net/netfilter/nft_tunnel.c

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #include <linux/kernel.h>
   3 #include <linux/init.h>
   4 #include <linux/module.h>
   5 #include <linux/seqlock.h>
   6 #include <linux/netlink.h>
   7 #include <linux/netfilter.h>
   8 #include <linux/netfilter/nf_tables.h>
   9 #include <net/netfilter/nf_tables.h>
  10 #include <net/dst_metadata.h>
  11 #include <net/ip_tunnels.h>
  12 #include <net/vxlan.h>
  13 #include <net/erspan.h>
  14 #include <net/geneve.h>
  15
  16 struct nft_tunnel {
  17         enum nft_tunnel_keys    key:8;
  18         u8                      dreg;
  19         enum nft_tunnel_mode    mode:8;
  20         u8                      len;
  21 };
  22
  23 static void nft_tunnel_get_eval(const struct nft_expr *expr,
  24                                 struct nft_regs *regs,
  25                                 const struct nft_pktinfo *pkt)
  26 {
  27         const struct nft_tunnel *priv = nft_expr_priv(expr);
  28         u32 *dest = &regs->data[priv->dreg];
  29         struct ip_tunnel_info *tun_info;
  30
  31         tun_info = skb_tunnel_info(pkt->skb);
  32
  33         switch (priv->key) {
  34         case NFT_TUNNEL_PATH:
  35                 if (!tun_info) {
  36                         nft_reg_store8(dest, false);
  37                         return;
  38                 }
  39                 if (priv->mode == NFT_TUNNEL_MODE_NONE ||
  40                     (priv->mode == NFT_TUNNEL_MODE_RX &&
  41                      !(tun_info->mode & IP_TUNNEL_INFO_TX)) ||
  42                     (priv->mode == NFT_TUNNEL_MODE_TX &&
  43                      (tun_info->mode & IP_TUNNEL_INFO_TX)))
  44                         nft_reg_store8(dest, true);
  45                 else
  46                         nft_reg_store8(dest, false);
  47                 break;
  48         case NFT_TUNNEL_ID:
  49                 if (!tun_info) {
  50                         regs->verdict.code = NFT_BREAK;
  51                         return;
  52                 }
  53                 if (priv->mode == NFT_TUNNEL_MODE_NONE ||
  54                     (priv->mode == NFT_TUNNEL_MODE_RX &&
  55                      !(tun_info->mode & IP_TUNNEL_INFO_TX)) ||
  56                     (priv->mode == NFT_TUNNEL_MODE_TX &&
  57                      (tun_info->mode & IP_TUNNEL_INFO_TX)))
  58                         *dest = ntohl(tunnel_id_to_key32(tun_info->key.tun_id));
  59                 else
  60                         regs->verdict.code = NFT_BREAK;
  61                 break;
  62         default:
  63                 WARN_ON(1);
  64                 regs->verdict.code = NFT_BREAK;
  65         }
  66 }
  67
  68 static const struct nla_policy nft_tunnel_policy[NFTA_TUNNEL_MAX + 1] = {
  69         [NFTA_TUNNEL_KEY]       = NLA_POLICY_MAX(NLA_BE32, 255),
  70         [NFTA_TUNNEL_DREG]      = { .type = NLA_U32 },
  71         [NFTA_TUNNEL_MODE]      = NLA_POLICY_MAX(NLA_BE32, 255),
  72 };
  73
  74 static int nft_tunnel_get_init(const struct nft_ctx *ctx,
  75                                const struct nft_expr *expr,
  76                                const struct nlattr * const tb[])
  77 {
  78         struct nft_tunnel *priv = nft_expr_priv(expr);
  79         u32 len;
  80
  81         if (!tb[NFTA_TUNNEL_KEY] ||
  82             !tb[NFTA_TUNNEL_DREG])
  83                 return -EINVAL;
  84
  85         priv->key = ntohl(nla_get_be32(tb[NFTA_TUNNEL_KEY]));
  86         switch (priv->key) {
  87         case NFT_TUNNEL_PATH:
  88                 len = sizeof(u8);
  89                 break;
  90         case NFT_TUNNEL_ID:
  91                 len = sizeof(u32);
  92                 break;
  93         default:
  94                 return -EOPNOTSUPP;
  95         }
  96
  97         if (tb[NFTA_TUNNEL_MODE]) {
  98                 priv->mode = ntohl(nla_get_be32(tb[NFTA_TUNNEL_MODE]));
  99                 if (priv->mode > NFT_TUNNEL_MODE_MAX)
 100                         return -EOPNOTSUPP;
 101         } else {
 102                 priv->mode = NFT_TUNNEL_MODE_NONE;
 103         }
 104
 105         priv->len = len;
 106         return nft_parse_register_store(ctx, tb[NFTA_TUNNEL_DREG], &priv->dreg,
 107                                         NULL, NFT_DATA_VALUE, len);
 108 }
 109
 110 static int nft_tunnel_get_dump(struct sk_buff *skb,
 111                                const struct nft_expr *expr, bool reset)
 112 {
 113         const struct nft_tunnel *priv = nft_expr_priv(expr);
 114
 115         if (nla_put_be32(skb, NFTA_TUNNEL_KEY, htonl(priv->key)))
 116                 goto nla_put_failure;
 117         if (nft_dump_register(skb, NFTA_TUNNEL_DREG, priv->dreg))
 118                 goto nla_put_failure;
 119         if (nla_put_be32(skb, NFTA_TUNNEL_MODE, htonl(priv->mode)))
 120                 goto nla_put_failure;
 121         return 0;
 122
 123 nla_put_failure:
 124         return -1;
 125 }
 126
 127 static bool nft_tunnel_get_reduce(struct nft_regs_track *track,
 128                                   const struct nft_expr *expr)
 129 {
 130         const struct nft_tunnel *priv = nft_expr_priv(expr);
 131         const struct nft_tunnel *tunnel;
 132
 133         if (!nft_reg_track_cmp(track, expr, priv->dreg)) {
 134                 nft_reg_track_update(track, expr, priv->dreg, priv->len);
 135                 return false;
 136         }
 137
 138         tunnel = nft_expr_priv(track->regs[priv->dreg].selector);
 139         if (priv->key != tunnel->key ||
 140             priv->dreg != tunnel->dreg ||
 141             priv->mode != tunnel->mode) {
 142                 nft_reg_track_update(track, expr, priv->dreg, priv->len);
 143                 return false;
 144         }
 145
 146         if (!track->regs[priv->dreg].bitwise)
 147                 return true;
 148
 149         return false;
 150 }
 151
 152 static struct nft_expr_type nft_tunnel_type;
 153 static const struct nft_expr_ops nft_tunnel_get_ops = {
 154         .type           = &nft_tunnel_type,
 155         .size           = NFT_EXPR_SIZE(sizeof(struct nft_tunnel)),
 156         .eval           = nft_tunnel_get_eval,
 157         .init           = nft_tunnel_get_init,
 158         .dump           = nft_tunnel_get_dump,
 159         .reduce         = nft_tunnel_get_reduce,
 160 };
 161
 162 static struct nft_expr_type nft_tunnel_type __read_mostly = {
 163         .name           = "tunnel",
 164         .family         = NFPROTO_NETDEV,
 165         .ops            = &nft_tunnel_get_ops,
 166         .policy         = nft_tunnel_policy,
 167         .maxattr        = NFTA_TUNNEL_MAX,
 168         .owner          = THIS_MODULE,
 169 };
 170
 171 struct nft_tunnel_opts {
 172         union {
 173                 struct vxlan_metadata   vxlan;
 174                 struct erspan_metadata  erspan;
 175                 u8      data[IP_TUNNEL_OPTS_MAX];
 176         } u;
 177         u32     len;
 178         __be16  flags;
 179 };
 180
 181 struct nft_tunnel_obj {
 182         struct metadata_dst     *md;
 183         struct nft_tunnel_opts  opts;
 184 };
 185
 186 static const struct nla_policy nft_tunnel_ip_policy[NFTA_TUNNEL_KEY_IP_MAX + 1] = {
 187         [NFTA_TUNNEL_KEY_IP_SRC]        = { .type = NLA_U32 },
 188         [NFTA_TUNNEL_KEY_IP_DST]        = { .type = NLA_U32 },
 189 };
 190
 191 static int nft_tunnel_obj_ip_init(const struct nft_ctx *ctx,
 192                                   const struct nlattr *attr,
 193                                   struct ip_tunnel_info *info)
 194 {
 195         struct nlattr *tb[NFTA_TUNNEL_KEY_IP_MAX + 1];
 196         int err;
 197
 198         err = nla_parse_nested_deprecated(tb, NFTA_TUNNEL_KEY_IP_MAX, attr,
 199                                           nft_tunnel_ip_policy, NULL);
 200         if (err < 0)
 201                 return err;
 202
 203         if (!tb[NFTA_TUNNEL_KEY_IP_DST])
 204                 return -EINVAL;
 205
 206         if (tb[NFTA_TUNNEL_KEY_IP_SRC])
 207                 info->key.u.ipv4.src = nla_get_be32(tb[NFTA_TUNNEL_KEY_IP_SRC]);
 208         if (tb[NFTA_TUNNEL_KEY_IP_DST])
 209                 info->key.u.ipv4.dst = nla_get_be32(tb[NFTA_TUNNEL_KEY_IP_DST]);
 210
 211         return 0;
 212 }
 213
 214 static const struct nla_policy nft_tunnel_ip6_policy[NFTA_TUNNEL_KEY_IP6_MAX + 1] = {
 215         [NFTA_TUNNEL_KEY_IP6_SRC]       = { .len = sizeof(struct in6_addr), },
 216         [NFTA_TUNNEL_KEY_IP6_DST]       = { .len = sizeof(struct in6_addr), },
 217         [NFTA_TUNNEL_KEY_IP6_FLOWLABEL] = { .type = NLA_U32, }
 218 };
 219
 220 static int nft_tunnel_obj_ip6_init(const struct nft_ctx *ctx,
 221                                    const struct nlattr *attr,
 222                                    struct ip_tunnel_info *info)
 223 {
 224         struct nlattr *tb[NFTA_TUNNEL_KEY_IP6_MAX + 1];
 225         int err;
 226
 227         err = nla_parse_nested_deprecated(tb, NFTA_TUNNEL_KEY_IP6_MAX, attr,
 228                                           nft_tunnel_ip6_policy, NULL);
 229         if (err < 0)
 230                 return err;
 231
 232         if (!tb[NFTA_TUNNEL_KEY_IP6_DST])
 233                 return -EINVAL;
 234
 235         if (tb[NFTA_TUNNEL_KEY_IP6_SRC]) {
 236                 memcpy(&info->key.u.ipv6.src,
 237                        nla_data(tb[NFTA_TUNNEL_KEY_IP6_SRC]),
 238                        sizeof(struct in6_addr));
 239         }
 240         if (tb[NFTA_TUNNEL_KEY_IP6_DST]) {
 241                 memcpy(&info->key.u.ipv6.dst,
 242                        nla_data(tb[NFTA_TUNNEL_KEY_IP6_DST]),
 243                        sizeof(struct in6_addr));
 244         }
 245         if (tb[NFTA_TUNNEL_KEY_IP6_FLOWLABEL])
 246                 info->key.label = nla_get_be32(tb[NFTA_TUNNEL_KEY_IP6_FLOWLABEL]);
 247
 248         info->mode |= IP_TUNNEL_INFO_IPV6;
 249
 250         return 0;
 251 }
 252
 253 static const struct nla_policy nft_tunnel_opts_vxlan_policy[NFTA_TUNNEL_KEY_VXLAN_MAX + 1] = {
 254         [NFTA_TUNNEL_KEY_VXLAN_GBP]     = { .type = NLA_U32 },
 255 };
 256
 257 static int nft_tunnel_obj_vxlan_init(const struct nlattr *attr,
 258                                      struct nft_tunnel_opts *opts)
 259 {
 260         struct nlattr *tb[NFTA_TUNNEL_KEY_VXLAN_MAX + 1];
 261         int err;
 262
 263         err = nla_parse_nested_deprecated(tb, NFTA_TUNNEL_KEY_VXLAN_MAX, attr,
 264                                           nft_tunnel_opts_vxlan_policy, NULL);
 265         if (err < 0)
 266                 return err;
 267
 268         if (!tb[NFTA_TUNNEL_KEY_VXLAN_GBP])
 269                 return -EINVAL;
 270
 271         opts->u.vxlan.gbp = ntohl(nla_get_be32(tb[NFTA_TUNNEL_KEY_VXLAN_GBP]));
 272
 273         opts->len       = sizeof(struct vxlan_metadata);
 274         opts->flags     = TUNNEL_VXLAN_OPT;
 275
 276         return 0;
 277 }
 278
 279 static const struct nla_policy nft_tunnel_opts_erspan_policy[NFTA_TUNNEL_KEY_ERSPAN_MAX + 1] = {
 280         [NFTA_TUNNEL_KEY_ERSPAN_VERSION]        = { .type = NLA_U32 },
 281         [NFTA_TUNNEL_KEY_ERSPAN_V1_INDEX]       = { .type = NLA_U32 },
 282         [NFTA_TUNNEL_KEY_ERSPAN_V2_DIR]         = { .type = NLA_U8 },
 283         [NFTA_TUNNEL_KEY_ERSPAN_V2_HWID]        = { .type = NLA_U8 },
 284 };
 285
 286 static int nft_tunnel_obj_erspan_init(const struct nlattr *attr,
 287                                       struct nft_tunnel_opts *opts)
 288 {
 289         struct nlattr *tb[NFTA_TUNNEL_KEY_ERSPAN_MAX + 1];
 290         uint8_t hwid, dir;
 291         int err, version;
 292
 293         err = nla_parse_nested_deprecated(tb, NFTA_TUNNEL_KEY_ERSPAN_MAX,
 294                                           attr, nft_tunnel_opts_erspan_policy,
 295                                           NULL);
 296         if (err < 0)
 297                 return err;
 298
 299         if (!tb[NFTA_TUNNEL_KEY_ERSPAN_VERSION])
 300                  return -EINVAL;
 301
 302         version = ntohl(nla_get_be32(tb[NFTA_TUNNEL_KEY_ERSPAN_VERSION]));
 303         switch (version) {
 304         case ERSPAN_VERSION:
 305                 if (!tb[NFTA_TUNNEL_KEY_ERSPAN_V1_INDEX])
 306                         return -EINVAL;
 307
 308                 opts->u.erspan.u.index =
 309                         nla_get_be32(tb[NFTA_TUNNEL_KEY_ERSPAN_V1_INDEX]);
 310                 break;
 311         case ERSPAN_VERSION2:
 312                 if (!tb[NFTA_TUNNEL_KEY_ERSPAN_V2_DIR] ||
 313                     !tb[NFTA_TUNNEL_KEY_ERSPAN_V2_HWID])
 314                         return -EINVAL;
 315
 316                 hwid = nla_get_u8(tb[NFTA_TUNNEL_KEY_ERSPAN_V2_HWID]);
 317                 dir = nla_get_u8(tb[NFTA_TUNNEL_KEY_ERSPAN_V2_DIR]);
 318
 319                 set_hwid(&opts->u.erspan.u.md2, hwid);
 320                 opts->u.erspan.u.md2.dir = dir;
 321                 break;
 322         default:
 323                 return -EOPNOTSUPP;
 324         }
 325         opts->u.erspan.version = version;
 326
 327         opts->len       = sizeof(struct erspan_metadata);
 328         opts->flags     = TUNNEL_ERSPAN_OPT;
 329
 330         return 0;
 331 }
 332
 333 static const struct nla_policy nft_tunnel_opts_geneve_policy[NFTA_TUNNEL_KEY_GENEVE_MAX + 1] = {
 334         [NFTA_TUNNEL_KEY_GENEVE_CLASS]  = { .type = NLA_U16 },
 335         [NFTA_TUNNEL_KEY_GENEVE_TYPE]   = { .type = NLA_U8 },
 336         [NFTA_TUNNEL_KEY_GENEVE_DATA]   = { .type = NLA_BINARY, .len = 128 },
 337 };
 338
 339 static int nft_tunnel_obj_geneve_init(const struct nlattr *attr,
 340                                       struct nft_tunnel_opts *opts)
 341 {
 342         struct geneve_opt *opt = (struct geneve_opt *)opts->u.data + opts->len;
 343         struct nlattr *tb[NFTA_TUNNEL_KEY_GENEVE_MAX + 1];
 344         int err, data_len;
 345
 346         err = nla_parse_nested(tb, NFTA_TUNNEL_KEY_GENEVE_MAX, attr,
 347                                nft_tunnel_opts_geneve_policy, NULL);
 348         if (err < 0)
 349                 return err;
 350
 351         if (!tb[NFTA_TUNNEL_KEY_GENEVE_CLASS] ||
 352             !tb[NFTA_TUNNEL_KEY_GENEVE_TYPE] ||
 353             !tb[NFTA_TUNNEL_KEY_GENEVE_DATA])
 354                 return -EINVAL;
 355
 356         attr = tb[NFTA_TUNNEL_KEY_GENEVE_DATA];
 357         data_len = nla_len(attr);
 358         if (data_len % 4)
 359                 return -EINVAL;
 360
 361         opts->len += sizeof(*opt) + data_len;
 362         if (opts->len > IP_TUNNEL_OPTS_MAX)
 363                 return -EINVAL;
 364
 365         memcpy(opt->opt_data, nla_data(attr), data_len);
 366         opt->length = data_len / 4;
 367         opt->opt_class = nla_get_be16(tb[NFTA_TUNNEL_KEY_GENEVE_CLASS]);
 368         opt->type = nla_get_u8(tb[NFTA_TUNNEL_KEY_GENEVE_TYPE]);
 369         opts->flags = TUNNEL_GENEVE_OPT;
 370
 371         return 0;
 372 }
 373
 374 static const struct nla_policy nft_tunnel_opts_policy[NFTA_TUNNEL_KEY_OPTS_MAX + 1] = {
 375         [NFTA_TUNNEL_KEY_OPTS_UNSPEC]   = {
 376                 .strict_start_type = NFTA_TUNNEL_KEY_OPTS_GENEVE },
 377         [NFTA_TUNNEL_KEY_OPTS_VXLAN]    = { .type = NLA_NESTED, },
 378         [NFTA_TUNNEL_KEY_OPTS_ERSPAN]   = { .type = NLA_NESTED, },
 379         [NFTA_TUNNEL_KEY_OPTS_GENEVE]   = { .type = NLA_NESTED, },
 380 };
 381
 382 static int nft_tunnel_obj_opts_init(const struct nft_ctx *ctx,
 383                                     const struct nlattr *attr,
 384                                     struct ip_tunnel_info *info,
 385                                     struct nft_tunnel_opts *opts)
 386 {
 387         struct nlattr *nla;
 388         __be16 type = 0;
 389         int err, rem;
 390
 391         err = nla_validate_nested_deprecated(attr, NFTA_TUNNEL_KEY_OPTS_MAX,
 392                                              nft_tunnel_opts_policy, NULL);
 393         if (err < 0)
 394                 return err;
 395
 396         nla_for_each_attr(nla, nla_data(attr), nla_len(attr), rem) {
 397                 switch (nla_type(nla)) {
 398                 case NFTA_TUNNEL_KEY_OPTS_VXLAN:
 399                         if (type)
 400                                 return -EINVAL;
 401                         err = nft_tunnel_obj_vxlan_init(nla, opts);
 402                         if (err)
 403                                 return err;
 404                         type = TUNNEL_VXLAN_OPT;
 405                         break;
 406                 case NFTA_TUNNEL_KEY_OPTS_ERSPAN:
 407                         if (type)
 408                                 return -EINVAL;
 409                         err = nft_tunnel_obj_erspan_init(nla, opts);
 410                         if (err)
 411                                 return err;
 412                         type = TUNNEL_ERSPAN_OPT;
 413                         break;
 414                 case NFTA_TUNNEL_KEY_OPTS_GENEVE:
 415                         if (type && type != TUNNEL_GENEVE_OPT)
 416                                 return -EINVAL;
 417                         err = nft_tunnel_obj_geneve_init(nla, opts);
 418                         if (err)
 419                                 return err;
 420                         type = TUNNEL_GENEVE_OPT;
 421                         break;
 422                 default:
 423                         return -EOPNOTSUPP;
 424                 }
 425         }
 426
 427         return err;
 428 }
 429
 430 static const struct nla_policy nft_tunnel_key_policy[NFTA_TUNNEL_KEY_MAX + 1] = {
 431         [NFTA_TUNNEL_KEY_IP]    = { .type = NLA_NESTED, },
 432         [NFTA_TUNNEL_KEY_IP6]   = { .type = NLA_NESTED, },
 433         [NFTA_TUNNEL_KEY_ID]    = { .type = NLA_U32, },
 434         [NFTA_TUNNEL_KEY_FLAGS] = { .type = NLA_U32, },
 435         [NFTA_TUNNEL_KEY_TOS]   = { .type = NLA_U8, },
 436         [NFTA_TUNNEL_KEY_TTL]   = { .type = NLA_U8, },
 437         [NFTA_TUNNEL_KEY_SPORT] = { .type = NLA_U16, },
 438         [NFTA_TUNNEL_KEY_DPORT] = { .type = NLA_U16, },
 439         [NFTA_TUNNEL_KEY_OPTS]  = { .type = NLA_NESTED, },
 440 };
 441
 442 static int nft_tunnel_obj_init(const struct nft_ctx *ctx,
 443                                const struct nlattr * const tb[],
 444                                struct nft_object *obj)
 445 {
 446         struct nft_tunnel_obj *priv = nft_obj_data(obj);
 447         struct ip_tunnel_info info;
 448         struct metadata_dst *md;
 449         int err;
 450
 451         if (!tb[NFTA_TUNNEL_KEY_ID])
 452                 return -EINVAL;
 453
 454         memset(&info, 0, sizeof(info));
 455         info.mode               = IP_TUNNEL_INFO_TX;
 456         info.key.tun_id         = key32_to_tunnel_id(nla_get_be32(tb[NFTA_TUNNEL_KEY_ID]));
 457         info.key.tun_flags      = TUNNEL_KEY | TUNNEL_CSUM | TUNNEL_NOCACHE;
 458
 459         if (tb[NFTA_TUNNEL_KEY_IP]) {
 460                 err = nft_tunnel_obj_ip_init(ctx, tb[NFTA_TUNNEL_KEY_IP], &info);
 461                 if (err < 0)
 462                         return err;
 463         } else if (tb[NFTA_TUNNEL_KEY_IP6]) {
 464                 err = nft_tunnel_obj_ip6_init(ctx, tb[NFTA_TUNNEL_KEY_IP6], &info);
 465                 if (err < 0)
 466                         return err;
 467         } else {
 468                 return -EINVAL;
 469         }
 470
 471         if (tb[NFTA_TUNNEL_KEY_SPORT]) {
 472                 info.key.tp_src = nla_get_be16(tb[NFTA_TUNNEL_KEY_SPORT]);
 473         }
 474         if (tb[NFTA_TUNNEL_KEY_DPORT]) {
 475                 info.key.tp_dst = nla_get_be16(tb[NFTA_TUNNEL_KEY_DPORT]);
 476         }
 477
 478         if (tb[NFTA_TUNNEL_KEY_FLAGS]) {
 479                 u32 tun_flags;
 480
 481                 tun_flags = ntohl(nla_get_be32(tb[NFTA_TUNNEL_KEY_FLAGS]));
 482                 if (tun_flags & ~NFT_TUNNEL_F_MASK)
 483                         return -EOPNOTSUPP;
 484
 485                 if (tun_flags & NFT_TUNNEL_F_ZERO_CSUM_TX)
 486                         info.key.tun_flags &= ~TUNNEL_CSUM;
 487                 if (tun_flags & NFT_TUNNEL_F_DONT_FRAGMENT)
 488                         info.key.tun_flags |= TUNNEL_DONT_FRAGMENT;
 489                 if (tun_flags & NFT_TUNNEL_F_SEQ_NUMBER)
 490                         info.key.tun_flags |= TUNNEL_SEQ;
 491         }
 492         if (tb[NFTA_TUNNEL_KEY_TOS])
 493                 info.key.tos = nla_get_u8(tb[NFTA_TUNNEL_KEY_TOS]);
 494         if (tb[NFTA_TUNNEL_KEY_TTL])
 495                 info.key.ttl = nla_get_u8(tb[NFTA_TUNNEL_KEY_TTL]);
 496         else
 497                 info.key.ttl = U8_MAX;
 498
 499         if (tb[NFTA_TUNNEL_KEY_OPTS]) {
 500                 err = nft_tunnel_obj_opts_init(ctx, tb[NFTA_TUNNEL_KEY_OPTS],
 501                                                &info, &priv->opts);
 502                 if (err < 0)
 503                         return err;
 504         }
 505
 506         md = metadata_dst_alloc(priv->opts.len, METADATA_IP_TUNNEL, GFP_KERNEL);
 507         if (!md)
 508                 return -ENOMEM;
 509
 510         memcpy(&md->u.tun_info, &info, sizeof(info));
 511 #ifdef CONFIG_DST_CACHE
 512         err = dst_cache_init(&md->u.tun_info.dst_cache, GFP_KERNEL);
 513         if (err < 0) {
 514                 metadata_dst_free(md);
 515                 return err;
 516         }
 517 #endif
 518         ip_tunnel_info_opts_set(&md->u.tun_info, &priv->opts.u, priv->opts.len,
 519                                 priv->opts.flags);
 520         priv->md = md;
 521
 522         return 0;
 523 }
 524
 525 static inline void nft_tunnel_obj_eval(struct nft_object *obj,
 526                                        struct nft_regs *regs,
 527                                        const struct nft_pktinfo *pkt)
 528 {
 529         struct nft_tunnel_obj *priv = nft_obj_data(obj);
 530         struct sk_buff *skb = pkt->skb;
 531
 532         skb_dst_drop(skb);
 533         dst_hold((struct dst_entry *) priv->md);
 534         skb_dst_set(skb, (struct dst_entry *) priv->md);
 535 }
 536
 537 static int nft_tunnel_ip_dump(struct sk_buff *skb, struct ip_tunnel_info *info)
 538 {
 539         struct nlattr *nest;
 540
 541         if (info->mode & IP_TUNNEL_INFO_IPV6) {
 542                 nest = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_IP6);
 543                 if (!nest)
 544                         return -1;
 545
 546                 if (nla_put_in6_addr(skb, NFTA_TUNNEL_KEY_IP6_SRC,
 547                                      &info->key.u.ipv6.src) < 0 ||
 548                     nla_put_in6_addr(skb, NFTA_TUNNEL_KEY_IP6_DST,
 549                                      &info->key.u.ipv6.dst) < 0 ||
 550                     nla_put_be32(skb, NFTA_TUNNEL_KEY_IP6_FLOWLABEL,
 551                                  info->key.label)) {
 552                         nla_nest_cancel(skb, nest);
 553                         return -1;
 554                 }
 555
 556                 nla_nest_end(skb, nest);
 557         } else {
 558                 nest = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_IP);
 559                 if (!nest)
 560                         return -1;
 561
 562                 if (nla_put_in_addr(skb, NFTA_TUNNEL_KEY_IP_SRC,
 563                                     info->key.u.ipv4.src) < 0 ||
 564                     nla_put_in_addr(skb, NFTA_TUNNEL_KEY_IP_DST,
 565                                     info->key.u.ipv4.dst) < 0) {
 566                         nla_nest_cancel(skb, nest);
 567                         return -1;
 568                 }
 569
 570                 nla_nest_end(skb, nest);
 571         }
 572
 573         return 0;
 574 }
 575
 576 static int nft_tunnel_opts_dump(struct sk_buff *skb,
 577                                 struct nft_tunnel_obj *priv)
 578 {
 579         struct nft_tunnel_opts *opts = &priv->opts;
 580         struct nlattr *nest, *inner;
 581
 582         nest = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_OPTS);
 583         if (!nest)
 584                 return -1;
 585
 586         if (opts->flags & TUNNEL_VXLAN_OPT) {
 587                 inner = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_OPTS_VXLAN);
 588                 if (!inner)
 589                         goto failure;
 590                 if (nla_put_be32(skb, NFTA_TUNNEL_KEY_VXLAN_GBP,
 591                                  htonl(opts->u.vxlan.gbp)))
 592                         goto inner_failure;
 593                 nla_nest_end(skb, inner);
 594         } else if (opts->flags & TUNNEL_ERSPAN_OPT) {
 595                 inner = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_OPTS_ERSPAN);
 596                 if (!inner)
 597                         goto failure;
 598                 if (nla_put_be32(skb, NFTA_TUNNEL_KEY_ERSPAN_VERSION,
 599                                  htonl(opts->u.erspan.version)))
 600                         goto inner_failure;
 601                 switch (opts->u.erspan.version) {
 602                 case ERSPAN_VERSION:
 603                         if (nla_put_be32(skb, NFTA_TUNNEL_KEY_ERSPAN_V1_INDEX,
 604                                          opts->u.erspan.u.index))
 605                                 goto inner_failure;
 606                         break;
 607                 case ERSPAN_VERSION2:
 608                         if (nla_put_u8(skb, NFTA_TUNNEL_KEY_ERSPAN_V2_HWID,
 609                                        get_hwid(&opts->u.erspan.u.md2)) ||
 610                             nla_put_u8(skb, NFTA_TUNNEL_KEY_ERSPAN_V2_DIR,
 611                                        opts->u.erspan.u.md2.dir))
 612                                 goto inner_failure;
 613                         break;
 614                 }
 615                 nla_nest_end(skb, inner);
 616         } else if (opts->flags & TUNNEL_GENEVE_OPT) {
 617                 struct geneve_opt *opt;
 618                 int offset = 0;
 619
 620                 inner = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_OPTS_GENEVE);
 621                 if (!inner)
 622                         goto failure;
 623                 while (opts->len > offset) {
 624                         opt = (struct geneve_opt *)opts->u.data + offset;
 625                         if (nla_put_be16(skb, NFTA_TUNNEL_KEY_GENEVE_CLASS,
 626                                          opt->opt_class) ||
 627                             nla_put_u8(skb, NFTA_TUNNEL_KEY_GENEVE_TYPE,
 628                                        opt->type) ||
 629                             nla_put(skb, NFTA_TUNNEL_KEY_GENEVE_DATA,
 630                                     opt->length * 4, opt->opt_data))
 631                                 goto inner_failure;
 632                         offset += sizeof(*opt) + opt->length * 4;
 633                 }
 634                 nla_nest_end(skb, inner);
 635         }
 636         nla_nest_end(skb, nest);
 637         return 0;
 638
 639 inner_failure:
 640         nla_nest_cancel(skb, inner);
 641 failure:
 642         nla_nest_cancel(skb, nest);
 643         return -1;
 644 }
 645
 646 static int nft_tunnel_ports_dump(struct sk_buff *skb,
 647                                  struct ip_tunnel_info *info)
 648 {
 649         if (nla_put_be16(skb, NFTA_TUNNEL_KEY_SPORT, info->key.tp_src) < 0 ||
 650             nla_put_be16(skb, NFTA_TUNNEL_KEY_DPORT, info->key.tp_dst) < 0)
 651                 return -1;
 652
 653         return 0;
 654 }
 655
 656 static int nft_tunnel_flags_dump(struct sk_buff *skb,
 657                                  struct ip_tunnel_info *info)
 658 {
 659         u32 flags = 0;
 660
 661         if (info->key.tun_flags & TUNNEL_DONT_FRAGMENT)
 662                 flags |= NFT_TUNNEL_F_DONT_FRAGMENT;
 663         if (!(info->key.tun_flags & TUNNEL_CSUM))
 664                 flags |= NFT_TUNNEL_F_ZERO_CSUM_TX;
 665         if (info->key.tun_flags & TUNNEL_SEQ)
 666                 flags |= NFT_TUNNEL_F_SEQ_NUMBER;
 667
 668         if (nla_put_be32(skb, NFTA_TUNNEL_KEY_FLAGS, htonl(flags)) < 0)
 669                 return -1;
 670
 671         return 0;
 672 }
 673
 674 static int nft_tunnel_obj_dump(struct sk_buff *skb,
 675                                struct nft_object *obj, bool reset)
 676 {
 677         struct nft_tunnel_obj *priv = nft_obj_data(obj);
 678         struct ip_tunnel_info *info = &priv->md->u.tun_info;
 679
 680         if (nla_put_be32(skb, NFTA_TUNNEL_KEY_ID,
 681                          tunnel_id_to_key32(info->key.tun_id)) ||
 682             nft_tunnel_ip_dump(skb, info) < 0 ||
 683             nft_tunnel_ports_dump(skb, info) < 0 ||
 684             nft_tunnel_flags_dump(skb, info) < 0 ||
 685             nla_put_u8(skb, NFTA_TUNNEL_KEY_TOS, info->key.tos) ||
 686             nla_put_u8(skb, NFTA_TUNNEL_KEY_TTL, info->key.ttl) ||
 687             nft_tunnel_opts_dump(skb, priv) < 0)
 688                 goto nla_put_failure;
 689
 690         return 0;
 691
 692 nla_put_failure:
 693         return -1;
 694 }
 695
 696 static void nft_tunnel_obj_destroy(const struct nft_ctx *ctx,
 697                                    struct nft_object *obj)
 698 {
 699         struct nft_tunnel_obj *priv = nft_obj_data(obj);
 700
 701         metadata_dst_free(priv->md);
 702 }
 703
 704 static struct nft_object_type nft_tunnel_obj_type;
 705 static const struct nft_object_ops nft_tunnel_obj_ops = {
 706         .type           = &nft_tunnel_obj_type,
 707         .size           = sizeof(struct nft_tunnel_obj),
 708         .eval           = nft_tunnel_obj_eval,
 709         .init           = nft_tunnel_obj_init,
 710         .destroy        = nft_tunnel_obj_destroy,
 711         .dump           = nft_tunnel_obj_dump,
 712 };
 713
 714 static struct nft_object_type nft_tunnel_obj_type __read_mostly = {
 715         .type           = NFT_OBJECT_TUNNEL,
 716         .family         = NFPROTO_NETDEV,
 717         .ops            = &nft_tunnel_obj_ops,
 718         .maxattr        = NFTA_TUNNEL_KEY_MAX,
 719         .policy         = nft_tunnel_key_policy,
 720         .owner          = THIS_MODULE,
 721 };
 722
 723 static int __init nft_tunnel_module_init(void)
 724 {
 725         int err;
 726
 727         err = nft_register_expr(&nft_tunnel_type);
 728         if (err < 0)
 729                 return err;
 730
 731         err = nft_register_obj(&nft_tunnel_obj_type);
 732         if (err < 0)
 733                 nft_unregister_expr(&nft_tunnel_type);
 734
 735         return err;
 736 }
 737
 738 static void __exit nft_tunnel_module_exit(void)
 739 {
 740         nft_unregister_obj(&nft_tunnel_obj_type);
 741         nft_unregister_expr(&nft_tunnel_type);
 742 }
 743
 744 module_init(nft_tunnel_module_init);
 745 module_exit(nft_tunnel_module_exit);
 746
 747 MODULE_LICENSE("GPL");
 748 MODULE_AUTHOR("Pablo Neira Ayuso <pablo@netfilter.org>");
 749 MODULE_ALIAS_NFT_EXPR("tunnel");
 750 MODULE_ALIAS_NFT_OBJ(NFT_OBJECT_TUNNEL);
 751 MODULE_DESCRIPTION("nftables tunnel expression support");