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 #include <asm/uaccess.h>
  17 #include <asm/system.h>
  18 #include <asm/bitops.h>
  19 #include <linux/config.h>
  20 #include <linux/types.h>
  21 #include <linux/kernel.h>
  22 #include <linux/sched.h>
  23 #include <linux/string.h>
  24 #include <linux/mm.h>
  25 #include <linux/socket.h>
  26 #include <linux/sockios.h>
  27 #include <linux/in.h>
  28 #include <linux/errno.h>
  29 #include <linux/interrupt.h>
  30 #include <linux/netdevice.h>
  31 #include <linux/skbuff.h>
  32 #include <linux/rtnetlink.h>
  33 #include <linux/init.h>
  34 #include <linux/kmod.h>
  35 #include <net/sock.h>
  36 #include <net/pkt_sched.h>
  37
  38 /* The list of all installed classifier types */
  39
  40 static struct tcf_proto_ops *tcf_proto_base;
  41
  42 /* Protects list of registered TC modules. It is pure SMP lock. */
  43 static rwlock_t cls_mod_lock = RW_LOCK_UNLOCKED;
  44
  45 /* Find classifier type by string name */
  46
  47 struct tcf_proto_ops * tcf_proto_lookup_ops(struct rtattr *kind)
  48 {
  49         struct tcf_proto_ops *t = NULL;
  50
  51         if (kind) {
  52                 read_lock(&cls_mod_lock);
  53                 for (t = tcf_proto_base; t; t = t->next) {
  54                         if (rtattr_strcmp(kind, t->kind) == 0)
  55                                 break;
  56                 }
  57                 read_unlock(&cls_mod_lock);
  58         }
  59         return t;
  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, **tp;
  67
  68         write_lock(&cls_mod_lock);
  69         for (tp = &tcf_proto_base; (t=*tp) != NULL; tp = &t->next) {
  70                 if (strcmp(ops->kind, t->kind) == 0) {
  71                         write_unlock(&cls_mod_lock);
  72                         return -EEXIST;
  73                 }
  74         }
  75
  76         ops->next = NULL;
  77         *tp = ops;
  78         write_unlock(&cls_mod_lock);
  79         return 0;
  80 }
  81
  82 int unregister_tcf_proto_ops(struct tcf_proto_ops *ops)
  83 {
  84         struct tcf_proto_ops *t, **tp;
  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                 write_unlock(&cls_mod_lock);
  93                 return -ENOENT;
  94         }
  95         *tp = t->next;
  96         write_unlock(&cls_mod_lock);
  97         return 0;
  98 }
  99
 100 #ifdef CONFIG_RTNETLINK
 101
 102 static int tfilter_notify(struct sk_buff *oskb, struct nlmsghdr *n,
 103                           struct tcf_proto *tp, 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, void *arg)
 121 {
 122         struct rtattr **tca = arg;
 123         struct tcmsg *t = NLMSG_DATA(n);
 124         u32 protocol = TC_H_MIN(t->tcm_info);
 125         u32 prio = TC_H_MAJ(t->tcm_info);
 126         u32 nprio = prio;
 127         u32 parent = t->tcm_parent;
 128         struct device *dev;
 129         struct Qdisc  *q;
 130         struct tcf_proto **back, **chain;
 131         struct tcf_proto *tp = NULL;
 132         struct tcf_proto_ops *tp_ops;
 133         struct Qdisc_class_ops *cops;
 134         unsigned long cl = 0;
 135         unsigned long fh;
 136         int err;
 137
 138         if (prio == 0) {
 139                 /* If no priority is given, user wants we allocated it. */
 140                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 141                         return -ENOENT;
 142                 prio = TC_H_MAKE(0x80000000U,0U);
 143         }
 144
 145         /* Find head of filter chain. */
 146
 147         /* Find link */
 148         if ((dev = dev_get_by_index(t->tcm_ifindex)) == NULL)
 149                 return -ENODEV;
 150
 151         /* Find qdisc */
 152         if (!parent) {
 153                 q = dev->qdisc_sleeping;
 154                 parent = q->handle;
 155         } else if ((q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent))) == NULL)
 156                 return -EINVAL;
 157
 158         /* Is it classful? */
 159         if ((cops = q->ops->cl_ops) == NULL)
 160                 return -EINVAL;
 161
 162         /* Do we search for filter, attached to class? */
 163         if (TC_H_MIN(parent)) {
 164                 cl = cops->get(q, parent);
 165                 if (cl == 0)
 166                         return -ENOENT;
 167         }
 168
 169         /* And the last stroke */
 170         chain = cops->tcf_chain(q, cl);
 171         err = -EINVAL;
 172         if (chain == NULL)
 173                 goto errout;
 174
 175         /* Check the chain for existence of proto-tcf with this priority */
 176         for (back = chain; (tp=*back) != NULL; back = &tp->next) {
 177                 if (tp->prio >= prio) {
 178                         if (tp->prio == prio) {
 179                                 if (!nprio || (tp->protocol != protocol && protocol))
 180                                         goto errout;
 181                         } else
 182                                 tp = NULL;
 183                         break;
 184                 }
 185         }
 186
 187         if (tp == NULL) {
 188                 /* Proto-tcf does not exist, create new one */
 189
 190                 if (tca[TCA_KIND-1] == NULL || !protocol)
 191                         goto errout;
 192
 193                 err = -ENOENT;
 194                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 195                         goto errout;
 196
 197
 198                 /* Create new proto tcf */
 199
 200                 err = -ENOBUFS;
 201                 if ((tp = kmalloc(sizeof(*tp), GFP_KERNEL)) == NULL)
 202                         goto errout;
 203                 tp_ops = tcf_proto_lookup_ops(tca[TCA_KIND-1]);
 204 #ifdef CONFIG_KMOD
 205                 if (tp_ops==NULL && tca[TCA_KIND-1] != NULL) {
 206                         struct rtattr *kind = tca[TCA_KIND-1];
 207                         char module_name[4 + IFNAMSIZ + 1];
 208
 209                         if (RTA_PAYLOAD(kind) <= IFNAMSIZ) {
 210                                 sprintf(module_name, "cls_%s", (char*)RTA_DATA(kind));
 211                                 request_module (module_name);
 212                                 tp_ops = tcf_proto_lookup_ops(kind);
 213                         }
 214                 }
 215 #endif
 216                 if (tp_ops == NULL) {
 217                         err = -EINVAL;
 218                         kfree(tp);
 219                         goto errout;
 220                 }
 221                 memset(tp, 0, sizeof(*tp));
 222                 tp->ops = tp_ops;
 223                 tp->protocol = protocol;
 224                 tp->prio = nprio ? : tcf_auto_prio(*back);
 225                 tp->q = q;
 226                 tp->classify = tp_ops->classify;
 227                 tp->classid = parent;
 228                 err = tp_ops->init(tp);
 229                 if (err) {
 230                         kfree(tp);
 231                         goto errout;
 232                 }
 233                 write_lock(&qdisc_tree_lock);
 234                 spin_lock_bh(&dev->queue_lock);
 235                 tp->next = *back;
 236                 *back = tp;
 237                 spin_unlock_bh(&dev->queue_lock);
 238                 write_unlock(&qdisc_tree_lock);
 239         } else if (tca[TCA_KIND-1] && rtattr_strcmp(tca[TCA_KIND-1], tp->ops->kind))
 240                 goto errout;
 241
 242         fh = tp->ops->get(tp, t->tcm_handle);
 243
 244         if (fh == 0) {
 245                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 246                         write_lock(&qdisc_tree_lock);
 247                         spin_lock_bh(&dev->queue_lock);
 248                         *back = tp->next;
 249                         spin_unlock_bh(&dev->queue_lock);
 250                         write_unlock(&qdisc_tree_lock);
 251
 252                         tp->ops->destroy(tp);
 253                         kfree(tp);
 254                         err = 0;
 255                         goto errout;
 256                 }
 257
 258                 err = -ENOENT;
 259                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 260                         goto errout;
 261         } else {
 262                 switch (n->nlmsg_type) {
 263                 case RTM_NEWTFILTER:
 264                         err = -EEXIST;
 265                         if (n->nlmsg_flags&NLM_F_EXCL)
 266                                 goto errout;
 267                         break;
 268                 case RTM_DELTFILTER:
 269                         err = tp->ops->delete(tp, fh);
 270                         goto errout;
 271                 case RTM_GETTFILTER:
 272                         err = tfilter_notify(skb, n, tp, fh, RTM_NEWTFILTER);
 273                         goto errout;
 274                 default:
 275                         err = -EINVAL;
 276                         goto errout;
 277                 }
 278         }
 279
 280         err = tp->ops->change(tp, cl, t->tcm_handle, tca, &fh);
 281         if (err == 0)
 282                 tfilter_notify(skb, n, tp, fh, RTM_NEWTFILTER);
 283
 284 errout:
 285         if (cl)
 286                 cops->put(q, cl);
 287         return err;
 288 }
 289
 290 static int
 291 tcf_fill_node(struct sk_buff *skb, struct tcf_proto *tp, unsigned long fh,
 292               u32 pid, u32 seq, unsigned flags, int event)
 293 {
 294         struct tcmsg *tcm;
 295         struct nlmsghdr  *nlh;
 296         unsigned char    *b = skb->tail;
 297
 298         nlh = NLMSG_PUT(skb, pid, seq, event, sizeof(*tcm));
 299         nlh->nlmsg_flags = flags;
 300         tcm = NLMSG_DATA(nlh);
 301         tcm->tcm_family = AF_UNSPEC;
 302         tcm->tcm_ifindex = tp->q->dev->ifindex;
 303         tcm->tcm_parent = tp->classid;
 304         tcm->tcm_handle = 0;
 305         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 306         RTA_PUT(skb, TCA_KIND, IFNAMSIZ, tp->ops->kind);
 307         if (tp->ops->dump && tp->ops->dump(tp, fh, skb, tcm) < 0)
 308                 goto rtattr_failure;
 309         nlh->nlmsg_len = skb->tail - b;
 310         return skb->len;
 311
 312 nlmsg_failure:
 313 rtattr_failure:
 314         skb_trim(skb, b - skb->data);
 315         return -1;
 316 }
 317
 318 static int tfilter_notify(struct sk_buff *oskb, struct nlmsghdr *n,
 319                           struct tcf_proto *tp, unsigned long fh, int event)
 320 {
 321         struct sk_buff *skb;
 322         u32 pid = oskb ? NETLINK_CB(oskb).pid : 0;
 323
 324         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 325         if (!skb)
 326                 return -ENOBUFS;
 327
 328         if (tcf_fill_node(skb, tp, fh, pid, n->nlmsg_seq, 0, event) <= 0) {
 329                 kfree_skb(skb);
 330                 return -EINVAL;
 331         }
 332
 333         return rtnetlink_send(skb, pid, RTMGRP_TC, n->nlmsg_flags&NLM_F_ECHO);
 334 }
 335
 336 struct tcf_dump_args
 337 {
 338         struct tcf_walker w;
 339         struct sk_buff *skb;
 340         struct netlink_callback *cb;
 341 };
 342
 343 static int tcf_node_dump(struct tcf_proto *tp, unsigned long n, struct tcf_walker *arg)
 344 {
 345         struct tcf_dump_args *a = (void*)arg;
 346
 347         return tcf_fill_node(a->skb, tp, n, NETLINK_CB(a->cb->skb).pid,
 348                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER);
 349 }
 350
 351 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
 352 {
 353         int t;
 354         int s_t;
 355         struct device *dev;
 356         struct Qdisc *q;
 357         struct tcf_proto *tp, **chain;
 358         struct tcmsg *tcm = (struct tcmsg*)NLMSG_DATA(cb->nlh);
 359         unsigned long cl = 0;
 360         struct Qdisc_class_ops *cops;
 361         struct tcf_dump_args arg;
 362
 363         if (cb->nlh->nlmsg_len < NLMSG_LENGTH(sizeof(*tcm)))
 364                 return skb->len;
 365         if ((dev = dev_get_by_index(tcm->tcm_ifindex)) == NULL)
 366                 return skb->len;
 367
 368         read_lock(&qdisc_tree_lock);
 369         if (!tcm->tcm_parent)
 370                 q = dev->qdisc_sleeping;
 371         else
 372                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
 373         if (q == NULL) {
 374                 read_unlock(&qdisc_tree_lock);
 375                 return skb->len;
 376         }
 377         if ((cops = q->ops->cl_ops) == NULL)
 378                 goto errout;
 379         if (TC_H_MIN(tcm->tcm_parent)) {
 380                 cl = cops->get(q, tcm->tcm_parent);
 381                 if (cl == 0)
 382                         goto errout;
 383         }
 384         chain = cops->tcf_chain(q, cl);
 385         if (chain == NULL)
 386                 goto errout;
 387
 388         s_t = cb->args[0];
 389
 390         for (tp=*chain, t=0; tp; tp = tp->next, t++) {
 391                 if (t < s_t) continue;
 392                 if (TC_H_MAJ(tcm->tcm_info) &&
 393                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 394                         continue;
 395                 if (TC_H_MIN(tcm->tcm_info) &&
 396                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 397                         continue;
 398                 if (t > s_t)
 399                         memset(&cb->args[1], 0, sizeof(cb->args)-sizeof(cb->args[0]));
 400                 if (cb->args[1] == 0) {
 401                         if (tcf_fill_node(skb, tp, 0, NETLINK_CB(cb->skb).pid,
 402                                           cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER) <= 0) {
 403                                 break;
 404                         }
 405                         cb->args[1] = 1;
 406                 }
 407                 if (tp->ops->walk == NULL)
 408                         continue;
 409                 arg.w.fn = tcf_node_dump;
 410                 arg.skb = skb;
 411                 arg.cb = cb;
 412                 arg.w.stop = 0;
 413                 arg.w.skip = cb->args[1]-1;
 414                 arg.w.count = 0;
 415                 tp->ops->walk(tp, &arg.w);
 416                 cb->args[1] = arg.w.count+1;
 417                 if (arg.w.stop)
 418                         break;
 419         }
 420
 421         cb->args[0] = t;
 422
 423 errout:
 424         if (cl)
 425                 cops->put(q, cl);
 426
 427         read_unlock(&qdisc_tree_lock);
 428         return skb->len;
 429 }
 430
 431 #endif
 432
 433
 434 __initfunc(int tc_filter_init(void))
 435 {
 436 #ifdef CONFIG_RTNETLINK
 437         struct rtnetlink_link *link_p = rtnetlink_links[PF_UNSPEC];
 438
 439         /* Setup rtnetlink links. It is made here to avoid
 440            exporting large number of public symbols.
 441          */
 442
 443         if (link_p) {
 444                 link_p[RTM_NEWTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 445                 link_p[RTM_DELTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 446                 link_p[RTM_GETTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 447                 link_p[RTM_GETTFILTER-RTM_BASE].dumpit = tc_dump_tfilter;
 448         }
 449 #endif
 450 #define INIT_TC_FILTER(name) { \
 451           extern struct tcf_proto_ops cls_##name##_ops; \
 452           register_tcf_proto_ops(&cls_##name##_ops); \
 453         }
 454
 455 #ifdef CONFIG_NET_CLS_U32
 456         INIT_TC_FILTER(u32);
 457 #endif
 458 #ifdef CONFIG_NET_CLS_ROUTE4
 459         INIT_TC_FILTER(route4);
 460 #endif
 461 #ifdef CONFIG_NET_CLS_FW
 462         INIT_TC_FILTER(fw);
 463 #endif
 464 #ifdef CONFIG_NET_CLS_RSVP
 465         INIT_TC_FILTER(rsvp);
 466 #endif
 467 #ifdef CONFIG_NET_CLS_RSVP6
 468         INIT_TC_FILTER(rsvp6);
 469 #endif
 470         return 0;
 471 }