Pileus Git - ~andy/linux/blob - 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/skbuff.h>
  23 #include <linux/init.h>
  24 #include <linux/kmod.h>
  25 #include <linux/err.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(struct nlattr *kind)
  42 {
  43         const struct tcf_proto_ops *t = NULL;
  44
  45         if (kind) {
  46                 read_lock(&cls_mod_lock);
  47                 list_for_each_entry(t, &tcf_proto_base, head) {
  48                         if (nla_strcmp(kind, t->kind) == 0) {
  49                                 if (!try_module_get(t->owner))
  50                                         t = NULL;
  51                                 break;
  52                         }
  53                 }
  54                 read_unlock(&cls_mod_lock);
  55         }
  56         return t;
  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 int unregister_tcf_proto_ops(struct tcf_proto_ops *ops)
  80 {
  81         struct tcf_proto_ops *t;
  82         int rc = -ENOENT;
  83
  84         write_lock(&cls_mod_lock);
  85         list_for_each_entry(t, &tcf_proto_base, head)
  86                 if (t == ops)
  87                         break;
  88
  89         if (!t)
  90                 goto out;
  91         list_del(&t->head);
  92         rc = 0;
  93 out:
  94         write_unlock(&cls_mod_lock);
  95         return rc;
  96 }
  97 EXPORT_SYMBOL(unregister_tcf_proto_ops);
  98
  99 static int tfilter_notify(struct net *net, struct sk_buff *oskb,
 100                           struct nlmsghdr *n, struct tcf_proto *tp,
 101                           unsigned long fh, int event);
 102
 103
 104 /* Select new prio value from the range, managed by kernel. */
 105
 106 static inline u32 tcf_auto_prio(struct tcf_proto *tp)
 107 {
 108         u32 first = TC_H_MAKE(0xC0000000U, 0U);
 109
 110         if (tp)
 111                 first = tp->prio - 1;
 112
 113         return first;
 114 }
 115
 116 /* Add/change/delete/get a filter node */
 117
 118 static int tc_ctl_tfilter(struct sk_buff *skb, struct nlmsghdr *n)
 119 {
 120         struct net *net = sock_net(skb->sk);
 121         struct nlattr *tca[TCA_MAX + 1];
 122         spinlock_t *root_lock;
 123         struct tcmsg *t;
 124         u32 protocol;
 125         u32 prio;
 126         u32 nprio;
 127         u32 parent;
 128         struct net_device *dev;
 129         struct Qdisc  *q;
 130         struct tcf_proto **back, **chain;
 131         struct tcf_proto *tp;
 132         const struct tcf_proto_ops *tp_ops;
 133         const struct Qdisc_class_ops *cops;
 134         unsigned long cl;
 135         unsigned long fh;
 136         int err;
 137         int tp_created = 0;
 138
 139         if ((n->nlmsg_type != RTM_GETTFILTER) && !capable(CAP_NET_ADMIN))
 140                 return -EPERM;
 141
 142 replay:
 143         err = nlmsg_parse(n, sizeof(*t), tca, TCA_MAX, NULL);
 144         if (err < 0)
 145                 return err;
 146
 147         t = nlmsg_data(n);
 148         protocol = TC_H_MIN(t->tcm_info);
 149         prio = TC_H_MAJ(t->tcm_info);
 150         nprio = prio;
 151         parent = t->tcm_parent;
 152         cl = 0;
 153
 154         if (prio == 0) {
 155                 /* If no priority is given, user wants we allocated it. */
 156                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 157                     !(n->nlmsg_flags & NLM_F_CREATE))
 158                         return -ENOENT;
 159                 prio = TC_H_MAKE(0x80000000U, 0U);
 160         }
 161
 162         /* Find head of filter chain. */
 163
 164         /* Find link */
 165         dev = __dev_get_by_index(net, t->tcm_ifindex);
 166         if (dev == NULL)
 167                 return -ENODEV;
 168
 169         /* Find qdisc */
 170         if (!parent) {
 171                 q = dev->qdisc;
 172                 parent = q->handle;
 173         } else {
 174                 q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent));
 175                 if (q == NULL)
 176                         return -EINVAL;
 177         }
 178
 179         /* Is it classful? */
 180         cops = q->ops->cl_ops;
 181         if (!cops)
 182                 return -EINVAL;
 183
 184         if (cops->tcf_chain == NULL)
 185                 return -EOPNOTSUPP;
 186
 187         /* Do we search for filter, attached to class? */
 188         if (TC_H_MIN(parent)) {
 189                 cl = cops->get(q, parent);
 190                 if (cl == 0)
 191                         return -ENOENT;
 192         }
 193
 194         /* And the last stroke */
 195         chain = cops->tcf_chain(q, cl);
 196         err = -EINVAL;
 197         if (chain == NULL)
 198                 goto errout;
 199
 200         /* Check the chain for existence of proto-tcf with this priority */
 201         for (back = chain; (tp = *back) != NULL; back = &tp->next) {
 202                 if (tp->prio >= prio) {
 203                         if (tp->prio == prio) {
 204                                 if (!nprio ||
 205                                     (tp->protocol != protocol && protocol))
 206                                         goto errout;
 207                         } else
 208                                 tp = NULL;
 209                         break;
 210                 }
 211         }
 212
 213         root_lock = qdisc_root_sleeping_lock(q);
 214
 215         if (tp == NULL) {
 216                 /* Proto-tcf does not exist, create new one */
 217
 218                 if (tca[TCA_KIND] == NULL || !protocol)
 219                         goto errout;
 220
 221                 err = -ENOENT;
 222                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 223                     !(n->nlmsg_flags & NLM_F_CREATE))
 224                         goto errout;
 225
 226
 227                 /* Create new proto tcf */
 228
 229                 err = -ENOBUFS;
 230                 tp = kzalloc(sizeof(*tp), GFP_KERNEL);
 231                 if (tp == NULL)
 232                         goto errout;
 233                 err = -ENOENT;
 234                 tp_ops = tcf_proto_lookup_ops(tca[TCA_KIND]);
 235                 if (tp_ops == NULL) {
 236 #ifdef CONFIG_MODULES
 237                         struct nlattr *kind = tca[TCA_KIND];
 238                         char name[IFNAMSIZ];
 239
 240                         if (kind != NULL &&
 241                             nla_strlcpy(name, kind, IFNAMSIZ) < IFNAMSIZ) {
 242                                 rtnl_unlock();
 243                                 request_module("cls_%s", name);
 244                                 rtnl_lock();
 245                                 tp_ops = tcf_proto_lookup_ops(kind);
 246                                 /* We dropped the RTNL semaphore in order to
 247                                  * perform the module load.  So, even if we
 248                                  * succeeded in loading the module we have to
 249                                  * replay the request.  We indicate this using
 250                                  * -EAGAIN.
 251                                  */
 252                                 if (tp_ops != NULL) {
 253                                         module_put(tp_ops->owner);
 254                                         err = -EAGAIN;
 255                                 }
 256                         }
 257 #endif
 258                         kfree(tp);
 259                         goto errout;
 260                 }
 261                 tp->ops = tp_ops;
 262                 tp->protocol = protocol;
 263                 tp->prio = nprio ? : TC_H_MAJ(tcf_auto_prio(*back));
 264                 tp->q = q;
 265                 tp->classify = tp_ops->classify;
 266                 tp->classid = parent;
 267
 268                 err = tp_ops->init(tp);
 269                 if (err != 0) {
 270                         module_put(tp_ops->owner);
 271                         kfree(tp);
 272                         goto errout;
 273                 }
 274
 275                 tp_created = 1;
 276
 277         } else if (tca[TCA_KIND] && nla_strcmp(tca[TCA_KIND], tp->ops->kind))
 278                 goto errout;
 279
 280         fh = tp->ops->get(tp, t->tcm_handle);
 281
 282         if (fh == 0) {
 283                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 284                         spin_lock_bh(root_lock);
 285                         *back = tp->next;
 286                         spin_unlock_bh(root_lock);
 287
 288                         tfilter_notify(net, skb, n, tp, fh, RTM_DELTFILTER);
 289                         tcf_destroy(tp);
 290                         err = 0;
 291                         goto errout;
 292                 }
 293
 294                 err = -ENOENT;
 295                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 296                     !(n->nlmsg_flags & NLM_F_CREATE))
 297                         goto errout;
 298         } else {
 299                 switch (n->nlmsg_type) {
 300                 case RTM_NEWTFILTER:
 301                         err = -EEXIST;
 302                         if (n->nlmsg_flags & NLM_F_EXCL) {
 303                                 if (tp_created)
 304                                         tcf_destroy(tp);
 305                                 goto errout;
 306                         }
 307                         break;
 308                 case RTM_DELTFILTER:
 309                         err = tp->ops->delete(tp, fh);
 310                         if (err == 0)
 311                                 tfilter_notify(net, skb, n, tp, fh, RTM_DELTFILTER);
 312                         goto errout;
 313                 case RTM_GETTFILTER:
 314                         err = tfilter_notify(net, skb, n, tp, fh, RTM_NEWTFILTER);
 315                         goto errout;
 316                 default:
 317                         err = -EINVAL;
 318                         goto errout;
 319                 }
 320         }
 321
 322         err = tp->ops->change(net, skb, tp, cl, t->tcm_handle, tca, &fh);
 323         if (err == 0) {
 324                 if (tp_created) {
 325                         spin_lock_bh(root_lock);
 326                         tp->next = *back;
 327                         *back = tp;
 328                         spin_unlock_bh(root_lock);
 329                 }
 330                 tfilter_notify(net, skb, n, tp, fh, RTM_NEWTFILTER);
 331         } else {
 332                 if (tp_created)
 333                         tcf_destroy(tp);
 334         }
 335
 336 errout:
 337         if (cl)
 338                 cops->put(q, cl);
 339         if (err == -EAGAIN)
 340                 /* Replay the request. */
 341                 goto replay;
 342         return err;
 343 }
 344
 345 static int tcf_fill_node(struct sk_buff *skb, struct tcf_proto *tp,
 346                          unsigned long fh, u32 portid, u32 seq, u16 flags, int event)
 347 {
 348         struct tcmsg *tcm;
 349         struct nlmsghdr  *nlh;
 350         unsigned char *b = skb_tail_pointer(skb);
 351
 352         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*tcm), flags);
 353         if (!nlh)
 354                 goto out_nlmsg_trim;
 355         tcm = nlmsg_data(nlh);
 356         tcm->tcm_family = AF_UNSPEC;
 357         tcm->tcm__pad1 = 0;
 358         tcm->tcm__pad2 = 0;
 359         tcm->tcm_ifindex = qdisc_dev(tp->q)->ifindex;
 360         tcm->tcm_parent = tp->classid;
 361         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 362         if (nla_put_string(skb, TCA_KIND, tp->ops->kind))
 363                 goto nla_put_failure;
 364         tcm->tcm_handle = fh;
 365         if (RTM_DELTFILTER != event) {
 366                 tcm->tcm_handle = 0;
 367                 if (tp->ops->dump && tp->ops->dump(tp, fh, skb, tcm) < 0)
 368                         goto nla_put_failure;
 369         }
 370         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 371         return skb->len;
 372
 373 out_nlmsg_trim:
 374 nla_put_failure:
 375         nlmsg_trim(skb, b);
 376         return -1;
 377 }
 378
 379 static int tfilter_notify(struct net *net, struct sk_buff *oskb,
 380                           struct nlmsghdr *n, struct tcf_proto *tp,
 381                           unsigned long fh, int event)
 382 {
 383         struct sk_buff *skb;
 384         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 385
 386         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 387         if (!skb)
 388                 return -ENOBUFS;
 389
 390         if (tcf_fill_node(skb, tp, fh, portid, n->nlmsg_seq, 0, event) <= 0) {
 391                 kfree_skb(skb);
 392                 return -EINVAL;
 393         }
 394
 395         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 396                               n->nlmsg_flags & NLM_F_ECHO);
 397 }
 398
 399 struct tcf_dump_args {
 400         struct tcf_walker w;
 401         struct sk_buff *skb;
 402         struct netlink_callback *cb;
 403 };
 404
 405 static int tcf_node_dump(struct tcf_proto *tp, unsigned long n,
 406                          struct tcf_walker *arg)
 407 {
 408         struct tcf_dump_args *a = (void *)arg;
 409
 410         return tcf_fill_node(a->skb, tp, n, NETLINK_CB(a->cb->skb).portid,
 411                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER);
 412 }
 413
 414 /* called with RTNL */
 415 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
 416 {
 417         struct net *net = sock_net(skb->sk);
 418         int t;
 419         int s_t;
 420         struct net_device *dev;
 421         struct Qdisc *q;
 422         struct tcf_proto *tp, **chain;
 423         struct tcmsg *tcm = nlmsg_data(cb->nlh);
 424         unsigned long cl = 0;
 425         const struct Qdisc_class_ops *cops;
 426         struct tcf_dump_args arg;
 427
 428         if (nlmsg_len(cb->nlh) < sizeof(*tcm))
 429                 return skb->len;
 430         dev = __dev_get_by_index(net, tcm->tcm_ifindex);
 431         if (!dev)
 432                 return skb->len;
 433
 434         if (!tcm->tcm_parent)
 435                 q = dev->qdisc;
 436         else
 437                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
 438         if (!q)
 439                 goto out;
 440         cops = q->ops->cl_ops;
 441         if (!cops)
 442                 goto errout;
 443         if (cops->tcf_chain == NULL)
 444                 goto errout;
 445         if (TC_H_MIN(tcm->tcm_parent)) {
 446                 cl = cops->get(q, tcm->tcm_parent);
 447                 if (cl == 0)
 448                         goto errout;
 449         }
 450         chain = cops->tcf_chain(q, cl);
 451         if (chain == NULL)
 452                 goto errout;
 453
 454         s_t = cb->args[0];
 455
 456         for (tp = *chain, t = 0; tp; tp = tp->next, t++) {
 457                 if (t < s_t)
 458                         continue;
 459                 if (TC_H_MAJ(tcm->tcm_info) &&
 460                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 461                         continue;
 462                 if (TC_H_MIN(tcm->tcm_info) &&
 463                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 464                         continue;
 465                 if (t > s_t)
 466                         memset(&cb->args[1], 0, sizeof(cb->args)-sizeof(cb->args[0]));
 467                 if (cb->args[1] == 0) {
 468                         if (tcf_fill_node(skb, tp, 0, NETLINK_CB(cb->skb).portid,
 469                                           cb->nlh->nlmsg_seq, NLM_F_MULTI,
 470                                           RTM_NEWTFILTER) <= 0)
 471                                 break;
 472
 473                         cb->args[1] = 1;
 474                 }
 475                 if (tp->ops->walk == NULL)
 476                         continue;
 477                 arg.w.fn = tcf_node_dump;
 478                 arg.skb = skb;
 479                 arg.cb = cb;
 480                 arg.w.stop = 0;
 481                 arg.w.skip = cb->args[1] - 1;
 482                 arg.w.count = 0;
 483                 tp->ops->walk(tp, &arg.w);
 484                 cb->args[1] = arg.w.count + 1;
 485                 if (arg.w.stop)
 486                         break;
 487         }
 488
 489         cb->args[0] = t;
 490
 491 errout:
 492         if (cl)
 493                 cops->put(q, cl);
 494 out:
 495         return skb->len;
 496 }
 497
 498 void tcf_exts_destroy(struct tcf_proto *tp, struct tcf_exts *exts)
 499 {
 500 #ifdef CONFIG_NET_CLS_ACT
 501         tcf_action_destroy(&exts->actions, TCA_ACT_UNBIND);
 502         INIT_LIST_HEAD(&exts->actions);
 503 #endif
 504 }
 505 EXPORT_SYMBOL(tcf_exts_destroy);
 506
 507 int tcf_exts_validate(struct net *net, struct tcf_proto *tp, struct nlattr **tb,
 508                   struct nlattr *rate_tlv, struct tcf_exts *exts)
 509 {
 510 #ifdef CONFIG_NET_CLS_ACT
 511         {
 512                 struct tc_action *act;
 513
 514                 INIT_LIST_HEAD(&exts->actions);
 515                 if (exts->police && tb[exts->police]) {
 516                         act = tcf_action_init_1(net, tb[exts->police], rate_tlv,
 517                                                 "police", TCA_ACT_NOREPLACE,
 518                                                 TCA_ACT_BIND);
 519                         if (IS_ERR(act))
 520                                 return PTR_ERR(act);
 521
 522                         act->type = exts->type = TCA_OLD_COMPAT;
 523                         list_add(&act->list, &exts->actions);
 524                 } else if (exts->action && tb[exts->action]) {
 525                         int err;
 526                         err = tcf_action_init(net, tb[exts->action], rate_tlv,
 527                                               NULL, TCA_ACT_NOREPLACE,
 528                                               TCA_ACT_BIND, &exts->actions);
 529                         if (err)
 530                                 return err;
 531                 }
 532         }
 533 #else
 534         if ((exts->action && tb[exts->action]) ||
 535             (exts->police && tb[exts->police]))
 536                 return -EOPNOTSUPP;
 537 #endif
 538
 539         return 0;
 540 }
 541 EXPORT_SYMBOL(tcf_exts_validate);
 542
 543 void tcf_exts_change(struct tcf_proto *tp, struct tcf_exts *dst,
 544                      struct tcf_exts *src)
 545 {
 546 #ifdef CONFIG_NET_CLS_ACT
 547         if (!list_empty(&src->actions)) {
 548                 LIST_HEAD(tmp);
 549                 tcf_tree_lock(tp);
 550                 list_splice_init(&dst->actions, &tmp);
 551                 list_splice(&src->actions, &dst->actions);
 552                 tcf_tree_unlock(tp);
 553                 tcf_action_destroy(&tmp, TCA_ACT_UNBIND);
 554         }
 555 #endif
 556 }
 557 EXPORT_SYMBOL(tcf_exts_change);
 558
 559 #define tcf_exts_first_act(ext) \
 560                 list_first_entry(&(exts)->actions, struct tc_action, list)
 561
 562 int tcf_exts_dump(struct sk_buff *skb, struct tcf_exts *exts)
 563 {
 564 #ifdef CONFIG_NET_CLS_ACT
 565         if (exts->action && !list_empty(&exts->actions)) {
 566                 /*
 567                  * again for backward compatible mode - we want
 568                  * to work with both old and new modes of entering
 569                  * tc data even if iproute2  was newer - jhs
 570                  */
 571                 struct nlattr *nest;
 572                 if (exts->type != TCA_OLD_COMPAT) {
 573                         nest = nla_nest_start(skb, exts->action);
 574                         if (nest == NULL)
 575                                 goto nla_put_failure;
 576                         if (tcf_action_dump(skb, &exts->actions, 0, 0) < 0)
 577                                 goto nla_put_failure;
 578                         nla_nest_end(skb, nest);
 579                 } else if (exts->police) {
 580                         struct tc_action *act = tcf_exts_first_act(exts);
 581                         nest = nla_nest_start(skb, exts->police);
 582                         if (nest == NULL)
 583                                 goto nla_put_failure;
 584                         if (tcf_action_dump_old(skb, act, 0, 0) < 0)
 585                                 goto nla_put_failure;
 586                         nla_nest_end(skb, nest);
 587                 }
 588         }
 589 #endif
 590         return 0;
 591 nla_put_failure: __attribute__ ((unused))
 592         return -1;
 593 }
 594 EXPORT_SYMBOL(tcf_exts_dump);
 595
 596
 597 int tcf_exts_dump_stats(struct sk_buff *skb, struct tcf_exts *exts)
 598 {
 599 #ifdef CONFIG_NET_CLS_ACT
 600         struct tc_action *a = tcf_exts_first_act(exts);
 601         if (tcf_action_copy_stats(skb, a, 1) < 0)
 602                 return -1;
 603 #endif
 604         return 0;
 605 }
 606 EXPORT_SYMBOL(tcf_exts_dump_stats);
 607
 608 static int __init tc_filter_init(void)
 609 {
 610         rtnl_register(PF_UNSPEC, RTM_NEWTFILTER, tc_ctl_tfilter, NULL, NULL);
 611         rtnl_register(PF_UNSPEC, RTM_DELTFILTER, tc_ctl_tfilter, NULL, NULL);
 612         rtnl_register(PF_UNSPEC, RTM_GETTFILTER, tc_ctl_tfilter,
 613                       tc_dump_tfilter, NULL);
 614
 615         return 0;
 616 }
 617
 618 subsys_initcall(tc_filter_init);