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 extern const struct nla_policy rtm_tca_policy[TCA_MAX + 1];
  35
  36 /* The list of all installed classifier types */
  37 static LIST_HEAD(tcf_proto_base);
  38
  39 /* Protects list of registered TC modules. It is pure SMP lock. */
  40 static DEFINE_RWLOCK(cls_mod_lock);
  41
  42 /* Find classifier type by string name */
  43
  44 static const struct tcf_proto_ops *tcf_proto_lookup_ops(const char *kind)
  45 {
  46         const struct tcf_proto_ops *t, *res = NULL;
  47
  48         if (kind) {
  49                 read_lock(&cls_mod_lock);
  50                 list_for_each_entry(t, &tcf_proto_base, head) {
  51                         if (strcmp(kind, t->kind) == 0) {
  52                                 if (try_module_get(t->owner))
  53                                         res = t;
  54                                 break;
  55                         }
  56                 }
  57                 read_unlock(&cls_mod_lock);
  58         }
  59         return res;
  60 }
  61
  62 /* Register(unregister) new classifier type */
  63
  64 int register_tcf_proto_ops(struct tcf_proto_ops *ops)
  65 {
  66         struct tcf_proto_ops *t;
  67         int rc = -EEXIST;
  68
  69         write_lock(&cls_mod_lock);
  70         list_for_each_entry(t, &tcf_proto_base, head)
  71                 if (!strcmp(ops->kind, t->kind))
  72                         goto out;
  73
  74         list_add_tail(&ops->head, &tcf_proto_base);
  75         rc = 0;
  76 out:
  77         write_unlock(&cls_mod_lock);
  78         return rc;
  79 }
  80 EXPORT_SYMBOL(register_tcf_proto_ops);
  81
  82 static struct workqueue_struct *tc_filter_wq;
  83
  84 int unregister_tcf_proto_ops(struct tcf_proto_ops *ops)
  85 {
  86         struct tcf_proto_ops *t;
  87         int rc = -ENOENT;
  88
  89         /* Wait for outstanding call_rcu()s, if any, from a
  90          * tcf_proto_ops's destroy() handler.
  91          */
  92         rcu_barrier();
  93         flush_workqueue(tc_filter_wq);
  94
  95         write_lock(&cls_mod_lock);
  96         list_for_each_entry(t, &tcf_proto_base, head) {
  97                 if (t == ops) {
  98                         list_del(&t->head);
  99                         rc = 0;
 100                         break;
 101                 }
 102         }
 103         write_unlock(&cls_mod_lock);
 104         return rc;
 105 }
 106 EXPORT_SYMBOL(unregister_tcf_proto_ops);
 107
 108 bool tcf_queue_work(struct work_struct *work)
 109 {
 110         return queue_work(tc_filter_wq, work);
 111 }
 112 EXPORT_SYMBOL(tcf_queue_work);
 113
 114 /* Select new prio value from the range, managed by kernel. */
 115
 116 static inline u32 tcf_auto_prio(struct tcf_proto *tp)
 117 {
 118         u32 first = TC_H_MAKE(0xC0000000U, 0U);
 119
 120         if (tp)
 121                 first = tp->prio - 1;
 122
 123         return TC_H_MAJ(first);
 124 }
 125
 126 static struct tcf_proto *tcf_proto_create(const char *kind, u32 protocol,
 127                                           u32 prio, u32 parent, struct Qdisc *q,
 128                                           struct tcf_chain *chain)
 129 {
 130         struct tcf_proto *tp;
 131         int err;
 132
 133         tp = kzalloc(sizeof(*tp), GFP_KERNEL);
 134         if (!tp)
 135                 return ERR_PTR(-ENOBUFS);
 136
 137         err = -ENOENT;
 138         tp->ops = tcf_proto_lookup_ops(kind);
 139         if (!tp->ops) {
 140 #ifdef CONFIG_MODULES
 141                 rtnl_unlock();
 142                 request_module("cls_%s", kind);
 143                 rtnl_lock();
 144                 tp->ops = tcf_proto_lookup_ops(kind);
 145                 /* We dropped the RTNL semaphore in order to perform
 146                  * the module load. So, even if we succeeded in loading
 147                  * the module we have to replay the request. We indicate
 148                  * this using -EAGAIN.
 149                  */
 150                 if (tp->ops) {
 151                         module_put(tp->ops->owner);
 152                         err = -EAGAIN;
 153                 } else {
 154                         err = -ENOENT;
 155                 }
 156 #endif
 157                 goto errout;
 158         }
 159         tp->classify = tp->ops->classify;
 160         tp->protocol = protocol;
 161         tp->prio = prio;
 162         tp->classid = parent;
 163         tp->q = q;
 164         tp->chain = chain;
 165
 166         err = tp->ops->init(tp);
 167         if (err) {
 168                 module_put(tp->ops->owner);
 169                 goto errout;
 170         }
 171         return tp;
 172
 173 errout:
 174         kfree(tp);
 175         return ERR_PTR(err);
 176 }
 177
 178 static void tcf_proto_destroy(struct tcf_proto *tp)
 179 {
 180         tp->ops->destroy(tp);
 181         module_put(tp->ops->owner);
 182         kfree_rcu(tp, rcu);
 183 }
 184
 185 static struct tcf_chain *tcf_chain_create(struct tcf_block *block,
 186                                           u32 chain_index)
 187 {
 188         struct tcf_chain *chain;
 189
 190         chain = kzalloc(sizeof(*chain), GFP_KERNEL);
 191         if (!chain)
 192                 return NULL;
 193         list_add_tail(&chain->list, &block->chain_list);
 194         chain->block = block;
 195         chain->index = chain_index;
 196         chain->refcnt = 1;
 197         return chain;
 198 }
 199
 200 static void tcf_chain_flush(struct tcf_chain *chain)
 201 {
 202         struct tcf_proto *tp = rtnl_dereference(chain->filter_chain);
 203
 204         if (chain->p_filter_chain)
 205                 RCU_INIT_POINTER(*chain->p_filter_chain, NULL);
 206         while (tp) {
 207                 RCU_INIT_POINTER(chain->filter_chain, tp->next);
 208                 tcf_proto_destroy(tp);
 209                 tp = rtnl_dereference(chain->filter_chain);
 210                 tcf_chain_put(chain);
 211         }
 212 }
 213
 214 static void tcf_chain_destroy(struct tcf_chain *chain)
 215 {
 216         struct tcf_block *block = chain->block;
 217
 218         list_del(&chain->list);
 219         kfree(chain);
 220         if (list_empty(&block->chain_list))
 221                 kfree(block);
 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
 253 tcf_chain_filter_chain_ptr_set(struct tcf_chain *chain,
 254                                struct tcf_proto __rcu **p_filter_chain)
 255 {
 256         chain->p_filter_chain = p_filter_chain;
 257 }
 258
 259 int tcf_block_get(struct tcf_block **p_block,
 260                   struct tcf_proto __rcu **p_filter_chain)
 261 {
 262         struct tcf_block *block = kzalloc(sizeof(*block), GFP_KERNEL);
 263         struct tcf_chain *chain;
 264         int err;
 265
 266         if (!block)
 267                 return -ENOMEM;
 268         INIT_LIST_HEAD(&block->chain_list);
 269         /* Create chain 0 by default, it has to be always present. */
 270         chain = tcf_chain_create(block, 0);
 271         if (!chain) {
 272                 err = -ENOMEM;
 273                 goto err_chain_create;
 274         }
 275         tcf_chain_filter_chain_ptr_set(chain, p_filter_chain);
 276         *p_block = block;
 277         return 0;
 278
 279 err_chain_create:
 280         kfree(block);
 281         return err;
 282 }
 283 EXPORT_SYMBOL(tcf_block_get);
 284
 285 /* XXX: Standalone actions are not allowed to jump to any chain, and bound
 286  * actions should be all removed after flushing.
 287  */
 288 void tcf_block_put(struct tcf_block *block)
 289 {
 290         struct tcf_chain *chain, *tmp;
 291
 292         if (!block)
 293                 return;
 294
 295         /* Hold a refcnt for all chains, so that they don't disappear
 296          * while we are iterating.
 297          */
 298         list_for_each_entry(chain, &block->chain_list, list)
 299                 tcf_chain_hold(chain);
 300
 301         list_for_each_entry(chain, &block->chain_list, list)
 302                 tcf_chain_flush(chain);
 303
 304         /* At this point, all the chains should have refcnt >= 1. */
 305         list_for_each_entry_safe(chain, tmp, &block->chain_list, list)
 306                 tcf_chain_put(chain);
 307
 308         /* Finally, put chain 0 and allow block to be freed. */
 309         chain = list_first_entry(&block->chain_list, struct tcf_chain, list);
 310         tcf_chain_put(chain);
 311 }
 312 EXPORT_SYMBOL(tcf_block_put);
 313
 314 /* Main classifier routine: scans classifier chain attached
 315  * to this qdisc, (optionally) tests for protocol and asks
 316  * specific classifiers.
 317  */
 318 int tcf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 319                  struct tcf_result *res, bool compat_mode)
 320 {
 321 #ifdef CONFIG_NET_CLS_ACT
 322         const int max_reclassify_loop = 4;
 323         const struct tcf_proto *orig_tp = tp;
 324         const struct tcf_proto *first_tp;
 325         int limit = 0;
 326
 327 reclassify:
 328 #endif
 329         for (; tp; tp = rcu_dereference_bh(tp->next)) {
 330                 __be16 protocol = tc_skb_protocol(skb);
 331                 int err;
 332
 333                 if (tp->protocol != protocol &&
 334                     tp->protocol != htons(ETH_P_ALL))
 335                         continue;
 336
 337                 err = tp->classify(skb, tp, res);
 338 #ifdef CONFIG_NET_CLS_ACT
 339                 if (unlikely(err == TC_ACT_RECLASSIFY && !compat_mode)) {
 340                         first_tp = orig_tp;
 341                         goto reset;
 342                 } else if (unlikely(TC_ACT_EXT_CMP(err, TC_ACT_GOTO_CHAIN))) {
 343                         first_tp = res->goto_tp;
 344                         goto reset;
 345                 }
 346 #endif
 347                 if (err >= 0)
 348                         return err;
 349         }
 350
 351         return TC_ACT_UNSPEC; /* signal: continue lookup */
 352 #ifdef CONFIG_NET_CLS_ACT
 353 reset:
 354         if (unlikely(limit++ >= max_reclassify_loop)) {
 355                 net_notice_ratelimited("%s: reclassify loop, rule prio %u, protocol %02x\n",
 356                                        tp->q->ops->id, tp->prio & 0xffff,
 357                                        ntohs(tp->protocol));
 358                 return TC_ACT_SHOT;
 359         }
 360
 361         tp = first_tp;
 362         goto reclassify;
 363 #endif
 364 }
 365 EXPORT_SYMBOL(tcf_classify);
 366
 367 struct tcf_chain_info {
 368         struct tcf_proto __rcu **pprev;
 369         struct tcf_proto __rcu *next;
 370 };
 371
 372 static struct tcf_proto *tcf_chain_tp_prev(struct tcf_chain_info *chain_info)
 373 {
 374         return rtnl_dereference(*chain_info->pprev);
 375 }
 376
 377 static void tcf_chain_tp_insert(struct tcf_chain *chain,
 378                                 struct tcf_chain_info *chain_info,
 379                                 struct tcf_proto *tp)
 380 {
 381         if (chain->p_filter_chain &&
 382             *chain_info->pprev == chain->filter_chain)
 383                 rcu_assign_pointer(*chain->p_filter_chain, tp);
 384         RCU_INIT_POINTER(tp->next, tcf_chain_tp_prev(chain_info));
 385         rcu_assign_pointer(*chain_info->pprev, tp);
 386         tcf_chain_hold(chain);
 387 }
 388
 389 static void tcf_chain_tp_remove(struct tcf_chain *chain,
 390                                 struct tcf_chain_info *chain_info,
 391                                 struct tcf_proto *tp)
 392 {
 393         struct tcf_proto *next = rtnl_dereference(chain_info->next);
 394
 395         if (chain->p_filter_chain && tp == chain->filter_chain)
 396                 RCU_INIT_POINTER(*chain->p_filter_chain, next);
 397         RCU_INIT_POINTER(*chain_info->pprev, next);
 398         tcf_chain_put(chain);
 399 }
 400
 401 static struct tcf_proto *tcf_chain_tp_find(struct tcf_chain *chain,
 402                                            struct tcf_chain_info *chain_info,
 403                                            u32 protocol, u32 prio,
 404                                            bool prio_allocate)
 405 {
 406         struct tcf_proto **pprev;
 407         struct tcf_proto *tp;
 408
 409         /* Check the chain for existence of proto-tcf with this priority */
 410         for (pprev = &chain->filter_chain;
 411              (tp = rtnl_dereference(*pprev)); pprev = &tp->next) {
 412                 if (tp->prio >= prio) {
 413                         if (tp->prio == prio) {
 414                                 if (prio_allocate ||
 415                                     (tp->protocol != protocol && protocol))
 416                                         return ERR_PTR(-EINVAL);
 417                         } else {
 418                                 tp = NULL;
 419                         }
 420                         break;
 421                 }
 422         }
 423         chain_info->pprev = pprev;
 424         chain_info->next = tp ? tp->next : NULL;
 425         return tp;
 426 }
 427
 428 static int tcf_fill_node(struct net *net, struct sk_buff *skb,
 429                          struct tcf_proto *tp, void *fh, u32 portid,
 430                          u32 seq, u16 flags, int event)
 431 {
 432         struct tcmsg *tcm;
 433         struct nlmsghdr  *nlh;
 434         unsigned char *b = skb_tail_pointer(skb);
 435
 436         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*tcm), flags);
 437         if (!nlh)
 438                 goto out_nlmsg_trim;
 439         tcm = nlmsg_data(nlh);
 440         tcm->tcm_family = AF_UNSPEC;
 441         tcm->tcm__pad1 = 0;
 442         tcm->tcm__pad2 = 0;
 443         tcm->tcm_ifindex = qdisc_dev(tp->q)->ifindex;
 444         tcm->tcm_parent = tp->classid;
 445         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 446         if (nla_put_string(skb, TCA_KIND, tp->ops->kind))
 447                 goto nla_put_failure;
 448         if (nla_put_u32(skb, TCA_CHAIN, tp->chain->index))
 449                 goto nla_put_failure;
 450         if (!fh) {
 451                 tcm->tcm_handle = 0;
 452         } else {
 453                 if (tp->ops->dump && tp->ops->dump(net, tp, fh, skb, tcm) < 0)
 454                         goto nla_put_failure;
 455         }
 456         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 457         return skb->len;
 458
 459 out_nlmsg_trim:
 460 nla_put_failure:
 461         nlmsg_trim(skb, b);
 462         return -1;
 463 }
 464
 465 static int tfilter_notify(struct net *net, struct sk_buff *oskb,
 466                           struct nlmsghdr *n, struct tcf_proto *tp,
 467                           void *fh, int event, bool unicast)
 468 {
 469         struct sk_buff *skb;
 470         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 471
 472         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 473         if (!skb)
 474                 return -ENOBUFS;
 475
 476         if (tcf_fill_node(net, skb, tp, fh, portid, n->nlmsg_seq,
 477                           n->nlmsg_flags, event) <= 0) {
 478                 kfree_skb(skb);
 479                 return -EINVAL;
 480         }
 481
 482         if (unicast)
 483                 return netlink_unicast(net->rtnl, skb, portid, MSG_DONTWAIT);
 484
 485         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 486                               n->nlmsg_flags & NLM_F_ECHO);
 487 }
 488
 489 static int tfilter_del_notify(struct net *net, struct sk_buff *oskb,
 490                               struct nlmsghdr *n, struct tcf_proto *tp,
 491                               void *fh, bool unicast, bool *last)
 492 {
 493         struct sk_buff *skb;
 494         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 495         int err;
 496
 497         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 498         if (!skb)
 499                 return -ENOBUFS;
 500
 501         if (tcf_fill_node(net, skb, tp, fh, portid, n->nlmsg_seq,
 502                           n->nlmsg_flags, RTM_DELTFILTER) <= 0) {
 503                 kfree_skb(skb);
 504                 return -EINVAL;
 505         }
 506
 507         err = tp->ops->delete(tp, fh, last);
 508         if (err) {
 509                 kfree_skb(skb);
 510                 return err;
 511         }
 512
 513         if (unicast)
 514                 return netlink_unicast(net->rtnl, skb, portid, MSG_DONTWAIT);
 515
 516         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 517                               n->nlmsg_flags & NLM_F_ECHO);
 518 }
 519
 520 static void tfilter_notify_chain(struct net *net, struct sk_buff *oskb,
 521                                  struct nlmsghdr *n,
 522                                  struct tcf_chain *chain, int event)
 523 {
 524         struct tcf_proto *tp;
 525
 526         for (tp = rtnl_dereference(chain->filter_chain);
 527              tp; tp = rtnl_dereference(tp->next))
 528                 tfilter_notify(net, oskb, n, tp, 0, event, false);
 529 }
 530
 531 /* Add/change/delete/get a filter node */
 532
 533 static int tc_ctl_tfilter(struct sk_buff *skb, struct nlmsghdr *n,
 534                           struct netlink_ext_ack *extack)
 535 {
 536         struct net *net = sock_net(skb->sk);
 537         struct nlattr *tca[TCA_MAX + 1];
 538         struct tcmsg *t;
 539         u32 protocol;
 540         u32 prio;
 541         bool prio_allocate;
 542         u32 parent;
 543         u32 chain_index;
 544         struct net_device *dev;
 545         struct Qdisc  *q;
 546         struct tcf_chain_info chain_info;
 547         struct tcf_chain *chain = NULL;
 548         struct tcf_block *block;
 549         struct tcf_proto *tp;
 550         const struct Qdisc_class_ops *cops;
 551         unsigned long cl;
 552         void *fh;
 553         int err;
 554         int tp_created;
 555
 556         if ((n->nlmsg_type != RTM_GETTFILTER) &&
 557             !netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN))
 558                 return -EPERM;
 559
 560 replay:
 561         tp_created = 0;
 562
 563         err = nlmsg_parse(n, sizeof(*t), tca, TCA_MAX, rtm_tca_policy, extack);
 564         if (err < 0)
 565                 return err;
 566
 567         t = nlmsg_data(n);
 568         protocol = TC_H_MIN(t->tcm_info);
 569         prio = TC_H_MAJ(t->tcm_info);
 570         prio_allocate = false;
 571         parent = t->tcm_parent;
 572         cl = 0;
 573
 574         if (prio == 0) {
 575                 switch (n->nlmsg_type) {
 576                 case RTM_DELTFILTER:
 577                         if (protocol || t->tcm_handle || tca[TCA_KIND])
 578                                 return -ENOENT;
 579                         break;
 580                 case RTM_NEWTFILTER:
 581                         /* If no priority is provided by the user,
 582                          * we allocate one.
 583                          */
 584                         if (n->nlmsg_flags & NLM_F_CREATE) {
 585                                 prio = TC_H_MAKE(0x80000000U, 0U);
 586                                 prio_allocate = true;
 587                                 break;
 588                         }
 589                         /* fall-through */
 590                 default:
 591                         return -ENOENT;
 592                 }
 593         }
 594
 595         /* Find head of filter chain. */
 596
 597         /* Find link */
 598         dev = __dev_get_by_index(net, t->tcm_ifindex);
 599         if (dev == NULL)
 600                 return -ENODEV;
 601
 602         /* Find qdisc */
 603         if (!parent) {
 604                 q = dev->qdisc;
 605                 parent = q->handle;
 606         } else {
 607                 q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent));
 608                 if (q == NULL)
 609                         return -EINVAL;
 610         }
 611
 612         /* Is it classful? */
 613         cops = q->ops->cl_ops;
 614         if (!cops)
 615                 return -EINVAL;
 616
 617         if (!cops->tcf_block)
 618                 return -EOPNOTSUPP;
 619
 620         /* Do we search for filter, attached to class? */
 621         if (TC_H_MIN(parent)) {
 622                 cl = cops->find(q, parent);
 623                 if (cl == 0)
 624                         return -ENOENT;
 625         }
 626
 627         /* And the last stroke */
 628         block = cops->tcf_block(q, cl);
 629         if (!block) {
 630                 err = -EINVAL;
 631                 goto errout;
 632         }
 633
 634         chain_index = tca[TCA_CHAIN] ? nla_get_u32(tca[TCA_CHAIN]) : 0;
 635         if (chain_index > TC_ACT_EXT_VAL_MASK) {
 636                 err = -EINVAL;
 637                 goto errout;
 638         }
 639         chain = tcf_chain_get(block, chain_index,
 640                               n->nlmsg_type == RTM_NEWTFILTER);
 641         if (!chain) {
 642                 err = n->nlmsg_type == RTM_NEWTFILTER ? -ENOMEM : -EINVAL;
 643                 goto errout;
 644         }
 645
 646         if (n->nlmsg_type == RTM_DELTFILTER && prio == 0) {
 647                 tfilter_notify_chain(net, skb, n, chain, RTM_DELTFILTER);
 648                 tcf_chain_flush(chain);
 649                 err = 0;
 650                 goto errout;
 651         }
 652
 653         tp = tcf_chain_tp_find(chain, &chain_info, protocol,
 654                                prio, prio_allocate);
 655         if (IS_ERR(tp)) {
 656                 err = PTR_ERR(tp);
 657                 goto errout;
 658         }
 659
 660         if (tp == NULL) {
 661                 /* Proto-tcf does not exist, create new one */
 662
 663                 if (tca[TCA_KIND] == NULL || !protocol) {
 664                         err = -EINVAL;
 665                         goto errout;
 666                 }
 667
 668                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 669                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 670                         err = -ENOENT;
 671                         goto errout;
 672                 }
 673
 674                 if (prio_allocate)
 675                         prio = tcf_auto_prio(tcf_chain_tp_prev(&chain_info));
 676
 677                 tp = tcf_proto_create(nla_data(tca[TCA_KIND]),
 678                                       protocol, prio, parent, q, chain);
 679                 if (IS_ERR(tp)) {
 680                         err = PTR_ERR(tp);
 681                         goto errout;
 682                 }
 683                 tp_created = 1;
 684         } else if (tca[TCA_KIND] && nla_strcmp(tca[TCA_KIND], tp->ops->kind)) {
 685                 err = -EINVAL;
 686                 goto errout;
 687         }
 688
 689         fh = tp->ops->get(tp, t->tcm_handle);
 690
 691         if (!fh) {
 692                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 693                         tcf_chain_tp_remove(chain, &chain_info, tp);
 694                         tfilter_notify(net, skb, n, tp, fh,
 695                                        RTM_DELTFILTER, false);
 696                         tcf_proto_destroy(tp);
 697                         err = 0;
 698                         goto errout;
 699                 }
 700
 701                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 702                     !(n->nlmsg_flags & NLM_F_CREATE)) {
 703                         err = -ENOENT;
 704                         goto errout;
 705                 }
 706         } else {
 707                 bool last;
 708
 709                 switch (n->nlmsg_type) {
 710                 case RTM_NEWTFILTER:
 711                         if (n->nlmsg_flags & NLM_F_EXCL) {
 712                                 if (tp_created)
 713                                         tcf_proto_destroy(tp);
 714                                 err = -EEXIST;
 715                                 goto errout;
 716                         }
 717                         break;
 718                 case RTM_DELTFILTER:
 719                         err = tfilter_del_notify(net, skb, n, tp, fh, false,
 720                                                  &last);
 721                         if (err)
 722                                 goto errout;
 723                         if (last) {
 724                                 tcf_chain_tp_remove(chain, &chain_info, tp);
 725                                 tcf_proto_destroy(tp);
 726                         }
 727                         goto errout;
 728                 case RTM_GETTFILTER:
 729                         err = tfilter_notify(net, skb, n, tp, fh,
 730                                              RTM_NEWTFILTER, true);
 731                         goto errout;
 732                 default:
 733                         err = -EINVAL;
 734                         goto errout;
 735                 }
 736         }
 737
 738         err = tp->ops->change(net, skb, tp, cl, t->tcm_handle, tca, &fh,
 739                               n->nlmsg_flags & NLM_F_CREATE ? TCA_ACT_NOREPLACE : TCA_ACT_REPLACE);
 740         if (err == 0) {
 741                 if (tp_created)
 742                         tcf_chain_tp_insert(chain, &chain_info, tp);
 743                 tfilter_notify(net, skb, n, tp, fh, RTM_NEWTFILTER, false);
 744         } else {
 745                 if (tp_created)
 746                         tcf_proto_destroy(tp);
 747         }
 748
 749 errout:
 750         if (chain)
 751                 tcf_chain_put(chain);
 752         if (err == -EAGAIN)
 753                 /* Replay the request. */
 754                 goto replay;
 755         return err;
 756 }
 757
 758 struct tcf_dump_args {
 759         struct tcf_walker w;
 760         struct sk_buff *skb;
 761         struct netlink_callback *cb;
 762 };
 763
 764 static int tcf_node_dump(struct tcf_proto *tp, void *n, struct tcf_walker *arg)
 765 {
 766         struct tcf_dump_args *a = (void *)arg;
 767         struct net *net = sock_net(a->skb->sk);
 768
 769         return tcf_fill_node(net, a->skb, tp, n, NETLINK_CB(a->cb->skb).portid,
 770                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI,
 771                              RTM_NEWTFILTER);
 772 }
 773
 774 static bool tcf_chain_dump(struct tcf_chain *chain, struct sk_buff *skb,
 775                            struct netlink_callback *cb,
 776                            long index_start, long *p_index)
 777 {
 778         struct net *net = sock_net(skb->sk);
 779         struct tcmsg *tcm = nlmsg_data(cb->nlh);
 780         struct tcf_dump_args arg;
 781         struct tcf_proto *tp;
 782
 783         for (tp = rtnl_dereference(chain->filter_chain);
 784              tp; tp = rtnl_dereference(tp->next), (*p_index)++) {
 785                 if (*p_index < index_start)
 786                         continue;
 787                 if (TC_H_MAJ(tcm->tcm_info) &&
 788                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 789                         continue;
 790                 if (TC_H_MIN(tcm->tcm_info) &&
 791                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 792                         continue;
 793                 if (*p_index > index_start)
 794                         memset(&cb->args[1], 0,
 795                                sizeof(cb->args) - sizeof(cb->args[0]));
 796                 if (cb->args[1] == 0) {
 797                         if (tcf_fill_node(net, skb, tp, 0,
 798                                           NETLINK_CB(cb->skb).portid,
 799                                           cb->nlh->nlmsg_seq, NLM_F_MULTI,
 800                                           RTM_NEWTFILTER) <= 0)
 801                                 return false;
 802
 803                         cb->args[1] = 1;
 804                 }
 805                 if (!tp->ops->walk)
 806                         continue;
 807                 arg.w.fn = tcf_node_dump;
 808                 arg.skb = skb;
 809                 arg.cb = cb;
 810                 arg.w.stop = 0;
 811                 arg.w.skip = cb->args[1] - 1;
 812                 arg.w.count = 0;
 813                 tp->ops->walk(tp, &arg.w);
 814                 cb->args[1] = arg.w.count + 1;
 815                 if (arg.w.stop)
 816                         return false;
 817         }
 818         return true;
 819 }
 820
 821 /* called with RTNL */
 822 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
 823 {
 824         struct net *net = sock_net(skb->sk);
 825         struct nlattr *tca[TCA_MAX + 1];
 826         struct net_device *dev;
 827         struct Qdisc *q;
 828         struct tcf_block *block;
 829         struct tcf_chain *chain;
 830         struct tcmsg *tcm = nlmsg_data(cb->nlh);
 831         unsigned long cl = 0;
 832         const struct Qdisc_class_ops *cops;
 833         long index_start;
 834         long index;
 835         int err;
 836
 837         if (nlmsg_len(cb->nlh) < sizeof(*tcm))
 838                 return skb->len;
 839
 840         err = nlmsg_parse(cb->nlh, sizeof(*tcm), tca, TCA_MAX, rtm_tca_policy,
 841                           NULL);
 842         if (err)
 843                 return err;
 844
 845         dev = __dev_get_by_index(net, tcm->tcm_ifindex);
 846         if (!dev)
 847                 return skb->len;
 848
 849         if (!tcm->tcm_parent)
 850                 q = dev->qdisc;
 851         else
 852                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
 853         if (!q)
 854                 goto out;
 855         cops = q->ops->cl_ops;
 856         if (!cops)
 857                 goto out;
 858         if (!cops->tcf_block)
 859                 goto out;
 860         if (TC_H_MIN(tcm->tcm_parent)) {
 861                 cl = cops->find(q, tcm->tcm_parent);
 862                 if (cl == 0)
 863                         goto out;
 864         }
 865         block = cops->tcf_block(q, cl);
 866         if (!block)
 867                 goto out;
 868
 869         index_start = cb->args[0];
 870         index = 0;
 871
 872         list_for_each_entry(chain, &block->chain_list, list) {
 873                 if (tca[TCA_CHAIN] &&
 874                     nla_get_u32(tca[TCA_CHAIN]) != chain->index)
 875                         continue;
 876                 if (!tcf_chain_dump(chain, skb, cb, index_start, &index)) {
 877                         err = -EMSGSIZE;
 878                         break;
 879                 }
 880         }
 881
 882         cb->args[0] = index;
 883
 884 out:
 885         /* If we did no progress, the error (EMSGSIZE) is real */
 886         if (skb->len == 0 && err)
 887                 return err;
 888         return skb->len;
 889 }
 890
 891 void tcf_exts_destroy(struct tcf_exts *exts)
 892 {
 893 #ifdef CONFIG_NET_CLS_ACT
 894         LIST_HEAD(actions);
 895
 896         ASSERT_RTNL();
 897         tcf_exts_to_list(exts, &actions);
 898         tcf_action_destroy(&actions, TCA_ACT_UNBIND);
 899         kfree(exts->actions);
 900         exts->nr_actions = 0;
 901 #endif
 902 }
 903 EXPORT_SYMBOL(tcf_exts_destroy);
 904
 905 int tcf_exts_validate(struct net *net, struct tcf_proto *tp, struct nlattr **tb,
 906                       struct nlattr *rate_tlv, struct tcf_exts *exts, bool ovr)
 907 {
 908 #ifdef CONFIG_NET_CLS_ACT
 909         {
 910                 struct tc_action *act;
 911
 912                 if (exts->police && tb[exts->police]) {
 913                         act = tcf_action_init_1(net, tp, tb[exts->police],
 914                                                 rate_tlv, "police", ovr,
 915                                                 TCA_ACT_BIND);
 916                         if (IS_ERR(act))
 917                                 return PTR_ERR(act);
 918
 919                         act->type = exts->type = TCA_OLD_COMPAT;
 920                         exts->actions[0] = act;
 921                         exts->nr_actions = 1;
 922                 } else if (exts->action && tb[exts->action]) {
 923                         LIST_HEAD(actions);
 924                         int err, i = 0;
 925
 926                         err = tcf_action_init(net, tp, tb[exts->action],
 927                                               rate_tlv, NULL, ovr, TCA_ACT_BIND,
 928                                               &actions);
 929                         if (err)
 930                                 return err;
 931                         list_for_each_entry(act, &actions, list)
 932                                 exts->actions[i++] = act;
 933                         exts->nr_actions = i;
 934                 }
 935                 exts->net = net;
 936         }
 937 #else
 938         if ((exts->action && tb[exts->action]) ||
 939             (exts->police && tb[exts->police]))
 940                 return -EOPNOTSUPP;
 941 #endif
 942
 943         return 0;
 944 }
 945 EXPORT_SYMBOL(tcf_exts_validate);
 946
 947 void tcf_exts_change(struct tcf_exts *dst, struct tcf_exts *src)
 948 {
 949 #ifdef CONFIG_NET_CLS_ACT
 950         struct tcf_exts old = *dst;
 951
 952         *dst = *src;
 953         tcf_exts_destroy(&old);
 954 #endif
 955 }
 956 EXPORT_SYMBOL(tcf_exts_change);
 957
 958 #ifdef CONFIG_NET_CLS_ACT
 959 static struct tc_action *tcf_exts_first_act(struct tcf_exts *exts)
 960 {
 961         if (exts->nr_actions == 0)
 962                 return NULL;
 963         else
 964                 return exts->actions[0];
 965 }
 966 #endif
 967
 968 int tcf_exts_dump(struct sk_buff *skb, struct tcf_exts *exts)
 969 {
 970 #ifdef CONFIG_NET_CLS_ACT
 971         struct nlattr *nest;
 972
 973         if (exts->action && tcf_exts_has_actions(exts)) {
 974                 /*
 975                  * again for backward compatible mode - we want
 976                  * to work with both old and new modes of entering
 977                  * tc data even if iproute2  was newer - jhs
 978                  */
 979                 if (exts->type != TCA_OLD_COMPAT) {
 980                         LIST_HEAD(actions);
 981
 982                         nest = nla_nest_start(skb, exts->action);
 983                         if (nest == NULL)
 984                                 goto nla_put_failure;
 985
 986                         tcf_exts_to_list(exts, &actions);
 987                         if (tcf_action_dump(skb, &actions, 0, 0) < 0)
 988                                 goto nla_put_failure;
 989                         nla_nest_end(skb, nest);
 990                 } else if (exts->police) {
 991                         struct tc_action *act = tcf_exts_first_act(exts);
 992                         nest = nla_nest_start(skb, exts->police);
 993                         if (nest == NULL || !act)
 994                                 goto nla_put_failure;
 995                         if (tcf_action_dump_old(skb, act, 0, 0) < 0)
 996                                 goto nla_put_failure;
 997                         nla_nest_end(skb, nest);
 998                 }
 999         }
1000         return 0;
1001
1002 nla_put_failure:
1003         nla_nest_cancel(skb, nest);
1004         return -1;
1005 #else
1006         return 0;
1007 #endif
1008 }
1009 EXPORT_SYMBOL(tcf_exts_dump);
1010
1011
1012 int tcf_exts_dump_stats(struct sk_buff *skb, struct tcf_exts *exts)
1013 {
1014 #ifdef CONFIG_NET_CLS_ACT
1015         struct tc_action *a = tcf_exts_first_act(exts);
1016         if (a != NULL && tcf_action_copy_stats(skb, a, 1) < 0)
1017                 return -1;
1018 #endif
1019         return 0;
1020 }
1021 EXPORT_SYMBOL(tcf_exts_dump_stats);
1022
1023 int tcf_exts_get_dev(struct net_device *dev, struct tcf_exts *exts,
1024                      struct net_device **hw_dev)
1025 {
1026 #ifdef CONFIG_NET_CLS_ACT
1027         const struct tc_action *a;
1028         LIST_HEAD(actions);
1029
1030         if (!tcf_exts_has_actions(exts))
1031                 return -EINVAL;
1032
1033         tcf_exts_to_list(exts, &actions);
1034         list_for_each_entry(a, &actions, list) {
1035                 if (a->ops->get_dev) {
1036                         a->ops->get_dev(a, dev_net(dev), hw_dev);
1037                         break;
1038                 }
1039         }
1040         if (*hw_dev)
1041                 return 0;
1042 #endif
1043         return -EOPNOTSUPP;
1044 }
1045 EXPORT_SYMBOL(tcf_exts_get_dev);
1046
1047 static int __init tc_filter_init(void)
1048 {
1049         tc_filter_wq = alloc_ordered_workqueue("tc_filter_workqueue", 0);
1050         if (!tc_filter_wq)
1051                 return -ENOMEM;
1052
1053         rtnl_register(PF_UNSPEC, RTM_NEWTFILTER, tc_ctl_tfilter, NULL, 0);
1054         rtnl_register(PF_UNSPEC, RTM_DELTFILTER, tc_ctl_tfilter, NULL, 0);
1055         rtnl_register(PF_UNSPEC, RTM_GETTFILTER, tc_ctl_tfilter,
1056                       tc_dump_tfilter, 0);
1057
1058         return 0;
1059 }
1060
1061 subsys_initcall(tc_filter_init);