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 unsigned 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  * under ife->tcf_lock for existing action
 271 */
 272 static int load_metaops_and_vet(struct tcf_ife_info *ife, u32 metaid,
 273                                 void *val, int len, bool exists)
 274 {
 275         struct tcf_meta_ops *ops = find_ife_oplist(metaid);
 276         int ret = 0;
 277
 278         if (!ops) {
 279                 ret = -ENOENT;
 280 #ifdef CONFIG_MODULES
 281                 if (exists)
 282                         spin_unlock_bh(&ife->tcf_lock);
 283                 rtnl_unlock();
 284                 request_module("ifemeta%u", metaid);
 285                 rtnl_lock();
 286                 if (exists)
 287                         spin_lock_bh(&ife->tcf_lock);
 288                 ops = find_ife_oplist(metaid);
 289 #endif
 290         }
 291
 292         if (ops) {
 293                 ret = 0;
 294                 if (len)
 295                         ret = ife_validate_metatype(ops, val, len);
 296
 297                 module_put(ops->owner);
 298         }
 299
 300         return ret;
 301 }
 302
 303 /* called when adding new meta information
 304  * under ife->tcf_lock for existing action
 305 */
 306 static int add_metainfo(struct tcf_ife_info *ife, u32 metaid, void *metaval,
 307                         int len, bool atomic)
 308 {
 309         struct tcf_meta_info *mi = NULL;
 310         struct tcf_meta_ops *ops = find_ife_oplist(metaid);
 311         int ret = 0;
 312
 313         if (!ops)
 314                 return -ENOENT;
 315
 316         mi = kzalloc(sizeof(*mi), atomic ? GFP_ATOMIC : GFP_KERNEL);
 317         if (!mi) {
 318                 /*put back what find_ife_oplist took */
 319                 module_put(ops->owner);
 320                 return -ENOMEM;
 321         }
 322
 323         mi->metaid = metaid;
 324         mi->ops = ops;
 325         if (len > 0) {
 326                 ret = ops->alloc(mi, metaval, atomic ? GFP_ATOMIC : GFP_KERNEL);
 327                 if (ret != 0) {
 328                         kfree(mi);
 329                         module_put(ops->owner);
 330                         return ret;
 331                 }
 332         }
 333
 334         list_add_tail(&mi->metalist, &ife->metalist);
 335
 336         return ret;
 337 }
 338
 339 static int use_all_metadata(struct tcf_ife_info *ife)
 340 {
 341         struct tcf_meta_ops *o;
 342         int rc = 0;
 343         int installed = 0;
 344
 345         read_lock(&ife_mod_lock);
 346         list_for_each_entry(o, &ifeoplist, list) {
 347                 rc = add_metainfo(ife, o->metaid, NULL, 0, true);
 348                 if (rc == 0)
 349                         installed += 1;
 350         }
 351         read_unlock(&ife_mod_lock);
 352
 353         if (installed)
 354                 return 0;
 355         else
 356                 return -EINVAL;
 357 }
 358
 359 static int dump_metalist(struct sk_buff *skb, struct tcf_ife_info *ife)
 360 {
 361         struct tcf_meta_info *e;
 362         struct nlattr *nest;
 363         unsigned char *b = skb_tail_pointer(skb);
 364         int total_encoded = 0;
 365
 366         /*can only happen on decode */
 367         if (list_empty(&ife->metalist))
 368                 return 0;
 369
 370         nest = nla_nest_start(skb, TCA_IFE_METALST);
 371         if (!nest)
 372                 goto out_nlmsg_trim;
 373
 374         list_for_each_entry(e, &ife->metalist, metalist) {
 375                 if (!e->ops->get(skb, e))
 376                         total_encoded += 1;
 377         }
 378
 379         if (!total_encoded)
 380                 goto out_nlmsg_trim;
 381
 382         nla_nest_end(skb, nest);
 383
 384         return 0;
 385
 386 out_nlmsg_trim:
 387         nlmsg_trim(skb, b);
 388         return -1;
 389 }
 390
 391 /* under ife->tcf_lock */
 392 static void _tcf_ife_cleanup(struct tc_action *a, int bind)
 393 {
 394         struct tcf_ife_info *ife = to_ife(a);
 395         struct tcf_meta_info *e, *n;
 396
 397         list_for_each_entry_safe(e, n, &ife->metalist, metalist) {
 398                 module_put(e->ops->owner);
 399                 list_del(&e->metalist);
 400                 if (e->metaval) {
 401                         if (e->ops->release)
 402                                 e->ops->release(e);
 403                         else
 404                                 kfree(e->metaval);
 405                 }
 406                 kfree(e);
 407         }
 408 }
 409
 410 static void tcf_ife_cleanup(struct tc_action *a, int bind)
 411 {
 412         struct tcf_ife_info *ife = to_ife(a);
 413
 414         spin_lock_bh(&ife->tcf_lock);
 415         _tcf_ife_cleanup(a, bind);
 416         spin_unlock_bh(&ife->tcf_lock);
 417 }
 418
 419 /* under ife->tcf_lock for existing action */
 420 static int populate_metalist(struct tcf_ife_info *ife, struct nlattr **tb,
 421                              bool exists)
 422 {
 423         int len = 0;
 424         int rc = 0;
 425         int i = 0;
 426         void *val;
 427
 428         for (i = 1; i < max_metacnt; i++) {
 429                 if (tb[i]) {
 430                         val = nla_data(tb[i]);
 431                         len = nla_len(tb[i]);
 432
 433                         rc = load_metaops_and_vet(ife, i, val, len, exists);
 434                         if (rc != 0)
 435                                 return rc;
 436
 437                         rc = add_metainfo(ife, i, val, len, exists);
 438                         if (rc)
 439                                 return rc;
 440                 }
 441         }
 442
 443         return rc;
 444 }
 445
 446 static int tcf_ife_init(struct net *net, struct nlattr *nla,
 447                         struct nlattr *est, struct tc_action **a,
 448                         int ovr, int bind)
 449 {
 450         struct tc_action_net *tn = net_generic(net, ife_net_id);
 451         struct nlattr *tb[TCA_IFE_MAX + 1];
 452         struct nlattr *tb2[IFE_META_MAX + 1];
 453         struct tcf_ife_info *ife;
 454         struct tc_ife *parm;
 455         u16 ife_type = 0;
 456         u8 *daddr = NULL;
 457         u8 *saddr = NULL;
 458         bool exists = false;
 459         int ret = 0;
 460         int err;
 461
 462         err = nla_parse_nested(tb, TCA_IFE_MAX, nla, ife_policy);
 463         if (err < 0)
 464                 return err;
 465
 466         if (!tb[TCA_IFE_PARMS])
 467                 return -EINVAL;
 468
 469         parm = nla_data(tb[TCA_IFE_PARMS]);
 470
 471         exists = tcf_hash_check(tn, parm->index, a, bind);
 472         if (exists && bind)
 473                 return 0;
 474
 475         if (parm->flags & IFE_ENCODE) {
 476                 /* Until we get issued the ethertype, we cant have
 477                  * a default..
 478                 **/
 479                 if (!tb[TCA_IFE_TYPE]) {
 480                         if (exists)
 481                                 tcf_hash_release(*a, bind);
 482                         pr_info("You MUST pass etherype for encoding\n");
 483                         return -EINVAL;
 484                 }
 485         }
 486
 487         if (!exists) {
 488                 ret = tcf_hash_create(tn, parm->index, est, a, &act_ife_ops,
 489                                       bind, false);
 490                 if (ret)
 491                         return ret;
 492                 ret = ACT_P_CREATED;
 493         } else {
 494                 tcf_hash_release(*a, bind);
 495                 if (!ovr)
 496                         return -EEXIST;
 497         }
 498
 499         ife = to_ife(*a);
 500         ife->flags = parm->flags;
 501
 502         if (parm->flags & IFE_ENCODE) {
 503                 ife_type = nla_get_u16(tb[TCA_IFE_TYPE]);
 504                 if (tb[TCA_IFE_DMAC])
 505                         daddr = nla_data(tb[TCA_IFE_DMAC]);
 506                 if (tb[TCA_IFE_SMAC])
 507                         saddr = nla_data(tb[TCA_IFE_SMAC]);
 508         }
 509
 510         if (exists)
 511                 spin_lock_bh(&ife->tcf_lock);
 512         ife->tcf_action = parm->action;
 513
 514         if (parm->flags & IFE_ENCODE) {
 515                 if (daddr)
 516                         ether_addr_copy(ife->eth_dst, daddr);
 517                 else
 518                         eth_zero_addr(ife->eth_dst);
 519
 520                 if (saddr)
 521                         ether_addr_copy(ife->eth_src, saddr);
 522                 else
 523                         eth_zero_addr(ife->eth_src);
 524
 525                 ife->eth_type = ife_type;
 526         }
 527
 528         if (ret == ACT_P_CREATED)
 529                 INIT_LIST_HEAD(&ife->metalist);
 530
 531         if (tb[TCA_IFE_METALST]) {
 532                 err = nla_parse_nested(tb2, IFE_META_MAX, tb[TCA_IFE_METALST],
 533                                        NULL);
 534                 if (err) {
 535 metadata_parse_err:
 536                         if (exists)
 537                                 tcf_hash_release(*a, bind);
 538                         if (ret == ACT_P_CREATED)
 539                                 _tcf_ife_cleanup(*a, bind);
 540
 541                         if (exists)
 542                                 spin_unlock_bh(&ife->tcf_lock);
 543                         return err;
 544                 }
 545
 546                 err = populate_metalist(ife, tb2, exists);
 547                 if (err)
 548                         goto metadata_parse_err;
 549
 550         } else {
 551                 /* if no passed metadata allow list or passed allow-all
 552                  * then here we process by adding as many supported metadatum
 553                  * as we can. You better have at least one else we are
 554                  * going to bail out
 555                  */
 556                 err = use_all_metadata(ife);
 557                 if (err) {
 558                         if (ret == ACT_P_CREATED)
 559                                 _tcf_ife_cleanup(*a, bind);
 560
 561                         if (exists)
 562                                 spin_unlock_bh(&ife->tcf_lock);
 563                         return err;
 564                 }
 565         }
 566
 567         if (exists)
 568                 spin_unlock_bh(&ife->tcf_lock);
 569
 570         if (ret == ACT_P_CREATED)
 571                 tcf_hash_insert(tn, *a);
 572
 573         return ret;
 574 }
 575
 576 static int tcf_ife_dump(struct sk_buff *skb, struct tc_action *a, int bind,
 577                         int ref)
 578 {
 579         unsigned char *b = skb_tail_pointer(skb);
 580         struct tcf_ife_info *ife = to_ife(a);
 581         struct tc_ife opt = {
 582                 .index = ife->tcf_index,
 583                 .refcnt = ife->tcf_refcnt - ref,
 584                 .bindcnt = ife->tcf_bindcnt - bind,
 585                 .action = ife->tcf_action,
 586                 .flags = ife->flags,
 587         };
 588         struct tcf_t t;
 589
 590         if (nla_put(skb, TCA_IFE_PARMS, sizeof(opt), &opt))
 591                 goto nla_put_failure;
 592
 593         tcf_tm_dump(&t, &ife->tcf_tm);
 594         if (nla_put_64bit(skb, TCA_IFE_TM, sizeof(t), &t, TCA_IFE_PAD))
 595                 goto nla_put_failure;
 596
 597         if (!is_zero_ether_addr(ife->eth_dst)) {
 598                 if (nla_put(skb, TCA_IFE_DMAC, ETH_ALEN, ife->eth_dst))
 599                         goto nla_put_failure;
 600         }
 601
 602         if (!is_zero_ether_addr(ife->eth_src)) {
 603                 if (nla_put(skb, TCA_IFE_SMAC, ETH_ALEN, ife->eth_src))
 604                         goto nla_put_failure;
 605         }
 606
 607         if (nla_put(skb, TCA_IFE_TYPE, 2, &ife->eth_type))
 608                 goto nla_put_failure;
 609
 610         if (dump_metalist(skb, ife)) {
 611                 /*ignore failure to dump metalist */
 612                 pr_info("Failed to dump metalist\n");
 613         }
 614
 615         return skb->len;
 616
 617 nla_put_failure:
 618         nlmsg_trim(skb, b);
 619         return -1;
 620 }
 621
 622 int find_decode_metaid(struct sk_buff *skb, struct tcf_ife_info *ife,
 623                        u16 metaid, u16 mlen, void *mdata)
 624 {
 625         struct tcf_meta_info *e;
 626
 627         /* XXX: use hash to speed up */
 628         list_for_each_entry(e, &ife->metalist, metalist) {
 629                 if (metaid == e->metaid) {
 630                         if (e->ops) {
 631                                 /* We check for decode presence already */
 632                                 return e->ops->decode(skb, mdata, mlen);
 633                         }
 634                 }
 635         }
 636
 637         return 0;
 638 }
 639
 640 struct ifeheadr {
 641         __be16 metalen;
 642         u8 tlv_data[];
 643 };
 644
 645 struct meta_tlvhdr {
 646         __be16 type;
 647         __be16 len;
 648 };
 649
 650 static int tcf_ife_decode(struct sk_buff *skb, const struct tc_action *a,
 651                           struct tcf_result *res)
 652 {
 653         struct tcf_ife_info *ife = to_ife(a);
 654         int action = ife->tcf_action;
 655         struct ifeheadr *ifehdr = (struct ifeheadr *)skb->data;
 656         int ifehdrln = (int)ifehdr->metalen;
 657         struct meta_tlvhdr *tlv = (struct meta_tlvhdr *)(ifehdr->tlv_data);
 658
 659         spin_lock(&ife->tcf_lock);
 660         bstats_update(&ife->tcf_bstats, skb);
 661         tcf_lastuse_update(&ife->tcf_tm);
 662         spin_unlock(&ife->tcf_lock);
 663
 664         ifehdrln = ntohs(ifehdrln);
 665         if (unlikely(!pskb_may_pull(skb, ifehdrln))) {
 666                 spin_lock(&ife->tcf_lock);
 667                 ife->tcf_qstats.drops++;
 668                 spin_unlock(&ife->tcf_lock);
 669                 return TC_ACT_SHOT;
 670         }
 671
 672         skb_set_mac_header(skb, ifehdrln);
 673         __skb_pull(skb, ifehdrln);
 674         skb->protocol = eth_type_trans(skb, skb->dev);
 675         ifehdrln -= IFE_METAHDRLEN;
 676
 677         while (ifehdrln > 0) {
 678                 u8 *tlvdata = (u8 *)tlv;
 679                 u16 mtype = tlv->type;
 680                 u16 mlen = tlv->len;
 681                 u16 alen;
 682
 683                 mtype = ntohs(mtype);
 684                 mlen = ntohs(mlen);
 685                 alen = NLA_ALIGN(mlen);
 686
 687                 if (find_decode_metaid(skb, ife, mtype, (mlen - NLA_HDRLEN),
 688                                        (void *)(tlvdata + NLA_HDRLEN))) {
 689                         /* abuse overlimits to count when we receive metadata
 690                          * but dont have an ops for it
 691                          */
 692                         pr_info_ratelimited("Unknown metaid %d alnlen %d\n",
 693                                             mtype, mlen);
 694                         ife->tcf_qstats.overlimits++;
 695                 }
 696
 697                 tlvdata += alen;
 698                 ifehdrln -= alen;
 699                 tlv = (struct meta_tlvhdr *)tlvdata;
 700         }
 701
 702         skb_reset_network_header(skb);
 703         return action;
 704 }
 705
 706 /*XXX: check if we can do this at install time instead of current
 707  * send data path
 708 **/
 709 static int ife_get_sz(struct sk_buff *skb, struct tcf_ife_info *ife)
 710 {
 711         struct tcf_meta_info *e, *n;
 712         int tot_run_sz = 0, run_sz = 0;
 713
 714         list_for_each_entry_safe(e, n, &ife->metalist, metalist) {
 715                 if (e->ops->check_presence) {
 716                         run_sz = e->ops->check_presence(skb, e);
 717                         tot_run_sz += run_sz;
 718                 }
 719         }
 720
 721         return tot_run_sz;
 722 }
 723
 724 static int tcf_ife_encode(struct sk_buff *skb, const struct tc_action *a,
 725                           struct tcf_result *res)
 726 {
 727         struct tcf_ife_info *ife = to_ife(a);
 728         int action = ife->tcf_action;
 729         struct ethhdr *oethh;   /* outer ether header */
 730         struct ethhdr *iethh;   /* inner eth header */
 731         struct tcf_meta_info *e;
 732         /*
 733            OUTERHDR:TOTMETALEN:{TLVHDR:Metadatum:TLVHDR..}:ORIGDATA
 734            where ORIGDATA = original ethernet header ...
 735          */
 736         u16 metalen = ife_get_sz(skb, ife);
 737         int hdrm = metalen + skb->dev->hard_header_len + IFE_METAHDRLEN;
 738         unsigned int skboff = skb->dev->hard_header_len;
 739         int new_len = skb->len + hdrm;
 740         bool exceed_mtu = false;
 741         int err;
 742
 743         if (!skb_at_tc_ingress(skb)) {
 744                 if (new_len > skb->dev->mtu)
 745                         exceed_mtu = true;
 746         }
 747
 748         spin_lock(&ife->tcf_lock);
 749         bstats_update(&ife->tcf_bstats, skb);
 750         tcf_lastuse_update(&ife->tcf_tm);
 751
 752         if (!metalen) {         /* no metadata to send */
 753                 /* abuse overlimits to count when we allow packet
 754                  * with no metadata
 755                  */
 756                 ife->tcf_qstats.overlimits++;
 757                 spin_unlock(&ife->tcf_lock);
 758                 return action;
 759         }
 760         /* could be stupid policy setup or mtu config
 761          * so lets be conservative.. */
 762         if ((action == TC_ACT_SHOT) || exceed_mtu) {
 763                 ife->tcf_qstats.drops++;
 764                 spin_unlock(&ife->tcf_lock);
 765                 return TC_ACT_SHOT;
 766         }
 767
 768         err = skb_cow_head(skb, hdrm);
 769         if (unlikely(err)) {
 770                 ife->tcf_qstats.drops++;
 771                 spin_unlock(&ife->tcf_lock);
 772                 return TC_ACT_SHOT;
 773         }
 774
 775         if (skb_at_tc_ingress(skb))
 776                 skb_push(skb, skb->dev->hard_header_len);
 777
 778         iethh = (struct ethhdr *)skb->data;
 779         __skb_push(skb, hdrm);
 780         memcpy(skb->data, iethh, skb->mac_len);
 781         skb_reset_mac_header(skb);
 782         oethh = eth_hdr(skb);
 783
 784         /*total metadata length */
 785         metalen += IFE_METAHDRLEN;
 786         metalen = htons(metalen);
 787         memcpy((skb->data + skboff), &metalen, IFE_METAHDRLEN);
 788         skboff += IFE_METAHDRLEN;
 789
 790         /* XXX: we dont have a clever way of telling encode to
 791          * not repeat some of the computations that are done by
 792          * ops->presence_check...
 793          */
 794         list_for_each_entry(e, &ife->metalist, metalist) {
 795                 if (e->ops->encode) {
 796                         err = e->ops->encode(skb, (void *)(skb->data + skboff),
 797                                              e);
 798                 }
 799                 if (err < 0) {
 800                         /* too corrupt to keep around if overwritten */
 801                         ife->tcf_qstats.drops++;
 802                         spin_unlock(&ife->tcf_lock);
 803                         return TC_ACT_SHOT;
 804                 }
 805                 skboff += err;
 806         }
 807
 808         if (!is_zero_ether_addr(ife->eth_src))
 809                 ether_addr_copy(oethh->h_source, ife->eth_src);
 810         else
 811                 ether_addr_copy(oethh->h_source, iethh->h_source);
 812         if (!is_zero_ether_addr(ife->eth_dst))
 813                 ether_addr_copy(oethh->h_dest, ife->eth_dst);
 814         else
 815                 ether_addr_copy(oethh->h_dest, iethh->h_dest);
 816         oethh->h_proto = htons(ife->eth_type);
 817
 818         if (skb_at_tc_ingress(skb))
 819                 skb_pull(skb, skb->dev->hard_header_len);
 820
 821         spin_unlock(&ife->tcf_lock);
 822
 823         return action;
 824 }
 825
 826 static int tcf_ife_act(struct sk_buff *skb, const struct tc_action *a,
 827                        struct tcf_result *res)
 828 {
 829         struct tcf_ife_info *ife = to_ife(a);
 830
 831         if (ife->flags & IFE_ENCODE)
 832                 return tcf_ife_encode(skb, a, res);
 833
 834         if (!(ife->flags & IFE_ENCODE))
 835                 return tcf_ife_decode(skb, a, res);
 836
 837         pr_info_ratelimited("unknown failure(policy neither de/encode\n");
 838         spin_lock(&ife->tcf_lock);
 839         bstats_update(&ife->tcf_bstats, skb);
 840         tcf_lastuse_update(&ife->tcf_tm);
 841         ife->tcf_qstats.drops++;
 842         spin_unlock(&ife->tcf_lock);
 843
 844         return TC_ACT_SHOT;
 845 }
 846
 847 static int tcf_ife_walker(struct net *net, struct sk_buff *skb,
 848                           struct netlink_callback *cb, int type,
 849                           const struct tc_action_ops *ops)
 850 {
 851         struct tc_action_net *tn = net_generic(net, ife_net_id);
 852
 853         return tcf_generic_walker(tn, skb, cb, type, ops);
 854 }
 855
 856 static int tcf_ife_search(struct net *net, struct tc_action **a, u32 index)
 857 {
 858         struct tc_action_net *tn = net_generic(net, ife_net_id);
 859
 860         return tcf_hash_search(tn, a, index);
 861 }
 862
 863 static struct tc_action_ops act_ife_ops = {
 864         .kind = "ife",
 865         .type = TCA_ACT_IFE,
 866         .owner = THIS_MODULE,
 867         .act = tcf_ife_act,
 868         .dump = tcf_ife_dump,
 869         .cleanup = tcf_ife_cleanup,
 870         .init = tcf_ife_init,
 871         .walk = tcf_ife_walker,
 872         .lookup = tcf_ife_search,
 873         .size = sizeof(struct tcf_ife_info),
 874 };
 875
 876 static __net_init int ife_init_net(struct net *net)
 877 {
 878         struct tc_action_net *tn = net_generic(net, ife_net_id);
 879
 880         return tc_action_net_init(tn, &act_ife_ops, IFE_TAB_MASK);
 881 }
 882
 883 static void __net_exit ife_exit_net(struct net *net)
 884 {
 885         struct tc_action_net *tn = net_generic(net, ife_net_id);
 886
 887         tc_action_net_exit(tn);
 888 }
 889
 890 static struct pernet_operations ife_net_ops = {
 891         .init = ife_init_net,
 892         .exit = ife_exit_net,
 893         .id   = &ife_net_id,
 894         .size = sizeof(struct tc_action_net),
 895 };
 896
 897 static int __init ife_init_module(void)
 898 {
 899         return tcf_register_action(&act_ife_ops, &ife_net_ops);
 900 }
 901
 902 static void __exit ife_cleanup_module(void)
 903 {
 904         tcf_unregister_action(&act_ife_ops, &ife_net_ops);
 905 }
 906
 907 module_init(ife_init_module);
 908 module_exit(ife_cleanup_module);
 909
 910 MODULE_AUTHOR("Jamal Hadi Salim(2015)");
 911 MODULE_DESCRIPTION("Inter-FE LFB action");
 912 MODULE_LICENSE("GPL");