net/sched/act_ife.c

   1 /*
   2  * net/sched/ife.c      Inter-FE action based on ForCES WG InterFE LFB
   3  *
   4  *              Refer to:
   5  *              draft-ietf-forces-interfelfb-03
   6  *              and
   7  *              netdev01 paper:
   8  *              "Distributing Linux Traffic Control Classifier-Action
   9  *              Subsystem"
  10  *              Authors: Jamal Hadi Salim and Damascene M. Joachimpillai
  11  *
  12  *              This program is free software; you can redistribute it and/or
  13  *              modify it under the terms of the GNU General Public License
  14  *              as published by the Free Software Foundation; either version
  15  *              2 of the License, or (at your option) any later version.
  16  *
  17  * copyright Jamal Hadi Salim (2015)
  18  *
  19 */
  20
  21 #include <linux/types.h>
  22 #include <linux/kernel.h>
  23 #include <linux/string.h>
  24 #include <linux/errno.h>
  25 #include <linux/skbuff.h>
  26 #include <linux/rtnetlink.h>
  27 #include <linux/module.h>
  28 #include <linux/init.h>
  29 #include <net/net_namespace.h>
  30 #include <net/netlink.h>
  31 #include <net/pkt_sched.h>
  32 #include <uapi/linux/tc_act/tc_ife.h>
  33 #include <net/tc_act/tc_ife.h>
  34 #include <linux/etherdevice.h>
  35
  36 #define IFE_TAB_MASK 15
  37
  38 static int ife_net_id;
  39 static int max_metacnt = IFE_META_MAX + 1;
  40
  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 /* Caller takes care of presenting data in network order
  49 */
  50 int ife_tlv_meta_encode(void *skbdata, u16 attrtype, u16 dlen, const void *dval)
  51 {
  52         u32 *tlv = (u32 *)(skbdata);
  53         u16 totlen = nla_total_size(dlen);      /*alignment + hdr */
  54         char *dptr = (char *)tlv + NLA_HDRLEN;
  55         u32 htlv = attrtype << 16 | totlen;
  56
  57         *tlv = htonl(htlv);
  58         memset(dptr, 0, totlen - NLA_HDRLEN);
  59         memcpy(dptr, dval, dlen);
  60
  61         return totlen;
  62 }
  63 EXPORT_SYMBOL_GPL(ife_tlv_meta_encode);
  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_encode_meta_u32(u32 metaval, void *skbdata, struct tcf_meta_info *mi)
  84 {
  85         u32 edata = metaval;
  86
  87         if (mi->metaval)
  88                 edata = *(u32 *)mi->metaval;
  89         else if (metaval)
  90                 edata = metaval;
  91
  92         if (!edata) /* will not encode */
  93                 return 0;
  94
  95         edata = htonl(edata);
  96         return ife_tlv_meta_encode(skbdata, mi->metaid, 4, &edata);
  97 }
  98 EXPORT_SYMBOL_GPL(ife_encode_meta_u32);
  99
 100 int ife_get_meta_u16(struct sk_buff *skb, struct tcf_meta_info *mi)
 101 {
 102         if (mi->metaval)
 103                 return nla_put_u16(skb, mi->metaid, *(u16 *)mi->metaval);
 104         else
 105                 return nla_put(skb, mi->metaid, 0, NULL);
 106 }
 107 EXPORT_SYMBOL_GPL(ife_get_meta_u16);
 108
 109 int ife_alloc_meta_u32(struct tcf_meta_info *mi, void *metaval, gfp_t gfp)
 110 {
 111         mi->metaval = kmemdup(metaval, sizeof(u32), gfp);
 112         if (!mi->metaval)
 113                 return -ENOMEM;
 114
 115         return 0;
 116 }
 117 EXPORT_SYMBOL_GPL(ife_alloc_meta_u32);
 118
 119 int ife_alloc_meta_u16(struct tcf_meta_info *mi, void *metaval, gfp_t gfp)
 120 {
 121         mi->metaval = kmemdup(metaval, sizeof(u16), gfp);
 122         if (!mi->metaval)
 123                 return -ENOMEM;
 124
 125         return 0;
 126 }
 127 EXPORT_SYMBOL_GPL(ife_alloc_meta_u16);
 128
 129 void ife_release_meta_gen(struct tcf_meta_info *mi)
 130 {
 131         kfree(mi->metaval);
 132 }
 133 EXPORT_SYMBOL_GPL(ife_release_meta_gen);
 134
 135 int ife_validate_meta_u32(void *val, int len)
 136 {
 137         if (len == 4)
 138                 return 0;
 139
 140         return -EINVAL;
 141 }
 142 EXPORT_SYMBOL_GPL(ife_validate_meta_u32);
 143
 144 int ife_validate_meta_u16(void *val, int len)
 145 {
 146         /* length will include padding */
 147         if (len == NLA_ALIGN(2))
 148                 return 0;
 149
 150         return -EINVAL;
 151 }
 152 EXPORT_SYMBOL_GPL(ife_validate_meta_u16);
 153
 154 static LIST_HEAD(ifeoplist);
 155 static DEFINE_RWLOCK(ife_mod_lock);
 156
 157 static struct tcf_meta_ops *find_ife_oplist(u16 metaid)
 158 {
 159         struct tcf_meta_ops *o;
 160
 161         read_lock(&ife_mod_lock);
 162         list_for_each_entry(o, &ifeoplist, list) {
 163                 if (o->metaid == metaid) {
 164                         if (!try_module_get(o->owner))
 165                                 o = NULL;
 166                         read_unlock(&ife_mod_lock);
 167                         return o;
 168                 }
 169         }
 170         read_unlock(&ife_mod_lock);
 171
 172         return NULL;
 173 }
 174
 175 int register_ife_op(struct tcf_meta_ops *mops)
 176 {
 177         struct tcf_meta_ops *m;
 178
 179         if (!mops->metaid || !mops->metatype || !mops->name ||
 180             !mops->check_presence || !mops->encode || !mops->decode ||
 181             !mops->get || !mops->alloc)
 182                 return -EINVAL;
 183
 184         write_lock(&ife_mod_lock);
 185
 186         list_for_each_entry(m, &ifeoplist, list) {
 187                 if (m->metaid == mops->metaid ||
 188                     (strcmp(mops->name, m->name) == 0)) {
 189                         write_unlock(&ife_mod_lock);
 190                         return -EEXIST;
 191                 }
 192         }
 193
 194         if (!mops->release)
 195                 mops->release = ife_release_meta_gen;
 196
 197         list_add_tail(&mops->list, &ifeoplist);
 198         write_unlock(&ife_mod_lock);
 199         return 0;
 200 }
 201 EXPORT_SYMBOL_GPL(unregister_ife_op);
 202
 203 int unregister_ife_op(struct tcf_meta_ops *mops)
 204 {
 205         struct tcf_meta_ops *m;
 206         int err = -ENOENT;
 207
 208         write_lock(&ife_mod_lock);
 209         list_for_each_entry(m, &ifeoplist, list) {
 210                 if (m->metaid == mops->metaid) {
 211                         list_del(&mops->list);
 212                         err = 0;
 213                         break;
 214                 }
 215         }
 216         write_unlock(&ife_mod_lock);
 217
 218         return err;
 219 }
 220 EXPORT_SYMBOL_GPL(register_ife_op);
 221
 222 static int ife_validate_metatype(struct tcf_meta_ops *ops, void *val, int len)
 223 {
 224         int ret = 0;
 225         /* XXX: unfortunately cant use nla_policy at this point
 226         * because a length of 0 is valid in the case of
 227         * "allow". "use" semantics do enforce for proper
 228         * length and i couldve use nla_policy but it makes it hard
 229         * to use it just for that..
 230         */
 231         if (ops->validate)
 232                 return ops->validate(val, len);
 233
 234         if (ops->metatype == NLA_U32)
 235                 ret = ife_validate_meta_u32(val, len);
 236         else if (ops->metatype == NLA_U16)
 237                 ret = ife_validate_meta_u16(val, len);
 238
 239         return ret;
 240 }
 241
 242 /* called when adding new meta information
 243  * under ife->tcf_lock for existing action
 244 */
 245 static int load_metaops_and_vet(struct tcf_ife_info *ife, u32 metaid,
 246                                 void *val, int len, bool exists)
 247 {
 248         struct tcf_meta_ops *ops = find_ife_oplist(metaid);
 249         int ret = 0;
 250
 251         if (!ops) {
 252                 ret = -ENOENT;
 253 #ifdef CONFIG_MODULES
 254                 if (exists)
 255                         spin_unlock_bh(&ife->tcf_lock);
 256                 rtnl_unlock();
 257                 request_module("ifemeta%u", metaid);
 258                 rtnl_lock();
 259                 if (exists)
 260                         spin_lock_bh(&ife->tcf_lock);
 261                 ops = find_ife_oplist(metaid);
 262 #endif
 263         }
 264
 265         if (ops) {
 266                 ret = 0;
 267                 if (len)
 268                         ret = ife_validate_metatype(ops, val, len);
 269
 270                 module_put(ops->owner);
 271         }
 272
 273         return ret;
 274 }
 275
 276 /* called when adding new meta information
 277  * under ife->tcf_lock for existing action
 278 */
 279 static int add_metainfo(struct tcf_ife_info *ife, u32 metaid, void *metaval,
 280                         int len, bool exists)
 281 {
 282         struct tcf_meta_info *mi = NULL;
 283         struct tcf_meta_ops *ops = find_ife_oplist(metaid);
 284         int ret = 0;
 285
 286         if (!ops)
 287                 return -ENOENT;
 288
 289         mi = kzalloc(sizeof(*mi), exists ? GFP_ATOMIC : GFP_KERNEL);
 290         if (!mi) {
 291                 /*put back what find_ife_oplist took */
 292                 module_put(ops->owner);
 293                 return -ENOMEM;
 294         }
 295
 296         mi->metaid = metaid;
 297         mi->ops = ops;
 298         if (len > 0) {
 299                 ret = ops->alloc(mi, metaval, exists ? GFP_ATOMIC : GFP_KERNEL);
 300                 if (ret != 0) {
 301                         kfree(mi);
 302                         module_put(ops->owner);
 303                         return ret;
 304                 }
 305         }
 306
 307         list_add_tail(&mi->metalist, &ife->metalist);
 308
 309         return ret;
 310 }
 311
 312 static int use_all_metadata(struct tcf_ife_info *ife, bool exists)
 313 {
 314         struct tcf_meta_ops *o;
 315         int rc = 0;
 316         int installed = 0;
 317
 318         list_for_each_entry(o, &ifeoplist, list) {
 319                 rc = add_metainfo(ife, o->metaid, NULL, 0, exists);
 320                 if (rc == 0)
 321                         installed += 1;
 322         }
 323
 324         if (installed)
 325                 return 0;
 326         else
 327                 return -EINVAL;
 328 }
 329
 330 static int dump_metalist(struct sk_buff *skb, struct tcf_ife_info *ife)
 331 {
 332         struct tcf_meta_info *e;
 333         struct nlattr *nest;
 334         unsigned char *b = skb_tail_pointer(skb);
 335         int total_encoded = 0;
 336
 337         /*can only happen on decode */
 338         if (list_empty(&ife->metalist))
 339                 return 0;
 340
 341         nest = nla_nest_start(skb, TCA_IFE_METALST);
 342         if (!nest)
 343                 goto out_nlmsg_trim;
 344
 345         list_for_each_entry(e, &ife->metalist, metalist) {
 346                 if (!e->ops->get(skb, e))
 347                         total_encoded += 1;
 348         }
 349
 350         if (!total_encoded)
 351                 goto out_nlmsg_trim;
 352
 353         nla_nest_end(skb, nest);
 354
 355         return 0;
 356
 357 out_nlmsg_trim:
 358         nlmsg_trim(skb, b);
 359         return -1;
 360 }
 361
 362 /* under ife->tcf_lock */
 363 static void _tcf_ife_cleanup(struct tc_action *a, int bind)
 364 {
 365         struct tcf_ife_info *ife = a->priv;
 366         struct tcf_meta_info *e, *n;
 367
 368         list_for_each_entry_safe(e, n, &ife->metalist, metalist) {
 369                 module_put(e->ops->owner);
 370                 list_del(&e->metalist);
 371                 if (e->metaval) {
 372                         if (e->ops->release)
 373                                 e->ops->release(e);
 374                         else
 375                                 kfree(e->metaval);
 376                 }
 377                 kfree(e);
 378         }
 379 }
 380
 381 static void tcf_ife_cleanup(struct tc_action *a, int bind)
 382 {
 383         struct tcf_ife_info *ife = a->priv;
 384
 385         spin_lock_bh(&ife->tcf_lock);
 386         _tcf_ife_cleanup(a, bind);
 387         spin_unlock_bh(&ife->tcf_lock);
 388 }
 389
 390 /* under ife->tcf_lock for existing action */
 391 static int populate_metalist(struct tcf_ife_info *ife, struct nlattr **tb,
 392                              bool exists)
 393 {
 394         int len = 0;
 395         int rc = 0;
 396         int i = 0;
 397         void *val;
 398
 399         for (i = 1; i < max_metacnt; i++) {
 400                 if (tb[i]) {
 401                         val = nla_data(tb[i]);
 402                         len = nla_len(tb[i]);
 403
 404                         rc = load_metaops_and_vet(ife, i, val, len, exists);
 405                         if (rc != 0)
 406                                 return rc;
 407
 408                         rc = add_metainfo(ife, i, val, len, exists);
 409                         if (rc)
 410                                 return rc;
 411                 }
 412         }
 413
 414         return rc;
 415 }
 416
 417 static int tcf_ife_init(struct net *net, struct nlattr *nla,
 418                         struct nlattr *est, struct tc_action *a,
 419                         int ovr, int bind)
 420 {
 421         struct tc_action_net *tn = net_generic(net, ife_net_id);
 422         struct nlattr *tb[TCA_IFE_MAX + 1];
 423         struct nlattr *tb2[IFE_META_MAX + 1];
 424         struct tcf_ife_info *ife;
 425         struct tc_ife *parm;
 426         u16 ife_type = 0;
 427         u8 *daddr = NULL;
 428         u8 *saddr = NULL;
 429         int ret = 0, exists = 0;
 430         int err;
 431
 432         err = nla_parse_nested(tb, TCA_IFE_MAX, nla, ife_policy);
 433         if (err < 0)
 434                 return err;
 435
 436         if (!tb[TCA_IFE_PARMS])
 437                 return -EINVAL;
 438
 439         parm = nla_data(tb[TCA_IFE_PARMS]);
 440
 441         exists = tcf_hash_check(tn, parm->index, a, bind);
 442         if (exists && bind)
 443                 return 0;
 444
 445         if (parm->flags & IFE_ENCODE) {
 446                 /* Until we get issued the ethertype, we cant have
 447                  * a default..
 448                 **/
 449                 if (!tb[TCA_IFE_TYPE]) {
 450                         if (exists)
 451                                 tcf_hash_release(a, bind);
 452                         pr_info("You MUST pass etherype for encoding\n");
 453                         return -EINVAL;
 454                 }
 455         }
 456
 457         if (!exists) {
 458                 ret = tcf_hash_create(tn, parm->index, est, a, sizeof(*ife),
 459                                       bind, false);
 460                 if (ret)
 461                         return ret;
 462                 ret = ACT_P_CREATED;
 463         } else {
 464                 tcf_hash_release(a, bind);
 465                 if (!ovr)
 466                         return -EEXIST;
 467         }
 468
 469         ife = to_ife(a);
 470         ife->flags = parm->flags;
 471
 472         if (parm->flags & IFE_ENCODE) {
 473                 ife_type = nla_get_u16(tb[TCA_IFE_TYPE]);
 474                 if (tb[TCA_IFE_DMAC])
 475                         daddr = nla_data(tb[TCA_IFE_DMAC]);
 476                 if (tb[TCA_IFE_SMAC])
 477                         saddr = nla_data(tb[TCA_IFE_SMAC]);
 478         }
 479
 480         if (exists)
 481                 spin_lock_bh(&ife->tcf_lock);
 482         ife->tcf_action = parm->action;
 483
 484         if (parm->flags & IFE_ENCODE) {
 485                 if (daddr)
 486                         ether_addr_copy(ife->eth_dst, daddr);
 487                 else
 488                         eth_zero_addr(ife->eth_dst);
 489
 490                 if (saddr)
 491                         ether_addr_copy(ife->eth_src, saddr);
 492                 else
 493                         eth_zero_addr(ife->eth_src);
 494
 495                 ife->eth_type = ife_type;
 496         }
 497
 498         if (ret == ACT_P_CREATED)
 499                 INIT_LIST_HEAD(&ife->metalist);
 500
 501         if (tb[TCA_IFE_METALST]) {
 502                 err = nla_parse_nested(tb2, IFE_META_MAX, tb[TCA_IFE_METALST],
 503                                        NULL);
 504                 if (err) {
 505 metadata_parse_err:
 506                         if (exists)
 507                                 tcf_hash_release(a, bind);
 508                         if (ret == ACT_P_CREATED)
 509                                 _tcf_ife_cleanup(a, bind);
 510
 511                         if (exists)
 512                                 spin_unlock_bh(&ife->tcf_lock);
 513                         return err;
 514                 }
 515
 516                 err = populate_metalist(ife, tb2, exists);
 517                 if (err)
 518                         goto metadata_parse_err;
 519
 520         } else {
 521                 /* if no passed metadata allow list or passed allow-all
 522                  * then here we process by adding as many supported metadatum
 523                  * as we can. You better have at least one else we are
 524                  * going to bail out
 525                  */
 526                 err = use_all_metadata(ife, exists);
 527                 if (err) {
 528                         if (ret == ACT_P_CREATED)
 529                                 _tcf_ife_cleanup(a, bind);
 530
 531                         if (exists)
 532                                 spin_unlock_bh(&ife->tcf_lock);
 533                         return err;
 534                 }
 535         }
 536
 537         if (exists)
 538                 spin_unlock_bh(&ife->tcf_lock);
 539
 540         if (ret == ACT_P_CREATED)
 541                 tcf_hash_insert(tn, a);
 542
 543         return ret;
 544 }
 545
 546 static int tcf_ife_dump(struct sk_buff *skb, struct tc_action *a, int bind,
 547                         int ref)
 548 {
 549         unsigned char *b = skb_tail_pointer(skb);
 550         struct tcf_ife_info *ife = a->priv;
 551         struct tc_ife opt = {
 552                 .index = ife->tcf_index,
 553                 .refcnt = ife->tcf_refcnt - ref,
 554                 .bindcnt = ife->tcf_bindcnt - bind,
 555                 .action = ife->tcf_action,
 556                 .flags = ife->flags,
 557         };
 558         struct tcf_t t;
 559
 560         if (nla_put(skb, TCA_IFE_PARMS, sizeof(opt), &opt))
 561                 goto nla_put_failure;
 562
 563         t.install = jiffies_to_clock_t(jiffies - ife->tcf_tm.install);
 564         t.lastuse = jiffies_to_clock_t(jiffies - ife->tcf_tm.lastuse);
 565         t.expires = jiffies_to_clock_t(ife->tcf_tm.expires);
 566         if (nla_put_64bit(skb, TCA_IFE_TM, sizeof(t), &t, TCA_IFE_PAD))
 567                 goto nla_put_failure;
 568
 569         if (!is_zero_ether_addr(ife->eth_dst)) {
 570                 if (nla_put(skb, TCA_IFE_DMAC, ETH_ALEN, ife->eth_dst))
 571                         goto nla_put_failure;
 572         }
 573
 574         if (!is_zero_ether_addr(ife->eth_src)) {
 575                 if (nla_put(skb, TCA_IFE_SMAC, ETH_ALEN, ife->eth_src))
 576                         goto nla_put_failure;
 577         }
 578
 579         if (nla_put(skb, TCA_IFE_TYPE, 2, &ife->eth_type))
 580                 goto nla_put_failure;
 581
 582         if (dump_metalist(skb, ife)) {
 583                 /*ignore failure to dump metalist */
 584                 pr_info("Failed to dump metalist\n");
 585         }
 586
 587         return skb->len;
 588
 589 nla_put_failure:
 590         nlmsg_trim(skb, b);
 591         return -1;
 592 }
 593
 594 int find_decode_metaid(struct sk_buff *skb, struct tcf_ife_info *ife,
 595                        u16 metaid, u16 mlen, void *mdata)
 596 {
 597         struct tcf_meta_info *e;
 598
 599         /* XXX: use hash to speed up */
 600         list_for_each_entry(e, &ife->metalist, metalist) {
 601                 if (metaid == e->metaid) {
 602                         if (e->ops) {
 603                                 /* We check for decode presence already */
 604                                 return e->ops->decode(skb, mdata, mlen);
 605                         }
 606                 }
 607         }
 608
 609         return 0;
 610 }
 611
 612 struct ifeheadr {
 613         __be16 metalen;
 614         u8 tlv_data[];
 615 };
 616
 617 struct meta_tlvhdr {
 618         __be16 type;
 619         __be16 len;
 620 };
 621
 622 static int tcf_ife_decode(struct sk_buff *skb, const struct tc_action *a,
 623                           struct tcf_result *res)
 624 {
 625         struct tcf_ife_info *ife = a->priv;
 626         int action = ife->tcf_action;
 627         struct ifeheadr *ifehdr = (struct ifeheadr *)skb->data;
 628         u16 ifehdrln = ifehdr->metalen;
 629         struct meta_tlvhdr *tlv = (struct meta_tlvhdr *)(ifehdr->tlv_data);
 630
 631         spin_lock(&ife->tcf_lock);
 632         bstats_update(&ife->tcf_bstats, skb);
 633         ife->tcf_tm.lastuse = jiffies;
 634         spin_unlock(&ife->tcf_lock);
 635
 636         ifehdrln = ntohs(ifehdrln);
 637         if (unlikely(!pskb_may_pull(skb, ifehdrln))) {
 638                 spin_lock(&ife->tcf_lock);
 639                 ife->tcf_qstats.drops++;
 640                 spin_unlock(&ife->tcf_lock);
 641                 return TC_ACT_SHOT;
 642         }
 643
 644         skb_set_mac_header(skb, ifehdrln);
 645         __skb_pull(skb, ifehdrln);
 646         skb->protocol = eth_type_trans(skb, skb->dev);
 647         ifehdrln -= IFE_METAHDRLEN;
 648
 649         while (ifehdrln > 0) {
 650                 u8 *tlvdata = (u8 *)tlv;
 651                 u16 mtype = tlv->type;
 652                 u16 mlen = tlv->len;
 653
 654                 mtype = ntohs(mtype);
 655                 mlen = ntohs(mlen);
 656
 657                 if (find_decode_metaid(skb, ife, mtype, (mlen - 4),
 658                                        (void *)(tlvdata + 4))) {
 659                         /* abuse overlimits to count when we receive metadata
 660                          * but dont have an ops for it
 661                          */
 662                         pr_info_ratelimited("Unknown metaid %d alnlen %d\n",
 663                                             mtype, mlen);
 664                         ife->tcf_qstats.overlimits++;
 665                 }
 666
 667                 tlvdata += mlen;
 668                 ifehdrln -= mlen;
 669                 tlv = (struct meta_tlvhdr *)tlvdata;
 670         }
 671
 672         skb_reset_network_header(skb);
 673         return action;
 674 }
 675
 676 /*XXX: check if we can do this at install time instead of current
 677  * send data path
 678 **/
 679 static int ife_get_sz(struct sk_buff *skb, struct tcf_ife_info *ife)
 680 {
 681         struct tcf_meta_info *e, *n;
 682         int tot_run_sz = 0, run_sz = 0;
 683
 684         list_for_each_entry_safe(e, n, &ife->metalist, metalist) {
 685                 if (e->ops->check_presence) {
 686                         run_sz = e->ops->check_presence(skb, e);
 687                         tot_run_sz += run_sz;
 688                 }
 689         }
 690
 691         return tot_run_sz;
 692 }
 693
 694 static int tcf_ife_encode(struct sk_buff *skb, const struct tc_action *a,
 695                           struct tcf_result *res)
 696 {
 697         struct tcf_ife_info *ife = a->priv;
 698         int action = ife->tcf_action;
 699         struct ethhdr *oethh;   /* outer ether header */
 700         struct ethhdr *iethh;   /* inner eth header */
 701         struct tcf_meta_info *e;
 702         /*
 703            OUTERHDR:TOTMETALEN:{TLVHDR:Metadatum:TLVHDR..}:ORIGDATA
 704            where ORIGDATA = original ethernet header ...
 705          */
 706         u16 metalen = ife_get_sz(skb, ife);
 707         int hdrm = metalen + skb->dev->hard_header_len + IFE_METAHDRLEN;
 708         unsigned int skboff = skb->dev->hard_header_len;
 709         u32 at = G_TC_AT(skb->tc_verd);
 710         int new_len = skb->len + hdrm;
 711         bool exceed_mtu = false;
 712         int err;
 713
 714         if (at & AT_EGRESS) {
 715                 if (new_len > skb->dev->mtu)
 716                         exceed_mtu = true;
 717         }
 718
 719         spin_lock(&ife->tcf_lock);
 720         bstats_update(&ife->tcf_bstats, skb);
 721         ife->tcf_tm.lastuse = jiffies;
 722
 723         if (!metalen) {         /* no metadata to send */
 724                 /* abuse overlimits to count when we allow packet
 725                  * with no metadata
 726                  */
 727                 ife->tcf_qstats.overlimits++;
 728                 spin_unlock(&ife->tcf_lock);
 729                 return action;
 730         }
 731         /* could be stupid policy setup or mtu config
 732          * so lets be conservative.. */
 733         if ((action == TC_ACT_SHOT) || exceed_mtu) {
 734                 ife->tcf_qstats.drops++;
 735                 spin_unlock(&ife->tcf_lock);
 736                 return TC_ACT_SHOT;
 737         }
 738
 739         iethh = eth_hdr(skb);
 740
 741         err = skb_cow_head(skb, hdrm);
 742         if (unlikely(err)) {
 743                 ife->tcf_qstats.drops++;
 744                 spin_unlock(&ife->tcf_lock);
 745                 return TC_ACT_SHOT;
 746         }
 747
 748         if (!(at & AT_EGRESS))
 749                 skb_push(skb, skb->dev->hard_header_len);
 750
 751         __skb_push(skb, hdrm);
 752         memcpy(skb->data, iethh, skb->mac_len);
 753         skb_reset_mac_header(skb);
 754         oethh = eth_hdr(skb);
 755
 756         /*total metadata length */
 757         metalen += IFE_METAHDRLEN;
 758         metalen = htons(metalen);
 759         memcpy((skb->data + skboff), &metalen, IFE_METAHDRLEN);
 760         skboff += IFE_METAHDRLEN;
 761
 762         /* XXX: we dont have a clever way of telling encode to
 763          * not repeat some of the computations that are done by
 764          * ops->presence_check...
 765          */
 766         list_for_each_entry(e, &ife->metalist, metalist) {
 767                 if (e->ops->encode) {
 768                         err = e->ops->encode(skb, (void *)(skb->data + skboff),
 769                                              e);
 770                 }
 771                 if (err < 0) {
 772                         /* too corrupt to keep around if overwritten */
 773                         ife->tcf_qstats.drops++;
 774                         spin_unlock(&ife->tcf_lock);
 775                         return TC_ACT_SHOT;
 776                 }
 777                 skboff += err;
 778         }
 779
 780         if (!is_zero_ether_addr(ife->eth_src))
 781                 ether_addr_copy(oethh->h_source, ife->eth_src);
 782         else
 783                 ether_addr_copy(oethh->h_source, iethh->h_source);
 784         if (!is_zero_ether_addr(ife->eth_dst))
 785                 ether_addr_copy(oethh->h_dest, ife->eth_dst);
 786         else
 787                 ether_addr_copy(oethh->h_dest, iethh->h_dest);
 788         oethh->h_proto = htons(ife->eth_type);
 789
 790         if (!(at & AT_EGRESS))
 791                 skb_pull(skb, skb->dev->hard_header_len);
 792
 793         spin_unlock(&ife->tcf_lock);
 794
 795         return action;
 796 }
 797
 798 static int tcf_ife_act(struct sk_buff *skb, const struct tc_action *a,
 799                        struct tcf_result *res)
 800 {
 801         struct tcf_ife_info *ife = a->priv;
 802
 803         if (ife->flags & IFE_ENCODE)
 804                 return tcf_ife_encode(skb, a, res);
 805
 806         if (!(ife->flags & IFE_ENCODE))
 807                 return tcf_ife_decode(skb, a, res);
 808
 809         pr_info_ratelimited("unknown failure(policy neither de/encode\n");
 810         spin_lock(&ife->tcf_lock);
 811         bstats_update(&ife->tcf_bstats, skb);
 812         ife->tcf_tm.lastuse = jiffies;
 813         ife->tcf_qstats.drops++;
 814         spin_unlock(&ife->tcf_lock);
 815
 816         return TC_ACT_SHOT;
 817 }
 818
 819 static int tcf_ife_walker(struct net *net, struct sk_buff *skb,
 820                           struct netlink_callback *cb, int type,
 821                           struct tc_action *a)
 822 {
 823         struct tc_action_net *tn = net_generic(net, ife_net_id);
 824
 825         return tcf_generic_walker(tn, skb, cb, type, a);
 826 }
 827
 828 static int tcf_ife_search(struct net *net, struct tc_action *a, u32 index)
 829 {
 830         struct tc_action_net *tn = net_generic(net, ife_net_id);
 831
 832         return tcf_hash_search(tn, a, index);
 833 }
 834
 835 static struct tc_action_ops act_ife_ops = {
 836         .kind = "ife",
 837         .type = TCA_ACT_IFE,
 838         .owner = THIS_MODULE,
 839         .act = tcf_ife_act,
 840         .dump = tcf_ife_dump,
 841         .cleanup = tcf_ife_cleanup,
 842         .init = tcf_ife_init,
 843         .walk = tcf_ife_walker,
 844         .lookup = tcf_ife_search,
 845 };
 846
 847 static __net_init int ife_init_net(struct net *net)
 848 {
 849         struct tc_action_net *tn = net_generic(net, ife_net_id);
 850
 851         return tc_action_net_init(tn, &act_ife_ops, IFE_TAB_MASK);
 852 }
 853
 854 static void __net_exit ife_exit_net(struct net *net)
 855 {
 856         struct tc_action_net *tn = net_generic(net, ife_net_id);
 857
 858         tc_action_net_exit(tn);
 859 }
 860
 861 static struct pernet_operations ife_net_ops = {
 862         .init = ife_init_net,
 863         .exit = ife_exit_net,
 864         .id   = &ife_net_id,
 865         .size = sizeof(struct tc_action_net),
 866 };
 867
 868 static int __init ife_init_module(void)
 869 {
 870         return tcf_register_action(&act_ife_ops, &ife_net_ops);
 871 }
 872
 873 static void __exit ife_cleanup_module(void)
 874 {
 875         tcf_unregister_action(&act_ife_ops, &ife_net_ops);
 876 }
 877
 878 module_init(ife_init_module);
 879 module_exit(ife_cleanup_module);
 880
 881 MODULE_AUTHOR("Jamal Hadi Salim(2015)");
 882 MODULE_DESCRIPTION("Inter-FE LFB action");
 883 MODULE_LICENSE("GPL");