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