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  55  * Interval to check for and remove caches that are no longer valid.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)
 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, port_hash(&lookup), port_cmp);
 341                         if (tbl_node)
 342                                 goto found;
 343                 }
 344
 345                 if (key_remote_ports) {
 346                         lookup.saddr = 0;
 347
 348                         tbl_node = tbl_lookup(table, &lookup, port_hash(&lookup), port_cmp);
 349                         if (tbl_node)
 350                                 goto found;
 351
 352                         lookup.saddr = saddr;
 353                 }
 354         }
 355
 356         if (tunnel_type & TNL_T_KEY_MATCH) {
 357                 lookup.key = 0;
 358                 lookup.tunnel_type = tunnel_type & ~TNL_T_KEY_EXACT;
 359
 360                 if (local_remote_ports) {
 361                         tbl_node = tbl_lookup(table, &lookup, port_hash(&lookup), port_cmp);
 362                         if (tbl_node)
 363                                 goto found;
 364                 }
 365
 366                 if (remote_ports) {
 367                         lookup.saddr = 0;
 368
 369                         tbl_node = tbl_lookup(table, &lookup, port_hash(&lookup), port_cmp);
 370                         if (tbl_node)
 371                                 goto found;
 372                 }
 373         }
 374
 375         return NULL;
 376
 377 found:
 378         *mutable = lookup.mutable;
 379         return tnl_vport_to_vport(tnl_vport_table_cast(tbl_node));
 380 }
 381
 382 static inline void ecn_decapsulate(struct sk_buff *skb)
 383 {
 384         /* This is accessing the outer IP header of the tunnel, which we've
 385          * already validated to be OK.  skb->data is currently set to the start
 386          * of the inner Ethernet header, and we've validated ETH_HLEN.
 387          */
 388         if (unlikely(INET_ECN_is_ce(ip_hdr(skb)->tos))) {
 389                 __be16 protocol = skb->protocol;
 390
 391                 skb_set_network_header(skb, ETH_HLEN);
 392
 393                 if (skb->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 /* Called with rcu_read_lock. */
 421 void tnl_rcv(struct vport *vport, struct sk_buff *skb)
 422 {
 423         /* Packets received by this function are in the following state:
 424          * - skb->data points to the inner Ethernet header.
 425          * - The inner Ethernet header is in the linear data area.
 426          * - skb->csum does not include the inner Ethernet header.
 427          * - The layer pointers point at the outer headers.
 428          */
 429
 430         struct ethhdr *eh = (struct ethhdr *)skb->data;
 431
 432         if (likely(ntohs(eh->h_proto) >= 1536))
 433                 skb->protocol = eh->h_proto;
 434         else
 435                 skb->protocol = htons(ETH_P_802_2);
 436
 437         skb_dst_drop(skb);
 438         nf_reset(skb);
 439         skb_clear_rxhash(skb);
 440         secpath_reset(skb);
 441
 442         ecn_decapsulate(skb);
 443         compute_ip_summed(skb, false);
 444         vlan_set_tci(skb, 0);
 445
 446         vport_receive(vport, skb);
 447 }
 448
 449 static bool check_ipv4_address(__be32 addr)
 450 {
 451         if (ipv4_is_multicast(addr) || ipv4_is_lbcast(addr)
 452             || ipv4_is_loopback(addr) || ipv4_is_zeronet(addr))
 453                 return false;
 454
 455         return true;
 456 }
 457
 458 static bool ipv4_should_icmp(struct sk_buff *skb)
 459 {
 460         struct iphdr *old_iph = ip_hdr(skb);
 461
 462         /* Don't respond to L2 broadcast. */
 463         if (is_multicast_ether_addr(eth_hdr(skb)->h_dest))
 464                 return false;
 465
 466         /* Don't respond to L3 broadcast or invalid addresses. */
 467         if (!check_ipv4_address(old_iph->daddr) ||
 468             !check_ipv4_address(old_iph->saddr))
 469                 return false;
 470
 471         /* Only respond to the first fragment. */
 472         if (old_iph->frag_off & htons(IP_OFFSET))
 473                 return false;
 474
 475         /* Don't respond to ICMP error messages. */
 476         if (old_iph->protocol == IPPROTO_ICMP) {
 477                 u8 icmp_type, *icmp_typep;
 478
 479                 icmp_typep = skb_header_pointer(skb, (u8 *)old_iph +
 480                                                 (old_iph->ihl << 2) +
 481                                                 offsetof(struct icmphdr, type) -
 482                                                 skb->data, sizeof(icmp_type),
 483                                                 &icmp_type);
 484
 485                 if (!icmp_typep)
 486                         return false;
 487
 488                 if (*icmp_typep > NR_ICMP_TYPES
 489                         || (*icmp_typep <= ICMP_PARAMETERPROB
 490                                 && *icmp_typep != ICMP_ECHOREPLY
 491                                 && *icmp_typep != ICMP_ECHO))
 492                         return false;
 493         }
 494
 495         return true;
 496 }
 497
 498 static void ipv4_build_icmp(struct sk_buff *skb, struct sk_buff *nskb,
 499                             unsigned int mtu, unsigned int payload_length)
 500 {
 501         struct iphdr *iph, *old_iph = ip_hdr(skb);
 502         struct icmphdr *icmph;
 503         u8 *payload;
 504
 505         iph = (struct iphdr *)skb_put(nskb, sizeof(struct iphdr));
 506         icmph = (struct icmphdr *)skb_put(nskb, sizeof(struct icmphdr));
 507         payload = skb_put(nskb, payload_length);
 508
 509         /* IP */
 510         iph->version            =       4;
 511         iph->ihl                =       sizeof(struct iphdr) >> 2;
 512         iph->tos                =       (old_iph->tos & IPTOS_TOS_MASK) |
 513                                         IPTOS_PREC_INTERNETCONTROL;
 514         iph->tot_len            =       htons(sizeof(struct iphdr)
 515                                               + sizeof(struct icmphdr)
 516                                               + payload_length);
 517         get_random_bytes(&iph->id, sizeof(iph->id));
 518         iph->frag_off           =       0;
 519         iph->ttl                =       IPDEFTTL;
 520         iph->protocol           =       IPPROTO_ICMP;
 521         iph->daddr              =       old_iph->saddr;
 522         iph->saddr              =       old_iph->daddr;
 523
 524         ip_send_check(iph);
 525
 526         /* ICMP */
 527         icmph->type             =       ICMP_DEST_UNREACH;
 528         icmph->code             =       ICMP_FRAG_NEEDED;
 529         icmph->un.gateway       =       htonl(mtu);
 530         icmph->checksum         =       0;
 531
 532         nskb->csum = csum_partial((u8 *)icmph, sizeof(struct icmphdr), 0);
 533         nskb->csum = skb_copy_and_csum_bits(skb, (u8 *)old_iph - skb->data,
 534                                             payload, payload_length,
 535                                             nskb->csum);
 536         icmph->checksum = csum_fold(nskb->csum);
 537 }
 538
 539 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 540 static bool ipv6_should_icmp(struct sk_buff *skb)
 541 {
 542         struct ipv6hdr *old_ipv6h = ipv6_hdr(skb);
 543         int addr_type;
 544         int payload_off = (u8 *)(old_ipv6h + 1) - skb->data;
 545         u8 nexthdr = ipv6_hdr(skb)->nexthdr;
 546
 547         /* Check source address is valid. */
 548         addr_type = ipv6_addr_type(&old_ipv6h->saddr);
 549         if (addr_type & IPV6_ADDR_MULTICAST || addr_type == IPV6_ADDR_ANY)
 550                 return false;
 551
 552         /* Don't reply to unspecified addresses. */
 553         if (ipv6_addr_type(&old_ipv6h->daddr) == IPV6_ADDR_ANY)
 554                 return false;
 555
 556         /* Don't respond to ICMP error messages. */
 557         payload_off = ipv6_skip_exthdr(skb, payload_off, &nexthdr);
 558         if (payload_off < 0)
 559                 return false;
 560
 561         if (nexthdr == NEXTHDR_ICMP) {
 562                 u8 icmp_type, *icmp_typep;
 563
 564                 icmp_typep = skb_header_pointer(skb, payload_off +
 565                                                 offsetof(struct icmp6hdr,
 566                                                         icmp6_type),
 567                                                 sizeof(icmp_type), &icmp_type);
 568
 569                 if (!icmp_typep || !(*icmp_typep & ICMPV6_INFOMSG_MASK))
 570                         return false;
 571         }
 572
 573         return true;
 574 }
 575
 576 static void ipv6_build_icmp(struct sk_buff *skb, struct sk_buff *nskb,
 577                             unsigned int mtu, unsigned int payload_length)
 578 {
 579         struct ipv6hdr *ipv6h, *old_ipv6h = ipv6_hdr(skb);
 580         struct icmp6hdr *icmp6h;
 581         u8 *payload;
 582
 583         ipv6h = (struct ipv6hdr *)skb_put(nskb, sizeof(struct ipv6hdr));
 584         icmp6h = (struct icmp6hdr *)skb_put(nskb, sizeof(struct icmp6hdr));
 585         payload = skb_put(nskb, payload_length);
 586
 587         /* IPv6 */
 588         ipv6h->version          =       6;
 589         ipv6h->priority         =       0;
 590         memset(&ipv6h->flow_lbl, 0, sizeof(ipv6h->flow_lbl));
 591         ipv6h->payload_len      =       htons(sizeof(struct icmp6hdr)
 592                                               + payload_length);
 593         ipv6h->nexthdr          =       NEXTHDR_ICMP;
 594         ipv6h->hop_limit        =       IPV6_DEFAULT_HOPLIMIT;
 595         ipv6_addr_copy(&ipv6h->daddr, &old_ipv6h->saddr);
 596         ipv6_addr_copy(&ipv6h->saddr, &old_ipv6h->daddr);
 597
 598         /* ICMPv6 */
 599         icmp6h->icmp6_type      =       ICMPV6_PKT_TOOBIG;
 600         icmp6h->icmp6_code      =       0;
 601         icmp6h->icmp6_cksum     =       0;
 602         icmp6h->icmp6_mtu       =       htonl(mtu);
 603
 604         nskb->csum = csum_partial((u8 *)icmp6h, sizeof(struct icmp6hdr), 0);
 605         nskb->csum = skb_copy_and_csum_bits(skb, (u8 *)old_ipv6h - skb->data,
 606                                             payload, payload_length,
 607                                             nskb->csum);
 608         icmp6h->icmp6_cksum = csum_ipv6_magic(&ipv6h->saddr, &ipv6h->daddr,
 609                                                 sizeof(struct icmp6hdr)
 610                                                 + payload_length,
 611                                                 ipv6h->nexthdr, nskb->csum);
 612 }
 613 #endif /* IPv6 */
 614
 615 bool tnl_frag_needed(struct vport *vport, const struct tnl_mutable_config *mutable,
 616                      struct sk_buff *skb, unsigned int mtu, __be64 flow_key)
 617 {
 618         unsigned int eth_hdr_len = ETH_HLEN;
 619         unsigned int total_length = 0, header_length = 0, payload_length;
 620         struct ethhdr *eh, *old_eh = eth_hdr(skb);
 621         struct sk_buff *nskb;
 622
 623         /* Sanity check */
 624         if (skb->protocol == htons(ETH_P_IP)) {
 625                 if (mtu < IP_MIN_MTU)
 626                         return false;
 627
 628                 if (!ipv4_should_icmp(skb))
 629                         return true;
 630         }
 631 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 632         else if (skb->protocol == htons(ETH_P_IPV6)) {
 633                 if (mtu < IPV6_MIN_MTU)
 634                         return false;
 635
 636                 /*
 637                  * In theory we should do PMTUD on IPv6 multicast messages but
 638                  * we don't have an address to send from so just fragment.
 639                  */
 640                 if (ipv6_addr_type(&ipv6_hdr(skb)->daddr) & IPV6_ADDR_MULTICAST)
 641                         return false;
 642
 643                 if (!ipv6_should_icmp(skb))
 644                         return true;
 645         }
 646 #endif
 647         else
 648                 return false;
 649
 650         /* Allocate */
 651         if (old_eh->h_proto == htons(ETH_P_8021Q))
 652                 eth_hdr_len = VLAN_ETH_HLEN;
 653
 654         payload_length = skb->len - eth_hdr_len;
 655         if (skb->protocol == htons(ETH_P_IP)) {
 656                 header_length = sizeof(struct iphdr) + sizeof(struct icmphdr);
 657                 total_length = min_t(unsigned int, header_length +
 658                                                    payload_length, 576);
 659         }
 660 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 661         else {
 662                 header_length = sizeof(struct ipv6hdr) +
 663                                 sizeof(struct icmp6hdr);
 664                 total_length = min_t(unsigned int, header_length +
 665                                                   payload_length, IPV6_MIN_MTU);
 666         }
 667 #endif
 668
 669         payload_length = total_length - header_length;
 670
 671         nskb = dev_alloc_skb(NET_IP_ALIGN + eth_hdr_len + header_length +
 672                              payload_length);
 673         if (!nskb)
 674                 return false;
 675
 676         skb_reserve(nskb, NET_IP_ALIGN);
 677
 678         /* Ethernet / VLAN */
 679         eh = (struct ethhdr *)skb_put(nskb, eth_hdr_len);
 680         memcpy(eh->h_dest, old_eh->h_source, ETH_ALEN);
 681         memcpy(eh->h_source, mutable->eth_addr, ETH_ALEN);
 682         nskb->protocol = eh->h_proto = old_eh->h_proto;
 683         if (old_eh->h_proto == htons(ETH_P_8021Q)) {
 684                 struct vlan_ethhdr *vh = (struct vlan_ethhdr *)eh;
 685
 686                 vh->h_vlan_TCI = vlan_eth_hdr(skb)->h_vlan_TCI;
 687                 vh->h_vlan_encapsulated_proto = skb->protocol;
 688         } else
 689                 vlan_set_tci(nskb, vlan_get_tci(skb));
 690         skb_reset_mac_header(nskb);
 691
 692         /* Protocol */
 693         if (skb->protocol == htons(ETH_P_IP))
 694                 ipv4_build_icmp(skb, nskb, mtu, payload_length);
 695 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 696         else
 697                 ipv6_build_icmp(skb, nskb, mtu, payload_length);
 698 #endif
 699
 700         /*
 701          * Assume that flow based keys are symmetric with respect to input
 702          * and output and use the key that we were going to put on the
 703          * outgoing packet for the fake received packet.  If the keys are
 704          * not symmetric then PMTUD needs to be disabled since we won't have
 705          * any way of synthesizing packets.
 706          */
 707         if ((mutable->flags & (TNL_F_IN_KEY_MATCH | TNL_F_OUT_KEY_ACTION)) ==
 708             (TNL_F_IN_KEY_MATCH | TNL_F_OUT_KEY_ACTION))
 709                 OVS_CB(nskb)->tun_id = flow_key;
 710
 711         compute_ip_summed(nskb, false);
 712         vport_receive(vport, nskb);
 713
 714         return true;
 715 }
 716
 717 static bool check_mtu(struct sk_buff *skb,
 718                       struct vport *vport,
 719                       const struct tnl_mutable_config *mutable,
 720                       const struct rtable *rt, __be16 *frag_offp)
 721 {
 722         bool pmtud = mutable->flags & TNL_F_PMTUD;
 723         __be16 frag_off = 0;
 724         int mtu = 0;
 725         unsigned int packet_length = skb->len - ETH_HLEN;
 726
 727         /* Allow for one level of tagging in the packet length. */
 728         if (!vlan_tx_tag_present(skb) &&
 729             eth_hdr(skb)->h_proto == htons(ETH_P_8021Q))
 730                 packet_length -= VLAN_HLEN;
 731
 732         if (pmtud) {
 733                 int vlan_header = 0;
 734
 735                 frag_off = htons(IP_DF);
 736
 737                 /* The tag needs to go in packet regardless of where it
 738                  * currently is, so subtract it from the MTU.
 739                  */
 740                 if (vlan_tx_tag_present(skb) ||
 741                     eth_hdr(skb)->h_proto == htons(ETH_P_8021Q))
 742                         vlan_header = VLAN_HLEN;
 743
 744                 mtu = dst_mtu(&rt_dst(rt))
 745                         - ETH_HLEN
 746                         - mutable->tunnel_hlen
 747                         - vlan_header;
 748         }
 749
 750         if (skb->protocol == htons(ETH_P_IP)) {
 751                 struct iphdr *iph = ip_hdr(skb);
 752
 753                 frag_off |= iph->frag_off & htons(IP_DF);
 754
 755                 if (pmtud && iph->frag_off & htons(IP_DF)) {
 756                         mtu = max(mtu, IP_MIN_MTU);
 757
 758                         if (packet_length > mtu &&
 759                             tnl_frag_needed(vport, mutable, skb, mtu,
 760                                             OVS_CB(skb)->tun_id))
 761                                 return false;
 762                 }
 763         }
 764 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 765         else if (skb->protocol == htons(ETH_P_IPV6)) {
 766                 /* IPv6 requires PMTUD if the packet is above the minimum MTU. */
 767                 if (packet_length > IPV6_MIN_MTU)
 768                         frag_off = htons(IP_DF);
 769
 770                 if (pmtud) {
 771                         mtu = max(mtu, IPV6_MIN_MTU);
 772
 773                         if (packet_length > mtu &&
 774                             tnl_frag_needed(vport, mutable, skb, mtu,
 775                                             OVS_CB(skb)->tun_id))
 776                                 return false;
 777                 }
 778         }
 779 #endif
 780
 781         *frag_offp = frag_off;
 782         return true;
 783 }
 784
 785 static void create_tunnel_header(const struct vport *vport,
 786                                  const struct tnl_mutable_config *mutable,
 787                                  const struct rtable *rt, void *header)
 788 {
 789         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 790         struct iphdr *iph = header;
 791
 792         iph->version    = 4;
 793         iph->ihl        = sizeof(struct iphdr) >> 2;
 794         iph->frag_off   = htons(IP_DF);
 795         iph->protocol   = tnl_vport->tnl_ops->ipproto;
 796         iph->tos        = mutable->tos;
 797         iph->daddr      = rt->rt_dst;
 798         iph->saddr      = rt->rt_src;
 799         iph->ttl        = mutable->ttl;
 800         if (!iph->ttl)
 801                 iph->ttl = dst_metric(&rt_dst(rt), RTAX_HOPLIMIT);
 802
 803         tnl_vport->tnl_ops->build_header(vport, mutable, iph + 1);
 804 }
 805
 806 static inline void *get_cached_header(const struct tnl_cache *cache)
 807 {
 808         return (void *)cache + ALIGN(sizeof(struct tnl_cache), CACHE_DATA_ALIGN);
 809 }
 810
 811 static inline bool check_cache_valid(const struct tnl_cache *cache,
 812                                      const struct tnl_mutable_config *mutable)
 813 {
 814         return cache &&
 815 #ifdef NEED_CACHE_TIMEOUT
 816                 time_before(jiffies, cache->expiration) &&
 817 #endif
 818 #ifdef HAVE_RT_GENID
 819                 atomic_read(&init_net.ipv4.rt_genid) == cache->rt->rt_genid &&
 820 #endif
 821 #ifdef HAVE_HH_SEQ
 822                 rt_dst(cache->rt).hh->hh_lock.sequence == cache->hh_seq &&
 823 #endif
 824                 mutable->seq == cache->mutable_seq &&
 825                 (!is_internal_dev(rt_dst(cache->rt).dev) ||
 826                 (cache->flow && !cache->flow->dead));
 827 }
 828
 829 static int cache_cleaner_cb(struct tbl_node *tbl_node, void *aux)
 830 {
 831         struct tnl_vport *tnl_vport = tnl_vport_table_cast(tbl_node);
 832         const struct tnl_mutable_config *mutable = rcu_dereference(tnl_vport->mutable);
 833         const struct tnl_cache *cache = rcu_dereference(tnl_vport->cache);
 834
 835         if (cache && !check_cache_valid(cache, mutable) &&
 836             spin_trylock_bh(&tnl_vport->cache_lock)) {
 837                 assign_cache_rcu(tnl_vport_to_vport(tnl_vport), NULL);
 838                 spin_unlock_bh(&tnl_vport->cache_lock);
 839         }
 840
 841         return 0;
 842 }
 843
 844 static void cache_cleaner(struct work_struct *work)
 845 {
 846         schedule_cache_cleaner();
 847
 848         rcu_read_lock();
 849         tbl_foreach(rcu_dereference(port_table), cache_cleaner_cb, NULL);
 850         rcu_read_unlock();
 851 }
 852
 853 static inline void create_eth_hdr(struct tnl_cache *cache,
 854                                   const struct rtable *rt)
 855 {
 856         void *cache_data = get_cached_header(cache);
 857         int hh_len = rt_dst(rt).hh->hh_len;
 858         int hh_off = HH_DATA_ALIGN(rt_dst(rt).hh->hh_len) - hh_len;
 859
 860 #ifdef HAVE_HH_SEQ
 861         unsigned hh_seq;
 862
 863         do {
 864                 hh_seq = read_seqbegin(&rt_dst(rt).hh->hh_lock);
 865                 memcpy(cache_data, (void *)rt_dst(rt).hh->hh_data + hh_off, hh_len);
 866         } while (read_seqretry(&rt_dst(rt).hh->hh_lock, hh_seq));
 867
 868         cache->hh_seq = hh_seq;
 869 #else
 870         read_lock_bh(&rt_dst(rt).hh->hh_lock);
 871         memcpy(cache_data, (void *)rt_dst(rt).hh->hh_data + hh_off, hh_len);
 872         read_unlock_bh(&rt_dst(rt).hh->hh_lock);
 873 #endif
 874 }
 875
 876 static struct tnl_cache *build_cache(struct vport *vport,
 877                                      const struct tnl_mutable_config *mutable,
 878                                      struct rtable *rt)
 879 {
 880         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 881         struct tnl_cache *cache;
 882         void *cache_data;
 883         int cache_len;
 884
 885         if (!(mutable->flags & TNL_F_HDR_CACHE))
 886                 return NULL;
 887
 888         /*
 889          * If there is no entry in the ARP cache or if this device does not
 890          * support hard header caching just fall back to the IP stack.
 891          */
 892         if (!rt_dst(rt).hh)
 893                 return NULL;
 894
 895         /*
 896          * If lock is contended fall back to directly building the header.
 897          * We're not going to help performance by sitting here spinning.
 898          */
 899         if (!spin_trylock_bh(&tnl_vport->cache_lock))
 900                 return NULL;
 901
 902         cache = cache_dereference(tnl_vport);
 903         if (check_cache_valid(cache, mutable))
 904                 goto unlock;
 905         else
 906                 cache = NULL;
 907
 908         cache_len = rt_dst(rt).hh->hh_len + mutable->tunnel_hlen;
 909
 910         cache = kzalloc(ALIGN(sizeof(struct tnl_cache), CACHE_DATA_ALIGN) +
 911                         cache_len, GFP_ATOMIC);
 912         if (!cache)
 913                 goto unlock;
 914
 915         cache->len = cache_len;
 916
 917         create_eth_hdr(cache, rt);
 918         cache_data = get_cached_header(cache) + rt_dst(rt).hh->hh_len;
 919
 920         create_tunnel_header(vport, mutable, rt, cache_data);
 921
 922         cache->mutable_seq = mutable->seq;
 923         cache->rt = rt;
 924 #ifdef NEED_CACHE_TIMEOUT
 925         cache->expiration = jiffies + tnl_vport->cache_exp_interval;
 926 #endif
 927
 928         if (is_internal_dev(rt_dst(rt).dev)) {
 929                 struct sw_flow_key flow_key;
 930                 struct tbl_node *flow_node;
 931                 struct vport *dst_vport;
 932                 struct sk_buff *skb;
 933                 bool is_frag;
 934                 int err;
 935
 936                 dst_vport = internal_dev_get_vport(rt_dst(rt).dev);
 937                 if (!dst_vport)
 938                         goto done;
 939
 940                 skb = alloc_skb(cache->len, GFP_ATOMIC);
 941                 if (!skb)
 942                         goto done;
 943
 944                 __skb_put(skb, cache->len);
 945                 memcpy(skb->data, get_cached_header(cache), cache->len);
 946
 947                 err = flow_extract(skb, dst_vport->port_no, &flow_key, &is_frag);
 948
 949                 kfree_skb(skb);
 950                 if (err || is_frag)
 951                         goto done;
 952
 953                 flow_node = tbl_lookup(rcu_dereference(dst_vport->dp->table),
 954                                        &flow_key, flow_hash(&flow_key),
 955                                        flow_cmp);
 956                 if (flow_node) {
 957                         struct sw_flow *flow = flow_cast(flow_node);
 958
 959                         cache->flow = flow;
 960                         flow_hold(flow);
 961                 }
 962         }
 963
 964 done:
 965         assign_cache_rcu(vport, cache);
 966
 967 unlock:
 968         spin_unlock_bh(&tnl_vport->cache_lock);
 969
 970         return cache;
 971 }
 972
 973 static struct rtable *find_route(struct vport *vport,
 974                                  const struct tnl_mutable_config *mutable,
 975                                  u8 tos, struct tnl_cache **cache)
 976 {
 977         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
 978         struct tnl_cache *cur_cache = rcu_dereference(tnl_vport->cache);
 979
 980         *cache = NULL;
 981         tos = RT_TOS(tos);
 982
 983         if (likely(tos == mutable->tos && check_cache_valid(cur_cache, mutable))) {
 984                 *cache = cur_cache;
 985                 return cur_cache->rt;
 986         } else {
 987                 struct rtable *rt;
 988                 struct flowi fl = { .nl_u = { .ip4_u =
 989                                               { .daddr = mutable->daddr,
 990                                                 .saddr = mutable->saddr,
 991                                                 .tos = tos } },
 992                                     .proto = tnl_vport->tnl_ops->ipproto };
 993
 994                 if (unlikely(ip_route_output_key(&init_net, &rt, &fl)))
 995                         return NULL;
 996
 997                 if (likely(tos == mutable->tos))
 998                         *cache = build_cache(vport, mutable, rt);
 999
1000                 return rt;
1001         }
1002 }
1003
1004 static struct sk_buff *check_headroom(struct sk_buff *skb, int headroom)
1005 {
1006         if (skb_headroom(skb) < headroom || skb_header_cloned(skb)) {
1007                 struct sk_buff *nskb = skb_realloc_headroom(skb, headroom + 16);
1008                 if (unlikely(!nskb)) {
1009                         kfree_skb(skb);
1010                         return ERR_PTR(-ENOMEM);
1011                 }
1012
1013                 set_skb_csum_bits(skb, nskb);
1014
1015                 if (skb->sk)
1016                         skb_set_owner_w(nskb, skb->sk);
1017
1018                 kfree_skb(skb);
1019                 return nskb;
1020         }
1021
1022         return skb;
1023 }
1024
1025 static inline bool need_linearize(const struct sk_buff *skb)
1026 {
1027         int i;
1028
1029         if (unlikely(skb_shinfo(skb)->frag_list))
1030                 return true;
1031
1032         /*
1033          * Generally speaking we should linearize if there are paged frags.
1034          * However, if all of the refcounts are 1 we know nobody else can
1035          * change them from underneath us and we can skip the linearization.
1036          */
1037         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
1038                 if (unlikely(page_count(skb_shinfo(skb)->frags[i].page) > 1))
1039                         return true;
1040
1041         return false;
1042 }
1043
1044 static struct sk_buff *handle_offloads(struct sk_buff *skb,
1045                                        const struct tnl_mutable_config *mutable,
1046                                        const struct rtable *rt)
1047 {
1048         int min_headroom;
1049         int err;
1050
1051         forward_ip_summed(skb);
1052
1053         err = vswitch_skb_checksum_setup(skb);
1054         if (unlikely(err))
1055                 goto error_free;
1056
1057         min_headroom = LL_RESERVED_SPACE(rt_dst(rt).dev) + rt_dst(rt).header_len
1058                         + mutable->tunnel_hlen
1059                         + (vlan_tx_tag_present(skb) ? VLAN_HLEN : 0);
1060
1061         skb = check_headroom(skb, min_headroom);
1062         if (IS_ERR(skb)) {
1063                 err = PTR_ERR(skb);
1064                 goto error;
1065         }
1066
1067         if (skb_is_gso(skb)) {
1068                 struct sk_buff *nskb;
1069
1070                 nskb = skb_gso_segment(skb, 0);
1071                 kfree_skb(skb);
1072                 if (IS_ERR(nskb)) {
1073                         err = PTR_ERR(nskb);
1074                         goto error;
1075                 }
1076
1077                 skb = nskb;
1078         } else if (skb->ip_summed == CHECKSUM_PARTIAL) {
1079                 /* Pages aren't locked and could change at any time.
1080                  * If this happens after we compute the checksum, the
1081                  * checksum will be wrong.  We linearize now to avoid
1082                  * this problem.
1083                  */
1084                 if (unlikely(need_linearize(skb))) {
1085                         err = __skb_linearize(skb);
1086                         if (unlikely(err))
1087                                 goto error_free;
1088                 }
1089
1090                 err = skb_checksum_help(skb);
1091                 if (unlikely(err))
1092                         goto error_free;
1093         } else if (skb->ip_summed == CHECKSUM_COMPLETE)
1094                 skb->ip_summed = CHECKSUM_NONE;
1095
1096         return skb;
1097
1098 error_free:
1099         kfree_skb(skb);
1100 error:
1101         return ERR_PTR(err);
1102 }
1103
1104 static int send_frags(struct sk_buff *skb,
1105                       const struct tnl_mutable_config *mutable)
1106 {
1107         int sent_len;
1108
1109         sent_len = 0;
1110         while (skb) {
1111                 struct sk_buff *next = skb->next;
1112                 int frag_len = skb->len - mutable->tunnel_hlen;
1113                 int err;
1114
1115                 skb->next = NULL;
1116                 memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
1117
1118                 err = ip_local_out(skb);
1119                 skb = next;
1120                 if (unlikely(net_xmit_eval(err)))
1121                         goto free_frags;
1122                 sent_len += frag_len;
1123         }
1124
1125         return sent_len;
1126
1127 free_frags:
1128         /*
1129          * There's no point in continuing to send fragments once one has been
1130          * dropped so just free the rest.  This may help improve the congestion
1131          * that caused the first packet to be dropped.
1132          */
1133         tnl_free_linked_skbs(skb);
1134         return sent_len;
1135 }
1136
1137 int tnl_send(struct vport *vport, struct sk_buff *skb)
1138 {
1139         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1140         const struct tnl_mutable_config *mutable = rcu_dereference(tnl_vport->mutable);
1141
1142         enum vport_err_type err = VPORT_E_TX_ERROR;
1143         struct rtable *rt;
1144         struct dst_entry *unattached_dst = NULL;
1145         struct tnl_cache *cache;
1146         int sent_len = 0;
1147         __be16 frag_off = 0;
1148         u8 ttl;
1149         u8 inner_tos;
1150         u8 tos;
1151
1152         /* Validate the protocol headers before we try to use them. */
1153         if (skb->protocol == htons(ETH_P_8021Q) &&
1154             !vlan_tx_tag_present(skb)) {
1155                 if (unlikely(!pskb_may_pull(skb, VLAN_ETH_HLEN)))
1156                         goto error_free;
1157
1158                 skb->protocol = vlan_eth_hdr(skb)->h_vlan_encapsulated_proto;
1159                 skb_set_network_header(skb, VLAN_ETH_HLEN);
1160         }
1161
1162         if (skb->protocol == htons(ETH_P_IP)) {
1163                 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
1164                     + sizeof(struct iphdr))))
1165                         skb->protocol = 0;
1166         }
1167 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1168         else if (skb->protocol == htons(ETH_P_IPV6)) {
1169                 if (unlikely(!pskb_may_pull(skb, skb_network_offset(skb)
1170                     + sizeof(struct ipv6hdr))))
1171                         skb->protocol = 0;
1172         }
1173 #endif
1174
1175         /* ToS */
1176         if (skb->protocol == htons(ETH_P_IP))
1177                 inner_tos = ip_hdr(skb)->tos;
1178 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1179         else if (skb->protocol == htons(ETH_P_IPV6))
1180                 inner_tos = ipv6_get_dsfield(ipv6_hdr(skb));
1181 #endif
1182         else
1183                 inner_tos = 0;
1184
1185         if (mutable->flags & TNL_F_TOS_INHERIT)
1186                 tos = inner_tos;
1187         else
1188                 tos = mutable->tos;
1189
1190         tos = INET_ECN_encapsulate(tos, inner_tos);
1191
1192         /* Route lookup */
1193         rt = find_route(vport, mutable, tos, &cache);
1194         if (unlikely(!rt))
1195                 goto error_free;
1196         if (unlikely(!cache))
1197                 unattached_dst = &rt_dst(rt);
1198
1199         /* Reset SKB */
1200         nf_reset(skb);
1201         secpath_reset(skb);
1202         skb_dst_drop(skb);
1203         skb_clear_rxhash(skb);
1204
1205         /* Offloading */
1206         skb = handle_offloads(skb, mutable, rt);
1207         if (IS_ERR(skb))
1208                 goto error;
1209
1210         /* MTU */
1211         if (unlikely(!check_mtu(skb, vport, mutable, rt, &frag_off))) {
1212                 err = VPORT_E_TX_DROPPED;
1213                 goto error_free;
1214         }
1215
1216         /*
1217          * If we are over the MTU, allow the IP stack to handle fragmentation.
1218          * Fragmentation is a slow path anyways.
1219          */
1220         if (unlikely(skb->len + mutable->tunnel_hlen > dst_mtu(&rt_dst(rt)) &&
1221                      cache)) {
1222                 unattached_dst = &rt_dst(rt);
1223                 dst_hold(unattached_dst);
1224                 cache = NULL;
1225         }
1226
1227         /* TTL */
1228         ttl = mutable->ttl;
1229         if (!ttl)
1230                 ttl = dst_metric(&rt_dst(rt), RTAX_HOPLIMIT);
1231
1232         if (mutable->flags & TNL_F_TTL_INHERIT) {
1233                 if (skb->protocol == htons(ETH_P_IP))
1234                         ttl = ip_hdr(skb)->ttl;
1235 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1236                 else if (skb->protocol == htons(ETH_P_IPV6))
1237                         ttl = ipv6_hdr(skb)->hop_limit;
1238 #endif
1239         }
1240
1241         while (skb) {
1242                 struct iphdr *iph;
1243                 struct sk_buff *next_skb = skb->next;
1244                 skb->next = NULL;
1245
1246                 if (unlikely(vlan_deaccel_tag(skb)))
1247                         goto next;
1248
1249                 if (likely(cache)) {
1250                         skb_push(skb, cache->len);
1251                         memcpy(skb->data, get_cached_header(cache), cache->len);
1252                         skb_reset_mac_header(skb);
1253                         skb_set_network_header(skb, rt_dst(rt).hh->hh_len);
1254
1255                 } else {
1256                         skb_push(skb, mutable->tunnel_hlen);
1257                         create_tunnel_header(vport, mutable, rt, skb->data);
1258                         skb_reset_network_header(skb);
1259
1260                         if (next_skb)
1261                                 skb_dst_set(skb, dst_clone(unattached_dst));
1262                         else {
1263                                 skb_dst_set(skb, unattached_dst);
1264                                 unattached_dst = NULL;
1265                         }
1266                 }
1267                 skb_set_transport_header(skb, skb_network_offset(skb) + sizeof(struct iphdr));
1268
1269                 iph = ip_hdr(skb);
1270                 iph->tos = tos;
1271                 iph->ttl = ttl;
1272                 iph->frag_off = frag_off;
1273                 ip_select_ident(iph, &rt_dst(rt), NULL);
1274
1275                 skb = tnl_vport->tnl_ops->update_header(vport, mutable, &rt_dst(rt), skb);
1276                 if (unlikely(!skb))
1277                         goto next;
1278
1279                 if (likely(cache)) {
1280                         int orig_len = skb->len - cache->len;
1281                         struct vport *cache_vport = internal_dev_get_vport(rt_dst(rt).dev);
1282
1283                         skb->protocol = htons(ETH_P_IP);
1284                         iph = ip_hdr(skb);
1285                         iph->tot_len = htons(skb->len - skb_network_offset(skb));
1286                         ip_send_check(iph);
1287
1288                         if (cache_vport) {
1289                                 OVS_CB(skb)->flow = cache->flow;
1290                                 compute_ip_summed(skb, true);
1291                                 vport_receive(cache_vport, skb);
1292                                 sent_len += orig_len;
1293                         } else {
1294                                 int xmit_err;
1295
1296                                 skb->dev = rt_dst(rt).dev;
1297                                 xmit_err = dev_queue_xmit(skb);
1298
1299                                 if (likely(net_xmit_eval(xmit_err) == 0))
1300                                         sent_len += orig_len;
1301                         }
1302                 } else
1303                         sent_len += send_frags(skb, mutable);
1304
1305 next:
1306                 skb = next_skb;
1307         }
1308
1309         if (unlikely(sent_len == 0))
1310                 vport_record_error(vport, VPORT_E_TX_DROPPED);
1311
1312         goto out;
1313
1314 error_free:
1315         tnl_free_linked_skbs(skb);
1316 error:
1317         dst_release(unattached_dst);
1318         vport_record_error(vport, err);
1319 out:
1320         return sent_len;
1321 }
1322
1323 static const struct nla_policy tnl_policy[ODP_TUNNEL_ATTR_MAX + 1] = {
1324         [ODP_TUNNEL_ATTR_FLAGS]    = { .type = NLA_U32 },
1325         [ODP_TUNNEL_ATTR_DST_IPV4] = { .type = NLA_U32 },
1326         [ODP_TUNNEL_ATTR_SRC_IPV4] = { .type = NLA_U32 },
1327         [ODP_TUNNEL_ATTR_OUT_KEY]  = { .type = NLA_U64 },
1328         [ODP_TUNNEL_ATTR_IN_KEY]   = { .type = NLA_U64 },
1329         [ODP_TUNNEL_ATTR_TOS]      = { .type = NLA_U8 },
1330         [ODP_TUNNEL_ATTR_TTL]      = { .type = NLA_U8 },
1331 };
1332
1333 /* Sets ODP_TUNNEL_ATTR_* fields in 'mutable', which must initially be zeroed. */
1334 static int tnl_set_config(struct nlattr *options, const struct tnl_ops *tnl_ops,
1335                           const struct vport *cur_vport,
1336                           struct tnl_mutable_config *mutable)
1337 {
1338         const struct vport *old_vport;
1339         const struct tnl_mutable_config *old_mutable;
1340         struct nlattr *a[ODP_TUNNEL_ATTR_MAX + 1];
1341         int err;
1342
1343         if (!options)
1344                 return -EINVAL;
1345
1346         err = nla_parse_nested(a, ODP_TUNNEL_ATTR_MAX, options, tnl_policy);
1347         if (err)
1348                 return err;
1349
1350         if (!a[ODP_TUNNEL_ATTR_FLAGS] || !a[ODP_TUNNEL_ATTR_DST_IPV4])
1351                 return -EINVAL;
1352
1353         mutable->flags = nla_get_u32(a[ODP_TUNNEL_ATTR_FLAGS]) & TNL_F_PUBLIC;
1354
1355         if (a[ODP_TUNNEL_ATTR_SRC_IPV4])
1356                 mutable->saddr = nla_get_be32(a[ODP_TUNNEL_ATTR_SRC_IPV4]);
1357         mutable->daddr = nla_get_be32(a[ODP_TUNNEL_ATTR_DST_IPV4]);
1358
1359         if (a[ODP_TUNNEL_ATTR_TOS]) {
1360                 mutable->tos = nla_get_u8(a[ODP_TUNNEL_ATTR_TOS]);
1361                 if (mutable->tos != RT_TOS(mutable->tos))
1362                         return -EINVAL;
1363         }
1364
1365         if (a[ODP_TUNNEL_ATTR_TTL])
1366                 mutable->ttl = nla_get_u8(a[ODP_TUNNEL_ATTR_TTL]);
1367
1368         mutable->tunnel_type = tnl_ops->tunnel_type;
1369         if (!a[ODP_TUNNEL_ATTR_IN_KEY]) {
1370                 mutable->tunnel_type |= TNL_T_KEY_MATCH;
1371                 mutable->flags |= TNL_F_IN_KEY_MATCH;
1372         } else {
1373                 mutable->tunnel_type |= TNL_T_KEY_EXACT;
1374                 mutable->in_key = nla_get_be64(a[ODP_TUNNEL_ATTR_IN_KEY]);
1375         }
1376
1377         if (!a[ODP_TUNNEL_ATTR_OUT_KEY])
1378                 mutable->flags |= TNL_F_OUT_KEY_ACTION;
1379         else
1380                 mutable->out_key = nla_get_be64(a[ODP_TUNNEL_ATTR_OUT_KEY]);
1381
1382         mutable->tunnel_hlen = tnl_ops->hdr_len(mutable);
1383         if (mutable->tunnel_hlen < 0)
1384                 return mutable->tunnel_hlen;
1385
1386         mutable->tunnel_hlen += sizeof(struct iphdr);
1387
1388         old_vport = tnl_find_port(mutable->saddr, mutable->daddr,
1389                                   mutable->in_key, mutable->tunnel_type,
1390                                   &old_mutable);
1391
1392         if (old_vport && old_vport != cur_vport)
1393                 return -EEXIST;
1394
1395         return 0;
1396 }
1397
1398 struct vport *tnl_create(const struct vport_parms *parms,
1399                          const struct vport_ops *vport_ops,
1400                          const struct tnl_ops *tnl_ops)
1401 {
1402         struct vport *vport;
1403         struct tnl_vport *tnl_vport;
1404         struct tnl_mutable_config *mutable;
1405         int initial_frag_id;
1406         int err;
1407
1408         vport = vport_alloc(sizeof(struct tnl_vport), vport_ops, parms);
1409         if (IS_ERR(vport)) {
1410                 err = PTR_ERR(vport);
1411                 goto error;
1412         }
1413
1414         tnl_vport = tnl_vport_priv(vport);
1415
1416         strcpy(tnl_vport->name, parms->name);
1417         tnl_vport->tnl_ops = tnl_ops;
1418
1419         mutable = kzalloc(sizeof(struct tnl_mutable_config), GFP_KERNEL);
1420         if (!mutable) {
1421                 err = -ENOMEM;
1422                 goto error_free_vport;
1423         }
1424
1425         vport_gen_rand_ether_addr(mutable->eth_addr);
1426
1427         get_random_bytes(&initial_frag_id, sizeof(int));
1428         atomic_set(&tnl_vport->frag_id, initial_frag_id);
1429
1430         err = tnl_set_config(parms->options, tnl_ops, NULL, mutable);
1431         if (err)
1432                 goto error_free_mutable;
1433
1434         spin_lock_init(&tnl_vport->cache_lock);
1435
1436 #ifdef NEED_CACHE_TIMEOUT
1437         tnl_vport->cache_exp_interval = MAX_CACHE_EXP -
1438                                        (net_random() % (MAX_CACHE_EXP / 2));
1439 #endif
1440
1441         rcu_assign_pointer(tnl_vport->mutable, mutable);
1442
1443         err = add_port(vport);
1444         if (err)
1445                 goto error_free_mutable;
1446
1447         return vport;
1448
1449 error_free_mutable:
1450         kfree(mutable);
1451 error_free_vport:
1452         vport_free(vport);
1453 error:
1454         return ERR_PTR(err);
1455 }
1456
1457 int tnl_set_options(struct vport *vport, struct nlattr *options)
1458 {
1459         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1460         const struct tnl_mutable_config *old_mutable;
1461         struct tnl_mutable_config *mutable;
1462         int err;
1463
1464         mutable = kzalloc(sizeof(struct tnl_mutable_config), GFP_KERNEL);
1465         if (!mutable) {
1466                 err = -ENOMEM;
1467                 goto error;
1468         }
1469
1470         /* Copy fields whose values should be retained. */
1471         old_mutable = rtnl_dereference(tnl_vport->mutable);
1472         mutable->seq = old_mutable->seq + 1;
1473         memcpy(mutable->eth_addr, old_mutable->eth_addr, ETH_ALEN);
1474
1475         /* Parse the others configured by userspace. */
1476         err = tnl_set_config(options, tnl_vport->tnl_ops, vport, mutable);
1477         if (err)
1478                 goto error_free;
1479
1480         err = move_port(vport, mutable);
1481         if (err)
1482                 goto error_free;
1483
1484         return 0;
1485
1486 error_free:
1487         kfree(mutable);
1488 error:
1489         return err;
1490 }
1491
1492 int tnl_get_options(const struct vport *vport, struct sk_buff *skb)
1493 {
1494         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1495         const struct tnl_mutable_config *mutable = rcu_dereference_rtnl(tnl_vport->mutable);
1496
1497         NLA_PUT_U32(skb, ODP_TUNNEL_ATTR_FLAGS, mutable->flags & TNL_F_PUBLIC);
1498         NLA_PUT_BE32(skb, ODP_TUNNEL_ATTR_DST_IPV4, mutable->daddr);
1499
1500         if (!(mutable->flags & TNL_F_IN_KEY_MATCH))
1501                 NLA_PUT_BE64(skb, ODP_TUNNEL_ATTR_IN_KEY, mutable->in_key);
1502         if (!(mutable->flags & TNL_F_OUT_KEY_ACTION))
1503                 NLA_PUT_BE64(skb, ODP_TUNNEL_ATTR_OUT_KEY, mutable->out_key);
1504         if (mutable->saddr)
1505                 NLA_PUT_BE32(skb, ODP_TUNNEL_ATTR_SRC_IPV4, mutable->saddr);
1506         if (mutable->tos)
1507                 NLA_PUT_U8(skb, ODP_TUNNEL_ATTR_TOS, mutable->tos);
1508         if (mutable->ttl)
1509                 NLA_PUT_U8(skb, ODP_TUNNEL_ATTR_TTL, mutable->ttl);
1510
1511         return 0;
1512
1513 nla_put_failure:
1514         return -EMSGSIZE;
1515 }
1516
1517 static void free_port_rcu(struct rcu_head *rcu)
1518 {
1519         struct tnl_vport *tnl_vport = container_of(rcu,
1520                                                    struct tnl_vport, rcu);
1521
1522         free_cache((struct tnl_cache __force *)tnl_vport->cache);
1523         kfree((struct tnl_mutable __force *)tnl_vport->mutable);
1524         vport_free(tnl_vport_to_vport(tnl_vport));
1525 }
1526
1527 int tnl_destroy(struct vport *vport)
1528 {
1529         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1530         const struct tnl_mutable_config *mutable, *old_mutable;
1531
1532         mutable = rtnl_dereference(tnl_vport->mutable);
1533
1534         if (vport == tnl_find_port(mutable->saddr, mutable->daddr,
1535                                    mutable->in_key, mutable->tunnel_type,
1536                                    &old_mutable))
1537                 del_port(vport);
1538
1539         call_rcu(&tnl_vport->rcu, free_port_rcu);
1540
1541         return 0;
1542 }
1543
1544 int tnl_set_addr(struct vport *vport, const unsigned char *addr)
1545 {
1546         struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1547         struct tnl_mutable_config *mutable;
1548
1549         mutable = kmemdup(rtnl_dereference(tnl_vport->mutable),
1550                           sizeof(struct tnl_mutable_config), GFP_KERNEL);
1551         if (!mutable)
1552                 return -ENOMEM;
1553
1554         memcpy(mutable->eth_addr, addr, ETH_ALEN);
1555         assign_config_rcu(vport, mutable);
1556
1557         return 0;
1558 }
1559
1560 const char *tnl_get_name(const struct vport *vport)
1561 {
1562         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1563         return tnl_vport->name;
1564 }
1565
1566 const unsigned char *tnl_get_addr(const struct vport *vport)
1567 {
1568         const struct tnl_vport *tnl_vport = tnl_vport_priv(vport);
1569         return rcu_dereference_rtnl(tnl_vport->mutable)->eth_addr;
1570 }
1571
1572 void tnl_free_linked_skbs(struct sk_buff *skb)
1573 {
1574         while (skb) {
1575                 struct sk_buff *next = skb->next;
1576                 kfree_skb(skb);
1577                 skb = next;
1578         }
1579 }