net/sched/act_ife.c

   1 /*
   2  * net/sched/ife.c      Inter-FE action based on ForCES WG InterFE LFB
   3  *
   4  *              Refer to:
   5  *              draft-ietf-forces-interfelfb-03
   6  *              and
   7  *              netdev01 paper:
   8  *              "Distributing Linux Traffic Control Classifier-Action
   9  *              Subsystem"
  10  *              Authors: Jamal Hadi Salim and Damascene M. Joachimpillai
  11  *
  12  *              This program is free software; you can redistribute it and/or
  13  *              modify it under the terms of the GNU General Public License
  14  *              as published by the Free Software Foundation; either version
  15  *              2 of the License, or (at your option) any later version.
  16  *
  17  * copyright Jamal Hadi Salim (2015)
  18  *
  19 */
  20
  21 #include <linux/types.h>
  22 #include <linux/kernel.h>
  23 #include <linux/string.h>
  24 #include <linux/errno.h>
  25 #include <linux/skbuff.h>
  26 #include <linux/rtnetlink.h>
  27 #include <linux/module.h>
  28 #include <linux/init.h>
  29 #include <net/net_namespace.h>
  30 #include <net/netlink.h>
  31 #include <net/pkt_sched.h>
  32 #include <uapi/linux/tc_act/tc_ife.h>
  33 #include <net/tc_act/tc_ife.h>
  34 #include <linux/etherdevice.h>
  35
  36 #define IFE_TAB_MASK 15
  37
  38 static int ife_net_id;
  39 static int max_metacnt = IFE_META_MAX + 1;
  40 static struct tc_action_ops act_ife_ops;
  41
  42 static const struct nla_policy ife_policy[TCA_IFE_MAX + 1] = {
  43         [TCA_IFE_PARMS] = { .len = sizeof(struct tc_ife)},
  44         [TCA_IFE_DMAC] = { .len = ETH_ALEN},
  45         [TCA_IFE_SMAC] = { .len = ETH_ALEN},
  46         [TCA_IFE_TYPE] = { .type = NLA_U16},
  47 };
  48
  49 /* Caller takes care of presenting data in network order
  50 */
  51 int ife_tlv_meta_encode(void *skbdata, u16 attrtype, u16 dlen, const void *dval)
  52 {
  53         u32 *tlv = (u32 *)(skbdata);
  54         u16 totlen = nla_total_size(dlen);      /*alignment + hdr */
  55         char *dptr = (char *)tlv + NLA_HDRLEN;
  56         u32 htlv = attrtype << 16 | (dlen + NLA_HDRLEN);
  57
  58         *tlv = htonl(htlv);
  59         memset(dptr, 0, totlen - NLA_HDRLEN);
  60         memcpy(dptr, dval, dlen);
  61
  62         return totlen;
  63 }
  64 EXPORT_SYMBOL_GPL(ife_tlv_meta_encode);
  65
  66 int ife_encode_meta_u16(u16 metaval, void *skbdata, struct tcf_meta_info *mi)
  67 {
  68         u16 edata = 0;
  69
  70         if (mi->metaval)
  71                 edata = *(u16 *)mi->metaval;
  72         else if (metaval)
  73                 edata = metaval;
  74
  75         if (!edata) /* will not encode */
  76                 return 0;
  77
  78         edata = htons(edata);
  79         return ife_tlv_meta_encode(skbdata, mi->metaid, 2, &edata);
  80 }
  81 EXPORT_SYMBOL_GPL(ife_encode_meta_u16);
  82
  83 int ife_get_meta_u32(struct sk_buff *skb, struct tcf_meta_info *mi)
  84 {
  85         if (mi->metaval)
  86                 return nla_put_u32(skb, mi->metaid, *(u32 *)mi->metaval);
  87         else
  88                 return nla_put(skb, mi->metaid, 0, NULL);
  89 }
  90 EXPORT_SYMBOL_GPL(ife_get_meta_u32);
  91
  92 int ife_check_meta_u32(u32 metaval, struct tcf_meta_info *mi)
  93 {
  94         if (metaval || mi->metaval)
  95                 return 8; /* T+L+V == 2+2+4 */
  96
  97         return 0;
  98 }
  99 EXPORT_SYMBOL_GPL(ife_check_meta_u32);
 100
 101 int ife_check_meta_u16(u16 metaval, struct tcf_meta_info *mi)
 102 {
 103         if (metaval || mi->metaval)
 104                 return 8; /* T+L+(V) == 2+2+(2+2bytepad) */
 105
 106         return 0;
 107 }
 108 EXPORT_SYMBOL_GPL(ife_check_meta_u16);
 109
 110 int ife_encode_meta_u32(u32 metaval, void *skbdata, struct tcf_meta_info *mi)
 111 {
 112         u32 edata = metaval;
 113
 114         if (mi->metaval)
 115                 edata = *(u32 *)mi->metaval;
 116         else if (metaval)
 117                 edata = metaval;
 118
 119         if (!edata) /* will not encode */
 120                 return 0;
 121
 122         edata = htonl(edata);
 123         return ife_tlv_meta_encode(skbdata, mi->metaid, 4, &edata);
 124 }
 125 EXPORT_SYMBOL_GPL(ife_encode_meta_u32);
 126
 127 int ife_get_meta_u16(struct sk_buff *skb, struct tcf_meta_info *mi)
 128 {
 129         if (mi->metaval)
 130                 return nla_put_u16(skb, mi->metaid, *(u16 *)mi->metaval);
 131         else
 132                 return nla_put(skb, mi->metaid, 0, NULL);
 133 }
 134 EXPORT_SYMBOL_GPL(ife_get_meta_u16);
 135
 136 int ife_alloc_meta_u32(struct tcf_meta_info *mi, void *metaval, gfp_t gfp)
 137 {
 138         mi->metaval = kmemdup(metaval, sizeof(u32), gfp);
 139         if (!mi->metaval)
 140                 return -ENOMEM;
 141
 142         return 0;
 143 }
 144 EXPORT_SYMBOL_GPL(ife_alloc_meta_u32);
 145
 146 int ife_alloc_meta_u16(struct tcf_meta_info *mi, void *metaval, gfp_t gfp)
 147 {
 148         mi->metaval = kmemdup(metaval, sizeof(u16), gfp);
 149         if (!mi->metaval)
 150                 return -ENOMEM;
 151
 152         return 0;
 153 }
 154 EXPORT_SYMBOL_GPL(ife_alloc_meta_u16);
 155
 156 void ife_release_meta_gen(struct tcf_meta_info *mi)
 157 {
 158         kfree(mi->metaval);
 159 }
 160 EXPORT_SYMBOL_GPL(ife_release_meta_gen);
 161
 162 int ife_validate_meta_u32(void *val, int len)
 163 {
 164         if (len == sizeof(u32))
 165                 return 0;
 166
 167         return -EINVAL;
 168 }
 169 EXPORT_SYMBOL_GPL(ife_validate_meta_u32);
 170
 171 int ife_validate_meta_u16(void *val, int len)
 172 {
 173         /* length will not include padding */
 174         if (len == sizeof(u16))
 175                 return 0;
 176
 177         return -EINVAL;
 178 }
 179 EXPORT_SYMBOL_GPL(ife_validate_meta_u16);
 180
 181 static LIST_HEAD(ifeoplist);
 182 static DEFINE_RWLOCK(ife_mod_lock);
 183
 184 static struct tcf_meta_ops *find_ife_oplist(u16 metaid)
 185 {
 186         struct tcf_meta_ops *o;
 187
 188         read_lock(&ife_mod_lock);
 189         list_for_each_entry(o, &ifeoplist, list) {
 190                 if (o->metaid == metaid) {
 191                         if (!try_module_get(o->owner))
 192                                 o = NULL;
 193                         read_unlock(&ife_mod_lock);
 194                         return o;
 195                 }
 196         }
 197         read_unlock(&ife_mod_lock);
 198
 199         return NULL;
 200 }
 201
 202 int register_ife_op(struct tcf_meta_ops *mops)
 203 {
 204         struct tcf_meta_ops *m;
 205
 206         if (!mops->metaid || !mops->metatype || !mops->name ||
 207             !mops->check_presence || !mops->encode || !mops->decode ||
 208             !mops->get || !mops->alloc)
 209                 return -EINVAL;
 210
 211         write_lock(&ife_mod_lock);
 212
 213         list_for_each_entry(m, &ifeoplist, list) {
 214                 if (m->metaid == mops->metaid ||
 215                     (strcmp(mops->name, m->name) == 0)) {
 216                         write_unlock(&ife_mod_lock);
 217                         return -EEXIST;
 218                 }
 219         }
 220
 221         if (!mops->release)
 222                 mops->release = ife_release_meta_gen;
 223
 224         list_add_tail(&mops->list, &ifeoplist);
 225         write_unlock(&ife_mod_lock);
 226         return 0;
 227 }
 228 EXPORT_SYMBOL_GPL(unregister_ife_op);
 229
 230 int unregister_ife_op(struct tcf_meta_ops *mops)
 231 {
 232         struct tcf_meta_ops *m;
 233         int err = -ENOENT;
 234
 235         write_lock(&ife_mod_lock);
 236         list_for_each_entry(m, &ifeoplist, list) {
 237                 if (m->metaid == mops->metaid) {
 238                         list_del(&mops->list);
 239                         err = 0;
 240                         break;
 241                 }
 242         }
 243         write_unlock(&ife_mod_lock);
 244
 245         return err;
 246 }
 247 EXPORT_SYMBOL_GPL(register_ife_op);
 248
 249 static int ife_validate_metatype(struct tcf_meta_ops *ops, void *val, int len)
 250 {
 251         int ret = 0;
 252         /* XXX: unfortunately cant use nla_policy at this point
 253         * because a length of 0 is valid in the case of
 254         * "allow". "use" semantics do enforce for proper
 255         * length and i couldve use nla_policy but it makes it hard
 256         * to use it just for that..
 257         */
 258         if (ops->validate)
 259                 return ops->validate(val, len);
 260
 261         if (ops->metatype == NLA_U32)
 262                 ret = ife_validate_meta_u32(val, len);
 263         else if (ops->metatype == NLA_U16)
 264                 ret = ife_validate_meta_u16(val, len);
 265
 266         return ret;
 267 }
 268
 269 /* called when adding new meta information
 270 */
 271 static int load_metaops_and_vet(u32 metaid, void *val, int len)
 272 {
 273         struct tcf_meta_ops *ops = find_ife_oplist(metaid);
 274         int ret = 0;
 275
 276         if (!ops) {
 277                 ret = -ENOENT;
 278 #ifdef CONFIG_MODULES
 279                 rtnl_unlock();
 280                 request_module("ifemeta%u", metaid);
 281                 rtnl_lock();
 282                 ops = find_ife_oplist(metaid);
 283 #endif
 284         }
 285
 286         if (ops) {
 287                 ret = 0;
 288                 if (len)
 289                         ret = ife_validate_metatype(ops, val, len);
 290
 291                 module_put(ops->owner);
 292         }
 293
 294         return ret;
 295 }
 296
 297 /* called when adding new meta information
 298 */
 299 static int __add_metainfo(const struct tcf_meta_ops *ops,
 300                           struct tcf_ife_info *ife, u32 metaid, void *metaval,
 301                           int len, bool atomic, bool exists)
 302 {
 303         struct tcf_meta_info *mi = NULL;
 304         int ret = 0;
 305
 306         mi = kzalloc(sizeof(*mi), atomic ? GFP_ATOMIC : GFP_KERNEL);
 307         if (!mi)
 308                 return -ENOMEM;
 309
 310         mi->metaid = metaid;
 311         mi->ops = ops;
 312         if (len > 0) {
 313                 ret = ops->alloc(mi, metaval, atomic ? GFP_ATOMIC : GFP_KERNEL);
 314                 if (ret != 0) {
 315                         kfree(mi);
 316                         return ret;
 317                 }
 318         }
 319
 320         if (exists)
 321                 spin_lock_bh(&ife->tcf_lock);
 322         list_add_tail(&mi->metalist, &ife->metalist);
 323         if (exists)
 324                 spin_unlock_bh(&ife->tcf_lock);
 325
 326         return ret;
 327 }
 328
 329 static int add_metainfo_and_get_ops(const struct tcf_meta_ops *ops,
 330                                     struct tcf_ife_info *ife, u32 metaid,
 331                                     bool exists)
 332 {
 333         int ret;
 334
 335         if (!try_module_get(ops->owner))
 336                 return -ENOENT;
 337         ret = __add_metainfo(ops, ife, metaid, NULL, 0, true, exists);
 338         if (ret)
 339                 module_put(ops->owner);
 340         return ret;
 341 }
 342
 343 static int add_metainfo(struct tcf_ife_info *ife, u32 metaid, void *metaval,
 344                         int len, bool exists)
 345 {
 346         const struct tcf_meta_ops *ops = find_ife_oplist(metaid);
 347         int ret;
 348
 349         if (!ops)
 350                 return -ENOENT;
 351         ret = __add_metainfo(ops, ife, metaid, metaval, len, false, exists);
 352         if (ret)
 353                 /*put back what find_ife_oplist took */
 354                 module_put(ops->owner);
 355         return ret;
 356 }
 357
 358 static int use_all_metadata(struct tcf_ife_info *ife, bool exists)
 359 {
 360         struct tcf_meta_ops *o;
 361         int rc = 0;
 362         int installed = 0;
 363
 364         read_lock(&ife_mod_lock);
 365         list_for_each_entry(o, &ifeoplist, list) {
 366                 rc = add_metainfo_and_get_ops(o, ife, o->metaid, exists);
 367                 if (rc == 0)
 368                         installed += 1;
 369         }
 370         read_unlock(&ife_mod_lock);
 371
 372         if (installed)
 373                 return 0;
 374         else
 375                 return -EINVAL;
 376 }
 377
 378 static int dump_metalist(struct sk_buff *skb, struct tcf_ife_info *ife)
 379 {
 380         struct tcf_meta_info *e;
 381         struct nlattr *nest;
 382         unsigned char *b = skb_tail_pointer(skb);
 383         int total_encoded = 0;
 384
 385         /*can only happen on decode */
 386         if (list_empty(&ife->metalist))
 387                 return 0;
 388
 389         nest = nla_nest_start(skb, TCA_IFE_METALST);
 390         if (!nest)
 391                 goto out_nlmsg_trim;
 392
 393         list_for_each_entry(e, &ife->metalist, metalist) {
 394                 if (!e->ops->get(skb, e))
 395                         total_encoded += 1;
 396         }
 397
 398         if (!total_encoded)
 399                 goto out_nlmsg_trim;
 400
 401         nla_nest_end(skb, nest);
 402
 403         return 0;
 404
 405 out_nlmsg_trim:
 406         nlmsg_trim(skb, b);
 407         return -1;
 408 }
 409
 410 /* under ife->tcf_lock */
 411 static void _tcf_ife_cleanup(struct tc_action *a, int bind)
 412 {
 413         struct tcf_ife_info *ife = to_ife(a);
 414         struct tcf_meta_info *e, *n;
 415
 416         list_for_each_entry_safe(e, n, &ife->metalist, metalist) {
 417                 list_del(&e->metalist);
 418                 if (e->metaval) {
 419                         if (e->ops->release)
 420                                 e->ops->release(e);
 421                         else
 422                                 kfree(e->metaval);
 423                 }
 424                 module_put(e->ops->owner);
 425                 kfree(e);
 426         }
 427 }
 428
 429 static void tcf_ife_cleanup(struct tc_action *a, int bind)
 430 {
 431         struct tcf_ife_info *ife = to_ife(a);
 432
 433         spin_lock_bh(&ife->tcf_lock);
 434         _tcf_ife_cleanup(a, bind);
 435         spin_unlock_bh(&ife->tcf_lock);
 436 }
 437
 438 static int populate_metalist(struct tcf_ife_info *ife, struct nlattr **tb,
 439                              bool exists)
 440 {
 441         int len = 0;
 442         int rc = 0;
 443         int i = 0;
 444         void *val;
 445
 446         for (i = 1; i < max_metacnt; i++) {
 447                 if (tb[i]) {
 448                         val = nla_data(tb[i]);
 449                         len = nla_len(tb[i]);
 450
 451                         rc = load_metaops_and_vet(i, val, len);
 452                         if (rc != 0)
 453                                 return rc;
 454
 455                         rc = add_metainfo(ife, i, val, len, exists);
 456                         if (rc)
 457                                 return rc;
 458                 }
 459         }
 460
 461         return rc;
 462 }
 463
 464 static int tcf_ife_init(struct net *net, struct nlattr *nla,
 465                         struct nlattr *est, struct tc_action **a,
 466                         int ovr, int bind)
 467 {
 468         struct tc_action_net *tn = net_generic(net, ife_net_id);
 469         struct nlattr *tb[TCA_IFE_MAX + 1];
 470         struct nlattr *tb2[IFE_META_MAX + 1];
 471         struct tcf_ife_info *ife;
 472         struct tc_ife *parm;
 473         u16 ife_type = 0;
 474         u8 *daddr = NULL;
 475         u8 *saddr = NULL;
 476         bool exists = false;
 477         int ret = 0;
 478         int err;
 479
 480         if (!nla)
 481                 return -EINVAL;
 482
 483         err = nla_parse_nested(tb, TCA_IFE_MAX, nla, ife_policy);
 484         if (err < 0)
 485                 return err;
 486
 487         if (!tb[TCA_IFE_PARMS])
 488                 return -EINVAL;
 489
 490         parm = nla_data(tb[TCA_IFE_PARMS]);
 491
 492         exists = tcf_hash_check(tn, parm->index, a, bind);
 493         if (exists && bind)
 494                 return 0;
 495
 496         if (parm->flags & IFE_ENCODE) {
 497                 /* Until we get issued the ethertype, we cant have
 498                  * a default..
 499                 **/
 500                 if (!tb[TCA_IFE_TYPE]) {
 501                         if (exists)
 502                                 tcf_hash_release(*a, bind);
 503                         pr_info("You MUST pass etherype for encoding\n");
 504                         return -EINVAL;
 505                 }
 506         }
 507
 508         if (!exists) {
 509                 ret = tcf_hash_create(tn, parm->index, est, a, &act_ife_ops,
 510                                       bind, false);
 511                 if (ret)
 512                         return ret;
 513                 ret = ACT_P_CREATED;
 514         } else {
 515                 tcf_hash_release(*a, bind);
 516                 if (!ovr)
 517                         return -EEXIST;
 518         }
 519
 520         ife = to_ife(*a);
 521         ife->flags = parm->flags;
 522
 523         if (parm->flags & IFE_ENCODE) {
 524                 ife_type = nla_get_u16(tb[TCA_IFE_TYPE]);
 525                 if (tb[TCA_IFE_DMAC])
 526                         daddr = nla_data(tb[TCA_IFE_DMAC]);
 527                 if (tb[TCA_IFE_SMAC])
 528                         saddr = nla_data(tb[TCA_IFE_SMAC]);
 529         }
 530
 531         if (exists)
 532                 spin_lock_bh(&ife->tcf_lock);
 533         ife->tcf_action = parm->action;
 534         if (exists)
 535                 spin_unlock_bh(&ife->tcf_lock);
 536
 537         if (parm->flags & IFE_ENCODE) {
 538                 if (daddr)
 539                         ether_addr_copy(ife->eth_dst, daddr);
 540                 else
 541                         eth_zero_addr(ife->eth_dst);
 542
 543                 if (saddr)
 544                         ether_addr_copy(ife->eth_src, saddr);
 545                 else
 546                         eth_zero_addr(ife->eth_src);
 547
 548                 ife->eth_type = ife_type;
 549         }
 550
 551         if (ret == ACT_P_CREATED)
 552                 INIT_LIST_HEAD(&ife->metalist);
 553
 554         if (tb[TCA_IFE_METALST]) {
 555                 err = nla_parse_nested(tb2, IFE_META_MAX, tb[TCA_IFE_METALST],
 556                                        NULL);
 557                 if (err) {
 558 metadata_parse_err:
 559                         if (exists)
 560                                 tcf_hash_release(*a, bind);
 561                         if (ret == ACT_P_CREATED)
 562                                 _tcf_ife_cleanup(*a, bind);
 563                         return err;
 564                 }
 565
 566                 err = populate_metalist(ife, tb2, exists);
 567                 if (err)
 568                         goto metadata_parse_err;
 569
 570         } else {
 571                 /* if no passed metadata allow list or passed allow-all
 572                  * then here we process by adding as many supported metadatum
 573                  * as we can. You better have at least one else we are
 574                  * going to bail out
 575                  */
 576                 err = use_all_metadata(ife, exists);
 577                 if (err) {
 578                         if (ret == ACT_P_CREATED)
 579                                 _tcf_ife_cleanup(*a, bind);
 580                         return err;
 581                 }
 582         }
 583
 584         if (ret == ACT_P_CREATED)
 585                 tcf_hash_insert(tn, *a);
 586
 587         return ret;
 588 }
 589
 590 static int tcf_ife_dump(struct sk_buff *skb, struct tc_action *a, int bind,
 591                         int ref)
 592 {
 593         unsigned char *b = skb_tail_pointer(skb);
 594         struct tcf_ife_info *ife = to_ife(a);
 595         struct tc_ife opt = {
 596                 .index = ife->tcf_index,
 597                 .refcnt = ife->tcf_refcnt - ref,
 598                 .bindcnt = ife->tcf_bindcnt - bind,
 599                 .action = ife->tcf_action,
 600                 .flags = ife->flags,
 601         };
 602         struct tcf_t t;
 603
 604         if (nla_put(skb, TCA_IFE_PARMS, sizeof(opt), &opt))
 605                 goto nla_put_failure;
 606
 607         tcf_tm_dump(&t, &ife->tcf_tm);
 608         if (nla_put_64bit(skb, TCA_IFE_TM, sizeof(t), &t, TCA_IFE_PAD))
 609                 goto nla_put_failure;
 610
 611         if (!is_zero_ether_addr(ife->eth_dst)) {
 612                 if (nla_put(skb, TCA_IFE_DMAC, ETH_ALEN, ife->eth_dst))
 613                         goto nla_put_failure;
 614         }
 615
 616         if (!is_zero_ether_addr(ife->eth_src)) {
 617                 if (nla_put(skb, TCA_IFE_SMAC, ETH_ALEN, ife->eth_src))
 618                         goto nla_put_failure;
 619         }
 620
 621         if (nla_put(skb, TCA_IFE_TYPE, 2, &ife->eth_type))
 622                 goto nla_put_failure;
 623
 624         if (dump_metalist(skb, ife)) {
 625                 /*ignore failure to dump metalist */
 626                 pr_info("Failed to dump metalist\n");
 627         }
 628
 629         return skb->len;
 630
 631 nla_put_failure:
 632         nlmsg_trim(skb, b);
 633         return -1;
 634 }
 635
 636 int find_decode_metaid(struct sk_buff *skb, struct tcf_ife_info *ife,
 637                        u16 metaid, u16 mlen, void *mdata)
 638 {
 639         struct tcf_meta_info *e;
 640
 641         /* XXX: use hash to speed up */
 642         list_for_each_entry(e, &ife->metalist, metalist) {
 643                 if (metaid == e->metaid) {
 644                         if (e->ops) {
 645                                 /* We check for decode presence already */
 646                                 return e->ops->decode(skb, mdata, mlen);
 647                         }
 648                 }
 649         }
 650
 651         return -ENOENT;
 652 }
 653
 654 struct ifeheadr {
 655         __be16 metalen;
 656         u8 tlv_data[];
 657 };
 658
 659 struct meta_tlvhdr {
 660         __be16 type;
 661         __be16 len;
 662 };
 663
 664 static int tcf_ife_decode(struct sk_buff *skb, const struct tc_action *a,
 665                           struct tcf_result *res)
 666 {
 667         struct tcf_ife_info *ife = to_ife(a);
 668         int action = ife->tcf_action;
 669         struct ifeheadr *ifehdr = (struct ifeheadr *)skb->data;
 670         int ifehdrln = (int)ifehdr->metalen;
 671         struct meta_tlvhdr *tlv = (struct meta_tlvhdr *)(ifehdr->tlv_data);
 672
 673         spin_lock(&ife->tcf_lock);
 674         bstats_update(&ife->tcf_bstats, skb);
 675         tcf_lastuse_update(&ife->tcf_tm);
 676         spin_unlock(&ife->tcf_lock);
 677
 678         ifehdrln = ntohs(ifehdrln);
 679         if (unlikely(!pskb_may_pull(skb, ifehdrln))) {
 680                 spin_lock(&ife->tcf_lock);
 681                 ife->tcf_qstats.drops++;
 682                 spin_unlock(&ife->tcf_lock);
 683                 return TC_ACT_SHOT;
 684         }
 685
 686         skb_set_mac_header(skb, ifehdrln);
 687         __skb_pull(skb, ifehdrln);
 688         skb->protocol = eth_type_trans(skb, skb->dev);
 689         ifehdrln -= IFE_METAHDRLEN;
 690
 691         while (ifehdrln > 0) {
 692                 u8 *tlvdata = (u8 *)tlv;
 693                 u16 mtype = tlv->type;
 694                 u16 mlen = tlv->len;
 695                 u16 alen;
 696
 697                 mtype = ntohs(mtype);
 698                 mlen = ntohs(mlen);
 699                 alen = NLA_ALIGN(mlen);
 700
 701                 if (find_decode_metaid(skb, ife, mtype, (mlen - NLA_HDRLEN),
 702                                        (void *)(tlvdata + NLA_HDRLEN))) {
 703                         /* abuse overlimits to count when we receive metadata
 704                          * but dont have an ops for it
 705                          */
 706                         pr_info_ratelimited("Unknown metaid %d alnlen %d\n",
 707                                             mtype, mlen);
 708                         ife->tcf_qstats.overlimits++;
 709                 }
 710
 711                 tlvdata += alen;
 712                 ifehdrln -= alen;
 713                 tlv = (struct meta_tlvhdr *)tlvdata;
 714         }
 715
 716         skb_reset_network_header(skb);
 717         return action;
 718 }
 719
 720 /*XXX: check if we can do this at install time instead of current
 721  * send data path
 722 **/
 723 static int ife_get_sz(struct sk_buff *skb, struct tcf_ife_info *ife)
 724 {
 725         struct tcf_meta_info *e, *n;
 726         int tot_run_sz = 0, run_sz = 0;
 727
 728         list_for_each_entry_safe(e, n, &ife->metalist, metalist) {
 729                 if (e->ops->check_presence) {
 730                         run_sz = e->ops->check_presence(skb, e);
 731                         tot_run_sz += run_sz;
 732                 }
 733         }
 734
 735         return tot_run_sz;
 736 }
 737
 738 static int tcf_ife_encode(struct sk_buff *skb, const struct tc_action *a,
 739                           struct tcf_result *res)
 740 {
 741         struct tcf_ife_info *ife = to_ife(a);
 742         int action = ife->tcf_action;
 743         struct ethhdr *oethh;   /* outer ether header */
 744         struct ethhdr *iethh;   /* inner eth header */
 745         struct tcf_meta_info *e;
 746         /*
 747            OUTERHDR:TOTMETALEN:{TLVHDR:Metadatum:TLVHDR..}:ORIGDATA
 748            where ORIGDATA = original ethernet header ...
 749          */
 750         u16 metalen = ife_get_sz(skb, ife);
 751         int hdrm = metalen + skb->dev->hard_header_len + IFE_METAHDRLEN;
 752         unsigned int skboff = skb->dev->hard_header_len;
 753         u32 at = G_TC_AT(skb->tc_verd);
 754         int new_len = skb->len + hdrm;
 755         bool exceed_mtu = false;
 756         int err;
 757
 758         if (at & AT_EGRESS) {
 759                 if (new_len > skb->dev->mtu)
 760                         exceed_mtu = true;
 761         }
 762
 763         spin_lock(&ife->tcf_lock);
 764         bstats_update(&ife->tcf_bstats, skb);
 765         tcf_lastuse_update(&ife->tcf_tm);
 766
 767         if (!metalen) {         /* no metadata to send */
 768                 /* abuse overlimits to count when we allow packet
 769                  * with no metadata
 770                  */
 771                 ife->tcf_qstats.overlimits++;
 772                 spin_unlock(&ife->tcf_lock);
 773                 return action;
 774         }
 775         /* could be stupid policy setup or mtu config
 776          * so lets be conservative.. */
 777         if ((action == TC_ACT_SHOT) || exceed_mtu) {
 778                 ife->tcf_qstats.drops++;
 779                 spin_unlock(&ife->tcf_lock);
 780                 return TC_ACT_SHOT;
 781         }
 782
 783         err = skb_cow_head(skb, hdrm);
 784         if (unlikely(err)) {
 785                 ife->tcf_qstats.drops++;
 786                 spin_unlock(&ife->tcf_lock);
 787                 return TC_ACT_SHOT;
 788         }
 789
 790         if (!(at & AT_EGRESS))
 791                 skb_push(skb, skb->dev->hard_header_len);
 792
 793         iethh = (struct ethhdr *)skb->data;
 794         __skb_push(skb, hdrm);
 795         memcpy(skb->data, iethh, skb->mac_len);
 796         skb_reset_mac_header(skb);
 797         oethh = eth_hdr(skb);
 798
 799         /*total metadata length */
 800         metalen += IFE_METAHDRLEN;
 801         metalen = htons(metalen);
 802         memcpy((skb->data + skboff), &metalen, IFE_METAHDRLEN);
 803         skboff += IFE_METAHDRLEN;
 804
 805         /* XXX: we dont have a clever way of telling encode to
 806          * not repeat some of the computations that are done by
 807          * ops->presence_check...
 808          */
 809         list_for_each_entry(e, &ife->metalist, metalist) {
 810                 if (e->ops->encode) {
 811                         err = e->ops->encode(skb, (void *)(skb->data + skboff),
 812                                              e);
 813                 }
 814                 if (err < 0) {
 815                         /* too corrupt to keep around if overwritten */
 816                         ife->tcf_qstats.drops++;
 817                         spin_unlock(&ife->tcf_lock);
 818                         return TC_ACT_SHOT;
 819                 }
 820                 skboff += err;
 821         }
 822
 823         if (!is_zero_ether_addr(ife->eth_src))
 824                 ether_addr_copy(oethh->h_source, ife->eth_src);
 825         else
 826                 ether_addr_copy(oethh->h_source, iethh->h_source);
 827         if (!is_zero_ether_addr(ife->eth_dst))
 828                 ether_addr_copy(oethh->h_dest, ife->eth_dst);
 829         else
 830                 ether_addr_copy(oethh->h_dest, iethh->h_dest);
 831         oethh->h_proto = htons(ife->eth_type);
 832
 833         if (!(at & AT_EGRESS))
 834                 skb_pull(skb, skb->dev->hard_header_len);
 835
 836         spin_unlock(&ife->tcf_lock);
 837
 838         return action;
 839 }
 840
 841 static int tcf_ife_act(struct sk_buff *skb, const struct tc_action *a,
 842                        struct tcf_result *res)
 843 {
 844         struct tcf_ife_info *ife = to_ife(a);
 845
 846         if (ife->flags & IFE_ENCODE)
 847                 return tcf_ife_encode(skb, a, res);
 848
 849         if (!(ife->flags & IFE_ENCODE))
 850                 return tcf_ife_decode(skb, a, res);
 851
 852         pr_info_ratelimited("unknown failure(policy neither de/encode\n");
 853         spin_lock(&ife->tcf_lock);
 854         bstats_update(&ife->tcf_bstats, skb);
 855         tcf_lastuse_update(&ife->tcf_tm);
 856         ife->tcf_qstats.drops++;
 857         spin_unlock(&ife->tcf_lock);
 858
 859         return TC_ACT_SHOT;
 860 }
 861
 862 static int tcf_ife_walker(struct net *net, struct sk_buff *skb,
 863                           struct netlink_callback *cb, int type,
 864                           const struct tc_action_ops *ops)
 865 {
 866         struct tc_action_net *tn = net_generic(net, ife_net_id);
 867
 868         return tcf_generic_walker(tn, skb, cb, type, ops);
 869 }
 870
 871 static int tcf_ife_search(struct net *net, struct tc_action **a, u32 index)
 872 {
 873         struct tc_action_net *tn = net_generic(net, ife_net_id);
 874
 875         return tcf_hash_search(tn, a, index);
 876 }
 877
 878 static struct tc_action_ops act_ife_ops = {
 879         .kind = "ife",
 880         .type = TCA_ACT_IFE,
 881         .owner = THIS_MODULE,
 882         .act = tcf_ife_act,
 883         .dump = tcf_ife_dump,
 884         .cleanup = tcf_ife_cleanup,
 885         .init = tcf_ife_init,
 886         .walk = tcf_ife_walker,
 887         .lookup = tcf_ife_search,
 888         .size = sizeof(struct tcf_ife_info),
 889 };
 890
 891 static __net_init int ife_init_net(struct net *net)
 892 {
 893         struct tc_action_net *tn = net_generic(net, ife_net_id);
 894
 895         return tc_action_net_init(tn, &act_ife_ops, IFE_TAB_MASK);
 896 }
 897
 898 static void __net_exit ife_exit_net(struct net *net)
 899 {
 900         struct tc_action_net *tn = net_generic(net, ife_net_id);
 901
 902         tc_action_net_exit(tn);
 903 }
 904
 905 static struct pernet_operations ife_net_ops = {
 906         .init = ife_init_net,
 907         .exit = ife_exit_net,
 908         .id   = &ife_net_id,
 909         .size = sizeof(struct tc_action_net),
 910 };
 911
 912 static int __init ife_init_module(void)
 913 {
 914         return tcf_register_action(&act_ife_ops, &ife_net_ops);
 915 }
 916
 917 static void __exit ife_cleanup_module(void)
 918 {
 919         tcf_unregister_action(&act_ife_ops, &ife_net_ops);
 920 }
 921
 922 module_init(ife_init_module);
 923 module_exit(ife_cleanup_module);
 924
 925 MODULE_AUTHOR("Jamal Hadi Salim(2015)");
 926 MODULE_DESCRIPTION("Inter-FE LFB action");
 927 MODULE_LICENSE("GPL");