net/core/fib_rules.c

   1 /*
   2  * net/core/fib_rules.c         Generic Routing Rules
   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 as
   6  *      published by the Free Software Foundation, version 2.
   7  *
   8  * Authors:     Thomas Graf <[email protected]>
   9  */
  10
  11 #include <linux/types.h>
  12 #include <linux/kernel.h>
  13 #include <linux/slab.h>
  14 #include <linux/list.h>
  15 #include <linux/module.h>
  16 #include <net/net_namespace.h>
  17 #include <net/sock.h>
  18 #include <net/fib_rules.h>
  19 #include <net/ip_tunnels.h>
  20
  21 static const struct fib_kuid_range fib_kuid_range_unset = {
  22         KUIDT_INIT(0),
  23         KUIDT_INIT(~0),
  24 };
  25
  26 bool fib_rule_matchall(const struct fib_rule *rule)
  27 {
  28         if (rule->iifindex || rule->oifindex || rule->mark || rule->tun_id ||
  29             rule->flags)
  30                 return false;
  31         if (rule->suppress_ifgroup != -1 || rule->suppress_prefixlen != -1)
  32                 return false;
  33         if (!uid_eq(rule->uid_range.start, fib_kuid_range_unset.start) ||
  34             !uid_eq(rule->uid_range.end, fib_kuid_range_unset.end))
  35                 return false;
  36         return true;
  37 }
  38 EXPORT_SYMBOL_GPL(fib_rule_matchall);
  39
  40 int fib_default_rule_add(struct fib_rules_ops *ops,
  41                          u32 pref, u32 table, u32 flags)
  42 {
  43         struct fib_rule *r;
  44
  45         r = kzalloc(ops->rule_size, GFP_KERNEL);
  46         if (r == NULL)
  47                 return -ENOMEM;
  48
  49         refcount_set(&r->refcnt, 1);
  50         r->action = FR_ACT_TO_TBL;
  51         r->pref = pref;
  52         r->table = table;
  53         r->flags = flags;
  54         r->fr_net = ops->fro_net;
  55         r->uid_range = fib_kuid_range_unset;
  56
  57         r->suppress_prefixlen = -1;
  58         r->suppress_ifgroup = -1;
  59
  60         /* The lock is not required here, the list in unreacheable
  61          * at the moment this function is called */
  62         list_add_tail(&r->list, &ops->rules_list);
  63         return 0;
  64 }
  65 EXPORT_SYMBOL(fib_default_rule_add);
  66
  67 static u32 fib_default_rule_pref(struct fib_rules_ops *ops)
  68 {
  69         struct list_head *pos;
  70         struct fib_rule *rule;
  71
  72         if (!list_empty(&ops->rules_list)) {
  73                 pos = ops->rules_list.next;
  74                 if (pos->next != &ops->rules_list) {
  75                         rule = list_entry(pos->next, struct fib_rule, list);
  76                         if (rule->pref)
  77                                 return rule->pref - 1;
  78                 }
  79         }
  80
  81         return 0;
  82 }
  83
  84 static void notify_rule_change(int event, struct fib_rule *rule,
  85                                struct fib_rules_ops *ops, struct nlmsghdr *nlh,
  86                                u32 pid);
  87
  88 static struct fib_rules_ops *lookup_rules_ops(struct net *net, int family)
  89 {
  90         struct fib_rules_ops *ops;
  91
  92         rcu_read_lock();
  93         list_for_each_entry_rcu(ops, &net->rules_ops, list) {
  94                 if (ops->family == family) {
  95                         if (!try_module_get(ops->owner))
  96                                 ops = NULL;
  97                         rcu_read_unlock();
  98                         return ops;
  99                 }
 100         }
 101         rcu_read_unlock();
 102
 103         return NULL;
 104 }
 105
 106 static void rules_ops_put(struct fib_rules_ops *ops)
 107 {
 108         if (ops)
 109                 module_put(ops->owner);
 110 }
 111
 112 static void flush_route_cache(struct fib_rules_ops *ops)
 113 {
 114         if (ops->flush_cache)
 115                 ops->flush_cache(ops);
 116 }
 117
 118 static int __fib_rules_register(struct fib_rules_ops *ops)
 119 {
 120         int err = -EEXIST;
 121         struct fib_rules_ops *o;
 122         struct net *net;
 123
 124         net = ops->fro_net;
 125
 126         if (ops->rule_size < sizeof(struct fib_rule))
 127                 return -EINVAL;
 128
 129         if (ops->match == NULL || ops->configure == NULL ||
 130             ops->compare == NULL || ops->fill == NULL ||
 131             ops->action == NULL)
 132                 return -EINVAL;
 133
 134         spin_lock(&net->rules_mod_lock);
 135         list_for_each_entry(o, &net->rules_ops, list)
 136                 if (ops->family == o->family)
 137                         goto errout;
 138
 139         list_add_tail_rcu(&ops->list, &net->rules_ops);
 140         err = 0;
 141 errout:
 142         spin_unlock(&net->rules_mod_lock);
 143
 144         return err;
 145 }
 146
 147 struct fib_rules_ops *
 148 fib_rules_register(const struct fib_rules_ops *tmpl, struct net *net)
 149 {
 150         struct fib_rules_ops *ops;
 151         int err;
 152
 153         ops = kmemdup(tmpl, sizeof(*ops), GFP_KERNEL);
 154         if (ops == NULL)
 155                 return ERR_PTR(-ENOMEM);
 156
 157         INIT_LIST_HEAD(&ops->rules_list);
 158         ops->fro_net = net;
 159
 160         err = __fib_rules_register(ops);
 161         if (err) {
 162                 kfree(ops);
 163                 ops = ERR_PTR(err);
 164         }
 165
 166         return ops;
 167 }
 168 EXPORT_SYMBOL_GPL(fib_rules_register);
 169
 170 static void fib_rules_cleanup_ops(struct fib_rules_ops *ops)
 171 {
 172         struct fib_rule *rule, *tmp;
 173
 174         list_for_each_entry_safe(rule, tmp, &ops->rules_list, list) {
 175                 list_del_rcu(&rule->list);
 176                 if (ops->delete)
 177                         ops->delete(rule);
 178                 fib_rule_put(rule);
 179         }
 180 }
 181
 182 void fib_rules_unregister(struct fib_rules_ops *ops)
 183 {
 184         struct net *net = ops->fro_net;
 185
 186         spin_lock(&net->rules_mod_lock);
 187         list_del_rcu(&ops->list);
 188         spin_unlock(&net->rules_mod_lock);
 189
 190         fib_rules_cleanup_ops(ops);
 191         kfree_rcu(ops, rcu);
 192 }
 193 EXPORT_SYMBOL_GPL(fib_rules_unregister);
 194
 195 static int uid_range_set(struct fib_kuid_range *range)
 196 {
 197         return uid_valid(range->start) && uid_valid(range->end);
 198 }
 199
 200 static struct fib_kuid_range nla_get_kuid_range(struct nlattr **tb)
 201 {
 202         struct fib_rule_uid_range *in;
 203         struct fib_kuid_range out;
 204
 205         in = (struct fib_rule_uid_range *)nla_data(tb[FRA_UID_RANGE]);
 206
 207         out.start = make_kuid(current_user_ns(), in->start);
 208         out.end = make_kuid(current_user_ns(), in->end);
 209
 210         return out;
 211 }
 212
 213 static int nla_put_uid_range(struct sk_buff *skb, struct fib_kuid_range *range)
 214 {
 215         struct fib_rule_uid_range out = {
 216                 from_kuid_munged(current_user_ns(), range->start),
 217                 from_kuid_munged(current_user_ns(), range->end)
 218         };
 219
 220         return nla_put(skb, FRA_UID_RANGE, sizeof(out), &out);
 221 }
 222
 223 static int fib_rule_match(struct fib_rule *rule, struct fib_rules_ops *ops,
 224                           struct flowi *fl, int flags,
 225                           struct fib_lookup_arg *arg)
 226 {
 227         int ret = 0;
 228
 229         if (rule->iifindex && (rule->iifindex != fl->flowi_iif))
 230                 goto out;
 231
 232         if (rule->oifindex && (rule->oifindex != fl->flowi_oif))
 233                 goto out;
 234
 235         if ((rule->mark ^ fl->flowi_mark) & rule->mark_mask)
 236                 goto out;
 237
 238         if (rule->tun_id && (rule->tun_id != fl->flowi_tun_key.tun_id))
 239                 goto out;
 240
 241         if (rule->l3mdev && !l3mdev_fib_rule_match(rule->fr_net, fl, arg))
 242                 goto out;
 243
 244         if (uid_lt(fl->flowi_uid, rule->uid_range.start) ||
 245             uid_gt(fl->flowi_uid, rule->uid_range.end))
 246                 goto out;
 247
 248         ret = ops->match(rule, fl, flags);
 249 out:
 250         return (rule->flags & FIB_RULE_INVERT) ? !ret : ret;
 251 }
 252
 253 int fib_rules_lookup(struct fib_rules_ops *ops, struct flowi *fl,
 254                      int flags, struct fib_lookup_arg *arg)
 255 {
 256         struct fib_rule *rule;
 257         int err;
 258
 259         rcu_read_lock();
 260
 261         list_for_each_entry_rcu(rule, &ops->rules_list, list) {
 262 jumped:
 263                 if (!fib_rule_match(rule, ops, fl, flags, arg))
 264                         continue;
 265
 266                 if (rule->action == FR_ACT_GOTO) {
 267                         struct fib_rule *target;
 268
 269                         target = rcu_dereference(rule->ctarget);
 270                         if (target == NULL) {
 271                                 continue;
 272                         } else {
 273                                 rule = target;
 274                                 goto jumped;
 275                         }
 276                 } else if (rule->action == FR_ACT_NOP)
 277                         continue;
 278                 else
 279                         err = ops->action(rule, fl, flags, arg);
 280
 281                 if (!err && ops->suppress && ops->suppress(rule, arg))
 282                         continue;
 283
 284                 if (err != -EAGAIN) {
 285                         if ((arg->flags & FIB_LOOKUP_NOREF) ||
 286                             likely(refcount_inc_not_zero(&rule->refcnt))) {
 287                                 arg->rule = rule;
 288                                 goto out;
 289                         }
 290                         break;
 291                 }
 292         }
 293
 294         err = -ESRCH;
 295 out:
 296         rcu_read_unlock();
 297
 298         return err;
 299 }
 300 EXPORT_SYMBOL_GPL(fib_rules_lookup);
 301
 302 static int validate_rulemsg(struct fib_rule_hdr *frh, struct nlattr **tb,
 303                             struct fib_rules_ops *ops)
 304 {
 305         int err = -EINVAL;
 306
 307         if (frh->src_len)
 308                 if (tb[FRA_SRC] == NULL ||
 309                     frh->src_len > (ops->addr_size * 8) ||
 310                     nla_len(tb[FRA_SRC]) != ops->addr_size)
 311                         goto errout;
 312
 313         if (frh->dst_len)
 314                 if (tb[FRA_DST] == NULL ||
 315                     frh->dst_len > (ops->addr_size * 8) ||
 316                     nla_len(tb[FRA_DST]) != ops->addr_size)
 317                         goto errout;
 318
 319         err = 0;
 320 errout:
 321         return err;
 322 }
 323
 324 static int rule_exists(struct fib_rules_ops *ops, struct fib_rule_hdr *frh,
 325                        struct nlattr **tb, struct fib_rule *rule)
 326 {
 327         struct fib_rule *r;
 328
 329         list_for_each_entry(r, &ops->rules_list, list) {
 330                 if (r->action != rule->action)
 331                         continue;
 332
 333                 if (r->table != rule->table)
 334                         continue;
 335
 336                 if (r->pref != rule->pref)
 337                         continue;
 338
 339                 if (memcmp(r->iifname, rule->iifname, IFNAMSIZ))
 340                         continue;
 341
 342                 if (memcmp(r->oifname, rule->oifname, IFNAMSIZ))
 343                         continue;
 344
 345                 if (r->mark != rule->mark)
 346                         continue;
 347
 348                 if (r->mark_mask != rule->mark_mask)
 349                         continue;
 350
 351                 if (r->tun_id != rule->tun_id)
 352                         continue;
 353
 354                 if (r->fr_net != rule->fr_net)
 355                         continue;
 356
 357                 if (r->l3mdev != rule->l3mdev)
 358                         continue;
 359
 360                 if (!uid_eq(r->uid_range.start, rule->uid_range.start) ||
 361                     !uid_eq(r->uid_range.end, rule->uid_range.end))
 362                         continue;
 363
 364                 if (!ops->compare(r, frh, tb))
 365                         continue;
 366                 return 1;
 367         }
 368         return 0;
 369 }
 370
 371 int fib_nl_newrule(struct sk_buff *skb, struct nlmsghdr *nlh,
 372                    struct netlink_ext_ack *extack)
 373 {
 374         struct net *net = sock_net(skb->sk);
 375         struct fib_rule_hdr *frh = nlmsg_data(nlh);
 376         struct fib_rules_ops *ops = NULL;
 377         struct fib_rule *rule, *r, *last = NULL;
 378         struct nlattr *tb[FRA_MAX+1];
 379         int err = -EINVAL, unresolved = 0;
 380
 381         if (nlh->nlmsg_len < nlmsg_msg_size(sizeof(*frh)))
 382                 goto errout;
 383
 384         ops = lookup_rules_ops(net, frh->family);
 385         if (ops == NULL) {
 386                 err = -EAFNOSUPPORT;
 387                 goto errout;
 388         }
 389
 390         err = nlmsg_parse(nlh, sizeof(*frh), tb, FRA_MAX, ops->policy, extack);
 391         if (err < 0)
 392                 goto errout;
 393
 394         err = validate_rulemsg(frh, tb, ops);
 395         if (err < 0)
 396                 goto errout;
 397
 398         rule = kzalloc(ops->rule_size, GFP_KERNEL);
 399         if (rule == NULL) {
 400                 err = -ENOMEM;
 401                 goto errout;
 402         }
 403         refcount_set(&rule->refcnt, 1);
 404         rule->fr_net = net;
 405
 406         rule->pref = tb[FRA_PRIORITY] ? nla_get_u32(tb[FRA_PRIORITY])
 407                                       : fib_default_rule_pref(ops);
 408
 409         if (tb[FRA_IIFNAME]) {
 410                 struct net_device *dev;
 411
 412                 rule->iifindex = -1;
 413                 nla_strlcpy(rule->iifname, tb[FRA_IIFNAME], IFNAMSIZ);
 414                 dev = __dev_get_by_name(net, rule->iifname);
 415                 if (dev)
 416                         rule->iifindex = dev->ifindex;
 417         }
 418
 419         if (tb[FRA_OIFNAME]) {
 420                 struct net_device *dev;
 421
 422                 rule->oifindex = -1;
 423                 nla_strlcpy(rule->oifname, tb[FRA_OIFNAME], IFNAMSIZ);
 424                 dev = __dev_get_by_name(net, rule->oifname);
 425                 if (dev)
 426                         rule->oifindex = dev->ifindex;
 427         }
 428
 429         if (tb[FRA_FWMARK]) {
 430                 rule->mark = nla_get_u32(tb[FRA_FWMARK]);
 431                 if (rule->mark)
 432                         /* compatibility: if the mark value is non-zero all bits
 433                          * are compared unless a mask is explicitly specified.
 434                          */
 435                         rule->mark_mask = 0xFFFFFFFF;
 436         }
 437
 438         if (tb[FRA_FWMASK])
 439                 rule->mark_mask = nla_get_u32(tb[FRA_FWMASK]);
 440
 441         if (tb[FRA_TUN_ID])
 442                 rule->tun_id = nla_get_be64(tb[FRA_TUN_ID]);
 443
 444         err = -EINVAL;
 445         if (tb[FRA_L3MDEV]) {
 446 #ifdef CONFIG_NET_L3_MASTER_DEV
 447                 rule->l3mdev = nla_get_u8(tb[FRA_L3MDEV]);
 448                 if (rule->l3mdev != 1)
 449 #endif
 450                         goto errout_free;
 451         }
 452
 453         rule->action = frh->action;
 454         rule->flags = frh->flags;
 455         rule->table = frh_get_table(frh, tb);
 456         if (tb[FRA_SUPPRESS_PREFIXLEN])
 457                 rule->suppress_prefixlen = nla_get_u32(tb[FRA_SUPPRESS_PREFIXLEN]);
 458         else
 459                 rule->suppress_prefixlen = -1;
 460
 461         if (tb[FRA_SUPPRESS_IFGROUP])
 462                 rule->suppress_ifgroup = nla_get_u32(tb[FRA_SUPPRESS_IFGROUP]);
 463         else
 464                 rule->suppress_ifgroup = -1;
 465
 466         if (tb[FRA_GOTO]) {
 467                 if (rule->action != FR_ACT_GOTO)
 468                         goto errout_free;
 469
 470                 rule->target = nla_get_u32(tb[FRA_GOTO]);
 471                 /* Backward jumps are prohibited to avoid endless loops */
 472                 if (rule->target <= rule->pref)
 473                         goto errout_free;
 474
 475                 list_for_each_entry(r, &ops->rules_list, list) {
 476                         if (r->pref == rule->target) {
 477                                 RCU_INIT_POINTER(rule->ctarget, r);
 478                                 break;
 479                         }
 480                 }
 481
 482                 if (rcu_dereference_protected(rule->ctarget, 1) == NULL)
 483                         unresolved = 1;
 484         } else if (rule->action == FR_ACT_GOTO)
 485                 goto errout_free;
 486
 487         if (rule->l3mdev && rule->table)
 488                 goto errout_free;
 489
 490         if (tb[FRA_UID_RANGE]) {
 491                 if (current_user_ns() != net->user_ns) {
 492                         err = -EPERM;
 493                         goto errout_free;
 494                 }
 495
 496                 rule->uid_range = nla_get_kuid_range(tb);
 497
 498                 if (!uid_range_set(&rule->uid_range) ||
 499                     !uid_lte(rule->uid_range.start, rule->uid_range.end))
 500                         goto errout_free;
 501         } else {
 502                 rule->uid_range = fib_kuid_range_unset;
 503         }
 504
 505         if ((nlh->nlmsg_flags & NLM_F_EXCL) &&
 506             rule_exists(ops, frh, tb, rule)) {
 507                 err = -EEXIST;
 508                 goto errout_free;
 509         }
 510
 511         err = ops->configure(rule, skb, frh, tb);
 512         if (err < 0)
 513                 goto errout_free;
 514
 515         list_for_each_entry(r, &ops->rules_list, list) {
 516                 if (r->pref > rule->pref)
 517                         break;
 518                 last = r;
 519         }
 520
 521         if (last)
 522                 list_add_rcu(&rule->list, &last->list);
 523         else
 524                 list_add_rcu(&rule->list, &ops->rules_list);
 525
 526         if (ops->unresolved_rules) {
 527                 /*
 528                  * There are unresolved goto rules in the list, check if
 529                  * any of them are pointing to this new rule.
 530                  */
 531                 list_for_each_entry(r, &ops->rules_list, list) {
 532                         if (r->action == FR_ACT_GOTO &&
 533                             r->target == rule->pref &&
 534                             rtnl_dereference(r->ctarget) == NULL) {
 535                                 rcu_assign_pointer(r->ctarget, rule);
 536                                 if (--ops->unresolved_rules == 0)
 537                                         break;
 538                         }
 539                 }
 540         }
 541
 542         if (rule->action == FR_ACT_GOTO)
 543                 ops->nr_goto_rules++;
 544
 545         if (unresolved)
 546                 ops->unresolved_rules++;
 547
 548         if (rule->tun_id)
 549                 ip_tunnel_need_metadata();
 550
 551         notify_rule_change(RTM_NEWRULE, rule, ops, nlh, NETLINK_CB(skb).portid);
 552         flush_route_cache(ops);
 553         rules_ops_put(ops);
 554         return 0;
 555
 556 errout_free:
 557         kfree(rule);
 558 errout:
 559         rules_ops_put(ops);
 560         return err;
 561 }
 562 EXPORT_SYMBOL_GPL(fib_nl_newrule);
 563
 564 int fib_nl_delrule(struct sk_buff *skb, struct nlmsghdr *nlh,
 565                    struct netlink_ext_ack *extack)
 566 {
 567         struct net *net = sock_net(skb->sk);
 568         struct fib_rule_hdr *frh = nlmsg_data(nlh);
 569         struct fib_rules_ops *ops = NULL;
 570         struct fib_rule *rule, *r;
 571         struct nlattr *tb[FRA_MAX+1];
 572         struct fib_kuid_range range;
 573         int err = -EINVAL;
 574
 575         if (nlh->nlmsg_len < nlmsg_msg_size(sizeof(*frh)))
 576                 goto errout;
 577
 578         ops = lookup_rules_ops(net, frh->family);
 579         if (ops == NULL) {
 580                 err = -EAFNOSUPPORT;
 581                 goto errout;
 582         }
 583
 584         err = nlmsg_parse(nlh, sizeof(*frh), tb, FRA_MAX, ops->policy, extack);
 585         if (err < 0)
 586                 goto errout;
 587
 588         err = validate_rulemsg(frh, tb, ops);
 589         if (err < 0)
 590                 goto errout;
 591
 592         if (tb[FRA_UID_RANGE]) {
 593                 range = nla_get_kuid_range(tb);
 594                 if (!uid_range_set(&range)) {
 595                         err = -EINVAL;
 596                         goto errout;
 597                 }
 598         } else {
 599                 range = fib_kuid_range_unset;
 600         }
 601
 602         list_for_each_entry(rule, &ops->rules_list, list) {
 603                 if (frh->action && (frh->action != rule->action))
 604                         continue;
 605
 606                 if (frh_get_table(frh, tb) &&
 607                     (frh_get_table(frh, tb) != rule->table))
 608                         continue;
 609
 610                 if (tb[FRA_PRIORITY] &&
 611                     (rule->pref != nla_get_u32(tb[FRA_PRIORITY])))
 612                         continue;
 613
 614                 if (tb[FRA_IIFNAME] &&
 615                     nla_strcmp(tb[FRA_IIFNAME], rule->iifname))
 616                         continue;
 617
 618                 if (tb[FRA_OIFNAME] &&
 619                     nla_strcmp(tb[FRA_OIFNAME], rule->oifname))
 620                         continue;
 621
 622                 if (tb[FRA_FWMARK] &&
 623                     (rule->mark != nla_get_u32(tb[FRA_FWMARK])))
 624                         continue;
 625
 626                 if (tb[FRA_FWMASK] &&
 627                     (rule->mark_mask != nla_get_u32(tb[FRA_FWMASK])))
 628                         continue;
 629
 630                 if (tb[FRA_TUN_ID] &&
 631                     (rule->tun_id != nla_get_be64(tb[FRA_TUN_ID])))
 632                         continue;
 633
 634                 if (tb[FRA_L3MDEV] &&
 635                     (rule->l3mdev != nla_get_u8(tb[FRA_L3MDEV])))
 636                         continue;
 637
 638                 if (uid_range_set(&range) &&
 639                     (!uid_eq(rule->uid_range.start, range.start) ||
 640                      !uid_eq(rule->uid_range.end, range.end)))
 641                         continue;
 642
 643                 if (!ops->compare(rule, frh, tb))
 644                         continue;
 645
 646                 if (rule->flags & FIB_RULE_PERMANENT) {
 647                         err = -EPERM;
 648                         goto errout;
 649                 }
 650
 651                 if (ops->delete) {
 652                         err = ops->delete(rule);
 653                         if (err)
 654                                 goto errout;
 655                 }
 656
 657                 if (rule->tun_id)
 658                         ip_tunnel_unneed_metadata();
 659
 660                 list_del_rcu(&rule->list);
 661
 662                 if (rule->action == FR_ACT_GOTO) {
 663                         ops->nr_goto_rules--;
 664                         if (rtnl_dereference(rule->ctarget) == NULL)
 665                                 ops->unresolved_rules--;
 666                 }
 667
 668                 /*
 669                  * Check if this rule is a target to any of them. If so,
 670                  * adjust to the next one with the same preference or
 671                  * disable them. As this operation is eventually very
 672                  * expensive, it is only performed if goto rules, except
 673                  * current if it is goto rule, have actually been added.
 674                  */
 675                 if (ops->nr_goto_rules > 0) {
 676                         struct fib_rule *n;
 677
 678                         n = list_next_entry(rule, list);
 679                         if (&n->list == &ops->rules_list || n->pref != rule->pref)
 680                                 n = NULL;
 681                         list_for_each_entry(r, &ops->rules_list, list) {
 682                                 if (rtnl_dereference(r->ctarget) != rule)
 683                                         continue;
 684                                 rcu_assign_pointer(r->ctarget, n);
 685                                 if (!n)
 686                                         ops->unresolved_rules++;
 687                         }
 688                 }
 689
 690                 notify_rule_change(RTM_DELRULE, rule, ops, nlh,
 691                                    NETLINK_CB(skb).portid);
 692                 fib_rule_put(rule);
 693                 flush_route_cache(ops);
 694                 rules_ops_put(ops);
 695                 return 0;
 696         }
 697
 698         err = -ENOENT;
 699 errout:
 700         rules_ops_put(ops);
 701         return err;
 702 }
 703 EXPORT_SYMBOL_GPL(fib_nl_delrule);
 704
 705 static inline size_t fib_rule_nlmsg_size(struct fib_rules_ops *ops,
 706                                          struct fib_rule *rule)
 707 {
 708         size_t payload = NLMSG_ALIGN(sizeof(struct fib_rule_hdr))
 709                          + nla_total_size(IFNAMSIZ) /* FRA_IIFNAME */
 710                          + nla_total_size(IFNAMSIZ) /* FRA_OIFNAME */
 711                          + nla_total_size(4) /* FRA_PRIORITY */
 712                          + nla_total_size(4) /* FRA_TABLE */
 713                          + nla_total_size(4) /* FRA_SUPPRESS_PREFIXLEN */
 714                          + nla_total_size(4) /* FRA_SUPPRESS_IFGROUP */
 715                          + nla_total_size(4) /* FRA_FWMARK */
 716                          + nla_total_size(4) /* FRA_FWMASK */
 717                          + nla_total_size_64bit(8) /* FRA_TUN_ID */
 718                          + nla_total_size(sizeof(struct fib_kuid_range));
 719
 720         if (ops->nlmsg_payload)
 721                 payload += ops->nlmsg_payload(rule);
 722
 723         return payload;
 724 }
 725
 726 static int fib_nl_fill_rule(struct sk_buff *skb, struct fib_rule *rule,
 727                             u32 pid, u32 seq, int type, int flags,
 728                             struct fib_rules_ops *ops)
 729 {
 730         struct nlmsghdr *nlh;
 731         struct fib_rule_hdr *frh;
 732
 733         nlh = nlmsg_put(skb, pid, seq, type, sizeof(*frh), flags);
 734         if (nlh == NULL)
 735                 return -EMSGSIZE;
 736
 737         frh = nlmsg_data(nlh);
 738         frh->family = ops->family;
 739         frh->table = rule->table;
 740         if (nla_put_u32(skb, FRA_TABLE, rule->table))
 741                 goto nla_put_failure;
 742         if (nla_put_u32(skb, FRA_SUPPRESS_PREFIXLEN, rule->suppress_prefixlen))
 743                 goto nla_put_failure;
 744         frh->res1 = 0;
 745         frh->res2 = 0;
 746         frh->action = rule->action;
 747         frh->flags = rule->flags;
 748
 749         if (rule->action == FR_ACT_GOTO &&
 750             rcu_access_pointer(rule->ctarget) == NULL)
 751                 frh->flags |= FIB_RULE_UNRESOLVED;
 752
 753         if (rule->iifname[0]) {
 754                 if (nla_put_string(skb, FRA_IIFNAME, rule->iifname))
 755                         goto nla_put_failure;
 756                 if (rule->iifindex == -1)
 757                         frh->flags |= FIB_RULE_IIF_DETACHED;
 758         }
 759
 760         if (rule->oifname[0]) {
 761                 if (nla_put_string(skb, FRA_OIFNAME, rule->oifname))
 762                         goto nla_put_failure;
 763                 if (rule->oifindex == -1)
 764                         frh->flags |= FIB_RULE_OIF_DETACHED;
 765         }
 766
 767         if ((rule->pref &&
 768              nla_put_u32(skb, FRA_PRIORITY, rule->pref)) ||
 769             (rule->mark &&
 770              nla_put_u32(skb, FRA_FWMARK, rule->mark)) ||
 771             ((rule->mark_mask || rule->mark) &&
 772              nla_put_u32(skb, FRA_FWMASK, rule->mark_mask)) ||
 773             (rule->target &&
 774              nla_put_u32(skb, FRA_GOTO, rule->target)) ||
 775             (rule->tun_id &&
 776              nla_put_be64(skb, FRA_TUN_ID, rule->tun_id, FRA_PAD)) ||
 777             (rule->l3mdev &&
 778              nla_put_u8(skb, FRA_L3MDEV, rule->l3mdev)) ||
 779             (uid_range_set(&rule->uid_range) &&
 780              nla_put_uid_range(skb, &rule->uid_range)))
 781                 goto nla_put_failure;
 782
 783         if (rule->suppress_ifgroup != -1) {
 784                 if (nla_put_u32(skb, FRA_SUPPRESS_IFGROUP, rule->suppress_ifgroup))
 785                         goto nla_put_failure;
 786         }
 787
 788         if (ops->fill(rule, skb, frh) < 0)
 789                 goto nla_put_failure;
 790
 791         nlmsg_end(skb, nlh);
 792         return 0;
 793
 794 nla_put_failure:
 795         nlmsg_cancel(skb, nlh);
 796         return -EMSGSIZE;
 797 }
 798
 799 static int dump_rules(struct sk_buff *skb, struct netlink_callback *cb,
 800                       struct fib_rules_ops *ops)
 801 {
 802         int idx = 0;
 803         struct fib_rule *rule;
 804         int err = 0;
 805
 806         rcu_read_lock();
 807         list_for_each_entry_rcu(rule, &ops->rules_list, list) {
 808                 if (idx < cb->args[1])
 809                         goto skip;
 810
 811                 err = fib_nl_fill_rule(skb, rule, NETLINK_CB(cb->skb).portid,
 812                                        cb->nlh->nlmsg_seq, RTM_NEWRULE,
 813                                        NLM_F_MULTI, ops);
 814                 if (err)
 815                         break;
 816 skip:
 817                 idx++;
 818         }
 819         rcu_read_unlock();
 820         cb->args[1] = idx;
 821         rules_ops_put(ops);
 822
 823         return err;
 824 }
 825
 826 static int fib_nl_dumprule(struct sk_buff *skb, struct netlink_callback *cb)
 827 {
 828         struct net *net = sock_net(skb->sk);
 829         struct fib_rules_ops *ops;
 830         int idx = 0, family;
 831
 832         family = rtnl_msg_family(cb->nlh);
 833         if (family != AF_UNSPEC) {
 834                 /* Protocol specific dump request */
 835                 ops = lookup_rules_ops(net, family);
 836                 if (ops == NULL)
 837                         return -EAFNOSUPPORT;
 838
 839                 dump_rules(skb, cb, ops);
 840
 841                 return skb->len;
 842         }
 843
 844         rcu_read_lock();
 845         list_for_each_entry_rcu(ops, &net->rules_ops, list) {
 846                 if (idx < cb->args[0] || !try_module_get(ops->owner))
 847                         goto skip;
 848
 849                 if (dump_rules(skb, cb, ops) < 0)
 850                         break;
 851
 852                 cb->args[1] = 0;
 853 skip:
 854                 idx++;
 855         }
 856         rcu_read_unlock();
 857         cb->args[0] = idx;
 858
 859         return skb->len;
 860 }
 861
 862 static void notify_rule_change(int event, struct fib_rule *rule,
 863                                struct fib_rules_ops *ops, struct nlmsghdr *nlh,
 864                                u32 pid)
 865 {
 866         struct net *net;
 867         struct sk_buff *skb;
 868         int err = -ENOBUFS;
 869
 870         net = ops->fro_net;
 871         skb = nlmsg_new(fib_rule_nlmsg_size(ops, rule), GFP_KERNEL);
 872         if (skb == NULL)
 873                 goto errout;
 874
 875         err = fib_nl_fill_rule(skb, rule, pid, nlh->nlmsg_seq, event, 0, ops);
 876         if (err < 0) {
 877                 /* -EMSGSIZE implies BUG in fib_rule_nlmsg_size() */
 878                 WARN_ON(err == -EMSGSIZE);
 879                 kfree_skb(skb);
 880                 goto errout;
 881         }
 882
 883         rtnl_notify(skb, net, pid, ops->nlgroup, nlh, GFP_KERNEL);
 884         return;
 885 errout:
 886         if (err < 0)
 887                 rtnl_set_sk_err(net, ops->nlgroup, err);
 888 }
 889
 890 static void attach_rules(struct list_head *rules, struct net_device *dev)
 891 {
 892         struct fib_rule *rule;
 893
 894         list_for_each_entry(rule, rules, list) {
 895                 if (rule->iifindex == -1 &&
 896                     strcmp(dev->name, rule->iifname) == 0)
 897                         rule->iifindex = dev->ifindex;
 898                 if (rule->oifindex == -1 &&
 899                     strcmp(dev->name, rule->oifname) == 0)
 900                         rule->oifindex = dev->ifindex;
 901         }
 902 }
 903
 904 static void detach_rules(struct list_head *rules, struct net_device *dev)
 905 {
 906         struct fib_rule *rule;
 907
 908         list_for_each_entry(rule, rules, list) {
 909                 if (rule->iifindex == dev->ifindex)
 910                         rule->iifindex = -1;
 911                 if (rule->oifindex == dev->ifindex)
 912                         rule->oifindex = -1;
 913         }
 914 }
 915
 916
 917 static int fib_rules_event(struct notifier_block *this, unsigned long event,
 918                            void *ptr)
 919 {
 920         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 921         struct net *net = dev_net(dev);
 922         struct fib_rules_ops *ops;
 923
 924         ASSERT_RTNL();
 925
 926         switch (event) {
 927         case NETDEV_REGISTER:
 928                 list_for_each_entry(ops, &net->rules_ops, list)
 929                         attach_rules(&ops->rules_list, dev);
 930                 break;
 931
 932         case NETDEV_CHANGENAME:
 933                 list_for_each_entry(ops, &net->rules_ops, list) {
 934                         detach_rules(&ops->rules_list, dev);
 935                         attach_rules(&ops->rules_list, dev);
 936                 }
 937                 break;
 938
 939         case NETDEV_UNREGISTER:
 940                 list_for_each_entry(ops, &net->rules_ops, list)
 941                         detach_rules(&ops->rules_list, dev);
 942                 break;
 943         }
 944
 945         return NOTIFY_DONE;
 946 }
 947
 948 static struct notifier_block fib_rules_notifier = {
 949         .notifier_call = fib_rules_event,
 950 };
 951
 952 static int __net_init fib_rules_net_init(struct net *net)
 953 {
 954         INIT_LIST_HEAD(&net->rules_ops);
 955         spin_lock_init(&net->rules_mod_lock);
 956         return 0;
 957 }
 958
 959 static struct pernet_operations fib_rules_net_ops = {
 960         .init = fib_rules_net_init,
 961 };
 962
 963 static int __init fib_rules_init(void)
 964 {
 965         int err;
 966         rtnl_register(PF_UNSPEC, RTM_NEWRULE, fib_nl_newrule, NULL, NULL);
 967         rtnl_register(PF_UNSPEC, RTM_DELRULE, fib_nl_delrule, NULL, NULL);
 968         rtnl_register(PF_UNSPEC, RTM_GETRULE, NULL, fib_nl_dumprule, NULL);
 969
 970         err = register_pernet_subsys(&fib_rules_net_ops);
 971         if (err < 0)
 972                 goto fail;
 973
 974         err = register_netdevice_notifier(&fib_rules_notifier);
 975         if (err < 0)
 976                 goto fail_unregister;
 977
 978         return 0;
 979
 980 fail_unregister:
 981         unregister_pernet_subsys(&fib_rules_net_ops);
 982 fail:
 983         rtnl_unregister(PF_UNSPEC, RTM_NEWRULE);
 984         rtnl_unregister(PF_UNSPEC, RTM_DELRULE);
 985         rtnl_unregister(PF_UNSPEC, RTM_GETRULE);
 986         return err;
 987 }
 988
 989 subsys_initcall(fib_rules_init);