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                 return RTE_PTYPE_TUNNEL_GRE;
 187         }
 188         case IPPROTO_IPIP:
 189                 *proto = rte_cpu_to_be_16(ETHER_TYPE_IPv4);
 190                 return RTE_PTYPE_TUNNEL_IP;
 191         case IPPROTO_IPV6:
 192                 *proto = rte_cpu_to_be_16(ETHER_TYPE_IPv6);
 193                 return RTE_PTYPE_TUNNEL_IP; /* IP is also valid for IPv6 */
 194         default:
 195                 return 0;
 196         }
 197 }
 198
 199 /* get the ipv4 header length */
 200 static uint8_t
 201 ip4_hlen(const struct ipv4_hdr *hdr)
 202 {
 203         return (hdr->version_ihl & 0xf) * 4;
 204 }
 205
 206 /* parse ipv6 extended headers, update offset and return next proto */
 207 static uint16_t
 208 skip_ip6_ext(uint16_t proto, const struct rte_mbuf *m, uint32_t *off,
 209         int *frag)
 210 {
 211         struct ext_hdr {
 212                 uint8_t next_hdr;
 213                 uint8_t len;
 214         };
 215         const struct ext_hdr *xh;
 216         struct ext_hdr xh_copy;
 217         unsigned int i;
 218
 219         *frag = 0;
 220
 221 #define MAX_EXT_HDRS 5
 222         for (i = 0; i < MAX_EXT_HDRS; i++) {
 223                 switch (proto) {
 224                 case IPPROTO_HOPOPTS:
 225                 case IPPROTO_ROUTING:
 226                 case IPPROTO_DSTOPTS:
 227                         xh = rte_pktmbuf_read(m, *off, sizeof(*xh),
 228                                 &xh_copy);
 229                         if (xh == NULL)
 230                                 return 0;
 231                         *off += (xh->len + 1) * 8;
 232                         proto = xh->next_hdr;
 233                         break;
 234                 case IPPROTO_FRAGMENT:
 235                         xh = rte_pktmbuf_read(m, *off, sizeof(*xh),
 236                                 &xh_copy);
 237                         if (xh == NULL)
 238                                 return 0;
 239                         *off += 8;
 240                         proto = xh->next_hdr;
 241                         *frag = 1;
 242                         return proto; /* this is always the last ext hdr */
 243                 case IPPROTO_NONE:
 244                         return 0;
 245                 default:
 246                         return proto;
 247                 }
 248         }
 249         return 0;
 250 }
 251
 252 /* parse mbuf data to get packet type */
 253 uint32_t rte_net_get_ptype(const struct rte_mbuf *m,
 254         struct rte_net_hdr_lens *hdr_lens)
 255 {
 256         struct rte_net_hdr_lens local_hdr_lens;
 257         const struct ether_hdr *eh;
 258         struct ether_hdr eh_copy;
 259         uint32_t pkt_type = RTE_PTYPE_L2_ETHER;
 260         uint32_t off = 0;
 261         uint16_t proto;
 262
 263         if (hdr_lens == NULL)
 264                 hdr_lens = &local_hdr_lens;
 265
 266         eh = rte_pktmbuf_read(m, off, sizeof(*eh), &eh_copy);
 267         if (unlikely(eh == NULL))
 268                 return 0;
 269         proto = eh->ether_type;
 270         off = sizeof(*eh);
 271         hdr_lens->l2_len = off;
 272
 273         if (proto == rte_cpu_to_be_16(ETHER_TYPE_IPv4))
 274                 goto l3; /* fast path if packet is IPv4 */
 275
 276         if (proto == rte_cpu_to_be_16(ETHER_TYPE_VLAN)) {
 277                 const struct vlan_hdr *vh;
 278                 struct vlan_hdr vh_copy;
 279
 280                 pkt_type = RTE_PTYPE_L2_ETHER_VLAN;
 281                 vh = rte_pktmbuf_read(m, off, sizeof(*vh), &vh_copy);
 282                 if (unlikely(vh == NULL))
 283                         return pkt_type;
 284                 off += sizeof(*vh);
 285                 hdr_lens->l2_len += sizeof(*vh);
 286                 proto = vh->eth_proto;
 287         } else if (proto == rte_cpu_to_be_16(ETHER_TYPE_QINQ)) {
 288                 const struct vlan_hdr *vh;
 289                 struct vlan_hdr vh_copy;
 290
 291                 pkt_type = RTE_PTYPE_L2_ETHER_QINQ;
 292                 vh = rte_pktmbuf_read(m, off + sizeof(*vh), sizeof(*vh),
 293                         &vh_copy);
 294                 if (unlikely(vh == NULL))
 295                         return pkt_type;
 296                 off += 2 * sizeof(*vh);
 297                 hdr_lens->l2_len += 2 * sizeof(*vh);
 298                 proto = vh->eth_proto;
 299         }
 300
 301  l3:
 302         if (proto == rte_cpu_to_be_16(ETHER_TYPE_IPv4)) {
 303                 const struct ipv4_hdr *ip4h;
 304                 struct ipv4_hdr ip4h_copy;
 305
 306                 ip4h = rte_pktmbuf_read(m, off, sizeof(*ip4h), &ip4h_copy);
 307                 if (unlikely(ip4h == NULL))
 308                         return pkt_type;
 309
 310                 pkt_type |= ptype_l3_ip(ip4h->version_ihl);
 311                 hdr_lens->l3_len = ip4_hlen(ip4h);
 312                 off += hdr_lens->l3_len;
 313                 if (ip4h->fragment_offset & rte_cpu_to_be_16(
 314                                 IPV4_HDR_OFFSET_MASK | IPV4_HDR_MF_FLAG)) {
 315                         pkt_type |= RTE_PTYPE_L4_FRAG;
 316                         hdr_lens->l4_len = 0;
 317                         return pkt_type;
 318                 }
 319                 proto = ip4h->next_proto_id;
 320                 pkt_type |= ptype_l4(proto);
 321         } else if (proto == rte_cpu_to_be_16(ETHER_TYPE_IPv6)) {
 322                 const struct ipv6_hdr *ip6h;
 323                 struct ipv6_hdr ip6h_copy;
 324                 int frag = 0;
 325
 326                 ip6h = rte_pktmbuf_read(m, off, sizeof(*ip6h), &ip6h_copy);
 327                 if (unlikely(ip6h == NULL))
 328                         return pkt_type;
 329
 330                 proto = ip6h->proto;
 331                 hdr_lens->l3_len = sizeof(*ip6h);
 332                 off += hdr_lens->l3_len;
 333                 pkt_type |= ptype_l3_ip6(proto);
 334                 if ((pkt_type & RTE_PTYPE_L3_MASK) == RTE_PTYPE_L3_IPV6_EXT) {
 335                         proto = skip_ip6_ext(proto, m, &off, &frag);
 336                         hdr_lens->l3_len = off - hdr_lens->l2_len;
 337                 }
 338                 if (proto == 0)
 339                         return pkt_type;
 340                 if (frag) {
 341                         pkt_type |= RTE_PTYPE_L4_FRAG;
 342                         hdr_lens->l4_len = 0;
 343                         return pkt_type;
 344                 }
 345                 pkt_type |= ptype_l4(proto);
 346         }
 347
 348         if ((pkt_type & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP) {
 349                 hdr_lens->l4_len = sizeof(struct udp_hdr);
 350                 return pkt_type;
 351         } else if ((pkt_type & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP) {
 352                 const struct tcp_hdr *th;
 353                 struct tcp_hdr th_copy;
 354
 355                 th = rte_pktmbuf_read(m, off, sizeof(*th), &th_copy);
 356                 if (unlikely(th == NULL))
 357                         return pkt_type & (RTE_PTYPE_L2_MASK |
 358                                 RTE_PTYPE_L3_MASK);
 359                 hdr_lens->l4_len = (th->data_off & 0xf0) >> 2;
 360                 return pkt_type;
 361         } else if ((pkt_type & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP) {
 362                 hdr_lens->l4_len = sizeof(struct sctp_hdr);
 363                 return pkt_type;
 364         } else {
 365                 uint32_t prev_off = off;
 366
 367                 hdr_lens->l4_len = 0;
 368                 pkt_type |= ptype_tunnel(&proto, m, &off);
 369                 hdr_lens->tunnel_len = off - prev_off;
 370         }
 371
 372         /* same job for inner header: we need to duplicate the code
 373          * because the packet types do not have the same value.
 374          */
 375         hdr_lens->inner_l2_len = 0;
 376
 377         if (proto == rte_cpu_to_be_16(ETHER_TYPE_IPv4)) {
 378                 const struct ipv4_hdr *ip4h;
 379                 struct ipv4_hdr ip4h_copy;
 380
 381                 ip4h = rte_pktmbuf_read(m, off, sizeof(*ip4h), &ip4h_copy);
 382                 if (unlikely(ip4h == NULL))
 383                         return pkt_type;
 384
 385                 pkt_type |= ptype_inner_l3_ip(ip4h->version_ihl);
 386                 hdr_lens->inner_l3_len = ip4_hlen(ip4h);
 387                 off += hdr_lens->inner_l3_len;
 388                 if (ip4h->fragment_offset &
 389                                 rte_cpu_to_be_16(IPV4_HDR_OFFSET_MASK |
 390                                         IPV4_HDR_MF_FLAG)) {
 391                         pkt_type |= RTE_PTYPE_INNER_L4_FRAG;
 392                         hdr_lens->inner_l4_len = 0;
 393                         return pkt_type;
 394                 }
 395                 proto = ip4h->next_proto_id;
 396                 pkt_type |= ptype_inner_l4(proto);
 397         } else if (proto == rte_cpu_to_be_16(ETHER_TYPE_IPv6)) {
 398                 const struct ipv6_hdr *ip6h;
 399                 struct ipv6_hdr ip6h_copy;
 400                 int frag = 0;
 401
 402                 ip6h = rte_pktmbuf_read(m, off, sizeof(*ip6h), &ip6h_copy);
 403                 if (unlikely(ip6h == NULL))
 404                         return pkt_type;
 405
 406                 proto = ip6h->proto;
 407                 hdr_lens->inner_l3_len = sizeof(*ip6h);
 408                 off += hdr_lens->inner_l3_len;
 409                 pkt_type |= ptype_inner_l3_ip6(proto);
 410                 if ((pkt_type & RTE_PTYPE_INNER_L3_MASK) ==
 411                                 RTE_PTYPE_INNER_L3_IPV6_EXT) {
 412                         uint32_t prev_off;
 413
 414                         prev_off = off;
 415                         proto = skip_ip6_ext(proto, m, &off, &frag);
 416                         hdr_lens->inner_l3_len += off - prev_off;
 417                 }
 418                 if (proto == 0)
 419                         return pkt_type;
 420                 if (frag) {
 421                         pkt_type |= RTE_PTYPE_INNER_L4_FRAG;
 422                         hdr_lens->inner_l4_len = 0;
 423                         return pkt_type;
 424                 }
 425                 pkt_type |= ptype_inner_l4(proto);
 426         }
 427
 428         if ((pkt_type & RTE_PTYPE_INNER_L4_MASK) == RTE_PTYPE_INNER_L4_UDP) {
 429                 hdr_lens->inner_l4_len = sizeof(struct udp_hdr);
 430         } else if ((pkt_type & RTE_PTYPE_INNER_L4_MASK) ==
 431                         RTE_PTYPE_INNER_L4_TCP) {
 432                 const struct tcp_hdr *th;
 433                 struct tcp_hdr th_copy;
 434
 435                 th = rte_pktmbuf_read(m, off, sizeof(*th), &th_copy);
 436                 if (unlikely(th == NULL))
 437                         return pkt_type & (RTE_PTYPE_INNER_L2_MASK |
 438                                 RTE_PTYPE_INNER_L3_MASK);
 439                 hdr_lens->inner_l4_len = (th->data_off & 0xf0) >> 2;
 440         } else if ((pkt_type & RTE_PTYPE_INNER_L4_MASK) ==
 441                         RTE_PTYPE_INNER_L4_SCTP) {
 442                 hdr_lens->inner_l4_len = sizeof(struct sctp_hdr);
 443         } else {
 444                 hdr_lens->inner_l4_len = 0;
 445         }
 446
 447         return pkt_type;
 448 }