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         list_del(&chain->list);
 219         tcf_chain_flush(chain);
 220         kfree(chain);
 221 }
 222
 223 struct tcf_chain *tcf_chain_get(struct tcf_block *block, u32 chain_index)
 224 {
 225         struct tcf_chain *chain;
 226
 227         list_for_each_entry(chain, &block->chain_list, list) {
 228                 if (chain->index == chain_index) {
 229                         chain->refcnt++;
 230                         return chain;
 231                 }
 232         }
 233         return tcf_chain_create(block, chain_index);
 234 }
 235 EXPORT_SYMBOL(tcf_chain_get);
 236
 237 void tcf_chain_put(struct tcf_chain *chain)
 238 {
 239         /* Destroy unused chain, with exception of chain 0, which is the
 240          * default one and has to be always present.
 241          */
 242         if (--chain->refcnt == 0 && !chain->filter_chain && chain->index != 0)
 243                 tcf_chain_destroy(chain);
 244 }
 245 EXPORT_SYMBOL(tcf_chain_put);
 246
 247 static void
 248 tcf_chain_filter_chain_ptr_set(struct tcf_chain *chain,
 249                                struct tcf_proto __rcu **p_filter_chain)
 250 {
 251         chain->p_filter_chain = p_filter_chain;
 252 }
 253
 254 int tcf_block_get(struct tcf_block **p_block,
 255                   struct tcf_proto __rcu **p_filter_chain)
 256 {
 257         struct tcf_block *block = kzalloc(sizeof(*block), GFP_KERNEL);
 258         struct tcf_chain *chain;
 259         int err;
 260
 261         if (!block)
 262                 return -ENOMEM;
 263         INIT_LIST_HEAD(&block->chain_list);
 264         /* Create chain 0 by default, it has to be always present. */
 265         chain = tcf_chain_create(block, 0);
 266         if (!chain) {
 267                 err = -ENOMEM;
 268                 goto err_chain_create;
 269         }
 270         tcf_chain_filter_chain_ptr_set(chain, p_filter_chain);
 271         *p_block = block;
 272         return 0;
 273
 274 err_chain_create:
 275         kfree(block);
 276         return err;
 277 }
 278 EXPORT_SYMBOL(tcf_block_get);
 279
 280 void tcf_block_put(struct tcf_block *block)
 281 {
 282         struct tcf_chain *chain, *tmp;
 283
 284         if (!block)
 285                 return;
 286
 287         list_for_each_entry_safe(chain, tmp, &block->chain_list, list)
 288                 tcf_chain_destroy(chain);
 289         kfree(block);
 290 }
 291 EXPORT_SYMBOL(tcf_block_put);
 292
 293 /* Main classifier routine: scans classifier chain attached
 294  * to this qdisc, (optionally) tests for protocol and asks
 295  * specific classifiers.
 296  */
 297 int tcf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 298                  struct tcf_result *res, bool compat_mode)
 299 {
 300         __be16 protocol = tc_skb_protocol(skb);
 301 #ifdef CONFIG_NET_CLS_ACT
 302         const int max_reclassify_loop = 4;
 303         const struct tcf_proto *orig_tp = tp;
 304         const struct tcf_proto *first_tp;
 305         int limit = 0;
 306
 307 reclassify:
 308 #endif
 309         for (; tp; tp = rcu_dereference_bh(tp->next)) {
 310                 int err;
 311
 312                 if (tp->protocol != protocol &&
 313                     tp->protocol != htons(ETH_P_ALL))
 314                         continue;
 315
 316                 err = tp->classify(skb, tp, res);
 317 #ifdef CONFIG_NET_CLS_ACT
 318                 if (unlikely(err == TC_ACT_RECLASSIFY && !compat_mode)) {
 319                         first_tp = orig_tp;
 320                         goto reset;
 321                 } else if (unlikely(TC_ACT_EXT_CMP(err, TC_ACT_GOTO_CHAIN))) {
 322                         first_tp = res->goto_tp;
 323                         goto reset;
 324                 }
 325 #endif
 326                 if (err >= 0)
 327                         return err;
 328         }
 329
 330         return TC_ACT_UNSPEC; /* signal: continue lookup */
 331 #ifdef CONFIG_NET_CLS_ACT
 332 reset:
 333         if (unlikely(limit++ >= max_reclassify_loop)) {
 334                 net_notice_ratelimited("%s: reclassify loop, rule prio %u, protocol %02x\n",
 335                                        tp->q->ops->id, tp->prio & 0xffff,
 336                                        ntohs(tp->protocol));
 337                 return TC_ACT_SHOT;
 338         }
 339
 340         tp = first_tp;
 341         protocol = tc_skb_protocol(skb);
 342         goto reclassify;
 343 #endif
 344 }
 345 EXPORT_SYMBOL(tcf_classify);
 346
 347 struct tcf_chain_info {
 348         struct tcf_proto __rcu **pprev;
 349         struct tcf_proto __rcu *next;
 350 };
 351
 352 static struct tcf_proto *tcf_chain_tp_prev(struct tcf_chain_info *chain_info)
 353 {
 354         return rtnl_dereference(*chain_info->pprev);
 355 }
 356
 357 static void tcf_chain_tp_insert(struct tcf_chain *chain,
 358                                 struct tcf_chain_info *chain_info,
 359                                 struct tcf_proto *tp)
 360 {
 361         if (chain->p_filter_chain &&
 362             *chain_info->pprev == chain->filter_chain)
 363                 rcu_assign_pointer(*chain->p_filter_chain, tp);
 364         RCU_INIT_POINTER(tp->next, tcf_chain_tp_prev(chain_info));
 365         rcu_assign_pointer(*chain_info->pprev, tp);
 366 }
 367
 368 static void tcf_chain_tp_remove(struct tcf_chain *chain,
 369                                 struct tcf_chain_info *chain_info,
 370                                 struct tcf_proto *tp)
 371 {
 372         struct tcf_proto *next = rtnl_dereference(chain_info->next);
 373
 374         if (chain->p_filter_chain && tp == chain->filter_chain)
 375                 RCU_INIT_POINTER(*chain->p_filter_chain, next);
 376         RCU_INIT_POINTER(*chain_info->pprev, next);
 377 }
 378
 379 static struct tcf_proto *tcf_chain_tp_find(struct tcf_chain *chain,
 380                                            struct tcf_chain_info *chain_info,
 381                                            u32 protocol, u32 prio,
 382                                            bool prio_allocate)
 383 {
 384         struct tcf_proto **pprev;
 385         struct tcf_proto *tp;
 386
 387         /* Check the chain for existence of proto-tcf with this priority */
 388         for (pprev = &chain->filter_chain;
 389              (tp = rtnl_dereference(*pprev)); pprev = &tp->next) {
 390                 if (tp->prio >= prio) {
 391                         if (tp->prio == prio) {
 392                                 if (prio_allocate ||
 393                                     (tp->protocol != protocol && protocol))
 394                                         return ERR_PTR(-EINVAL);
 395                         } else {
 396                                 tp = NULL;
 397                         }
 398                         break;
 399                 }
 400         }
 401         chain_info->pprev = pprev;
 402         chain_info->next = tp ? tp->next : NULL;
 403         return tp;
 404 }
 405
 406 /* Add/change/delete/get a filter node */
 407
 408 static int tc_ctl_tfilter(struct sk_buff *skb, struct nlmsghdr *n,
 409                           struct netlink_ext_ack *extack)
 410 {
 411         struct net *net = sock_net(skb->sk);
 412         struct nlattr *tca[TCA_MAX + 1];
 413         struct tcmsg *t;
 414         u32 protocol;
 415         u32 prio;
 416         bool prio_allocate;
 417         u32 parent;
 418         u32 chain_index;
 419         struct net_device *dev;
 420         struct Qdisc  *q;
 421         struct tcf_chain_info chain_info;
 422         struct tcf_chain *chain = NULL;
 423         struct tcf_block *block;
 424         struct tcf_proto *tp;
 425         const struct Qdisc_class_ops *cops;
 426         unsigned long cl;
 427         unsigned long fh;
 428         int err;
 429         int tp_created;
 430
 431         if ((n->nlmsg_type != RTM_GETTFILTER) &&
 432             !netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN))
 433                 return -EPERM;
 434
 435 replay:
 436         tp_created = 0;
 437
 438         err = nlmsg_parse(n, sizeof(*t), tca, TCA_MAX, NULL, extack);
 439         if (err < 0)
 440                 return err;
 441
 442         t = nlmsg_data(n);
 443         protocol = TC_H_MIN(t->tcm_info);
 444         prio = TC_H_MAJ(t->tcm_info);
 445         prio_allocate = false;
 446         parent = t->tcm_parent;
 447         cl = 0;
 448
 449         if (prio == 0) {
 450                 switch (n->nlmsg_type) {
 451                 case RTM_DELTFILTER:
 452                         if (protocol || t->tcm_handle || tca[TCA_KIND])
 453                                 return -ENOENT;
 454                         break;
 455                 case RTM_NEWTFILTER:
 456                         /* If no priority is provided by the user,
 457                          * we allocate one.
 458                          */
 459                         if (n->nlmsg_flags & NLM_F_CREATE) {
 460                                 prio = TC_H_MAKE(0x80000000U, 0U);
 461                                 prio_allocate = true;
 462                                 break;
 463                         }
 464                         /* fall-through */
 465                 default:
 466                         return -ENOENT;
 467                 }
 468         }
 469
 470         /* Find head of filter chain. */
 471
 472         /* Find link */
 473         dev = __dev_get_by_index(net, t->tcm_ifindex);
 474         if (dev == NULL)
 475                 return -ENODEV;
 476
 477         /* Find qdisc */
 478         if (!parent) {
 479                 q = dev->qdisc;
 480                 parent = q->handle;
 481         } else {
 482                 q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent));
 483                 if (q == NULL)
 484                         return -EINVAL;
 485         }
 486
 487         /* Is it classful? */
 488         cops = q->ops->cl_ops;
 489         if (!cops)
 490                 return -EINVAL;
 491
 492         if (!cops->tcf_block)
 493                 return -EOPNOTSUPP;
 494
 495         /* Do we search for filter, attached to class? */
 496         if (TC_H_MIN(parent)) {
 497                 cl = cops->get(q, parent);
 498                 if (cl == 0)
 499                         return -ENOENT;
 500         }
 501
 502         /* And the last stroke */
 503         block = cops->tcf_block(q, cl);
 504         if (!block) {
 505                 err = -EINVAL;
 506                 goto errout;
 507         }
 508
 509         chain_index = tca[TCA_CHAIN] ? nla_get_u32(tca[TCA_CHAIN]) : 0;
 510         if (chain_index > TC_ACT_EXT_VAL_MASK) {
 511                 err = -EINVAL;
 512                 goto errout;
 513         }
 514         chain = tcf_chain_get(block, chain_index);
 515         if (!chain) {
 516                 err = -ENOMEM;
 517                 goto errout;
 518         }
 519
 520         if (n->nlmsg_type == RTM_DELTFILTER && prio == 0) {
 521                 tfilter_notify_chain(net, skb, n, chain, RTM_DELTFILTER);
 522                 tcf_chain_flush(chain);
 523                 err = 0;
 524                 goto errout;
 525         }
 526
 527         tp = tcf_chain_tp_find(chain, &chain_info, protocol,
 528                                prio, prio_allocate);
 529         if (IS_ERR(tp)) {
 530                 err = PTR_ERR(tp);
 531                 goto errout;
 532         }
 533
 534         if (tp == NULL) {
 535                 /* Proto-tcf does not exist, create new one */
 536
 537                 if (tca[TCA_KIND] == NULL || !protocol) {
 538                         err = -EINVAL;
 539                         goto errout;
 540                 }
 541
 542                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 543                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 544                         err = -ENOENT;
 545                         goto errout;
 546                 }
 547
 548                 if (prio_allocate)
 549                         prio = tcf_auto_prio(tcf_chain_tp_prev(&chain_info));
 550
 551                 tp = tcf_proto_create(nla_data(tca[TCA_KIND]),
 552                                       protocol, prio, parent, q, chain);
 553                 if (IS_ERR(tp)) {
 554                         err = PTR_ERR(tp);
 555                         goto errout;
 556                 }
 557                 tp_created = 1;
 558         } else if (tca[TCA_KIND] && nla_strcmp(tca[TCA_KIND], tp->ops->kind)) {
 559                 err = -EINVAL;
 560                 goto errout;
 561         }
 562
 563         fh = tp->ops->get(tp, t->tcm_handle);
 564
 565         if (fh == 0) {
 566                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 567                         tcf_chain_tp_remove(chain, &chain_info, tp);
 568                         tfilter_notify(net, skb, n, tp, fh,
 569                                        RTM_DELTFILTER, false);
 570                         tcf_proto_destroy(tp);
 571                         err = 0;
 572                         goto errout;
 573                 }
 574
 575                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 576                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 577                         err = -ENOENT;
 578                         goto errout;
 579                 }
 580         } else {
 581                 bool last;
 582
 583                 switch (n->nlmsg_type) {
 584                 case RTM_NEWTFILTER:
 585                         if (n->nlmsg_flags & NLM_F_EXCL) {
 586                                 if (tp_created)
 587                                         tcf_proto_destroy(tp);
 588                                 err = -EEXIST;
 589                                 goto errout;
 590                         }
 591                         break;
 592                 case RTM_DELTFILTER:
 593                         err = tp->ops->delete(tp, fh, &last);
 594                         if (err)
 595                                 goto errout;
 596                         tfilter_notify(net, skb, n, tp, t->tcm_handle,
 597                                        RTM_DELTFILTER, false);
 598                         if (last) {
 599                                 tcf_chain_tp_remove(chain, &chain_info, tp);
 600                                 tcf_proto_destroy(tp);
 601                         }
 602                         goto errout;
 603                 case RTM_GETTFILTER:
 604                         err = tfilter_notify(net, skb, n, tp, fh,
 605                                              RTM_NEWTFILTER, true);
 606                         goto errout;
 607                 default:
 608                         err = -EINVAL;
 609                         goto errout;
 610                 }
 611         }
 612
 613         err = tp->ops->change(net, skb, tp, cl, t->tcm_handle, tca, &fh,
 614                               n->nlmsg_flags & NLM_F_CREATE ? TCA_ACT_NOREPLACE : TCA_ACT_REPLACE);
 615         if (err == 0) {
 616                 if (tp_created)
 617                         tcf_chain_tp_insert(chain, &chain_info, tp);
 618                 tfilter_notify(net, skb, n, tp, fh, RTM_NEWTFILTER, false);
 619         } else {
 620                 if (tp_created)
 621                         tcf_proto_destroy(tp);
 622         }
 623
 624 errout:
 625         if (chain)
 626                 tcf_chain_put(chain);
 627         if (cl)
 628                 cops->put(q, cl);
 629         if (err == -EAGAIN)
 630                 /* Replay the request. */
 631                 goto replay;
 632         return err;
 633 }
 634
 635 static int tcf_fill_node(struct net *net, struct sk_buff *skb,
 636                          struct tcf_proto *tp, unsigned long fh, u32 portid,
 637                          u32 seq, u16 flags, int event)
 638 {
 639         struct tcmsg *tcm;
 640         struct nlmsghdr  *nlh;
 641         unsigned char *b = skb_tail_pointer(skb);
 642
 643         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*tcm), flags);
 644         if (!nlh)
 645                 goto out_nlmsg_trim;
 646         tcm = nlmsg_data(nlh);
 647         tcm->tcm_family = AF_UNSPEC;
 648         tcm->tcm__pad1 = 0;
 649         tcm->tcm__pad2 = 0;
 650         tcm->tcm_ifindex = qdisc_dev(tp->q)->ifindex;
 651         tcm->tcm_parent = tp->classid;
 652         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 653         if (nla_put_string(skb, TCA_KIND, tp->ops->kind))
 654                 goto nla_put_failure;
 655         if (nla_put_u32(skb, TCA_CHAIN, tp->chain->index))
 656                 goto nla_put_failure;
 657         tcm->tcm_handle = fh;
 658         if (RTM_DELTFILTER != event) {
 659                 tcm->tcm_handle = 0;
 660                 if (tp->ops->dump && tp->ops->dump(net, tp, fh, skb, tcm) < 0)
 661                         goto nla_put_failure;
 662         }
 663         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 664         return skb->len;
 665
 666 out_nlmsg_trim:
 667 nla_put_failure:
 668         nlmsg_trim(skb, b);
 669         return -1;
 670 }
 671
 672 static int tfilter_notify(struct net *net, struct sk_buff *oskb,
 673                           struct nlmsghdr *n, struct tcf_proto *tp,
 674                           unsigned long fh, int event, bool unicast)
 675 {
 676         struct sk_buff *skb;
 677         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 678
 679         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 680         if (!skb)
 681                 return -ENOBUFS;
 682
 683         if (tcf_fill_node(net, skb, tp, fh, portid, n->nlmsg_seq,
 684                           n->nlmsg_flags, event) <= 0) {
 685                 kfree_skb(skb);
 686                 return -EINVAL;
 687         }
 688
 689         if (unicast)
 690                 return netlink_unicast(net->rtnl, skb, portid, MSG_DONTWAIT);
 691
 692         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 693                               n->nlmsg_flags & NLM_F_ECHO);
 694 }
 695
 696 struct tcf_dump_args {
 697         struct tcf_walker w;
 698         struct sk_buff *skb;
 699         struct netlink_callback *cb;
 700 };
 701
 702 static int tcf_node_dump(struct tcf_proto *tp, unsigned long n,
 703                          struct tcf_walker *arg)
 704 {
 705         struct tcf_dump_args *a = (void *)arg;
 706         struct net *net = sock_net(a->skb->sk);
 707
 708         return tcf_fill_node(net, a->skb, tp, n, NETLINK_CB(a->cb->skb).portid,
 709                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI,
 710                              RTM_NEWTFILTER);
 711 }
 712
 713 static bool tcf_chain_dump(struct tcf_chain *chain, struct sk_buff *skb,
 714                            struct netlink_callback *cb,
 715                            long index_start, long *p_index)
 716 {
 717         struct net *net = sock_net(skb->sk);
 718         struct tcmsg *tcm = nlmsg_data(cb->nlh);
 719         struct tcf_dump_args arg;
 720         struct tcf_proto *tp;
 721
 722         for (tp = rtnl_dereference(chain->filter_chain);
 723              tp; tp = rtnl_dereference(tp->next), (*p_index)++) {
 724                 if (*p_index < index_start)
 725                         continue;
 726                 if (TC_H_MAJ(tcm->tcm_info) &&
 727                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 728                         continue;
 729                 if (TC_H_MIN(tcm->tcm_info) &&
 730                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 731                         continue;
 732                 if (*p_index > index_start)
 733                         memset(&cb->args[1], 0,
 734                                sizeof(cb->args) - sizeof(cb->args[0]));
 735                 if (cb->args[1] == 0) {
 736                         if (tcf_fill_node(net, skb, tp, 0,
 737                                           NETLINK_CB(cb->skb).portid,
 738                                           cb->nlh->nlmsg_seq, NLM_F_MULTI,
 739                                           RTM_NEWTFILTER) <= 0)
 740                                 return false;
 741
 742                         cb->args[1] = 1;
 743                 }
 744                 if (!tp->ops->walk)
 745                         continue;
 746                 arg.w.fn = tcf_node_dump;
 747                 arg.skb = skb;
 748                 arg.cb = cb;
 749                 arg.w.stop = 0;
 750                 arg.w.skip = cb->args[1] - 1;
 751                 arg.w.count = 0;
 752                 tp->ops->walk(tp, &arg.w);
 753                 cb->args[1] = arg.w.count + 1;
 754                 if (arg.w.stop)
 755                         return false;
 756         }
 757         return true;
 758 }
 759
 760 /* called with RTNL */
 761 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
 762 {
 763         struct net *net = sock_net(skb->sk);
 764         struct nlattr *tca[TCA_MAX + 1];
 765         struct net_device *dev;
 766         struct Qdisc *q;
 767         struct tcf_block *block;
 768         struct tcf_chain *chain;
 769         struct tcmsg *tcm = nlmsg_data(cb->nlh);
 770         unsigned long cl = 0;
 771         const struct Qdisc_class_ops *cops;
 772         long index_start;
 773         long index;
 774         int err;
 775
 776         if (nlmsg_len(cb->nlh) < sizeof(*tcm))
 777                 return skb->len;
 778
 779         err = nlmsg_parse(cb->nlh, sizeof(*tcm), tca, TCA_MAX, NULL, NULL);
 780         if (err)
 781                 return err;
 782
 783         dev = __dev_get_by_index(net, tcm->tcm_ifindex);
 784         if (!dev)
 785                 return skb->len;
 786
 787         if (!tcm->tcm_parent)
 788                 q = dev->qdisc;
 789         else
 790                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
 791         if (!q)
 792                 goto out;
 793         cops = q->ops->cl_ops;
 794         if (!cops)
 795                 goto errout;
 796         if (!cops->tcf_block)
 797                 goto errout;
 798         if (TC_H_MIN(tcm->tcm_parent)) {
 799                 cl = cops->get(q, tcm->tcm_parent);
 800                 if (cl == 0)
 801                         goto errout;
 802         }
 803         block = cops->tcf_block(q, cl);
 804         if (!block)
 805                 goto errout;
 806
 807         index_start = cb->args[0];
 808         index = 0;
 809
 810         list_for_each_entry(chain, &block->chain_list, list) {
 811                 if (tca[TCA_CHAIN] &&
 812                     nla_get_u32(tca[TCA_CHAIN]) != chain->index)
 813                         continue;
 814                 if (!tcf_chain_dump(chain, skb, cb, index_start, &index))
 815                         break;
 816         }
 817
 818         cb->args[0] = index;
 819
 820 errout:
 821         if (cl)
 822                 cops->put(q, cl);
 823 out:
 824         return skb->len;
 825 }
 826
 827 void tcf_exts_destroy(struct tcf_exts *exts)
 828 {
 829 #ifdef CONFIG_NET_CLS_ACT
 830         LIST_HEAD(actions);
 831
 832         tcf_exts_to_list(exts, &actions);
 833         tcf_action_destroy(&actions, TCA_ACT_UNBIND);
 834         kfree(exts->actions);
 835         exts->nr_actions = 0;
 836 #endif
 837 }
 838 EXPORT_SYMBOL(tcf_exts_destroy);
 839
 840 int tcf_exts_validate(struct net *net, struct tcf_proto *tp, struct nlattr **tb,
 841                       struct nlattr *rate_tlv, struct tcf_exts *exts, bool ovr)
 842 {
 843 #ifdef CONFIG_NET_CLS_ACT
 844         {
 845                 struct tc_action *act;
 846
 847                 if (exts->police && tb[exts->police]) {
 848                         act = tcf_action_init_1(net, tp, tb[exts->police],
 849                                                 rate_tlv, "police", ovr,
 850                                                 TCA_ACT_BIND);
 851                         if (IS_ERR(act))
 852                                 return PTR_ERR(act);
 853
 854                         act->type = exts->type = TCA_OLD_COMPAT;
 855                         exts->actions[0] = act;
 856                         exts->nr_actions = 1;
 857                 } else if (exts->action && tb[exts->action]) {
 858                         LIST_HEAD(actions);
 859                         int err, i = 0;
 860
 861                         err = tcf_action_init(net, tp, tb[exts->action],
 862                                               rate_tlv, NULL, ovr, TCA_ACT_BIND,
 863                                               &actions);
 864                         if (err)
 865                                 return err;
 866                         list_for_each_entry(act, &actions, list)
 867                                 exts->actions[i++] = act;
 868                         exts->nr_actions = i;
 869                 }
 870         }
 871 #else
 872         if ((exts->action && tb[exts->action]) ||
 873             (exts->police && tb[exts->police]))
 874                 return -EOPNOTSUPP;
 875 #endif
 876
 877         return 0;
 878 }
 879 EXPORT_SYMBOL(tcf_exts_validate);
 880
 881 void tcf_exts_change(struct tcf_proto *tp, struct tcf_exts *dst,
 882                      struct tcf_exts *src)
 883 {
 884 #ifdef CONFIG_NET_CLS_ACT
 885         struct tcf_exts old = *dst;
 886
 887         tcf_tree_lock(tp);
 888         dst->nr_actions = src->nr_actions;
 889         dst->actions = src->actions;
 890         dst->type = src->type;
 891         tcf_tree_unlock(tp);
 892
 893         tcf_exts_destroy(&old);
 894 #endif
 895 }
 896 EXPORT_SYMBOL(tcf_exts_change);
 897
 898 #ifdef CONFIG_NET_CLS_ACT
 899 static struct tc_action *tcf_exts_first_act(struct tcf_exts *exts)
 900 {
 901         if (exts->nr_actions == 0)
 902                 return NULL;
 903         else
 904                 return exts->actions[0];
 905 }
 906 #endif
 907
 908 int tcf_exts_dump(struct sk_buff *skb, struct tcf_exts *exts)
 909 {
 910 #ifdef CONFIG_NET_CLS_ACT
 911         struct nlattr *nest;
 912
 913         if (exts->action && exts->nr_actions) {
 914                 /*
 915                  * again for backward compatible mode - we want
 916                  * to work with both old and new modes of entering
 917                  * tc data even if iproute2  was newer - jhs
 918                  */
 919                 if (exts->type != TCA_OLD_COMPAT) {
 920                         LIST_HEAD(actions);
 921
 922                         nest = nla_nest_start(skb, exts->action);
 923                         if (nest == NULL)
 924                                 goto nla_put_failure;
 925
 926                         tcf_exts_to_list(exts, &actions);
 927                         if (tcf_action_dump(skb, &actions, 0, 0) < 0)
 928                                 goto nla_put_failure;
 929                         nla_nest_end(skb, nest);
 930                 } else if (exts->police) {
 931                         struct tc_action *act = tcf_exts_first_act(exts);
 932                         nest = nla_nest_start(skb, exts->police);
 933                         if (nest == NULL || !act)
 934                                 goto nla_put_failure;
 935                         if (tcf_action_dump_old(skb, act, 0, 0) < 0)
 936                                 goto nla_put_failure;
 937                         nla_nest_end(skb, nest);
 938                 }
 939         }
 940         return 0;
 941
 942 nla_put_failure:
 943         nla_nest_cancel(skb, nest);
 944         return -1;
 945 #else
 946         return 0;
 947 #endif
 948 }
 949 EXPORT_SYMBOL(tcf_exts_dump);
 950
 951
 952 int tcf_exts_dump_stats(struct sk_buff *skb, struct tcf_exts *exts)
 953 {
 954 #ifdef CONFIG_NET_CLS_ACT
 955         struct tc_action *a = tcf_exts_first_act(exts);
 956         if (a != NULL && tcf_action_copy_stats(skb, a, 1) < 0)
 957                 return -1;
 958 #endif
 959         return 0;
 960 }
 961 EXPORT_SYMBOL(tcf_exts_dump_stats);
 962
 963 int tcf_exts_get_dev(struct net_device *dev, struct tcf_exts *exts,
 964                      struct net_device **hw_dev)
 965 {
 966 #ifdef CONFIG_NET_CLS_ACT
 967         const struct tc_action *a;
 968         LIST_HEAD(actions);
 969
 970         if (tc_no_actions(exts))
 971                 return -EINVAL;
 972
 973         tcf_exts_to_list(exts, &actions);
 974         list_for_each_entry(a, &actions, list) {
 975                 if (a->ops->get_dev) {
 976                         a->ops->get_dev(a, dev_net(dev), hw_dev);
 977                         break;
 978                 }
 979         }
 980         if (*hw_dev)
 981                 return 0;
 982 #endif
 983         return -EOPNOTSUPP;
 984 }
 985 EXPORT_SYMBOL(tcf_exts_get_dev);
 986
 987 static int __init tc_filter_init(void)
 988 {
 989         rtnl_register(PF_UNSPEC, RTM_NEWTFILTER, tc_ctl_tfilter, NULL, NULL);
 990         rtnl_register(PF_UNSPEC, RTM_DELTFILTER, tc_ctl_tfilter, NULL, NULL);
 991         rtnl_register(PF_UNSPEC, RTM_GETTFILTER, tc_ctl_tfilter,
 992                       tc_dump_tfilter, NULL);
 993
 994         return 0;
 995 }
 996
 997 subsys_initcall(tc_filter_init);