lib/librte_net/rte_net.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2016 6WIND S.A.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of 6WIND S.A. nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdint.h>
  35
  36 #include <rte_mbuf.h>
  37 #include <rte_mbuf_ptype.h>
  38 #include <rte_byteorder.h>
  39 #include <rte_ether.h>
  40 #include <rte_ip.h>
  41 #include <rte_tcp.h>
  42 #include <rte_udp.h>
  43 #include <rte_sctp.h>
  44 #include <rte_gre.h>
  45 #include <rte_net.h>
  46
  47 /* get l3 packet type from ip6 next protocol */
  48 static uint32_t
  49 ptype_l3_ip6(uint8_t ip6_proto)
  50 {
  51         static const uint32_t ip6_ext_proto_map[256] = {
  52                 [IPPROTO_HOPOPTS] = RTE_PTYPE_L3_IPV6_EXT - RTE_PTYPE_L3_IPV6,
  53                 [IPPROTO_ROUTING] = RTE_PTYPE_L3_IPV6_EXT - RTE_PTYPE_L3_IPV6,
  54                 [IPPROTO_FRAGMENT] = RTE_PTYPE_L3_IPV6_EXT - RTE_PTYPE_L3_IPV6,
  55                 [IPPROTO_ESP] = RTE_PTYPE_L3_IPV6_EXT - RTE_PTYPE_L3_IPV6,
  56                 [IPPROTO_AH] = RTE_PTYPE_L3_IPV6_EXT - RTE_PTYPE_L3_IPV6,
  57                 [IPPROTO_DSTOPTS] = RTE_PTYPE_L3_IPV6_EXT - RTE_PTYPE_L3_IPV6,
  58         };
  59
  60         return RTE_PTYPE_L3_IPV6 + ip6_ext_proto_map[ip6_proto];
  61 }
  62
  63 /* get l3 packet type from ip version and header length */
  64 static uint32_t
  65 ptype_l3_ip(uint8_t ipv_ihl)
  66 {
  67         static const uint32_t ptype_l3_ip_proto_map[256] = {
  68                 [0x45] = RTE_PTYPE_L3_IPV4,
  69                 [0x46] = RTE_PTYPE_L3_IPV4_EXT,
  70                 [0x47] = RTE_PTYPE_L3_IPV4_EXT,
  71                 [0x48] = RTE_PTYPE_L3_IPV4_EXT,
  72                 [0x49] = RTE_PTYPE_L3_IPV4_EXT,
  73                 [0x4A] = RTE_PTYPE_L3_IPV4_EXT,
  74                 [0x4B] = RTE_PTYPE_L3_IPV4_EXT,
  75                 [0x4C] = RTE_PTYPE_L3_IPV4_EXT,
  76                 [0x4D] = RTE_PTYPE_L3_IPV4_EXT,
  77                 [0x4E] = RTE_PTYPE_L3_IPV4_EXT,
  78                 [0x4F] = RTE_PTYPE_L3_IPV4_EXT,
  79         };
  80
  81         return ptype_l3_ip_proto_map[ipv_ihl];
  82 }
  83
  84 /* get l4 packet type from proto */
  85 static uint32_t
  86 ptype_l4(uint8_t proto)
  87 {
  88         static const uint32_t ptype_l4_proto[256] = {
  89                 [IPPROTO_UDP] = RTE_PTYPE_L4_UDP,
  90                 [IPPROTO_TCP] = RTE_PTYPE_L4_TCP,
  91                 [IPPROTO_SCTP] = RTE_PTYPE_L4_SCTP,
  92         };
  93
  94         return ptype_l4_proto[proto];
  95 }
  96
  97 /* get inner l3 packet type from ip6 next protocol */
  98 static uint32_t
  99 ptype_inner_l3_ip6(uint8_t ip6_proto)
 100 {
 101         static const uint32_t ptype_inner_ip6_ext_proto_map[256] = {
 102                 [IPPROTO_HOPOPTS] = RTE_PTYPE_INNER_L3_IPV6_EXT -
 103                         RTE_PTYPE_INNER_L3_IPV6,
 104                 [IPPROTO_ROUTING] = RTE_PTYPE_INNER_L3_IPV6_EXT -
 105                         RTE_PTYPE_INNER_L3_IPV6,
 106                 [IPPROTO_FRAGMENT] = RTE_PTYPE_INNER_L3_IPV6_EXT -
 107                         RTE_PTYPE_INNER_L3_IPV6,
 108                 [IPPROTO_ESP] = RTE_PTYPE_INNER_L3_IPV6_EXT -
 109                         RTE_PTYPE_INNER_L3_IPV6,
 110                 [IPPROTO_AH] = RTE_PTYPE_INNER_L3_IPV6_EXT -
 111                         RTE_PTYPE_INNER_L3_IPV6,
 112                 [IPPROTO_DSTOPTS] = RTE_PTYPE_INNER_L3_IPV6_EXT -
 113                         RTE_PTYPE_INNER_L3_IPV6,
 114         };
 115
 116         return RTE_PTYPE_INNER_L3_IPV6 +
 117                 ptype_inner_ip6_ext_proto_map[ip6_proto];
 118 }
 119
 120 /* get inner l3 packet type from ip version and header length */
 121 static uint32_t
 122 ptype_inner_l3_ip(uint8_t ipv_ihl)
 123 {
 124         static const uint32_t ptype_inner_l3_ip_proto_map[256] = {
 125                 [0x45] = RTE_PTYPE_INNER_L3_IPV4,
 126                 [0x46] = RTE_PTYPE_INNER_L3_IPV4_EXT,
 127                 [0x47] = RTE_PTYPE_INNER_L3_IPV4_EXT,
 128                 [0x48] = RTE_PTYPE_INNER_L3_IPV4_EXT,
 129                 [0x49] = RTE_PTYPE_INNER_L3_IPV4_EXT,
 130                 [0x4A] = RTE_PTYPE_INNER_L3_IPV4_EXT,
 131                 [0x4B] = RTE_PTYPE_INNER_L3_IPV4_EXT,
 132                 [0x4C] = RTE_PTYPE_INNER_L3_IPV4_EXT,
 133                 [0x4D] = RTE_PTYPE_INNER_L3_IPV4_EXT,
 134                 [0x4E] = RTE_PTYPE_INNER_L3_IPV4_EXT,
 135                 [0x4F] = RTE_PTYPE_INNER_L3_IPV4_EXT,
 136         };
 137
 138         return ptype_inner_l3_ip_proto_map[ipv_ihl];
 139 }
 140
 141 /* get inner l4 packet type from proto */
 142 static uint32_t
 143 ptype_inner_l4(uint8_t proto)
 144 {
 145         static const uint32_t ptype_inner_l4_proto[256] = {
 146                 [IPPROTO_UDP] = RTE_PTYPE_INNER_L4_UDP,
 147                 [IPPROTO_TCP] = RTE_PTYPE_INNER_L4_TCP,
 148                 [IPPROTO_SCTP] = RTE_PTYPE_INNER_L4_SCTP,
 149         };
 150
 151         return ptype_inner_l4_proto[proto];
 152 }
 153
 154 /* get the tunnel packet type if any, update proto and off. */
 155 static uint32_t
 156 ptype_tunnel(uint16_t *proto, const struct rte_mbuf *m,
 157         uint32_t *off)
 158 {
 159         switch (*proto) {
 160         case IPPROTO_GRE: {
 161                 static const uint8_t opt_len[16] = {
 162                         [0x0] = 4,
 163                         [0x1] = 8,
 164                         [0x2] = 8,
 165                         [0x8] = 8,
 166                         [0x3] = 12,
 167                         [0x9] = 12,
 168                         [0xa] = 12,
 169                         [0xb] = 16,
 170                 };
 171                 const struct gre_hdr *gh;
 172                 struct gre_hdr gh_copy;
 173                 uint16_t flags;
 174
 175                 gh = rte_pktmbuf_read(m, *off, sizeof(*gh), &gh_copy);
 176                 if (unlikely(gh == NULL))
 177                         return 0;
 178
 179                 flags = rte_be_to_cpu_16(*(const uint16_t *)gh);
 180                 flags >>= 12;
 181                 if (opt_len[flags] == 0)
 182                         return 0;
 183
 184                 *off += opt_len[flags];
 185                 *proto = gh->proto;
 186                 if (*proto == rte_cpu_to_be_16(ETHER_TYPE_TEB))
 187                         return RTE_PTYPE_TUNNEL_NVGRE;
 188                 else
 189                         return RTE_PTYPE_TUNNEL_GRE;
 190         }
 191         case IPPROTO_IPIP:
 192                 *proto = rte_cpu_to_be_16(ETHER_TYPE_IPv4);
 193                 return RTE_PTYPE_TUNNEL_IP;
 194         case IPPROTO_IPV6:
 195                 *proto = rte_cpu_to_be_16(ETHER_TYPE_IPv6);
 196                 return RTE_PTYPE_TUNNEL_IP; /* IP is also valid for IPv6 */
 197         default:
 198                 return 0;
 199         }
 200 }
 201
 202 /* get the ipv4 header length */
 203 static uint8_t
 204 ip4_hlen(const struct ipv4_hdr *hdr)
 205 {
 206         return (hdr->version_ihl & 0xf) * 4;
 207 }
 208
 209 /* parse ipv6 extended headers, update offset and return next proto */
 210 static uint16_t
 211 skip_ip6_ext(uint16_t proto, const struct rte_mbuf *m, uint32_t *off,
 212         int *frag)
 213 {
 214         struct ext_hdr {
 215                 uint8_t next_hdr;
 216                 uint8_t len;
 217         };
 218         const struct ext_hdr *xh;
 219         struct ext_hdr xh_copy;
 220         unsigned int i;
 221
 222         *frag = 0;
 223
 224 #define MAX_EXT_HDRS 5
 225         for (i = 0; i < MAX_EXT_HDRS; i++) {
 226                 switch (proto) {
 227                 case IPPROTO_HOPOPTS:
 228                 case IPPROTO_ROUTING:
 229                 case IPPROTO_DSTOPTS:
 230                         xh = rte_pktmbuf_read(m, *off, sizeof(*xh),
 231                                 &xh_copy);
 232                         if (xh == NULL)
 233                                 return 0;
 234                         *off += (xh->len + 1) * 8;
 235                         proto = xh->next_hdr;
 236                         break;
 237                 case IPPROTO_FRAGMENT:
 238                         xh = rte_pktmbuf_read(m, *off, sizeof(*xh),
 239                                 &xh_copy);
 240                         if (xh == NULL)
 241                                 return 0;
 242                         *off += 8;
 243                         proto = xh->next_hdr;
 244                         *frag = 1;
 245                         return proto; /* this is always the last ext hdr */
 246                 case IPPROTO_NONE:
 247                         return 0;
 248                 default:
 249                         return proto;
 250                 }
 251         }
 252         return 0;
 253 }
 254
 255 /* parse mbuf data to get packet type */
 256 uint32_t rte_net_get_ptype(const struct rte_mbuf *m,
 257         struct rte_net_hdr_lens *hdr_lens)
 258 {
 259         struct rte_net_hdr_lens local_hdr_lens;
 260         const struct ether_hdr *eh;
 261         struct ether_hdr eh_copy;
 262         uint32_t pkt_type = RTE_PTYPE_L2_ETHER;
 263         uint32_t off = 0;
 264         uint16_t proto;
 265
 266         if (hdr_lens == NULL)
 267                 hdr_lens = &local_hdr_lens;
 268
 269         eh = rte_pktmbuf_read(m, off, sizeof(*eh), &eh_copy);
 270         if (unlikely(eh == NULL))
 271                 return 0;
 272         proto = eh->ether_type;
 273         off = sizeof(*eh);
 274         hdr_lens->l2_len = off;
 275
 276         if (proto == rte_cpu_to_be_16(ETHER_TYPE_IPv4))
 277                 goto l3; /* fast path if packet is IPv4 */
 278
 279         if (proto == rte_cpu_to_be_16(ETHER_TYPE_VLAN)) {
 280                 const struct vlan_hdr *vh;
 281                 struct vlan_hdr vh_copy;
 282
 283                 pkt_type = RTE_PTYPE_L2_ETHER_VLAN;
 284                 vh = rte_pktmbuf_read(m, off, sizeof(*vh), &vh_copy);
 285                 if (unlikely(vh == NULL))
 286                         return pkt_type;
 287                 off += sizeof(*vh);
 288                 hdr_lens->l2_len += sizeof(*vh);
 289                 proto = vh->eth_proto;
 290         } else if (proto == rte_cpu_to_be_16(ETHER_TYPE_QINQ)) {
 291                 const struct vlan_hdr *vh;
 292                 struct vlan_hdr vh_copy;
 293
 294                 pkt_type = RTE_PTYPE_L2_ETHER_QINQ;
 295                 vh = rte_pktmbuf_read(m, off + sizeof(*vh), sizeof(*vh),
 296                         &vh_copy);
 297                 if (unlikely(vh == NULL))
 298                         return pkt_type;
 299                 off += 2 * sizeof(*vh);
 300                 hdr_lens->l2_len += 2 * sizeof(*vh);
 301                 proto = vh->eth_proto;
 302         }
 303
 304  l3:
 305         if (proto == rte_cpu_to_be_16(ETHER_TYPE_IPv4)) {
 306                 const struct ipv4_hdr *ip4h;
 307                 struct ipv4_hdr ip4h_copy;
 308
 309                 ip4h = rte_pktmbuf_read(m, off, sizeof(*ip4h), &ip4h_copy);
 310                 if (unlikely(ip4h == NULL))
 311                         return pkt_type;
 312
 313                 pkt_type |= ptype_l3_ip(ip4h->version_ihl);
 314                 hdr_lens->l3_len = ip4_hlen(ip4h);
 315                 off += hdr_lens->l3_len;
 316                 if (ip4h->fragment_offset & rte_cpu_to_be_16(
 317                                 IPV4_HDR_OFFSET_MASK | IPV4_HDR_MF_FLAG)) {
 318                         pkt_type |= RTE_PTYPE_L4_FRAG;
 319                         hdr_lens->l4_len = 0;
 320                         return pkt_type;
 321                 }
 322                 proto = ip4h->next_proto_id;
 323                 pkt_type |= ptype_l4(proto);
 324         } else if (proto == rte_cpu_to_be_16(ETHER_TYPE_IPv6)) {
 325                 const struct ipv6_hdr *ip6h;
 326                 struct ipv6_hdr ip6h_copy;
 327                 int frag = 0;
 328
 329                 ip6h = rte_pktmbuf_read(m, off, sizeof(*ip6h), &ip6h_copy);
 330                 if (unlikely(ip6h == NULL))
 331                         return pkt_type;
 332
 333                 proto = ip6h->proto;
 334                 hdr_lens->l3_len = sizeof(*ip6h);
 335                 off += hdr_lens->l3_len;
 336                 pkt_type |= ptype_l3_ip6(proto);
 337                 if ((pkt_type & RTE_PTYPE_L3_MASK) == RTE_PTYPE_L3_IPV6_EXT) {
 338                         proto = skip_ip6_ext(proto, m, &off, &frag);
 339                         hdr_lens->l3_len = off - hdr_lens->l2_len;
 340                 }
 341                 if (proto == 0)
 342                         return pkt_type;
 343                 if (frag) {
 344                         pkt_type |= RTE_PTYPE_L4_FRAG;
 345                         hdr_lens->l4_len = 0;
 346                         return pkt_type;
 347                 }
 348                 pkt_type |= ptype_l4(proto);
 349         }
 350
 351         if ((pkt_type & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP) {
 352                 hdr_lens->l4_len = sizeof(struct udp_hdr);
 353                 return pkt_type;
 354         } else if ((pkt_type & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP) {
 355                 const struct tcp_hdr *th;
 356                 struct tcp_hdr th_copy;
 357
 358                 th = rte_pktmbuf_read(m, off, sizeof(*th), &th_copy);
 359                 if (unlikely(th == NULL))
 360                         return pkt_type & (RTE_PTYPE_L2_MASK |
 361                                 RTE_PTYPE_L3_MASK);
 362                 hdr_lens->l4_len = (th->data_off & 0xf0) >> 2;
 363                 return pkt_type;
 364         } else if ((pkt_type & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP) {
 365                 hdr_lens->l4_len = sizeof(struct sctp_hdr);
 366                 return pkt_type;
 367         } else {
 368                 uint32_t prev_off = off;
 369
 370                 hdr_lens->l4_len = 0;
 371                 pkt_type |= ptype_tunnel(&proto, m, &off);
 372                 hdr_lens->tunnel_len = off - prev_off;
 373         }
 374
 375         /* same job for inner header: we need to duplicate the code
 376          * because the packet types do not have the same value.
 377          */
 378         if (proto == rte_cpu_to_be_16(ETHER_TYPE_TEB)) {
 379                 eh = rte_pktmbuf_read(m, off, sizeof(*eh), &eh_copy);
 380                 if (unlikely(eh == NULL))
 381                         return pkt_type;
 382                 pkt_type |= RTE_PTYPE_INNER_L2_ETHER;
 383                 proto = eh->ether_type;
 384                 off += sizeof(*eh);
 385                 hdr_lens->inner_l2_len = sizeof(*eh);
 386         }
 387
 388         if (proto == rte_cpu_to_be_16(ETHER_TYPE_VLAN)) {
 389                 const struct vlan_hdr *vh;
 390                 struct vlan_hdr vh_copy;
 391
 392                 pkt_type &= ~RTE_PTYPE_INNER_L2_MASK;
 393                 pkt_type |= RTE_PTYPE_INNER_L2_ETHER_VLAN;
 394                 vh = rte_pktmbuf_read(m, off, sizeof(*vh), &vh_copy);
 395                 if (unlikely(vh == NULL))
 396                         return pkt_type;
 397                 off += sizeof(*vh);
 398                 hdr_lens->inner_l2_len += sizeof(*vh);
 399                 proto = vh->eth_proto;
 400         } else if (proto == rte_cpu_to_be_16(ETHER_TYPE_QINQ)) {
 401                 const struct vlan_hdr *vh;
 402                 struct vlan_hdr vh_copy;
 403
 404                 pkt_type &= ~RTE_PTYPE_INNER_L2_MASK;
 405                 pkt_type |= RTE_PTYPE_INNER_L2_ETHER_QINQ;
 406                 vh = rte_pktmbuf_read(m, off + sizeof(*vh), sizeof(*vh),
 407                         &vh_copy);
 408                 if (unlikely(vh == NULL))
 409                         return pkt_type;
 410                 off += 2 * sizeof(*vh);
 411                 hdr_lens->inner_l2_len += 2 * sizeof(*vh);
 412                 proto = vh->eth_proto;
 413         }
 414
 415         if (proto == rte_cpu_to_be_16(ETHER_TYPE_IPv4)) {
 416                 const struct ipv4_hdr *ip4h;
 417                 struct ipv4_hdr ip4h_copy;
 418
 419                 ip4h = rte_pktmbuf_read(m, off, sizeof(*ip4h), &ip4h_copy);
 420                 if (unlikely(ip4h == NULL))
 421                         return pkt_type;
 422
 423                 pkt_type |= ptype_inner_l3_ip(ip4h->version_ihl);
 424                 hdr_lens->inner_l3_len = ip4_hlen(ip4h);
 425                 off += hdr_lens->inner_l3_len;
 426                 if (ip4h->fragment_offset &
 427                                 rte_cpu_to_be_16(IPV4_HDR_OFFSET_MASK |
 428                                         IPV4_HDR_MF_FLAG)) {
 429                         pkt_type |= RTE_PTYPE_INNER_L4_FRAG;
 430                         hdr_lens->inner_l4_len = 0;
 431                         return pkt_type;
 432                 }
 433                 proto = ip4h->next_proto_id;
 434                 pkt_type |= ptype_inner_l4(proto);
 435         } else if (proto == rte_cpu_to_be_16(ETHER_TYPE_IPv6)) {
 436                 const struct ipv6_hdr *ip6h;
 437                 struct ipv6_hdr ip6h_copy;
 438                 int frag = 0;
 439
 440                 ip6h = rte_pktmbuf_read(m, off, sizeof(*ip6h), &ip6h_copy);
 441                 if (unlikely(ip6h == NULL))
 442                         return pkt_type;
 443
 444                 proto = ip6h->proto;
 445                 hdr_lens->inner_l3_len = sizeof(*ip6h);
 446                 off += hdr_lens->inner_l3_len;
 447                 pkt_type |= ptype_inner_l3_ip6(proto);
 448                 if ((pkt_type & RTE_PTYPE_INNER_L3_MASK) ==
 449                                 RTE_PTYPE_INNER_L3_IPV6_EXT) {
 450                         uint32_t prev_off;
 451
 452                         prev_off = off;
 453                         proto = skip_ip6_ext(proto, m, &off, &frag);
 454                         hdr_lens->inner_l3_len += off - prev_off;
 455                 }
 456                 if (proto == 0)
 457                         return pkt_type;
 458                 if (frag) {
 459                         pkt_type |= RTE_PTYPE_INNER_L4_FRAG;
 460                         hdr_lens->inner_l4_len = 0;
 461                         return pkt_type;
 462                 }
 463                 pkt_type |= ptype_inner_l4(proto);
 464         }
 465
 466         if ((pkt_type & RTE_PTYPE_INNER_L4_MASK) == RTE_PTYPE_INNER_L4_UDP) {
 467                 hdr_lens->inner_l4_len = sizeof(struct udp_hdr);
 468         } else if ((pkt_type & RTE_PTYPE_INNER_L4_MASK) ==
 469                         RTE_PTYPE_INNER_L4_TCP) {
 470                 const struct tcp_hdr *th;
 471                 struct tcp_hdr th_copy;
 472
 473                 th = rte_pktmbuf_read(m, off, sizeof(*th), &th_copy);
 474                 if (unlikely(th == NULL))
 475                         return pkt_type & (RTE_PTYPE_INNER_L2_MASK |
 476                                 RTE_PTYPE_INNER_L3_MASK);
 477                 hdr_lens->inner_l4_len = (th->data_off & 0xf0) >> 2;
 478         } else if ((pkt_type & RTE_PTYPE_INNER_L4_MASK) ==
 479                         RTE_PTYPE_INNER_L4_SCTP) {
 480                 hdr_lens->inner_l4_len = sizeof(struct sctp_hdr);
 481         } else {
 482                 hdr_lens->inner_l4_len = 0;
 483         }
 484
 485         return pkt_type;
 486 }