopback(addr) || ipv4_is_zeronet(addr))
 467                 return false;
 468
 469         return true;
 470 }
 471
 472 static bool ipv4_should_icmp(struct sk_buff *skb)
 473 {
 474         struct iphdr *old_iph = ip_hdr(skb);
 475
 476         /* Don't respond to L2 broadcast. */
 477         if (is_multicast_ether_addr(eth_hdr(skb)->h_dest))
 478                 return false;
 479
 480         /* Don't respond to L3 broadcast or invalid addresses. */
 481         if (!check_ipv4_address(old_iph->daddr) ||
 482             !check_ipv4_address(old_iph->saddr))
 483                 return false;
 484
 485         /* Only respond to the first fragment. */
 486         if (old_iph->frag_off & htons(IP_OFFSET))
 487                 return false;
 488
 489         /* Don't respond to ICMP error messages. */
 490         if (old_iph->protocol == IPPROTO_ICMP) {
 491                 u8 icmp_type, *icmp_typep;
 492
 493                 icmp_typep = skb_header_pointer(skb, (u8 *)old_iph +
 494                                                 (old_iph->ihl << 2) +
 495                                                 offsetof(struct icmphdr, type) -
 496                                                 skb->data, sizeof(icmp_type),
 497                                                 &icmp_type);
 498
 499                 if (!icmp_typep)
 500                         return false;
 501
 502                 if (*icmp_typep > NR_ICMP_TYPES
 503                         || (*icmp_typep <= ICMP_PARAMETERPROB
 504                                 && *icmp_typep != ICMP_ECHOREPLY
 505                                 && *icmp_typep != ICMP_ECHO))
 506                         return false;
 507         }
 508
 509         return true;
 510 }
 511
 512 static void ipv4_build_icmp(struct sk_buff *skb, struct sk_buff *nskb,
 513                             unsigned int mtu, unsigned int payload_length)
 514 {
 515         struct iphdr *iph, *old_iph = ip_hdr(skb);
 516         struct icmphdr *icmph;
 517         u8 *payload;
 518
 519         iph = (struct iphdr *)skb_put(nskb, sizeof(struct iphdr));
 520         icmph = (struct icmphdr *)skb_put(nskb, sizeof(struct icmphdr));
 521         payload = skb_put(nskb, payload_length);
 522
 523         /* IP */
 524         iph->version            =       4;
 525         iph->ihl                =       sizeof(struct iphdr) >> 2;
 526         iph->tos                =       (old_iph->tos & IPTOS_TOS_MASK) |
 527                                         IPTOS_PREC_INTERNETCONTROL;
 528         iph->tot_len            =       htons(sizeof(struct iphdr)
 529                                               + sizeof(struct icmphdr)
 530                                               + payload_length);
 531         get_random_bytes(&iph->id, sizeof(iph->id));
 532         iph->frag_off           =       0;
 533         iph->ttl                =       IPDEFTTL;
 534         iph->protocol           =       IPPROTO_ICMP;
 535         iph->daddr              =       old_iph->saddr;
 536         iph->saddr              =       old_iph->daddr;
 537
 538         ip_send_check(iph);
 539
 540         /* ICMP */
 541         icmph->type             =       ICMP_DEST_UNREACH;
 542         icmph->code             =       ICMP_FRAG_NEEDED;
 543         icmph->un.gateway       =       htonl(mtu);
 544         icmph->checksum         =       0;
 545
 546         nskb->csum = csum_partial((u8 *)icmph, sizeof(struct icmphdr), 0);
 547         nskb->csum = skb_copy_and_csum_bits(skb, (u8 *)old_iph - skb->data,
 548                                             payload, payload_length,
 549                                             nskb->csum);
 550         icmph->checksum = csum_fold(nskb->csum);
 551 }
 552
 553 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 554 static bool ipv6_should_icmp(struct sk_buff *skb)
 555 {
 556         struct ipv6hdr *old_ipv6h = ipv6_hdr(skb);
 557         int addr_type;
 558         int payload_off = (u8 *)(old_ipv6h + 1) - skb->data;
 559         u8 nexthdr = ipv6_hdr(skb)->nexthdr;
 560
 561         /* Check source address is valid. */
 562         addr_type = ipv6_addr_type(&old_ipv6h->saddr);
 563         if (addr_type & IPV6_ADDR_MULTICAST || addr_type == IPV6_ADDR_ANY)
 564                 return false;
 565
 566         /* Don't reply to unspecified addresses. */
 567         if (ipv6_addr_type(&old_ipv6h->daddr) == IPV6_ADDR_ANY)
 568                 return false;
 569
 570         /* Don't respond to ICMP error messages. */
 571         payload_off = ipv6_skip_exthdr(skb, payload_off, &nexthdr);
 572         if (payload_off < 0)
 573                 return false;
 574
 575         if (nexthdr == NEXTHDR_ICMP) {
 576                 u8 icmp_type, *icmp_typep;
 577
 578                 icmp_typep = skb_header_pointer(skb, payload_off +
 579                                                 offsetof(struct icmp6hdr,
 580                                                         icmp6_type),
 581                                                 sizeof(icmp_type), &icmp_type);
 582
 583                 if (!icmp_typep || !(*icmp_typep & ICMPV6_INFOMSG_MASK))
 584                         return false;
 585         }
 586
 587         return true;
 588 }
 589
 590 static void ipv6_build_icmp(struct sk_buff *skb, struct sk_buff *nskb,
 591                             unsigned int mtu, unsigned int payload_length)
 592 {
 593         struct ipv6hdr *ipv6h, *old_ipv6h = ipv6_hdr(skb);
 594         struct icmp6hdr *icmp6h;
 595         u8 *payload;
 596
 597         ipv6h = (struct ipv6hdr *)skb_put(nskb, sizeof(struct ipv6hdr));
 598         icmp6h = (struct icmp6hdr *)skb_put(nskb, sizeof(struct icmp6hdr));
 599         payload = skb_put(nskb, payload_length);
 600
 601         /* IPv6 */
 602         ipv6h->version          =       6;
 603         ipv6h->priority         =       0;
 604         memset(&ipv6h->flow_lbl, 0, sizeof(ipv6h->flow_lbl));
 605         ipv6h->payload_len      =       htons(sizeof(struct icmp6hdr)
 606                                               + payload_length);
 607         ipv6h->nexthdr          =       NEXTHDR_ICMP;
 608         ipv6h->hop_limit        =       IPV6_DEFAULT_HOPLIMIT;
 609         ipv6_addr_copy(&ipv6h->daddr, &old_ipv6h->saddr);
 610         ipv6_addr_copy(&ipv6h->saddr, &old_ipv6h->daddr);
 611
 612         /* ICMPv6 */
 613         icmp6h->icmp6_type      =       ICMPV6_PKT_TOOBIG;
 614         icmp6h->icmp6_code      =       0;
 615         icmp6h->icmp6_cksum     =       0;
 616         icmp6h->icmp6_mtu       =       htonl(mtu);
 617
 618         nskb->csum = csum_partial((u8 *)icmp6h, sizeof(struct icmp6hdr), 0);
 619         nskb->csum = skb_copy_and_csum_bits(skb, (u8 *)old_ipv6h - skb->data,
 620                                             payload, payload_length,
 621                                             nskb->csum);
 622         icmp6h->icmp6_cksum = csum_ipv6_magic(&ipv6h->saddr, &ipv6h->daddr,
 623                                                 sizeof(struct icmp6hdr)
 624                                                 + payload_length,
 625                                                 ipv6h->nexthdr, nskb->csum);
 626 }
 627 #endif /* IPv6 */
 628
 629 bool tnl_frag_needed(struct vport *vport, const struct tnl_mutable_config *mutable,
 630                      struct sk_buff *skb, unsigned int mtu, __be64 flow_key)
 631 {
 632         unsigned int eth_hdr_len = ETH_HLEN;
 633         unsigned int total_length = 0, header_length = 0, payload_length;
 634         struct ethhdr *eh, *old_eh = eth_hdr(skb);
 635         struct sk_buff *nskb;
 636
 637         /* Sanity check */
 638         if (skb->protocol == htons(ETH_P_IP)) {
 639                 if (mtu < IP_MIN_MTU)
 640                         return false;
 641
 642                 if (!ipv4_should_icmp(skb))
 643                         return true;
 644         }
 645 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 646         else if (skb->protocol == htons(ETH_P_IPV6)) {
 647                 if (mtu < IPV6_MIN_MTU)
 648                         return false;
 649
 650                 /*
 651                  * In theory we should do PMTUD on IPv6 multicast messages but
 652                  * we don't have an address to send from so just fragment.
 653                  */
 654                 if (ipv6_addr_type(&ipv6_hdr(skb)->daddr) & IPV6_ADDR_MULTICAST)
 655                         return false;
 656
 657                 if (!ipv6_should_icmp(skb))
 658                         return true;
 659         }
 660 #endif
 661         else
 662                 return false;
 663
 664         /* Allocate */
 665         if (old_eh->h_proto == htons(ETH_P_8021Q))
 666                 eth_hdr_len = VLAN_ETH_HLEN;
 667
 668         payload_length = skb->len - eth_hdr_len;
 669         if (skb->protocol == htons(ETH_P_IP)) {
 670                 header_length = sizeof(struct iphdr) + sizeof(struct icmphdr);
 671                 total_length = min_t(unsigned int, header_length +
 672                                                    payload_length, 576);
 673         }
 674 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 675         else {
 676                 header_length = sizeof(struct ipv6hdr) +
 677                                 sizeof(struct icmp6hdr);
 678                 total_length = min_t(unsigned int, header_length +
 679                                                   payload_length, IPV6_MIN_MTU);
 680         }
 681 #endif
 682
 683         payload_length = total_length - header_length;
 684
 685         nskb = dev_alloc_skb(NET_IP_ALIGN + eth_hdr_len + header_length +
 686                              payload_length);
 687         if (!nskb)
 688                 return false;
 689
 690         skb_reserve(nskb, NET_IP_ALIGN);
 691
 692         /* Ethernet / VLAN */
 693         eh = (struct ethhdr *)skb_put(nskb, eth_hdr_len);
 694         memcpy(eh->h_dest, old_eh->h_source, ETH_ALEN);
 695         memcpy(eh->h_source, mutable->eth_addr, ETH_ALEN);
 696         nskb->protocol = eh->h_proto = old_eh->h_proto;
 697         if (old_eh->h_proto == htons(ETH_P_8021Q)) {
 698                 struct vlan_ethhdr *vh = (struct vlan_ethhdr *)eh;
 699
 700                 vh->h_vlan_TCI = vlan_eth_hdr(skb)->h_vlan_TCI;
 701                 vh->h_vlan_encapsulated_proto = skb->protocol;
 702         } else
 703                 vlan_set_tci(nskb, vlan_get_tci(skb));
 704         skb_reset_mac_header(nskb);
 705
 706         /* Protocol */
 707         if (skb->protocol == htons(ETH_P_IP))
 708                 ipv4_build_icmp(skb, nskb, mtu, payload_length);
 709 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 710         else
 711                 ipv6_build_icmp(skb, nskb, mtu, payload_length);
 712 #endif
 713
 714         /*
 715          * Assume that flow based keys are symmetric with respect to input
 716          * and output and use the key that we were going to put on the
 717          * outgoing packet for the fake received packet.  If the keys are
 718          * not symmetric then PMTUD needs to be disabled since we won't have
 719          * any way of synthesizing packets.
 720          */
 721         if ((mutable->flags & (TNL_F_IN_KEY_MATCH | TNL_F_OUT_KEY_ACTION)) ==
 722             (TNL_F_IN_KEY_MATCH | TNL_F_OUT_KEY_ACTION))
 723                 OVS_CB(nskb)->tun_id = flow_key;
 724
 725         if (unlikely(compute_ip_summed(nskb, false))) {
 726                 kfree_skb(nskb);
 727                 return false;
 728         }
 729
 730         vport_receive(vport, nskb);
 731
 732         return true;
 733 }
 734
 735 static bool check_mtu(struct sk_buff *skb,
 736                       struct vport *vport,
 737                       const struct tnl_mutable_config *mutable,
 738                       const struct rtable *rt, __be16 *frag_offp)
 739 {
 740         bool df_inherit = mutable->flags & TNL_F_DF_INHERIT;
 741         bool pmtud = mutable->flags & TNL_F_PMTUD;
 742         __be16 frag_off = mutable->flags & TNL_F_DF_DEFAULT ? htons(IP_DF) : 0;
 743         int mtu = 0;
 744         unsigned int packet_length = skb->len - ETH_HLEN;
 745
 746         /* Allow for one level of tagging in the packet length. */
 747         if (!vlan_tx_tag_present(skb) &&
 748             eth_hdr(skb)->h_proto == htons(ETH_P_8021Q))
 749                 packet_length -= VLAN_HLEN;
 750
 751         if (pmtud) {
 752                 int vlan_header = 0;
 753
 754                 /* The tag needs to go in packet regardless of where it
 755                  * currently is, so subtract it from the MTU.
 756                  */
 757                 if (vlan_tx_tag_present(skb) ||
 758                     eth_hdr(skb)->h_proto == htons(ETH_P_8021Q))
 759                         vlan_header = VLAN_HLEN;
 760
 761                 mtu = dst_mtu(&rt_dst(rt))
 762                         - ETH_HLEN
 763                         - mutable->tunnel_hlen
 764                         - vlan_header;
 765         }
 766
 767         if (skb->protocol == htons(ETH_P_IP)) {
 768                 struct iphdr *iph = ip_hdr(skb);
 769
 770                 if (df_inherit)
 771                         frag_off = iph->frag_off & htons(IP_DF);
 772
 773                 if (pmtud && iph->frag_off & htons(IP_DF)) {
 774                         mtu = max(mtu, IP_MIN_MTU);
 775
 776                         if (packet_length > mtu &&
 777                             tnl_frag_needed(vport, mutable, skb, mtu,
 778                                             OVS_CB(skb)->tun_id))
 779                                 return false;
 780                 }
 781         }
 782 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 783         else if (skb->protocol == htons(ETH_P_IPV6)) {
 784                 /* IPv6 requires end hosts to do fragmentation
 785                  * if the packet is above the minimum MTU.
 786                  */
 787                 if (df_inherit && packet_length > IPV6_MIN_MTU)
 788                         frag_off = htons(IP_DF);
 789
 790                 if (pmtud) {
 791                         mtu = max(mtu, IPV6_MIN_MTU);
 792
 793                         if (packet_length > mtu &&
 794                             tnl_frag_needed(vport, mutable, skb, mtu,
 795                                             OVS_CB(skb)->tun_id))
 796                                 return false;
 797                 }
 798         }
 799 #endif
 800
 801         *frag_offp = frag_off;
 802         return true;
 803 }
 804
 805 static void create_tunnel_header(const struct vport *vport,
 806                                  const struct tnl_mutable_config *mutable,
 807                                  const struct rtable *rt, void *header)
 808 {
 809         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 810         struct iphdr *iph = header;
 811
 812         iph->version    = 4;
 813         iph->ihl        = sizeof(struct iphdr) >> 2;
 814         iph->frag_off   = htons(IP_DF);
 815         iph->protocol   = tnl_vport->tnl_ops->ipproto;
 816         iph->tos        = mutable->tos;
 817         iph->daddr      = rt->rt_dst;
 818         iph->saddr      = rt->rt_src;
 819         iph->ttl        = mutable->ttl;
 820         if (!iph->ttl)
 821                 iph->ttl = ip4_dst_hoplimit(&rt_dst(rt));
 822
 823         tnl_vport->tnl_ops->build_header(vport, mutable, iph + 1);
 824 }
 825
 826 static inline void *get_cached_header(const struct tnl_cache *cache)
 827 {
 828         return (void *)cache + ALIGN(sizeof(struct tnl_cache), CACHE_DATA_ALIGN);
 829 }
 830
 831 static inline bool check_cache_valid(const struct tnl_cache *cache,
 832                                      const struct tnl_mutable_config *mutable)
 833 {
 834         return cache &&
 835 #ifdef NEED_CACHE_TIMEOUT
 836                 time_before(jiffies, cache->expiration) &&
 837 #endif
 838 #ifdef HAVE_RT_GENID
 839                 atomic_read(&init_net.ipv4.rt_genid) == cache->rt->rt_genid &&
 840 #endif
 841 #ifdef HAVE_HH_SEQ
 842                 rt_dst(cache->rt).hh->hh_lock.sequence == cache->hh_seq &&
 843 #endif
 844                 mutable->seq == cache->mutable_seq &&
 845                 (!is_internal_dev(rt_dst(cache->rt).dev) ||
 846                 (cache->flow && !cache->flow->dead));
 847 }
 848
 849 static int cache_cleaner_cb(struct tbl_node *tbl_node, void *aux)
 850 {
 851         struct tnl_vport *tnl_vport = tnl_vport_table_cast(tbl_node);
 852         const struct tnl_mutable_config *mutable = rcu_dereference(tnl_vport->mutable);
 853         const struct tnl_cache *cache = rcu_dereference(tnl_vport->cache);
 854
 855         if (cache && !check_cache_valid(cache, mutable) &&
 856             spin_trylock_bh(&tnl_vport->cache_lock)) {
 857                 assign_cache_rcu(tnl_vport_to_vport(tnl_vport), NULL);
 858                 spin_unlock_bh(&tnl_vport->cache_lock);
 859         }
 860
 861         return 0;
 862 }
 863
 864 static void cache_cleaner(struct work_struct *work)
 865 {
 866         schedule_cache_cleaner();
 867
 868         rcu_read_lock();
 869         tbl_foreach(rcu_dereference(port_table), cache_cleaner_cb, NULL);
 870         rcu_read_unlock();
 871 }
 872
 873 static inline void create_eth_hdr(struct tnl_cache *cache,
 874                                   const struct rtable *rt)
 875 {
 876         void *cache_data = get_cached_header(cache);
 877         int hh_len = rt_dst(rt).hh->hh_len;
 878         int hh_off = HH_DATA_ALIGN(rt_dst(rt).hh->hh_len) - hh_len;
 879
 880 #ifdef HAVE_HH_SEQ
 881         unsigned hh_seq;
 882
 883         do {
 884                 hh_seq = read_seqbegin(&rt_dst(rt).hh->hh_lock);
 885                 memcpy(cache_data, (void *)rt_dst(rt).hh->hh_data + hh_off, hh_len);
 886         } while (read_seqretry(&rt_dst(rt).hh->hh_lock, hh_seq));
 887
 888         cache->hh_seq = hh_seq;
 889 #else
 890         read_lock_bh(&rt_dst(rt).hh->hh_lock);
 891         memcpy(cache_data, (void *)rt_dst(rt).hh->hh_data + hh_off, hh_len);
 892         read_unlock_bh(&rt_dst(rt).hh->hh_lock);
 893 #endif
 894 }
 895
 896 static struct tnl_cache *build_cache(struct vport *vport,
 897                                      const struct tnl_mutable_config *mutable,
 898                                      struct rtable *rt)
 899 {
 900         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 901         struct tnl_cache *cache;
 902         void *cache_data;
 903         int cache_len;
 904
 905         if (!(mutable->flags & TNL_F_HDR_CACHE))
 906                 return NULL;
 907
 908         /*
 909          * If there is no entry in the ARP cache or if this device does not
 910          * support hard header caching just fall back to the IP stack.
 911          */
 912         if (!rt_dst(rt).hh)
 913                 return NULL;
 914
 915         /*
 916          * If lock is contended fall back to directly building the header.
 917          * We're not going to help performance by sitting here spinning.
 918          */
 919         if (!spin_trylock_bh(&tnl_vport->cache_lock))
 920                 return NULL;
 921
 922         cache = cache_dereference(tnl_vport);
 923         if (check_cache_valid(cache, mutable))
 924                 goto unlock;
 925         else
 926                 cache = NULL;
 927
 928         cache_len = rt_dst(rt).hh->hh_len + mutable->tunnel_hlen;
 929
 930         cache = kzalloc(ALIGN(sizeof(struct tnl_cache), CACHE_DATA_ALIGN) +
 931                         cache_len, GFP_ATOMIC);
 932         if (!cache)
 933                 goto unlock;
 934
 935         cache->len = cache_len;
 936
 937         create_eth_hdr(cache, rt);
 938         cache_data = get_cached_header(cache) + rt_dst(rt).hh->hh_len;
 939
 940         create_tunnel_header(vport, mutable, rt, cache_data);
 941
 942         cache->mutable_seq = mutable->seq;
 943         cache->rt = rt;
 944 #ifdef NEED_CACHE_TIMEOUT
 945         cache->expiration = jiffies + tnl_vport->cache_exp_interval;
 946 #endif
 947
 948         if (is_internal_dev(rt_dst(rt).dev)) {
 949                 struct sw_flow_key flow_key;
 950                 struct tbl_node *flow_node;
 951                 struct vport *dst_vport;
 952                 struct sk_buff *skb;
 953                 bool is_frag;
 954                 int err;
 955                 int flow_key_len;
 956
 957                 dst_vport = internal_dev_get_vport(rt_dst(rt).dev);
 958                 if (!dst_vport)
 959                         goto done;
 960
 961                 skb = alloc_skb(cache->len, GFP_ATOMIC);
 962                 if (!skb)
 963                         goto done;
 964
 965                 __skb_put(skb, cache->len);
 966                 memcpy(skb->data, get_cached_header(cache), cache->len);
 967
 968                 err = flow_extract(skb, dst_vport->port_no, &flow_key,
 969                                    &flow_key_len, &is_frag);
 970
 971                 kfree_skb(skb);
 972                 if (err || is_frag)
 973                         goto done;
 974
 975                 flow_node = tbl_lookup(rcu_dereference(dst_vport->dp->table),
 976                                        &flow_key, flow_key_len,
 977                                        flow_hash(&flow_key, flow_key_len),
 978                                        flow_cmp);
 979                 if (flow_node) {
 980                         struct sw_flow *flow = flow_cast(flow_node);
 981
 982                         cache->flow = flow;
 983                         flow_hold(flow);
 984                 }
 985         }
 986
 987 done:
 988         assign_cache_rcu(vport, cache);
 989
 990 unlock:
 991         spin_unlock_bh(&tnl_vport->cache_lock);
 992
 993         return cache;
 994 }
 995
 996 static struct rtable *find_route(struct vport *vport,
 997                                  const struct tnl_mutable_config *mutable,
 998                                  u8 tos, struct tnl_cache **cache)
 999 {
1000         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1001         struct tnl_cache *cur_cache = rcu_dereference(tnl_vport->cache);
1002
1003         *cache = NULL;
1004         tos = RT_TOS(tos);
1005
1006         if (likely(tos == mutable->tos && check_cache_valid(cur_cache, mutable))) {
1007                 *cache = cur_cache;
1008                 return cur_cache->rt;
1009         } else {
1010                 struct rtable *rt;
1011 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,39)
1012                 struct flowi fl = { .nl_u = { .ip4_u =
1013                                               { .daddr = mutable->daddr,
1014                                                 .saddr = mutable->saddr,
1015                                                 .tos = tos } },
1016                                     .proto = tnl_vport->tnl_ops->ipproto };
1017
1018                 if (unlikely(ip_route_output_key(&init_net, &rt, &fl)))
1019                         return NULL;
1020 #else
1021                 struct flowi4 fl = { .daddr = mutable->daddr,
1022                                      .saddr = mutable->saddr,
1023                                      .flowi4_tos = tos,
1024                                      .flowi4_proto = tnl_vport->tnl_ops->ipproto };
1025
1026                 rt = ip_route_output_key(&init_net, &fl);
1027                 if (IS_ERR(rt))
1028                         return NULL;
1029 #endif
1030
1031                 if (likely(tos == mutable->tos))
1032                         *cache = build_cache(vport, mutable, rt);
1033
1034                 return rt;
1035         }
1036 }
1037
1038 static inline bool need_linearize(const struct sk_buff *skb)
1039 {
1040         int i;
1041
1042         if (unlikely(skb_shinfo(skb)->frag_list))
1043                 return true;
1044
1045         /*
1046          * Generally speaking we should linearize if there are paged frags.
1047          * However, if all of the refcounts are 1 we know nobody else can
1048          * change them from underneath us and we can skip the linearization.
1049          */
1050         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
1051                 if (unlikely(page_count(skb_shinfo(skb)->frags[i].page) > 1))
1052                         return true;
1053
1054         return false;
1055 }
1056
1057 static struct sk_buff *handle_offloads(struct sk_buff *skb,
1058                                        const struct tnl_mutable_config *mutable,
1059                                        const struct rtable *rt)
1060 {
1061         int min_headroom;
1062         int err;
1063
1064         min_headroom = LL_RESERVED_SPACE(rt_dst(rt).dev) + rt_dst(rt).header_len
1065                         + mutable->tunnel_hlen
1066                         + (vlan_tx_tag_present(skb) ? VLAN_HLEN : 0);
1067
1068         if (skb_headroom(skb) < min_headroom || skb_header_cloned(skb)) {
1069                 int head_delta = SKB_DATA_ALIGN(min_headroom -
1070                                                 skb_headroom(skb) +
1071                                                 16);
1072                 err = pskb_expand_head(skb, max_t(int, head_delta, 0),
1073                                         0, GFP_ATOMIC);
1074                 if (unlikely(err))
1075                         goto error_free;
1076         }
1077
1078         forward_ip_summed(skb, true);
1079
1080         if (skb_is_gso(skb)) {
1081                 struct sk_buff *nskb;
1082
1083                 nskb = skb_gso_segment(skb, 0);
1084                 kfree_skb(skb);
1085                 if (IS_ERR(nskb)) {
1086                         err = PTR_ERR(nskb);
1087                         goto error;
1088                 }
1089
1090                 skb = nskb;
1091         } else if (get_ip_summed(skb) == OVS_CSUM_PARTIAL) {
1092                 /* Pages aren't locked and could change at any time.
1093                  * If this happens after we compute the checksum, the
1094                  * checksum will be wrong.  We linearize now to avoid
1095                  * this problem.
1096                  */
1097                 if (unlikely(need_linearize(skb))) {
1098                         err = __skb_linearize(skb);
1099                         if (unlikely(err))
1100                                 goto error_free;
1101                 }
1102
1103                 err = skb_checksum_help(skb);
1104                 if (unlikely(err))
1105                         goto error_free;
1106         }
1107
1108         set_ip_summed(skb, OVS_CSUM_NONE);
1109
1110         return skb;
1111
1112 error_free:
1113         kfree_skb(skb);
1114 error:
1115         return ERR_PTR(err);
1116 }
1117
1118 static int send_frags(struct sk_buff *skb,
1119                       const struct tnl_mutable_config *mutable)
1120 {
1121         int sent_len;
1122
1123         sent_len = 0;
1124         while (skb) {
1125                 struct sk_buff *next = skb->next;
1126                 int frag_len = skb->len - mutable->tunnel_hlen;
1127                 int err;
1128
1129                 skb->next = NULL;
1130                 memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
1131
1132                 err = ip_local_out(skb);
1133                 skb = next;
1134                 if (unlikely(net_xmit_eval(err)))
1135                         goto free_frags;
1136                 sent_len += frag_len;
1137         }
1138
1139         return sent_len;
1140
1141 free_frags:
1142         /*
1143          * There's no point in continuing to send fragments once one has been
1144          * dropped so just free the rest.  This may help improve the congestion
1145          * that caused the first packet to be dropped.
1146          */
1147         tnl_free_linked_skbs(skb);
1148         return sent_len;
1149 }
1150
1151 int tnl_send(struct vport *vport, struct sk_buff *skb)
1152 {
1153         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1154         const struct tnl_mutable_config *mutable = rcu_dereference(tnl_vport->mutable);
1155
1156         enum vport_err_type err = VPORT_E_TX_ERROR;
1157         struct rtable *rt;
1158         struct dst_entry *unattached_dst = NULL;
1159         struct tnl_cache *cache;
1160         int sent_len = 0;
1161         __be16 frag_off = 0;
1162         u8 ttl;
1163         u8 inner_tos;
1164         u8 tos;
1165
1166         /* Validate the protocol headers before we try to use them. */
1167         if (skb->protocol == htons(ETH_P_8021Q) &&
1168             !vlan_tx_tag_present(skb)) {
1169                 if (unlikely(!pskb_may_pull(skb, VLAN_ETH_HLEN)))
1170                         goto error_free;
1171
1172                 skb->protocol = vlan_eth_hdr(skb)->h_vlan_encapsulated_proto;
1173                 skb_set_network_header(skb, VLAN_ETH_HLEN);
1174         }
1175
1176         if (skb->protocol == htons(ETH_P_IP)) {
1177                 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
1178                     + sizeof(struct iphdr))))
1179                         skb->protocol = 0;
1180         }
1181 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1182         else if (skb->protocol == htons(ETH_P_IPV6)) {
1183                 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
1184                     + sizeof(struct ipv6hdr))))
1185                         skb->protocol = 0;
1186         }
1187 #endif
1188
1189         /* ToS */
1190         if (skb->protocol == htons(ETH_P_IP))
1191                 inner_tos = ip_hdr(skb)->tos;
1192 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1193         else if (skb->protocol == htons(ETH_P_IPV6))
1194                 inner_tos = ipv6_get_dsfield(ipv6_hdr(skb));
1195 #endif
1196         else
1197                 inner_tos = 0;
1198
1199         if (mutable->flags & TNL_F_TOS_INHERIT)
1200                 tos = inner_tos;
1201         else
1202                 tos = mutable->tos;
1203
1204         tos = INET_ECN_encapsulate(tos, inner_tos);
1205
1206         /* Route lookup */
1207         rt = find_route(vport, mutable, tos, &cache);
1208         if (unlikely(!rt))
1209                 goto error_free;
1210         if (unlikely(!cache))
1211                 unattached_dst = &rt_dst(rt);
1212
1213         /* Reset SKB */
1214         nf_reset(skb);
1215         secpath_reset(skb);
1216         skb_dst_drop(skb);
1217         skb_clear_rxhash(skb);
1218
1219         /* Offloading */
1220         skb = handle_offloads(skb, mutable, rt);
1221         if (IS_ERR(skb))
1222                 goto error;
1223
1224         /* MTU */
1225         if (unlikely(!check_mtu(skb, vport, mutable, rt, &frag_off))) {
1226                 err = VPORT_E_TX_DROPPED;
1227                 goto error_free;
1228         }
1229
1230         /*
1231          * If we are over the MTU, allow the IP stack to handle fragmentation.
1232          * Fragmentation is a slow path anyways.
1233          */
1234         if (unlikely(skb->len + mutable->tunnel_hlen > dst_mtu(&rt_dst(rt)) &&
1235                      cache)) {
1236                 unattached_dst = &rt_dst(rt);
1237                 dst_hold(unattached_dst);
1238                 cache = NULL;
1239         }
1240
1241         /* TTL */
1242         ttl = mutable->ttl;
1243         if (!ttl)
1244                 ttl = ip4_dst_hoplimit(&rt_dst(rt));
1245
1246         if (mutable->flags & TNL_F_TTL_INHERIT) {
1247                 if (skb->protocol == htons(ETH_P_IP))
1248                         ttl = ip_hdr(skb)->ttl;
1249 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1250                 else if (skb->protocol == htons(ETH_P_IPV6))
1251                         ttl = ipv6_hdr(skb)->hop_limit;
1252 #endif
1253         }
1254
1255         while (skb) {
1256                 struct iphdr *iph;
1257                 struct sk_buff *next_skb = skb->next;
1258                 skb->next = NULL;
1259
1260                 if (unlikely(vlan_deaccel_tag(skb)))
1261                         goto next;
1262
1263                 if (likely(cache)) {
1264                         skb_push(skb, cache->len);
1265                         memcpy(skb->data, get_cached_header(cache), cache->len);
1266                         skb_reset_mac_header(skb);
1267                         skb_set_network_header(skb, rt_dst(rt).hh->hh_len);
1268
1269                 } else {
1270                         skb_push(skb, mutable->tunnel_hlen);
1271                         create_tunnel_header(vport, mutable, rt, skb->data);
1272                         skb_reset_network_header(skb);
1273
1274                         if (next_skb)
1275                                 skb_dst_set(skb, dst_clone(unattached_dst));
1276                         else {
1277                                 skb_dst_set(skb, unattached_dst);
1278                                 unattached_dst = NULL;
1279                         }
1280                 }
1281                 skb_set_transport_header(skb, skb_network_offset(skb) + sizeof(struct iphdr));
1282
1283                 iph = ip_hdr(skb);
1284                 iph->tos = tos;
1285                 iph->ttl = ttl;
1286                 iph->frag_off = frag_off;
1287                 ip_select_ident(iph, &rt_dst(rt), NULL);
1288
1289                 skb = tnl_vport->tnl_ops->update_header(vport, mutable, &rt_dst(rt), skb);
1290                 if (unlikely(!skb))
1291                         goto next;
1292
1293                 if (likely(cache)) {
1294                         int orig_len = skb->len - cache->len;
1295                         struct vport *cache_vport = internal_dev_get_vport(rt_dst(rt).dev);
1296
1297                         skb->protocol = htons(ETH_P_IP);
1298                         iph = ip_hdr(skb);
1299                         iph->tot_len = htons(skb->len - skb_network_offset(skb));
1300                         ip_send_check(iph);
1301
1302                         if (cache_vport) {
1303                                 if (unlikely(compute_ip_summed(skb, true))) {
1304                                         kfree_skb(skb);
1305                                         goto next;
1306                                 }
1307
1308                                 OVS_CB(skb)->flow = cache->flow;
1309                                 vport_receive(cache_vport, skb);
1310                                 sent_len += orig_len;
1311                         } else {
1312                                 int xmit_err;
1313
1314                                 skb->dev = rt_dst(rt).dev;
1315                                 xmit_err = dev_queue_xmit(skb);
1316
1317                                 if (likely(net_xmit_eval(xmit_err) == 0))
1318                                         sent_len += orig_len;
1319                         }
1320                 } else
1321                         sent_len += send_frags(skb, mutable);
1322
1323 next:
1324                 skb = next_skb;
1325         }
1326
1327         if (unlikely(sent_len == 0))
1328                 vport_record_error(vport, VPORT_E_TX_DROPPED);
1329
1330         goto out;
1331
1332 error_free:
1333         tnl_free_linked_skbs(skb);
1334 error:
1335         vport_record_error(vport, err);
1336 out:
1337         dst_release(unattached_dst);
1338         return sent_len;
1339 }
1340
1341 static const struct nla_policy tnl_policy[ODP_TUNNEL_ATTR_MAX + 1] = {
1342         [ODP_TUNNEL_ATTR_FLAGS]    = { .type = NLA_U32 },
1343         [ODP_TUNNEL_ATTR_DST_IPV4] = { .type = NLA_U32 },
1344         [ODP_TUNNEL_ATTR_SRC_IPV4] = { .type = NLA_U32 },
1345         [ODP_TUNNEL_ATTR_OUT_KEY]  = { .type = NLA_U64 },
1346         [ODP_TUNNEL_ATTR_IN_KEY]   = { .type = NLA_U64 },
1347         [ODP_TUNNEL_ATTR_TOS]      = { .type = NLA_U8 },
1348         [ODP_TUNNEL_ATTR_TTL]      = { .type = NLA_U8 },
1349 };
1350
1351 /* Sets ODP_TUNNEL_ATTR_* fields in 'mutable', which must initially be zeroed. */
1352 static int tnl_set_config(struct nlattr *options, const struct tnl_ops *tnl_ops,
1353                           const struct vport *cur_vport,
1354                           struct tnl_mutable_config *mutable)
1355 {
1356         const struct vport *old_vport;
1357         const struct tnl_mutable_config *old_mutable;
1358         struct nlattr *a[ODP_TUNNEL_ATTR_MAX + 1];
1359         int err;
1360
1361         if (!options)
1362                 return -EINVAL;
1363
1364         err = nla_parse_nested(a, ODP_TUNNEL_ATTR_MAX, options, tnl_policy);
1365         if (err)
1366                 return err;
1367
1368         if (!a[ODP_TUNNEL_ATTR_FLAGS] || !a[ODP_TUNNEL_ATTR_DST_IPV4])
1369                 return -EINVAL;
1370
1371         mutable->flags = nla_get_u32(a[ODP_TUNNEL_ATTR_FLAGS]) & TNL_F_PUBLIC;
1372
1373         if (a[ODP_TUNNEL_ATTR_SRC_IPV4])
1374                 mutable->saddr = nla_get_be32(a[ODP_TUNNEL_ATTR_SRC_IPV4]);
1375         mutable->daddr = nla_get_be32(a[ODP_TUNNEL_ATTR_DST_IPV4]);
1376
1377         if (a[ODP_TUNNEL_ATTR_TOS]) {
1378                 mutable->tos = nla_get_u8(a[ODP_TUNNEL_ATTR_TOS]);
1379                 if (mutable->tos != RT_TOS(mutable->tos))
1380                         return -EINVAL;
1381         }
1382
1383         if (a[ODP_TUNNEL_ATTR_TTL])
1384                 mutable->ttl = nla_get_u8(a[ODP_TUNNEL_ATTR_TTL]);
1385
1386         mutable->tunnel_type = tnl_ops->tunnel_type;
1387         if (!a[ODP_TUNNEL_ATTR_IN_KEY]) {
1388                 mutable->tunnel_type |= TNL_T_KEY_MATCH;
1389                 mutable->flags |= TNL_F_IN_KEY_MATCH;
1390         } else {
1391                 mutable->tunnel_type |= TNL_T_KEY_EXACT;
1392                 mutable->in_key = nla_get_be64(a[ODP_TUNNEL_ATTR_IN_KEY]);
1393         }
1394
1395         if (!a[ODP_TUNNEL_ATTR_OUT_KEY])
1396                 mutable->flags |= TNL_F_OUT_KEY_ACTION;
1397         else
1398                 mutable->out_key = nla_get_be64(a[ODP_TUNNEL_ATTR_OUT_KEY]);
1399
1400         mutable->tunnel_hlen = tnl_ops->hdr_len(mutable);
1401         if (mutable->tunnel_hlen < 0)
1402                 return mutable->tunnel_hlen;
1403
1404         mutable->tunnel_hlen += sizeof(struct iphdr);
1405
1406         old_vport = tnl_find_port(mutable->saddr, mutable->daddr,
1407                                   mutable->in_key, mutable->tunnel_type,
1408                                   &old_mutable);
1409
1410         if (old_vport && old_vport != cur_vport)
1411                 return -EEXIST;
1412
1413         return 0;
1414 }
1415
1416 struct vport *tnl_create(const struct vport_parms *parms,
1417                          const struct vport_ops *vport_ops,
1418                          const struct tnl_ops *tnl_ops)
1419 {
1420         struct vport *vport;
1421         struct tnl_vport *tnl_vport;
1422         struct tnl_mutable_config *mutable;
1423         int initial_frag_id;
1424         int err;
1425
1426         vport = vport_alloc(sizeof(struct tnl_vport), vport_ops, parms);
1427         if (IS_ERR(vport)) {
1428                 err = PTR_ERR(vport);
1429                 goto error;
1430         }
1431
1432         tnl_vport = tnl_vport_priv(vport);
1433
1434         strcpy(tnl_vport->name, parms->name);
1435         tnl_vport->tnl_ops = tnl_ops;
1436
1437         mutable = kzalloc(sizeof(struct tnl_mutable_config), GFP_KERNEL);
1438         if (!mutable) {
1439                 err = -ENOMEM;
1440                 goto error_free_vport;
1441         }
1442
1443         vport_gen_rand_ether_addr(mutable->eth_addr);
1444
1445         get_random_bytes(&initial_frag_id, sizeof(int));
1446         atomic_set(&tnl_vport->frag_id, initial_frag_id);
1447
1448         err = tnl_set_config(parms->options, tnl_ops, NULL, mutable);
1449         if (err)
1450                 goto error_free_mutable;
1451
1452         spin_lock_init(&tnl_vport->cache_lock);
1453
1454 #ifdef NEED_CACHE_TIMEOUT
1455         tnl_vport->cache_exp_interval = MAX_CACHE_EXP -
1456                                        (net_random() % (MAX_CACHE_EXP / 2));
1457 #endif
1458
1459         rcu_assign_pointer(tnl_vport->mutable, mutable);
1460
1461         err = add_port(vport);
1462         if (err)
1463                 goto error_free_mutable;
1464
1465         return vport;
1466
1467 error_free_mutable:
1468         kfree(mutable);
1469 error_free_vport:
1470         vport_free(vport);
1471 error:
1472         return ERR_PTR(err);
1473 }
1474
1475 int tnl_set_options(struct vport *vport, struct nlattr *options)
1476 {
1477         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1478         const struct tnl_mutable_config *old_mutable;
1479         struct tnl_mutable_config *mutable;
1480         int err;
1481
1482         mutable = kzalloc(sizeof(struct tnl_mutable_config), GFP_KERNEL);
1483         if (!mutable) {
1484                 err = -ENOMEM;
1485                 goto error;
1486         }
1487
1488         /* Copy fields whose values should be retained. */
1489         old_mutable = rtnl_dereference(tnl_vport->mutable);
1490         mutable->seq = old_mutable->seq + 1;
1491         memcpy(mutable->eth_addr, old_mutable->eth_addr, ETH_ALEN);
1492
1493         /* Parse the others configured by userspace. */
1494         err = tnl_set_config(options, tnl_vport->tnl_ops, vport, mutable);
1495         if (err)
1496                 goto error_free;
1497
1498         err = move_port(vport, mutable);
1499         if (err)
1500                 goto error_free;
1501
1502         return 0;
1503
1504 error_free:
1505         kfree(mutable);
1506 error:
1507         return err;
1508 }
1509
1510 int tnl_get_options(const struct vport *vport, struct sk_buff *skb)
1511 {
1512         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1513         const struct tnl_mutable_config *mutable = rcu_dereference_rtnl(tnl_vport->mutable);
1514
1515         NLA_PUT_U32(skb, ODP_TUNNEL_ATTR_FLAGS, mutable->flags & TNL_F_PUBLIC);
1516         NLA_PUT_BE32(skb, ODP_TUNNEL_ATTR_DST_IPV4, mutable->daddr);
1517
1518         if (!(mutable->flags & TNL_F_IN_KEY_MATCH))
1519                 NLA_PUT_BE64(skb, ODP_TUNNEL_ATTR_IN_KEY, mutable->in_key);
1520         if (!(mutable->flags & TNL_F_OUT_KEY_ACTION))
1521                 NLA_PUT_BE64(skb, ODP_TUNNEL_ATTR_OUT_KEY, mutable->out_key);
1522         if (mutable->saddr)
1523                 NLA_PUT_BE32(skb, ODP_TUNNEL_ATTR_SRC_IPV4, mutable->saddr);
1524         if (mutable->tos)
1525                 NLA_PUT_U8(skb, ODP_TUNNEL_ATTR_TOS, mutable->tos);
1526         if (mutable->ttl)
1527                 NLA_PUT_U8(skb, ODP_TUNNEL_ATTR_TTL, mutable->ttl);
1528
1529         return 0;
1530
1531 nla_put_failure:
1532         return -EMSGSIZE;
1533 }
1534
1535 static void free_port_rcu(struct rcu_head *rcu)
1536 {
1537         struct tnl_vport *tnl_vport = container_of(rcu,
1538                                                    struct tnl_vport, rcu);
1539
1540         free_cache((struct tnl_cache __force *)tnl_vport->cache);
1541         kfree((struct tnl_mutable __force *)tnl_vport->mutable);
1542         vport_free(tnl_vport_to_vport(tnl_vport));
1543 }
1544
1545 int tnl_destroy(struct vport *vport)
1546 {
1547         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1548         const struct tnl_mutable_config *mutable, *old_mutable;
1549
1550         mutable = rtnl_dereference(tnl_vport->mutable);
1551
1552         if (vport == tnl_find_port(mutable->saddr, mutable->daddr,
1553                                    mutable->in_key, mutable->tunnel_type,
1554                                    &old_mutable))
1555                 del_port(vport);
1556
1557         call_rcu(&tnl_vport->rcu, free_port_rcu);
1558
1559         return 0;
1560 }
1561
1562 int tnl_set_addr(struct vport *vport, const unsigned char *addr)
1563 {
1564         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1565         struct tnl_mutable_config *mutable;
1566
1567         mutable = kmemdup(rtnl_dereference(tnl_vport->mutable),
1568                           sizeof(struct tnl_mutable_config), GFP_KERNEL);
1569         if (!mutable)
1570                 return -ENOMEM;
1571
1572         memcpy(mutable->eth_addr, addr, ETH_ALEN);
1573         assign_config_rcu(vport, mutable);
1574
1575         return 0;
1576 }
1577
1578 const char *tnl_get_name(const struct vport *vport)
1579 {
1580         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1581         return tnl_vport->name;
1582 }
1583
1584 const unsigned char *tnl_get_addr(const struct vport *vport)
1585 {
1586         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1587         return rcu_dereference_rtnl(tnl_vport->mutable)->eth_addr;
1588 }
1589
1590 void tnl_free_linked_skbs(struct sk_buff *skb)
1591 {
1592         while (skb) {
1593                 struct sk_buff *next = skb->next;
1594                 kfree_skb(skb);
1595                 skb = next;
1596         }
1597 }