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_head_change(struct tcf_chain *chain,
 199                                   struct tcf_proto *tp_head)
 200 {
 201         if (chain->chain_head_change)
 202                 chain->chain_head_change(tp_head,
 203                                          chain->chain_head_change_priv);
 204 }
 205
 206 static void tcf_chain_flush(struct tcf_chain *chain)
 207 {
 208         struct tcf_proto *tp;
 209
 210         tcf_chain_head_change(chain, NULL);
 211         while ((tp = rtnl_dereference(chain->filter_chain)) != NULL) {
 212                 RCU_INIT_POINTER(chain->filter_chain, tp->next);
 213                 tcf_chain_put(chain);
 214                 tcf_proto_destroy(tp);
 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         /* Only chain 0 should be still here. */
 339         list_for_each_entry_safe(chain, tmp, &block->chain_list, list)
 340                 tcf_chain_put(chain);
 341         rtnl_unlock();
 342         kfree(block);
 343 }
 344
 345 /* XXX: Standalone actions are not allowed to jump to any chain, and bound
 346  * actions should be all removed after flushing. However, filters are now
 347  * destroyed in tc filter workqueue with RTNL lock, they can not race here.
 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, *tmp;
 353
 354         list_for_each_entry_safe(chain, tmp, &block->chain_list, list)
 355                 tcf_chain_flush(chain);
 356
 357         tcf_block_offload_unbind(block, q, ei);
 358
 359         INIT_WORK(&block->work, tcf_block_put_final);
 360         /* Wait for existing RCU callbacks to cool down, make sure their works
 361          * have been queued before this. We can not flush pending works here
 362          * because we are holding the RTNL lock.
 363          */
 364         rcu_barrier();
 365         tcf_queue_work(&block->work);
 366 }
 367 EXPORT_SYMBOL(tcf_block_put_ext);
 368
 369 void tcf_block_put(struct tcf_block *block)
 370 {
 371         struct tcf_block_ext_info ei = {0, };
 372
 373         if (!block)
 374                 return;
 375         tcf_block_put_ext(block, block->q, &ei);
 376 }
 377
 378 EXPORT_SYMBOL(tcf_block_put);
 379
 380 struct tcf_block_cb {
 381         struct list_head list;
 382         tc_setup_cb_t *cb;
 383         void *cb_ident;
 384         void *cb_priv;
 385         unsigned int refcnt;
 386 };
 387
 388 void *tcf_block_cb_priv(struct tcf_block_cb *block_cb)
 389 {
 390         return block_cb->cb_priv;
 391 }
 392 EXPORT_SYMBOL(tcf_block_cb_priv);
 393
 394 struct tcf_block_cb *tcf_block_cb_lookup(struct tcf_block *block,
 395                                          tc_setup_cb_t *cb, void *cb_ident)
 396 {       struct tcf_block_cb *block_cb;
 397
 398         list_for_each_entry(block_cb, &block->cb_list, list)
 399                 if (block_cb->cb == cb && block_cb->cb_ident == cb_ident)
 400                         return block_cb;
 401         return NULL;
 402 }
 403 EXPORT_SYMBOL(tcf_block_cb_lookup);
 404
 405 void tcf_block_cb_incref(struct tcf_block_cb *block_cb)
 406 {
 407         block_cb->refcnt++;
 408 }
 409 EXPORT_SYMBOL(tcf_block_cb_incref);
 410
 411 unsigned int tcf_block_cb_decref(struct tcf_block_cb *block_cb)
 412 {
 413         return --block_cb->refcnt;
 414 }
 415 EXPORT_SYMBOL(tcf_block_cb_decref);
 416
 417 struct tcf_block_cb *__tcf_block_cb_register(struct tcf_block *block,
 418                                              tc_setup_cb_t *cb, void *cb_ident,
 419                                              void *cb_priv)
 420 {
 421         struct tcf_block_cb *block_cb;
 422
 423         block_cb = kzalloc(sizeof(*block_cb), GFP_KERNEL);
 424         if (!block_cb)
 425                 return NULL;
 426         block_cb->cb = cb;
 427         block_cb->cb_ident = cb_ident;
 428         block_cb->cb_priv = cb_priv;
 429         list_add(&block_cb->list, &block->cb_list);
 430         return block_cb;
 431 }
 432 EXPORT_SYMBOL(__tcf_block_cb_register);
 433
 434 int tcf_block_cb_register(struct tcf_block *block,
 435                           tc_setup_cb_t *cb, void *cb_ident,
 436                           void *cb_priv)
 437 {
 438         struct tcf_block_cb *block_cb;
 439
 440         block_cb = __tcf_block_cb_register(block, cb, cb_ident, cb_priv);
 441         return block_cb ? 0 : -ENOMEM;
 442 }
 443 EXPORT_SYMBOL(tcf_block_cb_register);
 444
 445 void __tcf_block_cb_unregister(struct tcf_block_cb *block_cb)
 446 {
 447         list_del(&block_cb->list);
 448         kfree(block_cb);
 449 }
 450 EXPORT_SYMBOL(__tcf_block_cb_unregister);
 451
 452 void tcf_block_cb_unregister(struct tcf_block *block,
 453                              tc_setup_cb_t *cb, void *cb_ident)
 454 {
 455         struct tcf_block_cb *block_cb;
 456
 457         block_cb = tcf_block_cb_lookup(block, cb, cb_ident);
 458         if (!block_cb)
 459                 return;
 460         __tcf_block_cb_unregister(block_cb);
 461 }
 462 EXPORT_SYMBOL(tcf_block_cb_unregister);
 463
 464 static int tcf_block_cb_call(struct tcf_block *block, enum tc_setup_type type,
 465                              void *type_data, bool err_stop)
 466 {
 467         struct tcf_block_cb *block_cb;
 468         int ok_count = 0;
 469         int err;
 470
 471         list_for_each_entry(block_cb, &block->cb_list, list) {
 472                 err = block_cb->cb(type, type_data, block_cb->cb_priv);
 473                 if (err) {
 474                         if (err_stop)
 475                                 return err;
 476                 } else {
 477                         ok_count++;
 478                 }
 479         }
 480         return ok_count;
 481 }
 482
 483 /* Main classifier routine: scans classifier chain attached
 484  * to this qdisc, (optionally) tests for protocol and asks
 485  * specific classifiers.
 486  */
 487 int tcf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 488                  struct tcf_result *res, bool compat_mode)
 489 {
 490         __be16 protocol = tc_skb_protocol(skb);
 491 #ifdef CONFIG_NET_CLS_ACT
 492         const int max_reclassify_loop = 4;
 493         const struct tcf_proto *orig_tp = tp;
 494         const struct tcf_proto *first_tp;
 495         int limit = 0;
 496
 497 reclassify:
 498 #endif
 499         for (; tp; tp = rcu_dereference_bh(tp->next)) {
 500                 int err;
 501
 502                 if (tp->protocol != protocol &&
 503                     tp->protocol != htons(ETH_P_ALL))
 504                         continue;
 505
 506                 err = tp->classify(skb, tp, res);
 507 #ifdef CONFIG_NET_CLS_ACT
 508                 if (unlikely(err == TC_ACT_RECLASSIFY && !compat_mode)) {
 509                         first_tp = orig_tp;
 510                         goto reset;
 511                 } else if (unlikely(TC_ACT_EXT_CMP(err, TC_ACT_GOTO_CHAIN))) {
 512                         first_tp = res->goto_tp;
 513                         goto reset;
 514                 }
 515 #endif
 516                 if (err >= 0)
 517                         return err;
 518         }
 519
 520         return TC_ACT_UNSPEC; /* signal: continue lookup */
 521 #ifdef CONFIG_NET_CLS_ACT
 522 reset:
 523         if (unlikely(limit++ >= max_reclassify_loop)) {
 524                 net_notice_ratelimited("%s: reclassify loop, rule prio %u, protocol %02x\n",
 525                                        tp->q->ops->id, tp->prio & 0xffff,
 526                                        ntohs(tp->protocol));
 527                 return TC_ACT_SHOT;
 528         }
 529
 530         tp = first_tp;
 531         protocol = tc_skb_protocol(skb);
 532         goto reclassify;
 533 #endif
 534 }
 535 EXPORT_SYMBOL(tcf_classify);
 536
 537 struct tcf_chain_info {
 538         struct tcf_proto __rcu **pprev;
 539         struct tcf_proto __rcu *next;
 540 };
 541
 542 static struct tcf_proto *tcf_chain_tp_prev(struct tcf_chain_info *chain_info)
 543 {
 544         return rtnl_dereference(*chain_info->pprev);
 545 }
 546
 547 static void tcf_chain_tp_insert(struct tcf_chain *chain,
 548                                 struct tcf_chain_info *chain_info,
 549                                 struct tcf_proto *tp)
 550 {
 551         if (*chain_info->pprev == chain->filter_chain)
 552                 tcf_chain_head_change(chain, tp);
 553         RCU_INIT_POINTER(tp->next, tcf_chain_tp_prev(chain_info));
 554         rcu_assign_pointer(*chain_info->pprev, tp);
 555         tcf_chain_hold(chain);
 556 }
 557
 558 static void tcf_chain_tp_remove(struct tcf_chain *chain,
 559                                 struct tcf_chain_info *chain_info,
 560                                 struct tcf_proto *tp)
 561 {
 562         struct tcf_proto *next = rtnl_dereference(chain_info->next);
 563
 564         if (tp == chain->filter_chain)
 565                 tcf_chain_head_change(chain, next);
 566         RCU_INIT_POINTER(*chain_info->pprev, next);
 567         tcf_chain_put(chain);
 568 }
 569
 570 static struct tcf_proto *tcf_chain_tp_find(struct tcf_chain *chain,
 571                                            struct tcf_chain_info *chain_info,
 572                                            u32 protocol, u32 prio,
 573                                            bool prio_allocate)
 574 {
 575         struct tcf_proto **pprev;
 576         struct tcf_proto *tp;
 577
 578         /* Check the chain for existence of proto-tcf with this priority */
 579         for (pprev = &chain->filter_chain;
 580              (tp = rtnl_dereference(*pprev)); pprev = &tp->next) {
 581                 if (tp->prio >= prio) {
 582                         if (tp->prio == prio) {
 583                                 if (prio_allocate ||
 584                                     (tp->protocol != protocol && protocol))
 585                                         return ERR_PTR(-EINVAL);
 586                         } else {
 587                                 tp = NULL;
 588                         }
 589                         break;
 590                 }
 591         }
 592         chain_info->pprev = pprev;
 593         chain_info->next = tp ? tp->next : NULL;
 594         return tp;
 595 }
 596
 597 static int tcf_fill_node(struct net *net, struct sk_buff *skb,
 598                          struct tcf_proto *tp, struct Qdisc *q, u32 parent,
 599                          void *fh, u32 portid, u32 seq, u16 flags, int event)
 600 {
 601         struct tcmsg *tcm;
 602         struct nlmsghdr  *nlh;
 603         unsigned char *b = skb_tail_pointer(skb);
 604
 605         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*tcm), flags);
 606         if (!nlh)
 607                 goto out_nlmsg_trim;
 608         tcm = nlmsg_data(nlh);
 609         tcm->tcm_family = AF_UNSPEC;
 610         tcm->tcm__pad1 = 0;
 611         tcm->tcm__pad2 = 0;
 612         tcm->tcm_ifindex = qdisc_dev(q)->ifindex;
 613         tcm->tcm_parent = parent;
 614         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 615         if (nla_put_string(skb, TCA_KIND, tp->ops->kind))
 616                 goto nla_put_failure;
 617         if (nla_put_u32(skb, TCA_CHAIN, tp->chain->index))
 618                 goto nla_put_failure;
 619         if (!fh) {
 620                 tcm->tcm_handle = 0;
 621         } else {
 622                 if (tp->ops->dump && tp->ops->dump(net, tp, fh, skb, tcm) < 0)
 623                         goto nla_put_failure;
 624         }
 625         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 626         return skb->len;
 627
 628 out_nlmsg_trim:
 629 nla_put_failure:
 630         nlmsg_trim(skb, b);
 631         return -1;
 632 }
 633
 634 static int tfilter_notify(struct net *net, struct sk_buff *oskb,
 635                           struct nlmsghdr *n, struct tcf_proto *tp,
 636                           struct Qdisc *q, u32 parent,
 637                           void *fh, int event, bool unicast)
 638 {
 639         struct sk_buff *skb;
 640         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 641
 642         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 643         if (!skb)
 644                 return -ENOBUFS;
 645
 646         if (tcf_fill_node(net, skb, tp, q, parent, fh, portid, n->nlmsg_seq,
 647                           n->nlmsg_flags, event) <= 0) {
 648                 kfree_skb(skb);
 649                 return -EINVAL;
 650         }
 651
 652         if (unicast)
 653                 return netlink_unicast(net->rtnl, skb, portid, MSG_DONTWAIT);
 654
 655         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 656                               n->nlmsg_flags & NLM_F_ECHO);
 657 }
 658
 659 static int tfilter_del_notify(struct net *net, struct sk_buff *oskb,
 660                               struct nlmsghdr *n, struct tcf_proto *tp,
 661                               struct Qdisc *q, u32 parent,
 662                               void *fh, bool unicast, bool *last)
 663 {
 664         struct sk_buff *skb;
 665         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 666         int err;
 667
 668         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 669         if (!skb)
 670                 return -ENOBUFS;
 671
 672         if (tcf_fill_node(net, skb, tp, q, parent, fh, portid, n->nlmsg_seq,
 673                           n->nlmsg_flags, RTM_DELTFILTER) <= 0) {
 674                 kfree_skb(skb);
 675                 return -EINVAL;
 676         }
 677
 678         err = tp->ops->delete(tp, fh, last);
 679         if (err) {
 680                 kfree_skb(skb);
 681                 return err;
 682         }
 683
 684         if (unicast)
 685                 return netlink_unicast(net->rtnl, skb, portid, MSG_DONTWAIT);
 686
 687         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 688                               n->nlmsg_flags & NLM_F_ECHO);
 689 }
 690
 691 static void tfilter_notify_chain(struct net *net, struct sk_buff *oskb,
 692                                  struct Qdisc *q, u32 parent,
 693                                  struct nlmsghdr *n,
 694                                  struct tcf_chain *chain, int event)
 695 {
 696         struct tcf_proto *tp;
 697
 698         for (tp = rtnl_dereference(chain->filter_chain);
 699              tp; tp = rtnl_dereference(tp->next))
 700                 tfilter_notify(net, oskb, n, tp, q, parent, 0, event, false);
 701 }
 702
 703 /* Add/change/delete/get a filter node */
 704
 705 static int tc_ctl_tfilter(struct sk_buff *skb, struct nlmsghdr *n,
 706                           struct netlink_ext_ack *extack)
 707 {
 708         struct net *net = sock_net(skb->sk);
 709         struct nlattr *tca[TCA_MAX + 1];
 710         struct tcmsg *t;
 711         u32 protocol;
 712         u32 prio;
 713         bool prio_allocate;
 714         u32 parent;
 715         u32 chain_index;
 716         struct net_device *dev;
 717         struct Qdisc  *q;
 718         struct tcf_chain_info chain_info;
 719         struct tcf_chain *chain = NULL;
 720         struct tcf_block *block;
 721         struct tcf_proto *tp;
 722         const struct Qdisc_class_ops *cops;
 723         unsigned long cl;
 724         void *fh;
 725         int err;
 726         int tp_created;
 727
 728         if ((n->nlmsg_type != RTM_GETTFILTER) &&
 729             !netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN))
 730                 return -EPERM;
 731
 732 replay:
 733         tp_created = 0;
 734
 735         err = nlmsg_parse(n, sizeof(*t), tca, TCA_MAX, NULL, extack);
 736         if (err < 0)
 737                 return err;
 738
 739         t = nlmsg_data(n);
 740         protocol = TC_H_MIN(t->tcm_info);
 741         prio = TC_H_MAJ(t->tcm_info);
 742         prio_allocate = false;
 743         parent = t->tcm_parent;
 744         cl = 0;
 745
 746         if (prio == 0) {
 747                 switch (n->nlmsg_type) {
 748                 case RTM_DELTFILTER:
 749                         if (protocol || t->tcm_handle || tca[TCA_KIND])
 750                                 return -ENOENT;
 751                         break;
 752                 case RTM_NEWTFILTER:
 753                         /* If no priority is provided by the user,
 754                          * we allocate one.
 755                          */
 756                         if (n->nlmsg_flags & NLM_F_CREATE) {
 757                                 prio = TC_H_MAKE(0x80000000U, 0U);
 758                                 prio_allocate = true;
 759                                 break;
 760                         }
 761                         /* fall-through */
 762                 default:
 763                         return -ENOENT;
 764                 }
 765         }
 766
 767         /* Find head of filter chain. */
 768
 769         /* Find link */
 770         dev = __dev_get_by_index(net, t->tcm_ifindex);
 771         if (dev == NULL)
 772                 return -ENODEV;
 773
 774         /* Find qdisc */
 775         if (!parent) {
 776                 q = dev->qdisc;
 777                 parent = q->handle;
 778         } else {
 779                 q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent));
 780                 if (q == NULL)
 781                         return -EINVAL;
 782         }
 783
 784         /* Is it classful? */
 785         cops = q->ops->cl_ops;
 786         if (!cops)
 787                 return -EINVAL;
 788
 789         if (!cops->tcf_block)
 790                 return -EOPNOTSUPP;
 791
 792         /* Do we search for filter, attached to class? */
 793         if (TC_H_MIN(parent)) {
 794                 cl = cops->find(q, parent);
 795                 if (cl == 0)
 796                         return -ENOENT;
 797         }
 798
 799         /* And the last stroke */
 800         block = cops->tcf_block(q, cl);
 801         if (!block) {
 802                 err = -EINVAL;
 803                 goto errout;
 804         }
 805
 806         chain_index = tca[TCA_CHAIN] ? nla_get_u32(tca[TCA_CHAIN]) : 0;
 807         if (chain_index > TC_ACT_EXT_VAL_MASK) {
 808                 err = -EINVAL;
 809                 goto errout;
 810         }
 811         chain = tcf_chain_get(block, chain_index,
 812                               n->nlmsg_type == RTM_NEWTFILTER);
 813         if (!chain) {
 814                 err = n->nlmsg_type == RTM_NEWTFILTER ? -ENOMEM : -EINVAL;
 815                 goto errout;
 816         }
 817
 818         if (n->nlmsg_type == RTM_DELTFILTER && prio == 0) {
 819                 tfilter_notify_chain(net, skb, q, parent, n,
 820                                      chain, RTM_DELTFILTER);
 821                 tcf_chain_flush(chain);
 822                 err = 0;
 823                 goto errout;
 824         }
 825
 826         tp = tcf_chain_tp_find(chain, &chain_info, protocol,
 827                                prio, prio_allocate);
 828         if (IS_ERR(tp)) {
 829                 err = PTR_ERR(tp);
 830                 goto errout;
 831         }
 832
 833         if (tp == NULL) {
 834                 /* Proto-tcf does not exist, create new one */
 835
 836                 if (tca[TCA_KIND] == NULL || !protocol) {
 837                         err = -EINVAL;
 838                         goto errout;
 839                 }
 840
 841                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 842                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 843                         err = -ENOENT;
 844                         goto errout;
 845                 }
 846
 847                 if (prio_allocate)
 848                         prio = tcf_auto_prio(tcf_chain_tp_prev(&chain_info));
 849
 850                 tp = tcf_proto_create(nla_data(tca[TCA_KIND]),
 851                                       protocol, prio, parent, q, chain);
 852                 if (IS_ERR(tp)) {
 853                         err = PTR_ERR(tp);
 854                         goto errout;
 855                 }
 856                 tp_created = 1;
 857         } else if (tca[TCA_KIND] && nla_strcmp(tca[TCA_KIND], tp->ops->kind)) {
 858                 err = -EINVAL;
 859                 goto errout;
 860         }
 861
 862         fh = tp->ops->get(tp, t->tcm_handle);
 863
 864         if (!fh) {
 865                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 866                         tcf_chain_tp_remove(chain, &chain_info, tp);
 867                         tfilter_notify(net, skb, n, tp, q, parent, fh,
 868                                        RTM_DELTFILTER, false);
 869                         tcf_proto_destroy(tp);
 870                         err = 0;
 871                         goto errout;
 872                 }
 873
 874                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 875                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 876                         err = -ENOENT;
 877                         goto errout;
 878                 }
 879         } else {
 880                 bool last;
 881
 882                 switch (n->nlmsg_type) {
 883                 case RTM_NEWTFILTER:
 884                         if (n->nlmsg_flags & NLM_F_EXCL) {
 885                                 if (tp_created)
 886                                         tcf_proto_destroy(tp);
 887                                 err = -EEXIST;
 888                                 goto errout;
 889                         }
 890                         break;
 891                 case RTM_DELTFILTER:
 892                         err = tfilter_del_notify(net, skb, n, tp, q, parent,
 893                                                  fh, false, &last);
 894                         if (err)
 895                                 goto errout;
 896                         if (last) {
 897                                 tcf_chain_tp_remove(chain, &chain_info, tp);
 898                                 tcf_proto_destroy(tp);
 899                         }
 900                         goto errout;
 901                 case RTM_GETTFILTER:
 902                         err = tfilter_notify(net, skb, n, tp, q, parent, fh,
 903                                              RTM_NEWTFILTER, true);
 904                         goto errout;
 905                 default:
 906                         err = -EINVAL;
 907                         goto errout;
 908                 }
 909         }
 910
 911         err = tp->ops->change(net, skb, tp, cl, t->tcm_handle, tca, &fh,
 912                               n->nlmsg_flags & NLM_F_CREATE ? TCA_ACT_NOREPLACE : TCA_ACT_REPLACE);
 913         if (err == 0) {
 914                 if (tp_created)
 915                         tcf_chain_tp_insert(chain, &chain_info, tp);
 916                 tfilter_notify(net, skb, n, tp, q, parent, fh,
 917                                RTM_NEWTFILTER, false);
 918         } else {
 919                 if (tp_created)
 920                         tcf_proto_destroy(tp);
 921         }
 922
 923 errout:
 924         if (chain)
 925                 tcf_chain_put(chain);
 926         if (err == -EAGAIN)
 927                 /* Replay the request. */
 928                 goto replay;
 929         return err;
 930 }
 931
 932 struct tcf_dump_args {
 933         struct tcf_walker w;
 934         struct sk_buff *skb;
 935         struct netlink_callback *cb;
 936         struct Qdisc *q;
 937         u32 parent;
 938 };
 939
 940 static int tcf_node_dump(struct tcf_proto *tp, void *n, struct tcf_walker *arg)
 941 {
 942         struct tcf_dump_args *a = (void *)arg;
 943         struct net *net = sock_net(a->skb->sk);
 944
 945         return tcf_fill_node(net, a->skb, tp, a->q, a->parent,
 946                              n, NETLINK_CB(a->cb->skb).portid,
 947                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI,
 948                              RTM_NEWTFILTER);
 949 }
 950
 951 static bool tcf_chain_dump(struct tcf_chain *chain, struct Qdisc *q, u32 parent,
 952                            struct sk_buff *skb, struct netlink_callback *cb,
 953                            long index_start, long *p_index)
 954 {
 955         struct net *net = sock_net(skb->sk);
 956         struct tcmsg *tcm = nlmsg_data(cb->nlh);
 957         struct tcf_dump_args arg;
 958         struct tcf_proto *tp;
 959
 960         for (tp = rtnl_dereference(chain->filter_chain);
 961              tp; tp = rtnl_dereference(tp->next), (*p_index)++) {
 962                 if (*p_index < index_start)
 963                         continue;
 964                 if (TC_H_MAJ(tcm->tcm_info) &&
 965                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 966                         continue;
 967                 if (TC_H_MIN(tcm->tcm_info) &&
 968                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 969                         continue;
 970                 if (*p_index > index_start)
 971                         memset(&cb->args[1], 0,
 972                                sizeof(cb->args) - sizeof(cb->args[0]));
 973                 if (cb->args[1] == 0) {
 974                         if (tcf_fill_node(net, skb, tp, q, parent, 0,
 975                                           NETLINK_CB(cb->skb).portid,
 976                                           cb->nlh->nlmsg_seq, NLM_F_MULTI,
 977                                           RTM_NEWTFILTER) <= 0)
 978                                 return false;
 979
 980                         cb->args[1] = 1;
 981                 }
 982                 if (!tp->ops->walk)
 983                         continue;
 984                 arg.w.fn = tcf_node_dump;
 985                 arg.skb = skb;
 986                 arg.cb = cb;
 987                 arg.q = q;
 988                 arg.parent = parent;
 989                 arg.w.stop = 0;
 990                 arg.w.skip = cb->args[1] - 1;
 991                 arg.w.count = 0;
 992                 tp->ops->walk(tp, &arg.w);
 993                 cb->args[1] = arg.w.count + 1;
 994                 if (arg.w.stop)
 995                         return false;
 996         }
 997         return true;
 998 }
 999
1000 /* called with RTNL */
1001 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
1002 {
1003         struct net *net = sock_net(skb->sk);
1004         struct nlattr *tca[TCA_MAX + 1];
1005         struct net_device *dev;
1006         struct Qdisc *q;
1007         struct tcf_block *block;
1008         struct tcf_chain *chain;
1009         struct tcmsg *tcm = nlmsg_data(cb->nlh);
1010         unsigned long cl = 0;
1011         const struct Qdisc_class_ops *cops;
1012         long index_start;
1013         long index;
1014         u32 parent;
1015         int err;
1016
1017         if (nlmsg_len(cb->nlh) < sizeof(*tcm))
1018                 return skb->len;
1019
1020         err = nlmsg_parse(cb->nlh, sizeof(*tcm), tca, TCA_MAX, NULL, NULL);
1021         if (err)
1022                 return err;
1023
1024         dev = __dev_get_by_index(net, tcm->tcm_ifindex);
1025         if (!dev)
1026                 return skb->len;
1027
1028         parent = tcm->tcm_parent;
1029         if (!parent) {
1030                 q = dev->qdisc;
1031                 parent = q->handle;
1032         } else {
1033                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
1034         }
1035         if (!q)
1036                 goto out;
1037         cops = q->ops->cl_ops;
1038         if (!cops)
1039                 goto out;
1040         if (!cops->tcf_block)
1041                 goto out;
1042         if (TC_H_MIN(tcm->tcm_parent)) {
1043                 cl = cops->find(q, tcm->tcm_parent);
1044                 if (cl == 0)
1045                         goto out;
1046         }
1047         block = cops->tcf_block(q, cl);
1048         if (!block)
1049                 goto out;
1050
1051         index_start = cb->args[0];
1052         index = 0;
1053
1054         list_for_each_entry(chain, &block->chain_list, list) {
1055                 if (tca[TCA_CHAIN] &&
1056                     nla_get_u32(tca[TCA_CHAIN]) != chain->index)
1057                         continue;
1058                 if (!tcf_chain_dump(chain, q, parent, skb, cb,
1059                                     index_start, &index))
1060                         break;
1061         }
1062
1063         cb->args[0] = index;
1064
1065 out:
1066         return skb->len;
1067 }
1068
1069 void tcf_exts_destroy(struct tcf_exts *exts)
1070 {
1071 #ifdef CONFIG_NET_CLS_ACT
1072         LIST_HEAD(actions);
1073
1074         ASSERT_RTNL();
1075         tcf_exts_to_list(exts, &actions);
1076         tcf_action_destroy(&actions, TCA_ACT_UNBIND);
1077         kfree(exts->actions);
1078         exts->nr_actions = 0;
1079 #endif
1080 }
1081 EXPORT_SYMBOL(tcf_exts_destroy);
1082
1083 int tcf_exts_validate(struct net *net, struct tcf_proto *tp, struct nlattr **tb,
1084                       struct nlattr *rate_tlv, struct tcf_exts *exts, bool ovr)
1085 {
1086 #ifdef CONFIG_NET_CLS_ACT
1087         {
1088                 struct tc_action *act;
1089
1090                 if (exts->police && tb[exts->police]) {
1091                         act = tcf_action_init_1(net, tp, tb[exts->police],
1092                                                 rate_tlv, "police", ovr,
1093                                                 TCA_ACT_BIND);
1094                         if (IS_ERR(act))
1095                                 return PTR_ERR(act);
1096
1097                         act->type = exts->type = TCA_OLD_COMPAT;
1098                         exts->actions[0] = act;
1099                         exts->nr_actions = 1;
1100                 } else if (exts->action && tb[exts->action]) {
1101                         LIST_HEAD(actions);
1102                         int err, i = 0;
1103
1104                         err = tcf_action_init(net, tp, tb[exts->action],
1105                                               rate_tlv, NULL, ovr, TCA_ACT_BIND,
1106                                               &actions);
1107                         if (err)
1108                                 return err;
1109                         list_for_each_entry(act, &actions, list)
1110                                 exts->actions[i++] = act;
1111                         exts->nr_actions = i;
1112                 }
1113                 exts->net = net;
1114         }
1115 #else
1116         if ((exts->action && tb[exts->action]) ||
1117             (exts->police && tb[exts->police]))
1118                 return -EOPNOTSUPP;
1119 #endif
1120
1121         return 0;
1122 }
1123 EXPORT_SYMBOL(tcf_exts_validate);
1124
1125 void tcf_exts_change(struct tcf_exts *dst, struct tcf_exts *src)
1126 {
1127 #ifdef CONFIG_NET_CLS_ACT
1128         struct tcf_exts old = *dst;
1129
1130         *dst = *src;
1131         tcf_exts_destroy(&old);
1132 #endif
1133 }
1134 EXPORT_SYMBOL(tcf_exts_change);
1135
1136 #ifdef CONFIG_NET_CLS_ACT
1137 static struct tc_action *tcf_exts_first_act(struct tcf_exts *exts)
1138 {
1139         if (exts->nr_actions == 0)
1140                 return NULL;
1141         else
1142                 return exts->actions[0];
1143 }
1144 #endif
1145
1146 int tcf_exts_dump(struct sk_buff *skb, struct tcf_exts *exts)
1147 {
1148 #ifdef CONFIG_NET_CLS_ACT
1149         struct nlattr *nest;
1150
1151         if (exts->action && tcf_exts_has_actions(exts)) {
1152                 /*
1153                  * again for backward compatible mode - we want
1154                  * to work with both old and new modes of entering
1155                  * tc data even if iproute2  was newer - jhs
1156                  */
1157                 if (exts->type != TCA_OLD_COMPAT) {
1158                         LIST_HEAD(actions);
1159
1160                         nest = nla_nest_start(skb, exts->action);
1161                         if (nest == NULL)
1162                                 goto nla_put_failure;
1163
1164                         tcf_exts_to_list(exts, &actions);
1165                         if (tcf_action_dump(skb, &actions, 0, 0) < 0)
1166                                 goto nla_put_failure;
1167                         nla_nest_end(skb, nest);
1168                 } else if (exts->police) {
1169                         struct tc_action *act = tcf_exts_first_act(exts);
1170                         nest = nla_nest_start(skb, exts->police);
1171                         if (nest == NULL || !act)
1172                                 goto nla_put_failure;
1173                         if (tcf_action_dump_old(skb, act, 0, 0) < 0)
1174                                 goto nla_put_failure;
1175                         nla_nest_end(skb, nest);
1176                 }
1177         }
1178         return 0;
1179
1180 nla_put_failure:
1181         nla_nest_cancel(skb, nest);
1182         return -1;
1183 #else
1184         return 0;
1185 #endif
1186 }
1187 EXPORT_SYMBOL(tcf_exts_dump);
1188
1189
1190 int tcf_exts_dump_stats(struct sk_buff *skb, struct tcf_exts *exts)
1191 {
1192 #ifdef CONFIG_NET_CLS_ACT
1193         struct tc_action *a = tcf_exts_first_act(exts);
1194         if (a != NULL && tcf_action_copy_stats(skb, a, 1) < 0)
1195                 return -1;
1196 #endif
1197         return 0;
1198 }
1199 EXPORT_SYMBOL(tcf_exts_dump_stats);
1200
1201 static int tc_exts_setup_cb_egdev_call(struct tcf_exts *exts,
1202                                        enum tc_setup_type type,
1203                                        void *type_data, bool err_stop)
1204 {
1205         int ok_count = 0;
1206 #ifdef CONFIG_NET_CLS_ACT
1207         const struct tc_action *a;
1208         struct net_device *dev;
1209         int i, ret;
1210
1211         if (!tcf_exts_has_actions(exts))
1212                 return 0;
1213
1214         for (i = 0; i < exts->nr_actions; i++) {
1215                 a = exts->actions[i];
1216                 if (!a->ops->get_dev)
1217                         continue;
1218                 dev = a->ops->get_dev(a);
1219                 if (!dev)
1220                         continue;
1221                 ret = tc_setup_cb_egdev_call(dev, type, type_data, err_stop);
1222                 if (ret < 0)
1223                         return ret;
1224                 ok_count += ret;
1225         }
1226 #endif
1227         return ok_count;
1228 }
1229
1230 int tc_setup_cb_call(struct tcf_block *block, struct tcf_exts *exts,
1231                      enum tc_setup_type type, void *type_data, bool err_stop)
1232 {
1233         int ok_count;
1234         int ret;
1235
1236         ret = tcf_block_cb_call(block, type, type_data, err_stop);
1237         if (ret < 0)
1238                 return ret;
1239         ok_count = ret;
1240
1241         if (!exts)
1242                 return ok_count;
1243         ret = tc_exts_setup_cb_egdev_call(exts, type, type_data, err_stop);
1244         if (ret < 0)
1245                 return ret;
1246         ok_count += ret;
1247
1248         return ok_count;
1249 }
1250 EXPORT_SYMBOL(tc_setup_cb_call);
1251
1252 static int __init tc_filter_init(void)
1253 {
1254         tc_filter_wq = alloc_ordered_workqueue("tc_filter_workqueue", 0);
1255         if (!tc_filter_wq)
1256                 return -ENOMEM;
1257
1258         rtnl_register(PF_UNSPEC, RTM_NEWTFILTER, tc_ctl_tfilter, NULL, 0);
1259         rtnl_register(PF_UNSPEC, RTM_DELTFILTER, tc_ctl_tfilter, NULL, 0);
1260         rtnl_register(PF_UNSPEC, RTM_GETTFILTER, tc_ctl_tfilter,
1261                       tc_dump_tfilter, 0);
1262
1263         return 0;
1264 }
1265
1266 subsys_initcall(tc_filter_init);