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 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         if (fib_rule_port_range_set(&rule->sport_range))
  37                 return false;
  38         if (fib_rule_port_range_set(&rule->dport_range))
  39                 return false;
  40         return true;
  41 }
  42 EXPORT_SYMBOL_GPL(fib_rule_matchall);
  43
  44 int fib_default_rule_add(struct fib_rules_ops *ops,
  45                          u32 pref, u32 table, u32 flags)
  46 {
  47         struct fib_rule *r;
  48
  49         r = kzalloc(ops->rule_size, GFP_KERNEL);
  50         if (r == NULL)
  51                 return -ENOMEM;
  52
  53         refcount_set(&r->refcnt, 1);
  54         r->action = FR_ACT_TO_TBL;
  55         r->pref = pref;
  56         r->table = table;
  57         r->flags = flags;
  58         r->proto = RTPROT_KERNEL;
  59         r->fr_net = ops->fro_net;
  60         r->uid_range = fib_kuid_range_unset;
  61
  62         r->suppress_prefixlen = -1;
  63         r->suppress_ifgroup = -1;
  64
  65         /* The lock is not required here, the list in unreacheable
  66          * at the moment this function is called */
  67         list_add_tail(&r->list, &ops->rules_list);
  68         return 0;
  69 }
  70 EXPORT_SYMBOL(fib_default_rule_add);
  71
  72 static u32 fib_default_rule_pref(struct fib_rules_ops *ops)
  73 {
  74         struct list_head *pos;
  75         struct fib_rule *rule;
  76
  77         if (!list_empty(&ops->rules_list)) {
  78                 pos = ops->rules_list.next;
  79                 if (pos->next != &ops->rules_list) {
  80                         rule = list_entry(pos->next, struct fib_rule, list);
  81                         if (rule->pref)
  82                                 return rule->pref - 1;
  83                 }
  84         }
  85
  86         return 0;
  87 }
  88
  89 static void notify_rule_change(int event, struct fib_rule *rule,
  90                                struct fib_rules_ops *ops, struct nlmsghdr *nlh,
  91                                u32 pid);
  92
  93 static struct fib_rules_ops *lookup_rules_ops(struct net *net, int family)
  94 {
  95         struct fib_rules_ops *ops;
  96
  97         rcu_read_lock();
  98         list_for_each_entry_rcu(ops, &net->rules_ops, list) {
  99                 if (ops->family == family) {
 100                         if (!try_module_get(ops->owner))
 101                                 ops = NULL;
 102                         rcu_read_unlock();
 103                         return ops;
 104                 }
 105         }
 106         rcu_read_unlock();
 107
 108         return NULL;
 109 }
 110
 111 static void rules_ops_put(struct fib_rules_ops *ops)
 112 {
 113         if (ops)
 114                 module_put(ops->owner);
 115 }
 116
 117 static void flush_route_cache(struct fib_rules_ops *ops)
 118 {
 119         if (ops->flush_cache)
 120                 ops->flush_cache(ops);
 121 }
 122
 123 static int __fib_rules_register(struct fib_rules_ops *ops)
 124 {
 125         int err = -EEXIST;
 126         struct fib_rules_ops *o;
 127         struct net *net;
 128
 129         net = ops->fro_net;
 130
 131         if (ops->rule_size < sizeof(struct fib_rule))
 132                 return -EINVAL;
 133
 134         if (ops->match == NULL || ops->configure == NULL ||
 135             ops->compare == NULL || ops->fill == NULL ||
 136             ops->action == NULL)
 137                 return -EINVAL;
 138
 139         spin_lock(&net->rules_mod_lock);
 140         list_for_each_entry(o, &net->rules_ops, list)
 141                 if (ops->family == o->family)
 142                         goto errout;
 143
 144         list_add_tail_rcu(&ops->list, &net->rules_ops);
 145         err = 0;
 146 errout:
 147         spin_unlock(&net->rules_mod_lock);
 148
 149         return err;
 150 }
 151
 152 struct fib_rules_ops *
 153 fib_rules_register(const struct fib_rules_ops *tmpl, struct net *net)
 154 {
 155         struct fib_rules_ops *ops;
 156         int err;
 157
 158         ops = kmemdup(tmpl, sizeof(*ops), GFP_KERNEL);
 159         if (ops == NULL)
 160                 return ERR_PTR(-ENOMEM);
 161
 162         INIT_LIST_HEAD(&ops->rules_list);
 163         ops->fro_net = net;
 164
 165         err = __fib_rules_register(ops);
 166         if (err) {
 167                 kfree(ops);
 168                 ops = ERR_PTR(err);
 169         }
 170
 171         return ops;
 172 }
 173 EXPORT_SYMBOL_GPL(fib_rules_register);
 174
 175 static void fib_rules_cleanup_ops(struct fib_rules_ops *ops)
 176 {
 177         struct fib_rule *rule, *tmp;
 178
 179         list_for_each_entry_safe(rule, tmp, &ops->rules_list, list) {
 180                 list_del_rcu(&rule->list);
 181                 if (ops->delete)
 182                         ops->delete(rule);
 183                 fib_rule_put(rule);
 184         }
 185 }
 186
 187 void fib_rules_unregister(struct fib_rules_ops *ops)
 188 {
 189         struct net *net = ops->fro_net;
 190
 191         spin_lock(&net->rules_mod_lock);
 192         list_del_rcu(&ops->list);
 193         spin_unlock(&net->rules_mod_lock);
 194
 195         fib_rules_cleanup_ops(ops);
 196         kfree_rcu(ops, rcu);
 197 }
 198 EXPORT_SYMBOL_GPL(fib_rules_unregister);
 199
 200 static int uid_range_set(struct fib_kuid_range *range)
 201 {
 202         return uid_valid(range->start) && uid_valid(range->end);
 203 }
 204
 205 static struct fib_kuid_range nla_get_kuid_range(struct nlattr **tb)
 206 {
 207         struct fib_rule_uid_range *in;
 208         struct fib_kuid_range out;
 209
 210         in = (struct fib_rule_uid_range *)nla_data(tb[FRA_UID_RANGE]);
 211
 212         out.start = make_kuid(current_user_ns(), in->start);
 213         out.end = make_kuid(current_user_ns(), in->end);
 214
 215         return out;
 216 }
 217
 218 static int nla_put_uid_range(struct sk_buff *skb, struct fib_kuid_range *range)
 219 {
 220         struct fib_rule_uid_range out = {
 221                 from_kuid_munged(current_user_ns(), range->start),
 222                 from_kuid_munged(current_user_ns(), range->end)
 223         };
 224
 225         return nla_put(skb, FRA_UID_RANGE, sizeof(out), &out);
 226 }
 227
 228 static int nla_get_port_range(struct nlattr *pattr,
 229                               struct fib_rule_port_range *port_range)
 230 {
 231         const struct fib_rule_port_range *pr = nla_data(pattr);
 232
 233         if (!fib_rule_port_range_valid(pr))
 234                 return -EINVAL;
 235
 236         port_range->start = pr->start;
 237         port_range->end = pr->end;
 238
 239         return 0;
 240 }
 241
 242 static int nla_put_port_range(struct sk_buff *skb, int attrtype,
 243                               struct fib_rule_port_range *range)
 244 {
 245         return nla_put(skb, attrtype, sizeof(*range), range);
 246 }
 247
 248 static int fib_rule_match(struct fib_rule *rule, struct fib_rules_ops *ops,
 249                           struct flowi *fl, int flags,
 250                           struct fib_lookup_arg *arg)
 251 {
 252         int ret = 0;
 253
 254         if (rule->iifindex && (rule->iifindex != fl->flowi_iif))
 255                 goto out;
 256
 257         if (rule->oifindex && (rule->oifindex != fl->flowi_oif))
 258                 goto out;
 259
 260         if ((rule->mark ^ fl->flowi_mark) & rule->mark_mask)
 261                 goto out;
 262
 263         if (rule->tun_id && (rule->tun_id != fl->flowi_tun_key.tun_id))
 264                 goto out;
 265
 266         if (rule->l3mdev && !l3mdev_fib_rule_match(rule->fr_net, fl, arg))
 267                 goto out;
 268
 269         if (uid_lt(fl->flowi_uid, rule->uid_range.start) ||
 270             uid_gt(fl->flowi_uid, rule->uid_range.end))
 271                 goto out;
 272
 273         ret = ops->match(rule, fl, flags);
 274 out:
 275         return (rule->flags & FIB_RULE_INVERT) ? !ret : ret;
 276 }
 277
 278 int fib_rules_lookup(struct fib_rules_ops *ops, struct flowi *fl,
 279                      int flags, struct fib_lookup_arg *arg)
 280 {
 281         struct fib_rule *rule;
 282         int err;
 283
 284         rcu_read_lock();
 285
 286         list_for_each_entry_rcu(rule, &ops->rules_list, list) {
 287 jumped:
 288                 if (!fib_rule_match(rule, ops, fl, flags, arg))
 289                         continue;
 290
 291                 if (rule->action == FR_ACT_GOTO) {
 292                         struct fib_rule *target;
 293
 294                         target = rcu_dereference(rule->ctarget);
 295                         if (target == NULL) {
 296                                 continue;
 297                         } else {
 298                                 rule = target;
 299                                 goto jumped;
 300                         }
 301                 } else if (rule->action == FR_ACT_NOP)
 302                         continue;
 303                 else
 304                         err = ops->action(rule, fl, flags, arg);
 305
 306                 if (!err && ops->suppress && ops->suppress(rule, arg))
 307                         continue;
 308
 309                 if (err != -EAGAIN) {
 310                         if ((arg->flags & FIB_LOOKUP_NOREF) ||
 311                             likely(refcount_inc_not_zero(&rule->refcnt))) {
 312                                 arg->rule = rule;
 313                                 goto out;
 314                         }
 315                         break;
 316                 }
 317         }
 318
 319         err = -ESRCH;
 320 out:
 321         rcu_read_unlock();
 322
 323         return err;
 324 }
 325 EXPORT_SYMBOL_GPL(fib_rules_lookup);
 326
 327 static int call_fib_rule_notifier(struct notifier_block *nb, struct net *net,
 328                                   enum fib_event_type event_type,
 329                                   struct fib_rule *rule, int family)
 330 {
 331         struct fib_rule_notifier_info info = {
 332                 .info.family = family,
 333                 .rule = rule,
 334         };
 335
 336         return call_fib_notifier(nb, net, event_type, &info.info);
 337 }
 338
 339 static int call_fib_rule_notifiers(struct net *net,
 340                                    enum fib_event_type event_type,
 341                                    struct fib_rule *rule,
 342                                    struct fib_rules_ops *ops,
 343                                    struct netlink_ext_ack *extack)
 344 {
 345         struct fib_rule_notifier_info info = {
 346                 .info.family = ops->family,
 347                 .info.extack = extack,
 348                 .rule = rule,
 349         };
 350
 351         ops->fib_rules_seq++;
 352         return call_fib_notifiers(net, event_type, &info.info);
 353 }
 354
 355 /* Called with rcu_read_lock() */
 356 int fib_rules_dump(struct net *net, struct notifier_block *nb, int family)
 357 {
 358         struct fib_rules_ops *ops;
 359         struct fib_rule *rule;
 360
 361         ops = lookup_rules_ops(net, family);
 362         if (!ops)
 363                 return -EAFNOSUPPORT;
 364         list_for_each_entry_rcu(rule, &ops->rules_list, list)
 365                 call_fib_rule_notifier(nb, net, FIB_EVENT_RULE_ADD, rule,
 366                                        family);
 367         rules_ops_put(ops);
 368
 369         return 0;
 370 }
 371 EXPORT_SYMBOL_GPL(fib_rules_dump);
 372
 373 unsigned int fib_rules_seq_read(struct net *net, int family)
 374 {
 375         unsigned int fib_rules_seq;
 376         struct fib_rules_ops *ops;
 377
 378         ASSERT_RTNL();
 379
 380         ops = lookup_rules_ops(net, family);
 381         if (!ops)
 382                 return 0;
 383         fib_rules_seq = ops->fib_rules_seq;
 384         rules_ops_put(ops);
 385
 386         return fib_rules_seq;
 387 }
 388 EXPORT_SYMBOL_GPL(fib_rules_seq_read);
 389
 390 static struct fib_rule *rule_find(struct fib_rules_ops *ops,
 391                                   struct fib_rule_hdr *frh,
 392                                   struct nlattr **tb,
 393                                   struct fib_rule *rule,
 394                                   bool user_priority)
 395 {
 396         struct fib_rule *r;
 397
 398         list_for_each_entry(r, &ops->rules_list, list) {
 399                 if (rule->action && r->action != rule->action)
 400                         continue;
 401
 402                 if (rule->table && r->table != rule->table)
 403                         continue;
 404
 405                 if (user_priority && r->pref != rule->pref)
 406                         continue;
 407
 408                 if (rule->iifname[0] &&
 409                     memcmp(r->iifname, rule->iifname, IFNAMSIZ))
 410                         continue;
 411
 412                 if (rule->oifname[0] &&
 413                     memcmp(r->oifname, rule->oifname, IFNAMSIZ))
 414                         continue;
 415
 416                 if (rule->mark && r->mark != rule->mark)
 417                         continue;
 418
 419                 if (rule->suppress_ifgroup != -1 &&
 420                     r->suppress_ifgroup != rule->suppress_ifgroup)
 421                         continue;
 422
 423                 if (rule->suppress_prefixlen != -1 &&
 424                     r->suppress_prefixlen != rule->suppress_prefixlen)
 425                         continue;
 426
 427                 if (rule->mark_mask && r->mark_mask != rule->mark_mask)
 428                         continue;
 429
 430                 if (rule->tun_id && r->tun_id != rule->tun_id)
 431                         continue;
 432
 433                 if (r->fr_net != rule->fr_net)
 434                         continue;
 435
 436                 if (rule->l3mdev && r->l3mdev != rule->l3mdev)
 437                         continue;
 438
 439                 if (uid_range_set(&rule->uid_range) &&
 440                     (!uid_eq(r->uid_range.start, rule->uid_range.start) ||
 441                     !uid_eq(r->uid_range.end, rule->uid_range.end)))
 442                         continue;
 443
 444                 if (rule->ip_proto && r->ip_proto != rule->ip_proto)
 445                         continue;
 446
 447                 if (fib_rule_port_range_set(&rule->sport_range) &&
 448                     !fib_rule_port_range_compare(&r->sport_range,
 449                                                  &rule->sport_range))
 450                         continue;
 451
 452                 if (fib_rule_port_range_set(&rule->dport_range) &&
 453                     !fib_rule_port_range_compare(&r->dport_range,
 454                                                  &rule->dport_range))
 455                         continue;
 456
 457                 if (!ops->compare(r, frh, tb))
 458                         continue;
 459                 return r;
 460         }
 461
 462         return NULL;
 463 }
 464
 465 #ifdef CONFIG_NET_L3_MASTER_DEV
 466 static int fib_nl2rule_l3mdev(struct nlattr *nla, struct fib_rule *nlrule,
 467                               struct netlink_ext_ack *extack)
 468 {
 469         nlrule->l3mdev = nla_get_u8(nla);
 470         if (nlrule->l3mdev != 1) {
 471                 NL_SET_ERR_MSG(extack, "Invalid l3mdev attribute");
 472                 return -1;
 473         }
 474
 475         return 0;
 476 }
 477 #else
 478 static int fib_nl2rule_l3mdev(struct nlattr *nla, struct fib_rule *nlrule,
 479                               struct netlink_ext_ack *extack)
 480 {
 481         NL_SET_ERR_MSG(extack, "l3mdev support is not enabled in kernel");
 482         return -1;
 483 }
 484 #endif
 485
 486 static int fib_nl2rule(struct sk_buff *skb, struct nlmsghdr *nlh,
 487                        struct netlink_ext_ack *extack,
 488                        struct fib_rules_ops *ops,
 489                        struct nlattr *tb[],
 490                        struct fib_rule **rule,
 491                        bool *user_priority)
 492 {
 493         struct net *net = sock_net(skb->sk);
 494         struct fib_rule_hdr *frh = nlmsg_data(nlh);
 495         struct fib_rule *nlrule = NULL;
 496         int err = -EINVAL;
 497
 498         if (frh->src_len)
 499                 if (!tb[FRA_SRC] ||
 500                     frh->src_len > (ops->addr_size * 8) ||
 501                     nla_len(tb[FRA_SRC]) != ops->addr_size) {
 502                         NL_SET_ERR_MSG(extack, "Invalid source address");
 503                         goto errout;
 504         }
 505
 506         if (frh->dst_len)
 507                 if (!tb[FRA_DST] ||
 508                     frh->dst_len > (ops->addr_size * 8) ||
 509                     nla_len(tb[FRA_DST]) != ops->addr_size) {
 510                         NL_SET_ERR_MSG(extack, "Invalid dst address");
 511                         goto errout;
 512         }
 513
 514         nlrule = kzalloc(ops->rule_size, GFP_KERNEL);
 515         if (!nlrule) {
 516                 err = -ENOMEM;
 517                 goto errout;
 518         }
 519         refcount_set(&nlrule->refcnt, 1);
 520         nlrule->fr_net = net;
 521
 522         if (tb[FRA_PRIORITY]) {
 523                 nlrule->pref = nla_get_u32(tb[FRA_PRIORITY]);
 524                 *user_priority = true;
 525         } else {
 526                 nlrule->pref = fib_default_rule_pref(ops);
 527         }
 528
 529         nlrule->proto = tb[FRA_PROTOCOL] ?
 530                 nla_get_u8(tb[FRA_PROTOCOL]) : RTPROT_UNSPEC;
 531
 532         if (tb[FRA_IIFNAME]) {
 533                 struct net_device *dev;
 534
 535                 nlrule->iifindex = -1;
 536                 nla_strlcpy(nlrule->iifname, tb[FRA_IIFNAME], IFNAMSIZ);
 537                 dev = __dev_get_by_name(net, nlrule->iifname);
 538                 if (dev)
 539                         nlrule->iifindex = dev->ifindex;
 540         }
 541
 542         if (tb[FRA_OIFNAME]) {
 543                 struct net_device *dev;
 544
 545                 nlrule->oifindex = -1;
 546                 nla_strlcpy(nlrule->oifname, tb[FRA_OIFNAME], IFNAMSIZ);
 547                 dev = __dev_get_by_name(net, nlrule->oifname);
 548                 if (dev)
 549                         nlrule->oifindex = dev->ifindex;
 550         }
 551
 552         if (tb[FRA_FWMARK]) {
 553                 nlrule->mark = nla_get_u32(tb[FRA_FWMARK]);
 554                 if (nlrule->mark)
 555                         /* compatibility: if the mark value is non-zero all bits
 556                          * are compared unless a mask is explicitly specified.
 557                          */
 558                         nlrule->mark_mask = 0xFFFFFFFF;
 559         }
 560
 561         if (tb[FRA_FWMASK])
 562                 nlrule->mark_mask = nla_get_u32(tb[FRA_FWMASK]);
 563
 564         if (tb[FRA_TUN_ID])
 565                 nlrule->tun_id = nla_get_be64(tb[FRA_TUN_ID]);
 566
 567         err = -EINVAL;
 568         if (tb[FRA_L3MDEV] &&
 569             fib_nl2rule_l3mdev(tb[FRA_L3MDEV], nlrule, extack) < 0)
 570                 goto errout_free;
 571
 572         nlrule->action = frh->action;
 573         nlrule->flags = frh->flags;
 574         nlrule->table = frh_get_table(frh, tb);
 575         if (tb[FRA_SUPPRESS_PREFIXLEN])
 576                 nlrule->suppress_prefixlen = nla_get_u32(tb[FRA_SUPPRESS_PREFIXLEN]);
 577         else
 578                 nlrule->suppress_prefixlen = -1;
 579
 580         if (tb[FRA_SUPPRESS_IFGROUP])
 581                 nlrule->suppress_ifgroup = nla_get_u32(tb[FRA_SUPPRESS_IFGROUP]);
 582         else
 583                 nlrule->suppress_ifgroup = -1;
 584
 585         if (tb[FRA_GOTO]) {
 586                 if (nlrule->action != FR_ACT_GOTO) {
 587                         NL_SET_ERR_MSG(extack, "Unexpected goto");
 588                         goto errout_free;
 589                 }
 590
 591                 nlrule->target = nla_get_u32(tb[FRA_GOTO]);
 592                 /* Backward jumps are prohibited to avoid endless loops */
 593                 if (nlrule->target <= nlrule->pref) {
 594                         NL_SET_ERR_MSG(extack, "Backward goto not supported");
 595                         goto errout_free;
 596                 }
 597         } else if (nlrule->action == FR_ACT_GOTO) {
 598                 NL_SET_ERR_MSG(extack, "Missing goto target for action goto");
 599                 goto errout_free;
 600         }
 601
 602         if (nlrule->l3mdev && nlrule->table) {
 603                 NL_SET_ERR_MSG(extack, "l3mdev and table are mutually exclusive");
 604                 goto errout_free;
 605         }
 606
 607         if (tb[FRA_UID_RANGE]) {
 608                 if (current_user_ns() != net->user_ns) {
 609                         err = -EPERM;
 610                         NL_SET_ERR_MSG(extack, "No permission to set uid");
 611                         goto errout_free;
 612                 }
 613
 614                 nlrule->uid_range = nla_get_kuid_range(tb);
 615
 616                 if (!uid_range_set(&nlrule->uid_range) ||
 617                     !uid_lte(nlrule->uid_range.start, nlrule->uid_range.end)) {
 618                         NL_SET_ERR_MSG(extack, "Invalid uid range");
 619                         goto errout_free;
 620                 }
 621         } else {
 622                 nlrule->uid_range = fib_kuid_range_unset;
 623         }
 624
 625         if (tb[FRA_IP_PROTO])
 626                 nlrule->ip_proto = nla_get_u8(tb[FRA_IP_PROTO]);
 627
 628         if (tb[FRA_SPORT_RANGE]) {
 629                 err = nla_get_port_range(tb[FRA_SPORT_RANGE],
 630                                          &nlrule->sport_range);
 631                 if (err) {
 632                         NL_SET_ERR_MSG(extack, "Invalid sport range");
 633                         goto errout_free;
 634                 }
 635         }
 636
 637         if (tb[FRA_DPORT_RANGE]) {
 638                 err = nla_get_port_range(tb[FRA_DPORT_RANGE],
 639                                          &nlrule->dport_range);
 640                 if (err) {
 641                         NL_SET_ERR_MSG(extack, "Invalid dport range");
 642                         goto errout_free;
 643                 }
 644         }
 645
 646         *rule = nlrule;
 647
 648         return 0;
 649
 650 errout_free:
 651         kfree(nlrule);
 652 errout:
 653         return err;
 654 }
 655
 656 int fib_nl_newrule(struct sk_buff *skb, struct nlmsghdr *nlh,
 657                    struct netlink_ext_ack *extack)
 658 {
 659         struct net *net = sock_net(skb->sk);
 660         struct fib_rule_hdr *frh = nlmsg_data(nlh);
 661         struct fib_rules_ops *ops = NULL;
 662         struct fib_rule *rule = NULL, *r, *last = NULL;
 663         struct nlattr *tb[FRA_MAX + 1];
 664         int err = -EINVAL, unresolved = 0;
 665         bool user_priority = false;
 666
 667         if (nlh->nlmsg_len < nlmsg_msg_size(sizeof(*frh))) {
 668                 NL_SET_ERR_MSG(extack, "Invalid msg length");
 669                 goto errout;
 670         }
 671
 672         ops = lookup_rules_ops(net, frh->family);
 673         if (!ops) {
 674                 err = -EAFNOSUPPORT;
 675                 NL_SET_ERR_MSG(extack, "Rule family not supported");
 676                 goto errout;
 677         }
 678
 679         err = nlmsg_parse(nlh, sizeof(*frh), tb, FRA_MAX, ops->policy, extack);
 680         if (err < 0) {
 681                 NL_SET_ERR_MSG(extack, "Error parsing msg");
 682                 goto errout;
 683         }
 684
 685         err = fib_nl2rule(skb, nlh, extack, ops, tb, &rule, &user_priority);
 686         if (err)
 687                 goto errout;
 688
 689         if ((nlh->nlmsg_flags & NLM_F_EXCL) &&
 690             rule_find(ops, frh, tb, rule, user_priority)) {
 691                 err = -EEXIST;
 692                 goto errout_free;
 693         }
 694
 695         err = ops->configure(rule, skb, frh, tb, extack);
 696         if (err < 0)
 697                 goto errout_free;
 698
 699         err = call_fib_rule_notifiers(net, FIB_EVENT_RULE_ADD, rule, ops,
 700                                       extack);
 701         if (err < 0)
 702                 goto errout_free;
 703
 704         list_for_each_entry(r, &ops->rules_list, list) {
 705                 if (r->pref == rule->target) {
 706                         RCU_INIT_POINTER(rule->ctarget, r);
 707                         break;
 708                 }
 709         }
 710
 711         if (rcu_dereference_protected(rule->ctarget, 1) == NULL)
 712                 unresolved = 1;
 713
 714         list_for_each_entry(r, &ops->rules_list, list) {
 715                 if (r->pref > rule->pref)
 716                         break;
 717                 last = r;
 718         }
 719
 720         if (last)
 721                 list_add_rcu(&rule->list, &last->list);
 722         else
 723                 list_add_rcu(&rule->list, &ops->rules_list);
 724
 725         if (ops->unresolved_rules) {
 726                 /*
 727                  * There are unresolved goto rules in the list, check if
 728                  * any of them are pointing to this new rule.
 729                  */
 730                 list_for_each_entry(r, &ops->rules_list, list) {
 731                         if (r->action == FR_ACT_GOTO &&
 732                             r->target == rule->pref &&
 733                             rtnl_dereference(r->ctarget) == NULL) {
 734                                 rcu_assign_pointer(r->ctarget, rule);
 735                                 if (--ops->unresolved_rules == 0)
 736                                         break;
 737                         }
 738                 }
 739         }
 740
 741         if (rule->action == FR_ACT_GOTO)
 742                 ops->nr_goto_rules++;
 743
 744         if (unresolved)
 745                 ops->unresolved_rules++;
 746
 747         if (rule->tun_id)
 748                 ip_tunnel_need_metadata();
 749
 750         notify_rule_change(RTM_NEWRULE, rule, ops, nlh, NETLINK_CB(skb).portid);
 751         flush_route_cache(ops);
 752         rules_ops_put(ops);
 753         return 0;
 754
 755 errout_free:
 756         kfree(rule);
 757 errout:
 758         rules_ops_put(ops);
 759         return err;
 760 }
 761 EXPORT_SYMBOL_GPL(fib_nl_newrule);
 762
 763 int fib_nl_delrule(struct sk_buff *skb, struct nlmsghdr *nlh,
 764                    struct netlink_ext_ack *extack)
 765 {
 766         struct net *net = sock_net(skb->sk);
 767         struct fib_rule_hdr *frh = nlmsg_data(nlh);
 768         struct fib_rules_ops *ops = NULL;
 769         struct fib_rule *rule = NULL, *r, *nlrule = NULL;
 770         struct nlattr *tb[FRA_MAX+1];
 771         int err = -EINVAL;
 772         bool user_priority = false;
 773
 774         if (nlh->nlmsg_len < nlmsg_msg_size(sizeof(*frh))) {
 775                 NL_SET_ERR_MSG(extack, "Invalid msg length");
 776                 goto errout;
 777         }
 778
 779         ops = lookup_rules_ops(net, frh->family);
 780         if (ops == NULL) {
 781                 err = -EAFNOSUPPORT;
 782                 NL_SET_ERR_MSG(extack, "Rule family not supported");
 783                 goto errout;
 784         }
 785
 786         err = nlmsg_parse(nlh, sizeof(*frh), tb, FRA_MAX, ops->policy, extack);
 787         if (err < 0) {
 788                 NL_SET_ERR_MSG(extack, "Error parsing msg");
 789                 goto errout;
 790         }
 791
 792         err = fib_nl2rule(skb, nlh, extack, ops, tb, &nlrule, &user_priority);
 793         if (err)
 794                 goto errout;
 795
 796         rule = rule_find(ops, frh, tb, nlrule, user_priority);
 797         if (!rule) {
 798                 err = -ENOENT;
 799                 goto errout;
 800         }
 801
 802         if (rule->flags & FIB_RULE_PERMANENT) {
 803                 err = -EPERM;
 804                 goto errout;
 805         }
 806
 807         if (ops->delete) {
 808                 err = ops->delete(rule);
 809                 if (err)
 810                         goto errout;
 811         }
 812
 813         if (rule->tun_id)
 814                 ip_tunnel_unneed_metadata();
 815
 816         list_del_rcu(&rule->list);
 817
 818         if (rule->action == FR_ACT_GOTO) {
 819                 ops->nr_goto_rules--;
 820                 if (rtnl_dereference(rule->ctarget) == NULL)
 821                         ops->unresolved_rules--;
 822         }
 823
 824         /*
 825          * Check if this rule is a target to any of them. If so,
 826          * adjust to the next one with the same preference or
 827          * disable them. As this operation is eventually very
 828          * expensive, it is only performed if goto rules, except
 829          * current if it is goto rule, have actually been added.
 830          */
 831         if (ops->nr_goto_rules > 0) {
 832                 struct fib_rule *n;
 833
 834                 n = list_next_entry(rule, list);
 835                 if (&n->list == &ops->rules_list || n->pref != rule->pref)
 836                         n = NULL;
 837                 list_for_each_entry(r, &ops->rules_list, list) {
 838                         if (rtnl_dereference(r->ctarget) != rule)
 839                                 continue;
 840                         rcu_assign_pointer(r->ctarget, n);
 841                         if (!n)
 842                                 ops->unresolved_rules++;
 843                 }
 844         }
 845
 846         call_fib_rule_notifiers(net, FIB_EVENT_RULE_DEL, rule, ops,
 847                                 NULL);
 848         notify_rule_change(RTM_DELRULE, rule, ops, nlh,
 849                            NETLINK_CB(skb).portid);
 850         fib_rule_put(rule);
 851         flush_route_cache(ops);
 852         rules_ops_put(ops);
 853         kfree(nlrule);
 854         return 0;
 855
 856 errout:
 857         if (nlrule)
 858                 kfree(nlrule);
 859         rules_ops_put(ops);
 860         return err;
 861 }
 862 EXPORT_SYMBOL_GPL(fib_nl_delrule);
 863
 864 static inline size_t fib_rule_nlmsg_size(struct fib_rules_ops *ops,
 865                                          struct fib_rule *rule)
 866 {
 867         size_t payload = NLMSG_ALIGN(sizeof(struct fib_rule_hdr))
 868                          + nla_total_size(IFNAMSIZ) /* FRA_IIFNAME */
 869                          + nla_total_size(IFNAMSIZ) /* FRA_OIFNAME */
 870                          + nla_total_size(4) /* FRA_PRIORITY */
 871                          + nla_total_size(4) /* FRA_TABLE */
 872                          + nla_total_size(4) /* FRA_SUPPRESS_PREFIXLEN */
 873                          + nla_total_size(4) /* FRA_SUPPRESS_IFGROUP */
 874                          + nla_total_size(4) /* FRA_FWMARK */
 875                          + nla_total_size(4) /* FRA_FWMASK */
 876                          + nla_total_size_64bit(8) /* FRA_TUN_ID */
 877                          + nla_total_size(sizeof(struct fib_kuid_range))
 878                          + nla_total_size(1) /* FRA_PROTOCOL */
 879                          + nla_total_size(1) /* FRA_IP_PROTO */
 880                          + nla_total_size(sizeof(struct fib_rule_port_range)) /* FRA_SPORT_RANGE */
 881                          + nla_total_size(sizeof(struct fib_rule_port_range)); /* FRA_DPORT_RANGE */
 882
 883         if (ops->nlmsg_payload)
 884                 payload += ops->nlmsg_payload(rule);
 885
 886         return payload;
 887 }
 888
 889 static int fib_nl_fill_rule(struct sk_buff *skb, struct fib_rule *rule,
 890                             u32 pid, u32 seq, int type, int flags,
 891                             struct fib_rules_ops *ops)
 892 {
 893         struct nlmsghdr *nlh;
 894         struct fib_rule_hdr *frh;
 895
 896         nlh = nlmsg_put(skb, pid, seq, type, sizeof(*frh), flags);
 897         if (nlh == NULL)
 898                 return -EMSGSIZE;
 899
 900         frh = nlmsg_data(nlh);
 901         frh->family = ops->family;
 902         frh->table = rule->table;
 903         if (nla_put_u32(skb, FRA_TABLE, rule->table))
 904                 goto nla_put_failure;
 905         if (nla_put_u32(skb, FRA_SUPPRESS_PREFIXLEN, rule->suppress_prefixlen))
 906                 goto nla_put_failure;
 907         frh->res1 = 0;
 908         frh->res2 = 0;
 909         frh->action = rule->action;
 910         frh->flags = rule->flags;
 911
 912         if (nla_put_u8(skb, FRA_PROTOCOL, rule->proto))
 913                 goto nla_put_failure;
 914
 915         if (rule->action == FR_ACT_GOTO &&
 916             rcu_access_pointer(rule->ctarget) == NULL)
 917                 frh->flags |= FIB_RULE_UNRESOLVED;
 918
 919         if (rule->iifname[0]) {
 920                 if (nla_put_string(skb, FRA_IIFNAME, rule->iifname))
 921                         goto nla_put_failure;
 922                 if (rule->iifindex == -1)
 923                         frh->flags |= FIB_RULE_IIF_DETACHED;
 924         }
 925
 926         if (rule->oifname[0]) {
 927                 if (nla_put_string(skb, FRA_OIFNAME, rule->oifname))
 928                         goto nla_put_failure;
 929                 if (rule->oifindex == -1)
 930                         frh->flags |= FIB_RULE_OIF_DETACHED;
 931         }
 932
 933         if ((rule->pref &&
 934              nla_put_u32(skb, FRA_PRIORITY, rule->pref)) ||
 935             (rule->mark &&
 936              nla_put_u32(skb, FRA_FWMARK, rule->mark)) ||
 937             ((rule->mark_mask || rule->mark) &&
 938              nla_put_u32(skb, FRA_FWMASK, rule->mark_mask)) ||
 939             (rule->target &&
 940              nla_put_u32(skb, FRA_GOTO, rule->target)) ||
 941             (rule->tun_id &&
 942              nla_put_be64(skb, FRA_TUN_ID, rule->tun_id, FRA_PAD)) ||
 943             (rule->l3mdev &&
 944              nla_put_u8(skb, FRA_L3MDEV, rule->l3mdev)) ||
 945             (uid_range_set(&rule->uid_range) &&
 946              nla_put_uid_range(skb, &rule->uid_range)) ||
 947             (fib_rule_port_range_set(&rule->sport_range) &&
 948              nla_put_port_range(skb, FRA_SPORT_RANGE, &rule->sport_range)) ||
 949             (fib_rule_port_range_set(&rule->dport_range) &&
 950              nla_put_port_range(skb, FRA_DPORT_RANGE, &rule->dport_range)) ||
 951             (rule->ip_proto && nla_put_u8(skb, FRA_IP_PROTO, rule->ip_proto)))
 952                 goto nla_put_failure;
 953
 954         if (rule->suppress_ifgroup != -1) {
 955                 if (nla_put_u32(skb, FRA_SUPPRESS_IFGROUP, rule->suppress_ifgroup))
 956                         goto nla_put_failure;
 957         }
 958
 959         if (ops->fill(rule, skb, frh) < 0)
 960                 goto nla_put_failure;
 961
 962         nlmsg_end(skb, nlh);
 963         return 0;
 964
 965 nla_put_failure:
 966         nlmsg_cancel(skb, nlh);
 967         return -EMSGSIZE;
 968 }
 969
 970 static int dump_rules(struct sk_buff *skb, struct netlink_callback *cb,
 971                       struct fib_rules_ops *ops)
 972 {
 973         int idx = 0;
 974         struct fib_rule *rule;
 975         int err = 0;
 976
 977         rcu_read_lock();
 978         list_for_each_entry_rcu(rule, &ops->rules_list, list) {
 979                 if (idx < cb->args[1])
 980                         goto skip;
 981
 982                 err = fib_nl_fill_rule(skb, rule, NETLINK_CB(cb->skb).portid,
 983                                        cb->nlh->nlmsg_seq, RTM_NEWRULE,
 984                                        NLM_F_MULTI, ops);
 985                 if (err)
 986                         break;
 987 skip:
 988                 idx++;
 989         }
 990         rcu_read_unlock();
 991         cb->args[1] = idx;
 992         rules_ops_put(ops);
 993
 994         return err;
 995 }
 996
 997 static int fib_nl_dumprule(struct sk_buff *skb, struct netlink_callback *cb)
 998 {
 999         struct net *net = sock_net(skb->sk);
1000         struct fib_rules_ops *ops;
1001         int idx = 0, family;
1002
1003         family = rtnl_msg_family(cb->nlh);
1004         if (family != AF_UNSPEC) {
1005                 /* Protocol specific dump request */
1006                 ops = lookup_rules_ops(net, family);
1007                 if (ops == NULL)
1008                         return -EAFNOSUPPORT;
1009
1010                 dump_rules(skb, cb, ops);
1011
1012                 return skb->len;
1013         }
1014
1015         rcu_read_lock();
1016         list_for_each_entry_rcu(ops, &net->rules_ops, list) {
1017                 if (idx < cb->args[0] || !try_module_get(ops->owner))
1018                         goto skip;
1019
1020                 if (dump_rules(skb, cb, ops) < 0)
1021                         break;
1022
1023                 cb->args[1] = 0;
1024 skip:
1025                 idx++;
1026         }
1027         rcu_read_unlock();
1028         cb->args[0] = idx;
1029
1030         return skb->len;
1031 }
1032
1033 static void notify_rule_change(int event, struct fib_rule *rule,
1034                                struct fib_rules_ops *ops, struct nlmsghdr *nlh,
1035                                u32 pid)
1036 {
1037         struct net *net;
1038         struct sk_buff *skb;
1039         int err = -ENOBUFS;
1040
1041         net = ops->fro_net;
1042         skb = nlmsg_new(fib_rule_nlmsg_size(ops, rule), GFP_KERNEL);
1043         if (skb == NULL)
1044                 goto errout;
1045
1046         err = fib_nl_fill_rule(skb, rule, pid, nlh->nlmsg_seq, event, 0, ops);
1047         if (err < 0) {
1048                 /* -EMSGSIZE implies BUG in fib_rule_nlmsg_size() */
1049                 WARN_ON(err == -EMSGSIZE);
1050                 kfree_skb(skb);
1051                 goto errout;
1052         }
1053
1054         rtnl_notify(skb, net, pid, ops->nlgroup, nlh, GFP_KERNEL);
1055         return;
1056 errout:
1057         if (err < 0)
1058                 rtnl_set_sk_err(net, ops->nlgroup, err);
1059 }
1060
1061 static void attach_rules(struct list_head *rules, struct net_device *dev)
1062 {
1063         struct fib_rule *rule;
1064
1065         list_for_each_entry(rule, rules, list) {
1066                 if (rule->iifindex == -1 &&
1067                     strcmp(dev->name, rule->iifname) == 0)
1068                         rule->iifindex = dev->ifindex;
1069                 if (rule->oifindex == -1 &&
1070                     strcmp(dev->name, rule->oifname) == 0)
1071                         rule->oifindex = dev->ifindex;
1072         }
1073 }
1074
1075 static void detach_rules(struct list_head *rules, struct net_device *dev)
1076 {
1077         struct fib_rule *rule;
1078
1079         list_for_each_entry(rule, rules, list) {
1080                 if (rule->iifindex == dev->ifindex)
1081                         rule->iifindex = -1;
1082                 if (rule->oifindex == dev->ifindex)
1083                         rule->oifindex = -1;
1084         }
1085 }
1086
1087
1088 static int fib_rules_event(struct notifier_block *this, unsigned long event,
1089                            void *ptr)
1090 {
1091         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1092         struct net *net = dev_net(dev);
1093         struct fib_rules_ops *ops;
1094
1095         ASSERT_RTNL();
1096
1097         switch (event) {
1098         case NETDEV_REGISTER:
1099                 list_for_each_entry(ops, &net->rules_ops, list)
1100                         attach_rules(&ops->rules_list, dev);
1101                 break;
1102
1103         case NETDEV_CHANGENAME:
1104                 list_for_each_entry(ops, &net->rules_ops, list) {
1105                         detach_rules(&ops->rules_list, dev);
1106                         attach_rules(&ops->rules_list, dev);
1107                 }
1108                 break;
1109
1110         case NETDEV_UNREGISTER:
1111                 list_for_each_entry(ops, &net->rules_ops, list)
1112                         detach_rules(&ops->rules_list, dev);
1113                 break;
1114         }
1115
1116         return NOTIFY_DONE;
1117 }
1118
1119 static struct notifier_block fib_rules_notifier = {
1120         .notifier_call = fib_rules_event,
1121 };
1122
1123 static int __net_init fib_rules_net_init(struct net *net)
1124 {
1125         INIT_LIST_HEAD(&net->rules_ops);
1126         spin_lock_init(&net->rules_mod_lock);
1127         return 0;
1128 }
1129
1130 static void __net_exit fib_rules_net_exit(struct net *net)
1131 {
1132         WARN_ON_ONCE(!list_empty(&net->rules_ops));
1133 }
1134
1135 static struct pernet_operations fib_rules_net_ops = {
1136         .init = fib_rules_net_init,
1137         .exit = fib_rules_net_exit,
1138 };
1139
1140 static int __init fib_rules_init(void)
1141 {
1142         int err;
1143         rtnl_register(PF_UNSPEC, RTM_NEWRULE, fib_nl_newrule, NULL, 0);
1144         rtnl_register(PF_UNSPEC, RTM_DELRULE, fib_nl_delrule, NULL, 0);
1145         rtnl_register(PF_UNSPEC, RTM_GETRULE, NULL, fib_nl_dumprule, 0);
1146
1147         err = register_pernet_subsys(&fib_rules_net_ops);
1148         if (err < 0)
1149                 goto fail;
1150
1151         err = register_netdevice_notifier(&fib_rules_notifier);
1152         if (err < 0)
1153                 goto fail_unregister;
1154
1155         return 0;
1156
1157 fail_unregister:
1158         unregister_pernet_subsys(&fib_rules_net_ops);
1159 fail:
1160         rtnl_unregister(PF_UNSPEC, RTM_NEWRULE);
1161         rtnl_unregister(PF_UNSPEC, RTM_DELRULE);
1162         rtnl_unregister(PF_UNSPEC, RTM_GETRULE);
1163         return err;
1164 }
1165
1166 subsys_initcall(fib_rules_init);