net/sched/cls_api.c

   1 /*
   2  * net/sched/cls_api.c  Packet classifier API.
   3  *
   4  *              This program is free software; you can redistribute it and/or
   5  *              modify it under the terms of the GNU General Public License
   6  *              as published by the Free Software Foundation; either version
   7  *              2 of the License, or (at your option) any later version.
   8  *
   9  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  10  *
  11  * Changes:
  12  *
  13  * Eduardo J. Blanco <ejbs@netlabs.com.uy> :990222: kmod support
  14  *
  15  */
  16
  17 #include <linux/module.h>
  18 #include <linux/types.h>
  19 #include <linux/kernel.h>
  20 #include <linux/string.h>
  21 #include <linux/errno.h>
  22 #include <linux/err.h>
  23 #include <linux/skbuff.h>
  24 #include <linux/init.h>
  25 #include <linux/kmod.h>
  26 #include <linux/err.h>
  27 #include <linux/slab.h>
  28 #include <net/net_namespace.h>
  29 #include <net/sock.h>
  30 #include <net/netlink.h>
  31 #include <net/pkt_sched.h>
  32 #include <net/pkt_cls.h>
  33
  34 /* The list of all installed classifier types */
  35 static LIST_HEAD(tcf_proto_base);
  36
  37 /* Protects list of registered TC modules. It is pure SMP lock. */
  38 static DEFINE_RWLOCK(cls_mod_lock);
  39
  40 /* Find classifier type by string name */
  41
  42 static const struct tcf_proto_ops *tcf_proto_lookup_ops(const char *kind)
  43 {
  44         const struct tcf_proto_ops *t, *res = NULL;
  45
  46         if (kind) {
  47                 read_lock(&cls_mod_lock);
  48                 list_for_each_entry(t, &tcf_proto_base, head) {
  49                         if (strcmp(kind, t->kind) == 0) {
  50                                 if (try_module_get(t->owner))
  51                                         res = t;
  52                                 break;
  53                         }
  54                 }
  55                 read_unlock(&cls_mod_lock);
  56         }
  57         return res;
  58 }
  59
  60 /* Register(unregister) new classifier type */
  61
  62 int register_tcf_proto_ops(struct tcf_proto_ops *ops)
  63 {
  64         struct tcf_proto_ops *t;
  65         int rc = -EEXIST;
  66
  67         write_lock(&cls_mod_lock);
  68         list_for_each_entry(t, &tcf_proto_base, head)
  69                 if (!strcmp(ops->kind, t->kind))
  70                         goto out;
  71
  72         list_add_tail(&ops->head, &tcf_proto_base);
  73         rc = 0;
  74 out:
  75         write_unlock(&cls_mod_lock);
  76         return rc;
  77 }
  78 EXPORT_SYMBOL(register_tcf_proto_ops);
  79
  80 int unregister_tcf_proto_ops(struct tcf_proto_ops *ops)
  81 {
  82         struct tcf_proto_ops *t;
  83         int rc = -ENOENT;
  84
  85         /* Wait for outstanding call_rcu()s, if any, from a
  86          * tcf_proto_ops's destroy() handler.
  87          */
  88         rcu_barrier();
  89
  90         write_lock(&cls_mod_lock);
  91         list_for_each_entry(t, &tcf_proto_base, head) {
  92                 if (t == ops) {
  93                         list_del(&t->head);
  94                         rc = 0;
  95                         break;
  96                 }
  97         }
  98         write_unlock(&cls_mod_lock);
  99         return rc;
 100 }
 101 EXPORT_SYMBOL(unregister_tcf_proto_ops);
 102
 103 static int tfilter_notify(struct net *net, struct sk_buff *oskb,
 104                           struct nlmsghdr *n, struct tcf_proto *tp,
 105                           unsigned long fh, int event, bool unicast);
 106
 107 static void tfilter_notify_chain(struct net *net, struct sk_buff *oskb,
 108                                  struct nlmsghdr *n,
 109                                  struct tcf_chain *chain, int event)
 110 {
 111         struct tcf_proto *tp;
 112
 113         for (tp = rtnl_dereference(chain->filter_chain);
 114              tp; tp = rtnl_dereference(tp->next))
 115                 tfilter_notify(net, oskb, n, tp, 0, event, false);
 116 }
 117
 118 /* Select new prio value from the range, managed by kernel. */
 119
 120 static inline u32 tcf_auto_prio(struct tcf_proto *tp)
 121 {
 122         u32 first = TC_H_MAKE(0xC0000000U, 0U);
 123
 124         if (tp)
 125                 first = tp->prio - 1;
 126
 127         return TC_H_MAJ(first);
 128 }
 129
 130 static struct tcf_proto *tcf_proto_create(const char *kind, u32 protocol,
 131                                           u32 prio, u32 parent, struct Qdisc *q,
 132                                           struct tcf_chain *chain)
 133 {
 134         struct tcf_proto *tp;
 135         int err;
 136
 137         tp = kzalloc(sizeof(*tp), GFP_KERNEL);
 138         if (!tp)
 139                 return ERR_PTR(-ENOBUFS);
 140
 141         err = -ENOENT;
 142         tp->ops = tcf_proto_lookup_ops(kind);
 143         if (!tp->ops) {
 144 #ifdef CONFIG_MODULES
 145                 rtnl_unlock();
 146                 request_module("cls_%s", kind);
 147                 rtnl_lock();
 148                 tp->ops = tcf_proto_lookup_ops(kind);
 149                 /* We dropped the RTNL semaphore in order to perform
 150                  * the module load. So, even if we succeeded in loading
 151                  * the module we have to replay the request. We indicate
 152                  * this using -EAGAIN.
 153                  */
 154                 if (tp->ops) {
 155                         module_put(tp->ops->owner);
 156                         err = -EAGAIN;
 157                 } else {
 158                         err = -ENOENT;
 159                 }
 160                 goto errout;
 161 #endif
 162         }
 163         tp->classify = tp->ops->classify;
 164         tp->protocol = protocol;
 165         tp->prio = prio;
 166         tp->classid = parent;
 167         tp->q = q;
 168         tp->chain = chain;
 169
 170         err = tp->ops->init(tp);
 171         if (err) {
 172                 module_put(tp->ops->owner);
 173                 goto errout;
 174         }
 175         return tp;
 176
 177 errout:
 178         kfree(tp);
 179         return ERR_PTR(err);
 180 }
 181
 182 static void tcf_proto_destroy(struct tcf_proto *tp)
 183 {
 184         tp->ops->destroy(tp);
 185         module_put(tp->ops->owner);
 186         kfree_rcu(tp, rcu);
 187 }
 188
 189 static struct tcf_chain *tcf_chain_create(struct tcf_block *block,
 190                                           u32 chain_index)
 191 {
 192         struct tcf_chain *chain;
 193
 194         chain = kzalloc(sizeof(*chain), GFP_KERNEL);
 195         if (!chain)
 196                 return NULL;
 197         list_add_tail(&chain->list, &block->chain_list);
 198         chain->block = block;
 199         chain->index = chain_index;
 200         chain->refcnt = 1;
 201         return chain;
 202 }
 203
 204 static void tcf_chain_flush(struct tcf_chain *chain)
 205 {
 206         struct tcf_proto *tp;
 207
 208         if (chain->p_filter_chain)
 209                 RCU_INIT_POINTER(*chain->p_filter_chain, NULL);
 210         while ((tp = rtnl_dereference(chain->filter_chain)) != NULL) {
 211                 RCU_INIT_POINTER(chain->filter_chain, tp->next);
 212                 tcf_proto_destroy(tp);
 213         }
 214 }
 215
 216 static void tcf_chain_destroy(struct tcf_chain *chain)
 217 {
 218         /* May be already removed from the list by the previous call. */
 219         if (!list_empty(&chain->list))
 220                 list_del_init(&chain->list);
 221
 222         tcf_chain_flush(chain);
 223
 224         /* There might still be a reference held when we got here from
 225          * tcf_block_put. Wait for the user to drop reference before free.
 226          */
 227         if (!chain->refcnt)
 228                 kfree(chain);
 229 }
 230
 231 struct tcf_chain *tcf_chain_get(struct tcf_block *block, u32 chain_index,
 232                                 bool create)
 233 {
 234         struct tcf_chain *chain;
 235
 236         list_for_each_entry(chain, &block->chain_list, list) {
 237                 if (chain->index == chain_index) {
 238                         chain->refcnt++;
 239                         return chain;
 240                 }
 241         }
 242         if (create)
 243                 return tcf_chain_create(block, chain_index);
 244         else
 245                 return NULL;
 246 }
 247 EXPORT_SYMBOL(tcf_chain_get);
 248
 249 void tcf_chain_put(struct tcf_chain *chain)
 250 {
 251         /* Destroy unused chain, with exception of chain 0, which is the
 252          * default one and has to be always present.
 253          */
 254         if (--chain->refcnt == 0 && !chain->filter_chain && chain->index != 0)
 255                 tcf_chain_destroy(chain);
 256 }
 257 EXPORT_SYMBOL(tcf_chain_put);
 258
 259 static void
 260 tcf_chain_filter_chain_ptr_set(struct tcf_chain *chain,
 261                                struct tcf_proto __rcu **p_filter_chain)
 262 {
 263         chain->p_filter_chain = p_filter_chain;
 264 }
 265
 266 int tcf_block_get(struct tcf_block **p_block,
 267                   struct tcf_proto __rcu **p_filter_chain)
 268 {
 269         struct tcf_block *block = kzalloc(sizeof(*block), GFP_KERNEL);
 270         struct tcf_chain *chain;
 271         int err;
 272
 273         if (!block)
 274                 return -ENOMEM;
 275         INIT_LIST_HEAD(&block->chain_list);
 276         /* Create chain 0 by default, it has to be always present. */
 277         chain = tcf_chain_create(block, 0);
 278         if (!chain) {
 279                 err = -ENOMEM;
 280                 goto err_chain_create;
 281         }
 282         tcf_chain_filter_chain_ptr_set(chain, p_filter_chain);
 283         *p_block = block;
 284         return 0;
 285
 286 err_chain_create:
 287         kfree(block);
 288         return err;
 289 }
 290 EXPORT_SYMBOL(tcf_block_get);
 291
 292 void tcf_block_put(struct tcf_block *block)
 293 {
 294         struct tcf_chain *chain, *tmp;
 295
 296         if (!block)
 297                 return;
 298
 299         list_for_each_entry_safe(chain, tmp, &block->chain_list, list)
 300                 tcf_chain_destroy(chain);
 301         kfree(block);
 302 }
 303 EXPORT_SYMBOL(tcf_block_put);
 304
 305 /* Main classifier routine: scans classifier chain attached
 306  * to this qdisc, (optionally) tests for protocol and asks
 307  * specific classifiers.
 308  */
 309 int tcf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 310                  struct tcf_result *res, bool compat_mode)
 311 {
 312         __be16 protocol = tc_skb_protocol(skb);
 313 #ifdef CONFIG_NET_CLS_ACT
 314         const int max_reclassify_loop = 4;
 315         const struct tcf_proto *orig_tp = tp;
 316         const struct tcf_proto *first_tp;
 317         int limit = 0;
 318
 319 reclassify:
 320 #endif
 321         for (; tp; tp = rcu_dereference_bh(tp->next)) {
 322                 int err;
 323
 324                 if (tp->protocol != protocol &&
 325                     tp->protocol != htons(ETH_P_ALL))
 326                         continue;
 327
 328                 err = tp->classify(skb, tp, res);
 329 #ifdef CONFIG_NET_CLS_ACT
 330                 if (unlikely(err == TC_ACT_RECLASSIFY && !compat_mode)) {
 331                         first_tp = orig_tp;
 332                         goto reset;
 333                 } else if (unlikely(TC_ACT_EXT_CMP(err, TC_ACT_GOTO_CHAIN))) {
 334                         first_tp = res->goto_tp;
 335                         goto reset;
 336                 }
 337 #endif
 338                 if (err >= 0)
 339                         return err;
 340         }
 341
 342         return TC_ACT_UNSPEC; /* signal: continue lookup */
 343 #ifdef CONFIG_NET_CLS_ACT
 344 reset:
 345         if (unlikely(limit++ >= max_reclassify_loop)) {
 346                 net_notice_ratelimited("%s: reclassify loop, rule prio %u, protocol %02x\n",
 347                                        tp->q->ops->id, tp->prio & 0xffff,
 348                                        ntohs(tp->protocol));
 349                 return TC_ACT_SHOT;
 350         }
 351
 352         tp = first_tp;
 353         protocol = tc_skb_protocol(skb);
 354         goto reclassify;
 355 #endif
 356 }
 357 EXPORT_SYMBOL(tcf_classify);
 358
 359 struct tcf_chain_info {
 360         struct tcf_proto __rcu **pprev;
 361         struct tcf_proto __rcu *next;
 362 };
 363
 364 static struct tcf_proto *tcf_chain_tp_prev(struct tcf_chain_info *chain_info)
 365 {
 366         return rtnl_dereference(*chain_info->pprev);
 367 }
 368
 369 static void tcf_chain_tp_insert(struct tcf_chain *chain,
 370                                 struct tcf_chain_info *chain_info,
 371                                 struct tcf_proto *tp)
 372 {
 373         if (chain->p_filter_chain &&
 374             *chain_info->pprev == chain->filter_chain)
 375                 rcu_assign_pointer(*chain->p_filter_chain, tp);
 376         RCU_INIT_POINTER(tp->next, tcf_chain_tp_prev(chain_info));
 377         rcu_assign_pointer(*chain_info->pprev, tp);
 378 }
 379
 380 static void tcf_chain_tp_remove(struct tcf_chain *chain,
 381                                 struct tcf_chain_info *chain_info,
 382                                 struct tcf_proto *tp)
 383 {
 384         struct tcf_proto *next = rtnl_dereference(chain_info->next);
 385
 386         if (chain->p_filter_chain && tp == chain->filter_chain)
 387                 RCU_INIT_POINTER(*chain->p_filter_chain, next);
 388         RCU_INIT_POINTER(*chain_info->pprev, next);
 389 }
 390
 391 static struct tcf_proto *tcf_chain_tp_find(struct tcf_chain *chain,
 392                                            struct tcf_chain_info *chain_info,
 393                                            u32 protocol, u32 prio,
 394                                            bool prio_allocate)
 395 {
 396         struct tcf_proto **pprev;
 397         struct tcf_proto *tp;
 398
 399         /* Check the chain for existence of proto-tcf with this priority */
 400         for (pprev = &chain->filter_chain;
 401              (tp = rtnl_dereference(*pprev)); pprev = &tp->next) {
 402                 if (tp->prio >= prio) {
 403                         if (tp->prio == prio) {
 404                                 if (prio_allocate ||
 405                                     (tp->protocol != protocol && protocol))
 406                                         return ERR_PTR(-EINVAL);
 407                         } else {
 408                                 tp = NULL;
 409                         }
 410                         break;
 411                 }
 412         }
 413         chain_info->pprev = pprev;
 414         chain_info->next = tp ? tp->next : NULL;
 415         return tp;
 416 }
 417
 418 /* Add/change/delete/get a filter node */
 419
 420 static int tc_ctl_tfilter(struct sk_buff *skb, struct nlmsghdr *n,
 421                           struct netlink_ext_ack *extack)
 422 {
 423         struct net *net = sock_net(skb->sk);
 424         struct nlattr *tca[TCA_MAX + 1];
 425         struct tcmsg *t;
 426         u32 protocol;
 427         u32 prio;
 428         bool prio_allocate;
 429         u32 parent;
 430         u32 chain_index;
 431         struct net_device *dev;
 432         struct Qdisc  *q;
 433         struct tcf_chain_info chain_info;
 434         struct tcf_chain *chain = NULL;
 435         struct tcf_block *block;
 436         struct tcf_proto *tp;
 437         const struct Qdisc_class_ops *cops;
 438         unsigned long cl;
 439         unsigned long fh;
 440         int err;
 441         int tp_created;
 442
 443         if ((n->nlmsg_type != RTM_GETTFILTER) &&
 444             !netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN))
 445                 return -EPERM;
 446
 447 replay:
 448         tp_created = 0;
 449
 450         err = nlmsg_parse(n, sizeof(*t), tca, TCA_MAX, NULL, extack);
 451         if (err < 0)
 452                 return err;
 453
 454         t = nlmsg_data(n);
 455         protocol = TC_H_MIN(t->tcm_info);
 456         prio = TC_H_MAJ(t->tcm_info);
 457         prio_allocate = false;
 458         parent = t->tcm_parent;
 459         cl = 0;
 460
 461         if (prio == 0) {
 462                 switch (n->nlmsg_type) {
 463                 case RTM_DELTFILTER:
 464                         if (protocol || t->tcm_handle || tca[TCA_KIND])
 465                                 return -ENOENT;
 466                         break;
 467                 case RTM_NEWTFILTER:
 468                         /* If no priority is provided by the user,
 469                          * we allocate one.
 470                          */
 471                         if (n->nlmsg_flags & NLM_F_CREATE) {
 472                                 prio = TC_H_MAKE(0x80000000U, 0U);
 473                                 prio_allocate = true;
 474                                 break;
 475                         }
 476                         /* fall-through */
 477                 default:
 478                         return -ENOENT;
 479                 }
 480         }
 481
 482         /* Find head of filter chain. */
 483
 484         /* Find link */
 485         dev = __dev_get_by_index(net, t->tcm_ifindex);
 486         if (dev == NULL)
 487                 return -ENODEV;
 488
 489         /* Find qdisc */
 490         if (!parent) {
 491                 q = dev->qdisc;
 492                 parent = q->handle;
 493         } else {
 494                 q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent));
 495                 if (q == NULL)
 496                         return -EINVAL;
 497         }
 498
 499         /* Is it classful? */
 500         cops = q->ops->cl_ops;
 501         if (!cops)
 502                 return -EINVAL;
 503
 504         if (!cops->tcf_block)
 505                 return -EOPNOTSUPP;
 506
 507         /* Do we search for filter, attached to class? */
 508         if (TC_H_MIN(parent)) {
 509                 cl = cops->get(q, parent);
 510                 if (cl == 0)
 511                         return -ENOENT;
 512         }
 513
 514         /* And the last stroke */
 515         block = cops->tcf_block(q, cl);
 516         if (!block) {
 517                 err = -EINVAL;
 518                 goto errout;
 519         }
 520
 521         chain_index = tca[TCA_CHAIN] ? nla_get_u32(tca[TCA_CHAIN]) : 0;
 522         if (chain_index > TC_ACT_EXT_VAL_MASK) {
 523                 err = -EINVAL;
 524                 goto errout;
 525         }
 526         chain = tcf_chain_get(block, chain_index,
 527                               n->nlmsg_type == RTM_NEWTFILTER);
 528         if (!chain) {
 529                 err = n->nlmsg_type == RTM_NEWTFILTER ? -ENOMEM : -EINVAL;
 530                 goto errout;
 531         }
 532
 533         if (n->nlmsg_type == RTM_DELTFILTER && prio == 0) {
 534                 tfilter_notify_chain(net, skb, n, chain, RTM_DELTFILTER);
 535                 tcf_chain_flush(chain);
 536                 err = 0;
 537                 goto errout;
 538         }
 539
 540         tp = tcf_chain_tp_find(chain, &chain_info, protocol,
 541                                prio, prio_allocate);
 542         if (IS_ERR(tp)) {
 543                 err = PTR_ERR(tp);
 544                 goto errout;
 545         }
 546
 547         if (tp == NULL) {
 548                 /* Proto-tcf does not exist, create new one */
 549
 550                 if (tca[TCA_KIND] == NULL || !protocol) {
 551                         err = -EINVAL;
 552                         goto errout;
 553                 }
 554
 555                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 556                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 557                         err = -ENOENT;
 558                         goto errout;
 559                 }
 560
 561                 if (prio_allocate)
 562                         prio = tcf_auto_prio(tcf_chain_tp_prev(&chain_info));
 563
 564                 tp = tcf_proto_create(nla_data(tca[TCA_KIND]),
 565                                       protocol, prio, parent, q, chain);
 566                 if (IS_ERR(tp)) {
 567                         err = PTR_ERR(tp);
 568                         goto errout;
 569                 }
 570                 tp_created = 1;
 571         } else if (tca[TCA_KIND] && nla_strcmp(tca[TCA_KIND], tp->ops->kind)) {
 572                 err = -EINVAL;
 573                 goto errout;
 574         }
 575
 576         fh = tp->ops->get(tp, t->tcm_handle);
 577
 578         if (fh == 0) {
 579                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 580                         tcf_chain_tp_remove(chain, &chain_info, tp);
 581                         tfilter_notify(net, skb, n, tp, fh,
 582                                        RTM_DELTFILTER, false);
 583                         tcf_proto_destroy(tp);
 584                         err = 0;
 585                         goto errout;
 586                 }
 587
 588                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 589                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 590                         err = -ENOENT;
 591                         goto errout;
 592                 }
 593         } else {
 594                 bool last;
 595
 596                 switch (n->nlmsg_type) {
 597                 case RTM_NEWTFILTER:
 598                         if (n->nlmsg_flags & NLM_F_EXCL) {
 599                                 if (tp_created)
 600                                         tcf_proto_destroy(tp);
 601                                 err = -EEXIST;
 602                                 goto errout;
 603                         }
 604                         break;
 605                 case RTM_DELTFILTER:
 606                         err = tp->ops->delete(tp, fh, &last);
 607                         if (err)
 608                                 goto errout;
 609                         tfilter_notify(net, skb, n, tp, t->tcm_handle,
 610                                        RTM_DELTFILTER, false);
 611                         if (last) {
 612                                 tcf_chain_tp_remove(chain, &chain_info, tp);
 613                                 tcf_proto_destroy(tp);
 614                         }
 615                         goto errout;
 616                 case RTM_GETTFILTER:
 617                         err = tfilter_notify(net, skb, n, tp, fh,
 618                                              RTM_NEWTFILTER, true);
 619                         goto errout;
 620                 default:
 621                         err = -EINVAL;
 622                         goto errout;
 623                 }
 624         }
 625
 626         err = tp->ops->change(net, skb, tp, cl, t->tcm_handle, tca, &fh,
 627                               n->nlmsg_flags & NLM_F_CREATE ? TCA_ACT_NOREPLACE : TCA_ACT_REPLACE);
 628         if (err == 0) {
 629                 if (tp_created)
 630                         tcf_chain_tp_insert(chain, &chain_info, tp);
 631                 tfilter_notify(net, skb, n, tp, fh, RTM_NEWTFILTER, false);
 632         } else {
 633                 if (tp_created)
 634                         tcf_proto_destroy(tp);
 635         }
 636
 637 errout:
 638         if (chain)
 639                 tcf_chain_put(chain);
 640         if (cl)
 641                 cops->put(q, cl);
 642         if (err == -EAGAIN)
 643                 /* Replay the request. */
 644                 goto replay;
 645         return err;
 646 }
 647
 648 static int tcf_fill_node(struct net *net, struct sk_buff *skb,
 649                          struct tcf_proto *tp, unsigned long fh, u32 portid,
 650                          u32 seq, u16 flags, int event)
 651 {
 652         struct tcmsg *tcm;
 653         struct nlmsghdr  *nlh;
 654         unsigned char *b = skb_tail_pointer(skb);
 655
 656         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*tcm), flags);
 657         if (!nlh)
 658                 goto out_nlmsg_trim;
 659         tcm = nlmsg_data(nlh);
 660         tcm->tcm_family = AF_UNSPEC;
 661         tcm->tcm__pad1 = 0;
 662         tcm->tcm__pad2 = 0;
 663         tcm->tcm_ifindex = qdisc_dev(tp->q)->ifindex;
 664         tcm->tcm_parent = tp->classid;
 665         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 666         if (nla_put_string(skb, TCA_KIND, tp->ops->kind))
 667                 goto nla_put_failure;
 668         if (nla_put_u32(skb, TCA_CHAIN, tp->chain->index))
 669                 goto nla_put_failure;
 670         tcm->tcm_handle = fh;
 671         if (RTM_DELTFILTER != event) {
 672                 tcm->tcm_handle = 0;
 673                 if (tp->ops->dump && tp->ops->dump(net, tp, fh, skb, tcm) < 0)
 674                         goto nla_put_failure;
 675         }
 676         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 677         return skb->len;
 678
 679 out_nlmsg_trim:
 680 nla_put_failure:
 681         nlmsg_trim(skb, b);
 682         return -1;
 683 }
 684
 685 static int tfilter_notify(struct net *net, struct sk_buff *oskb,
 686                           struct nlmsghdr *n, struct tcf_proto *tp,
 687                           unsigned long fh, int event, bool unicast)
 688 {
 689         struct sk_buff *skb;
 690         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 691
 692         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 693         if (!skb)
 694                 return -ENOBUFS;
 695
 696         if (tcf_fill_node(net, skb, tp, fh, portid, n->nlmsg_seq,
 697                           n->nlmsg_flags, event) <= 0) {
 698                 kfree_skb(skb);
 699                 return -EINVAL;
 700         }
 701
 702         if (unicast)
 703                 return netlink_unicast(net->rtnl, skb, portid, MSG_DONTWAIT);
 704
 705         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 706                               n->nlmsg_flags & NLM_F_ECHO);
 707 }
 708
 709 struct tcf_dump_args {
 710         struct tcf_walker w;
 711         struct sk_buff *skb;
 712         struct netlink_callback *cb;
 713 };
 714
 715 static int tcf_node_dump(struct tcf_proto *tp, unsigned long n,
 716                          struct tcf_walker *arg)
 717 {
 718         struct tcf_dump_args *a = (void *)arg;
 719         struct net *net = sock_net(a->skb->sk);
 720
 721         return tcf_fill_node(net, a->skb, tp, n, NETLINK_CB(a->cb->skb).portid,
 722                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI,
 723                              RTM_NEWTFILTER);
 724 }
 725
 726 static bool tcf_chain_dump(struct tcf_chain *chain, struct sk_buff *skb,
 727                            struct netlink_callback *cb,
 728                            long index_start, long *p_index)
 729 {
 730         struct net *net = sock_net(skb->sk);
 731         struct tcmsg *tcm = nlmsg_data(cb->nlh);
 732         struct tcf_dump_args arg;
 733         struct tcf_proto *tp;
 734
 735         for (tp = rtnl_dereference(chain->filter_chain);
 736              tp; tp = rtnl_dereference(tp->next), (*p_index)++) {
 737                 if (*p_index < index_start)
 738                         continue;
 739                 if (TC_H_MAJ(tcm->tcm_info) &&
 740                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 741                         continue;
 742                 if (TC_H_MIN(tcm->tcm_info) &&
 743                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 744                         continue;
 745                 if (*p_index > index_start)
 746                         memset(&cb->args[1], 0,
 747                                sizeof(cb->args) - sizeof(cb->args[0]));
 748                 if (cb->args[1] == 0) {
 749                         if (tcf_fill_node(net, skb, tp, 0,
 750                                           NETLINK_CB(cb->skb).portid,
 751                                           cb->nlh->nlmsg_seq, NLM_F_MULTI,
 752                                           RTM_NEWTFILTER) <= 0)
 753                                 return false;
 754
 755                         cb->args[1] = 1;
 756                 }
 757                 if (!tp->ops->walk)
 758                         continue;
 759                 arg.w.fn = tcf_node_dump;
 760                 arg.skb = skb;
 761                 arg.cb = cb;
 762                 arg.w.stop = 0;
 763                 arg.w.skip = cb->args[1] - 1;
 764                 arg.w.count = 0;
 765                 tp->ops->walk(tp, &arg.w);
 766                 cb->args[1] = arg.w.count + 1;
 767                 if (arg.w.stop)
 768                         return false;
 769         }
 770         return true;
 771 }
 772
 773 /* called with RTNL */
 774 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
 775 {
 776         struct net *net = sock_net(skb->sk);
 777         struct nlattr *tca[TCA_MAX + 1];
 778         struct net_device *dev;
 779         struct Qdisc *q;
 780         struct tcf_block *block;
 781         struct tcf_chain *chain;
 782         struct tcmsg *tcm = nlmsg_data(cb->nlh);
 783         unsigned long cl = 0;
 784         const struct Qdisc_class_ops *cops;
 785         long index_start;
 786         long index;
 787         int err;
 788
 789         if (nlmsg_len(cb->nlh) < sizeof(*tcm))
 790                 return skb->len;
 791
 792         err = nlmsg_parse(cb->nlh, sizeof(*tcm), tca, TCA_MAX, NULL, NULL);
 793         if (err)
 794                 return err;
 795
 796         dev = __dev_get_by_index(net, tcm->tcm_ifindex);
 797         if (!dev)
 798                 return skb->len;
 799
 800         if (!tcm->tcm_parent)
 801                 q = dev->qdisc;
 802         else
 803                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
 804         if (!q)
 805                 goto out;
 806         cops = q->ops->cl_ops;
 807         if (!cops)
 808                 goto errout;
 809         if (!cops->tcf_block)
 810                 goto errout;
 811         if (TC_H_MIN(tcm->tcm_parent)) {
 812                 cl = cops->get(q, tcm->tcm_parent);
 813                 if (cl == 0)
 814                         goto errout;
 815         }
 816         block = cops->tcf_block(q, cl);
 817         if (!block)
 818                 goto errout;
 819
 820         index_start = cb->args[0];
 821         index = 0;
 822
 823         list_for_each_entry(chain, &block->chain_list, list) {
 824                 if (tca[TCA_CHAIN] &&
 825                     nla_get_u32(tca[TCA_CHAIN]) != chain->index)
 826                         continue;
 827                 if (!tcf_chain_dump(chain, skb, cb, index_start, &index))
 828                         break;
 829         }
 830
 831         cb->args[0] = index;
 832
 833 errout:
 834         if (cl)
 835                 cops->put(q, cl);
 836 out:
 837         return skb->len;
 838 }
 839
 840 void tcf_exts_destroy(struct tcf_exts *exts)
 841 {
 842 #ifdef CONFIG_NET_CLS_ACT
 843         LIST_HEAD(actions);
 844
 845         tcf_exts_to_list(exts, &actions);
 846         tcf_action_destroy(&actions, TCA_ACT_UNBIND);
 847         kfree(exts->actions);
 848         exts->nr_actions = 0;
 849 #endif
 850 }
 851 EXPORT_SYMBOL(tcf_exts_destroy);
 852
 853 int tcf_exts_validate(struct net *net, struct tcf_proto *tp, struct nlattr **tb,
 854                       struct nlattr *rate_tlv, struct tcf_exts *exts, bool ovr)
 855 {
 856 #ifdef CONFIG_NET_CLS_ACT
 857         {
 858                 struct tc_action *act;
 859
 860                 if (exts->police && tb[exts->police]) {
 861                         act = tcf_action_init_1(net, tp, tb[exts->police],
 862                                                 rate_tlv, "police", ovr,
 863                                                 TCA_ACT_BIND);
 864                         if (IS_ERR(act))
 865                                 return PTR_ERR(act);
 866
 867                         act->type = exts->type = TCA_OLD_COMPAT;
 868                         exts->actions[0] = act;
 869                         exts->nr_actions = 1;
 870                 } else if (exts->action && tb[exts->action]) {
 871                         LIST_HEAD(actions);
 872                         int err, i = 0;
 873
 874                         err = tcf_action_init(net, tp, tb[exts->action],
 875                                               rate_tlv, NULL, ovr, TCA_ACT_BIND,
 876                                               &actions);
 877                         if (err)
 878                                 return err;
 879                         list_for_each_entry(act, &actions, list)
 880                                 exts->actions[i++] = act;
 881                         exts->nr_actions = i;
 882                 }
 883         }
 884 #else
 885         if ((exts->action && tb[exts->action]) ||
 886             (exts->police && tb[exts->police]))
 887                 return -EOPNOTSUPP;
 888 #endif
 889
 890         return 0;
 891 }
 892 EXPORT_SYMBOL(tcf_exts_validate);
 893
 894 void tcf_exts_change(struct tcf_proto *tp, struct tcf_exts *dst,
 895                      struct tcf_exts *src)
 896 {
 897 #ifdef CONFIG_NET_CLS_ACT
 898         struct tcf_exts old = *dst;
 899
 900         tcf_tree_lock(tp);
 901         dst->nr_actions = src->nr_actions;
 902         dst->actions = src->actions;
 903         dst->type = src->type;
 904         tcf_tree_unlock(tp);
 905
 906         tcf_exts_destroy(&old);
 907 #endif
 908 }
 909 EXPORT_SYMBOL(tcf_exts_change);
 910
 911 #ifdef CONFIG_NET_CLS_ACT
 912 static struct tc_action *tcf_exts_first_act(struct tcf_exts *exts)
 913 {
 914         if (exts->nr_actions == 0)
 915                 return NULL;
 916         else
 917                 return exts->actions[0];
 918 }
 919 #endif
 920
 921 int tcf_exts_dump(struct sk_buff *skb, struct tcf_exts *exts)
 922 {
 923 #ifdef CONFIG_NET_CLS_ACT
 924         struct nlattr *nest;
 925
 926         if (exts->action && exts->nr_actions) {
 927                 /*
 928                  * again for backward compatible mode - we want
 929                  * to work with both old and new modes of entering
 930                  * tc data even if iproute2  was newer - jhs
 931                  */
 932                 if (exts->type != TCA_OLD_COMPAT) {
 933                         LIST_HEAD(actions);
 934
 935                         nest = nla_nest_start(skb, exts->action);
 936                         if (nest == NULL)
 937                                 goto nla_put_failure;
 938
 939                         tcf_exts_to_list(exts, &actions);
 940                         if (tcf_action_dump(skb, &actions, 0, 0) < 0)
 941                                 goto nla_put_failure;
 942                         nla_nest_end(skb, nest);
 943                 } else if (exts->police) {
 944                         struct tc_action *act = tcf_exts_first_act(exts);
 945                         nest = nla_nest_start(skb, exts->police);
 946                         if (nest == NULL || !act)
 947                                 goto nla_put_failure;
 948                         if (tcf_action_dump_old(skb, act, 0, 0) < 0)
 949                                 goto nla_put_failure;
 950                         nla_nest_end(skb, nest);
 951                 }
 952         }
 953         return 0;
 954
 955 nla_put_failure:
 956         nla_nest_cancel(skb, nest);
 957         return -1;
 958 #else
 959         return 0;
 960 #endif
 961 }
 962 EXPORT_SYMBOL(tcf_exts_dump);
 963
 964
 965 int tcf_exts_dump_stats(struct sk_buff *skb, struct tcf_exts *exts)
 966 {
 967 #ifdef CONFIG_NET_CLS_ACT
 968         struct tc_action *a = tcf_exts_first_act(exts);
 969         if (a != NULL && tcf_action_copy_stats(skb, a, 1) < 0)
 970                 return -1;
 971 #endif
 972         return 0;
 973 }
 974 EXPORT_SYMBOL(tcf_exts_dump_stats);
 975
 976 int tcf_exts_get_dev(struct net_device *dev, struct tcf_exts *exts,
 977                      struct net_device **hw_dev)
 978 {
 979 #ifdef CONFIG_NET_CLS_ACT
 980         const struct tc_action *a;
 981         LIST_HEAD(actions);
 982
 983         if (tc_no_actions(exts))
 984                 return -EINVAL;
 985
 986         tcf_exts_to_list(exts, &actions);
 987         list_for_each_entry(a, &actions, list) {
 988                 if (a->ops->get_dev) {
 989                         a->ops->get_dev(a, dev_net(dev), hw_dev);
 990                         break;
 991                 }
 992         }
 993         if (*hw_dev)
 994                 return 0;
 995 #endif
 996         return -EOPNOTSUPP;
 997 }
 998 EXPORT_SYMBOL(tcf_exts_get_dev);
 999
1000 static int __init tc_filter_init(void)
1001 {
1002         rtnl_register(PF_UNSPEC, RTM_NEWTFILTER, tc_ctl_tfilter, NULL, NULL);
1003         rtnl_register(PF_UNSPEC, RTM_DELTFILTER, tc_ctl_tfilter, NULL, NULL);
1004         rtnl_register(PF_UNSPEC, RTM_GETTFILTER, tc_ctl_tfilter,
1005                       tc_dump_tfilter, NULL);
1006
1007         return 0;
1008 }
1009
1010 subsys_initcall(tc_filter_init);