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 #ifdef CONFIG_MODULES
 252 static const char *ife_meta_id2name(u32 metaid)
 253 {
 254         switch (metaid) {
 255         case IFE_META_SKBMARK:
 256                 return "skbmark";
 257         case IFE_META_PRIO:
 258                 return "skbprio";
 259         case IFE_META_TCINDEX:
 260                 return "tcindex";
 261         default:
 262                 return "unknown";
 263         }
 264 }
 265 #endif
 266
 267 /* called when adding new meta information
 268 */
 269 static int load_metaops_and_vet(u32 metaid, void *val, int len, bool rtnl_held)
 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 (rtnl_held)
 278                         rtnl_unlock();
 279                 request_module("ife-meta-%s", ife_meta_id2name(metaid));
 280                 if (rtnl_held)
 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(struct tcf_ife_info *ife, u32 metaid, void *metaval,
 300                         int len, bool atomic, bool exists)
 301 {
 302         struct tcf_meta_info *mi = NULL;
 303         struct tcf_meta_ops *ops = find_ife_oplist(metaid);
 304         int ret = 0;
 305
 306         if (!ops)
 307                 return -ENOENT;
 308
 309         mi = kzalloc(sizeof(*mi), atomic ? GFP_ATOMIC : GFP_KERNEL);
 310         if (!mi) {
 311                 /*put back what find_ife_oplist took */
 312                 module_put(ops->owner);
 313                 return -ENOMEM;
 314         }
 315
 316         mi->metaid = metaid;
 317         mi->ops = ops;
 318         if (len > 0) {
 319                 ret = ops->alloc(mi, metaval, atomic ? GFP_ATOMIC : GFP_KERNEL);
 320                 if (ret != 0) {
 321                         kfree(mi);
 322                         module_put(ops->owner);
 323                         return ret;
 324                 }
 325         }
 326
 327         if (exists)
 328                 spin_lock_bh(&ife->tcf_lock);
 329         list_add_tail(&mi->metalist, &ife->metalist);
 330         if (exists)
 331                 spin_unlock_bh(&ife->tcf_lock);
 332
 333         return ret;
 334 }
 335
 336 static int use_all_metadata(struct tcf_ife_info *ife, bool exists)
 337 {
 338         struct tcf_meta_ops *o;
 339         int rc = 0;
 340         int installed = 0;
 341
 342         read_lock(&ife_mod_lock);
 343         list_for_each_entry(o, &ifeoplist, list) {
 344                 rc = add_metainfo(ife, o->metaid, NULL, 0, true, exists);
 345                 if (rc == 0)
 346                         installed += 1;
 347         }
 348         read_unlock(&ife_mod_lock);
 349
 350         if (installed)
 351                 return 0;
 352         else
 353                 return -EINVAL;
 354 }
 355
 356 static int dump_metalist(struct sk_buff *skb, struct tcf_ife_info *ife)
 357 {
 358         struct tcf_meta_info *e;
 359         struct nlattr *nest;
 360         unsigned char *b = skb_tail_pointer(skb);
 361         int total_encoded = 0;
 362
 363         /*can only happen on decode */
 364         if (list_empty(&ife->metalist))
 365                 return 0;
 366
 367         nest = nla_nest_start(skb, TCA_IFE_METALST);
 368         if (!nest)
 369                 goto out_nlmsg_trim;
 370
 371         list_for_each_entry(e, &ife->metalist, metalist) {
 372                 if (!e->ops->get(skb, e))
 373                         total_encoded += 1;
 374         }
 375
 376         if (!total_encoded)
 377                 goto out_nlmsg_trim;
 378
 379         nla_nest_end(skb, nest);
 380
 381         return 0;
 382
 383 out_nlmsg_trim:
 384         nlmsg_trim(skb, b);
 385         return -1;
 386 }
 387
 388 /* under ife->tcf_lock */
 389 static void _tcf_ife_cleanup(struct tc_action *a)
 390 {
 391         struct tcf_ife_info *ife = to_ife(a);
 392         struct tcf_meta_info *e, *n;
 393
 394         list_for_each_entry_safe(e, n, &ife->metalist, metalist) {
 395                 module_put(e->ops->owner);
 396                 list_del(&e->metalist);
 397                 if (e->metaval) {
 398                         if (e->ops->release)
 399                                 e->ops->release(e);
 400                         else
 401                                 kfree(e->metaval);
 402                 }
 403                 kfree(e);
 404         }
 405 }
 406
 407 static void tcf_ife_cleanup(struct tc_action *a)
 408 {
 409         struct tcf_ife_info *ife = to_ife(a);
 410         struct tcf_ife_params *p;
 411
 412         spin_lock_bh(&ife->tcf_lock);
 413         _tcf_ife_cleanup(a);
 414         spin_unlock_bh(&ife->tcf_lock);
 415
 416         p = rcu_dereference_protected(ife->params, 1);
 417         if (p)
 418                 kfree_rcu(p, rcu);
 419 }
 420
 421 static int populate_metalist(struct tcf_ife_info *ife, struct nlattr **tb,
 422                              bool exists, bool rtnl_held)
 423 {
 424         int len = 0;
 425         int rc = 0;
 426         int i = 0;
 427         void *val;
 428
 429         for (i = 1; i < max_metacnt; i++) {
 430                 if (tb[i]) {
 431                         val = nla_data(tb[i]);
 432                         len = nla_len(tb[i]);
 433
 434                         rc = load_metaops_and_vet(i, val, len, rtnl_held);
 435                         if (rc != 0)
 436                                 return rc;
 437
 438                         rc = add_metainfo(ife, i, val, len, false, exists);
 439                         if (rc)
 440                                 return rc;
 441                 }
 442         }
 443
 444         return rc;
 445 }
 446
 447 static int tcf_ife_init(struct net *net, struct nlattr *nla,
 448                         struct nlattr *est, struct tc_action **a,
 449                         int ovr, int bind, bool rtnl_held,
 450                         struct netlink_ext_ack *extack)
 451 {
 452         struct tc_action_net *tn = net_generic(net, ife_net_id);
 453         struct nlattr *tb[TCA_IFE_MAX + 1];
 454         struct nlattr *tb2[IFE_META_MAX + 1];
 455         struct tcf_ife_params *p;
 456         struct tcf_ife_info *ife;
 457         u16 ife_type = ETH_P_IFE;
 458         struct tc_ife *parm;
 459         u8 *daddr = NULL;
 460         u8 *saddr = NULL;
 461         bool exists = false;
 462         int ret = 0;
 463         int err;
 464
 465         err = nla_parse_nested(tb, TCA_IFE_MAX, nla, ife_policy, NULL);
 466         if (err < 0)
 467                 return err;
 468
 469         if (!tb[TCA_IFE_PARMS])
 470                 return -EINVAL;
 471
 472         parm = nla_data(tb[TCA_IFE_PARMS]);
 473
 474         /* IFE_DECODE is 0 and indicates the opposite of IFE_ENCODE because
 475          * they cannot run as the same time. Check on all other values which
 476          * are not supported right now.
 477          */
 478         if (parm->flags & ~IFE_ENCODE)
 479                 return -EINVAL;
 480
 481         p = kzalloc(sizeof(*p), GFP_KERNEL);
 482         if (!p)
 483                 return -ENOMEM;
 484
 485         err = tcf_idr_check_alloc(tn, &parm->index, a, bind);
 486         if (err < 0) {
 487                 kfree(p);
 488                 return err;
 489         }
 490         exists = err;
 491         if (exists && bind) {
 492                 kfree(p);
 493                 return 0;
 494         }
 495
 496         if (!exists) {
 497                 ret = tcf_idr_create(tn, parm->index, est, a, &act_ife_ops,
 498                                      bind, true);
 499                 if (ret) {
 500                         tcf_idr_cleanup(tn, parm->index);
 501                         kfree(p);
 502                         return ret;
 503                 }
 504                 ret = ACT_P_CREATED;
 505         } else if (!ovr) {
 506                 tcf_idr_release(*a, bind);
 507                 kfree(p);
 508                 return -EEXIST;
 509         }
 510
 511         ife = to_ife(*a);
 512         p->flags = parm->flags;
 513
 514         if (parm->flags & IFE_ENCODE) {
 515                 if (tb[TCA_IFE_TYPE])
 516                         ife_type = nla_get_u16(tb[TCA_IFE_TYPE]);
 517                 if (tb[TCA_IFE_DMAC])
 518                         daddr = nla_data(tb[TCA_IFE_DMAC]);
 519                 if (tb[TCA_IFE_SMAC])
 520                         saddr = nla_data(tb[TCA_IFE_SMAC]);
 521         }
 522
 523         if (parm->flags & IFE_ENCODE) {
 524                 if (daddr)
 525                         ether_addr_copy(p->eth_dst, daddr);
 526                 else
 527                         eth_zero_addr(p->eth_dst);
 528
 529                 if (saddr)
 530                         ether_addr_copy(p->eth_src, saddr);
 531                 else
 532                         eth_zero_addr(p->eth_src);
 533
 534                 p->eth_type = ife_type;
 535         }
 536
 537
 538         if (ret == ACT_P_CREATED)
 539                 INIT_LIST_HEAD(&ife->metalist);
 540
 541         if (tb[TCA_IFE_METALST]) {
 542                 err = nla_parse_nested(tb2, IFE_META_MAX, tb[TCA_IFE_METALST],
 543                                        NULL, NULL);
 544                 if (err) {
 545 metadata_parse_err:
 546                         tcf_idr_release(*a, bind);
 547                         kfree(p);
 548                         return err;
 549                 }
 550
 551                 err = populate_metalist(ife, tb2, exists, rtnl_held);
 552                 if (err)
 553                         goto metadata_parse_err;
 554
 555         } else {
 556                 /* if no passed metadata allow list or passed allow-all
 557                  * then here we process by adding as many supported metadatum
 558                  * as we can. You better have at least one else we are
 559                  * going to bail out
 560                  */
 561                 err = use_all_metadata(ife, exists);
 562                 if (err) {
 563                         tcf_idr_release(*a, bind);
 564                         kfree(p);
 565                         return err;
 566                 }
 567         }
 568
 569         if (exists)
 570                 spin_lock_bh(&ife->tcf_lock);
 571         ife->tcf_action = parm->action;
 572         /* protected by tcf_lock when modifying existing action */
 573         rcu_swap_protected(ife->params, p, 1);
 574
 575         if (exists)
 576                 spin_unlock_bh(&ife->tcf_lock);
 577         if (p)
 578                 kfree_rcu(p, rcu);
 579
 580         if (ret == ACT_P_CREATED)
 581                 tcf_idr_insert(tn, *a);
 582
 583         return ret;
 584 }
 585
 586 static int tcf_ife_dump(struct sk_buff *skb, struct tc_action *a, int bind,
 587                         int ref)
 588 {
 589         unsigned char *b = skb_tail_pointer(skb);
 590         struct tcf_ife_info *ife = to_ife(a);
 591         struct tcf_ife_params *p;
 592         struct tc_ife opt = {
 593                 .index = ife->tcf_index,
 594                 .refcnt = refcount_read(&ife->tcf_refcnt) - ref,
 595                 .bindcnt = atomic_read(&ife->tcf_bindcnt) - bind,
 596         };
 597         struct tcf_t t;
 598
 599         spin_lock_bh(&ife->tcf_lock);
 600         opt.action = ife->tcf_action;
 601         p = rcu_dereference_protected(ife->params,
 602                                       lockdep_is_held(&ife->tcf_lock));
 603         opt.flags = p->flags;
 604
 605         if (nla_put(skb, TCA_IFE_PARMS, sizeof(opt), &opt))
 606                 goto nla_put_failure;
 607
 608         tcf_tm_dump(&t, &ife->tcf_tm);
 609         if (nla_put_64bit(skb, TCA_IFE_TM, sizeof(t), &t, TCA_IFE_PAD))
 610                 goto nla_put_failure;
 611
 612         if (!is_zero_ether_addr(p->eth_dst)) {
 613                 if (nla_put(skb, TCA_IFE_DMAC, ETH_ALEN, p->eth_dst))
 614                         goto nla_put_failure;
 615         }
 616
 617         if (!is_zero_ether_addr(p->eth_src)) {
 618                 if (nla_put(skb, TCA_IFE_SMAC, ETH_ALEN, p->eth_src))
 619                         goto nla_put_failure;
 620         }
 621
 622         if (nla_put(skb, TCA_IFE_TYPE, 2, &p->eth_type))
 623                 goto nla_put_failure;
 624
 625         if (dump_metalist(skb, ife)) {
 626                 /*ignore failure to dump metalist */
 627                 pr_info("Failed to dump metalist\n");
 628         }
 629
 630         spin_unlock_bh(&ife->tcf_lock);
 631         return skb->len;
 632
 633 nla_put_failure:
 634         spin_unlock_bh(&ife->tcf_lock);
 635         nlmsg_trim(skb, b);
 636         return -1;
 637 }
 638
 639 static int find_decode_metaid(struct sk_buff *skb, struct tcf_ife_info *ife,
 640                               u16 metaid, u16 mlen, void *mdata)
 641 {
 642         struct tcf_meta_info *e;
 643
 644         /* XXX: use hash to speed up */
 645         list_for_each_entry(e, &ife->metalist, metalist) {
 646                 if (metaid == e->metaid) {
 647                         if (e->ops) {
 648                                 /* We check for decode presence already */
 649                                 return e->ops->decode(skb, mdata, mlen);
 650                         }
 651                 }
 652         }
 653
 654         return -ENOENT;
 655 }
 656
 657 static int tcf_ife_decode(struct sk_buff *skb, const struct tc_action *a,
 658                           struct tcf_result *res)
 659 {
 660         struct tcf_ife_info *ife = to_ife(a);
 661         int action = ife->tcf_action;
 662         u8 *ifehdr_end;
 663         u8 *tlv_data;
 664         u16 metalen;
 665
 666         bstats_cpu_update(this_cpu_ptr(ife->common.cpu_bstats), skb);
 667         tcf_lastuse_update(&ife->tcf_tm);
 668
 669         if (skb_at_tc_ingress(skb))
 670                 skb_push(skb, skb->dev->hard_header_len);
 671
 672         tlv_data = ife_decode(skb, &metalen);
 673         if (unlikely(!tlv_data)) {
 674                 qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 675                 return TC_ACT_SHOT;
 676         }
 677
 678         ifehdr_end = tlv_data + metalen;
 679         for (; tlv_data < ifehdr_end; tlv_data = ife_tlv_meta_next(tlv_data)) {
 680                 u8 *curr_data;
 681                 u16 mtype;
 682                 u16 dlen;
 683
 684                 curr_data = ife_tlv_meta_decode(tlv_data, ifehdr_end, &mtype,
 685                                                 &dlen, NULL);
 686                 if (!curr_data) {
 687                         qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 688                         return TC_ACT_SHOT;
 689                 }
 690
 691                 if (find_decode_metaid(skb, ife, mtype, dlen, curr_data)) {
 692                         /* abuse overlimits to count when we receive metadata
 693                          * but dont have an ops for it
 694                          */
 695                         pr_info_ratelimited("Unknown metaid %d dlen %d\n",
 696                                             mtype, dlen);
 697                         qstats_overlimit_inc(this_cpu_ptr(ife->common.cpu_qstats));
 698                 }
 699         }
 700
 701         if (WARN_ON(tlv_data != ifehdr_end)) {
 702                 qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 703                 return TC_ACT_SHOT;
 704         }
 705
 706         skb->protocol = eth_type_trans(skb, skb->dev);
 707         skb_reset_network_header(skb);
 708
 709         return action;
 710 }
 711
 712 /*XXX: check if we can do this at install time instead of current
 713  * send data path
 714 **/
 715 static int ife_get_sz(struct sk_buff *skb, struct tcf_ife_info *ife)
 716 {
 717         struct tcf_meta_info *e, *n;
 718         int tot_run_sz = 0, run_sz = 0;
 719
 720         list_for_each_entry_safe(e, n, &ife->metalist, metalist) {
 721                 if (e->ops->check_presence) {
 722                         run_sz = e->ops->check_presence(skb, e);
 723                         tot_run_sz += run_sz;
 724                 }
 725         }
 726
 727         return tot_run_sz;
 728 }
 729
 730 static int tcf_ife_encode(struct sk_buff *skb, const struct tc_action *a,
 731                           struct tcf_result *res, struct tcf_ife_params *p)
 732 {
 733         struct tcf_ife_info *ife = to_ife(a);
 734         int action = ife->tcf_action;
 735         struct ethhdr *oethh;   /* outer ether header */
 736         struct tcf_meta_info *e;
 737         /*
 738            OUTERHDR:TOTMETALEN:{TLVHDR:Metadatum:TLVHDR..}:ORIGDATA
 739            where ORIGDATA = original ethernet header ...
 740          */
 741         u16 metalen = ife_get_sz(skb, ife);
 742         int hdrm = metalen + skb->dev->hard_header_len + IFE_METAHDRLEN;
 743         unsigned int skboff = 0;
 744         int new_len = skb->len + hdrm;
 745         bool exceed_mtu = false;
 746         void *ife_meta;
 747         int err = 0;
 748
 749         if (!skb_at_tc_ingress(skb)) {
 750                 if (new_len > skb->dev->mtu)
 751                         exceed_mtu = true;
 752         }
 753
 754         bstats_cpu_update(this_cpu_ptr(ife->common.cpu_bstats), skb);
 755         tcf_lastuse_update(&ife->tcf_tm);
 756
 757         if (!metalen) {         /* no metadata to send */
 758                 /* abuse overlimits to count when we allow packet
 759                  * with no metadata
 760                  */
 761                 qstats_overlimit_inc(this_cpu_ptr(ife->common.cpu_qstats));
 762                 return action;
 763         }
 764         /* could be stupid policy setup or mtu config
 765          * so lets be conservative.. */
 766         if ((action == TC_ACT_SHOT) || exceed_mtu) {
 767                 qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 768                 return TC_ACT_SHOT;
 769         }
 770
 771         if (skb_at_tc_ingress(skb))
 772                 skb_push(skb, skb->dev->hard_header_len);
 773
 774         ife_meta = ife_encode(skb, metalen);
 775
 776         spin_lock(&ife->tcf_lock);
 777
 778         /* XXX: we dont have a clever way of telling encode to
 779          * not repeat some of the computations that are done by
 780          * ops->presence_check...
 781          */
 782         list_for_each_entry(e, &ife->metalist, metalist) {
 783                 if (e->ops->encode) {
 784                         err = e->ops->encode(skb, (void *)(ife_meta + skboff),
 785                                              e);
 786                 }
 787                 if (err < 0) {
 788                         /* too corrupt to keep around if overwritten */
 789                         spin_unlock(&ife->tcf_lock);
 790                         qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 791                         return TC_ACT_SHOT;
 792                 }
 793                 skboff += err;
 794         }
 795         spin_unlock(&ife->tcf_lock);
 796         oethh = (struct ethhdr *)skb->data;
 797
 798         if (!is_zero_ether_addr(p->eth_src))
 799                 ether_addr_copy(oethh->h_source, p->eth_src);
 800         if (!is_zero_ether_addr(p->eth_dst))
 801                 ether_addr_copy(oethh->h_dest, p->eth_dst);
 802         oethh->h_proto = htons(p->eth_type);
 803
 804         if (skb_at_tc_ingress(skb))
 805                 skb_pull(skb, skb->dev->hard_header_len);
 806
 807         return action;
 808 }
 809
 810 static int tcf_ife_act(struct sk_buff *skb, const struct tc_action *a,
 811                        struct tcf_result *res)
 812 {
 813         struct tcf_ife_info *ife = to_ife(a);
 814         struct tcf_ife_params *p;
 815         int ret;
 816
 817         p = rcu_dereference_bh(ife->params);
 818         if (p->flags & IFE_ENCODE) {
 819                 ret = tcf_ife_encode(skb, a, res, p);
 820                 return ret;
 821         }
 822
 823         return tcf_ife_decode(skb, a, res);
 824 }
 825
 826 static int tcf_ife_walker(struct net *net, struct sk_buff *skb,
 827                           struct netlink_callback *cb, int type,
 828                           const struct tc_action_ops *ops,
 829                           struct netlink_ext_ack *extack)
 830 {
 831         struct tc_action_net *tn = net_generic(net, ife_net_id);
 832
 833         return tcf_generic_walker(tn, skb, cb, type, ops, extack);
 834 }
 835
 836 static int tcf_ife_search(struct net *net, struct tc_action **a, u32 index,
 837                           struct netlink_ext_ack *extack)
 838 {
 839         struct tc_action_net *tn = net_generic(net, ife_net_id);
 840
 841         return tcf_idr_search(tn, a, index);
 842 }
 843
 844 static struct tc_action_ops act_ife_ops = {
 845         .kind = "ife",
 846         .type = TCA_ACT_IFE,
 847         .owner = THIS_MODULE,
 848         .act = tcf_ife_act,
 849         .dump = tcf_ife_dump,
 850         .cleanup = tcf_ife_cleanup,
 851         .init = tcf_ife_init,
 852         .walk = tcf_ife_walker,
 853         .lookup = tcf_ife_search,
 854         .size = sizeof(struct tcf_ife_info),
 855 };
 856
 857 static __net_init int ife_init_net(struct net *net)
 858 {
 859         struct tc_action_net *tn = net_generic(net, ife_net_id);
 860
 861         return tc_action_net_init(tn, &act_ife_ops);
 862 }
 863
 864 static void __net_exit ife_exit_net(struct list_head *net_list)
 865 {
 866         tc_action_net_exit(net_list, ife_net_id);
 867 }
 868
 869 static struct pernet_operations ife_net_ops = {
 870         .init = ife_init_net,
 871         .exit_batch = ife_exit_net,
 872         .id   = &ife_net_id,
 873         .size = sizeof(struct tc_action_net),
 874 };
 875
 876 static int __init ife_init_module(void)
 877 {
 878         return tcf_register_action(&act_ife_ops, &ife_net_ops);
 879 }
 880
 881 static void __exit ife_cleanup_module(void)
 882 {
 883         tcf_unregister_action(&act_ife_ops, &ife_net_ops);
 884 }
 885
 886 module_init(ife_init_module);
 887 module_exit(ife_cleanup_module);
 888
 889 MODULE_AUTHOR("Jamal Hadi Salim(2015)");
 890 MODULE_DESCRIPTION("Inter-FE LFB action");
 891 MODULE_LICENSE("GPL");