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 <asm/bitops.h>
  20 #include <linux/config.h>
  21 #include <linux/module.h>
  22 #include <linux/types.h>
  23 #include <linux/kernel.h>
  24 #include <linux/sched.h>
  25 #include <linux/string.h>
  26 #include <linux/mm.h>
  27 #include <linux/socket.h>
  28 #include <linux/sockios.h>
  29 #include <linux/in.h>
  30 #include <linux/errno.h>
  31 #include <linux/interrupt.h>
  32 #include <linux/netdevice.h>
  33 #include <linux/skbuff.h>
  34 #include <linux/rtnetlink.h>
  35 #include <linux/init.h>
  36 #include <linux/kmod.h>
  37 #include <net/sock.h>
  38 #include <net/pkt_sched.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 rwlock_t cls_mod_lock = RW_LOCK_UNLOCKED;
  52
  53 /* Find classifier type by string name */
  54
  55 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                                 break;
  64                 }
  65                 read_unlock(&cls_mod_lock);
  66         }
  67         return t;
  68 }
  69
  70 /* Register(unregister) new classifier type */
  71
  72 int register_tcf_proto_ops(struct tcf_proto_ops *ops)
  73 {
  74         struct tcf_proto_ops *t, **tp;
  75         int rc = -EEXIST;
  76
  77         write_lock(&cls_mod_lock);
  78         for (tp = &tcf_proto_base; (t = *tp) != NULL; tp = &t->next)
  79                 if (!strcmp(ops->kind, t->kind))
  80                         goto out;
  81
  82         ops->next = NULL;
  83         *tp = ops;
  84         rc = 0;
  85 out:
  86         write_unlock(&cls_mod_lock);
  87         return rc;
  88 }
  89
  90 int unregister_tcf_proto_ops(struct tcf_proto_ops *ops)
  91 {
  92         struct tcf_proto_ops *t, **tp;
  93         int rc = -ENOENT;
  94
  95         write_lock(&cls_mod_lock);
  96         for (tp = &tcf_proto_base; (t=*tp) != NULL; tp = &t->next)
  97                 if (t == ops)
  98                         break;
  99
 100         if (!t)
 101                 goto out;
 102         *tp = t->next;
 103         rc = 0;
 104 out:
 105         write_unlock(&cls_mod_lock);
 106         return rc;
 107 }
 108
 109 static int tfilter_notify(struct sk_buff *oskb, struct nlmsghdr *n,
 110                           struct tcf_proto *tp, unsigned long fh, int event);
 111
 112
 113 /* Select new prio value from the range, managed by kernel. */
 114
 115 static __inline__ u32 tcf_auto_prio(struct tcf_proto *tp)
 116 {
 117         u32 first = TC_H_MAKE(0xC0000000U,0U);
 118
 119         if (tp)
 120                 first = tp->prio-1;
 121
 122         return first;
 123 }
 124
 125 /* Add/change/delete/get a filter node */
 126
 127 static int tc_ctl_tfilter(struct sk_buff *skb, struct nlmsghdr *n, void *arg)
 128 {
 129         struct rtattr **tca = arg;
 130         struct tcmsg *t = NLMSG_DATA(n);
 131         u32 protocol = TC_H_MIN(t->tcm_info);
 132         u32 prio = TC_H_MAJ(t->tcm_info);
 133         u32 nprio = prio;
 134         u32 parent = t->tcm_parent;
 135         struct net_device *dev;
 136         struct Qdisc  *q;
 137         struct tcf_proto **back, **chain;
 138         struct tcf_proto *tp = NULL;
 139         struct tcf_proto_ops *tp_ops;
 140         struct Qdisc_class_ops *cops;
 141         unsigned long cl = 0;
 142         unsigned long fh, fh_s;
 143         int err;
 144
 145         if (prio == 0) {
 146                 /* If no priority is given, user wants we allocated it. */
 147                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 148                         return -ENOENT;
 149                 prio = TC_H_MAKE(0x80000000U,0U);
 150         }
 151
 152         /* Find head of filter chain. */
 153
 154         /* Find link */
 155         if ((dev = __dev_get_by_index(t->tcm_ifindex)) == NULL)
 156                 return -ENODEV;
 157
 158         /* Find qdisc */
 159         if (!parent) {
 160                 q = dev->qdisc_sleeping;
 161                 parent = q->handle;
 162         } else if ((q = qdisc_lookup(dev, TC_H_MAJ(t->tcm_parent))) == NULL)
 163                 return -EINVAL;
 164
 165         /* Is it classful? */
 166         if ((cops = q->ops->cl_ops) == NULL)
 167                 return -EINVAL;
 168
 169         /* Do we search for filter, attached to class? */
 170         if (TC_H_MIN(parent)) {
 171                 cl = cops->get(q, parent);
 172                 if (cl == 0)
 173                         return -ENOENT;
 174         }
 175
 176         /* And the last stroke */
 177         chain = cops->tcf_chain(q, cl);
 178         err = -EINVAL;
 179         if (chain == NULL)
 180                 goto errout;
 181
 182         /* Check the chain for existence of proto-tcf with this priority */
 183         for (back = chain; (tp=*back) != NULL; back = &tp->next) {
 184                 if (tp->prio >= prio) {
 185                         if (tp->prio == prio) {
 186                                 if (!nprio || (tp->protocol != protocol && protocol))
 187                                         goto errout;
 188                         } else
 189                                 tp = NULL;
 190                         break;
 191                 }
 192         }
 193
 194         if (tp == NULL) {
 195                 /* Proto-tcf does not exist, create new one */
 196
 197                 if (tca[TCA_KIND-1] == NULL || !protocol)
 198                         goto errout;
 199
 200                 err = -ENOENT;
 201                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 202                         goto errout;
 203
 204
 205                 /* Create new proto tcf */
 206
 207                 err = -ENOBUFS;
 208                 if ((tp = kmalloc(sizeof(*tp), GFP_KERNEL)) == NULL)
 209                         goto errout;
 210                 tp_ops = tcf_proto_lookup_ops(tca[TCA_KIND-1]);
 211 #ifdef CONFIG_KMOD
 212                 if (tp_ops==NULL && tca[TCA_KIND-1] != NULL) {
 213                         struct rtattr *kind = tca[TCA_KIND-1];
 214
 215                         if (RTA_PAYLOAD(kind) <= IFNAMSIZ) {
 216                                 request_module("cls_%s", (char*)RTA_DATA(kind));
 217                                 tp_ops = tcf_proto_lookup_ops(kind);
 218                         }
 219                 }
 220 #endif
 221                 if (tp_ops == NULL) {
 222                         err = -EINVAL;
 223                         kfree(tp);
 224                         goto errout;
 225                 }
 226                 memset(tp, 0, sizeof(*tp));
 227                 tp->ops = tp_ops;
 228                 tp->protocol = protocol;
 229                 tp->prio = nprio ? : tcf_auto_prio(*back);
 230                 tp->q = q;
 231                 tp->classify = tp_ops->classify;
 232                 tp->classid = parent;
 233                 err = -EBUSY;
 234                 if (!try_module_get(tp_ops->owner) ||
 235                     (err = tp_ops->init(tp)) != 0) {
 236                         kfree(tp);
 237                         goto errout;
 238                 }
 239
 240                 qdisc_lock_tree(dev);
 241                 tp->next = *back;
 242                 *back = tp;
 243                 qdisc_unlock_tree(dev);
 244
 245         } else if (tca[TCA_KIND-1] && rtattr_strcmp(tca[TCA_KIND-1], tp->ops->kind))
 246                 goto errout;
 247
 248         fh_s = fh = tp->ops->get(tp, t->tcm_handle);
 249
 250         if (fh == 0) {
 251                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 252                         qdisc_lock_tree(dev);
 253                         *back = tp->next;
 254                         qdisc_unlock_tree(dev);
 255
 256                         tfilter_notify(skb, n, tp, fh_s, RTM_DELTFILTER);
 257                         tcf_destroy(tp);
 258                         err = 0;
 259                         goto errout;
 260                 }
 261
 262                 err = -ENOENT;
 263                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 264                         goto errout;
 265         } else {
 266                 switch (n->nlmsg_type) {
 267                 case RTM_NEWTFILTER:
 268                         err = -EEXIST;
 269                         if (n->nlmsg_flags&NLM_F_EXCL)
 270                                 goto errout;
 271                         break;
 272                 case RTM_DELTFILTER:
 273                         err = tp->ops->delete(tp, fh);
 274                         if (err == 0)
 275                                 tfilter_notify(skb, n, tp, fh_s, RTM_DELTFILTER);
 276                         goto errout;
 277                 case RTM_GETTFILTER:
 278                         err = tfilter_notify(skb, n, tp, fh, RTM_NEWTFILTER);
 279                         goto errout;
 280                 default:
 281                         err = -EINVAL;
 282                         goto errout;
 283                 }
 284         }
 285
 286         err = tp->ops->change(tp, cl, t->tcm_handle, tca, &fh);
 287         if (err == 0)
 288                 tfilter_notify(skb, n, tp, fh, RTM_NEWTFILTER);
 289
 290 errout:
 291         if (cl)
 292                 cops->put(q, cl);
 293         return err;
 294 }
 295
 296 unsigned long tcf_set_class(struct tcf_proto *tp, unsigned long *clp,
 297                             unsigned long cl)
 298 {
 299         unsigned long old_cl;
 300
 301         tcf_tree_lock(tp);
 302         old_cl = __cls_set_class(clp, cl);
 303         tcf_tree_unlock(tp);
 304
 305         return old_cl;
 306 }
 307
 308
 309 static int
 310 tcf_fill_node(struct sk_buff *skb, struct tcf_proto *tp, unsigned long fh,
 311               u32 pid, u32 seq, unsigned flags, int event)
 312 {
 313         struct tcmsg *tcm;
 314         struct nlmsghdr  *nlh;
 315         unsigned char    *b = skb->tail;
 316
 317         nlh = NLMSG_PUT(skb, pid, seq, event, sizeof(*tcm));
 318         nlh->nlmsg_flags = flags;
 319         tcm = NLMSG_DATA(nlh);
 320         tcm->tcm_family = AF_UNSPEC;
 321         tcm->tcm_ifindex = tp->q->dev->ifindex;
 322         tcm->tcm_parent = tp->classid;
 323         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 324         RTA_PUT(skb, TCA_KIND, IFNAMSIZ, tp->ops->kind);
 325         tcm->tcm_handle = fh;
 326         if (RTM_DELTFILTER != event) {
 327                 tcm->tcm_handle = 0;
 328                 if (tp->ops->dump && tp->ops->dump(tp, fh, skb, tcm) < 0)
 329                         goto rtattr_failure;
 330         }
 331         nlh->nlmsg_len = skb->tail - b;
 332         return skb->len;
 333
 334 nlmsg_failure:
 335 rtattr_failure:
 336         skb_trim(skb, b - skb->data);
 337         return -1;
 338 }
 339
 340 static int tfilter_notify(struct sk_buff *oskb, struct nlmsghdr *n,
 341                           struct tcf_proto *tp, unsigned long fh, int event)
 342 {
 343         struct sk_buff *skb;
 344         u32 pid = oskb ? NETLINK_CB(oskb).pid : 0;
 345
 346         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 347         if (!skb)
 348                 return -ENOBUFS;
 349
 350         if (tcf_fill_node(skb, tp, fh, pid, n->nlmsg_seq, 0, event) <= 0) {
 351                 kfree_skb(skb);
 352                 return -EINVAL;
 353         }
 354
 355         return rtnetlink_send(skb, pid, RTMGRP_TC, n->nlmsg_flags&NLM_F_ECHO);
 356 }
 357
 358 struct tcf_dump_args
 359 {
 360         struct tcf_walker w;
 361         struct sk_buff *skb;
 362         struct netlink_callback *cb;
 363 };
 364
 365 static int tcf_node_dump(struct tcf_proto *tp, unsigned long n, struct tcf_walker *arg)
 366 {
 367         struct tcf_dump_args *a = (void*)arg;
 368
 369         return tcf_fill_node(a->skb, tp, n, NETLINK_CB(a->cb->skb).pid,
 370                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER);
 371 }
 372
 373 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
 374 {
 375         int t;
 376         int s_t;
 377         struct net_device *dev;
 378         struct Qdisc *q;
 379         struct tcf_proto *tp, **chain;
 380         struct tcmsg *tcm = (struct tcmsg*)NLMSG_DATA(cb->nlh);
 381         unsigned long cl = 0;
 382         struct Qdisc_class_ops *cops;
 383         struct tcf_dump_args arg;
 384
 385         if (cb->nlh->nlmsg_len < NLMSG_LENGTH(sizeof(*tcm)))
 386                 return skb->len;
 387         if ((dev = dev_get_by_index(tcm->tcm_ifindex)) == NULL)
 388                 return skb->len;
 389
 390         read_lock_bh(&qdisc_tree_lock);
 391         if (!tcm->tcm_parent)
 392                 q = dev->qdisc_sleeping;
 393         else
 394                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
 395         if (!q)
 396                 goto out;
 397         if ((cops = q->ops->cl_ops) == NULL)
 398                 goto errout;
 399         if (TC_H_MIN(tcm->tcm_parent)) {
 400                 cl = cops->get(q, tcm->tcm_parent);
 401                 if (cl == 0)
 402                         goto errout;
 403         }
 404         chain = cops->tcf_chain(q, cl);
 405         if (chain == NULL)
 406                 goto errout;
 407
 408         s_t = cb->args[0];
 409
 410         for (tp=*chain, t=0; tp; tp = tp->next, t++) {
 411                 if (t < s_t) continue;
 412                 if (TC_H_MAJ(tcm->tcm_info) &&
 413                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 414                         continue;
 415                 if (TC_H_MIN(tcm->tcm_info) &&
 416                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 417                         continue;
 418                 if (t > s_t)
 419                         memset(&cb->args[1], 0, sizeof(cb->args)-sizeof(cb->args[0]));
 420                 if (cb->args[1] == 0) {
 421                         if (tcf_fill_node(skb, tp, 0, NETLINK_CB(cb->skb).pid,
 422                                           cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER) <= 0) {
 423                                 break;
 424                         }
 425                         cb->args[1] = 1;
 426                 }
 427                 if (tp->ops->walk == NULL)
 428                         continue;
 429                 arg.w.fn = tcf_node_dump;
 430                 arg.skb = skb;
 431                 arg.cb = cb;
 432                 arg.w.stop = 0;
 433                 arg.w.skip = cb->args[1]-1;
 434                 arg.w.count = 0;
 435                 tp->ops->walk(tp, &arg.w);
 436                 cb->args[1] = arg.w.count+1;
 437                 if (arg.w.stop)
 438                         break;
 439         }
 440
 441         cb->args[0] = t;
 442
 443 errout:
 444         if (cl)
 445                 cops->put(q, cl);
 446 out:
 447         read_unlock_bh(&qdisc_tree_lock);
 448         dev_put(dev);
 449         return skb->len;
 450 }
 451
 452
 453 static int __init tc_filter_init(void)
 454 {
 455         struct rtnetlink_link *link_p = rtnetlink_links[PF_UNSPEC];
 456
 457         /* Setup rtnetlink links. It is made here to avoid
 458            exporting large number of public symbols.
 459          */
 460
 461         if (link_p) {
 462                 link_p[RTM_NEWTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 463                 link_p[RTM_DELTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 464                 link_p[RTM_GETTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 465                 link_p[RTM_GETTFILTER-RTM_BASE].dumpit = tc_dump_tfilter;
 466         }
 467         return 0;
 468 }
 469
 470 subsys_initcall(tc_filter_init);
 471
 472 EXPORT_SYMBOL(register_tcf_proto_ops);
 473 EXPORT_SYMBOL(unregister_tcf_proto_ops);
 474 EXPORT_SYMBOL(tcf_set_class);