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 static struct workqueue_struct *tc_filter_wq;
  81
  82 int unregister_tcf_proto_ops(struct tcf_proto_ops *ops)
  83 {
  84         struct tcf_proto_ops *t;
  85         int rc = -ENOENT;
  86
  87         /* Wait for outstanding call_rcu()s, if any, from a
  88          * tcf_proto_ops's destroy() handler.
  89          */
  90         rcu_barrier();
  91         flush_workqueue(tc_filter_wq);
  92
  93         write_lock(&cls_mod_lock);
  94         list_for_each_entry(t, &tcf_proto_base, head) {
  95                 if (t == ops) {
  96                         list_del(&t->head);
  97                         rc = 0;
  98                         break;
  99                 }
 100         }
 101         write_unlock(&cls_mod_lock);
 102         return rc;
 103 }
 104 EXPORT_SYMBOL(unregister_tcf_proto_ops);
 105
 106 bool tcf_queue_work(struct work_struct *work)
 107 {
 108         return queue_work(tc_filter_wq, work);
 109 }
 110 EXPORT_SYMBOL(tcf_queue_work);
 111
 112 /* Select new prio value from the range, managed by kernel. */
 113
 114 static inline u32 tcf_auto_prio(struct tcf_proto *tp)
 115 {
 116         u32 first = TC_H_MAKE(0xC0000000U, 0U);
 117
 118         if (tp)
 119                 first = tp->prio - 1;
 120
 121         return TC_H_MAJ(first);
 122 }
 123
 124 static struct tcf_proto *tcf_proto_create(const char *kind, u32 protocol,
 125                                           u32 prio, u32 parent, struct Qdisc *q,
 126                                           struct tcf_chain *chain)
 127 {
 128         struct tcf_proto *tp;
 129         int err;
 130
 131         tp = kzalloc(sizeof(*tp), GFP_KERNEL);
 132         if (!tp)
 133                 return ERR_PTR(-ENOBUFS);
 134
 135         err = -ENOENT;
 136         tp->ops = tcf_proto_lookup_ops(kind);
 137         if (!tp->ops) {
 138 #ifdef CONFIG_MODULES
 139                 rtnl_unlock();
 140                 request_module("cls_%s", kind);
 141                 rtnl_lock();
 142                 tp->ops = tcf_proto_lookup_ops(kind);
 143                 /* We dropped the RTNL semaphore in order to perform
 144                  * the module load. So, even if we succeeded in loading
 145                  * the module we have to replay the request. We indicate
 146                  * this using -EAGAIN.
 147                  */
 148                 if (tp->ops) {
 149                         module_put(tp->ops->owner);
 150                         err = -EAGAIN;
 151                 } else {
 152                         err = -ENOENT;
 153                 }
 154                 goto errout;
 155 #endif
 156         }
 157         tp->classify = tp->ops->classify;
 158         tp->protocol = protocol;
 159         tp->prio = prio;
 160         tp->classid = parent;
 161         tp->q = q;
 162         tp->chain = chain;
 163
 164         err = tp->ops->init(tp);
 165         if (err) {
 166                 module_put(tp->ops->owner);
 167                 goto errout;
 168         }
 169         return tp;
 170
 171 errout:
 172         kfree(tp);
 173         return ERR_PTR(err);
 174 }
 175
 176 static void tcf_proto_destroy(struct tcf_proto *tp)
 177 {
 178         tp->ops->destroy(tp);
 179         module_put(tp->ops->owner);
 180         kfree_rcu(tp, rcu);
 181 }
 182
 183 static struct tcf_chain *tcf_chain_create(struct tcf_block *block,
 184                                           u32 chain_index)
 185 {
 186         struct tcf_chain *chain;
 187
 188         chain = kzalloc(sizeof(*chain), GFP_KERNEL);
 189         if (!chain)
 190                 return NULL;
 191         list_add_tail(&chain->list, &block->chain_list);
 192         chain->block = block;
 193         chain->index = chain_index;
 194         chain->refcnt = 1;
 195         return chain;
 196 }
 197
 198 static void tcf_chain_flush(struct tcf_chain *chain)
 199 {
 200         struct tcf_proto *tp;
 201
 202         if (chain->p_filter_chain)
 203                 RCU_INIT_POINTER(*chain->p_filter_chain, NULL);
 204         while ((tp = rtnl_dereference(chain->filter_chain)) != NULL) {
 205                 RCU_INIT_POINTER(chain->filter_chain, tp->next);
 206                 tcf_chain_put(chain);
 207                 tcf_proto_destroy(tp);
 208         }
 209 }
 210
 211 static void tcf_chain_destroy(struct tcf_chain *chain)
 212 {
 213         list_del(&chain->list);
 214         kfree(chain);
 215 }
 216
 217 static void tcf_chain_hold(struct tcf_chain *chain)
 218 {
 219         ++chain->refcnt;
 220 }
 221
 222 struct tcf_chain *tcf_chain_get(struct tcf_block *block, u32 chain_index,
 223                                 bool create)
 224 {
 225         struct tcf_chain *chain;
 226
 227         list_for_each_entry(chain, &block->chain_list, list) {
 228                 if (chain->index == chain_index) {
 229                         tcf_chain_hold(chain);
 230                         return chain;
 231                 }
 232         }
 233
 234         return create ? tcf_chain_create(block, chain_index) : NULL;
 235 }
 236 EXPORT_SYMBOL(tcf_chain_get);
 237
 238 void tcf_chain_put(struct tcf_chain *chain)
 239 {
 240         if (--chain->refcnt == 0)
 241                 tcf_chain_destroy(chain);
 242 }
 243 EXPORT_SYMBOL(tcf_chain_put);
 244
 245 static void
 246 tcf_chain_filter_chain_ptr_set(struct tcf_chain *chain,
 247                                struct tcf_proto __rcu **p_filter_chain)
 248 {
 249         chain->p_filter_chain = p_filter_chain;
 250 }
 251
 252 static void tcf_block_offload_cmd(struct tcf_block *block, struct Qdisc *q,
 253                                   struct tcf_block_ext_info *ei,
 254                                   enum tc_block_command command)
 255 {
 256         struct net_device *dev = q->dev_queue->dev;
 257         struct tc_block_offload bo = {};
 258
 259         if (!dev->netdev_ops->ndo_setup_tc)
 260                 return;
 261         bo.command = command;
 262         bo.binder_type = ei->binder_type;
 263         bo.block = block;
 264         dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_BLOCK, &bo);
 265 }
 266
 267 static void tcf_block_offload_bind(struct tcf_block *block, struct Qdisc *q,
 268                                    struct tcf_block_ext_info *ei)
 269 {
 270         tcf_block_offload_cmd(block, q, ei, TC_BLOCK_BIND);
 271 }
 272
 273 static void tcf_block_offload_unbind(struct tcf_block *block, struct Qdisc *q,
 274                                      struct tcf_block_ext_info *ei)
 275 {
 276         tcf_block_offload_cmd(block, q, ei, TC_BLOCK_UNBIND);
 277 }
 278
 279 int tcf_block_get_ext(struct tcf_block **p_block,
 280                       struct tcf_proto __rcu **p_filter_chain, struct Qdisc *q,
 281                       struct tcf_block_ext_info *ei)
 282 {
 283         struct tcf_block *block = kzalloc(sizeof(*block), GFP_KERNEL);
 284         struct tcf_chain *chain;
 285         int err;
 286
 287         if (!block)
 288                 return -ENOMEM;
 289         INIT_LIST_HEAD(&block->chain_list);
 290         INIT_LIST_HEAD(&block->cb_list);
 291
 292         /* Create chain 0 by default, it has to be always present. */
 293         chain = tcf_chain_create(block, 0);
 294         if (!chain) {
 295                 err = -ENOMEM;
 296                 goto err_chain_create;
 297         }
 298         tcf_chain_filter_chain_ptr_set(chain, p_filter_chain);
 299         block->net = qdisc_net(q);
 300         block->q = q;
 301         tcf_block_offload_bind(block, q, ei);
 302         *p_block = block;
 303         return 0;
 304
 305 err_chain_create:
 306         kfree(block);
 307         return err;
 308 }
 309 EXPORT_SYMBOL(tcf_block_get_ext);
 310
 311 int tcf_block_get(struct tcf_block **p_block,
 312                   struct tcf_proto __rcu **p_filter_chain, struct Qdisc *q)
 313 {
 314         struct tcf_block_ext_info ei = {0, };
 315
 316         return tcf_block_get_ext(p_block, p_filter_chain, q, &ei);
 317 }
 318 EXPORT_SYMBOL(tcf_block_get);
 319
 320 static void tcf_block_put_final(struct work_struct *work)
 321 {
 322         struct tcf_block *block = container_of(work, struct tcf_block, work);
 323         struct tcf_chain *chain, *tmp;
 324
 325         rtnl_lock();
 326         /* Only chain 0 should be still here. */
 327         list_for_each_entry_safe(chain, tmp, &block->chain_list, list)
 328                 tcf_chain_put(chain);
 329         rtnl_unlock();
 330         kfree(block);
 331 }
 332
 333 /* XXX: Standalone actions are not allowed to jump to any chain, and bound
 334  * actions should be all removed after flushing. However, filters are now
 335  * destroyed in tc filter workqueue with RTNL lock, they can not race here.
 336  */
 337 void tcf_block_put_ext(struct tcf_block *block,
 338                        struct tcf_proto __rcu **p_filter_chain, struct Qdisc *q,
 339                        struct tcf_block_ext_info *ei)
 340 {
 341         struct tcf_chain *chain, *tmp;
 342
 343         if (!block)
 344                 return;
 345
 346         tcf_block_offload_unbind(block, q, ei);
 347
 348         list_for_each_entry_safe(chain, tmp, &block->chain_list, list)
 349                 tcf_chain_flush(chain);
 350
 351         INIT_WORK(&block->work, tcf_block_put_final);
 352         /* Wait for existing RCU callbacks to cool down, make sure their works
 353          * have been queued before this. We can not flush pending works here
 354          * because we are holding the RTNL lock.
 355          */
 356         rcu_barrier();
 357         tcf_queue_work(&block->work);
 358 }
 359 EXPORT_SYMBOL(tcf_block_put_ext);
 360
 361 void tcf_block_put(struct tcf_block *block)
 362 {
 363         struct tcf_block_ext_info ei = {0, };
 364
 365         tcf_block_put_ext(block, NULL, block->q, &ei);
 366 }
 367
 368 EXPORT_SYMBOL(tcf_block_put);
 369
 370 struct tcf_block_cb {
 371         struct list_head list;
 372         tc_setup_cb_t *cb;
 373         void *cb_ident;
 374         void *cb_priv;
 375         unsigned int refcnt;
 376 };
 377
 378 void *tcf_block_cb_priv(struct tcf_block_cb *block_cb)
 379 {
 380         return block_cb->cb_priv;
 381 }
 382 EXPORT_SYMBOL(tcf_block_cb_priv);
 383
 384 struct tcf_block_cb *tcf_block_cb_lookup(struct tcf_block *block,
 385                                          tc_setup_cb_t *cb, void *cb_ident)
 386 {       struct tcf_block_cb *block_cb;
 387
 388         list_for_each_entry(block_cb, &block->cb_list, list)
 389                 if (block_cb->cb == cb && block_cb->cb_ident == cb_ident)
 390                         return block_cb;
 391         return NULL;
 392 }
 393 EXPORT_SYMBOL(tcf_block_cb_lookup);
 394
 395 void tcf_block_cb_incref(struct tcf_block_cb *block_cb)
 396 {
 397         block_cb->refcnt++;
 398 }
 399 EXPORT_SYMBOL(tcf_block_cb_incref);
 400
 401 unsigned int tcf_block_cb_decref(struct tcf_block_cb *block_cb)
 402 {
 403         return --block_cb->refcnt;
 404 }
 405 EXPORT_SYMBOL(tcf_block_cb_decref);
 406
 407 struct tcf_block_cb *__tcf_block_cb_register(struct tcf_block *block,
 408                                              tc_setup_cb_t *cb, void *cb_ident,
 409                                              void *cb_priv)
 410 {
 411         struct tcf_block_cb *block_cb;
 412
 413         block_cb = kzalloc(sizeof(*block_cb), GFP_KERNEL);
 414         if (!block_cb)
 415                 return NULL;
 416         block_cb->cb = cb;
 417         block_cb->cb_ident = cb_ident;
 418         block_cb->cb_priv = cb_priv;
 419         list_add(&block_cb->list, &block->cb_list);
 420         return block_cb;
 421 }
 422 EXPORT_SYMBOL(__tcf_block_cb_register);
 423
 424 int tcf_block_cb_register(struct tcf_block *block,
 425                           tc_setup_cb_t *cb, void *cb_ident,
 426                           void *cb_priv)
 427 {
 428         struct tcf_block_cb *block_cb;
 429
 430         block_cb = __tcf_block_cb_register(block, cb, cb_ident, cb_priv);
 431         return block_cb ? 0 : -ENOMEM;
 432 }
 433 EXPORT_SYMBOL(tcf_block_cb_register);
 434
 435 void __tcf_block_cb_unregister(struct tcf_block_cb *block_cb)
 436 {
 437         list_del(&block_cb->list);
 438         kfree(block_cb);
 439 }
 440 EXPORT_SYMBOL(__tcf_block_cb_unregister);
 441
 442 void tcf_block_cb_unregister(struct tcf_block *block,
 443                              tc_setup_cb_t *cb, void *cb_ident)
 444 {
 445         struct tcf_block_cb *block_cb;
 446
 447         block_cb = tcf_block_cb_lookup(block, cb, cb_ident);
 448         if (!block_cb)
 449                 return;
 450         __tcf_block_cb_unregister(block_cb);
 451 }
 452 EXPORT_SYMBOL(tcf_block_cb_unregister);
 453
 454 static int tcf_block_cb_call(struct tcf_block *block, enum tc_setup_type type,
 455                              void *type_data, bool err_stop)
 456 {
 457         struct tcf_block_cb *block_cb;
 458         int ok_count = 0;
 459         int err;
 460
 461         list_for_each_entry(block_cb, &block->cb_list, list) {
 462                 err = block_cb->cb(type, type_data, block_cb->cb_priv);
 463                 if (err) {
 464                         if (err_stop)
 465                                 return err;
 466                 } else {
 467                         ok_count++;
 468                 }
 469         }
 470         return ok_count;
 471 }
 472
 473 /* Main classifier routine: scans classifier chain attached
 474  * to this qdisc, (optionally) tests for protocol and asks
 475  * specific classifiers.
 476  */
 477 int tcf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 478                  struct tcf_result *res, bool compat_mode)
 479 {
 480         __be16 protocol = tc_skb_protocol(skb);
 481 #ifdef CONFIG_NET_CLS_ACT
 482         const int max_reclassify_loop = 4;
 483         const struct tcf_proto *orig_tp = tp;
 484         const struct tcf_proto *first_tp;
 485         int limit = 0;
 486
 487 reclassify:
 488 #endif
 489         for (; tp; tp = rcu_dereference_bh(tp->next)) {
 490                 int err;
 491
 492                 if (tp->protocol != protocol &&
 493                     tp->protocol != htons(ETH_P_ALL))
 494                         continue;
 495
 496                 err = tp->classify(skb, tp, res);
 497 #ifdef CONFIG_NET_CLS_ACT
 498                 if (unlikely(err == TC_ACT_RECLASSIFY && !compat_mode)) {
 499                         first_tp = orig_tp;
 500                         goto reset;
 501                 } else if (unlikely(TC_ACT_EXT_CMP(err, TC_ACT_GOTO_CHAIN))) {
 502                         first_tp = res->goto_tp;
 503                         goto reset;
 504                 }
 505 #endif
 506                 if (err >= 0)
 507                         return err;
 508         }
 509
 510         return TC_ACT_UNSPEC; /* signal: continue lookup */
 511 #ifdef CONFIG_NET_CLS_ACT
 512 reset:
 513         if (unlikely(limit++ >= max_reclassify_loop)) {
 514                 net_notice_ratelimited("%s: reclassify loop, rule prio %u, protocol %02x\n",
 515                                        tp->q->ops->id, tp->prio & 0xffff,
 516                                        ntohs(tp->protocol));
 517                 return TC_ACT_SHOT;
 518         }
 519
 520         tp = first_tp;
 521         protocol = tc_skb_protocol(skb);
 522         goto reclassify;
 523 #endif
 524 }
 525 EXPORT_SYMBOL(tcf_classify);
 526
 527 struct tcf_chain_info {
 528         struct tcf_proto __rcu **pprev;
 529         struct tcf_proto __rcu *next;
 530 };
 531
 532 static struct tcf_proto *tcf_chain_tp_prev(struct tcf_chain_info *chain_info)
 533 {
 534         return rtnl_dereference(*chain_info->pprev);
 535 }
 536
 537 static void tcf_chain_tp_insert(struct tcf_chain *chain,
 538                                 struct tcf_chain_info *chain_info,
 539                                 struct tcf_proto *tp)
 540 {
 541         if (chain->p_filter_chain &&
 542             *chain_info->pprev == chain->filter_chain)
 543                 rcu_assign_pointer(*chain->p_filter_chain, tp);
 544         RCU_INIT_POINTER(tp->next, tcf_chain_tp_prev(chain_info));
 545         rcu_assign_pointer(*chain_info->pprev, tp);
 546         tcf_chain_hold(chain);
 547 }
 548
 549 static void tcf_chain_tp_remove(struct tcf_chain *chain,
 550                                 struct tcf_chain_info *chain_info,
 551                                 struct tcf_proto *tp)
 552 {
 553         struct tcf_proto *next = rtnl_dereference(chain_info->next);
 554
 555         if (chain->p_filter_chain && tp == chain->filter_chain)
 556                 RCU_INIT_POINTER(*chain->p_filter_chain, next);
 557         RCU_INIT_POINTER(*chain_info->pprev, next);
 558         tcf_chain_put(chain);
 559 }
 560
 561 static struct tcf_proto *tcf_chain_tp_find(struct tcf_chain *chain,
 562                                            struct tcf_chain_info *chain_info,
 563                                            u32 protocol, u32 prio,
 564                                            bool prio_allocate)
 565 {
 566         struct tcf_proto **pprev;
 567         struct tcf_proto *tp;
 568
 569         /* Check the chain for existence of proto-tcf with this priority */
 570         for (pprev = &chain->filter_chain;
 571              (tp = rtnl_dereference(*pprev)); pprev = &tp->next) {
 572                 if (tp->prio >= prio) {
 573                         if (tp->prio == prio) {
 574                                 if (prio_allocate ||
 575                                     (tp->protocol != protocol && protocol))
 576                                         return ERR_PTR(-EINVAL);
 577                         } else {
 578                                 tp = NULL;
 579                         }
 580                         break;
 581                 }
 582         }
 583         chain_info->pprev = pprev;
 584         chain_info->next = tp ? tp->next : NULL;
 585         return tp;
 586 }
 587
 588 static int tcf_fill_node(struct net *net, struct sk_buff *skb,
 589                          struct tcf_proto *tp, struct Qdisc *q, u32 parent,
 590                          void *fh, u32 portid, u32 seq, u16 flags, int event)
 591 {
 592         struct tcmsg *tcm;
 593         struct nlmsghdr  *nlh;
 594         unsigned char *b = skb_tail_pointer(skb);
 595
 596         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*tcm), flags);
 597         if (!nlh)
 598                 goto out_nlmsg_trim;
 599         tcm = nlmsg_data(nlh);
 600         tcm->tcm_family = AF_UNSPEC;
 601         tcm->tcm__pad1 = 0;
 602         tcm->tcm__pad2 = 0;
 603         tcm->tcm_ifindex = qdisc_dev(q)->ifindex;
 604         tcm->tcm_parent = parent;
 605         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 606         if (nla_put_string(skb, TCA_KIND, tp->ops->kind))
 607                 goto nla_put_failure;
 608         if (nla_put_u32(skb, TCA_CHAIN, tp->chain->index))
 609                 goto nla_put_failure;
 610         if (!fh) {
 611                 tcm->tcm_handle = 0;
 612         } else {
 613                 if (tp->ops->dump && tp->ops->dump(net, tp, fh, skb, tcm) < 0)
 614                         goto nla_put_failure;
 615         }
 616         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 617         return skb->len;
 618
 619 out_nlmsg_trim:
 620 nla_put_failure:
 621         nlmsg_trim(skb, b);
 622         return -1;
 623 }
 624
 625 static int tfilter_notify(struct net *net, struct sk_buff *oskb,
 626                           struct nlmsghdr *n, struct tcf_proto *tp,
 627                           struct Qdisc *q, u32 parent,
 628                           void *fh, int event, bool unicast)
 629 {
 630         struct sk_buff *skb;
 631         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 632
 633         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 634         if (!skb)
 635                 return -ENOBUFS;
 636
 637         if (tcf_fill_node(net, skb, tp, q, parent, fh, portid, n->nlmsg_seq,
 638                           n->nlmsg_flags, event) <= 0) {
 639                 kfree_skb(skb);
 640                 return -EINVAL;
 641         }
 642
 643         if (unicast)
 644                 return netlink_unicast(net->rtnl, skb, portid, MSG_DONTWAIT);
 645
 646         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 647                               n->nlmsg_flags & NLM_F_ECHO);
 648 }
 649
 650 static int tfilter_del_notify(struct net *net, struct sk_buff *oskb,
 651                               struct nlmsghdr *n, struct tcf_proto *tp,
 652                               struct Qdisc *q, u32 parent,
 653                               void *fh, bool unicast, bool *last)
 654 {
 655         struct sk_buff *skb;
 656         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 657         int err;
 658
 659         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 660         if (!skb)
 661                 return -ENOBUFS;
 662
 663         if (tcf_fill_node(net, skb, tp, q, parent, fh, portid, n->nlmsg_seq,
 664                           n->nlmsg_flags, RTM_DELTFILTER) <= 0) {
 665                 kfree_skb(skb);
 666                 return -EINVAL;
 667         }
 668
 669         err = tp->ops->delete(tp, fh, last);
 670         if (err) {
 671                 kfree_skb(skb);
 672                 return err;
 673         }
 674
 675         if (unicast)
 676                 return netlink_unicast(net->rtnl, skb, portid, MSG_DONTWAIT);
 677
 678         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 679                               n->nlmsg_flags & NLM_F_ECHO);
 680 }
 681
 682 static void tfilter_notify_chain(struct net *net, struct sk_buff *oskb,
 683                                  struct Qdisc *q, u32 parent,
 684                                  struct nlmsghdr *n,
 685                                  struct tcf_chain *chain, int event)
 686 {
 687         struct tcf_proto *tp;
 688
 689         for (tp = rtnl_dereference(chain->filter_chain);
 690              tp; tp = rtnl_dereference(tp->next))
 691                 tfilter_notify(net, oskb, n, tp, q, parent, 0, event, false);
 692 }
 693
 694 /* Add/change/delete/get a filter node */
 695
 696 static int tc_ctl_tfilter(struct sk_buff *skb, struct nlmsghdr *n,
 697                           struct netlink_ext_ack *extack)
 698 {
 699         struct net *net = sock_net(skb->sk);
 700         struct nlattr *tca[TCA_MAX + 1];
 701         struct tcmsg *t;
 702         u32 protocol;
 703         u32 prio;
 704         bool prio_allocate;
 705         u32 parent;
 706         u32 chain_index;
 707         struct net_device *dev;
 708         struct Qdisc  *q;
 709         struct tcf_chain_info chain_info;
 710         struct tcf_chain *chain = NULL;
 711         struct tcf_block *block;
 712         struct tcf_proto *tp;
 713         const struct Qdisc_class_ops *cops;
 714         unsigned long cl;
 715         void *fh;
 716         int err;
 717         int tp_created;
 718
 719         if ((n->nlmsg_type != RTM_GETTFILTER) &&
 720             !netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN))
 721                 return -EPERM;
 722
 723 replay:
 724         tp_created = 0;
 725
 726         err = nlmsg_parse(n, sizeof(*t), tca, TCA_MAX, NULL, extack);
 727         if (err < 0)
 728                 return err;
 729
 730         t = nlmsg_data(n);
 731         protocol = TC_H_MIN(t->tcm_info);
 732         prio = TC_H_MAJ(t->tcm_info);
 733         prio_allocate = false;
 734         parent = t->tcm_parent;
 735         cl = 0;
 736
 737         if (prio == 0) {
 738                 switch (n->nlmsg_type) {
 739                 case RTM_DELTFILTER:
 740                         if (protocol || t->tcm_handle || tca[TCA_KIND])
 741                                 return -ENOENT;
 742                         break;
 743                 case RTM_NEWTFILTER:
 744                         /* If no priority is provided by the user,
 745                          * we allocate one.
 746                          */
 747                         if (n->nlmsg_flags & NLM_F_CREATE) {
 748                                 prio = TC_H_MAKE(0x80000000U, 0U);
 749                                 prio_allocate = true;
 750                                 break;
 751                         }
 752                         /* fall-through */
 753                 default:
 754                         return -ENOENT;
 755                 }
 756         }
 757
 758         /* Find head of filter chain. */
 759
 760         /* Find link */
 761         dev = __dev_get_by_index(net, t->tcm_ifindex);
 762         if (dev == NULL)
 763                 return -ENODEV;
 764
 765         /* Find qdisc */
 766         if (!parent) {
 767                 q = dev->qdisc;
 768                 parent = q->handle;
 769         } else {
 770                 q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent));
 771                 if (q == NULL)
 772                         return -EINVAL;
 773         }
 774
 775         /* Is it classful? */
 776         cops = q->ops->cl_ops;
 777         if (!cops)
 778                 return -EINVAL;
 779
 780         if (!cops->tcf_block)
 781                 return -EOPNOTSUPP;
 782
 783         /* Do we search for filter, attached to class? */
 784         if (TC_H_MIN(parent)) {
 785                 cl = cops->find(q, parent);
 786                 if (cl == 0)
 787                         return -ENOENT;
 788         }
 789
 790         /* And the last stroke */
 791         block = cops->tcf_block(q, cl);
 792         if (!block) {
 793                 err = -EINVAL;
 794                 goto errout;
 795         }
 796
 797         chain_index = tca[TCA_CHAIN] ? nla_get_u32(tca[TCA_CHAIN]) : 0;
 798         if (chain_index > TC_ACT_EXT_VAL_MASK) {
 799                 err = -EINVAL;
 800                 goto errout;
 801         }
 802         chain = tcf_chain_get(block, chain_index,
 803                               n->nlmsg_type == RTM_NEWTFILTER);
 804         if (!chain) {
 805                 err = n->nlmsg_type == RTM_NEWTFILTER ? -ENOMEM : -EINVAL;
 806                 goto errout;
 807         }
 808
 809         if (n->nlmsg_type == RTM_DELTFILTER && prio == 0) {
 810                 tfilter_notify_chain(net, skb, q, parent, n,
 811                                      chain, RTM_DELTFILTER);
 812                 tcf_chain_flush(chain);
 813                 err = 0;
 814                 goto errout;
 815         }
 816
 817         tp = tcf_chain_tp_find(chain, &chain_info, protocol,
 818                                prio, prio_allocate);
 819         if (IS_ERR(tp)) {
 820                 err = PTR_ERR(tp);
 821                 goto errout;
 822         }
 823
 824         if (tp == NULL) {
 825                 /* Proto-tcf does not exist, create new one */
 826
 827                 if (tca[TCA_KIND] == NULL || !protocol) {
 828                         err = -EINVAL;
 829                         goto errout;
 830                 }
 831
 832                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 833                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 834                         err = -ENOENT;
 835                         goto errout;
 836                 }
 837
 838                 if (prio_allocate)
 839                         prio = tcf_auto_prio(tcf_chain_tp_prev(&chain_info));
 840
 841                 tp = tcf_proto_create(nla_data(tca[TCA_KIND]),
 842                                       protocol, prio, parent, q, chain);
 843                 if (IS_ERR(tp)) {
 844                         err = PTR_ERR(tp);
 845                         goto errout;
 846                 }
 847                 tp_created = 1;
 848         } else if (tca[TCA_KIND] && nla_strcmp(tca[TCA_KIND], tp->ops->kind)) {
 849                 err = -EINVAL;
 850                 goto errout;
 851         }
 852
 853         fh = tp->ops->get(tp, t->tcm_handle);
 854
 855         if (!fh) {
 856                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 857                         tcf_chain_tp_remove(chain, &chain_info, tp);
 858                         tfilter_notify(net, skb, n, tp, q, parent, fh,
 859                                        RTM_DELTFILTER, false);
 860                         tcf_proto_destroy(tp);
 861                         err = 0;
 862                         goto errout;
 863                 }
 864
 865                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 866                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 867                         err = -ENOENT;
 868                         goto errout;
 869                 }
 870         } else {
 871                 bool last;
 872
 873                 switch (n->nlmsg_type) {
 874                 case RTM_NEWTFILTER:
 875                         if (n->nlmsg_flags & NLM_F_EXCL) {
 876                                 if (tp_created)
 877                                         tcf_proto_destroy(tp);
 878                                 err = -EEXIST;
 879                                 goto errout;
 880                         }
 881                         break;
 882                 case RTM_DELTFILTER:
 883                         err = tfilter_del_notify(net, skb, n, tp, q, parent,
 884                                                  fh, false, &last);
 885                         if (err)
 886                                 goto errout;
 887                         if (last) {
 888                                 tcf_chain_tp_remove(chain, &chain_info, tp);
 889                                 tcf_proto_destroy(tp);
 890                         }
 891                         goto errout;
 892                 case RTM_GETTFILTER:
 893                         err = tfilter_notify(net, skb, n, tp, q, parent, fh,
 894                                              RTM_NEWTFILTER, true);
 895                         goto errout;
 896                 default:
 897                         err = -EINVAL;
 898                         goto errout;
 899                 }
 900         }
 901
 902         err = tp->ops->change(net, skb, tp, cl, t->tcm_handle, tca, &fh,
 903                               n->nlmsg_flags & NLM_F_CREATE ? TCA_ACT_NOREPLACE : TCA_ACT_REPLACE);
 904         if (err == 0) {
 905                 if (tp_created)
 906                         tcf_chain_tp_insert(chain, &chain_info, tp);
 907                 tfilter_notify(net, skb, n, tp, q, parent, fh,
 908                                RTM_NEWTFILTER, false);
 909         } else {
 910                 if (tp_created)
 911                         tcf_proto_destroy(tp);
 912         }
 913
 914 errout:
 915         if (chain)
 916                 tcf_chain_put(chain);
 917         if (err == -EAGAIN)
 918                 /* Replay the request. */
 919                 goto replay;
 920         return err;
 921 }
 922
 923 struct tcf_dump_args {
 924         struct tcf_walker w;
 925         struct sk_buff *skb;
 926         struct netlink_callback *cb;
 927         struct Qdisc *q;
 928         u32 parent;
 929 };
 930
 931 static int tcf_node_dump(struct tcf_proto *tp, void *n, struct tcf_walker *arg)
 932 {
 933         struct tcf_dump_args *a = (void *)arg;
 934         struct net *net = sock_net(a->skb->sk);
 935
 936         return tcf_fill_node(net, a->skb, tp, a->q, a->parent,
 937                              n, NETLINK_CB(a->cb->skb).portid,
 938                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI,
 939                              RTM_NEWTFILTER);
 940 }
 941
 942 static bool tcf_chain_dump(struct tcf_chain *chain, struct Qdisc *q, u32 parent,
 943                            struct sk_buff *skb, struct netlink_callback *cb,
 944                            long index_start, long *p_index)
 945 {
 946         struct net *net = sock_net(skb->sk);
 947         struct tcmsg *tcm = nlmsg_data(cb->nlh);
 948         struct tcf_dump_args arg;
 949         struct tcf_proto *tp;
 950
 951         for (tp = rtnl_dereference(chain->filter_chain);
 952              tp; tp = rtnl_dereference(tp->next), (*p_index)++) {
 953                 if (*p_index < index_start)
 954                         continue;
 955                 if (TC_H_MAJ(tcm->tcm_info) &&
 956                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 957                         continue;
 958                 if (TC_H_MIN(tcm->tcm_info) &&
 959                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 960                         continue;
 961                 if (*p_index > index_start)
 962                         memset(&cb->args[1], 0,
 963                                sizeof(cb->args) - sizeof(cb->args[0]));
 964                 if (cb->args[1] == 0) {
 965                         if (tcf_fill_node(net, skb, tp, q, parent, 0,
 966                                           NETLINK_CB(cb->skb).portid,
 967                                           cb->nlh->nlmsg_seq, NLM_F_MULTI,
 968                                           RTM_NEWTFILTER) <= 0)
 969                                 return false;
 970
 971                         cb->args[1] = 1;
 972                 }
 973                 if (!tp->ops->walk)
 974                         continue;
 975                 arg.w.fn = tcf_node_dump;
 976                 arg.skb = skb;
 977                 arg.cb = cb;
 978                 arg.q = q;
 979                 arg.parent = parent;
 980                 arg.w.stop = 0;
 981                 arg.w.skip = cb->args[1] - 1;
 982                 arg.w.count = 0;
 983                 tp->ops->walk(tp, &arg.w);
 984                 cb->args[1] = arg.w.count + 1;
 985                 if (arg.w.stop)
 986                         return false;
 987         }
 988         return true;
 989 }
 990
 991 /* called with RTNL */
 992 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
 993 {
 994         struct net *net = sock_net(skb->sk);
 995         struct nlattr *tca[TCA_MAX + 1];
 996         struct net_device *dev;
 997         struct Qdisc *q;
 998         struct tcf_block *block;
 999         struct tcf_chain *chain;
1000         struct tcmsg *tcm = nlmsg_data(cb->nlh);
1001         unsigned long cl = 0;
1002         const struct Qdisc_class_ops *cops;
1003         long index_start;
1004         long index;
1005         u32 parent;
1006         int err;
1007
1008         if (nlmsg_len(cb->nlh) < sizeof(*tcm))
1009                 return skb->len;
1010
1011         err = nlmsg_parse(cb->nlh, sizeof(*tcm), tca, TCA_MAX, NULL, NULL);
1012         if (err)
1013                 return err;
1014
1015         dev = __dev_get_by_index(net, tcm->tcm_ifindex);
1016         if (!dev)
1017                 return skb->len;
1018
1019         parent = tcm->tcm_parent;
1020         if (!parent) {
1021                 q = dev->qdisc;
1022                 parent = q->handle;
1023         } else {
1024                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
1025         }
1026         if (!q)
1027                 goto out;
1028         cops = q->ops->cl_ops;
1029         if (!cops)
1030                 goto out;
1031         if (!cops->tcf_block)
1032                 goto out;
1033         if (TC_H_MIN(tcm->tcm_parent)) {
1034                 cl = cops->find(q, tcm->tcm_parent);
1035                 if (cl == 0)
1036                         goto out;
1037         }
1038         block = cops->tcf_block(q, cl);
1039         if (!block)
1040                 goto out;
1041
1042         index_start = cb->args[0];
1043         index = 0;
1044
1045         list_for_each_entry(chain, &block->chain_list, list) {
1046                 if (tca[TCA_CHAIN] &&
1047                     nla_get_u32(tca[TCA_CHAIN]) != chain->index)
1048                         continue;
1049                 if (!tcf_chain_dump(chain, q, parent, skb, cb,
1050                                     index_start, &index))
1051                         break;
1052         }
1053
1054         cb->args[0] = index;
1055
1056 out:
1057         return skb->len;
1058 }
1059
1060 void tcf_exts_destroy(struct tcf_exts *exts)
1061 {
1062 #ifdef CONFIG_NET_CLS_ACT
1063         LIST_HEAD(actions);
1064
1065         ASSERT_RTNL();
1066         tcf_exts_to_list(exts, &actions);
1067         tcf_action_destroy(&actions, TCA_ACT_UNBIND);
1068         kfree(exts->actions);
1069         exts->nr_actions = 0;
1070 #endif
1071 }
1072 EXPORT_SYMBOL(tcf_exts_destroy);
1073
1074 int tcf_exts_validate(struct net *net, struct tcf_proto *tp, struct nlattr **tb,
1075                       struct nlattr *rate_tlv, struct tcf_exts *exts, bool ovr)
1076 {
1077 #ifdef CONFIG_NET_CLS_ACT
1078         {
1079                 struct tc_action *act;
1080
1081                 if (exts->police && tb[exts->police]) {
1082                         act = tcf_action_init_1(net, tp, tb[exts->police],
1083                                                 rate_tlv, "police", ovr,
1084                                                 TCA_ACT_BIND);
1085                         if (IS_ERR(act))
1086                                 return PTR_ERR(act);
1087
1088                         act->type = exts->type = TCA_OLD_COMPAT;
1089                         exts->actions[0] = act;
1090                         exts->nr_actions = 1;
1091                 } else if (exts->action && tb[exts->action]) {
1092                         LIST_HEAD(actions);
1093                         int err, i = 0;
1094
1095                         err = tcf_action_init(net, tp, tb[exts->action],
1096                                               rate_tlv, NULL, ovr, TCA_ACT_BIND,
1097                                               &actions);
1098                         if (err)
1099                                 return err;
1100                         list_for_each_entry(act, &actions, list)
1101                                 exts->actions[i++] = act;
1102                         exts->nr_actions = i;
1103                 }
1104         }
1105 #else
1106         if ((exts->action && tb[exts->action]) ||
1107             (exts->police && tb[exts->police]))
1108                 return -EOPNOTSUPP;
1109 #endif
1110
1111         return 0;
1112 }
1113 EXPORT_SYMBOL(tcf_exts_validate);
1114
1115 void tcf_exts_change(struct tcf_exts *dst, struct tcf_exts *src)
1116 {
1117 #ifdef CONFIG_NET_CLS_ACT
1118         struct tcf_exts old = *dst;
1119
1120         *dst = *src;
1121         tcf_exts_destroy(&old);
1122 #endif
1123 }
1124 EXPORT_SYMBOL(tcf_exts_change);
1125
1126 #ifdef CONFIG_NET_CLS_ACT
1127 static struct tc_action *tcf_exts_first_act(struct tcf_exts *exts)
1128 {
1129         if (exts->nr_actions == 0)
1130                 return NULL;
1131         else
1132                 return exts->actions[0];
1133 }
1134 #endif
1135
1136 int tcf_exts_dump(struct sk_buff *skb, struct tcf_exts *exts)
1137 {
1138 #ifdef CONFIG_NET_CLS_ACT
1139         struct nlattr *nest;
1140
1141         if (exts->action && tcf_exts_has_actions(exts)) {
1142                 /*
1143                  * again for backward compatible mode - we want
1144                  * to work with both old and new modes of entering
1145                  * tc data even if iproute2  was newer - jhs
1146                  */
1147                 if (exts->type != TCA_OLD_COMPAT) {
1148                         LIST_HEAD(actions);
1149
1150                         nest = nla_nest_start(skb, exts->action);
1151                         if (nest == NULL)
1152                                 goto nla_put_failure;
1153
1154                         tcf_exts_to_list(exts, &actions);
1155                         if (tcf_action_dump(skb, &actions, 0, 0) < 0)
1156                                 goto nla_put_failure;
1157                         nla_nest_end(skb, nest);
1158                 } else if (exts->police) {
1159                         struct tc_action *act = tcf_exts_first_act(exts);
1160                         nest = nla_nest_start(skb, exts->police);
1161                         if (nest == NULL || !act)
1162                                 goto nla_put_failure;
1163                         if (tcf_action_dump_old(skb, act, 0, 0) < 0)
1164                                 goto nla_put_failure;
1165                         nla_nest_end(skb, nest);
1166                 }
1167         }
1168         return 0;
1169
1170 nla_put_failure:
1171         nla_nest_cancel(skb, nest);
1172         return -1;
1173 #else
1174         return 0;
1175 #endif
1176 }
1177 EXPORT_SYMBOL(tcf_exts_dump);
1178
1179
1180 int tcf_exts_dump_stats(struct sk_buff *skb, struct tcf_exts *exts)
1181 {
1182 #ifdef CONFIG_NET_CLS_ACT
1183         struct tc_action *a = tcf_exts_first_act(exts);
1184         if (a != NULL && tcf_action_copy_stats(skb, a, 1) < 0)
1185                 return -1;
1186 #endif
1187         return 0;
1188 }
1189 EXPORT_SYMBOL(tcf_exts_dump_stats);
1190
1191 static int tc_exts_setup_cb_egdev_call(struct tcf_exts *exts,
1192                                        enum tc_setup_type type,
1193                                        void *type_data, bool err_stop)
1194 {
1195         int ok_count = 0;
1196 #ifdef CONFIG_NET_CLS_ACT
1197         const struct tc_action *a;
1198         struct net_device *dev;
1199         int i, ret;
1200
1201         if (!tcf_exts_has_actions(exts))
1202                 return 0;
1203
1204         for (i = 0; i < exts->nr_actions; i++) {
1205                 a = exts->actions[i];
1206                 if (!a->ops->get_dev)
1207                         continue;
1208                 dev = a->ops->get_dev(a);
1209                 if (!dev || !tc_can_offload(dev))
1210                         continue;
1211                 ret = tc_setup_cb_egdev_call(dev, type, type_data, err_stop);
1212                 if (ret < 0)
1213                         return ret;
1214                 ok_count += ret;
1215         }
1216 #endif
1217         return ok_count;
1218 }
1219
1220 int tc_setup_cb_call(struct tcf_block *block, struct tcf_exts *exts,
1221                      enum tc_setup_type type, void *type_data, bool err_stop)
1222 {
1223         int ok_count;
1224         int ret;
1225
1226         ret = tcf_block_cb_call(block, type, type_data, err_stop);
1227         if (ret < 0)
1228                 return ret;
1229         ok_count = ret;
1230
1231         if (!exts)
1232                 return ok_count;
1233         ret = tc_exts_setup_cb_egdev_call(exts, type, type_data, err_stop);
1234         if (ret < 0)
1235                 return ret;
1236         ok_count += ret;
1237
1238         return ok_count;
1239 }
1240 EXPORT_SYMBOL(tc_setup_cb_call);
1241
1242 static int __init tc_filter_init(void)
1243 {
1244         tc_filter_wq = alloc_ordered_workqueue("tc_filter_workqueue", 0);
1245         if (!tc_filter_wq)
1246                 return -ENOMEM;
1247
1248         rtnl_register(PF_UNSPEC, RTM_NEWTFILTER, tc_ctl_tfilter, NULL, 0);
1249         rtnl_register(PF_UNSPEC, RTM_DELTFILTER, tc_ctl_tfilter, NULL, 0);
1250         rtnl_register(PF_UNSPEC, RTM_GETTFILTER, tc_ctl_tfilter,
1251                       tc_dump_tfilter, 0);
1252
1253         return 0;
1254 }
1255
1256 subsys_initcall(tc_filter_init);