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 #include <net/ife.h>
  36
  37 static unsigned int ife_net_id;
  38 static int max_metacnt = IFE_META_MAX + 1;
  39 static struct tc_action_ops act_ife_ops;
  40
  41 static const struct nla_policy ife_policy[TCA_IFE_MAX + 1] = {
  42         [TCA_IFE_PARMS] = { .len = sizeof(struct tc_ife)},
  43         [TCA_IFE_DMAC] = { .len = ETH_ALEN},
  44         [TCA_IFE_SMAC] = { .len = ETH_ALEN},
  45         [TCA_IFE_TYPE] = { .type = NLA_U16},
  46 };
  47
  48 int ife_encode_meta_u16(u16 metaval, void *skbdata, struct tcf_meta_info *mi)
  49 {
  50         u16 edata = 0;
  51
  52         if (mi->metaval)
  53                 edata = *(u16 *)mi->metaval;
  54         else if (metaval)
  55                 edata = metaval;
  56
  57         if (!edata) /* will not encode */
  58                 return 0;
  59
  60         edata = htons(edata);
  61         return ife_tlv_meta_encode(skbdata, mi->metaid, 2, &edata);
  62 }
  63 EXPORT_SYMBOL_GPL(ife_encode_meta_u16);
  64
  65 int ife_get_meta_u32(struct sk_buff *skb, struct tcf_meta_info *mi)
  66 {
  67         if (mi->metaval)
  68                 return nla_put_u32(skb, mi->metaid, *(u32 *)mi->metaval);
  69         else
  70                 return nla_put(skb, mi->metaid, 0, NULL);
  71 }
  72 EXPORT_SYMBOL_GPL(ife_get_meta_u32);
  73
  74 int ife_check_meta_u32(u32 metaval, struct tcf_meta_info *mi)
  75 {
  76         if (metaval || mi->metaval)
  77                 return 8; /* T+L+V == 2+2+4 */
  78
  79         return 0;
  80 }
  81 EXPORT_SYMBOL_GPL(ife_check_meta_u32);
  82
  83 int ife_check_meta_u16(u16 metaval, struct tcf_meta_info *mi)
  84 {
  85         if (metaval || mi->metaval)
  86                 return 8; /* T+L+(V) == 2+2+(2+2bytepad) */
  87
  88         return 0;
  89 }
  90 EXPORT_SYMBOL_GPL(ife_check_meta_u16);
  91
  92 int ife_encode_meta_u32(u32 metaval, void *skbdata, struct tcf_meta_info *mi)
  93 {
  94         u32 edata = metaval;
  95
  96         if (mi->metaval)
  97                 edata = *(u32 *)mi->metaval;
  98         else if (metaval)
  99                 edata = metaval;
 100
 101         if (!edata) /* will not encode */
 102                 return 0;
 103
 104         edata = htonl(edata);
 105         return ife_tlv_meta_encode(skbdata, mi->metaid, 4, &edata);
 106 }
 107 EXPORT_SYMBOL_GPL(ife_encode_meta_u32);
 108
 109 int ife_get_meta_u16(struct sk_buff *skb, struct tcf_meta_info *mi)
 110 {
 111         if (mi->metaval)
 112                 return nla_put_u16(skb, mi->metaid, *(u16 *)mi->metaval);
 113         else
 114                 return nla_put(skb, mi->metaid, 0, NULL);
 115 }
 116 EXPORT_SYMBOL_GPL(ife_get_meta_u16);
 117
 118 int ife_alloc_meta_u32(struct tcf_meta_info *mi, void *metaval, gfp_t gfp)
 119 {
 120         mi->metaval = kmemdup(metaval, sizeof(u32), gfp);
 121         if (!mi->metaval)
 122                 return -ENOMEM;
 123
 124         return 0;
 125 }
 126 EXPORT_SYMBOL_GPL(ife_alloc_meta_u32);
 127
 128 int ife_alloc_meta_u16(struct tcf_meta_info *mi, void *metaval, gfp_t gfp)
 129 {
 130         mi->metaval = kmemdup(metaval, sizeof(u16), gfp);
 131         if (!mi->metaval)
 132                 return -ENOMEM;
 133
 134         return 0;
 135 }
 136 EXPORT_SYMBOL_GPL(ife_alloc_meta_u16);
 137
 138 void ife_release_meta_gen(struct tcf_meta_info *mi)
 139 {
 140         kfree(mi->metaval);
 141 }
 142 EXPORT_SYMBOL_GPL(ife_release_meta_gen);
 143
 144 int ife_validate_meta_u32(void *val, int len)
 145 {
 146         if (len == sizeof(u32))
 147                 return 0;
 148
 149         return -EINVAL;
 150 }
 151 EXPORT_SYMBOL_GPL(ife_validate_meta_u32);
 152
 153 int ife_validate_meta_u16(void *val, int len)
 154 {
 155         /* length will not include padding */
 156         if (len == sizeof(u16))
 157                 return 0;
 158
 159         return -EINVAL;
 160 }
 161 EXPORT_SYMBOL_GPL(ife_validate_meta_u16);
 162
 163 static LIST_HEAD(ifeoplist);
 164 static DEFINE_RWLOCK(ife_mod_lock);
 165
 166 static struct tcf_meta_ops *find_ife_oplist(u16 metaid)
 167 {
 168         struct tcf_meta_ops *o;
 169
 170         read_lock(&ife_mod_lock);
 171         list_for_each_entry(o, &ifeoplist, list) {
 172                 if (o->metaid == metaid) {
 173                         if (!try_module_get(o->owner))
 174                                 o = NULL;
 175                         read_unlock(&ife_mod_lock);
 176                         return o;
 177                 }
 178         }
 179         read_unlock(&ife_mod_lock);
 180
 181         return NULL;
 182 }
 183
 184 int register_ife_op(struct tcf_meta_ops *mops)
 185 {
 186         struct tcf_meta_ops *m;
 187
 188         if (!mops->metaid || !mops->metatype || !mops->name ||
 189             !mops->check_presence || !mops->encode || !mops->decode ||
 190             !mops->get || !mops->alloc)
 191                 return -EINVAL;
 192
 193         write_lock(&ife_mod_lock);
 194
 195         list_for_each_entry(m, &ifeoplist, list) {
 196                 if (m->metaid == mops->metaid ||
 197                     (strcmp(mops->name, m->name) == 0)) {
 198                         write_unlock(&ife_mod_lock);
 199                         return -EEXIST;
 200                 }
 201         }
 202
 203         if (!mops->release)
 204                 mops->release = ife_release_meta_gen;
 205
 206         list_add_tail(&mops->list, &ifeoplist);
 207         write_unlock(&ife_mod_lock);
 208         return 0;
 209 }
 210 EXPORT_SYMBOL_GPL(unregister_ife_op);
 211
 212 int unregister_ife_op(struct tcf_meta_ops *mops)
 213 {
 214         struct tcf_meta_ops *m;
 215         int err = -ENOENT;
 216
 217         write_lock(&ife_mod_lock);
 218         list_for_each_entry(m, &ifeoplist, list) {
 219                 if (m->metaid == mops->metaid) {
 220                         list_del(&mops->list);
 221                         err = 0;
 222                         break;
 223                 }
 224         }
 225         write_unlock(&ife_mod_lock);
 226
 227         return err;
 228 }
 229 EXPORT_SYMBOL_GPL(register_ife_op);
 230
 231 static int ife_validate_metatype(struct tcf_meta_ops *ops, void *val, int len)
 232 {
 233         int ret = 0;
 234         /* XXX: unfortunately cant use nla_policy at this point
 235         * because a length of 0 is valid in the case of
 236         * "allow". "use" semantics do enforce for proper
 237         * length and i couldve use nla_policy but it makes it hard
 238         * to use it just for that..
 239         */
 240         if (ops->validate)
 241                 return ops->validate(val, len);
 242
 243         if (ops->metatype == NLA_U32)
 244                 ret = ife_validate_meta_u32(val, len);
 245         else if (ops->metatype == NLA_U16)
 246                 ret = ife_validate_meta_u16(val, len);
 247
 248         return ret;
 249 }
 250
 251 static const char *ife_meta_id2name(u32 metaid)
 252 {
 253         switch (metaid) {
 254         case IFE_META_SKBMARK:
 255                 return "skbmark";
 256         case IFE_META_PRIO:
 257                 return "skbprio";
 258         case IFE_META_TCINDEX:
 259                 return "tcindex";
 260         default:
 261                 return "unknown";
 262         }
 263 }
 264
 265 /* called when adding new meta information
 266  * under ife->tcf_lock for existing action
 267 */
 268 static int load_metaops_and_vet(struct tcf_ife_info *ife, u32 metaid,
 269                                 void *val, int len, bool exists)
 270 {
 271         struct tcf_meta_ops *ops = find_ife_oplist(metaid);
 272         int ret = 0;
 273
 274         if (!ops) {
 275                 ret = -ENOENT;
 276 #ifdef CONFIG_MODULES
 277                 if (exists)
 278                         spin_unlock_bh(&ife->tcf_lock);
 279                 rtnl_unlock();
 280                 request_module("ife-meta-%s", ife_meta_id2name(metaid));
 281                 rtnl_lock();
 282                 if (exists)
 283                         spin_lock_bh(&ife->tcf_lock);
 284                 ops = find_ife_oplist(metaid);
 285 #endif
 286         }
 287
 288         if (ops) {
 289                 ret = 0;
 290                 if (len)
 291                         ret = ife_validate_metatype(ops, val, len);
 292
 293                 module_put(ops->owner);
 294         }
 295
 296         return ret;
 297 }
 298
 299 /* called when adding new meta information
 300  * under ife->tcf_lock for existing action
 301 */
 302 static int add_metainfo(struct tcf_ife_info *ife, u32 metaid, void *metaval,
 303                         int len, bool atomic)
 304 {
 305         struct tcf_meta_info *mi = NULL;
 306         struct tcf_meta_ops *ops = find_ife_oplist(metaid);
 307         int ret = 0;
 308
 309         if (!ops)
 310                 return -ENOENT;
 311
 312         mi = kzalloc(sizeof(*mi), atomic ? GFP_ATOMIC : GFP_KERNEL);
 313         if (!mi) {
 314                 /*put back what find_ife_oplist took */
 315                 module_put(ops->owner);
 316                 return -ENOMEM;
 317         }
 318
 319         mi->metaid = metaid;
 320         mi->ops = ops;
 321         if (len > 0) {
 322                 ret = ops->alloc(mi, metaval, atomic ? GFP_ATOMIC : GFP_KERNEL);
 323                 if (ret != 0) {
 324                         kfree(mi);
 325                         module_put(ops->owner);
 326                         return ret;
 327                 }
 328         }
 329
 330         list_add_tail(&mi->metalist, &ife->metalist);
 331
 332         return ret;
 333 }
 334
 335 static int use_all_metadata(struct tcf_ife_info *ife)
 336 {
 337         struct tcf_meta_ops *o;
 338         int rc = 0;
 339         int installed = 0;
 340
 341         read_lock(&ife_mod_lock);
 342         list_for_each_entry(o, &ifeoplist, list) {
 343                 rc = add_metainfo(ife, o->metaid, NULL, 0, true);
 344                 if (rc == 0)
 345                         installed += 1;
 346         }
 347         read_unlock(&ife_mod_lock);
 348
 349         if (installed)
 350                 return 0;
 351         else
 352                 return -EINVAL;
 353 }
 354
 355 static int dump_metalist(struct sk_buff *skb, struct tcf_ife_info *ife)
 356 {
 357         struct tcf_meta_info *e;
 358         struct nlattr *nest;
 359         unsigned char *b = skb_tail_pointer(skb);
 360         int total_encoded = 0;
 361
 362         /*can only happen on decode */
 363         if (list_empty(&ife->metalist))
 364                 return 0;
 365
 366         nest = nla_nest_start(skb, TCA_IFE_METALST);
 367         if (!nest)
 368                 goto out_nlmsg_trim;
 369
 370         list_for_each_entry(e, &ife->metalist, metalist) {
 371                 if (!e->ops->get(skb, e))
 372                         total_encoded += 1;
 373         }
 374
 375         if (!total_encoded)
 376                 goto out_nlmsg_trim;
 377
 378         nla_nest_end(skb, nest);
 379
 380         return 0;
 381
 382 out_nlmsg_trim:
 383         nlmsg_trim(skb, b);
 384         return -1;
 385 }
 386
 387 /* under ife->tcf_lock */
 388 static void _tcf_ife_cleanup(struct tc_action *a, int bind)
 389 {
 390         struct tcf_ife_info *ife = to_ife(a);
 391         struct tcf_meta_info *e, *n;
 392
 393         list_for_each_entry_safe(e, n, &ife->metalist, metalist) {
 394                 module_put(e->ops->owner);
 395                 list_del(&e->metalist);
 396                 if (e->metaval) {
 397                         if (e->ops->release)
 398                                 e->ops->release(e);
 399                         else
 400                                 kfree(e->metaval);
 401                 }
 402                 kfree(e);
 403         }
 404 }
 405
 406 static void tcf_ife_cleanup(struct tc_action *a, int bind)
 407 {
 408         struct tcf_ife_info *ife = to_ife(a);
 409         struct tcf_ife_params *p;
 410
 411         spin_lock_bh(&ife->tcf_lock);
 412         _tcf_ife_cleanup(a, bind);
 413         spin_unlock_bh(&ife->tcf_lock);
 414
 415         p = rcu_dereference_protected(ife->params, 1);
 416         kfree_rcu(p, rcu);
 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_params *p, *p_old;
 454         struct tcf_ife_info *ife;
 455         u16 ife_type = ETH_P_IFE;
 456         struct tc_ife *parm;
 457         u8 *daddr = NULL;
 458         u8 *saddr = NULL;
 459         bool exists = false;
 460         int ret = 0;
 461         int err;
 462
 463         err = nla_parse_nested(tb, TCA_IFE_MAX, nla, ife_policy, NULL);
 464         if (err < 0)
 465                 return err;
 466
 467         if (!tb[TCA_IFE_PARMS])
 468                 return -EINVAL;
 469
 470         parm = nla_data(tb[TCA_IFE_PARMS]);
 471
 472         /* IFE_DECODE is 0 and indicates the opposite of IFE_ENCODE because
 473          * they cannot run as the same time. Check on all other values which
 474          * are not supported right now.
 475          */
 476         if (parm->flags & ~IFE_ENCODE)
 477                 return -EINVAL;
 478
 479         p = kzalloc(sizeof(*p), GFP_KERNEL);
 480         if (!p)
 481                 return -ENOMEM;
 482
 483         exists = tcf_idr_check(tn, parm->index, a, bind);
 484         if (exists && bind) {
 485                 kfree(p);
 486                 return 0;
 487         }
 488
 489         if (!exists) {
 490                 ret = tcf_idr_create(tn, parm->index, est, a, &act_ife_ops,
 491                                      bind, true);
 492                 if (ret) {
 493                         kfree(p);
 494                         return ret;
 495                 }
 496                 ret = ACT_P_CREATED;
 497         } else {
 498                 tcf_idr_release(*a, bind);
 499                 if (!ovr) {
 500                         kfree(p);
 501                         return -EEXIST;
 502                 }
 503         }
 504
 505         ife = to_ife(*a);
 506         p->flags = parm->flags;
 507
 508         if (parm->flags & IFE_ENCODE) {
 509                 if (tb[TCA_IFE_TYPE])
 510                         ife_type = nla_get_u16(tb[TCA_IFE_TYPE]);
 511                 if (tb[TCA_IFE_DMAC])
 512                         daddr = nla_data(tb[TCA_IFE_DMAC]);
 513                 if (tb[TCA_IFE_SMAC])
 514                         saddr = nla_data(tb[TCA_IFE_SMAC]);
 515         }
 516
 517         ife->tcf_action = parm->action;
 518
 519         if (parm->flags & IFE_ENCODE) {
 520                 if (daddr)
 521                         ether_addr_copy(p->eth_dst, daddr);
 522                 else
 523                         eth_zero_addr(p->eth_dst);
 524
 525                 if (saddr)
 526                         ether_addr_copy(p->eth_src, saddr);
 527                 else
 528                         eth_zero_addr(p->eth_src);
 529
 530                 p->eth_type = ife_type;
 531         }
 532
 533         if (exists)
 534                 spin_lock_bh(&ife->tcf_lock);
 535
 536         if (ret == ACT_P_CREATED)
 537                 INIT_LIST_HEAD(&ife->metalist);
 538
 539         if (tb[TCA_IFE_METALST]) {
 540                 err = nla_parse_nested(tb2, IFE_META_MAX, tb[TCA_IFE_METALST],
 541                                        NULL, NULL);
 542                 if (err) {
 543 metadata_parse_err:
 544                         if (exists)
 545                                 tcf_idr_release(*a, bind);
 546                         if (ret == ACT_P_CREATED)
 547                                 _tcf_ife_cleanup(*a, bind);
 548
 549                         if (exists)
 550                                 spin_unlock_bh(&ife->tcf_lock);
 551                         kfree(p);
 552                         return err;
 553                 }
 554
 555                 err = populate_metalist(ife, tb2, exists);
 556                 if (err)
 557                         goto metadata_parse_err;
 558
 559         } else {
 560                 /* if no passed metadata allow list or passed allow-all
 561                  * then here we process by adding as many supported metadatum
 562                  * as we can. You better have at least one else we are
 563                  * going to bail out
 564                  */
 565                 err = use_all_metadata(ife);
 566                 if (err) {
 567                         if (ret == ACT_P_CREATED)
 568                                 _tcf_ife_cleanup(*a, bind);
 569
 570                         if (exists)
 571                                 spin_unlock_bh(&ife->tcf_lock);
 572                         kfree(p);
 573                         return err;
 574                 }
 575         }
 576
 577         if (exists)
 578                 spin_unlock_bh(&ife->tcf_lock);
 579
 580         p_old = rtnl_dereference(ife->params);
 581         rcu_assign_pointer(ife->params, p);
 582         if (p_old)
 583                 kfree_rcu(p_old, rcu);
 584
 585         if (ret == ACT_P_CREATED)
 586                 tcf_idr_insert(tn, *a);
 587
 588         return ret;
 589 }
 590
 591 static int tcf_ife_dump(struct sk_buff *skb, struct tc_action *a, int bind,
 592                         int ref)
 593 {
 594         unsigned char *b = skb_tail_pointer(skb);
 595         struct tcf_ife_info *ife = to_ife(a);
 596         struct tcf_ife_params *p = rtnl_dereference(ife->params);
 597         struct tc_ife opt = {
 598                 .index = ife->tcf_index,
 599                 .refcnt = ife->tcf_refcnt - ref,
 600                 .bindcnt = ife->tcf_bindcnt - bind,
 601                 .action = ife->tcf_action,
 602                 .flags = p->flags,
 603         };
 604         struct tcf_t t;
 605
 606         if (nla_put(skb, TCA_IFE_PARMS, sizeof(opt), &opt))
 607                 goto nla_put_failure;
 608
 609         tcf_tm_dump(&t, &ife->tcf_tm);
 610         if (nla_put_64bit(skb, TCA_IFE_TM, sizeof(t), &t, TCA_IFE_PAD))
 611                 goto nla_put_failure;
 612
 613         if (!is_zero_ether_addr(p->eth_dst)) {
 614                 if (nla_put(skb, TCA_IFE_DMAC, ETH_ALEN, p->eth_dst))
 615                         goto nla_put_failure;
 616         }
 617
 618         if (!is_zero_ether_addr(p->eth_src)) {
 619                 if (nla_put(skb, TCA_IFE_SMAC, ETH_ALEN, p->eth_src))
 620                         goto nla_put_failure;
 621         }
 622
 623         if (nla_put(skb, TCA_IFE_TYPE, 2, &p->eth_type))
 624                 goto nla_put_failure;
 625
 626         if (dump_metalist(skb, ife)) {
 627                 /*ignore failure to dump metalist */
 628                 pr_info("Failed to dump metalist\n");
 629         }
 630
 631         return skb->len;
 632
 633 nla_put_failure:
 634         nlmsg_trim(skb, b);
 635         return -1;
 636 }
 637
 638 static int find_decode_metaid(struct sk_buff *skb, struct tcf_ife_info *ife,
 639                               u16 metaid, u16 mlen, void *mdata)
 640 {
 641         struct tcf_meta_info *e;
 642
 643         /* XXX: use hash to speed up */
 644         list_for_each_entry(e, &ife->metalist, metalist) {
 645                 if (metaid == e->metaid) {
 646                         if (e->ops) {
 647                                 /* We check for decode presence already */
 648                                 return e->ops->decode(skb, mdata, mlen);
 649                         }
 650                 }
 651         }
 652
 653         return 0;
 654 }
 655
 656 static int tcf_ife_decode(struct sk_buff *skb, const struct tc_action *a,
 657                           struct tcf_result *res)
 658 {
 659         struct tcf_ife_info *ife = to_ife(a);
 660         int action = ife->tcf_action;
 661         u8 *ifehdr_end;
 662         u8 *tlv_data;
 663         u16 metalen;
 664
 665         bstats_cpu_update(this_cpu_ptr(ife->common.cpu_bstats), skb);
 666         tcf_lastuse_update(&ife->tcf_tm);
 667
 668         if (skb_at_tc_ingress(skb))
 669                 skb_push(skb, skb->dev->hard_header_len);
 670
 671         tlv_data = ife_decode(skb, &metalen);
 672         if (unlikely(!tlv_data)) {
 673                 qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 674                 return TC_ACT_SHOT;
 675         }
 676
 677         ifehdr_end = tlv_data + metalen;
 678         for (; tlv_data < ifehdr_end; tlv_data = ife_tlv_meta_next(tlv_data)) {
 679                 u8 *curr_data;
 680                 u16 mtype;
 681                 u16 dlen;
 682
 683                 curr_data = ife_tlv_meta_decode(tlv_data, &mtype, &dlen, NULL);
 684
 685                 if (find_decode_metaid(skb, ife, mtype, dlen, curr_data)) {
 686                         /* abuse overlimits to count when we receive metadata
 687                          * but dont have an ops for it
 688                          */
 689                         pr_info_ratelimited("Unknown metaid %d dlen %d\n",
 690                                             mtype, dlen);
 691                         qstats_overlimit_inc(this_cpu_ptr(ife->common.cpu_qstats));
 692                 }
 693         }
 694
 695         if (WARN_ON(tlv_data != ifehdr_end)) {
 696                 qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 697                 return TC_ACT_SHOT;
 698         }
 699
 700         skb->protocol = eth_type_trans(skb, skb->dev);
 701         skb_reset_network_header(skb);
 702
 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, struct tcf_ife_params *p)
 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 tcf_meta_info *e;
 731         /*
 732            OUTERHDR:TOTMETALEN:{TLVHDR:Metadatum:TLVHDR..}:ORIGDATA
 733            where ORIGDATA = original ethernet header ...
 734          */
 735         u16 metalen = ife_get_sz(skb, ife);
 736         int hdrm = metalen + skb->dev->hard_header_len + IFE_METAHDRLEN;
 737         unsigned int skboff = 0;
 738         int new_len = skb->len + hdrm;
 739         bool exceed_mtu = false;
 740         void *ife_meta;
 741         int err = 0;
 742
 743         if (!skb_at_tc_ingress(skb)) {
 744                 if (new_len > skb->dev->mtu)
 745                         exceed_mtu = true;
 746         }
 747
 748         bstats_cpu_update(this_cpu_ptr(ife->common.cpu_bstats), skb);
 749         tcf_lastuse_update(&ife->tcf_tm);
 750
 751         if (!metalen) {         /* no metadata to send */
 752                 /* abuse overlimits to count when we allow packet
 753                  * with no metadata
 754                  */
 755                 qstats_overlimit_inc(this_cpu_ptr(ife->common.cpu_qstats));
 756                 return action;
 757         }
 758         /* could be stupid policy setup or mtu config
 759          * so lets be conservative.. */
 760         if ((action == TC_ACT_SHOT) || exceed_mtu) {
 761                 qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 762                 return TC_ACT_SHOT;
 763         }
 764
 765         if (skb_at_tc_ingress(skb))
 766                 skb_push(skb, skb->dev->hard_header_len);
 767
 768         ife_meta = ife_encode(skb, metalen);
 769
 770         spin_lock(&ife->tcf_lock);
 771
 772         /* XXX: we dont have a clever way of telling encode to
 773          * not repeat some of the computations that are done by
 774          * ops->presence_check...
 775          */
 776         list_for_each_entry(e, &ife->metalist, metalist) {
 777                 if (e->ops->encode) {
 778                         err = e->ops->encode(skb, (void *)(ife_meta + skboff),
 779                                              e);
 780                 }
 781                 if (err < 0) {
 782                         /* too corrupt to keep around if overwritten */
 783                         spin_unlock(&ife->tcf_lock);
 784                         qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 785                         return TC_ACT_SHOT;
 786                 }
 787                 skboff += err;
 788         }
 789         spin_unlock(&ife->tcf_lock);
 790         oethh = (struct ethhdr *)skb->data;
 791
 792         if (!is_zero_ether_addr(p->eth_src))
 793                 ether_addr_copy(oethh->h_source, p->eth_src);
 794         if (!is_zero_ether_addr(p->eth_dst))
 795                 ether_addr_copy(oethh->h_dest, p->eth_dst);
 796         oethh->h_proto = htons(p->eth_type);
 797
 798         if (skb_at_tc_ingress(skb))
 799                 skb_pull(skb, skb->dev->hard_header_len);
 800
 801         return action;
 802 }
 803
 804 static int tcf_ife_act(struct sk_buff *skb, const struct tc_action *a,
 805                        struct tcf_result *res)
 806 {
 807         struct tcf_ife_info *ife = to_ife(a);
 808         struct tcf_ife_params *p;
 809         int ret;
 810
 811         rcu_read_lock();
 812         p = rcu_dereference(ife->params);
 813         if (p->flags & IFE_ENCODE) {
 814                 ret = tcf_ife_encode(skb, a, res, p);
 815                 rcu_read_unlock();
 816                 return ret;
 817         }
 818         rcu_read_unlock();
 819
 820         return tcf_ife_decode(skb, a, res);
 821 }
 822
 823 static int tcf_ife_walker(struct net *net, struct sk_buff *skb,
 824                           struct netlink_callback *cb, int type,
 825                           const struct tc_action_ops *ops)
 826 {
 827         struct tc_action_net *tn = net_generic(net, ife_net_id);
 828
 829         return tcf_generic_walker(tn, skb, cb, type, ops);
 830 }
 831
 832 static int tcf_ife_search(struct net *net, struct tc_action **a, u32 index)
 833 {
 834         struct tc_action_net *tn = net_generic(net, ife_net_id);
 835
 836         return tcf_idr_search(tn, a, index);
 837 }
 838
 839 static struct tc_action_ops act_ife_ops = {
 840         .kind = "ife",
 841         .type = TCA_ACT_IFE,
 842         .owner = THIS_MODULE,
 843         .act = tcf_ife_act,
 844         .dump = tcf_ife_dump,
 845         .cleanup = tcf_ife_cleanup,
 846         .init = tcf_ife_init,
 847         .walk = tcf_ife_walker,
 848         .lookup = tcf_ife_search,
 849         .size = sizeof(struct tcf_ife_info),
 850 };
 851
 852 static __net_init int ife_init_net(struct net *net)
 853 {
 854         struct tc_action_net *tn = net_generic(net, ife_net_id);
 855
 856         return tc_action_net_init(tn, &act_ife_ops);
 857 }
 858
 859 static void __net_exit ife_exit_net(struct net *net)
 860 {
 861         struct tc_action_net *tn = net_generic(net, ife_net_id);
 862
 863         tc_action_net_exit(tn);
 864 }
 865
 866 static struct pernet_operations ife_net_ops = {
 867         .init = ife_init_net,
 868         .exit = ife_exit_net,
 869         .id   = &ife_net_id,
 870         .size = sizeof(struct tc_action_net),
 871 };
 872
 873 static int __init ife_init_module(void)
 874 {
 875         return tcf_register_action(&act_ife_ops, &ife_net_ops);
 876 }
 877
 878 static void __exit ife_cleanup_module(void)
 879 {
 880         tcf_unregister_action(&act_ife_ops, &ife_net_ops);
 881 }
 882
 883 module_init(ife_init_module);
 884 module_exit(ife_cleanup_module);
 885
 886 MODULE_AUTHOR("Jamal Hadi Salim(2015)");
 887 MODULE_DESCRIPTION("Inter-FE LFB action");
 888 MODULE_LICENSE("GPL");