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 <asm/uaccess.h>
  18 #include <asm/system.h>
  19 #include <linux/bitops.h>
  20 #include <linux/module.h>
  21 #include <linux/types.h>
  22 #include <linux/kernel.h>
  23 #include <linux/sched.h>
  24 #include <linux/string.h>
  25 #include <linux/mm.h>
  26 #include <linux/socket.h>
  27 #include <linux/sockios.h>
  28 #include <linux/in.h>
  29 #include <linux/errno.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/netdevice.h>
  32 #include <linux/skbuff.h>
  33 #include <linux/rtnetlink.h>
  34 #include <linux/init.h>
  35 #include <linux/kmod.h>
  36 #include <net/sock.h>
  37 #include <net/pkt_sched.h>
  38 #include <net/pkt_cls.h>
  39
  40 #if 0 /* control */
  41 #define DPRINTK(format,args...) printk(KERN_DEBUG format,##args)
  42 #else
  43 #define DPRINTK(format,args...)
  44 #endif
  45
  46 /* The list of all installed classifier types */
  47
  48 static struct tcf_proto_ops *tcf_proto_base;
  49
  50 /* Protects list of registered TC modules. It is pure SMP lock. */
  51 static DEFINE_RWLOCK(cls_mod_lock);
  52
  53 /* Find classifier type by string name */
  54
  55 static struct tcf_proto_ops * tcf_proto_lookup_ops(struct rtattr *kind)
  56 {
  57         struct tcf_proto_ops *t = NULL;
  58
  59         if (kind) {
  60                 read_lock(&cls_mod_lock);
  61                 for (t = tcf_proto_base; t; t = t->next) {
  62                         if (rtattr_strcmp(kind, t->kind) == 0) {
  63                                 if (!try_module_get(t->owner))
  64                                         t = NULL;
  65                                 break;
  66                         }
  67                 }
  68                 read_unlock(&cls_mod_lock);
  69         }
  70         return t;
  71 }
  72
  73 /* Register(unregister) new classifier type */
  74
  75 int register_tcf_proto_ops(struct tcf_proto_ops *ops)
  76 {
  77         struct tcf_proto_ops *t, **tp;
  78         int rc = -EEXIST;
  79
  80         write_lock(&cls_mod_lock);
  81         for (tp = &tcf_proto_base; (t = *tp) != NULL; tp = &t->next)
  82                 if (!strcmp(ops->kind, t->kind))
  83                         goto out;
  84
  85         ops->next = NULL;
  86         *tp = ops;
  87         rc = 0;
  88 out:
  89         write_unlock(&cls_mod_lock);
  90         return rc;
  91 }
  92
  93 int unregister_tcf_proto_ops(struct tcf_proto_ops *ops)
  94 {
  95         struct tcf_proto_ops *t, **tp;
  96         int rc = -ENOENT;
  97
  98         write_lock(&cls_mod_lock);
  99         for (tp = &tcf_proto_base; (t=*tp) != NULL; tp = &t->next)
 100                 if (t == ops)
 101                         break;
 102
 103         if (!t)
 104                 goto out;
 105         *tp = t->next;
 106         rc = 0;
 107 out:
 108         write_unlock(&cls_mod_lock);
 109         return rc;
 110 }
 111
 112 static int tfilter_notify(struct sk_buff *oskb, struct nlmsghdr *n,
 113                           struct tcf_proto *tp, unsigned long fh, int event);
 114
 115
 116 /* Select new prio value from the range, managed by kernel. */
 117
 118 static __inline__ u32 tcf_auto_prio(struct tcf_proto *tp)
 119 {
 120         u32 first = TC_H_MAKE(0xC0000000U,0U);
 121
 122         if (tp)
 123                 first = tp->prio-1;
 124
 125         return first;
 126 }
 127
 128 /* Add/change/delete/get a filter node */
 129
 130 static int tc_ctl_tfilter(struct sk_buff *skb, struct nlmsghdr *n, void *arg)
 131 {
 132         struct rtattr **tca;
 133         struct tcmsg *t;
 134         u32 protocol;
 135         u32 prio;
 136         u32 nprio;
 137         u32 parent;
 138         struct net_device *dev;
 139         struct Qdisc  *q;
 140         struct tcf_proto **back, **chain;
 141         struct tcf_proto *tp;
 142         struct tcf_proto_ops *tp_ops;
 143         struct Qdisc_class_ops *cops;
 144         unsigned long cl;
 145         unsigned long fh;
 146         int err;
 147
 148 replay:
 149         tca = arg;
 150         t = NLMSG_DATA(n);
 151         protocol = TC_H_MIN(t->tcm_info);
 152         prio = TC_H_MAJ(t->tcm_info);
 153         nprio = prio;
 154         parent = t->tcm_parent;
 155         cl = 0;
 156
 157         if (prio == 0) {
 158                 /* If no priority is given, user wants we allocated it. */
 159                 if (n->nlmsg_type != RTM_NEWTFILTER || !(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         if ((dev = __dev_get_by_index(t->tcm_ifindex)) == NULL)
 168                 return -ENODEV;
 169
 170         /* Find qdisc */
 171         if (!parent) {
 172                 q = dev->qdisc_sleeping;
 173                 parent = q->handle;
 174         } else if ((q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent))) == NULL)
 175                 return -EINVAL;
 176
 177         /* Is it classful? */
 178         if ((cops = q->ops->cl_ops) == NULL)
 179                 return -EINVAL;
 180
 181         /* Do we search for filter, attached to class? */
 182         if (TC_H_MIN(parent)) {
 183                 cl = cops->get(q, parent);
 184                 if (cl == 0)
 185                         return -ENOENT;
 186         }
 187
 188         /* And the last stroke */
 189         chain = cops->tcf_chain(q, cl);
 190         err = -EINVAL;
 191         if (chain == NULL)
 192                 goto errout;
 193
 194         /* Check the chain for existence of proto-tcf with this priority */
 195         for (back = chain; (tp=*back) != NULL; back = &tp->next) {
 196                 if (tp->prio >= prio) {
 197                         if (tp->prio == prio) {
 198                                 if (!nprio || (tp->protocol != protocol && protocol))
 199                                         goto errout;
 200                         } else
 201                                 tp = NULL;
 202                         break;
 203                 }
 204         }
 205
 206         if (tp == NULL) {
 207                 /* Proto-tcf does not exist, create new one */
 208
 209                 if (tca[TCA_KIND-1] == NULL || !protocol)
 210                         goto errout;
 211
 212                 err = -ENOENT;
 213                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 214                         goto errout;
 215
 216
 217                 /* Create new proto tcf */
 218
 219                 err = -ENOBUFS;
 220                 if ((tp = kzalloc(sizeof(*tp), GFP_KERNEL)) == NULL)
 221                         goto errout;
 222                 err = -EINVAL;
 223                 tp_ops = tcf_proto_lookup_ops(tca[TCA_KIND-1]);
 224                 if (tp_ops == NULL) {
 225 #ifdef CONFIG_KMOD
 226                         struct rtattr *kind = tca[TCA_KIND-1];
 227                         char name[IFNAMSIZ];
 228
 229                         if (kind != NULL &&
 230                             rtattr_strlcpy(name, kind, IFNAMSIZ) < IFNAMSIZ) {
 231                                 rtnl_unlock();
 232                                 request_module("cls_%s", name);
 233                                 rtnl_lock();
 234                                 tp_ops = tcf_proto_lookup_ops(kind);
 235                                 /* We dropped the RTNL semaphore in order to
 236                                  * perform the module load.  So, even if we
 237                                  * succeeded in loading the module we have to
 238                                  * replay the request.  We indicate this using
 239                                  * -EAGAIN.
 240                                  */
 241                                 if (tp_ops != NULL) {
 242                                         module_put(tp_ops->owner);
 243                                         err = -EAGAIN;
 244                                 }
 245                         }
 246 #endif
 247                         kfree(tp);
 248                         goto errout;
 249                 }
 250                 tp->ops = tp_ops;
 251                 tp->protocol = protocol;
 252                 tp->prio = nprio ? : tcf_auto_prio(*back);
 253                 tp->q = q;
 254                 tp->classify = tp_ops->classify;
 255                 tp->classid = parent;
 256                 if ((err = tp_ops->init(tp)) != 0) {
 257                         module_put(tp_ops->owner);
 258                         kfree(tp);
 259                         goto errout;
 260                 }
 261
 262                 qdisc_lock_tree(dev);
 263                 tp->next = *back;
 264                 *back = tp;
 265                 qdisc_unlock_tree(dev);
 266
 267         } else if (tca[TCA_KIND-1] && rtattr_strcmp(tca[TCA_KIND-1], tp->ops->kind))
 268                 goto errout;
 269
 270         fh = tp->ops->get(tp, t->tcm_handle);
 271
 272         if (fh == 0) {
 273                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 274                         qdisc_lock_tree(dev);
 275                         *back = tp->next;
 276                         qdisc_unlock_tree(dev);
 277
 278                         tfilter_notify(skb, n, tp, fh, RTM_DELTFILTER);
 279                         tcf_destroy(tp);
 280                         err = 0;
 281                         goto errout;
 282                 }
 283
 284                 err = -ENOENT;
 285                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 286                         goto errout;
 287         } else {
 288                 switch (n->nlmsg_type) {
 289                 case RTM_NEWTFILTER:
 290                         err = -EEXIST;
 291                         if (n->nlmsg_flags&NLM_F_EXCL)
 292                                 goto errout;
 293                         break;
 294                 case RTM_DELTFILTER:
 295                         err = tp->ops->delete(tp, fh);
 296                         if (err == 0)
 297                                 tfilter_notify(skb, n, tp, fh, RTM_DELTFILTER);
 298                         goto errout;
 299                 case RTM_GETTFILTER:
 300                         err = tfilter_notify(skb, n, tp, fh, RTM_NEWTFILTER);
 301                         goto errout;
 302                 default:
 303                         err = -EINVAL;
 304                         goto errout;
 305                 }
 306         }
 307
 308         err = tp->ops->change(tp, cl, t->tcm_handle, tca, &fh);
 309         if (err == 0)
 310                 tfilter_notify(skb, n, tp, fh, RTM_NEWTFILTER);
 311
 312 errout:
 313         if (cl)
 314                 cops->put(q, cl);
 315         if (err == -EAGAIN)
 316                 /* Replay the request. */
 317                 goto replay;
 318         return err;
 319 }
 320
 321 static int
 322 tcf_fill_node(struct sk_buff *skb, struct tcf_proto *tp, unsigned long fh,
 323               u32 pid, u32 seq, u16 flags, int event)
 324 {
 325         struct tcmsg *tcm;
 326         struct nlmsghdr  *nlh;
 327         unsigned char    *b = skb->tail;
 328
 329         nlh = NLMSG_NEW(skb, pid, seq, event, sizeof(*tcm), flags);
 330         tcm = NLMSG_DATA(nlh);
 331         tcm->tcm_family = AF_UNSPEC;
 332         tcm->tcm__pad1 = 0;
 333         tcm->tcm__pad1 = 0;
 334         tcm->tcm_ifindex = tp->q->dev->ifindex;
 335         tcm->tcm_parent = tp->classid;
 336         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 337         RTA_PUT(skb, TCA_KIND, IFNAMSIZ, tp->ops->kind);
 338         tcm->tcm_handle = fh;
 339         if (RTM_DELTFILTER != event) {
 340                 tcm->tcm_handle = 0;
 341                 if (tp->ops->dump && tp->ops->dump(tp, fh, skb, tcm) < 0)
 342                         goto rtattr_failure;
 343         }
 344         nlh->nlmsg_len = skb->tail - b;
 345         return skb->len;
 346
 347 nlmsg_failure:
 348 rtattr_failure:
 349         skb_trim(skb, b - skb->data);
 350         return -1;
 351 }
 352
 353 static int tfilter_notify(struct sk_buff *oskb, struct nlmsghdr *n,
 354                           struct tcf_proto *tp, unsigned long fh, int event)
 355 {
 356         struct sk_buff *skb;
 357         u32 pid = oskb ? NETLINK_CB(oskb).pid : 0;
 358
 359         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 360         if (!skb)
 361                 return -ENOBUFS;
 362
 363         if (tcf_fill_node(skb, tp, fh, pid, n->nlmsg_seq, 0, event) <= 0) {
 364                 kfree_skb(skb);
 365                 return -EINVAL;
 366         }
 367
 368         return rtnetlink_send(skb, pid, RTNLGRP_TC, n->nlmsg_flags&NLM_F_ECHO);
 369 }
 370
 371 struct tcf_dump_args
 372 {
 373         struct tcf_walker w;
 374         struct sk_buff *skb;
 375         struct netlink_callback *cb;
 376 };
 377
 378 static int tcf_node_dump(struct tcf_proto *tp, unsigned long n, struct tcf_walker *arg)
 379 {
 380         struct tcf_dump_args *a = (void*)arg;
 381
 382         return tcf_fill_node(a->skb, tp, n, NETLINK_CB(a->cb->skb).pid,
 383                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER);
 384 }
 385
 386 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
 387 {
 388         int t;
 389         int s_t;
 390         struct net_device *dev;
 391         struct Qdisc *q;
 392         struct tcf_proto *tp, **chain;
 393         struct tcmsg *tcm = (struct tcmsg*)NLMSG_DATA(cb->nlh);
 394         unsigned long cl = 0;
 395         struct Qdisc_class_ops *cops;
 396         struct tcf_dump_args arg;
 397
 398         if (cb->nlh->nlmsg_len < NLMSG_LENGTH(sizeof(*tcm)))
 399                 return skb->len;
 400         if ((dev = dev_get_by_index(tcm->tcm_ifindex)) == NULL)
 401                 return skb->len;
 402
 403         read_lock(&qdisc_tree_lock);
 404         if (!tcm->tcm_parent)
 405                 q = dev->qdisc_sleeping;
 406         else
 407                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
 408         if (!q)
 409                 goto out;
 410         if ((cops = q->ops->cl_ops) == NULL)
 411                 goto errout;
 412         if (TC_H_MIN(tcm->tcm_parent)) {
 413                 cl = cops->get(q, tcm->tcm_parent);
 414                 if (cl == 0)
 415                         goto errout;
 416         }
 417         chain = cops->tcf_chain(q, cl);
 418         if (chain == NULL)
 419                 goto errout;
 420
 421         s_t = cb->args[0];
 422
 423         for (tp=*chain, t=0; tp; tp = tp->next, t++) {
 424                 if (t < s_t) continue;
 425                 if (TC_H_MAJ(tcm->tcm_info) &&
 426                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 427                         continue;
 428                 if (TC_H_MIN(tcm->tcm_info) &&
 429                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 430                         continue;
 431                 if (t > s_t)
 432                         memset(&cb->args[1], 0, sizeof(cb->args)-sizeof(cb->args[0]));
 433                 if (cb->args[1] == 0) {
 434                         if (tcf_fill_node(skb, tp, 0, NETLINK_CB(cb->skb).pid,
 435                                           cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER) <= 0) {
 436                                 break;
 437                         }
 438                         cb->args[1] = 1;
 439                 }
 440                 if (tp->ops->walk == NULL)
 441                         continue;
 442                 arg.w.fn = tcf_node_dump;
 443                 arg.skb = skb;
 444                 arg.cb = cb;
 445                 arg.w.stop = 0;
 446                 arg.w.skip = cb->args[1]-1;
 447                 arg.w.count = 0;
 448                 tp->ops->walk(tp, &arg.w);
 449                 cb->args[1] = arg.w.count+1;
 450                 if (arg.w.stop)
 451                         break;
 452         }
 453
 454         cb->args[0] = t;
 455
 456 errout:
 457         if (cl)
 458                 cops->put(q, cl);
 459 out:
 460         read_unlock(&qdisc_tree_lock);
 461         dev_put(dev);
 462         return skb->len;
 463 }
 464
 465 void
 466 tcf_exts_destroy(struct tcf_proto *tp, struct tcf_exts *exts)
 467 {
 468 #ifdef CONFIG_NET_CLS_ACT
 469         if (exts->action) {
 470                 tcf_action_destroy(exts->action, TCA_ACT_UNBIND);
 471                 exts->action = NULL;
 472         }
 473 #elif defined CONFIG_NET_CLS_POLICE
 474         if (exts->police) {
 475                 tcf_police_release(exts->police, TCA_ACT_UNBIND);
 476                 exts->police = NULL;
 477         }
 478 #endif
 479 }
 480
 481
 482 int
 483 tcf_exts_validate(struct tcf_proto *tp, struct rtattr **tb,
 484                   struct rtattr *rate_tlv, struct tcf_exts *exts,
 485                   struct tcf_ext_map *map)
 486 {
 487         memset(exts, 0, sizeof(*exts));
 488
 489 #ifdef CONFIG_NET_CLS_ACT
 490         {
 491                 int err;
 492                 struct tc_action *act;
 493
 494                 if (map->police && tb[map->police-1]) {
 495                         act = tcf_action_init_1(tb[map->police-1], rate_tlv, "police",
 496                                 TCA_ACT_NOREPLACE, TCA_ACT_BIND, &err);
 497                         if (act == NULL)
 498                                 return err;
 499
 500                         act->type = TCA_OLD_COMPAT;
 501                         exts->action = act;
 502                 } else if (map->action && tb[map->action-1]) {
 503                         act = tcf_action_init(tb[map->action-1], rate_tlv, NULL,
 504                                 TCA_ACT_NOREPLACE, TCA_ACT_BIND, &err);
 505                         if (act == NULL)
 506                                 return err;
 507
 508                         exts->action = act;
 509                 }
 510         }
 511 #elif defined CONFIG_NET_CLS_POLICE
 512         if (map->police && tb[map->police-1]) {
 513                 struct tcf_police *p;
 514
 515                 p = tcf_police_locate(tb[map->police-1], rate_tlv);
 516                 if (p == NULL)
 517                         return -EINVAL;
 518
 519                 exts->police = p;
 520         } else if (map->action && tb[map->action-1])
 521                 return -EOPNOTSUPP;
 522 #else
 523         if ((map->action && tb[map->action-1]) ||
 524             (map->police && tb[map->police-1]))
 525                 return -EOPNOTSUPP;
 526 #endif
 527
 528         return 0;
 529 }
 530
 531 void
 532 tcf_exts_change(struct tcf_proto *tp, struct tcf_exts *dst,
 533                 struct tcf_exts *src)
 534 {
 535 #ifdef CONFIG_NET_CLS_ACT
 536         if (src->action) {
 537                 struct tc_action *act;
 538                 tcf_tree_lock(tp);
 539                 act = xchg(&dst->action, src->action);
 540                 tcf_tree_unlock(tp);
 541                 if (act)
 542                         tcf_action_destroy(act, TCA_ACT_UNBIND);
 543         }
 544 #elif defined CONFIG_NET_CLS_POLICE
 545         if (src->police) {
 546                 struct tcf_police *p;
 547                 tcf_tree_lock(tp);
 548                 p = xchg(&dst->police, src->police);
 549                 tcf_tree_unlock(tp);
 550                 if (p)
 551                         tcf_police_release(p, TCA_ACT_UNBIND);
 552         }
 553 #endif
 554 }
 555
 556 int
 557 tcf_exts_dump(struct sk_buff *skb, struct tcf_exts *exts,
 558               struct tcf_ext_map *map)
 559 {
 560 #ifdef CONFIG_NET_CLS_ACT
 561         if (map->action && exts->action) {
 562                 /*
 563                  * again for backward compatible mode - we want
 564                  * to work with both old and new modes of entering
 565                  * tc data even if iproute2  was newer - jhs
 566                  */
 567                 struct rtattr * p_rta = (struct rtattr*) skb->tail;
 568
 569                 if (exts->action->type != TCA_OLD_COMPAT) {
 570                         RTA_PUT(skb, map->action, 0, NULL);
 571                         if (tcf_action_dump(skb, exts->action, 0, 0) < 0)
 572                                 goto rtattr_failure;
 573                         p_rta->rta_len = skb->tail - (u8*)p_rta;
 574                 } else if (map->police) {
 575                         RTA_PUT(skb, map->police, 0, NULL);
 576                         if (tcf_action_dump_old(skb, exts->action, 0, 0) < 0)
 577                                 goto rtattr_failure;
 578                         p_rta->rta_len = skb->tail - (u8*)p_rta;
 579                 }
 580         }
 581 #elif defined CONFIG_NET_CLS_POLICE
 582         if (map->police && exts->police) {
 583                 struct rtattr * p_rta = (struct rtattr*) skb->tail;
 584
 585                 RTA_PUT(skb, map->police, 0, NULL);
 586
 587                 if (tcf_police_dump(skb, exts->police) < 0)
 588                         goto rtattr_failure;
 589
 590                 p_rta->rta_len = skb->tail - (u8*)p_rta;
 591         }
 592 #endif
 593         return 0;
 594 rtattr_failure: __attribute__ ((unused))
 595         return -1;
 596 }
 597
 598 int
 599 tcf_exts_dump_stats(struct sk_buff *skb, struct tcf_exts *exts,
 600                     struct tcf_ext_map *map)
 601 {
 602 #ifdef CONFIG_NET_CLS_ACT
 603         if (exts->action)
 604                 if (tcf_action_copy_stats(skb, exts->action, 1) < 0)
 605                         goto rtattr_failure;
 606 #elif defined CONFIG_NET_CLS_POLICE
 607         if (exts->police)
 608                 if (tcf_police_dump_stats(skb, exts->police) < 0)
 609                         goto rtattr_failure;
 610 #endif
 611         return 0;
 612 rtattr_failure: __attribute__ ((unused))
 613         return -1;
 614 }
 615
 616 static int __init tc_filter_init(void)
 617 {
 618         struct rtnetlink_link *link_p = rtnetlink_links[PF_UNSPEC];
 619
 620         /* Setup rtnetlink links. It is made here to avoid
 621            exporting large number of public symbols.
 622          */
 623
 624         if (link_p) {
 625                 link_p[RTM_NEWTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 626                 link_p[RTM_DELTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 627                 link_p[RTM_GETTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 628                 link_p[RTM_GETTFILTER-RTM_BASE].dumpit = tc_dump_tfilter;
 629         }
 630         return 0;
 631 }
 632
 633 subsys_initcall(tc_filter_init);
 634
 635 EXPORT_SYMBOL(register_tcf_proto_ops);
 636 EXPORT_SYMBOL(unregister_tcf_proto_ops);
 637 EXPORT_SYMBOL(tcf_exts_validate);
 638 EXPORT_SYMBOL(tcf_exts_destroy);
 639 EXPORT_SYMBOL(tcf_exts_change);
 640 EXPORT_SYMBOL(tcf_exts_dump);
 641 EXPORT_SYMBOL(tcf_exts_dump_stats);