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
  35 static struct tcf_proto_ops *tcf_proto_base __read_mostly;
  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(struct nlattr *kind)
  43 {
  44         const struct tcf_proto_ops *t = NULL;
  45
  46         if (kind) {
  47                 read_lock(&cls_mod_lock);
  48                 for (t = tcf_proto_base; t; t = t->next) {
  49                         if (nla_strcmp(kind, t->kind) == 0) {
  50                                 if (!try_module_get(t->owner))
  51                                         t = NULL;
  52                                 break;
  53                         }
  54                 }
  55                 read_unlock(&cls_mod_lock);
  56         }
  57         return t;
  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, **tp;
  65         int rc = -EEXIST;
  66
  67         write_lock(&cls_mod_lock);
  68         for (tp = &tcf_proto_base; (t = *tp) != NULL; tp = &t->next)
  69                 if (!strcmp(ops->kind, t->kind))
  70                         goto out;
  71
  72         ops->next = NULL;
  73         *tp = ops;
  74         rc = 0;
  75 out:
  76         write_unlock(&cls_mod_lock);
  77         return rc;
  78 }
  79 EXPORT_SYMBOL(register_tcf_proto_ops);
  80
  81 int unregister_tcf_proto_ops(struct tcf_proto_ops *ops)
  82 {
  83         struct tcf_proto_ops *t, **tp;
  84         int rc = -ENOENT;
  85
  86         write_lock(&cls_mod_lock);
  87         for (tp = &tcf_proto_base; (t = *tp) != NULL; tp = &t->next)
  88                 if (t == ops)
  89                         break;
  90
  91         if (!t)
  92                 goto out;
  93         *tp = t->next;
  94         rc = 0;
  95 out:
  96         write_unlock(&cls_mod_lock);
  97         return rc;
  98 }
  99 EXPORT_SYMBOL(unregister_tcf_proto_ops);
 100
 101 static int tfilter_notify(struct net *net, struct sk_buff *oskb,
 102                           struct nlmsghdr *n, struct tcf_proto *tp,
 103                           unsigned long fh, int event);
 104
 105
 106 /* Select new prio value from the range, managed by kernel. */
 107
 108 static inline u32 tcf_auto_prio(struct tcf_proto *tp)
 109 {
 110         u32 first = TC_H_MAKE(0xC0000000U, 0U);
 111
 112         if (tp)
 113                 first = tp->prio - 1;
 114
 115         return first;
 116 }
 117
 118 /* Add/change/delete/get a filter node */
 119
 120 static int tc_ctl_tfilter(struct sk_buff *skb, struct nlmsghdr *n)
 121 {
 122         struct net *net = sock_net(skb->sk);
 123         struct nlattr *tca[TCA_MAX + 1];
 124         spinlock_t *root_lock;
 125         struct tcmsg *t;
 126         u32 protocol;
 127         u32 prio;
 128         u32 nprio;
 129         u32 parent;
 130         struct net_device *dev;
 131         struct Qdisc  *q;
 132         struct tcf_proto **back, **chain;
 133         struct tcf_proto *tp;
 134         const struct tcf_proto_ops *tp_ops;
 135         const struct Qdisc_class_ops *cops;
 136         unsigned long cl;
 137         unsigned long fh;
 138         int err;
 139         int tp_created = 0;
 140
 141         if ((n->nlmsg_type != RTM_GETTFILTER) && !netlink_capable(skb, CAP_NET_ADMIN))
 142                 return -EPERM;
 143
 144 replay:
 145         err = nlmsg_parse(n, sizeof(*t), tca, TCA_MAX, NULL);
 146         if (err < 0)
 147                 return err;
 148
 149         t = nlmsg_data(n);
 150         protocol = TC_H_MIN(t->tcm_info);
 151         prio = TC_H_MAJ(t->tcm_info);
 152         nprio = prio;
 153         parent = t->tcm_parent;
 154         cl = 0;
 155
 156         if (prio == 0) {
 157                 /* If no priority is given, user wants we allocated it. */
 158                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 159                     !(n->nlmsg_flags & NLM_F_CREATE))
 160                         return -ENOENT;
 161                 prio = TC_H_MAKE(0x80000000U, 0U);
 162         }
 163
 164         /* Find head of filter chain. */
 165
 166         /* Find link */
 167         dev = __dev_get_by_index(net, t->tcm_ifindex);
 168         if (dev == NULL)
 169                 return -ENODEV;
 170
 171         /* Find qdisc */
 172         if (!parent) {
 173                 q = dev->qdisc;
 174                 parent = q->handle;
 175         } else {
 176                 q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent));
 177                 if (q == NULL)
 178                         return -EINVAL;
 179         }
 180
 181         /* Is it classful? */
 182         cops = q->ops->cl_ops;
 183         if (!cops)
 184                 return -EINVAL;
 185
 186         if (cops->tcf_chain == NULL)
 187                 return -EOPNOTSUPP;
 188
 189         /* Do we search for filter, attached to class? */
 190         if (TC_H_MIN(parent)) {
 191                 cl = cops->get(q, parent);
 192                 if (cl == 0)
 193                         return -ENOENT;
 194         }
 195
 196         /* And the last stroke */
 197         chain = cops->tcf_chain(q, cl);
 198         err = -EINVAL;
 199         if (chain == NULL)
 200                 goto errout;
 201
 202         /* Check the chain for existence of proto-tcf with this priority */
 203         for (back = chain; (tp = *back) != NULL; back = &tp->next) {
 204                 if (tp->prio >= prio) {
 205                         if (tp->prio == prio) {
 206                                 if (!nprio ||
 207                                     (tp->protocol != protocol && protocol))
 208                                         goto errout;
 209                         } else
 210                                 tp = NULL;
 211                         break;
 212                 }
 213         }
 214
 215         root_lock = qdisc_root_sleeping_lock(q);
 216
 217         if (tp == NULL) {
 218                 /* Proto-tcf does not exist, create new one */
 219
 220                 if (tca[TCA_KIND] == NULL || !protocol)
 221                         goto errout;
 222
 223                 err = -ENOENT;
 224                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 225                     !(n->nlmsg_flags & NLM_F_CREATE))
 226                         goto errout;
 227
 228
 229                 /* Create new proto tcf */
 230
 231                 err = -ENOBUFS;
 232                 tp = kzalloc(sizeof(*tp), GFP_KERNEL);
 233                 if (tp == NULL)
 234                         goto errout;
 235                 err = -ENOENT;
 236                 tp_ops = tcf_proto_lookup_ops(tca[TCA_KIND]);
 237                 if (tp_ops == NULL) {
 238 #ifdef CONFIG_MODULES
 239                         struct nlattr *kind = tca[TCA_KIND];
 240                         char name[IFNAMSIZ];
 241
 242                         if (kind != NULL &&
 243                             nla_strlcpy(name, kind, IFNAMSIZ) < IFNAMSIZ) {
 244                                 rtnl_unlock();
 245                                 request_module("cls_%s", name);
 246                                 rtnl_lock();
 247                                 tp_ops = tcf_proto_lookup_ops(kind);
 248                                 /* We dropped the RTNL semaphore in order to
 249                                  * perform the module load.  So, even if we
 250                                  * succeeded in loading the module we have to
 251                                  * replay the request.  We indicate this using
 252                                  * -EAGAIN.
 253                                  */
 254                                 if (tp_ops != NULL) {
 255                                         module_put(tp_ops->owner);
 256                                         err = -EAGAIN;
 257                                 }
 258                         }
 259 #endif
 260                         kfree(tp);
 261                         goto errout;
 262                 }
 263                 tp->ops = tp_ops;
 264                 tp->protocol = protocol;
 265                 tp->prio = nprio ? : TC_H_MAJ(tcf_auto_prio(*back));
 266                 tp->q = q;
 267                 tp->classify = tp_ops->classify;
 268                 tp->classid = parent;
 269
 270                 err = tp_ops->init(tp);
 271                 if (err != 0) {
 272                         module_put(tp_ops->owner);
 273                         kfree(tp);
 274                         goto errout;
 275                 }
 276
 277                 tp_created = 1;
 278
 279         } else if (tca[TCA_KIND] && nla_strcmp(tca[TCA_KIND], tp->ops->kind))
 280                 goto errout;
 281
 282         fh = tp->ops->get(tp, t->tcm_handle);
 283
 284         if (fh == 0) {
 285                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 286                         spin_lock_bh(root_lock);
 287                         *back = tp->next;
 288                         spin_unlock_bh(root_lock);
 289
 290                         tfilter_notify(net, skb, n, tp, fh, RTM_DELTFILTER);
 291                         tcf_destroy(tp);
 292                         err = 0;
 293                         goto errout;
 294                 }
 295
 296                 err = -ENOENT;
 297                 if (n->nlmsg_type != RTM_NEWTFILTER ||
 298                     !(n->nlmsg_flags & NLM_F_CREATE))
 299                         goto errout;
 300         } else {
 301                 switch (n->nlmsg_type) {
 302                 case RTM_NEWTFILTER:
 303                         err = -EEXIST;
 304                         if (n->nlmsg_flags & NLM_F_EXCL) {
 305                                 if (tp_created)
 306                                         tcf_destroy(tp);
 307                                 goto errout;
 308                         }
 309                         break;
 310                 case RTM_DELTFILTER:
 311                         err = tp->ops->delete(tp, fh);
 312                         if (err == 0)
 313                                 tfilter_notify(net, skb, n, tp, fh, RTM_DELTFILTER);
 314                         goto errout;
 315                 case RTM_GETTFILTER:
 316                         err = tfilter_notify(net, skb, n, tp, fh, RTM_NEWTFILTER);
 317                         goto errout;
 318                 default:
 319                         err = -EINVAL;
 320                         goto errout;
 321                 }
 322         }
 323
 324         err = tp->ops->change(net, skb, tp, cl, t->tcm_handle, tca, &fh);
 325         if (err == 0) {
 326                 if (tp_created) {
 327                         spin_lock_bh(root_lock);
 328                         tp->next = *back;
 329                         *back = tp;
 330                         spin_unlock_bh(root_lock);
 331                 }
 332                 tfilter_notify(net, skb, n, tp, fh, RTM_NEWTFILTER);
 333         } else {
 334                 if (tp_created)
 335                         tcf_destroy(tp);
 336         }
 337
 338 errout:
 339         if (cl)
 340                 cops->put(q, cl);
 341         if (err == -EAGAIN)
 342                 /* Replay the request. */
 343                 goto replay;
 344         return err;
 345 }
 346
 347 static int tcf_fill_node(struct sk_buff *skb, struct tcf_proto *tp,
 348                          unsigned long fh, u32 portid, u32 seq, u16 flags, int event)
 349 {
 350         struct tcmsg *tcm;
 351         struct nlmsghdr  *nlh;
 352         unsigned char *b = skb_tail_pointer(skb);
 353
 354         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*tcm), flags);
 355         if (!nlh)
 356                 goto out_nlmsg_trim;
 357         tcm = nlmsg_data(nlh);
 358         tcm->tcm_family = AF_UNSPEC;
 359         tcm->tcm__pad1 = 0;
 360         tcm->tcm__pad2 = 0;
 361         tcm->tcm_ifindex = qdisc_dev(tp->q)->ifindex;
 362         tcm->tcm_parent = tp->classid;
 363         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 364         if (nla_put_string(skb, TCA_KIND, tp->ops->kind))
 365                 goto nla_put_failure;
 366         tcm->tcm_handle = fh;
 367         if (RTM_DELTFILTER != event) {
 368                 tcm->tcm_handle = 0;
 369                 if (tp->ops->dump && tp->ops->dump(tp, fh, skb, tcm) < 0)
 370                         goto nla_put_failure;
 371         }
 372         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 373         return skb->len;
 374
 375 out_nlmsg_trim:
 376 nla_put_failure:
 377         nlmsg_trim(skb, b);
 378         return -1;
 379 }
 380
 381 static int tfilter_notify(struct net *net, struct sk_buff *oskb,
 382                           struct nlmsghdr *n, struct tcf_proto *tp,
 383                           unsigned long fh, int event)
 384 {
 385         struct sk_buff *skb;
 386         u32 portid = oskb ? NETLINK_CB(oskb).portid : 0;
 387
 388         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 389         if (!skb)
 390                 return -ENOBUFS;
 391
 392         if (tcf_fill_node(skb, tp, fh, portid, n->nlmsg_seq, 0, event) <= 0) {
 393                 kfree_skb(skb);
 394                 return -EINVAL;
 395         }
 396
 397         return rtnetlink_send(skb, net, portid, RTNLGRP_TC,
 398                               n->nlmsg_flags & NLM_F_ECHO);
 399 }
 400
 401 struct tcf_dump_args {
 402         struct tcf_walker w;
 403         struct sk_buff *skb;
 404         struct netlink_callback *cb;
 405 };
 406
 407 static int tcf_node_dump(struct tcf_proto *tp, unsigned long n,
 408                          struct tcf_walker *arg)
 409 {
 410         struct tcf_dump_args *a = (void *)arg;
 411
 412         return tcf_fill_node(a->skb, tp, n, NETLINK_CB(a->cb->skb).portid,
 413                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER);
 414 }
 415
 416 /* called with RTNL */
 417 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
 418 {
 419         struct net *net = sock_net(skb->sk);
 420         int t;
 421         int s_t;
 422         struct net_device *dev;
 423         struct Qdisc *q;
 424         struct tcf_proto *tp, **chain;
 425         struct tcmsg *tcm = nlmsg_data(cb->nlh);
 426         unsigned long cl = 0;
 427         const struct Qdisc_class_ops *cops;
 428         struct tcf_dump_args arg;
 429
 430         if (nlmsg_len(cb->nlh) < sizeof(*tcm))
 431                 return skb->len;
 432         dev = __dev_get_by_index(net, tcm->tcm_ifindex);
 433         if (!dev)
 434                 return skb->len;
 435
 436         if (!tcm->tcm_parent)
 437                 q = dev->qdisc;
 438         else
 439                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
 440         if (!q)
 441                 goto out;
 442         cops = q->ops->cl_ops;
 443         if (!cops)
 444                 goto errout;
 445         if (cops->tcf_chain == NULL)
 446                 goto errout;
 447         if (TC_H_MIN(tcm->tcm_parent)) {
 448                 cl = cops->get(q, tcm->tcm_parent);
 449                 if (cl == 0)
 450                         goto errout;
 451         }
 452         chain = cops->tcf_chain(q, cl);
 453         if (chain == NULL)
 454                 goto errout;
 455
 456         s_t = cb->args[0];
 457
 458         for (tp = *chain, t = 0; tp; tp = tp->next, t++) {
 459                 if (t < s_t)
 460                         continue;
 461                 if (TC_H_MAJ(tcm->tcm_info) &&
 462                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 463                         continue;
 464                 if (TC_H_MIN(tcm->tcm_info) &&
 465                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 466                         continue;
 467                 if (t > s_t)
 468                         memset(&cb->args[1], 0, sizeof(cb->args)-sizeof(cb->args[0]));
 469                 if (cb->args[1] == 0) {
 470                         if (tcf_fill_node(skb, tp, 0, NETLINK_CB(cb->skb).portid,
 471                                           cb->nlh->nlmsg_seq, NLM_F_MULTI,
 472                                           RTM_NEWTFILTER) <= 0)
 473                                 break;
 474
 475                         cb->args[1] = 1;
 476                 }
 477                 if (tp->ops->walk == NULL)
 478                         continue;
 479                 arg.w.fn = tcf_node_dump;
 480                 arg.skb = skb;
 481                 arg.cb = cb;
 482                 arg.w.stop = 0;
 483                 arg.w.skip = cb->args[1] - 1;
 484                 arg.w.count = 0;
 485                 tp->ops->walk(tp, &arg.w);
 486                 cb->args[1] = arg.w.count + 1;
 487                 if (arg.w.stop)
 488                         break;
 489         }
 490
 491         cb->args[0] = t;
 492
 493 errout:
 494         if (cl)
 495                 cops->put(q, cl);
 496 out:
 497         return skb->len;
 498 }
 499
 500 void tcf_exts_destroy(struct tcf_proto *tp, struct tcf_exts *exts)
 501 {
 502 #ifdef CONFIG_NET_CLS_ACT
 503         if (exts->action) {
 504                 tcf_action_destroy(exts->action, TCA_ACT_UNBIND);
 505                 exts->action = NULL;
 506         }
 507 #endif
 508 }
 509 EXPORT_SYMBOL(tcf_exts_destroy);
 510
 511 int tcf_exts_validate(struct net *net, struct tcf_proto *tp, struct nlattr **tb,
 512                   struct nlattr *rate_tlv, struct tcf_exts *exts,
 513                   const struct tcf_ext_map *map)
 514 {
 515         memset(exts, 0, sizeof(*exts));
 516
 517 #ifdef CONFIG_NET_CLS_ACT
 518         {
 519                 struct tc_action *act;
 520
 521                 if (map->police && tb[map->police]) {
 522                         act = tcf_action_init_1(net, tb[map->police], rate_tlv,
 523                                                 "police", TCA_ACT_NOREPLACE,
 524                                                 TCA_ACT_BIND);
 525                         if (IS_ERR(act))
 526                                 return PTR_ERR(act);
 527
 528                         act->type = TCA_OLD_COMPAT;
 529                         exts->action = act;
 530                 } else if (map->action && tb[map->action]) {
 531                         act = tcf_action_init(net, tb[map->action], rate_tlv,
 532                                               NULL, TCA_ACT_NOREPLACE,
 533                                               TCA_ACT_BIND);
 534                         if (IS_ERR(act))
 535                                 return PTR_ERR(act);
 536
 537                         exts->action = act;
 538                 }
 539         }
 540 #else
 541         if ((map->action && tb[map->action]) ||
 542             (map->police && tb[map->police]))
 543                 return -EOPNOTSUPP;
 544 #endif
 545
 546         return 0;
 547 }
 548 EXPORT_SYMBOL(tcf_exts_validate);
 549
 550 void tcf_exts_change(struct tcf_proto *tp, struct tcf_exts *dst,
 551                      struct tcf_exts *src)
 552 {
 553 #ifdef CONFIG_NET_CLS_ACT
 554         if (src->action) {
 555                 struct tc_action *act;
 556                 tcf_tree_lock(tp);
 557                 act = dst->action;
 558                 dst->action = src->action;
 559                 tcf_tree_unlock(tp);
 560                 if (act)
 561                         tcf_action_destroy(act, TCA_ACT_UNBIND);
 562         }
 563 #endif
 564 }
 565 EXPORT_SYMBOL(tcf_exts_change);
 566
 567 int tcf_exts_dump(struct sk_buff *skb, struct tcf_exts *exts,
 568                   const struct tcf_ext_map *map)
 569 {
 570 #ifdef CONFIG_NET_CLS_ACT
 571         if (map->action && exts->action) {
 572                 /*
 573                  * again for backward compatible mode - we want
 574                  * to work with both old and new modes of entering
 575                  * tc data even if iproute2  was newer - jhs
 576                  */
 577                 struct nlattr *nest;
 578
 579                 if (exts->action->type != TCA_OLD_COMPAT) {
 580                         nest = nla_nest_start(skb, map->action);
 581                         if (nest == NULL)
 582                                 goto nla_put_failure;
 583                         if (tcf_action_dump(skb, exts->action, 0, 0) < 0)
 584                                 goto nla_put_failure;
 585                         nla_nest_end(skb, nest);
 586                 } else if (map->police) {
 587                         nest = nla_nest_start(skb, map->police);
 588                         if (nest == NULL)
 589                                 goto nla_put_failure;
 590                         if (tcf_action_dump_old(skb, exts->action, 0, 0) < 0)
 591                                 goto nla_put_failure;
 592                         nla_nest_end(skb, nest);
 593                 }
 594         }
 595 #endif
 596         return 0;
 597 nla_put_failure: __attribute__ ((unused))
 598         return -1;
 599 }
 600 EXPORT_SYMBOL(tcf_exts_dump);
 601
 602
 603 int tcf_exts_dump_stats(struct sk_buff *skb, struct tcf_exts *exts,
 604                         const struct tcf_ext_map *map)
 605 {
 606 #ifdef CONFIG_NET_CLS_ACT
 607         if (exts->action)
 608                 if (tcf_action_copy_stats(skb, exts->action, 1) < 0)
 609                         goto nla_put_failure;
 610 #endif
 611         return 0;
 612 nla_put_failure: __attribute__ ((unused))
 613         return -1;
 614 }
 615 EXPORT_SYMBOL(tcf_exts_dump_stats);
 616
 617 static int __init tc_filter_init(void)
 618 {
 619         rtnl_register(PF_UNSPEC, RTM_NEWTFILTER, tc_ctl_tfilter, NULL, NULL);
 620         rtnl_register(PF_UNSPEC, RTM_DELTFILTER, tc_ctl_tfilter, NULL, NULL);
 621         rtnl_register(PF_UNSPEC, RTM_GETTFILTER, tc_ctl_tfilter,
 622                       tc_dump_tfilter, NULL);
 623
 624         return 0;
 625 }
 626
 627 subsys_initcall(tc_filter_init);