net/eth.c

   1 /*
   2  * QEMU network structures definitions and helper functions
   3  *
   4  * Copyright (c) 2012 Ravello Systems LTD (http://ravellosystems.com)
   5  *
   6  * Developed by Daynix Computing LTD (http://www.daynix.com)
   7  *
   8  * Authors:
   9  * Dmitry Fleytman <dmitry@daynix.com>
  10  * Tamir Shomer <tamirs@daynix.com>
  11  * Yan Vugenfirer <yan@daynix.com>
  12  *
  13  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  14  * See the COPYING file in the top-level directory.
  15  *
  16  */
  17
  18 #include "qemu/osdep.h"
  19 #include "qemu/log.h"
  20 #include "net/eth.h"
  21 #include "net/checksum.h"
  22 #include "net/tap.h"
  23
  24 void eth_setup_vlan_headers(struct eth_header *ehdr, size_t *ehdr_size,
  25                             uint16_t vlan_tag, uint16_t vlan_ethtype)
  26 {
  27     struct vlan_header *vhdr = PKT_GET_VLAN_HDR(ehdr);
  28
  29     memmove(vhdr + 1, vhdr, *ehdr_size - ETH_HLEN);
  30     vhdr->h_tci = cpu_to_be16(vlan_tag);
  31     vhdr->h_proto = ehdr->h_proto;
  32     ehdr->h_proto = cpu_to_be16(vlan_ethtype);
  33     *ehdr_size += sizeof(*vhdr);
  34 }
  35
  36 uint8_t
  37 eth_get_gso_type(uint16_t l3_proto, uint8_t *l3_hdr, uint8_t l4proto)
  38 {
  39     uint8_t ecn_state = 0;
  40
  41     if (l3_proto == ETH_P_IP) {
  42         struct ip_header *iphdr = (struct ip_header *) l3_hdr;
  43
  44         if (IP_HEADER_VERSION(iphdr) == IP_HEADER_VERSION_4) {
  45             if (IPTOS_ECN(iphdr->ip_tos) == IPTOS_ECN_CE) {
  46                 ecn_state = VIRTIO_NET_HDR_GSO_ECN;
  47             }
  48             if (l4proto == IP_PROTO_TCP) {
  49                 return VIRTIO_NET_HDR_GSO_TCPV4 | ecn_state;
  50             } else if (l4proto == IP_PROTO_UDP) {
  51                 return VIRTIO_NET_HDR_GSO_UDP | ecn_state;
  52             }
  53         }
  54     } else if (l3_proto == ETH_P_IPV6) {
  55         struct ip6_header *ip6hdr = (struct ip6_header *) l3_hdr;
  56
  57         if (IP6_ECN(ip6hdr->ip6_ecn_acc) == IP6_ECN_CE) {
  58             ecn_state = VIRTIO_NET_HDR_GSO_ECN;
  59         }
  60
  61         if (l4proto == IP_PROTO_TCP) {
  62             return VIRTIO_NET_HDR_GSO_TCPV6 | ecn_state;
  63         }
  64     }
  65     qemu_log_mask(LOG_UNIMP, "%s: probably not GSO frame, "
  66         "unknown L3 protocol: 0x%04"PRIx16"\n", __func__, l3_proto);
  67
  68     return VIRTIO_NET_HDR_GSO_NONE | ecn_state;
  69 }
  70
  71 uint16_t
  72 eth_get_l3_proto(const struct iovec *l2hdr_iov, int iovcnt, size_t l2hdr_len)
  73 {
  74     uint16_t proto;
  75     size_t copied;
  76     size_t size = iov_size(l2hdr_iov, iovcnt);
  77     size_t proto_offset = l2hdr_len - sizeof(proto);
  78
  79     if (size < proto_offset) {
  80         return ETH_P_UNKNOWN;
  81     }
  82
  83     copied = iov_to_buf(l2hdr_iov, iovcnt, proto_offset,
  84                         &proto, sizeof(proto));
  85
  86     return (copied == sizeof(proto)) ? be16_to_cpu(proto) : ETH_P_UNKNOWN;
  87 }
  88
  89 static bool
  90 _eth_copy_chunk(size_t input_size,
  91                 const struct iovec *iov, int iovcnt,
  92                 size_t offset, size_t length,
  93                 void *buffer)
  94 {
  95     size_t copied;
  96
  97     if (input_size < offset) {
  98         return false;
  99     }
 100
 101     copied = iov_to_buf(iov, iovcnt, offset, buffer, length);
 102
 103     if (copied < length) {
 104         return false;
 105     }
 106
 107     return true;
 108 }
 109
 110 static bool
 111 _eth_tcp_has_data(bool is_ip4,
 112                   const struct ip_header  *ip4_hdr,
 113                   const struct ip6_header *ip6_hdr,
 114                   size_t full_ip6hdr_len,
 115                   const struct tcp_header *tcp)
 116 {
 117     uint32_t l4len;
 118
 119     if (is_ip4) {
 120         l4len = be16_to_cpu(ip4_hdr->ip_len) - IP_HDR_GET_LEN(ip4_hdr);
 121     } else {
 122         size_t opts_len = full_ip6hdr_len - sizeof(struct ip6_header);
 123         l4len = be16_to_cpu(ip6_hdr->ip6_ctlun.ip6_un1.ip6_un1_plen) - opts_len;
 124     }
 125
 126     return l4len > TCP_HEADER_DATA_OFFSET(tcp);
 127 }
 128
 129 void eth_get_protocols(const struct iovec *iov, size_t iovcnt, size_t iovoff,
 130                        bool *hasip4, bool *hasip6,
 131                        size_t *l3hdr_off,
 132                        size_t *l4hdr_off,
 133                        size_t *l5hdr_off,
 134                        eth_ip6_hdr_info *ip6hdr_info,
 135                        eth_ip4_hdr_info *ip4hdr_info,
 136                        eth_l4_hdr_info  *l4hdr_info)
 137 {
 138     int proto;
 139     bool fragment = false;
 140     size_t input_size = iov_size(iov, iovcnt);
 141     size_t copied;
 142     uint8_t ip_p;
 143
 144     *hasip4 = *hasip6 = false;
 145     *l3hdr_off = iovoff + eth_get_l2_hdr_length_iov(iov, iovcnt, iovoff);
 146     l4hdr_info->proto = ETH_L4_HDR_PROTO_INVALID;
 147
 148     proto = eth_get_l3_proto(iov, iovcnt, *l3hdr_off);
 149
 150     if (proto == ETH_P_IP) {
 151         struct ip_header *iphdr = &ip4hdr_info->ip4_hdr;
 152
 153         if (input_size < *l3hdr_off) {
 154             return;
 155         }
 156
 157         copied = iov_to_buf(iov, iovcnt, *l3hdr_off, iphdr, sizeof(*iphdr));
 158         if (copied < sizeof(*iphdr) ||
 159             IP_HEADER_VERSION(iphdr) != IP_HEADER_VERSION_4) {
 160             return;
 161         }
 162
 163         *hasip4 = true;
 164         ip_p = iphdr->ip_p;
 165         ip4hdr_info->fragment = IP4_IS_FRAGMENT(iphdr);
 166         *l4hdr_off = *l3hdr_off + IP_HDR_GET_LEN(iphdr);
 167
 168         fragment = ip4hdr_info->fragment;
 169     } else if (proto == ETH_P_IPV6) {
 170         if (!eth_parse_ipv6_hdr(iov, iovcnt, *l3hdr_off, ip6hdr_info)) {
 171             return;
 172         }
 173
 174         *hasip6 = true;
 175         ip_p = ip6hdr_info->l4proto;
 176         *l4hdr_off = *l3hdr_off + ip6hdr_info->full_hdr_len;
 177         fragment = ip6hdr_info->fragment;
 178     } else {
 179         return;
 180     }
 181
 182     if (fragment) {
 183         return;
 184     }
 185
 186     switch (ip_p) {
 187     case IP_PROTO_TCP:
 188         if (_eth_copy_chunk(input_size,
 189                             iov, iovcnt,
 190                             *l4hdr_off, sizeof(l4hdr_info->hdr.tcp),
 191                             &l4hdr_info->hdr.tcp)) {
 192             l4hdr_info->proto = ETH_L4_HDR_PROTO_TCP;
 193             *l5hdr_off = *l4hdr_off +
 194                 TCP_HEADER_DATA_OFFSET(&l4hdr_info->hdr.tcp);
 195
 196             l4hdr_info->has_tcp_data =
 197                 _eth_tcp_has_data(proto == ETH_P_IP,
 198                                   &ip4hdr_info->ip4_hdr,
 199                                   &ip6hdr_info->ip6_hdr,
 200                                   *l4hdr_off - *l3hdr_off,
 201                                   &l4hdr_info->hdr.tcp);
 202         }
 203         break;
 204
 205     case IP_PROTO_UDP:
 206         if (_eth_copy_chunk(input_size,
 207                             iov, iovcnt,
 208                             *l4hdr_off, sizeof(l4hdr_info->hdr.udp),
 209                             &l4hdr_info->hdr.udp)) {
 210             l4hdr_info->proto = ETH_L4_HDR_PROTO_UDP;
 211             *l5hdr_off = *l4hdr_off + sizeof(l4hdr_info->hdr.udp);
 212         }
 213         break;
 214
 215     case IP_PROTO_SCTP:
 216         l4hdr_info->proto = ETH_L4_HDR_PROTO_SCTP;
 217         break;
 218     }
 219 }
 220
 221 size_t
 222 eth_strip_vlan(const struct iovec *iov, int iovcnt, size_t iovoff,
 223                void *new_ehdr_buf,
 224                uint16_t *payload_offset, uint16_t *tci)
 225 {
 226     struct vlan_header vlan_hdr;
 227     struct eth_header *new_ehdr = new_ehdr_buf;
 228
 229     size_t copied = iov_to_buf(iov, iovcnt, iovoff,
 230                                new_ehdr, sizeof(*new_ehdr));
 231
 232     if (copied < sizeof(*new_ehdr)) {
 233         return 0;
 234     }
 235
 236     switch (be16_to_cpu(new_ehdr->h_proto)) {
 237     case ETH_P_VLAN:
 238     case ETH_P_DVLAN:
 239         copied = iov_to_buf(iov, iovcnt, iovoff + sizeof(*new_ehdr),
 240                             &vlan_hdr, sizeof(vlan_hdr));
 241
 242         if (copied < sizeof(vlan_hdr)) {
 243             return 0;
 244         }
 245
 246         new_ehdr->h_proto = vlan_hdr.h_proto;
 247
 248         *tci = be16_to_cpu(vlan_hdr.h_tci);
 249         *payload_offset = iovoff + sizeof(*new_ehdr) + sizeof(vlan_hdr);
 250
 251         if (be16_to_cpu(new_ehdr->h_proto) == ETH_P_VLAN) {
 252
 253             copied = iov_to_buf(iov, iovcnt, *payload_offset,
 254                                 PKT_GET_VLAN_HDR(new_ehdr), sizeof(vlan_hdr));
 255
 256             if (copied < sizeof(vlan_hdr)) {
 257                 return 0;
 258             }
 259
 260             *payload_offset += sizeof(vlan_hdr);
 261
 262             return sizeof(struct eth_header) + sizeof(struct vlan_header);
 263         } else {
 264             return sizeof(struct eth_header);
 265         }
 266     default:
 267         return 0;
 268     }
 269 }
 270
 271 size_t
 272 eth_strip_vlan_ex(const struct iovec *iov, int iovcnt, size_t iovoff, int index,
 273                   uint16_t vet, uint16_t vet_ext, void *new_ehdr_buf,
 274                   uint16_t *payload_offset, uint16_t *tci)
 275 {
 276     struct vlan_header vlan_hdr;
 277     uint16_t *new_ehdr_proto;
 278     size_t new_ehdr_size;
 279     size_t copied;
 280
 281     switch (index) {
 282     case 0:
 283         new_ehdr_proto = &PKT_GET_ETH_HDR(new_ehdr_buf)->h_proto;
 284         new_ehdr_size = sizeof(struct eth_header);
 285         copied = iov_to_buf(iov, iovcnt, iovoff, new_ehdr_buf, new_ehdr_size);
 286         break;
 287
 288     case 1:
 289         new_ehdr_proto = &PKT_GET_VLAN_HDR(new_ehdr_buf)->h_proto;
 290         new_ehdr_size = sizeof(struct eth_header) + sizeof(struct vlan_header);
 291         copied = iov_to_buf(iov, iovcnt, iovoff, new_ehdr_buf, new_ehdr_size);
 292         if (be16_to_cpu(PKT_GET_ETH_HDR(new_ehdr_buf)->h_proto) != vet_ext) {
 293             return 0;
 294         }
 295         break;
 296
 297     default:
 298         return 0;
 299     }
 300
 301     if (copied < new_ehdr_size || be16_to_cpu(*new_ehdr_proto) != vet) {
 302         return 0;
 303     }
 304
 305     copied = iov_to_buf(iov, iovcnt, iovoff + new_ehdr_size,
 306                         &vlan_hdr, sizeof(vlan_hdr));
 307     if (copied < sizeof(vlan_hdr)) {
 308         return 0;
 309     }
 310
 311     *new_ehdr_proto = vlan_hdr.h_proto;
 312     *payload_offset = iovoff + new_ehdr_size + sizeof(vlan_hdr);
 313     *tci = be16_to_cpu(vlan_hdr.h_tci);
 314
 315     return new_ehdr_size;
 316 }
 317
 318 void
 319 eth_fix_ip4_checksum(void *l3hdr, size_t l3hdr_len)
 320 {
 321     struct ip_header *iphdr = (struct ip_header *) l3hdr;
 322     iphdr->ip_sum = 0;
 323     iphdr->ip_sum = cpu_to_be16(net_raw_checksum(l3hdr, l3hdr_len));
 324 }
 325
 326 uint32_t
 327 eth_calc_ip4_pseudo_hdr_csum(struct ip_header *iphdr,
 328                              uint16_t csl,
 329                              uint32_t *cso)
 330 {
 331     struct ip_pseudo_header ipph;
 332     ipph.ip_src = iphdr->ip_src;
 333     ipph.ip_dst = iphdr->ip_dst;
 334     ipph.ip_payload = cpu_to_be16(csl);
 335     ipph.ip_proto = iphdr->ip_p;
 336     ipph.zeros = 0;
 337     *cso = sizeof(ipph);
 338     return net_checksum_add(*cso, (uint8_t *) &ipph);
 339 }
 340
 341 uint32_t
 342 eth_calc_ip6_pseudo_hdr_csum(struct ip6_header *iphdr,
 343                              uint16_t csl,
 344                              uint8_t l4_proto,
 345                              uint32_t *cso)
 346 {
 347     struct ip6_pseudo_header ipph;
 348     ipph.ip6_src = iphdr->ip6_src;
 349     ipph.ip6_dst = iphdr->ip6_dst;
 350     ipph.len = cpu_to_be16(csl);
 351     ipph.zero[0] = 0;
 352     ipph.zero[1] = 0;
 353     ipph.zero[2] = 0;
 354     ipph.next_hdr = l4_proto;
 355     *cso = sizeof(ipph);
 356     return net_checksum_add(*cso, (uint8_t *)&ipph);
 357 }
 358
 359 static bool
 360 eth_is_ip6_extension_header_type(uint8_t hdr_type)
 361 {
 362     switch (hdr_type) {
 363     case IP6_HOP_BY_HOP:
 364     case IP6_ROUTING:
 365     case IP6_FRAGMENT:
 366     case IP6_AUTHENTICATION:
 367     case IP6_DESTINATON:
 368     case IP6_MOBILITY:
 369         return true;
 370     default:
 371         return false;
 372     }
 373 }
 374
 375 static bool
 376 _eth_get_rss_ex_dst_addr(const struct iovec *pkt, int pkt_frags,
 377                         size_t ext_hdr_offset,
 378                         struct ip6_ext_hdr *ext_hdr,
 379                         struct in6_address *dst_addr)
 380 {
 381     struct ip6_ext_hdr_routing rt_hdr;
 382     size_t input_size = iov_size(pkt, pkt_frags);
 383     size_t bytes_read;
 384
 385     if (input_size < ext_hdr_offset + sizeof(rt_hdr) + sizeof(*dst_addr)) {
 386         return false;
 387     }
 388
 389     bytes_read = iov_to_buf(pkt, pkt_frags, ext_hdr_offset,
 390                             &rt_hdr, sizeof(rt_hdr));
 391     assert(bytes_read == sizeof(rt_hdr));
 392     if ((rt_hdr.rtype != 2) || (rt_hdr.segleft != 1)) {
 393         return false;
 394     }
 395     bytes_read = iov_to_buf(pkt, pkt_frags, ext_hdr_offset + sizeof(rt_hdr),
 396                             dst_addr, sizeof(*dst_addr));
 397     assert(bytes_read == sizeof(*dst_addr));
 398
 399     return true;
 400 }
 401
 402 static bool
 403 _eth_get_rss_ex_src_addr(const struct iovec *pkt, int pkt_frags,
 404                         size_t dsthdr_offset,
 405                         struct ip6_ext_hdr *ext_hdr,
 406                         struct in6_address *src_addr)
 407 {
 408     size_t bytes_left = (ext_hdr->ip6r_len + 1) * 8 - sizeof(*ext_hdr);
 409     struct ip6_option_hdr opthdr;
 410     size_t opt_offset = dsthdr_offset + sizeof(*ext_hdr);
 411
 412     while (bytes_left > sizeof(opthdr)) {
 413         size_t input_size = iov_size(pkt, pkt_frags);
 414         size_t bytes_read, optlen;
 415
 416         if (input_size < opt_offset) {
 417             return false;
 418         }
 419
 420         bytes_read = iov_to_buf(pkt, pkt_frags, opt_offset,
 421                                 &opthdr, sizeof(opthdr));
 422
 423         if (bytes_read != sizeof(opthdr)) {
 424             return false;
 425         }
 426
 427         optlen = (opthdr.type == IP6_OPT_PAD1) ? 1
 428                                                : (opthdr.len + sizeof(opthdr));
 429
 430         if (optlen > bytes_left) {
 431             return false;
 432         }
 433
 434         if (opthdr.type == IP6_OPT_HOME) {
 435             if (input_size < opt_offset + sizeof(opthdr)) {
 436                 return false;
 437             }
 438
 439             bytes_read = iov_to_buf(pkt, pkt_frags,
 440                                     opt_offset + sizeof(opthdr),
 441                                     src_addr, sizeof(*src_addr));
 442
 443             return bytes_read == sizeof(*src_addr);
 444         }
 445
 446         opt_offset += optlen;
 447         bytes_left -= optlen;
 448     }
 449
 450     return false;
 451 }
 452
 453 bool eth_parse_ipv6_hdr(const struct iovec *pkt, int pkt_frags,
 454                         size_t ip6hdr_off, eth_ip6_hdr_info *info)
 455 {
 456     struct ip6_ext_hdr ext_hdr;
 457     size_t bytes_read;
 458     uint8_t curr_ext_hdr_type;
 459     size_t input_size = iov_size(pkt, pkt_frags);
 460
 461     info->rss_ex_dst_valid = false;
 462     info->rss_ex_src_valid = false;
 463     info->fragment = false;
 464
 465     if (input_size < ip6hdr_off) {
 466         return false;
 467     }
 468
 469     bytes_read = iov_to_buf(pkt, pkt_frags, ip6hdr_off,
 470                             &info->ip6_hdr, sizeof(info->ip6_hdr));
 471     if (bytes_read < sizeof(info->ip6_hdr)) {
 472         return false;
 473     }
 474
 475     info->full_hdr_len = sizeof(struct ip6_header);
 476
 477     curr_ext_hdr_type = info->ip6_hdr.ip6_nxt;
 478
 479     if (!eth_is_ip6_extension_header_type(curr_ext_hdr_type)) {
 480         info->l4proto = info->ip6_hdr.ip6_nxt;
 481         info->has_ext_hdrs = false;
 482         return true;
 483     }
 484
 485     info->has_ext_hdrs = true;
 486
 487     do {
 488         if (input_size < ip6hdr_off + info->full_hdr_len) {
 489             return false;
 490         }
 491
 492         bytes_read = iov_to_buf(pkt, pkt_frags, ip6hdr_off + info->full_hdr_len,
 493                                 &ext_hdr, sizeof(ext_hdr));
 494
 495         if (bytes_read < sizeof(ext_hdr)) {
 496             return false;
 497         }
 498
 499         if (curr_ext_hdr_type == IP6_ROUTING) {
 500             if (ext_hdr.ip6r_len == sizeof(struct in6_address) / 8) {
 501                 info->rss_ex_dst_valid =
 502                     _eth_get_rss_ex_dst_addr(pkt, pkt_frags,
 503                                              ip6hdr_off + info->full_hdr_len,
 504                                              &ext_hdr, &info->rss_ex_dst);
 505             }
 506         } else if (curr_ext_hdr_type == IP6_DESTINATON) {
 507             info->rss_ex_src_valid =
 508                 _eth_get_rss_ex_src_addr(pkt, pkt_frags,
 509                                          ip6hdr_off + info->full_hdr_len,
 510                                          &ext_hdr, &info->rss_ex_src);
 511         } else if (curr_ext_hdr_type == IP6_FRAGMENT) {
 512             info->fragment = true;
 513         }
 514
 515         info->full_hdr_len += (ext_hdr.ip6r_len + 1) * IP6_EXT_GRANULARITY;
 516         curr_ext_hdr_type = ext_hdr.ip6r_nxt;
 517     } while (eth_is_ip6_extension_header_type(curr_ext_hdr_type));
 518
 519     info->l4proto = ext_hdr.ip6r_nxt;
 520     return true;
 521 }
 522
 523 bool eth_pad_short_frame(uint8_t *padded_pkt, size_t *padded_buflen,
 524                          const void *pkt, size_t pkt_size)
 525 {
 526     assert(padded_buflen && *padded_buflen >= ETH_ZLEN);
 527
 528     if (pkt_size >= ETH_ZLEN) {
 529         return false;
 530     }
 531
 532     /* pad to minimum Ethernet frame length */
 533     memcpy(padded_pkt, pkt, pkt_size);
 534     memset(&padded_pkt[pkt_size], 0, ETH_ZLEN - pkt_size);
 535     *padded_buflen = ETH_ZLEN;
 536
 537     return true;
 538 }