sched/act_ife.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/ife.c      Inter-FE action based on ForCES WG InterFE LFB
   4  *
   5  *              Refer to:
   6  *              draft-ietf-forces-interfelfb-03
   7  *              and
   8  *              netdev01 paper:
   9  *              "Distributing Linux Traffic Control Classifier-Action
  10  *              Subsystem"
  11  *              Authors: Jamal Hadi Salim and Damascene M. Joachimpillai
  12  *
  13  * copyright Jamal Hadi Salim (2015)
  14 */
  15
  16 #include <linux/types.h>
  17 #include <linux/kernel.h>
  18 #include <linux/string.h>
  19 #include <linux/errno.h>
  20 #include <linux/skbuff.h>
  21 #include <linux/rtnetlink.h>
  22 #include <linux/module.h>
  23 #include <linux/init.h>
  24 #include <net/net_namespace.h>
  25 #include <net/netlink.h>
  26 #include <net/pkt_sched.h>
  27 #include <net/pkt_cls.h>
  28 #include <uapi/linux/tc_act/tc_ife.h>
  29 #include <net/tc_act/tc_ife.h>
  30 #include <linux/etherdevice.h>
  31 #include <net/ife.h>
  32
  33 static int max_metacnt = IFE_META_MAX + 1;
  34 static struct tc_action_ops act_ife_ops;
  35
  36 static const struct nla_policy ife_policy[TCA_IFE_MAX + 1] = {
  37         [TCA_IFE_PARMS] = { .len = sizeof(struct tc_ife)},
  38         [TCA_IFE_DMAC] = { .len = ETH_ALEN},
  39         [TCA_IFE_SMAC] = { .len = ETH_ALEN},
  40         [TCA_IFE_TYPE] = { .type = NLA_U16},
  41 };
  42
  43 int ife_encode_meta_u16(u16 metaval, void *skbdata, struct tcf_meta_info *mi)
  44 {
  45         u16 edata = 0;
  46
  47         if (mi->metaval)
  48                 edata = *(u16 *)mi->metaval;
  49         else if (metaval)
  50                 edata = metaval;
  51
  52         if (!edata) /* will not encode */
  53                 return 0;
  54
  55         edata = htons(edata);
  56         return ife_tlv_meta_encode(skbdata, mi->metaid, 2, &edata);
  57 }
  58 EXPORT_SYMBOL_GPL(ife_encode_meta_u16);
  59
  60 int ife_get_meta_u32(struct sk_buff *skb, struct tcf_meta_info *mi)
  61 {
  62         if (mi->metaval)
  63                 return nla_put_u32(skb, mi->metaid, *(u32 *)mi->metaval);
  64         else
  65                 return nla_put(skb, mi->metaid, 0, NULL);
  66 }
  67 EXPORT_SYMBOL_GPL(ife_get_meta_u32);
  68
  69 int ife_check_meta_u32(u32 metaval, struct tcf_meta_info *mi)
  70 {
  71         if (metaval || mi->metaval)
  72                 return 8; /* T+L+V == 2+2+4 */
  73
  74         return 0;
  75 }
  76 EXPORT_SYMBOL_GPL(ife_check_meta_u32);
  77
  78 int ife_check_meta_u16(u16 metaval, struct tcf_meta_info *mi)
  79 {
  80         if (metaval || mi->metaval)
  81                 return 8; /* T+L+(V) == 2+2+(2+2bytepad) */
  82
  83         return 0;
  84 }
  85 EXPORT_SYMBOL_GPL(ife_check_meta_u16);
  86
  87 int ife_encode_meta_u32(u32 metaval, void *skbdata, struct tcf_meta_info *mi)
  88 {
  89         u32 edata = metaval;
  90
  91         if (mi->metaval)
  92                 edata = *(u32 *)mi->metaval;
  93         else if (metaval)
  94                 edata = metaval;
  95
  96         if (!edata) /* will not encode */
  97                 return 0;
  98
  99         edata = htonl(edata);
 100         return ife_tlv_meta_encode(skbdata, mi->metaid, 4, &edata);
 101 }
 102 EXPORT_SYMBOL_GPL(ife_encode_meta_u32);
 103
 104 int ife_get_meta_u16(struct sk_buff *skb, struct tcf_meta_info *mi)
 105 {
 106         if (mi->metaval)
 107                 return nla_put_u16(skb, mi->metaid, *(u16 *)mi->metaval);
 108         else
 109                 return nla_put(skb, mi->metaid, 0, NULL);
 110 }
 111 EXPORT_SYMBOL_GPL(ife_get_meta_u16);
 112
 113 int ife_alloc_meta_u32(struct tcf_meta_info *mi, void *metaval, gfp_t gfp)
 114 {
 115         mi->metaval = kmemdup(metaval, sizeof(u32), gfp);
 116         if (!mi->metaval)
 117                 return -ENOMEM;
 118
 119         return 0;
 120 }
 121 EXPORT_SYMBOL_GPL(ife_alloc_meta_u32);
 122
 123 int ife_alloc_meta_u16(struct tcf_meta_info *mi, void *metaval, gfp_t gfp)
 124 {
 125         mi->metaval = kmemdup(metaval, sizeof(u16), gfp);
 126         if (!mi->metaval)
 127                 return -ENOMEM;
 128
 129         return 0;
 130 }
 131 EXPORT_SYMBOL_GPL(ife_alloc_meta_u16);
 132
 133 void ife_release_meta_gen(struct tcf_meta_info *mi)
 134 {
 135         kfree(mi->metaval);
 136 }
 137 EXPORT_SYMBOL_GPL(ife_release_meta_gen);
 138
 139 int ife_validate_meta_u32(void *val, int len)
 140 {
 141         if (len == sizeof(u32))
 142                 return 0;
 143
 144         return -EINVAL;
 145 }
 146 EXPORT_SYMBOL_GPL(ife_validate_meta_u32);
 147
 148 int ife_validate_meta_u16(void *val, int len)
 149 {
 150         /* length will not include padding */
 151         if (len == sizeof(u16))
 152                 return 0;
 153
 154         return -EINVAL;
 155 }
 156 EXPORT_SYMBOL_GPL(ife_validate_meta_u16);
 157
 158 static LIST_HEAD(ifeoplist);
 159 static DEFINE_RWLOCK(ife_mod_lock);
 160
 161 static struct tcf_meta_ops *find_ife_oplist(u16 metaid)
 162 {
 163         struct tcf_meta_ops *o;
 164
 165         read_lock(&ife_mod_lock);
 166         list_for_each_entry(o, &ifeoplist, list) {
 167                 if (o->metaid == metaid) {
 168                         if (!try_module_get(o->owner))
 169                                 o = NULL;
 170                         read_unlock(&ife_mod_lock);
 171                         return o;
 172                 }
 173         }
 174         read_unlock(&ife_mod_lock);
 175
 176         return NULL;
 177 }
 178
 179 int register_ife_op(struct tcf_meta_ops *mops)
 180 {
 181         struct tcf_meta_ops *m;
 182
 183         if (!mops->metaid || !mops->metatype || !mops->name ||
 184             !mops->check_presence || !mops->encode || !mops->decode ||
 185             !mops->get || !mops->alloc)
 186                 return -EINVAL;
 187
 188         write_lock(&ife_mod_lock);
 189
 190         list_for_each_entry(m, &ifeoplist, list) {
 191                 if (m->metaid == mops->metaid ||
 192                     (strcmp(mops->name, m->name) == 0)) {
 193                         write_unlock(&ife_mod_lock);
 194                         return -EEXIST;
 195                 }
 196         }
 197
 198         if (!mops->release)
 199                 mops->release = ife_release_meta_gen;
 200
 201         list_add_tail(&mops->list, &ifeoplist);
 202         write_unlock(&ife_mod_lock);
 203         return 0;
 204 }
 205 EXPORT_SYMBOL_GPL(unregister_ife_op);
 206
 207 int unregister_ife_op(struct tcf_meta_ops *mops)
 208 {
 209         struct tcf_meta_ops *m;
 210         int err = -ENOENT;
 211
 212         write_lock(&ife_mod_lock);
 213         list_for_each_entry(m, &ifeoplist, list) {
 214                 if (m->metaid == mops->metaid) {
 215                         list_del(&mops->list);
 216                         err = 0;
 217                         break;
 218                 }
 219         }
 220         write_unlock(&ife_mod_lock);
 221
 222         return err;
 223 }
 224 EXPORT_SYMBOL_GPL(register_ife_op);
 225
 226 static int ife_validate_metatype(struct tcf_meta_ops *ops, void *val, int len)
 227 {
 228         int ret = 0;
 229         /* XXX: unfortunately cant use nla_policy at this point
 230         * because a length of 0 is valid in the case of
 231         * "allow". "use" semantics do enforce for proper
 232         * length and i couldve use nla_policy but it makes it hard
 233         * to use it just for that..
 234         */
 235         if (ops->validate)
 236                 return ops->validate(val, len);
 237
 238         if (ops->metatype == NLA_U32)
 239                 ret = ife_validate_meta_u32(val, len);
 240         else if (ops->metatype == NLA_U16)
 241                 ret = ife_validate_meta_u16(val, len);
 242
 243         return ret;
 244 }
 245
 246 #ifdef CONFIG_MODULES
 247 static const char *ife_meta_id2name(u32 metaid)
 248 {
 249         switch (metaid) {
 250         case IFE_META_SKBMARK:
 251                 return "skbmark";
 252         case IFE_META_PRIO:
 253                 return "skbprio";
 254         case IFE_META_TCINDEX:
 255                 return "tcindex";
 256         default:
 257                 return "unknown";
 258         }
 259 }
 260 #endif
 261
 262 /* called when adding new meta information
 263 */
 264 static int load_metaops_and_vet(u32 metaid, void *val, int len, bool rtnl_held)
 265 {
 266         struct tcf_meta_ops *ops = find_ife_oplist(metaid);
 267         int ret = 0;
 268
 269         if (!ops) {
 270                 ret = -ENOENT;
 271 #ifdef CONFIG_MODULES
 272                 if (rtnl_held)
 273                         rtnl_unlock();
 274                 request_module("ife-meta-%s", ife_meta_id2name(metaid));
 275                 if (rtnl_held)
 276                         rtnl_lock();
 277                 ops = find_ife_oplist(metaid);
 278 #endif
 279         }
 280
 281         if (ops) {
 282                 ret = 0;
 283                 if (len)
 284                         ret = ife_validate_metatype(ops, val, len);
 285
 286                 module_put(ops->owner);
 287         }
 288
 289         return ret;
 290 }
 291
 292 /* called when adding new meta information
 293 */
 294 static int __add_metainfo(const struct tcf_meta_ops *ops,
 295                           struct tcf_ife_info *ife, u32 metaid, void *metaval,
 296                           int len, bool atomic, bool exists)
 297 {
 298         struct tcf_meta_info *mi = NULL;
 299         int ret = 0;
 300
 301         mi = kzalloc(sizeof(*mi), atomic ? GFP_ATOMIC : GFP_KERNEL);
 302         if (!mi)
 303                 return -ENOMEM;
 304
 305         mi->metaid = metaid;
 306         mi->ops = ops;
 307         if (len > 0) {
 308                 ret = ops->alloc(mi, metaval, atomic ? GFP_ATOMIC : GFP_KERNEL);
 309                 if (ret != 0) {
 310                         kfree(mi);
 311                         return ret;
 312                 }
 313         }
 314
 315         if (exists)
 316                 spin_lock_bh(&ife->tcf_lock);
 317         list_add_tail(&mi->metalist, &ife->metalist);
 318         if (exists)
 319                 spin_unlock_bh(&ife->tcf_lock);
 320
 321         return ret;
 322 }
 323
 324 static int add_metainfo_and_get_ops(const struct tcf_meta_ops *ops,
 325                                     struct tcf_ife_info *ife, u32 metaid,
 326                                     bool exists)
 327 {
 328         int ret;
 329
 330         if (!try_module_get(ops->owner))
 331                 return -ENOENT;
 332         ret = __add_metainfo(ops, ife, metaid, NULL, 0, true, exists);
 333         if (ret)
 334                 module_put(ops->owner);
 335         return ret;
 336 }
 337
 338 static int add_metainfo(struct tcf_ife_info *ife, u32 metaid, void *metaval,
 339                         int len, bool exists)
 340 {
 341         const struct tcf_meta_ops *ops = find_ife_oplist(metaid);
 342         int ret;
 343
 344         if (!ops)
 345                 return -ENOENT;
 346         ret = __add_metainfo(ops, ife, metaid, metaval, len, false, exists);
 347         if (ret)
 348                 /*put back what find_ife_oplist took */
 349                 module_put(ops->owner);
 350         return ret;
 351 }
 352
 353 static int use_all_metadata(struct tcf_ife_info *ife, bool exists)
 354 {
 355         struct tcf_meta_ops *o;
 356         int rc = 0;
 357         int installed = 0;
 358
 359         read_lock(&ife_mod_lock);
 360         list_for_each_entry(o, &ifeoplist, list) {
 361                 rc = add_metainfo_and_get_ops(o, ife, o->metaid, exists);
 362                 if (rc == 0)
 363                         installed += 1;
 364         }
 365         read_unlock(&ife_mod_lock);
 366
 367         if (installed)
 368                 return 0;
 369         else
 370                 return -EINVAL;
 371 }
 372
 373 static int dump_metalist(struct sk_buff *skb, struct tcf_ife_info *ife)
 374 {
 375         struct tcf_meta_info *e;
 376         struct nlattr *nest;
 377         unsigned char *b = skb_tail_pointer(skb);
 378         int total_encoded = 0;
 379
 380         /*can only happen on decode */
 381         if (list_empty(&ife->metalist))
 382                 return 0;
 383
 384         nest = nla_nest_start_noflag(skb, TCA_IFE_METALST);
 385         if (!nest)
 386                 goto out_nlmsg_trim;
 387
 388         list_for_each_entry(e, &ife->metalist, metalist) {
 389                 if (!e->ops->get(skb, e))
 390                         total_encoded += 1;
 391         }
 392
 393         if (!total_encoded)
 394                 goto out_nlmsg_trim;
 395
 396         nla_nest_end(skb, nest);
 397
 398         return 0;
 399
 400 out_nlmsg_trim:
 401         nlmsg_trim(skb, b);
 402         return -1;
 403 }
 404
 405 /* under ife->tcf_lock */
 406 static void _tcf_ife_cleanup(struct tc_action *a)
 407 {
 408         struct tcf_ife_info *ife = to_ife(a);
 409         struct tcf_meta_info *e, *n;
 410
 411         list_for_each_entry_safe(e, n, &ife->metalist, metalist) {
 412                 list_del(&e->metalist);
 413                 if (e->metaval) {
 414                         if (e->ops->release)
 415                                 e->ops->release(e);
 416                         else
 417                                 kfree(e->metaval);
 418                 }
 419                 module_put(e->ops->owner);
 420                 kfree(e);
 421         }
 422 }
 423
 424 static void tcf_ife_cleanup(struct tc_action *a)
 425 {
 426         struct tcf_ife_info *ife = to_ife(a);
 427         struct tcf_ife_params *p;
 428
 429         spin_lock_bh(&ife->tcf_lock);
 430         _tcf_ife_cleanup(a);
 431         spin_unlock_bh(&ife->tcf_lock);
 432
 433         p = rcu_dereference_protected(ife->params, 1);
 434         if (p)
 435                 kfree_rcu(p, rcu);
 436 }
 437
 438 static int load_metalist(struct nlattr **tb, bool rtnl_held)
 439 {
 440         int i;
 441
 442         for (i = 1; i < max_metacnt; i++) {
 443                 if (tb[i]) {
 444                         void *val = nla_data(tb[i]);
 445                         int len = nla_len(tb[i]);
 446                         int rc;
 447
 448                         rc = load_metaops_and_vet(i, val, len, rtnl_held);
 449                         if (rc != 0)
 450                                 return rc;
 451                 }
 452         }
 453
 454         return 0;
 455 }
 456
 457 static int populate_metalist(struct tcf_ife_info *ife, struct nlattr **tb,
 458                              bool exists, bool rtnl_held)
 459 {
 460         int len = 0;
 461         int rc = 0;
 462         int i = 0;
 463         void *val;
 464
 465         for (i = 1; i < max_metacnt; i++) {
 466                 if (tb[i]) {
 467                         val = nla_data(tb[i]);
 468                         len = nla_len(tb[i]);
 469
 470                         rc = add_metainfo(ife, i, val, len, exists);
 471                         if (rc)
 472                                 return rc;
 473                 }
 474         }
 475
 476         return rc;
 477 }
 478
 479 static int tcf_ife_init(struct net *net, struct nlattr *nla,
 480                         struct nlattr *est, struct tc_action **a,
 481                         struct tcf_proto *tp, u32 flags,
 482                         struct netlink_ext_ack *extack)
 483 {
 484         struct tc_action_net *tn = net_generic(net, act_ife_ops.net_id);
 485         bool bind = flags & TCA_ACT_FLAGS_BIND;
 486         struct nlattr *tb[TCA_IFE_MAX + 1];
 487         struct nlattr *tb2[IFE_META_MAX + 1];
 488         struct tcf_chain *goto_ch = NULL;
 489         struct tcf_ife_params *p;
 490         struct tcf_ife_info *ife;
 491         u16 ife_type = ETH_P_IFE;
 492         struct tc_ife *parm;
 493         u8 *daddr = NULL;
 494         u8 *saddr = NULL;
 495         bool exists = false;
 496         int ret = 0;
 497         u32 index;
 498         int err;
 499
 500         if (!nla) {
 501                 NL_SET_ERR_MSG_MOD(extack, "IFE requires attributes to be passed");
 502                 return -EINVAL;
 503         }
 504
 505         err = nla_parse_nested_deprecated(tb, TCA_IFE_MAX, nla, ife_policy,
 506                                           NULL);
 507         if (err < 0)
 508                 return err;
 509
 510         if (!tb[TCA_IFE_PARMS])
 511                 return -EINVAL;
 512
 513         parm = nla_data(tb[TCA_IFE_PARMS]);
 514
 515         /* IFE_DECODE is 0 and indicates the opposite of IFE_ENCODE because
 516          * they cannot run as the same time. Check on all other values which
 517          * are not supported right now.
 518          */
 519         if (parm->flags & ~IFE_ENCODE)
 520                 return -EINVAL;
 521
 522         p = kzalloc(sizeof(*p), GFP_KERNEL);
 523         if (!p)
 524                 return -ENOMEM;
 525
 526         if (tb[TCA_IFE_METALST]) {
 527                 err = nla_parse_nested_deprecated(tb2, IFE_META_MAX,
 528                                                   tb[TCA_IFE_METALST], NULL,
 529                                                   NULL);
 530                 if (err) {
 531                         kfree(p);
 532                         return err;
 533                 }
 534                 err = load_metalist(tb2, !(flags & TCA_ACT_FLAGS_NO_RTNL));
 535                 if (err) {
 536                         kfree(p);
 537                         return err;
 538                 }
 539         }
 540
 541         index = parm->index;
 542         err = tcf_idr_check_alloc(tn, &index, a, bind);
 543         if (err < 0) {
 544                 kfree(p);
 545                 return err;
 546         }
 547         exists = err;
 548         if (exists && bind) {
 549                 kfree(p);
 550                 return 0;
 551         }
 552
 553         if (!exists) {
 554                 ret = tcf_idr_create(tn, index, est, a, &act_ife_ops,
 555                                      bind, true, flags);
 556                 if (ret) {
 557                         tcf_idr_cleanup(tn, index);
 558                         kfree(p);
 559                         return ret;
 560                 }
 561                 ret = ACT_P_CREATED;
 562         } else if (!(flags & TCA_ACT_FLAGS_REPLACE)) {
 563                 tcf_idr_release(*a, bind);
 564                 kfree(p);
 565                 return -EEXIST;
 566         }
 567
 568         ife = to_ife(*a);
 569         if (ret == ACT_P_CREATED)
 570                 INIT_LIST_HEAD(&ife->metalist);
 571
 572         err = tcf_action_check_ctrlact(parm->action, tp, &goto_ch, extack);
 573         if (err < 0)
 574                 goto release_idr;
 575
 576         p->flags = parm->flags;
 577
 578         if (parm->flags & IFE_ENCODE) {
 579                 if (tb[TCA_IFE_TYPE])
 580                         ife_type = nla_get_u16(tb[TCA_IFE_TYPE]);
 581                 if (tb[TCA_IFE_DMAC])
 582                         daddr = nla_data(tb[TCA_IFE_DMAC]);
 583                 if (tb[TCA_IFE_SMAC])
 584                         saddr = nla_data(tb[TCA_IFE_SMAC]);
 585         }
 586
 587         if (parm->flags & IFE_ENCODE) {
 588                 if (daddr)
 589                         ether_addr_copy(p->eth_dst, daddr);
 590                 else
 591                         eth_zero_addr(p->eth_dst);
 592
 593                 if (saddr)
 594                         ether_addr_copy(p->eth_src, saddr);
 595                 else
 596                         eth_zero_addr(p->eth_src);
 597
 598                 p->eth_type = ife_type;
 599         }
 600
 601         if (tb[TCA_IFE_METALST]) {
 602                 err = populate_metalist(ife, tb2, exists,
 603                                         !(flags & TCA_ACT_FLAGS_NO_RTNL));
 604                 if (err)
 605                         goto metadata_parse_err;
 606         } else {
 607                 /* if no passed metadata allow list or passed allow-all
 608                  * then here we process by adding as many supported metadatum
 609                  * as we can. You better have at least one else we are
 610                  * going to bail out
 611                  */
 612                 err = use_all_metadata(ife, exists);
 613                 if (err)
 614                         goto metadata_parse_err;
 615         }
 616
 617         if (exists)
 618                 spin_lock_bh(&ife->tcf_lock);
 619         /* protected by tcf_lock when modifying existing action */
 620         goto_ch = tcf_action_set_ctrlact(*a, parm->action, goto_ch);
 621         p = rcu_replace_pointer(ife->params, p, 1);
 622
 623         if (exists)
 624                 spin_unlock_bh(&ife->tcf_lock);
 625         if (goto_ch)
 626                 tcf_chain_put_by_act(goto_ch);
 627         if (p)
 628                 kfree_rcu(p, rcu);
 629
 630         return ret;
 631 metadata_parse_err:
 632         if (goto_ch)
 633                 tcf_chain_put_by_act(goto_ch);
 634 release_idr:
 635         kfree(p);
 636         tcf_idr_release(*a, bind);
 637         return err;
 638 }
 639
 640 static int tcf_ife_dump(struct sk_buff *skb, struct tc_action *a, int bind,
 641                         int ref)
 642 {
 643         unsigned char *b = skb_tail_pointer(skb);
 644         struct tcf_ife_info *ife = to_ife(a);
 645         struct tcf_ife_params *p;
 646         struct tc_ife opt = {
 647                 .index = ife->tcf_index,
 648                 .refcnt = refcount_read(&ife->tcf_refcnt) - ref,
 649                 .bindcnt = atomic_read(&ife->tcf_bindcnt) - bind,
 650         };
 651         struct tcf_t t;
 652
 653         spin_lock_bh(&ife->tcf_lock);
 654         opt.action = ife->tcf_action;
 655         p = rcu_dereference_protected(ife->params,
 656                                       lockdep_is_held(&ife->tcf_lock));
 657         opt.flags = p->flags;
 658
 659         if (nla_put(skb, TCA_IFE_PARMS, sizeof(opt), &opt))
 660                 goto nla_put_failure;
 661
 662         tcf_tm_dump(&t, &ife->tcf_tm);
 663         if (nla_put_64bit(skb, TCA_IFE_TM, sizeof(t), &t, TCA_IFE_PAD))
 664                 goto nla_put_failure;
 665
 666         if (!is_zero_ether_addr(p->eth_dst)) {
 667                 if (nla_put(skb, TCA_IFE_DMAC, ETH_ALEN, p->eth_dst))
 668                         goto nla_put_failure;
 669         }
 670
 671         if (!is_zero_ether_addr(p->eth_src)) {
 672                 if (nla_put(skb, TCA_IFE_SMAC, ETH_ALEN, p->eth_src))
 673                         goto nla_put_failure;
 674         }
 675
 676         if (nla_put(skb, TCA_IFE_TYPE, 2, &p->eth_type))
 677                 goto nla_put_failure;
 678
 679         if (dump_metalist(skb, ife)) {
 680                 /*ignore failure to dump metalist */
 681                 pr_info("Failed to dump metalist\n");
 682         }
 683
 684         spin_unlock_bh(&ife->tcf_lock);
 685         return skb->len;
 686
 687 nla_put_failure:
 688         spin_unlock_bh(&ife->tcf_lock);
 689         nlmsg_trim(skb, b);
 690         return -1;
 691 }
 692
 693 static int find_decode_metaid(struct sk_buff *skb, struct tcf_ife_info *ife,
 694                               u16 metaid, u16 mlen, void *mdata)
 695 {
 696         struct tcf_meta_info *e;
 697
 698         /* XXX: use hash to speed up */
 699         list_for_each_entry(e, &ife->metalist, metalist) {
 700                 if (metaid == e->metaid) {
 701                         if (e->ops) {
 702                                 /* We check for decode presence already */
 703                                 return e->ops->decode(skb, mdata, mlen);
 704                         }
 705                 }
 706         }
 707
 708         return -ENOENT;
 709 }
 710
 711 static int tcf_ife_decode(struct sk_buff *skb, const struct tc_action *a,
 712                           struct tcf_result *res)
 713 {
 714         struct tcf_ife_info *ife = to_ife(a);
 715         int action = ife->tcf_action;
 716         u8 *ifehdr_end;
 717         u8 *tlv_data;
 718         u16 metalen;
 719
 720         bstats_update(this_cpu_ptr(ife->common.cpu_bstats), skb);
 721         tcf_lastuse_update(&ife->tcf_tm);
 722
 723         if (skb_at_tc_ingress(skb))
 724                 skb_push(skb, skb->dev->hard_header_len);
 725
 726         tlv_data = ife_decode(skb, &metalen);
 727         if (unlikely(!tlv_data)) {
 728                 qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 729                 return TC_ACT_SHOT;
 730         }
 731
 732         ifehdr_end = tlv_data + metalen;
 733         for (; tlv_data < ifehdr_end; tlv_data = ife_tlv_meta_next(tlv_data)) {
 734                 u8 *curr_data;
 735                 u16 mtype;
 736                 u16 dlen;
 737
 738                 curr_data = ife_tlv_meta_decode(tlv_data, ifehdr_end, &mtype,
 739                                                 &dlen, NULL);
 740                 if (!curr_data) {
 741                         qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 742                         return TC_ACT_SHOT;
 743                 }
 744
 745                 if (find_decode_metaid(skb, ife, mtype, dlen, curr_data)) {
 746                         /* abuse overlimits to count when we receive metadata
 747                          * but dont have an ops for it
 748                          */
 749                         pr_info_ratelimited("Unknown metaid %d dlen %d\n",
 750                                             mtype, dlen);
 751                         qstats_overlimit_inc(this_cpu_ptr(ife->common.cpu_qstats));
 752                 }
 753         }
 754
 755         if (WARN_ON(tlv_data != ifehdr_end)) {
 756                 qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 757                 return TC_ACT_SHOT;
 758         }
 759
 760         skb->protocol = eth_type_trans(skb, skb->dev);
 761         skb_reset_network_header(skb);
 762
 763         return action;
 764 }
 765
 766 /*XXX: check if we can do this at install time instead of current
 767  * send data path
 768 **/
 769 static int ife_get_sz(struct sk_buff *skb, struct tcf_ife_info *ife)
 770 {
 771         struct tcf_meta_info *e, *n;
 772         int tot_run_sz = 0, run_sz = 0;
 773
 774         list_for_each_entry_safe(e, n, &ife->metalist, metalist) {
 775                 if (e->ops->check_presence) {
 776                         run_sz = e->ops->check_presence(skb, e);
 777                         tot_run_sz += run_sz;
 778                 }
 779         }
 780
 781         return tot_run_sz;
 782 }
 783
 784 static int tcf_ife_encode(struct sk_buff *skb, const struct tc_action *a,
 785                           struct tcf_result *res, struct tcf_ife_params *p)
 786 {
 787         struct tcf_ife_info *ife = to_ife(a);
 788         int action = ife->tcf_action;
 789         struct ethhdr *oethh;   /* outer ether header */
 790         struct tcf_meta_info *e;
 791         /*
 792            OUTERHDR:TOTMETALEN:{TLVHDR:Metadatum:TLVHDR..}:ORIGDATA
 793            where ORIGDATA = original ethernet header ...
 794          */
 795         u16 metalen = ife_get_sz(skb, ife);
 796         int hdrm = metalen + skb->dev->hard_header_len + IFE_METAHDRLEN;
 797         unsigned int skboff = 0;
 798         int new_len = skb->len + hdrm;
 799         bool exceed_mtu = false;
 800         void *ife_meta;
 801         int err = 0;
 802
 803         if (!skb_at_tc_ingress(skb)) {
 804                 if (new_len > skb->dev->mtu)
 805                         exceed_mtu = true;
 806         }
 807
 808         bstats_update(this_cpu_ptr(ife->common.cpu_bstats), skb);
 809         tcf_lastuse_update(&ife->tcf_tm);
 810
 811         if (!metalen) {         /* no metadata to send */
 812                 /* abuse overlimits to count when we allow packet
 813                  * with no metadata
 814                  */
 815                 qstats_overlimit_inc(this_cpu_ptr(ife->common.cpu_qstats));
 816                 return action;
 817         }
 818         /* could be stupid policy setup or mtu config
 819          * so lets be conservative.. */
 820         if ((action == TC_ACT_SHOT) || exceed_mtu) {
 821                 qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 822                 return TC_ACT_SHOT;
 823         }
 824
 825         if (skb_at_tc_ingress(skb))
 826                 skb_push(skb, skb->dev->hard_header_len);
 827
 828         ife_meta = ife_encode(skb, metalen);
 829
 830         spin_lock(&ife->tcf_lock);
 831
 832         /* XXX: we dont have a clever way of telling encode to
 833          * not repeat some of the computations that are done by
 834          * ops->presence_check...
 835          */
 836         list_for_each_entry(e, &ife->metalist, metalist) {
 837                 if (e->ops->encode) {
 838                         err = e->ops->encode(skb, (void *)(ife_meta + skboff),
 839                                              e);
 840                 }
 841                 if (err < 0) {
 842                         /* too corrupt to keep around if overwritten */
 843                         spin_unlock(&ife->tcf_lock);
 844                         qstats_drop_inc(this_cpu_ptr(ife->common.cpu_qstats));
 845                         return TC_ACT_SHOT;
 846                 }
 847                 skboff += err;
 848         }
 849         spin_unlock(&ife->tcf_lock);
 850         oethh = (struct ethhdr *)skb->data;
 851
 852         if (!is_zero_ether_addr(p->eth_src))
 853                 ether_addr_copy(oethh->h_source, p->eth_src);
 854         if (!is_zero_ether_addr(p->eth_dst))
 855                 ether_addr_copy(oethh->h_dest, p->eth_dst);
 856         oethh->h_proto = htons(p->eth_type);
 857
 858         if (skb_at_tc_ingress(skb))
 859                 skb_pull(skb, skb->dev->hard_header_len);
 860
 861         return action;
 862 }
 863
 864 static int tcf_ife_act(struct sk_buff *skb, const struct tc_action *a,
 865                        struct tcf_result *res)
 866 {
 867         struct tcf_ife_info *ife = to_ife(a);
 868         struct tcf_ife_params *p;
 869         int ret;
 870
 871         p = rcu_dereference_bh(ife->params);
 872         if (p->flags & IFE_ENCODE) {
 873                 ret = tcf_ife_encode(skb, a, res, p);
 874                 return ret;
 875         }
 876
 877         return tcf_ife_decode(skb, a, res);
 878 }
 879
 880 static struct tc_action_ops act_ife_ops = {
 881         .kind = "ife",
 882         .id = TCA_ID_IFE,
 883         .owner = THIS_MODULE,
 884         .act = tcf_ife_act,
 885         .dump = tcf_ife_dump,
 886         .cleanup = tcf_ife_cleanup,
 887         .init = tcf_ife_init,
 888         .size = sizeof(struct tcf_ife_info),
 889 };
 890
 891 static __net_init int ife_init_net(struct net *net)
 892 {
 893         struct tc_action_net *tn = net_generic(net, act_ife_ops.net_id);
 894
 895         return tc_action_net_init(net, tn, &act_ife_ops);
 896 }
 897
 898 static void __net_exit ife_exit_net(struct list_head *net_list)
 899 {
 900         tc_action_net_exit(net_list, act_ife_ops.net_id);
 901 }
 902
 903 static struct pernet_operations ife_net_ops = {
 904         .init = ife_init_net,
 905         .exit_batch = ife_exit_net,
 906         .id   = &act_ife_ops.net_id,
 907         .size = sizeof(struct tc_action_net),
 908 };
 909
 910 static int __init ife_init_module(void)
 911 {
 912         return tcf_register_action(&act_ife_ops, &ife_net_ops);
 913 }
 914
 915 static void __exit ife_cleanup_module(void)
 916 {
 917         tcf_unregister_action(&act_ife_ops, &ife_net_ops);
 918 }
 919
 920 module_init(ife_init_module);
 921 module_exit(ife_cleanup_module);
 922
 923 MODULE_AUTHOR("Jamal Hadi Salim(2015)");
 924 MODULE_DESCRIPTION("Inter-FE LFB action");
 925 MODULE_LICENSE("GPL");