datapath/flow.c

   1 /*
   2  * Copyright (c) 2007-2014 Nicira, Inc.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of version 2 of the GNU General Public
   6  * License as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful, but
   9  * WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public License
  14  * along with this program; if not, write to the Free Software
  15  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
  16  * 02110-1301, USA
  17  */
  18
  19 #include "flow.h"
  20 #include "datapath.h"
  21 #include <linux/uaccess.h>
  22 #include <linux/netdevice.h>
  23 #include <linux/etherdevice.h>
  24 #include <linux/if_ether.h>
  25 #include <linux/if_vlan.h>
  26 #include <net/llc_pdu.h>
  27 #include <linux/kernel.h>
  28 #include <linux/jhash.h>
  29 #include <linux/jiffies.h>
  30 #include <linux/llc.h>
  31 #include <linux/module.h>
  32 #include <linux/in.h>
  33 #include <linux/rcupdate.h>
  34 #include <linux/if_arp.h>
  35 #include <linux/ip.h>
  36 #include <linux/ipv6.h>
  37 #include <linux/sctp.h>
  38 #include <linux/smp.h>
  39 #include <linux/tcp.h>
  40 #include <linux/udp.h>
  41 #include <linux/icmp.h>
  42 #include <linux/icmpv6.h>
  43 #include <linux/rculist.h>
  44 #include <net/ip.h>
  45 #include <net/ipv6.h>
  46 #include <net/ndisc.h>
  47
  48 #include "mpls.h"
  49 #include "vlan.h"
  50
  51 u64 ovs_flow_used_time(unsigned long flow_jiffies)
  52 {
  53         struct timespec cur_ts;
  54         u64 cur_ms, idle_ms;
  55
  56         ktime_get_ts(&cur_ts);
  57         idle_ms = jiffies_to_msecs(jiffies - flow_jiffies);
  58         cur_ms = (u64)cur_ts.tv_sec * MSEC_PER_SEC +
  59                  cur_ts.tv_nsec / NSEC_PER_MSEC;
  60
  61         return cur_ms - idle_ms;
  62 }
  63
  64 #define TCP_FLAGS_BE16(tp) (*(__be16 *)&tcp_flag_word(tp) & htons(0x0FFF))
  65
  66 void ovs_flow_stats_update(struct sw_flow *flow, __be16 tcp_flags,
  67                            struct sk_buff *skb)
  68 {
  69         struct flow_stats *stats;
  70         int node = numa_node_id();
  71
  72         stats = rcu_dereference(flow->stats[node]);
  73
  74         /* Check if already have node-specific stats. */
  75         if (likely(stats)) {
  76                 spin_lock(&stats->lock);
  77                 /* Mark if we write on the pre-allocated stats. */
  78                 if (node == 0 && unlikely(flow->stats_last_writer != node))
  79                         flow->stats_last_writer = node;
  80         } else {
  81                 stats = rcu_dereference(flow->stats[0]); /* Pre-allocated. */
  82                 spin_lock(&stats->lock);
  83
  84                 /* If the current NUMA-node is the only writer on the
  85                  * pre-allocated stats keep using them.
  86                  */
  87                 if (unlikely(flow->stats_last_writer != node)) {
  88                         /* A previous locker may have already allocated the
  89                          * stats, so we need to check again.  If node-specific
  90                          * stats were already allocated, we update the pre-
  91                          * allocated stats as we have already locked them.
  92                          */
  93                         if (likely(flow->stats_last_writer != NUMA_NO_NODE)
  94                             && likely(!rcu_dereference(flow->stats[node]))) {
  95                                 /* Try to allocate node-specific stats. */
  96                                 struct flow_stats *new_stats;
  97
  98                                 new_stats =
  99                                         kmem_cache_alloc_node(flow_stats_cache,
 100                                                               GFP_THISNODE |
 101                                                               __GFP_NOMEMALLOC,
 102                                                               node);
 103                                 if (likely(new_stats)) {
 104                                         new_stats->used = jiffies;
 105                                         new_stats->packet_count = 1;
 106                                         new_stats->byte_count = skb->len;
 107                                         new_stats->tcp_flags = tcp_flags;
 108                                         spin_lock_init(&new_stats->lock);
 109
 110                                         rcu_assign_pointer(flow->stats[node],
 111                                                            new_stats);
 112                                         goto unlock;
 113                                 }
 114                         }
 115                         flow->stats_last_writer = node;
 116                 }
 117         }
 118
 119         stats->used = jiffies;
 120         stats->packet_count++;
 121         stats->byte_count += skb->len;
 122         stats->tcp_flags |= tcp_flags;
 123 unlock:
 124         spin_unlock(&stats->lock);
 125 }
 126
 127 /* Must be called with rcu_read_lock or ovs_mutex. */
 128 void ovs_flow_stats_get(const struct sw_flow *flow,
 129                         struct ovs_flow_stats *ovs_stats,
 130                         unsigned long *used, __be16 *tcp_flags)
 131 {
 132         int node;
 133
 134         *used = 0;
 135         *tcp_flags = 0;
 136         memset(ovs_stats, 0, sizeof(*ovs_stats));
 137
 138         for_each_node(node) {
 139                 struct flow_stats *stats = rcu_dereference_ovsl(flow->stats[node]);
 140
 141                 if (stats) {
 142                         /* Local CPU may write on non-local stats, so we must
 143                          * block bottom-halves here.
 144                          */
 145                         spin_lock_bh(&stats->lock);
 146                         if (!*used || time_after(stats->used, *used))
 147                                 *used = stats->used;
 148                         *tcp_flags |= stats->tcp_flags;
 149                         ovs_stats->n_packets += stats->packet_count;
 150                         ovs_stats->n_bytes += stats->byte_count;
 151                         spin_unlock_bh(&stats->lock);
 152                 }
 153         }
 154 }
 155
 156 /* Called with ovs_mutex. */
 157 void ovs_flow_stats_clear(struct sw_flow *flow)
 158 {
 159         int node;
 160
 161         for_each_node(node) {
 162                 struct flow_stats *stats = ovsl_dereference(flow->stats[node]);
 163
 164                 if (stats) {
 165                         spin_lock_bh(&stats->lock);
 166                         stats->used = 0;
 167                         stats->packet_count = 0;
 168                         stats->byte_count = 0;
 169                         stats->tcp_flags = 0;
 170                         spin_unlock_bh(&stats->lock);
 171                 }
 172         }
 173 }
 174
 175 static int check_header(struct sk_buff *skb, int len)
 176 {
 177         if (unlikely(skb->len < len))
 178                 return -EINVAL;
 179         if (unlikely(!pskb_may_pull(skb, len)))
 180                 return -ENOMEM;
 181         return 0;
 182 }
 183
 184 static bool arphdr_ok(struct sk_buff *skb)
 185 {
 186         return pskb_may_pull(skb, skb_network_offset(skb) +
 187                                   sizeof(struct arp_eth_header));
 188 }
 189
 190 static int check_iphdr(struct sk_buff *skb)
 191 {
 192         unsigned int nh_ofs = skb_network_offset(skb);
 193         unsigned int ip_len;
 194         int err;
 195
 196         err = check_header(skb, nh_ofs + sizeof(struct iphdr));
 197         if (unlikely(err))
 198                 return err;
 199
 200         ip_len = ip_hdrlen(skb);
 201         if (unlikely(ip_len < sizeof(struct iphdr) ||
 202                      skb->len < nh_ofs + ip_len))
 203                 return -EINVAL;
 204
 205         skb_set_transport_header(skb, nh_ofs + ip_len);
 206         return 0;
 207 }
 208
 209 static bool tcphdr_ok(struct sk_buff *skb)
 210 {
 211         int th_ofs = skb_transport_offset(skb);
 212         int tcp_len;
 213
 214         if (unlikely(!pskb_may_pull(skb, th_ofs + sizeof(struct tcphdr))))
 215                 return false;
 216
 217         tcp_len = tcp_hdrlen(skb);
 218         if (unlikely(tcp_len < sizeof(struct tcphdr) ||
 219                      skb->len < th_ofs + tcp_len))
 220                 return false;
 221
 222         return true;
 223 }
 224
 225 static bool udphdr_ok(struct sk_buff *skb)
 226 {
 227         return pskb_may_pull(skb, skb_transport_offset(skb) +
 228                                   sizeof(struct udphdr));
 229 }
 230
 231 static bool sctphdr_ok(struct sk_buff *skb)
 232 {
 233         return pskb_may_pull(skb, skb_transport_offset(skb) +
 234                                   sizeof(struct sctphdr));
 235 }
 236
 237 static bool icmphdr_ok(struct sk_buff *skb)
 238 {
 239         return pskb_may_pull(skb, skb_transport_offset(skb) +
 240                                   sizeof(struct icmphdr));
 241 }
 242
 243 static int parse_ipv6hdr(struct sk_buff *skb, struct sw_flow_key *key)
 244 {
 245         unsigned int nh_ofs = skb_network_offset(skb);
 246         unsigned int nh_len;
 247         int payload_ofs;
 248         struct ipv6hdr *nh;
 249         uint8_t nexthdr;
 250         __be16 frag_off;
 251         int err;
 252
 253         err = check_header(skb, nh_ofs + sizeof(*nh));
 254         if (unlikely(err))
 255                 return err;
 256
 257         nh = ipv6_hdr(skb);
 258         nexthdr = nh->nexthdr;
 259         payload_ofs = (u8 *)(nh + 1) - skb->data;
 260
 261         key->ip.proto = NEXTHDR_NONE;
 262         key->ip.tos = ipv6_get_dsfield(nh);
 263         key->ip.ttl = nh->hop_limit;
 264         key->ipv6.label = *(__be32 *)nh & htonl(IPV6_FLOWINFO_FLOWLABEL);
 265         key->ipv6.addr.src = nh->saddr;
 266         key->ipv6.addr.dst = nh->daddr;
 267
 268         payload_ofs = ipv6_skip_exthdr(skb, payload_ofs, &nexthdr, &frag_off);
 269         if (unlikely(payload_ofs < 0))
 270                 return -EINVAL;
 271
 272         if (frag_off) {
 273                 if (frag_off & htons(~0x7))
 274                         key->ip.frag = OVS_FRAG_TYPE_LATER;
 275                 else
 276                         key->ip.frag = OVS_FRAG_TYPE_FIRST;
 277         } else {
 278                 key->ip.frag = OVS_FRAG_TYPE_NONE;
 279         }
 280
 281         nh_len = payload_ofs - nh_ofs;
 282         skb_set_transport_header(skb, nh_ofs + nh_len);
 283         key->ip.proto = nexthdr;
 284         return nh_len;
 285 }
 286
 287 static bool icmp6hdr_ok(struct sk_buff *skb)
 288 {
 289         return pskb_may_pull(skb, skb_transport_offset(skb) +
 290                                   sizeof(struct icmp6hdr));
 291 }
 292
 293 static int parse_vlan(struct sk_buff *skb, struct sw_flow_key *key)
 294 {
 295         struct qtag_prefix {
 296                 __be16 eth_type; /* ETH_P_8021Q */
 297                 __be16 tci;
 298         };
 299         struct qtag_prefix *qp;
 300
 301         if (unlikely(skb->len < sizeof(struct qtag_prefix) + sizeof(__be16)))
 302                 return 0;
 303
 304         if (unlikely(!pskb_may_pull(skb, sizeof(struct qtag_prefix) +
 305                                          sizeof(__be16))))
 306                 return -ENOMEM;
 307
 308         qp = (struct qtag_prefix *) skb->data;
 309         key->eth.tci = qp->tci | htons(VLAN_TAG_PRESENT);
 310         __skb_pull(skb, sizeof(struct qtag_prefix));
 311
 312         return 0;
 313 }
 314
 315 static __be16 parse_ethertype(struct sk_buff *skb)
 316 {
 317         struct llc_snap_hdr {
 318                 u8  dsap;  /* Always 0xAA */
 319                 u8  ssap;  /* Always 0xAA */
 320                 u8  ctrl;
 321                 u8  oui[3];
 322                 __be16 ethertype;
 323         };
 324         struct llc_snap_hdr *llc;
 325         __be16 proto;
 326
 327         proto = *(__be16 *) skb->data;
 328         __skb_pull(skb, sizeof(__be16));
 329
 330         if (ntohs(proto) >= ETH_P_802_3_MIN)
 331                 return proto;
 332
 333         if (skb->len < sizeof(struct llc_snap_hdr))
 334                 return htons(ETH_P_802_2);
 335
 336         if (unlikely(!pskb_may_pull(skb, sizeof(struct llc_snap_hdr))))
 337                 return htons(0);
 338
 339         llc = (struct llc_snap_hdr *) skb->data;
 340         if (llc->dsap != LLC_SAP_SNAP ||
 341             llc->ssap != LLC_SAP_SNAP ||
 342             (llc->oui[0] | llc->oui[1] | llc->oui[2]) != 0)
 343                 return htons(ETH_P_802_2);
 344
 345         __skb_pull(skb, sizeof(struct llc_snap_hdr));
 346
 347         if (ntohs(llc->ethertype) >= ETH_P_802_3_MIN)
 348                 return llc->ethertype;
 349
 350         return htons(ETH_P_802_2);
 351 }
 352
 353 static int parse_icmpv6(struct sk_buff *skb, struct sw_flow_key *key,
 354                         int nh_len)
 355 {
 356         struct icmp6hdr *icmp = icmp6_hdr(skb);
 357
 358         /* The ICMPv6 type and code fields use the 16-bit transport port
 359          * fields, so we need to store them in 16-bit network byte order.
 360          */
 361         key->tp.src = htons(icmp->icmp6_type);
 362         key->tp.dst = htons(icmp->icmp6_code);
 363         memset(&key->ipv6.nd, 0, sizeof(key->ipv6.nd));
 364
 365         if (icmp->icmp6_code == 0 &&
 366             (icmp->icmp6_type == NDISC_NEIGHBOUR_SOLICITATION ||
 367              icmp->icmp6_type == NDISC_NEIGHBOUR_ADVERTISEMENT)) {
 368                 int icmp_len = skb->len - skb_transport_offset(skb);
 369                 struct nd_msg *nd;
 370                 int offset;
 371
 372                 /* In order to process neighbor discovery options, we need the
 373                  * entire packet.
 374                  */
 375                 if (unlikely(icmp_len < sizeof(*nd)))
 376                         return 0;
 377
 378                 if (unlikely(skb_linearize(skb)))
 379                         return -ENOMEM;
 380
 381                 nd = (struct nd_msg *)skb_transport_header(skb);
 382                 key->ipv6.nd.target = nd->target;
 383
 384                 icmp_len -= sizeof(*nd);
 385                 offset = 0;
 386                 while (icmp_len >= 8) {
 387                         struct nd_opt_hdr *nd_opt =
 388                                  (struct nd_opt_hdr *)(nd->opt + offset);
 389                         int opt_len = nd_opt->nd_opt_len * 8;
 390
 391                         if (unlikely(!opt_len || opt_len > icmp_len))
 392                                 return 0;
 393
 394                         /* Store the link layer address if the appropriate
 395                          * option is provided.  It is considered an error if
 396                          * the same link layer option is specified twice.
 397                          */
 398                         if (nd_opt->nd_opt_type == ND_OPT_SOURCE_LL_ADDR
 399                             && opt_len == 8) {
 400                                 if (unlikely(!is_zero_ether_addr(key->ipv6.nd.sll)))
 401                                         goto invalid;
 402                                 ether_addr_copy(key->ipv6.nd.sll,
 403                                     &nd->opt[offset+sizeof(*nd_opt)]);
 404                         } else if (nd_opt->nd_opt_type == ND_OPT_TARGET_LL_ADDR
 405                                    && opt_len == 8) {
 406                                 if (unlikely(!is_zero_ether_addr(key->ipv6.nd.tll)))
 407                                         goto invalid;
 408                                 ether_addr_copy(key->ipv6.nd.tll,
 409                                     &nd->opt[offset+sizeof(*nd_opt)]);
 410                         }
 411
 412                         icmp_len -= opt_len;
 413                         offset += opt_len;
 414                 }
 415         }
 416
 417         return 0;
 418
 419 invalid:
 420         memset(&key->ipv6.nd.target, 0, sizeof(key->ipv6.nd.target));
 421         memset(key->ipv6.nd.sll, 0, sizeof(key->ipv6.nd.sll));
 422         memset(key->ipv6.nd.tll, 0, sizeof(key->ipv6.nd.tll));
 423
 424         return 0;
 425 }
 426
 427 /**
 428  * ovs_flow_extract - extracts a flow key from an Ethernet frame.
 429  * @skb: sk_buff that contains the frame, with skb->data pointing to the
 430  * Ethernet header
 431  * @in_port: port number on which @skb was received.
 432  * @key: output flow key
 433  *
 434  * The caller must ensure that skb->len >= ETH_HLEN.
 435  *
 436  * Returns 0 if successful, otherwise a negative errno value.
 437  *
 438  * Initializes @skb header pointers as follows:
 439  *
 440  *    - skb->mac_header: the Ethernet header.
 441  *
 442  *    - skb->network_header: just past the Ethernet header, or just past the
 443  *      VLAN header, to the first byte of the Ethernet payload.
 444  *
 445  *    - skb->transport_header: If key->eth.type is ETH_P_IP or ETH_P_IPV6
 446  *      on output, then just past the IP header, if one is present and
 447  *      of a correct length, otherwise the same as skb->network_header.
 448  *      For other key->eth.type values it is left untouched.
 449  */
 450 int ovs_flow_extract(struct sk_buff *skb, u16 in_port, struct sw_flow_key *key)
 451 {
 452         int error;
 453         struct ethhdr *eth;
 454
 455         if (OVS_CB(skb)->tun_info) {
 456                 struct ovs_tunnel_info *tun_info = OVS_CB(skb)->tun_info;
 457                 memcpy(&key->tun_key, &tun_info->tunnel,
 458                         sizeof(key->tun_key));
 459                 if (tun_info->options) {
 460                         BUILD_BUG_ON((1 << (sizeof(tun_info->options_len) * 8)) - 1
 461                                         > sizeof(key->tun_opts));
 462                         memcpy(GENEVE_OPTS(key, tun_info->options_len),
 463                                 tun_info->options, tun_info->options_len);
 464                         key->tun_opts_len = tun_info->options_len;
 465                 } else {
 466                         key->tun_opts_len = 0;
 467                 }
 468         } else {
 469                 key->tun_opts_len = 0;
 470                 memset(&key->tun_key, 0, sizeof(key->tun_key));
 471         }
 472
 473         key->phy.priority = skb->priority;
 474         key->phy.in_port = in_port;
 475         key->phy.skb_mark = skb->mark;
 476         key->ovs_flow_hash = 0;
 477         key->recirc_id = 0;
 478
 479         /* Flags are always used as part of stats. */
 480         key->tp.flags = 0;
 481
 482         skb_reset_mac_header(skb);
 483
 484         /* Link layer.  We are guaranteed to have at least the 14 byte Ethernet
 485          * header in the linear data area.
 486          */
 487         eth = eth_hdr(skb);
 488         ether_addr_copy(key->eth.src, eth->h_source);
 489         ether_addr_copy(key->eth.dst, eth->h_dest);
 490
 491         __skb_pull(skb, 2 * ETH_ALEN);
 492         /* We are going to push all headers that we pull, so no need to
 493          * update skb->csum here. */
 494
 495         key->eth.tci = 0;
 496         if (vlan_tx_tag_present(skb))
 497                 key->eth.tci = htons(vlan_get_tci(skb));
 498         else if (eth->h_proto == htons(ETH_P_8021Q))
 499                 if (unlikely(parse_vlan(skb, key)))
 500                         return -ENOMEM;
 501
 502         key->eth.type = parse_ethertype(skb);
 503         if (unlikely(key->eth.type == htons(0)))
 504                 return -ENOMEM;
 505
 506         skb_reset_network_header(skb);
 507         skb_reset_mac_len(skb);
 508         __skb_push(skb, skb->data - skb_mac_header(skb));
 509
 510         /* Network layer. */
 511         if (key->eth.type == htons(ETH_P_IP)) {
 512                 struct iphdr *nh;
 513                 __be16 offset;
 514
 515                 error = check_iphdr(skb);
 516                 if (unlikely(error)) {
 517                         memset(&key->ip, 0, sizeof(key->ip));
 518                         memset(&key->ipv4, 0, sizeof(key->ipv4));
 519                         if (error == -EINVAL) {
 520                                 skb->transport_header = skb->network_header;
 521                                 error = 0;
 522                         }
 523                         return error;
 524                 }
 525
 526                 nh = ip_hdr(skb);
 527                 key->ipv4.addr.src = nh->saddr;
 528                 key->ipv4.addr.dst = nh->daddr;
 529
 530                 key->ip.proto = nh->protocol;
 531                 key->ip.tos = nh->tos;
 532                 key->ip.ttl = nh->ttl;
 533
 534                 offset = nh->frag_off & htons(IP_OFFSET);
 535                 if (offset) {
 536                         key->ip.frag = OVS_FRAG_TYPE_LATER;
 537                         return 0;
 538                 }
 539                 if (nh->frag_off & htons(IP_MF) ||
 540                          skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 541                         key->ip.frag = OVS_FRAG_TYPE_FIRST;
 542                 else
 543                         key->ip.frag = OVS_FRAG_TYPE_NONE;
 544
 545                 /* Transport layer. */
 546                 if (key->ip.proto == IPPROTO_TCP) {
 547                         if (tcphdr_ok(skb)) {
 548                                 struct tcphdr *tcp = tcp_hdr(skb);
 549                                 key->tp.src = tcp->source;
 550                                 key->tp.dst = tcp->dest;
 551                                 key->tp.flags = TCP_FLAGS_BE16(tcp);
 552                         } else {
 553                                 memset(&key->tp, 0, sizeof(key->tp));
 554                         }
 555                 } else if (key->ip.proto == IPPROTO_UDP) {
 556                         if (udphdr_ok(skb)) {
 557                                 struct udphdr *udp = udp_hdr(skb);
 558                                 key->tp.src = udp->source;
 559                                 key->tp.dst = udp->dest;
 560                         } else {
 561                                 memset(&key->tp, 0, sizeof(key->tp));
 562                         }
 563                 } else if (key->ip.proto == IPPROTO_SCTP) {
 564                         if (sctphdr_ok(skb)) {
 565                                 struct sctphdr *sctp = sctp_hdr(skb);
 566                                 key->tp.src = sctp->source;
 567                                 key->tp.dst = sctp->dest;
 568                         } else {
 569                                 memset(&key->tp, 0, sizeof(key->tp));
 570                         }
 571                 } else if (key->ip.proto == IPPROTO_ICMP) {
 572                         if (icmphdr_ok(skb)) {
 573                                 struct icmphdr *icmp = icmp_hdr(skb);
 574                                 /* The ICMP type and code fields use the 16-bit
 575                                  * transport port fields, so we need to store
 576                                  * them in 16-bit network byte order. */
 577                                 key->tp.src = htons(icmp->type);
 578                                 key->tp.dst = htons(icmp->code);
 579                         } else {
 580                                 memset(&key->tp, 0, sizeof(key->tp));
 581                         }
 582                 }
 583
 584         } else if (key->eth.type == htons(ETH_P_ARP) ||
 585                    key->eth.type == htons(ETH_P_RARP)) {
 586                 struct arp_eth_header *arp;
 587
 588                 arp = (struct arp_eth_header *)skb_network_header(skb);
 589
 590                 if (arphdr_ok(skb)
 591                                 && arp->ar_hrd == htons(ARPHRD_ETHER)
 592                                 && arp->ar_pro == htons(ETH_P_IP)
 593                                 && arp->ar_hln == ETH_ALEN
 594                                 && arp->ar_pln == 4) {
 595
 596                         /* We only match on the lower 8 bits of the opcode. */
 597                         if (ntohs(arp->ar_op) <= 0xff)
 598                                 key->ip.proto = ntohs(arp->ar_op);
 599                         else
 600                                 key->ip.proto = 0;
 601
 602                         memcpy(&key->ipv4.addr.src, arp->ar_sip, sizeof(key->ipv4.addr.src));
 603                         memcpy(&key->ipv4.addr.dst, arp->ar_tip, sizeof(key->ipv4.addr.dst));
 604                         ether_addr_copy(key->ipv4.arp.sha, arp->ar_sha);
 605                         ether_addr_copy(key->ipv4.arp.tha, arp->ar_tha);
 606                 } else {
 607                         memset(&key->ip, 0, sizeof(key->ip));
 608                         memset(&key->ipv4, 0, sizeof(key->ipv4));
 609                 }
 610         } else if (eth_p_mpls(key->eth.type)) {
 611                 size_t stack_len = MPLS_HLEN;
 612
 613                 /* In the presence of an MPLS label stack the end of the L2
 614                  * header and the beginning of the L3 header differ.
 615                  *
 616                  * Advance network_header to the beginning of the L3
 617                  * header. mac_len corresponds to the end of the L2 header.
 618                  */
 619                 while (1) {
 620                         __be32 lse;
 621
 622                         error = check_header(skb, skb->mac_len + stack_len);
 623                         if (unlikely(error))
 624                                 return 0;
 625
 626                         memcpy(&lse, skb_network_header(skb), MPLS_HLEN);
 627
 628                         if (stack_len == MPLS_HLEN)
 629                                 memcpy(&key->mpls.top_lse, &lse, MPLS_HLEN);
 630
 631                         skb_set_network_header(skb, skb->mac_len + stack_len);
 632                         if (lse & htonl(MPLS_BOS_MASK))
 633                                 break;
 634
 635                         stack_len += MPLS_HLEN;
 636                 }
 637         } else if (key->eth.type == htons(ETH_P_IPV6)) {
 638                 int nh_len;             /* IPv6 Header + Extensions */
 639
 640                 nh_len = parse_ipv6hdr(skb, key);
 641                 if (unlikely(nh_len < 0)) {
 642                         memset(&key->ip, 0, sizeof(key->ip));
 643                         memset(&key->ipv6.addr, 0, sizeof(key->ipv6.addr));
 644                         if (nh_len == -EINVAL) {
 645                                 skb->transport_header = skb->network_header;
 646                                 error = 0;
 647                         } else {
 648                                 error = nh_len;
 649                         }
 650                         return error;
 651                 }
 652
 653                 if (key->ip.frag == OVS_FRAG_TYPE_LATER)
 654                         return 0;
 655                 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 656                         key->ip.frag = OVS_FRAG_TYPE_FIRST;
 657
 658                 /* Transport layer. */
 659                 if (key->ip.proto == NEXTHDR_TCP) {
 660                         if (tcphdr_ok(skb)) {
 661                                 struct tcphdr *tcp = tcp_hdr(skb);
 662                                 key->tp.src = tcp->source;
 663                                 key->tp.dst = tcp->dest;
 664                                 key->tp.flags = TCP_FLAGS_BE16(tcp);
 665                         } else {
 666                                 memset(&key->tp, 0, sizeof(key->tp));
 667                         }
 668                 } else if (key->ip.proto == NEXTHDR_UDP) {
 669                         if (udphdr_ok(skb)) {
 670                                 struct udphdr *udp = udp_hdr(skb);
 671                                 key->tp.src = udp->source;
 672                                 key->tp.dst = udp->dest;
 673                         } else {
 674                                 memset(&key->tp, 0, sizeof(key->tp));
 675                         }
 676                 } else if (key->ip.proto == NEXTHDR_SCTP) {
 677                         if (sctphdr_ok(skb)) {
 678                                 struct sctphdr *sctp = sctp_hdr(skb);
 679                                 key->tp.src = sctp->source;
 680                                 key->tp.dst = sctp->dest;
 681                         } else {
 682                                 memset(&key->tp, 0, sizeof(key->tp));
 683                         }
 684                 } else if (key->ip.proto == NEXTHDR_ICMP) {
 685                         if (icmp6hdr_ok(skb)) {
 686                                 error = parse_icmpv6(skb, key, nh_len);
 687                                 if (error)
 688                                         return error;
 689                         } else {
 690                                 memset(&key->tp, 0, sizeof(key->tp));
 691                         }
 692                 }
 693         }
 694
 695         return 0;
 696 }