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