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 static int tfilter_notify(struct sk_buff *oskb, struct nlmsghdr *n,
 101                           struct tcf_proto *tp, 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, void *arg)
 119 {
 120         struct rtattr **tca = arg;
 121         struct tcmsg *t = NLMSG_DATA(n);
 122         u32 protocol = TC_H_MIN(t->tcm_info);
 123         u32 prio = TC_H_MAJ(t->tcm_info);
 124         u32 nprio = prio;
 125         u32 parent = t->tcm_parent;
 126         struct net_device *dev;
 127         struct Qdisc  *q;
 128         struct tcf_proto **back, **chain;
 129         struct tcf_proto *tp = NULL;
 130         struct tcf_proto_ops *tp_ops;
 131         struct Qdisc_class_ops *cops;
 132         unsigned long cl = 0;
 133         unsigned long fh;
 134         int err;
 135
 136         if (prio == 0) {
 137                 /* If no priority is given, user wants we allocated it. */
 138                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 139                         return -ENOENT;
 140                 prio = TC_H_MAKE(0x80000000U,0U);
 141         }
 142
 143         /* Find head of filter chain. */
 144
 145         /* Find link */
 146         if ((dev = __dev_get_by_index(t->tcm_ifindex)) == NULL)
 147                 return -ENODEV;
 148
 149         /* Find qdisc */
 150         if (!parent) {
 151                 q = dev->qdisc_sleeping;
 152                 parent = q->handle;
 153         } else if ((q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent))) == NULL)
 154                 return -EINVAL;
 155
 156         /* Is it classful? */
 157         if ((cops = q->ops->cl_ops) == NULL)
 158                 return -EINVAL;
 159
 160         /* Do we search for filter, attached to class? */
 161         if (TC_H_MIN(parent)) {
 162                 cl = cops->get(q, parent);
 163                 if (cl == 0)
 164                         return -ENOENT;
 165         }
 166
 167         /* And the last stroke */
 168         chain = cops->tcf_chain(q, cl);
 169         err = -EINVAL;
 170         if (chain == NULL)
 171                 goto errout;
 172
 173         /* Check the chain for existence of proto-tcf with this priority */
 174         for (back = chain; (tp=*back) != NULL; back = &tp->next) {
 175                 if (tp->prio >= prio) {
 176                         if (tp->prio == prio) {
 177                                 if (!nprio || (tp->protocol != protocol && protocol))
 178                                         goto errout;
 179                         } else
 180                                 tp = NULL;
 181                         break;
 182                 }
 183         }
 184
 185         if (tp == NULL) {
 186                 /* Proto-tcf does not exist, create new one */
 187
 188                 if (tca[TCA_KIND-1] == NULL || !protocol)
 189                         goto errout;
 190
 191                 err = -ENOENT;
 192                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 193                         goto errout;
 194
 195
 196                 /* Create new proto tcf */
 197
 198                 err = -ENOBUFS;
 199                 if ((tp = kmalloc(sizeof(*tp), GFP_KERNEL)) == NULL)
 200                         goto errout;
 201                 tp_ops = tcf_proto_lookup_ops(tca[TCA_KIND-1]);
 202 #ifdef CONFIG_KMOD
 203                 if (tp_ops==NULL && tca[TCA_KIND-1] != NULL) {
 204                         struct rtattr *kind = tca[TCA_KIND-1];
 205                         char module_name[4 + IFNAMSIZ + 1];
 206
 207                         if (RTA_PAYLOAD(kind) <= IFNAMSIZ) {
 208                                 sprintf(module_name, "cls_%s", (char*)RTA_DATA(kind));
 209                                 request_module (module_name);
 210                                 tp_ops = tcf_proto_lookup_ops(kind);
 211                         }
 212                 }
 213 #endif
 214                 if (tp_ops == NULL) {
 215                         err = -EINVAL;
 216                         kfree(tp);
 217                         goto errout;
 218                 }
 219                 memset(tp, 0, sizeof(*tp));
 220                 tp->ops = tp_ops;
 221                 tp->protocol = protocol;
 222                 tp->prio = nprio ? : tcf_auto_prio(*back);
 223                 tp->q = q;
 224                 tp->classify = tp_ops->classify;
 225                 tp->classid = parent;
 226                 err = tp_ops->init(tp);
 227                 if (err) {
 228                         kfree(tp);
 229                         goto errout;
 230                 }
 231                 write_lock(&qdisc_tree_lock);
 232                 spin_lock_bh(&dev->queue_lock);
 233                 tp->next = *back;
 234                 *back = tp;
 235                 spin_unlock_bh(&dev->queue_lock);
 236                 write_unlock(&qdisc_tree_lock);
 237         } else if (tca[TCA_KIND-1] && rtattr_strcmp(tca[TCA_KIND-1], tp->ops->kind))
 238                 goto errout;
 239
 240         fh = tp->ops->get(tp, t->tcm_handle);
 241
 242         if (fh == 0) {
 243                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 244                         write_lock(&qdisc_tree_lock);
 245                         spin_lock_bh(&dev->queue_lock);
 246                         *back = tp->next;
 247                         spin_unlock_bh(&dev->queue_lock);
 248                         write_unlock(&qdisc_tree_lock);
 249                         tcf_destroy(tp);
 250                         err = 0;
 251                         goto errout;
 252                 }
 253
 254                 err = -ENOENT;
 255                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 256                         goto errout;
 257         } else {
 258                 switch (n->nlmsg_type) {
 259                 case RTM_NEWTFILTER:
 260                         err = -EEXIST;
 261                         if (n->nlmsg_flags&NLM_F_EXCL)
 262                                 goto errout;
 263                         break;
 264                 case RTM_DELTFILTER:
 265                         err = tp->ops->delete(tp, fh);
 266                         goto errout;
 267                 case RTM_GETTFILTER:
 268                         err = tfilter_notify(skb, n, tp, fh, RTM_NEWTFILTER);
 269                         goto errout;
 270                 default:
 271                         err = -EINVAL;
 272                         goto errout;
 273                 }
 274         }
 275
 276         err = tp->ops->change(tp, cl, t->tcm_handle, tca, &fh);
 277         if (err == 0)
 278                 tfilter_notify(skb, n, tp, fh, RTM_NEWTFILTER);
 279
 280 errout:
 281         if (cl)
 282                 cops->put(q, cl);
 283         return err;
 284 }
 285
 286 static int
 287 tcf_fill_node(struct sk_buff *skb, struct tcf_proto *tp, unsigned long fh,
 288               u32 pid, u32 seq, unsigned flags, int event)
 289 {
 290         struct tcmsg *tcm;
 291         struct nlmsghdr  *nlh;
 292         unsigned char    *b = skb->tail;
 293
 294         nlh = NLMSG_PUT(skb, pid, seq, event, sizeof(*tcm));
 295         nlh->nlmsg_flags = flags;
 296         tcm = NLMSG_DATA(nlh);
 297         tcm->tcm_family = AF_UNSPEC;
 298         tcm->tcm_ifindex = tp->q->dev->ifindex;
 299         tcm->tcm_parent = tp->classid;
 300         tcm->tcm_handle = 0;
 301         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 302         RTA_PUT(skb, TCA_KIND, IFNAMSIZ, tp->ops->kind);
 303         if (tp->ops->dump && tp->ops->dump(tp, fh, skb, tcm) < 0)
 304                 goto rtattr_failure;
 305         nlh->nlmsg_len = skb->tail - b;
 306         return skb->len;
 307
 308 nlmsg_failure:
 309 rtattr_failure:
 310         skb_trim(skb, b - skb->data);
 311         return -1;
 312 }
 313
 314 static int tfilter_notify(struct sk_buff *oskb, struct nlmsghdr *n,
 315                           struct tcf_proto *tp, unsigned long fh, int event)
 316 {
 317         struct sk_buff *skb;
 318         u32 pid = oskb ? NETLINK_CB(oskb).pid : 0;
 319
 320         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 321         if (!skb)
 322                 return -ENOBUFS;
 323
 324         if (tcf_fill_node(skb, tp, fh, pid, n->nlmsg_seq, 0, event) <= 0) {
 325                 kfree_skb(skb);
 326                 return -EINVAL;
 327         }
 328
 329         return rtnetlink_send(skb, pid, RTMGRP_TC, n->nlmsg_flags&NLM_F_ECHO);
 330 }
 331
 332 struct tcf_dump_args
 333 {
 334         struct tcf_walker w;
 335         struct sk_buff *skb;
 336         struct netlink_callback *cb;
 337 };
 338
 339 static int tcf_node_dump(struct tcf_proto *tp, unsigned long n, struct tcf_walker *arg)
 340 {
 341         struct tcf_dump_args *a = (void*)arg;
 342
 343         return tcf_fill_node(a->skb, tp, n, NETLINK_CB(a->cb->skb).pid,
 344                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER);
 345 }
 346
 347 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
 348 {
 349         int t;
 350         int s_t;
 351         struct net_device *dev;
 352         struct Qdisc *q;
 353         struct tcf_proto *tp, **chain;
 354         struct tcmsg *tcm = (struct tcmsg*)NLMSG_DATA(cb->nlh);
 355         unsigned long cl = 0;
 356         struct Qdisc_class_ops *cops;
 357         struct tcf_dump_args arg;
 358
 359         if (cb->nlh->nlmsg_len < NLMSG_LENGTH(sizeof(*tcm)))
 360                 return skb->len;
 361         if ((dev = dev_get_by_index(tcm->tcm_ifindex)) == NULL)
 362                 return skb->len;
 363
 364         read_lock(&qdisc_tree_lock);
 365         if (!tcm->tcm_parent)
 366                 q = dev->qdisc_sleeping;
 367         else
 368                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
 369         if (q == NULL) {
 370                 read_unlock(&qdisc_tree_lock);
 371                 dev_put(dev);
 372                 return skb->len;
 373         }
 374         if ((cops = q->ops->cl_ops) == NULL)
 375                 goto errout;
 376         if (TC_H_MIN(tcm->tcm_parent)) {
 377                 cl = cops->get(q, tcm->tcm_parent);
 378                 if (cl == 0)
 379                         goto errout;
 380         }
 381         chain = cops->tcf_chain(q, cl);
 382         if (chain == NULL)
 383                 goto errout;
 384
 385         s_t = cb->args[0];
 386
 387         for (tp=*chain, t=0; tp; tp = tp->next, t++) {
 388                 if (t < s_t) continue;
 389                 if (TC_H_MAJ(tcm->tcm_info) &&
 390                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 391                         continue;
 392                 if (TC_H_MIN(tcm->tcm_info) &&
 393                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 394                         continue;
 395                 if (t > s_t)
 396                         memset(&cb->args[1], 0, sizeof(cb->args)-sizeof(cb->args[0]));
 397                 if (cb->args[1] == 0) {
 398                         if (tcf_fill_node(skb, tp, 0, NETLINK_CB(cb->skb).pid,
 399                                           cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER) <= 0) {
 400                                 break;
 401                         }
 402                         cb->args[1] = 1;
 403                 }
 404                 if (tp->ops->walk == NULL)
 405                         continue;
 406                 arg.w.fn = tcf_node_dump;
 407                 arg.skb = skb;
 408                 arg.cb = cb;
 409                 arg.w.stop = 0;
 410                 arg.w.skip = cb->args[1]-1;
 411                 arg.w.count = 0;
 412                 tp->ops->walk(tp, &arg.w);
 413                 cb->args[1] = arg.w.count+1;
 414                 if (arg.w.stop)
 415                         break;
 416         }
 417
 418         cb->args[0] = t;
 419
 420 errout:
 421         if (cl)
 422                 cops->put(q, cl);
 423
 424         read_unlock(&qdisc_tree_lock);
 425         dev_put(dev);
 426         return skb->len;
 427 }
 428
 429
 430 int __init tc_filter_init(void)
 431 {
 432         struct rtnetlink_link *link_p = rtnetlink_links[PF_UNSPEC];
 433
 434         /* Setup rtnetlink links. It is made here to avoid
 435            exporting large number of public symbols.
 436          */
 437
 438         if (link_p) {
 439                 link_p[RTM_NEWTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 440                 link_p[RTM_DELTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 441                 link_p[RTM_GETTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 442                 link_p[RTM_GETTFILTER-RTM_BASE].dumpit = tc_dump_tfilter;
 443         }
 444 #define INIT_TC_FILTER(name) { \
 445           extern struct tcf_proto_ops cls_##name##_ops; \
 446           register_tcf_proto_ops(&cls_##name##_ops); \
 447         }
 448
 449 #ifdef CONFIG_NET_CLS_U32
 450         INIT_TC_FILTER(u32);
 451 #endif
 452 #ifdef CONFIG_NET_CLS_ROUTE4
 453         INIT_TC_FILTER(route4);
 454 #endif
 455 #ifdef CONFIG_NET_CLS_FW
 456         INIT_TC_FILTER(fw);
 457 #endif
 458 #ifdef CONFIG_NET_CLS_RSVP
 459         INIT_TC_FILTER(rsvp);
 460 #endif
 461 #ifdef CONFIG_NET_CLS_TCINDEX
 462         INIT_TC_FILTER(tcindex);
 463 #endif
 464 #ifdef CONFIG_NET_CLS_RSVP6
 465         INIT_TC_FILTER(rsvp6);
 466 #endif
 467         return 0;
 468 }