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/slab.h>
  27 #include <net/net_namespace.h>
  28 #include <net/sock.h>
  29 #include <net/netlink.h>
  30 #include <net/pkt_sched.h>
  31 #include <net/pkt_cls.h>
  32
  33 /* The list of all installed classifier types */
  34 static LIST_HEAD(tcf_proto_base);
  35
  36 /* Protects list of registered TC modules. It is pure SMP lock. */
  37 static DEFINE_RWLOCK(cls_mod_lock);
  38
  39 /* Find classifier type by string name */
  40
  41 static const struct tcf_proto_ops *tcf_proto_lookup_ops(const char *kind)
  42 {
  43         const struct tcf_proto_ops *t, *res = NULL;
  44
  45         if (kind) {
  46                 read_lock(&cls_mod_lock);
  47                 list_for_each_entry(t, &tcf_proto_base, head) {
  48                         if (strcmp(kind, t->kind) == 0) {
  49                                 if (try_module_get(t->owner))
  50                                         res = t;
  51                                 break;
  52                         }
  53                 }
  54                 read_unlock(&cls_mod_lock);
  55         }
  56         return res;
  57 }
  58
  59 /* Register(unregister) new classifier type */
  60
  61 int register_tcf_proto_ops(struct tcf_proto_ops *ops)
  62 {
  63         struct tcf_proto_ops *t;
  64         int rc = -EEXIST;
  65
  66         write_lock(&cls_mod_lock);
  67         list_for_each_entry(t, &tcf_proto_base, head)
  68                 if (!strcmp(ops->kind, t->kind))
  69                         goto out;
  70
  71         list_add_tail(&ops->head, &tcf_proto_base);
  72         rc = 0;
  73 out:
  74         write_unlock(&cls_mod_lock);
  75         return rc;
  76 }
  77 EXPORT_SYMBOL(register_tcf_proto_ops);
  78
  79 static struct workqueue_struct *tc_filter_wq;
  80
  81 int unregister_tcf_proto_ops(struct tcf_proto_ops *ops)
  82 {
  83         struct tcf_proto_ops *t;
  84         int rc = -ENOENT;
  85
  86         /* Wait for outstanding call_rcu()s, if any, from a
  87          * tcf_proto_ops's destroy() handler.
  88          */
  89         rcu_barrier();
  90         flush_workqueue(tc_filter_wq);
  91
  92         write_lock(&cls_mod_lock);
  93         list_for_each_entry(t, &tcf_proto_base, head) {
  94                 if (t == ops) {
  95                         list_del(&t->head);
  96                         rc = 0;
  97                         break;
  98                 }
  99         }
 100         write_unlock(&cls_mod_lock);
 101         return rc;
 102 }
 103 EXPORT_SYMBOL(unregister_tcf_proto_ops);
 104
 105 bool tcf_queue_work(struct work_struct *work)
 106 {
 107         return queue_work(tc_filter_wq, work);
 108 }
 109 EXPORT_SYMBOL(tcf_queue_work);
 110
 111 /* Select new prio value from the range, managed by kernel. */
 112
 113 static inline u32 tcf_auto_prio(struct tcf_proto *tp)
 114 {
 115         u32 first = TC_H_MAKE(0xC0000000U, 0U);
 116
 117         if (tp)
 118                 first = tp->prio - 1;
 119
 120         return TC_H_MAJ(first);
 121 }
 122
 123 static struct tcf_proto *tcf_proto_create(const char *kind, u32 protocol,
 124                                           u32 prio, u32 parent, struct Qdisc *q,
 125                                           struct tcf_chain *chain)
 126 {
 127         struct tcf_proto *tp;
 128         int err;
 129
 130         tp = kzalloc(sizeof(*tp), GFP_KERNEL);
 131         if (!tp)
 132                 return ERR_PTR(-ENOBUFS);
 133
 134         err = -ENOENT;
 135         tp->ops = tcf_proto_lookup_ops(kind);
 136         if (!tp->ops) {
 137 #ifdef CONFIG_MODULES
 138                 rtnl_unlock();
 139                 request_module("cls_%s", kind);
 140                 rtnl_lock();
 141                 tp->ops = tcf_proto_lookup_ops(kind);
 142                 /* We dropped the RTNL semaphore in order to perform
 143                  * the module load. So, even if we succeeded in loading
 144                  * the module we have to replay the request. We indicate
 145                  * this using -EAGAIN.
 146                  */
 147                 if (tp->ops) {
 148                         module_put(tp->ops->owner);
 149                         err = -EAGAIN;
 150                 } else {
 151                         err = -ENOENT;
 152                 }
 153                 goto errout;
 154 #endif
 155         }
 156         tp->classify = tp->ops->classify;
 157         tp->protocol = protocol;
 158         tp->prio = prio;
 159         tp->classid = parent;
 160         tp->q = q;
 161         tp->chain = chain;
 162
 163         err = tp->ops->init(tp);
 164         if (err) {
 165                 module_put(tp->ops->owner);
 166                 goto errout;
 167         }
 168         return tp;
 169
 170 errout:
 171         kfree(tp);
 172         return ERR_PTR(err);
 173 }
 174
 175 static void tcf_proto_destroy(struct tcf_proto *tp)
 176 {
 177         tp->ops->destroy(tp);
 178         module_put(tp->ops->owner);
 179         kfree_rcu(tp, rcu);
 180 }
 181
 182 static struct tcf_chain *tcf_chain_create(struct tcf_block *block,
 183                                           u32 chain_index)
 184 {
 185         struct tcf_chain *chain;
 186
 187         chain = kzalloc(sizeof(*chain), GFP_KERNEL);
 188         if (!chain)
 189                 return NULL;
 190         list_add_tail(&chain->list, &block->chain_list);
 191         chain->block = block;
 192         chain->index = chain_index;
 193         chain->refcnt = 1;
 194         return chain;
 195 }
 196
 197 static void tcf_chain_head_change(struct tcf_chain *chain,
 198                                   struct tcf_proto *tp_head)
 199 {
 200         if (chain->chain_head_change)
 201                 chain->chain_head_change(tp_head,
 202                                          chain->chain_head_change_priv);
 203 }
 204
 205 static void tcf_chain_flush(struct tcf_chain *chain)
 206 {
 207         struct tcf_proto *tp = rtnl_dereference(chain->filter_chain);
 208
 209         tcf_chain_head_change(chain, NULL);
 210         while (tp) {
 211                 RCU_INIT_POINTER(chain->filter_chain, tp->next);
 212                 tcf_proto_destroy(tp);
 213                 tp = rtnl_dereference(chain->filter_chain);
 214                 tcf_chain_put(chain);
 215         }
 216 }
 217
 218 static void tcf_chain_destroy(struct tcf_chain *chain)
 219 {
 220         list_del(&chain->list);
 221         kfree(chain);
 222 }
 223
 224 static void tcf_chain_hold(struct tcf_chain *chain)
 225 {
 226         ++chain->refcnt;
 227 }
 228
 229 struct tcf_chain *tcf_chain_get(struct tcf_block *block, u32 chain_index,
 230                                 bool create)
 231 {
 232         struct tcf_chain *chain;
 233
 234         list_for_each_entry(chain, &block->chain_list, list) {
 235                 if (chain->index == chain_index) {
 236                         tcf_chain_hold(chain);
 237                         return chain;
 238                 }
 239         }
 240
 241         return create ? tcf_chain_create(block, chain_index) : NULL;
 242 }
 243 EXPORT_SYMBOL(tcf_chain_get);
 244
 245 void tcf_chain_put(struct tcf_chain *chain)
 246 {
 247         if (--chain->refcnt == 0)
 248                 tcf_chain_destroy(chain);
 249 }
 250 EXPORT_SYMBOL(tcf_chain_put);
 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, struct Qdisc *q,
 280                       struct tcf_block_ext_info *ei)
 281 {
 282         struct tcf_block *block = kzalloc(sizeof(*block), GFP_KERNEL);
 283         struct tcf_chain *chain;
 284         int err;
 285
 286         if (!block)
 287                 return -ENOMEM;
 288         INIT_LIST_HEAD(&block->chain_list);
 289         INIT_LIST_HEAD(&block->cb_list);
 290
 291         /* Create chain 0 by default, it has to be always present. */
 292         chain = tcf_chain_create(block, 0);
 293         if (!chain) {
 294                 err = -ENOMEM;
 295                 goto err_chain_create;
 296         }
 297         WARN_ON(!ei->chain_head_change);
 298         chain->chain_head_change = ei->chain_head_change;
 299         chain->chain_head_change_priv = ei->chain_head_change_priv;
 300         block->net = qdisc_net(q);
 301         block->q = q;
 302         tcf_block_offload_bind(block, q, ei);
 303         *p_block = block;
 304         return 0;
 305
 306 err_chain_create:
 307         kfree(block);
 308         return err;
 309 }
 310 EXPORT_SYMBOL(tcf_block_get_ext);
 311
 312 static void tcf_chain_head_change_dflt(struct tcf_proto *tp_head, void *priv)
 313 {
 314         struct tcf_proto __rcu **p_filter_chain = priv;
 315
 316         rcu_assign_pointer(*p_filter_chain, tp_head);
 317 }
 318
 319 int tcf_block_get(struct tcf_block **p_block,
 320                   struct tcf_proto __rcu **p_filter_chain, struct Qdisc *q)
 321 {
 322         struct tcf_block_ext_info ei = {
 323                 .chain_head_change = tcf_chain_head_change_dflt,
 324                 .chain_head_change_priv = p_filter_chain,
 325         };
 326
 327         WARN_ON(!p_filter_chain);
 328         return tcf_block_get_ext(p_block, q, &ei);
 329 }
 330 EXPORT_SYMBOL(tcf_block_get);
 331
 332 static void tcf_block_put_final(struct work_struct *work)
 333 {
 334         struct tcf_block *block = container_of(work, struct tcf_block, work);
 335         struct tcf_chain *chain, *tmp;
 336
 337         rtnl_lock();
 338
 339         /* At this point, all the chains should have refcnt == 1. */
 340         list_for_each_entry_safe(chain, tmp, &block->chain_list, list)
 341                 tcf_chain_put(chain);
 342         rtnl_unlock();
 343         kfree(block);
 344 }
 345
 346 /* XXX: Standalone actions are not allowed to jump to any chain, and bound
 347  * actions should be all removed after flushing.
 348  */
 349 void tcf_block_put_ext(struct tcf_block *block, struct Qdisc *q,
 350                        struct tcf_block_ext_info *ei)
 351 {
 352         struct tcf_chain *chain;
 353
 354         if (!block)
 355                 return;
 356         /* Hold a refcnt for all chains, except 0, so that they don't disappear
 357          * while we are iterating.
 358          */
 359         list_for_each_entry(chain, &block->chain_list, list)
 360                 if (chain->index)
 361                         tcf_chain_hold(chain);
 362
 363         list_for_each_entry(chain, &block->chain_list, list)
 364                 tcf_chain_flush(chain);
 365
 366         tcf_block_offload_unbind(block, q, ei);
 367
 368         INIT_WORK(&block->work, tcf_block_put_final);
 369         /* Wait for existing RCU callbacks to cool down, make sure their works
 370          * have been queued before this. We can not flush pending works here
 371          * because we are holding the RTNL lock.
 372          */
 373         rcu_barrier();
 374         tcf_queue_work(&block->work);
 375 }
 376 EXPORT_SYMBOL(tcf_block_put_ext);
 377
 378 void tcf_block_put(struct tcf_block *block)
 379 {
 380         struct tcf_block_ext_info ei = {0, };
 381
 382         if (!block)
 383                 return;
 384         tcf_block_put_ext(block, block->q, &ei);
 385 }
 386
 387 EXPORT_SYMBOL(tcf_block_put);
 388
 389 struct tcf_block_cb {
 390         struct list_head list;
 391         tc_setup_cb_t *cb;
 392         void *cb_ident;
 393         void *cb_priv;
 394         unsigned int refcnt;
 395 };
 396
 397 void *tcf_block_cb_priv(struct tcf_block_cb *block_cb)
 398 {
 399         return block_cb->cb_priv;
 400 }
 401 EXPORT_SYMBOL(tcf_block_cb_priv);
 402
 403 struct tcf_block_cb *tcf_block_cb_lookup(struct tcf_block *block,
 404                                          tc_setup_cb_t *cb, void *cb_ident)
 405 {       struct tcf_block_cb *block_cb;
 406
 407         list_for_each_entry(block_cb, &block->cb_list, list)
 408                 if (block_cb->cb == cb && block_cb->cb_ident == cb_ident)
 409                         return block_cb;
 410         return NULL;
 411 }
 412 EXPORT_SYMBOL(tcf_block_cb_lookup);
 413
 414 void tcf_block_cb_incref(struct tcf_block_cb *block_cb)
 415 {
 416         block_cb->refcnt++;
 417 }
 418 EXPORT_SYMBOL(tcf_block_cb_incref);
 419
 420 unsigned int tcf_block_cb_decref(struct tcf_block_cb *block_cb)
 421 {
 422         return --block_cb->refcnt;
 423 }
 424 EXPORT_SYMBOL(tcf_block_cb_decref);
 425
 426 struct tcf_block_cb *__tcf_block_cb_register(struct tcf_block *block,
 427                                              tc_setup_cb_t *cb, void *cb_ident,
 428                                              void *cb_priv)
 429 {
 430         struct tcf_block_cb *block_cb;
 431
 432         block_cb = kzalloc(sizeof(*block_cb), GFP_KERNEL);
 433         if (!block_cb)
 434                 return NULL;
 435         block_cb->cb = cb;
 436         block_cb->cb_ident = cb_ident;
 437         block_cb->cb_priv = cb_priv;
 438         list_add(&block_cb->list, &block->cb_list);
 439         return block_cb;
 440 }
 441 EXPORT_SYMBOL(__tcf_block_cb_register);
 442
 443 int tcf_block_cb_register(struct tcf_block *block,
 444                           tc_setup_cb_t *cb, void *cb_ident,
 445                           void *cb_priv)
 446 {
 447         struct tcf_block_cb *block_cb;
 448
 449         block_cb = __tcf_block_cb_register(block, cb, cb_ident, cb_priv);
 450         return block_cb ? 0 : -ENOMEM;
 451 }
 452 EXPORT_SYMBOL(tcf_block_cb_register);
 453
 454 void __tcf_block_cb_unregister(struct tcf_block_cb *block_cb)
 455 {
 456         list_del(&block_cb->list);
 457         kfree(block_cb);
 458 }
 459 EXPORT_SYMBOL(__tcf_block_cb_unregister);
 460
 461 void tcf_block_cb_unregister(struct tcf_block *block,
 462                              tc_setup_cb_t *cb, void *cb_ident)
 463 {
 464         struct tcf_block_cb *block_cb;
 465
 466         block_cb = tcf_block_cb_lookup(block, cb, cb_ident);
 467         if (!block_cb)
 468                 return;
 469         __tcf_block_cb_unregister(block_cb);
 470 }
 471 EXPORT_SYMBOL(tcf_block_cb_unregister);
 472
 473 static int tcf_block_cb_call(struct tcf_block *block, enum tc_setup_type type,
 474                              void *type_data, bool err_stop)
 475 {
 476         struct tcf_block_cb *block_cb;
 477         int ok_count = 0;
 478         int err;
 479
 480         list_for_each_entry(block_cb, &block->cb_list, list) {
 481                 err = block_cb->cb(type, type_data, block_cb->cb_priv);
 482                 if (err) {
 483                         if (err_stop)
 484                                 return err;
 485                 } else {
 486                         ok_count++;
 487                 }
 488         }
 489         return ok_count;
 490 }
 491
 492 /* Main classifier routine: scans classifier chain attached
 493  * to this qdisc, (optionally) tests for protocol and asks
 494  * specific classifiers.
 495  */
 496 int tcf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 497                  struct tcf_result *res, bool compat_mode)
 498 {
 499         __be16 protocol = tc_skb_protocol(skb);
 500 #ifdef CONFIG_NET_CLS_ACT
 501         const int max_reclassify_loop = 4;
 502         const struct tcf_proto *orig_tp = tp;
 503         const struct tcf_proto *first_tp;
 504         int limit = 0;
 505
 506 reclassify:
 507 #endif
 508         for (; tp; tp = rcu_dereference_bh(tp->next)) {
 509                 int err;
 510
 511                 if (tp->protocol != protocol &&
 512                     tp->protocol != htons(ETH_P_ALL))
 513                         continue;
 514
 515                 err = tp->classify(skb, tp, res);
 516 #ifdef CONFIG_NET_CLS_ACT
 517                 if (unlikely(err == TC_ACT_RECLASSIFY && !compat_mode)) {
 518                         first_tp = orig_tp;
 519                         goto reset;
 520                 } else if (unlikely(TC_ACT_EXT_CMP(err, TC_ACT_GOTO_CHAIN))) {
 521                         first_tp = res->goto_tp;
 522                         goto reset;
 523                 }
 524 #endif
 525                 if (err >= 0)
 526                         return err;
 527         }
 528
 529         return TC_ACT_UNSPEC; /* signal: continue lookup */
 530 #ifdef CONFIG_NET_CLS_ACT
 531 reset:
 532         if (unlikely(limit++ >= max_reclassify_loop)) {
 533                 net_notice_ratelimited("%s: reclassify loop, rule prio %u, protocol %02x\n",
 534                                        tp->q->ops->id, tp->prio & 0xffff,
 535                                        ntohs(tp->protocol));
 536                 return TC_ACT_SHOT;
 537         }
 538
 539         tp = first_tp;
 540         protocol = tc_skb_protocol(skb);
 541         goto reclassify;
 542 #endif
 543 }
 544 EXPORT_SYMBOL(tcf_classify);
 545
 546 struct tcf_chain_info {
 547         struct tcf_proto __rcu **pprev;
 548         struct tcf_proto __rcu *next;
 549 };
 550
 551 static struct tcf_proto *tcf_chain_tp_prev(struct tcf_chain_info *chain_info)
 552 {
 553         return rtnl_dereference(*chain_info->pprev);
 554 }
 555
 556 static void tcf_chain_tp_insert(struct tcf_chain *chain,
 557                                 struct tcf_chain_info *chain_info,
 558                                 struct tcf_proto *tp)
 559 {
 560         if (*chain_info->pprev == chain->filter_chain)
 561                 tcf_chain_head_change(chain, tp);
 562         RCU_INIT_POINTER(tp->next, tcf_chain_tp_prev(chain_info));
 563         rcu_assign_pointer(*chain_info->pprev, tp);
 564         tcf_chain_hold(chain);
 565 }
 566
 567 static void tcf_chain_tp_remove(struct tcf_chain *chain,
 568                                 struct tcf_chain_info *chain_info,
 569                                 struct tcf_proto *tp)
 570 {
 571         struct tcf_proto *next = rtnl_dereference(chain_info->next);
 572
 573         if (tp == chain->filter_chain)
 574                 tcf_chain_head_change(chain, next);
 575         RCU_INIT_POINTER(*chain_info->pprev, next);
 576         tcf_chain_put(chain);
 577 }
 578
 579 static struct tcf_proto *tcf_chain_tp_find(struct tcf_chain *chain,
 580                                            struct tcf_chain_info *chain_info,
 581                                            u32 protocol, u32 prio,
 582                                            bool prio_allocate)
 583 {
 584         struct tcf_proto **pprev;
 585         struct tcf_proto *tp;
 586
 587         /* Check the chain for existence of proto-tcf with this priority */
 588         for (pprev = &chain->filter_chain;
 589              (tp = rtnl_dereference(*pprev)); pprev = &tp->next) {
 590                 if (tp->prio >= prio) {
 591                         if (tp->prio == prio) {
 592                                 if (prio_allocate ||
 593                                     (tp->protocol != protocol && protocol))
 594                                         return ERR_PTR(-EINVAL);
 595                         } else {
 596                                 tp = NULL;
 597                         }
 598                         break;
 599                 }
 600         }
 601         chain_info->pprev = pprev;
 602         chain_info->next = tp ? tp->next : NULL;
 603         return tp;
 604 }
 605
 606 static int tcf_fill_node(struct net *net, struct sk_buff *skb,
 607                          struct tcf_proto *tp, struct Qdisc *q, u32 parent,
 608                          void *fh, u32 portid, u32 seq, u16 flags, int event)
 609 {
 610         struct tcmsg *tcm;
 611         struct nlmsghdr  *nlh;
 612         unsigned char *b = skb_tail_pointer(skb);
 613
 614         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*tcm), flags);
 615         if (!nlh)
 616                 goto out_nlmsg_trim;
 617         tcm = nlmsg_data(nlh);
 618         tcm->tcm_family = AF_UNSPEC;
 619         tcm->tcm__pad1 = 0;
 620         tcm->tcm__pad2 = 0;
 621         tcm->tcm_ifindex = qdisc_dev(q)->ifindex;
 622         tcm->tcm_parent = parent;
 623         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 624         if (nla_put_string(skb, TCA_KIND, tp->ops->kind))
 625                 goto nla_put_failure;
 626         if (nla_put_u32(skb, TCA_CHAIN, tp->chain->index))
 627                 goto nla_put_failure;
 628         if (!fh) {
 629                 tcm->tcm_handle = 0;
 630         } else {
 631                 if (tp->ops->dump && tp->ops->dump(net, tp, fh, skb, tcm) < 0)
 632                         goto nla_put_failure;
 633         }
 634         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 635         return skb->len;
 636
 637 out_nlmsg_trim:
 638 nla_put_failure:
 639         nlmsg_trim(skb, b);
 640         return -1;
 641 }
 642
 643 static int tfilter_notify(struct net *net, struct sk_buff *oskb,
 644                           struct nlmsghdr *n, struct tcf_proto *tp,
 645                           struct Qdisc *q, u32 parent,
 646                           void *fh, int event, bool unicast)
 647 {
 648         struct sk_buff *skb;
 649         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 650
 651         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 652         if (!skb)
 653                 return -ENOBUFS;
 654
 655         if (tcf_fill_node(net, skb, tp, q, parent, fh, portid, n->nlmsg_seq,
 656                           n->nlmsg_flags, event) <= 0) {
 657                 kfree_skb(skb);
 658                 return -EINVAL;
 659         }
 660
 661         if (unicast)
 662                 return netlink_unicast(net->rtnl, skb, portid, MSG_DONTWAIT);
 663
 664         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 665                               n->nlmsg_flags & NLM_F_ECHO);
 666 }
 667
 668 static int tfilter_del_notify(struct net *net, struct sk_buff *oskb,
 669                               struct nlmsghdr *n, struct tcf_proto *tp,
 670                               struct Qdisc *q, u32 parent,
 671                               void *fh, bool unicast, bool *last)
 672 {
 673         struct sk_buff *skb;
 674         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 675         int err;
 676
 677         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 678         if (!skb)
 679                 return -ENOBUFS;
 680
 681         if (tcf_fill_node(net, skb, tp, q, parent, fh, portid, n->nlmsg_seq,
 682                           n->nlmsg_flags, RTM_DELTFILTER) <= 0) {
 683                 kfree_skb(skb);
 684                 return -EINVAL;
 685         }
 686
 687         err = tp->ops->delete(tp, fh, last);
 688         if (err) {
 689                 kfree_skb(skb);
 690                 return err;
 691         }
 692
 693         if (unicast)
 694                 return netlink_unicast(net->rtnl, skb, portid, MSG_DONTWAIT);
 695
 696         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 697                               n->nlmsg_flags & NLM_F_ECHO);
 698 }
 699
 700 static void tfilter_notify_chain(struct net *net, struct sk_buff *oskb,
 701                                  struct Qdisc *q, u32 parent,
 702                                  struct nlmsghdr *n,
 703                                  struct tcf_chain *chain, int event)
 704 {
 705         struct tcf_proto *tp;
 706
 707         for (tp = rtnl_dereference(chain->filter_chain);
 708              tp; tp = rtnl_dereference(tp->next))
 709                 tfilter_notify(net, oskb, n, tp, q, parent, 0, event, false);
 710 }
 711
 712 /* Add/change/delete/get a filter node */
 713
 714 static int tc_ctl_tfilter(struct sk_buff *skb, struct nlmsghdr *n,
 715                           struct netlink_ext_ack *extack)
 716 {
 717         struct net *net = sock_net(skb->sk);
 718         struct nlattr *tca[TCA_MAX + 1];
 719         struct tcmsg *t;
 720         u32 protocol;
 721         u32 prio;
 722         bool prio_allocate;
 723         u32 parent;
 724         u32 chain_index;
 725         struct net_device *dev;
 726         struct Qdisc  *q;
 727         struct tcf_chain_info chain_info;
 728         struct tcf_chain *chain = NULL;
 729         struct tcf_block *block;
 730         struct tcf_proto *tp;
 731         const struct Qdisc_class_ops *cops;
 732         unsigned long cl;
 733         void *fh;
 734         int err;
 735         int tp_created;
 736
 737         if ((n->nlmsg_type != RTM_GETTFILTER) &&
 738             !netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN))
 739                 return -EPERM;
 740
 741 replay:
 742         tp_created = 0;
 743
 744         err = nlmsg_parse(n, sizeof(*t), tca, TCA_MAX, NULL, extack);
 745         if (err < 0)
 746                 return err;
 747
 748         t = nlmsg_data(n);
 749         protocol = TC_H_MIN(t->tcm_info);
 750         prio = TC_H_MAJ(t->tcm_info);
 751         prio_allocate = false;
 752         parent = t->tcm_parent;
 753         cl = 0;
 754
 755         if (prio == 0) {
 756                 switch (n->nlmsg_type) {
 757                 case RTM_DELTFILTER:
 758                         if (protocol || t->tcm_handle || tca[TCA_KIND])
 759                                 return -ENOENT;
 760                         break;
 761                 case RTM_NEWTFILTER:
 762                         /* If no priority is provided by the user,
 763                          * we allocate one.
 764                          */
 765                         if (n->nlmsg_flags & NLM_F_CREATE) {
 766                                 prio = TC_H_MAKE(0x80000000U, 0U);
 767                                 prio_allocate = true;
 768                                 break;
 769                         }
 770                         /* fall-through */
 771                 default:
 772                         return -ENOENT;
 773                 }
 774         }
 775
 776         /* Find head of filter chain. */
 777
 778         /* Find link */
 779         dev = __dev_get_by_index(net, t->tcm_ifindex);
 780         if (dev == NULL)
 781                 return -ENODEV;
 782
 783         /* Find qdisc */
 784         if (!parent) {
 785                 q = dev->qdisc;
 786                 parent = q->handle;
 787         } else {
 788                 q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent));
 789                 if (q == NULL)
 790                         return -EINVAL;
 791         }
 792
 793         /* Is it classful? */
 794         cops = q->ops->cl_ops;
 795         if (!cops)
 796                 return -EINVAL;
 797
 798         if (!cops->tcf_block)
 799                 return -EOPNOTSUPP;
 800
 801         /* Do we search for filter, attached to class? */
 802         if (TC_H_MIN(parent)) {
 803                 cl = cops->find(q, parent);
 804                 if (cl == 0)
 805                         return -ENOENT;
 806         }
 807
 808         /* And the last stroke */
 809         block = cops->tcf_block(q, cl);
 810         if (!block) {
 811                 err = -EINVAL;
 812                 goto errout;
 813         }
 814
 815         chain_index = tca[TCA_CHAIN] ? nla_get_u32(tca[TCA_CHAIN]) : 0;
 816         if (chain_index > TC_ACT_EXT_VAL_MASK) {
 817                 err = -EINVAL;
 818                 goto errout;
 819         }
 820         chain = tcf_chain_get(block, chain_index,
 821                               n->nlmsg_type == RTM_NEWTFILTER);
 822         if (!chain) {
 823                 err = n->nlmsg_type == RTM_NEWTFILTER ? -ENOMEM : -EINVAL;
 824                 goto errout;
 825         }
 826
 827         if (n->nlmsg_type == RTM_DELTFILTER && prio == 0) {
 828                 tfilter_notify_chain(net, skb, q, parent, n,
 829                                      chain, RTM_DELTFILTER);
 830                 tcf_chain_flush(chain);
 831                 err = 0;
 832                 goto errout;
 833         }
 834
 835         tp = tcf_chain_tp_find(chain, &chain_info, protocol,
 836                                prio, prio_allocate);
 837         if (IS_ERR(tp)) {
 838                 err = PTR_ERR(tp);
 839                 goto errout;
 840         }
 841
 842         if (tp == NULL) {
 843                 /* Proto-tcf does not exist, create new one */
 844
 845                 if (tca[TCA_KIND] == NULL || !protocol) {
 846                         err = -EINVAL;
 847                         goto errout;
 848                 }
 849
 850                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 851                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 852                         err = -ENOENT;
 853                         goto errout;
 854                 }
 855
 856                 if (prio_allocate)
 857                         prio = tcf_auto_prio(tcf_chain_tp_prev(&chain_info));
 858
 859                 tp = tcf_proto_create(nla_data(tca[TCA_KIND]),
 860                                       protocol, prio, parent, q, chain);
 861                 if (IS_ERR(tp)) {
 862                         err = PTR_ERR(tp);
 863                         goto errout;
 864                 }
 865                 tp_created = 1;
 866         } else if (tca[TCA_KIND] && nla_strcmp(tca[TCA_KIND], tp->ops->kind)) {
 867                 err = -EINVAL;
 868                 goto errout;
 869         }
 870
 871         fh = tp->ops->get(tp, t->tcm_handle);
 872
 873         if (!fh) {
 874                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 875                         tcf_chain_tp_remove(chain, &chain_info, tp);
 876                         tfilter_notify(net, skb, n, tp, q, parent, fh,
 877                                        RTM_DELTFILTER, false);
 878                         tcf_proto_destroy(tp);
 879                         err = 0;
 880                         goto errout;
 881                 }
 882
 883                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 884                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 885                         err = -ENOENT;
 886                         goto errout;
 887                 }
 888         } else {
 889                 bool last;
 890
 891                 switch (n->nlmsg_type) {
 892                 case RTM_NEWTFILTER:
 893                         if (n->nlmsg_flags & NLM_F_EXCL) {
 894                                 if (tp_created)
 895                                         tcf_proto_destroy(tp);
 896                                 err = -EEXIST;
 897                                 goto errout;
 898                         }
 899                         break;
 900                 case RTM_DELTFILTER:
 901                         err = tfilter_del_notify(net, skb, n, tp, q, parent,
 902                                                  fh, false, &last);
 903                         if (err)
 904                                 goto errout;
 905                         if (last) {
 906                                 tcf_chain_tp_remove(chain, &chain_info, tp);
 907                                 tcf_proto_destroy(tp);
 908                         }
 909                         goto errout;
 910                 case RTM_GETTFILTER:
 911                         err = tfilter_notify(net, skb, n, tp, q, parent, fh,
 912                                              RTM_NEWTFILTER, true);
 913                         goto errout;
 914                 default:
 915                         err = -EINVAL;
 916                         goto errout;
 917                 }
 918         }
 919
 920         err = tp->ops->change(net, skb, tp, cl, t->tcm_handle, tca, &fh,
 921                               n->nlmsg_flags & NLM_F_CREATE ? TCA_ACT_NOREPLACE : TCA_ACT_REPLACE);
 922         if (err == 0) {
 923                 if (tp_created)
 924                         tcf_chain_tp_insert(chain, &chain_info, tp);
 925                 tfilter_notify(net, skb, n, tp, q, parent, fh,
 926                                RTM_NEWTFILTER, false);
 927         } else {
 928                 if (tp_created)
 929                         tcf_proto_destroy(tp);
 930         }
 931
 932 errout:
 933         if (chain)
 934                 tcf_chain_put(chain);
 935         if (err == -EAGAIN)
 936                 /* Replay the request. */
 937                 goto replay;
 938         return err;
 939 }
 940
 941 struct tcf_dump_args {
 942         struct tcf_walker w;
 943         struct sk_buff *skb;
 944         struct netlink_callback *cb;
 945         struct Qdisc *q;
 946         u32 parent;
 947 };
 948
 949 static int tcf_node_dump(struct tcf_proto *tp, void *n, struct tcf_walker *arg)
 950 {
 951         struct tcf_dump_args *a = (void *)arg;
 952         struct net *net = sock_net(a->skb->sk);
 953
 954         return tcf_fill_node(net, a->skb, tp, a->q, a->parent,
 955                              n, NETLINK_CB(a->cb->skb).portid,
 956                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI,
 957                              RTM_NEWTFILTER);
 958 }
 959
 960 static bool tcf_chain_dump(struct tcf_chain *chain, struct Qdisc *q, u32 parent,
 961                            struct sk_buff *skb, struct netlink_callback *cb,
 962                            long index_start, long *p_index)
 963 {
 964         struct net *net = sock_net(skb->sk);
 965         struct tcmsg *tcm = nlmsg_data(cb->nlh);
 966         struct tcf_dump_args arg;
 967         struct tcf_proto *tp;
 968
 969         for (tp = rtnl_dereference(chain->filter_chain);
 970              tp; tp = rtnl_dereference(tp->next), (*p_index)++) {
 971                 if (*p_index < index_start)
 972                         continue;
 973                 if (TC_H_MAJ(tcm->tcm_info) &&
 974                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 975                         continue;
 976                 if (TC_H_MIN(tcm->tcm_info) &&
 977                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 978                         continue;
 979                 if (*p_index > index_start)
 980                         memset(&cb->args[1], 0,
 981                                sizeof(cb->args) - sizeof(cb->args[0]));
 982                 if (cb->args[1] == 0) {
 983                         if (tcf_fill_node(net, skb, tp, q, parent, 0,
 984                                           NETLINK_CB(cb->skb).portid,
 985                                           cb->nlh->nlmsg_seq, NLM_F_MULTI,
 986                                           RTM_NEWTFILTER) <= 0)
 987                                 return false;
 988
 989                         cb->args[1] = 1;
 990                 }
 991                 if (!tp->ops->walk)
 992                         continue;
 993                 arg.w.fn = tcf_node_dump;
 994                 arg.skb = skb;
 995                 arg.cb = cb;
 996                 arg.q = q;
 997                 arg.parent = parent;
 998                 arg.w.stop = 0;
 999                 arg.w.skip = cb->args[1] - 1;
1000                 arg.w.count = 0;
1001                 tp->ops->walk(tp, &arg.w);
1002                 cb->args[1] = arg.w.count + 1;
1003                 if (arg.w.stop)
1004                         return false;
1005         }
1006         return true;
1007 }
1008
1009 /* called with RTNL */
1010 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
1011 {
1012         struct net *net = sock_net(skb->sk);
1013         struct nlattr *tca[TCA_MAX + 1];
1014         struct net_device *dev;
1015         struct Qdisc *q;
1016         struct tcf_block *block;
1017         struct tcf_chain *chain;
1018         struct tcmsg *tcm = nlmsg_data(cb->nlh);
1019         unsigned long cl = 0;
1020         const struct Qdisc_class_ops *cops;
1021         long index_start;
1022         long index;
1023         u32 parent;
1024         int err;
1025
1026         if (nlmsg_len(cb->nlh) < sizeof(*tcm))
1027                 return skb->len;
1028
1029         err = nlmsg_parse(cb->nlh, sizeof(*tcm), tca, TCA_MAX, NULL, NULL);
1030         if (err)
1031                 return err;
1032
1033         dev = __dev_get_by_index(net, tcm->tcm_ifindex);
1034         if (!dev)
1035                 return skb->len;
1036
1037         parent = tcm->tcm_parent;
1038         if (!parent) {
1039                 q = dev->qdisc;
1040                 parent = q->handle;
1041         } else {
1042                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
1043         }
1044         if (!q)
1045                 goto out;
1046         cops = q->ops->cl_ops;
1047         if (!cops)
1048                 goto out;
1049         if (!cops->tcf_block)
1050                 goto out;
1051         if (TC_H_MIN(tcm->tcm_parent)) {
1052                 cl = cops->find(q, tcm->tcm_parent);
1053                 if (cl == 0)
1054                         goto out;
1055         }
1056         block = cops->tcf_block(q, cl);
1057         if (!block)
1058                 goto out;
1059
1060         index_start = cb->args[0];
1061         index = 0;
1062
1063         list_for_each_entry(chain, &block->chain_list, list) {
1064                 if (tca[TCA_CHAIN] &&
1065                     nla_get_u32(tca[TCA_CHAIN]) != chain->index)
1066                         continue;
1067                 if (!tcf_chain_dump(chain, q, parent, skb, cb,
1068                                     index_start, &index))
1069                         break;
1070         }
1071
1072         cb->args[0] = index;
1073
1074 out:
1075         return skb->len;
1076 }
1077
1078 void tcf_exts_destroy(struct tcf_exts *exts)
1079 {
1080 #ifdef CONFIG_NET_CLS_ACT
1081         LIST_HEAD(actions);
1082
1083         ASSERT_RTNL();
1084         tcf_exts_to_list(exts, &actions);
1085         tcf_action_destroy(&actions, TCA_ACT_UNBIND);
1086         kfree(exts->actions);
1087         exts->nr_actions = 0;
1088 #endif
1089 }
1090 EXPORT_SYMBOL(tcf_exts_destroy);
1091
1092 int tcf_exts_validate(struct net *net, struct tcf_proto *tp, struct nlattr **tb,
1093                       struct nlattr *rate_tlv, struct tcf_exts *exts, bool ovr)
1094 {
1095 #ifdef CONFIG_NET_CLS_ACT
1096         {
1097                 struct tc_action *act;
1098
1099                 if (exts->police && tb[exts->police]) {
1100                         act = tcf_action_init_1(net, tp, tb[exts->police],
1101                                                 rate_tlv, "police", ovr,
1102                                                 TCA_ACT_BIND);
1103                         if (IS_ERR(act))
1104                                 return PTR_ERR(act);
1105
1106                         act->type = exts->type = TCA_OLD_COMPAT;
1107                         exts->actions[0] = act;
1108                         exts->nr_actions = 1;
1109                 } else if (exts->action && tb[exts->action]) {
1110                         LIST_HEAD(actions);
1111                         int err, i = 0;
1112
1113                         err = tcf_action_init(net, tp, tb[exts->action],
1114                                               rate_tlv, NULL, ovr, TCA_ACT_BIND,
1115                                               &actions);
1116                         if (err)
1117                                 return err;
1118                         list_for_each_entry(act, &actions, list)
1119                                 exts->actions[i++] = act;
1120                         exts->nr_actions = i;
1121                 }
1122                 exts->net = net;
1123         }
1124 #else
1125         if ((exts->action && tb[exts->action]) ||
1126             (exts->police && tb[exts->police]))
1127                 return -EOPNOTSUPP;
1128 #endif
1129
1130         return 0;
1131 }
1132 EXPORT_SYMBOL(tcf_exts_validate);
1133
1134 void tcf_exts_change(struct tcf_exts *dst, struct tcf_exts *src)
1135 {
1136 #ifdef CONFIG_NET_CLS_ACT
1137         struct tcf_exts old = *dst;
1138
1139         *dst = *src;
1140         tcf_exts_destroy(&old);
1141 #endif
1142 }
1143 EXPORT_SYMBOL(tcf_exts_change);
1144
1145 #ifdef CONFIG_NET_CLS_ACT
1146 static struct tc_action *tcf_exts_first_act(struct tcf_exts *exts)
1147 {
1148         if (exts->nr_actions == 0)
1149                 return NULL;
1150         else
1151                 return exts->actions[0];
1152 }
1153 #endif
1154
1155 int tcf_exts_dump(struct sk_buff *skb, struct tcf_exts *exts)
1156 {
1157 #ifdef CONFIG_NET_CLS_ACT
1158         struct nlattr *nest;
1159
1160         if (exts->action && tcf_exts_has_actions(exts)) {
1161                 /*
1162                  * again for backward compatible mode - we want
1163                  * to work with both old and new modes of entering
1164                  * tc data even if iproute2  was newer - jhs
1165                  */
1166                 if (exts->type != TCA_OLD_COMPAT) {
1167                         LIST_HEAD(actions);
1168
1169                         nest = nla_nest_start(skb, exts->action);
1170                         if (nest == NULL)
1171                                 goto nla_put_failure;
1172
1173                         tcf_exts_to_list(exts, &actions);
1174                         if (tcf_action_dump(skb, &actions, 0, 0) < 0)
1175                                 goto nla_put_failure;
1176                         nla_nest_end(skb, nest);
1177                 } else if (exts->police) {
1178                         struct tc_action *act = tcf_exts_first_act(exts);
1179                         nest = nla_nest_start(skb, exts->police);
1180                         if (nest == NULL || !act)
1181                                 goto nla_put_failure;
1182                         if (tcf_action_dump_old(skb, act, 0, 0) < 0)
1183                                 goto nla_put_failure;
1184                         nla_nest_end(skb, nest);
1185                 }
1186         }
1187         return 0;
1188
1189 nla_put_failure:
1190         nla_nest_cancel(skb, nest);
1191         return -1;
1192 #else
1193         return 0;
1194 #endif
1195 }
1196 EXPORT_SYMBOL(tcf_exts_dump);
1197
1198
1199 int tcf_exts_dump_stats(struct sk_buff *skb, struct tcf_exts *exts)
1200 {
1201 #ifdef CONFIG_NET_CLS_ACT
1202         struct tc_action *a = tcf_exts_first_act(exts);
1203         if (a != NULL && tcf_action_copy_stats(skb, a, 1) < 0)
1204                 return -1;
1205 #endif
1206         return 0;
1207 }
1208 EXPORT_SYMBOL(tcf_exts_dump_stats);
1209
1210 static int tc_exts_setup_cb_egdev_call(struct tcf_exts *exts,
1211                                        enum tc_setup_type type,
1212                                        void *type_data, bool err_stop)
1213 {
1214         int ok_count = 0;
1215 #ifdef CONFIG_NET_CLS_ACT
1216         const struct tc_action *a;
1217         struct net_device *dev;
1218         int i, ret;
1219
1220         if (!tcf_exts_has_actions(exts))
1221                 return 0;
1222
1223         for (i = 0; i < exts->nr_actions; i++) {
1224                 a = exts->actions[i];
1225                 if (!a->ops->get_dev)
1226                         continue;
1227                 dev = a->ops->get_dev(a);
1228                 if (!dev)
1229                         continue;
1230                 ret = tc_setup_cb_egdev_call(dev, type, type_data, err_stop);
1231                 if (ret < 0)
1232                         return ret;
1233                 ok_count += ret;
1234         }
1235 #endif
1236         return ok_count;
1237 }
1238
1239 int tc_setup_cb_call(struct tcf_block *block, struct tcf_exts *exts,
1240                      enum tc_setup_type type, void *type_data, bool err_stop)
1241 {
1242         int ok_count;
1243         int ret;
1244
1245         ret = tcf_block_cb_call(block, type, type_data, err_stop);
1246         if (ret < 0)
1247                 return ret;
1248         ok_count = ret;
1249
1250         if (!exts)
1251                 return ok_count;
1252         ret = tc_exts_setup_cb_egdev_call(exts, type, type_data, err_stop);
1253         if (ret < 0)
1254                 return ret;
1255         ok_count += ret;
1256
1257         return ok_count;
1258 }
1259 EXPORT_SYMBOL(tc_setup_cb_call);
1260
1261 static int __init tc_filter_init(void)
1262 {
1263         tc_filter_wq = alloc_ordered_workqueue("tc_filter_workqueue", 0);
1264         if (!tc_filter_wq)
1265                 return -ENOMEM;
1266
1267         rtnl_register(PF_UNSPEC, RTM_NEWTFILTER, tc_ctl_tfilter, NULL, 0);
1268         rtnl_register(PF_UNSPEC, RTM_DELTFILTER, tc_ctl_tfilter, NULL, 0);
1269         rtnl_register(PF_UNSPEC, RTM_GETTFILTER, tc_ctl_tfilter,
1270                       tc_dump_tfilter, 0);
1271
1272         return 0;
1273 }
1274
1275 subsys_initcall(tc_filter_init);