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/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
  39 /* The list of all installed classifier types */
  40
  41 static struct tcf_proto_ops *tcf_proto_base;
  42
  43 /* Protects list of registered TC modules. It is pure SMP lock. */
  44 static rwlock_t cls_mod_lock = RW_LOCK_UNLOCKED;
  45
  46 /* Find classifier type by string name */
  47
  48 struct tcf_proto_ops * tcf_proto_lookup_ops(struct rtattr *kind)
  49 {
  50         struct tcf_proto_ops *t = NULL;
  51
  52         if (kind) {
  53                 read_lock(&cls_mod_lock);
  54                 for (t = tcf_proto_base; t; t = t->next) {
  55                         if (rtattr_strcmp(kind, t->kind) == 0)
  56                                 break;
  57                 }
  58                 read_unlock(&cls_mod_lock);
  59         }
  60         return t;
  61 }
  62
  63 /* Register(unregister) new classifier type */
  64
  65 int register_tcf_proto_ops(struct tcf_proto_ops *ops)
  66 {
  67         struct tcf_proto_ops *t, **tp;
  68         int rc = -EEXIST;
  69
  70         write_lock(&cls_mod_lock);
  71         for (tp = &tcf_proto_base; (t = *tp) != NULL; tp = &t->next)
  72                 if (!strcmp(ops->kind, t->kind))
  73                         goto out;
  74
  75         ops->next = NULL;
  76         *tp = ops;
  77         rc = 0;
  78 out:
  79         write_unlock(&cls_mod_lock);
  80         return rc;
  81 }
  82
  83 int unregister_tcf_proto_ops(struct tcf_proto_ops *ops)
  84 {
  85         struct tcf_proto_ops *t, **tp;
  86         int rc = -ENOENT;
  87
  88         write_lock(&cls_mod_lock);
  89         for (tp = &tcf_proto_base; (t=*tp) != NULL; tp = &t->next)
  90                 if (t == ops)
  91                         break;
  92
  93         if (!t)
  94                 goto out;
  95         *tp = t->next;
  96         rc = 0;
  97 out:
  98         write_unlock(&cls_mod_lock);
  99         return rc;
 100 }
 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 net_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
 208                         if (RTA_PAYLOAD(kind) <= IFNAMSIZ) {
 209                                 request_module("cls_%s", (char*)RTA_DATA(kind));
 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 = -EBUSY;
 227                 if (!try_module_get(tp_ops->owner) ||
 228                     (err = tp_ops->init(tp)) != 0) {
 229                         kfree(tp);
 230                         goto errout;
 231                 }
 232                 write_lock(&qdisc_tree_lock);
 233                 spin_lock_bh(&dev->queue_lock);
 234                 tp->next = *back;
 235                 *back = tp;
 236                 spin_unlock_bh(&dev->queue_lock);
 237                 write_unlock(&qdisc_tree_lock);
 238         } else if (tca[TCA_KIND-1] && rtattr_strcmp(tca[TCA_KIND-1], tp->ops->kind))
 239                 goto errout;
 240
 241         fh = tp->ops->get(tp, t->tcm_handle);
 242
 243         if (fh == 0) {
 244                 if (n->nlmsg_type == RTM_DELTFILTER && t->tcm_handle == 0) {
 245                         write_lock(&qdisc_tree_lock);
 246                         spin_lock_bh(&dev->queue_lock);
 247                         *back = tp->next;
 248                         spin_unlock_bh(&dev->queue_lock);
 249                         write_unlock(&qdisc_tree_lock);
 250                         tcf_destroy(tp);
 251                         err = 0;
 252                         goto errout;
 253                 }
 254
 255                 err = -ENOENT;
 256                 if (n->nlmsg_type != RTM_NEWTFILTER || !(n->nlmsg_flags&NLM_F_CREATE))
 257                         goto errout;
 258         } else {
 259                 switch (n->nlmsg_type) {
 260                 case RTM_NEWTFILTER:
 261                         err = -EEXIST;
 262                         if (n->nlmsg_flags&NLM_F_EXCL)
 263                                 goto errout;
 264                         break;
 265                 case RTM_DELTFILTER:
 266                         err = tp->ops->delete(tp, fh);
 267                         goto errout;
 268                 case RTM_GETTFILTER:
 269                         err = tfilter_notify(skb, n, tp, fh, RTM_NEWTFILTER);
 270                         goto errout;
 271                 default:
 272                         err = -EINVAL;
 273                         goto errout;
 274                 }
 275         }
 276
 277         err = tp->ops->change(tp, cl, t->tcm_handle, tca, &fh);
 278         if (err == 0)
 279                 tfilter_notify(skb, n, tp, fh, RTM_NEWTFILTER);
 280
 281 errout:
 282         if (cl)
 283                 cops->put(q, cl);
 284         return err;
 285 }
 286
 287 static int
 288 tcf_fill_node(struct sk_buff *skb, struct tcf_proto *tp, unsigned long fh,
 289               u32 pid, u32 seq, unsigned flags, int event)
 290 {
 291         struct tcmsg *tcm;
 292         struct nlmsghdr  *nlh;
 293         unsigned char    *b = skb->tail;
 294
 295         nlh = NLMSG_PUT(skb, pid, seq, event, sizeof(*tcm));
 296         nlh->nlmsg_flags = flags;
 297         tcm = NLMSG_DATA(nlh);
 298         tcm->tcm_family = AF_UNSPEC;
 299         tcm->tcm_ifindex = tp->q->dev->ifindex;
 300         tcm->tcm_parent = tp->classid;
 301         tcm->tcm_handle = 0;
 302         tcm->tcm_info = TC_H_MAKE(tp->prio, tp->protocol);
 303         RTA_PUT(skb, TCA_KIND, IFNAMSIZ, tp->ops->kind);
 304         if (tp->ops->dump && tp->ops->dump(tp, fh, skb, tcm) < 0)
 305                 goto rtattr_failure;
 306         nlh->nlmsg_len = skb->tail - b;
 307         return skb->len;
 308
 309 nlmsg_failure:
 310 rtattr_failure:
 311         skb_trim(skb, b - skb->data);
 312         return -1;
 313 }
 314
 315 static int tfilter_notify(struct sk_buff *oskb, struct nlmsghdr *n,
 316                           struct tcf_proto *tp, unsigned long fh, int event)
 317 {
 318         struct sk_buff *skb;
 319         u32 pid = oskb ? NETLINK_CB(oskb).pid : 0;
 320
 321         skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
 322         if (!skb)
 323                 return -ENOBUFS;
 324
 325         if (tcf_fill_node(skb, tp, fh, pid, n->nlmsg_seq, 0, event) <= 0) {
 326                 kfree_skb(skb);
 327                 return -EINVAL;
 328         }
 329
 330         return rtnetlink_send(skb, pid, RTMGRP_TC, n->nlmsg_flags&NLM_F_ECHO);
 331 }
 332
 333 struct tcf_dump_args
 334 {
 335         struct tcf_walker w;
 336         struct sk_buff *skb;
 337         struct netlink_callback *cb;
 338 };
 339
 340 static int tcf_node_dump(struct tcf_proto *tp, unsigned long n, struct tcf_walker *arg)
 341 {
 342         struct tcf_dump_args *a = (void*)arg;
 343
 344         return tcf_fill_node(a->skb, tp, n, NETLINK_CB(a->cb->skb).pid,
 345                              a->cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER);
 346 }
 347
 348 static int tc_dump_tfilter(struct sk_buff *skb, struct netlink_callback *cb)
 349 {
 350         int t;
 351         int s_t;
 352         struct net_device *dev;
 353         struct Qdisc *q;
 354         struct tcf_proto *tp, **chain;
 355         struct tcmsg *tcm = (struct tcmsg*)NLMSG_DATA(cb->nlh);
 356         unsigned long cl = 0;
 357         struct Qdisc_class_ops *cops;
 358         struct tcf_dump_args arg;
 359
 360         if (cb->nlh->nlmsg_len < NLMSG_LENGTH(sizeof(*tcm)))
 361                 return skb->len;
 362         if ((dev = dev_get_by_index(tcm->tcm_ifindex)) == NULL)
 363                 return skb->len;
 364
 365         read_lock(&qdisc_tree_lock);
 366         if (!tcm->tcm_parent)
 367                 q = dev->qdisc_sleeping;
 368         else
 369                 q = qdisc_lookup(dev, TC_H_MAJ(tcm->tcm_parent));
 370         if (!q)
 371                 goto out;
 372         if ((cops = q->ops->cl_ops) == NULL)
 373                 goto errout;
 374         if (TC_H_MIN(tcm->tcm_parent)) {
 375                 cl = cops->get(q, tcm->tcm_parent);
 376                 if (cl == 0)
 377                         goto errout;
 378         }
 379         chain = cops->tcf_chain(q, cl);
 380         if (chain == NULL)
 381                 goto errout;
 382
 383         s_t = cb->args[0];
 384
 385         for (tp=*chain, t=0; tp; tp = tp->next, t++) {
 386                 if (t < s_t) continue;
 387                 if (TC_H_MAJ(tcm->tcm_info) &&
 388                     TC_H_MAJ(tcm->tcm_info) != tp->prio)
 389                         continue;
 390                 if (TC_H_MIN(tcm->tcm_info) &&
 391                     TC_H_MIN(tcm->tcm_info) != tp->protocol)
 392                         continue;
 393                 if (t > s_t)
 394                         memset(&cb->args[1], 0, sizeof(cb->args)-sizeof(cb->args[0]));
 395                 if (cb->args[1] == 0) {
 396                         if (tcf_fill_node(skb, tp, 0, NETLINK_CB(cb->skb).pid,
 397                                           cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTFILTER) <= 0) {
 398                                 break;
 399                         }
 400                         cb->args[1] = 1;
 401                 }
 402                 if (tp->ops->walk == NULL)
 403                         continue;
 404                 arg.w.fn = tcf_node_dump;
 405                 arg.skb = skb;
 406                 arg.cb = cb;
 407                 arg.w.stop = 0;
 408                 arg.w.skip = cb->args[1]-1;
 409                 arg.w.count = 0;
 410                 tp->ops->walk(tp, &arg.w);
 411                 cb->args[1] = arg.w.count+1;
 412                 if (arg.w.stop)
 413                         break;
 414         }
 415
 416         cb->args[0] = t;
 417
 418 errout:
 419         if (cl)
 420                 cops->put(q, cl);
 421 out:
 422         read_unlock(&qdisc_tree_lock);
 423         dev_put(dev);
 424         return skb->len;
 425 }
 426
 427
 428 static int __init tc_filter_init(void)
 429 {
 430         struct rtnetlink_link *link_p = rtnetlink_links[PF_UNSPEC];
 431
 432         /* Setup rtnetlink links. It is made here to avoid
 433            exporting large number of public symbols.
 434          */
 435
 436         if (link_p) {
 437                 link_p[RTM_NEWTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 438                 link_p[RTM_DELTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 439                 link_p[RTM_GETTFILTER-RTM_BASE].doit = tc_ctl_tfilter;
 440                 link_p[RTM_GETTFILTER-RTM_BASE].dumpit = tc_dump_tfilter;
 441         }
 442         return 0;
 443 }
 444
 445 subsys_initcall(tc_filter_init);
 446
 447 EXPORT_SYMBOL(register_tcf_proto_ops);
 448 EXPORT_SYMBOL(unregister_tcf_proto_ops);