linux/net/core/fib_rules.c
<<
>>
Prefs
   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
  21static const struct fib_kuid_range fib_kuid_range_unset = {
  22        KUIDT_INIT(0),
  23        KUIDT_INIT(~0),
  24};
  25
  26bool 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}
  42EXPORT_SYMBOL_GPL(fib_rule_matchall);
  43
  44int 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}
  70EXPORT_SYMBOL(fib_default_rule_add);
  71
  72static 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
  89static void notify_rule_change(int event, struct fib_rule *rule,
  90                               struct fib_rules_ops *ops, struct nlmsghdr *nlh,
  91                               u32 pid);
  92
  93static 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
 111static void rules_ops_put(struct fib_rules_ops *ops)
 112{
 113        if (ops)
 114                module_put(ops->owner);
 115}
 116
 117static void flush_route_cache(struct fib_rules_ops *ops)
 118{
 119        if (ops->flush_cache)
 120                ops->flush_cache(ops);
 121}
 122
 123static 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;
 146errout:
 147        spin_unlock(&net->rules_mod_lock);
 148
 149        return err;
 150}
 151
 152struct fib_rules_ops *
 153fib_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}
 173EXPORT_SYMBOL_GPL(fib_rules_register);
 174
 175static 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
 187void 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}
 198EXPORT_SYMBOL_GPL(fib_rules_unregister);
 199
 200static int uid_range_set(struct fib_kuid_range *range)
 201{
 202        return uid_valid(range->start) && uid_valid(range->end);
 203}
 204
 205static 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
 218static 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
 228static 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
 242static 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
 248static 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);
 274out:
 275        return (rule->flags & FIB_RULE_INVERT) ? !ret : ret;
 276}
 277
 278int 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) {
 287jumped:
 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;
 320out:
 321        rcu_read_unlock();
 322
 323        return err;
 324}
 325EXPORT_SYMBOL_GPL(fib_rules_lookup);
 326
 327static int call_fib_rule_notifier(struct notifier_block *nb,
 328                                  enum fib_event_type event_type,
 329                                  struct fib_rule *rule, int family,
 330                                  struct netlink_ext_ack *extack)
 331{
 332        struct fib_rule_notifier_info info = {
 333                .info.family = family,
 334                .info.extack = extack,
 335                .rule = rule,
 336        };
 337
 338        return call_fib_notifier(nb, event_type, &info.info);
 339}
 340
 341static int call_fib_rule_notifiers(struct net *net,
 342                                   enum fib_event_type event_type,
 343                                   struct fib_rule *rule,
 344                                   struct fib_rules_ops *ops,
 345                                   struct netlink_ext_ack *extack)
 346{
 347        struct fib_rule_notifier_info info = {
 348                .info.family = ops->family,
 349                .info.extack = extack,
 350                .rule = rule,
 351        };
 352
 353        ops->fib_rules_seq++;
 354        return call_fib_notifiers(net, event_type, &info.info);
 355}
 356
 357/* Called with rcu_read_lock() */
 358int fib_rules_dump(struct net *net, struct notifier_block *nb, int family,
 359                   struct netlink_ext_ack *extack)
 360{
 361        struct fib_rules_ops *ops;
 362        struct fib_rule *rule;
 363        int err = 0;
 364
 365        ops = lookup_rules_ops(net, family);
 366        if (!ops)
 367                return -EAFNOSUPPORT;
 368        list_for_each_entry_rcu(rule, &ops->rules_list, list) {
 369                err = call_fib_rule_notifier(nb, FIB_EVENT_RULE_ADD,
 370                                             rule, family, extack);
 371                if (err)
 372                        break;
 373        }
 374        rules_ops_put(ops);
 375
 376        return err;
 377}
 378EXPORT_SYMBOL_GPL(fib_rules_dump);
 379
 380unsigned int fib_rules_seq_read(struct net *net, int family)
 381{
 382        unsigned int fib_rules_seq;
 383        struct fib_rules_ops *ops;
 384
 385        ASSERT_RTNL();
 386
 387        ops = lookup_rules_ops(net, family);
 388        if (!ops)
 389                return 0;
 390        fib_rules_seq = ops->fib_rules_seq;
 391        rules_ops_put(ops);
 392
 393        return fib_rules_seq;
 394}
 395EXPORT_SYMBOL_GPL(fib_rules_seq_read);
 396
 397static struct fib_rule *rule_find(struct fib_rules_ops *ops,
 398                                  struct fib_rule_hdr *frh,
 399                                  struct nlattr **tb,
 400                                  struct fib_rule *rule,
 401                                  bool user_priority)
 402{
 403        struct fib_rule *r;
 404
 405        list_for_each_entry(r, &ops->rules_list, list) {
 406                if (rule->action && r->action != rule->action)
 407                        continue;
 408
 409                if (rule->table && r->table != rule->table)
 410                        continue;
 411
 412                if (user_priority && r->pref != rule->pref)
 413                        continue;
 414
 415                if (rule->iifname[0] &&
 416                    memcmp(r->iifname, rule->iifname, IFNAMSIZ))
 417                        continue;
 418
 419                if (rule->oifname[0] &&
 420                    memcmp(r->oifname, rule->oifname, IFNAMSIZ))
 421                        continue;
 422
 423                if (rule->mark && r->mark != rule->mark)
 424                        continue;
 425
 426                if (rule->suppress_ifgroup != -1 &&
 427                    r->suppress_ifgroup != rule->suppress_ifgroup)
 428                        continue;
 429
 430                if (rule->suppress_prefixlen != -1 &&
 431                    r->suppress_prefixlen != rule->suppress_prefixlen)
 432                        continue;
 433
 434                if (rule->mark_mask && r->mark_mask != rule->mark_mask)
 435                        continue;
 436
 437                if (rule->tun_id && r->tun_id != rule->tun_id)
 438                        continue;
 439
 440                if (r->fr_net != rule->fr_net)
 441                        continue;
 442
 443                if (rule->l3mdev && r->l3mdev != rule->l3mdev)
 444                        continue;
 445
 446                if (uid_range_set(&rule->uid_range) &&
 447                    (!uid_eq(r->uid_range.start, rule->uid_range.start) ||
 448                    !uid_eq(r->uid_range.end, rule->uid_range.end)))
 449                        continue;
 450
 451                if (rule->ip_proto && r->ip_proto != rule->ip_proto)
 452                        continue;
 453
 454                if (rule->proto && r->proto != rule->proto)
 455                        continue;
 456
 457                if (fib_rule_port_range_set(&rule->sport_range) &&
 458                    !fib_rule_port_range_compare(&r->sport_range,
 459                                                 &rule->sport_range))
 460                        continue;
 461
 462                if (fib_rule_port_range_set(&rule->dport_range) &&
 463                    !fib_rule_port_range_compare(&r->dport_range,
 464                                                 &rule->dport_range))
 465                        continue;
 466
 467                if (!ops->compare(r, frh, tb))
 468                        continue;
 469                return r;
 470        }
 471
 472        return NULL;
 473}
 474
 475#ifdef CONFIG_NET_L3_MASTER_DEV
 476static int fib_nl2rule_l3mdev(struct nlattr *nla, struct fib_rule *nlrule,
 477                              struct netlink_ext_ack *extack)
 478{
 479        nlrule->l3mdev = nla_get_u8(nla);
 480        if (nlrule->l3mdev != 1) {
 481                NL_SET_ERR_MSG(extack, "Invalid l3mdev attribute");
 482                return -1;
 483        }
 484
 485        return 0;
 486}
 487#else
 488static int fib_nl2rule_l3mdev(struct nlattr *nla, struct fib_rule *nlrule,
 489                              struct netlink_ext_ack *extack)
 490{
 491        NL_SET_ERR_MSG(extack, "l3mdev support is not enabled in kernel");
 492        return -1;
 493}
 494#endif
 495
 496static int fib_nl2rule(struct sk_buff *skb, struct nlmsghdr *nlh,
 497                       struct netlink_ext_ack *extack,
 498                       struct fib_rules_ops *ops,
 499                       struct nlattr *tb[],
 500                       struct fib_rule **rule,
 501                       bool *user_priority)
 502{
 503        struct net *net = sock_net(skb->sk);
 504        struct fib_rule_hdr *frh = nlmsg_data(nlh);
 505        struct fib_rule *nlrule = NULL;
 506        int err = -EINVAL;
 507
 508        if (frh->src_len)
 509                if (!tb[FRA_SRC] ||
 510                    frh->src_len > (ops->addr_size * 8) ||
 511                    nla_len(tb[FRA_SRC]) != ops->addr_size) {
 512                        NL_SET_ERR_MSG(extack, "Invalid source address");
 513                        goto errout;
 514        }
 515
 516        if (frh->dst_len)
 517                if (!tb[FRA_DST] ||
 518                    frh->dst_len > (ops->addr_size * 8) ||
 519                    nla_len(tb[FRA_DST]) != ops->addr_size) {
 520                        NL_SET_ERR_MSG(extack, "Invalid dst address");
 521                        goto errout;
 522        }
 523
 524        nlrule = kzalloc(ops->rule_size, GFP_KERNEL);
 525        if (!nlrule) {
 526                err = -ENOMEM;
 527                goto errout;
 528        }
 529        refcount_set(&nlrule->refcnt, 1);
 530        nlrule->fr_net = net;
 531
 532        if (tb[FRA_PRIORITY]) {
 533                nlrule->pref = nla_get_u32(tb[FRA_PRIORITY]);
 534                *user_priority = true;
 535        } else {
 536                nlrule->pref = fib_default_rule_pref(ops);
 537        }
 538
 539        nlrule->proto = tb[FRA_PROTOCOL] ?
 540                nla_get_u8(tb[FRA_PROTOCOL]) : RTPROT_UNSPEC;
 541
 542        if (tb[FRA_IIFNAME]) {
 543                struct net_device *dev;
 544
 545                nlrule->iifindex = -1;
 546                nla_strscpy(nlrule->iifname, tb[FRA_IIFNAME], IFNAMSIZ);
 547                dev = __dev_get_by_name(net, nlrule->iifname);
 548                if (dev)
 549                        nlrule->iifindex = dev->ifindex;
 550        }
 551
 552        if (tb[FRA_OIFNAME]) {
 553                struct net_device *dev;
 554
 555                nlrule->oifindex = -1;
 556                nla_strscpy(nlrule->oifname, tb[FRA_OIFNAME], IFNAMSIZ);
 557                dev = __dev_get_by_name(net, nlrule->oifname);
 558                if (dev)
 559                        nlrule->oifindex = dev->ifindex;
 560        }
 561
 562        if (tb[FRA_FWMARK]) {
 563                nlrule->mark = nla_get_u32(tb[FRA_FWMARK]);
 564                if (nlrule->mark)
 565                        /* compatibility: if the mark value is non-zero all bits
 566                         * are compared unless a mask is explicitly specified.
 567                         */
 568                        nlrule->mark_mask = 0xFFFFFFFF;
 569        }
 570
 571        if (tb[FRA_FWMASK])
 572                nlrule->mark_mask = nla_get_u32(tb[FRA_FWMASK]);
 573
 574        if (tb[FRA_TUN_ID])
 575                nlrule->tun_id = nla_get_be64(tb[FRA_TUN_ID]);
 576
 577        err = -EINVAL;
 578        if (tb[FRA_L3MDEV] &&
 579            fib_nl2rule_l3mdev(tb[FRA_L3MDEV], nlrule, extack) < 0)
 580                goto errout_free;
 581
 582        nlrule->action = frh->action;
 583        nlrule->flags = frh->flags;
 584        nlrule->table = frh_get_table(frh, tb);
 585        if (tb[FRA_SUPPRESS_PREFIXLEN])
 586                nlrule->suppress_prefixlen = nla_get_u32(tb[FRA_SUPPRESS_PREFIXLEN]);
 587        else
 588                nlrule->suppress_prefixlen = -1;
 589
 590        if (tb[FRA_SUPPRESS_IFGROUP])
 591                nlrule->suppress_ifgroup = nla_get_u32(tb[FRA_SUPPRESS_IFGROUP]);
 592        else
 593                nlrule->suppress_ifgroup = -1;
 594
 595        if (tb[FRA_GOTO]) {
 596                if (nlrule->action != FR_ACT_GOTO) {
 597                        NL_SET_ERR_MSG(extack, "Unexpected goto");
 598                        goto errout_free;
 599                }
 600
 601                nlrule->target = nla_get_u32(tb[FRA_GOTO]);
 602                /* Backward jumps are prohibited to avoid endless loops */
 603                if (nlrule->target <= nlrule->pref) {
 604                        NL_SET_ERR_MSG(extack, "Backward goto not supported");
 605                        goto errout_free;
 606                }
 607        } else if (nlrule->action == FR_ACT_GOTO) {
 608                NL_SET_ERR_MSG(extack, "Missing goto target for action goto");
 609                goto errout_free;
 610        }
 611
 612        if (nlrule->l3mdev && nlrule->table) {
 613                NL_SET_ERR_MSG(extack, "l3mdev and table are mutually exclusive");
 614                goto errout_free;
 615        }
 616
 617        if (tb[FRA_UID_RANGE]) {
 618                if (current_user_ns() != net->user_ns) {
 619                        err = -EPERM;
 620                        NL_SET_ERR_MSG(extack, "No permission to set uid");
 621                        goto errout_free;
 622                }
 623
 624                nlrule->uid_range = nla_get_kuid_range(tb);
 625
 626                if (!uid_range_set(&nlrule->uid_range) ||
 627                    !uid_lte(nlrule->uid_range.start, nlrule->uid_range.end)) {
 628                        NL_SET_ERR_MSG(extack, "Invalid uid range");
 629                        goto errout_free;
 630                }
 631        } else {
 632                nlrule->uid_range = fib_kuid_range_unset;
 633        }
 634
 635        if (tb[FRA_IP_PROTO])
 636                nlrule->ip_proto = nla_get_u8(tb[FRA_IP_PROTO]);
 637
 638        if (tb[FRA_SPORT_RANGE]) {
 639                err = nla_get_port_range(tb[FRA_SPORT_RANGE],
 640                                         &nlrule->sport_range);
 641                if (err) {
 642                        NL_SET_ERR_MSG(extack, "Invalid sport range");
 643                        goto errout_free;
 644                }
 645        }
 646
 647        if (tb[FRA_DPORT_RANGE]) {
 648                err = nla_get_port_range(tb[FRA_DPORT_RANGE],
 649                                         &nlrule->dport_range);
 650                if (err) {
 651                        NL_SET_ERR_MSG(extack, "Invalid dport range");
 652                        goto errout_free;
 653                }
 654        }
 655
 656        *rule = nlrule;
 657
 658        return 0;
 659
 660errout_free:
 661        kfree(nlrule);
 662errout:
 663        return err;
 664}
 665
 666static int rule_exists(struct fib_rules_ops *ops, struct fib_rule_hdr *frh,
 667                       struct nlattr **tb, struct fib_rule *rule)
 668{
 669        struct fib_rule *r;
 670
 671        list_for_each_entry(r, &ops->rules_list, list) {
 672                if (r->action != rule->action)
 673                        continue;
 674
 675                if (r->table != rule->table)
 676                        continue;
 677
 678                if (r->pref != rule->pref)
 679                        continue;
 680
 681                if (memcmp(r->iifname, rule->iifname, IFNAMSIZ))
 682                        continue;
 683
 684                if (memcmp(r->oifname, rule->oifname, IFNAMSIZ))
 685                        continue;
 686
 687                if (r->mark != rule->mark)
 688                        continue;
 689
 690                if (r->suppress_ifgroup != rule->suppress_ifgroup)
 691                        continue;
 692
 693                if (r->suppress_prefixlen != rule->suppress_prefixlen)
 694                        continue;
 695
 696                if (r->mark_mask != rule->mark_mask)
 697                        continue;
 698
 699                if (r->tun_id != rule->tun_id)
 700                        continue;
 701
 702                if (r->fr_net != rule->fr_net)
 703                        continue;
 704
 705                if (r->l3mdev != rule->l3mdev)
 706                        continue;
 707
 708                if (!uid_eq(r->uid_range.start, rule->uid_range.start) ||
 709                    !uid_eq(r->uid_range.end, rule->uid_range.end))
 710                        continue;
 711
 712                if (r->ip_proto != rule->ip_proto)
 713                        continue;
 714
 715                if (r->proto != rule->proto)
 716                        continue;
 717
 718                if (!fib_rule_port_range_compare(&r->sport_range,
 719                                                 &rule->sport_range))
 720                        continue;
 721
 722                if (!fib_rule_port_range_compare(&r->dport_range,
 723                                                 &rule->dport_range))
 724                        continue;
 725
 726                if (!ops->compare(r, frh, tb))
 727                        continue;
 728                return 1;
 729        }
 730        return 0;
 731}
 732
 733int fib_nl_newrule(struct sk_buff *skb, struct nlmsghdr *nlh,
 734                   struct netlink_ext_ack *extack)
 735{
 736        struct net *net = sock_net(skb->sk);
 737        struct fib_rule_hdr *frh = nlmsg_data(nlh);
 738        struct fib_rules_ops *ops = NULL;
 739        struct fib_rule *rule = NULL, *r, *last = NULL;
 740        struct nlattr *tb[FRA_MAX + 1];
 741        int err = -EINVAL, unresolved = 0;
 742        bool user_priority = false;
 743
 744        if (nlh->nlmsg_len < nlmsg_msg_size(sizeof(*frh))) {
 745                NL_SET_ERR_MSG(extack, "Invalid msg length");
 746                goto errout;
 747        }
 748
 749        ops = lookup_rules_ops(net, frh->family);
 750        if (!ops) {
 751                err = -EAFNOSUPPORT;
 752                NL_SET_ERR_MSG(extack, "Rule family not supported");
 753                goto errout;
 754        }
 755
 756        err = nlmsg_parse_deprecated(nlh, sizeof(*frh), tb, FRA_MAX,
 757                                     ops->policy, extack);
 758        if (err < 0) {
 759                NL_SET_ERR_MSG(extack, "Error parsing msg");
 760                goto errout;
 761        }
 762
 763        err = fib_nl2rule(skb, nlh, extack, ops, tb, &rule, &user_priority);
 764        if (err)
 765                goto errout;
 766
 767        if ((nlh->nlmsg_flags & NLM_F_EXCL) &&
 768            rule_exists(ops, frh, tb, rule)) {
 769                err = -EEXIST;
 770                goto errout_free;
 771        }
 772
 773        err = ops->configure(rule, skb, frh, tb, extack);
 774        if (err < 0)
 775                goto errout_free;
 776
 777        err = call_fib_rule_notifiers(net, FIB_EVENT_RULE_ADD, rule, ops,
 778                                      extack);
 779        if (err < 0)
 780                goto errout_free;
 781
 782        list_for_each_entry(r, &ops->rules_list, list) {
 783                if (r->pref == rule->target) {
 784                        RCU_INIT_POINTER(rule->ctarget, r);
 785                        break;
 786                }
 787        }
 788
 789        if (rcu_dereference_protected(rule->ctarget, 1) == NULL)
 790                unresolved = 1;
 791
 792        list_for_each_entry(r, &ops->rules_list, list) {
 793                if (r->pref > rule->pref)
 794                        break;
 795                last = r;
 796        }
 797
 798        if (last)
 799                list_add_rcu(&rule->list, &last->list);
 800        else
 801                list_add_rcu(&rule->list, &ops->rules_list);
 802
 803        if (ops->unresolved_rules) {
 804                /*
 805                 * There are unresolved goto rules in the list, check if
 806                 * any of them are pointing to this new rule.
 807                 */
 808                list_for_each_entry(r, &ops->rules_list, list) {
 809                        if (r->action == FR_ACT_GOTO &&
 810                            r->target == rule->pref &&
 811                            rtnl_dereference(r->ctarget) == NULL) {
 812                                rcu_assign_pointer(r->ctarget, rule);
 813                                if (--ops->unresolved_rules == 0)
 814                                        break;
 815                        }
 816                }
 817        }
 818
 819        if (rule->action == FR_ACT_GOTO)
 820                ops->nr_goto_rules++;
 821
 822        if (unresolved)
 823                ops->unresolved_rules++;
 824
 825        if (rule->tun_id)
 826                ip_tunnel_need_metadata();
 827
 828        notify_rule_change(RTM_NEWRULE, rule, ops, nlh, NETLINK_CB(skb).portid);
 829        flush_route_cache(ops);
 830        rules_ops_put(ops);
 831        return 0;
 832
 833errout_free:
 834        kfree(rule);
 835errout:
 836        rules_ops_put(ops);
 837        return err;
 838}
 839EXPORT_SYMBOL_GPL(fib_nl_newrule);
 840
 841int fib_nl_delrule(struct sk_buff *skb, struct nlmsghdr *nlh,
 842                   struct netlink_ext_ack *extack)
 843{
 844        struct net *net = sock_net(skb->sk);
 845        struct fib_rule_hdr *frh = nlmsg_data(nlh);
 846        struct fib_rules_ops *ops = NULL;
 847        struct fib_rule *rule = NULL, *r, *nlrule = NULL;
 848        struct nlattr *tb[FRA_MAX+1];
 849        int err = -EINVAL;
 850        bool user_priority = false;
 851
 852        if (nlh->nlmsg_len < nlmsg_msg_size(sizeof(*frh))) {
 853                NL_SET_ERR_MSG(extack, "Invalid msg length");
 854                goto errout;
 855        }
 856
 857        ops = lookup_rules_ops(net, frh->family);
 858        if (ops == NULL) {
 859                err = -EAFNOSUPPORT;
 860                NL_SET_ERR_MSG(extack, "Rule family not supported");
 861                goto errout;
 862        }
 863
 864        err = nlmsg_parse_deprecated(nlh, sizeof(*frh), tb, FRA_MAX,
 865                                     ops->policy, extack);
 866        if (err < 0) {
 867                NL_SET_ERR_MSG(extack, "Error parsing msg");
 868                goto errout;
 869        }
 870
 871        err = fib_nl2rule(skb, nlh, extack, ops, tb, &nlrule, &user_priority);
 872        if (err)
 873                goto errout;
 874
 875        rule = rule_find(ops, frh, tb, nlrule, user_priority);
 876        if (!rule) {
 877                err = -ENOENT;
 878                goto errout;
 879        }
 880
 881        if (rule->flags & FIB_RULE_PERMANENT) {
 882                err = -EPERM;
 883                goto errout;
 884        }
 885
 886        if (ops->delete) {
 887                err = ops->delete(rule);
 888                if (err)
 889                        goto errout;
 890        }
 891
 892        if (rule->tun_id)
 893                ip_tunnel_unneed_metadata();
 894
 895        list_del_rcu(&rule->list);
 896
 897        if (rule->action == FR_ACT_GOTO) {
 898                ops->nr_goto_rules--;
 899                if (rtnl_dereference(rule->ctarget) == NULL)
 900                        ops->unresolved_rules--;
 901        }
 902
 903        /*
 904         * Check if this rule is a target to any of them. If so,
 905         * adjust to the next one with the same preference or
 906         * disable them. As this operation is eventually very
 907         * expensive, it is only performed if goto rules, except
 908         * current if it is goto rule, have actually been added.
 909         */
 910        if (ops->nr_goto_rules > 0) {
 911                struct fib_rule *n;
 912
 913                n = list_next_entry(rule, list);
 914                if (&n->list == &ops->rules_list || n->pref != rule->pref)
 915                        n = NULL;
 916                list_for_each_entry(r, &ops->rules_list, list) {
 917                        if (rtnl_dereference(r->ctarget) != rule)
 918                                continue;
 919                        rcu_assign_pointer(r->ctarget, n);
 920                        if (!n)
 921                                ops->unresolved_rules++;
 922                }
 923        }
 924
 925        call_fib_rule_notifiers(net, FIB_EVENT_RULE_DEL, rule, ops,
 926                                NULL);
 927        notify_rule_change(RTM_DELRULE, rule, ops, nlh,
 928                           NETLINK_CB(skb).portid);
 929        fib_rule_put(rule);
 930        flush_route_cache(ops);
 931        rules_ops_put(ops);
 932        kfree(nlrule);
 933        return 0;
 934
 935errout:
 936        if (nlrule)
 937                kfree(nlrule);
 938        rules_ops_put(ops);
 939        return err;
 940}
 941EXPORT_SYMBOL_GPL(fib_nl_delrule);
 942
 943static inline size_t fib_rule_nlmsg_size(struct fib_rules_ops *ops,
 944                                         struct fib_rule *rule)
 945{
 946        size_t payload = NLMSG_ALIGN(sizeof(struct fib_rule_hdr))
 947                         + nla_total_size(IFNAMSIZ) /* FRA_IIFNAME */
 948                         + nla_total_size(IFNAMSIZ) /* FRA_OIFNAME */
 949                         + nla_total_size(4) /* FRA_PRIORITY */
 950                         + nla_total_size(4) /* FRA_TABLE */
 951                         + nla_total_size(4) /* FRA_SUPPRESS_PREFIXLEN */
 952                         + nla_total_size(4) /* FRA_SUPPRESS_IFGROUP */
 953                         + nla_total_size(4) /* FRA_FWMARK */
 954                         + nla_total_size(4) /* FRA_FWMASK */
 955                         + nla_total_size_64bit(8) /* FRA_TUN_ID */
 956                         + nla_total_size(sizeof(struct fib_kuid_range))
 957                         + nla_total_size(1) /* FRA_PROTOCOL */
 958                         + nla_total_size(1) /* FRA_IP_PROTO */
 959                         + nla_total_size(sizeof(struct fib_rule_port_range)) /* FRA_SPORT_RANGE */
 960                         + nla_total_size(sizeof(struct fib_rule_port_range)); /* FRA_DPORT_RANGE */
 961
 962        if (ops->nlmsg_payload)
 963                payload += ops->nlmsg_payload(rule);
 964
 965        return payload;
 966}
 967
 968static int fib_nl_fill_rule(struct sk_buff *skb, struct fib_rule *rule,
 969                            u32 pid, u32 seq, int type, int flags,
 970                            struct fib_rules_ops *ops)
 971{
 972        struct nlmsghdr *nlh;
 973        struct fib_rule_hdr *frh;
 974
 975        nlh = nlmsg_put(skb, pid, seq, type, sizeof(*frh), flags);
 976        if (nlh == NULL)
 977                return -EMSGSIZE;
 978
 979        frh = nlmsg_data(nlh);
 980        frh->family = ops->family;
 981        frh->table = rule->table;
 982        if (nla_put_u32(skb, FRA_TABLE, rule->table))
 983                goto nla_put_failure;
 984        if (nla_put_u32(skb, FRA_SUPPRESS_PREFIXLEN, rule->suppress_prefixlen))
 985                goto nla_put_failure;
 986        frh->res1 = 0;
 987        frh->res2 = 0;
 988        frh->action = rule->action;
 989        frh->flags = rule->flags;
 990
 991        if (nla_put_u8(skb, FRA_PROTOCOL, rule->proto))
 992                goto nla_put_failure;
 993
 994        if (rule->action == FR_ACT_GOTO &&
 995            rcu_access_pointer(rule->ctarget) == NULL)
 996                frh->flags |= FIB_RULE_UNRESOLVED;
 997
 998        if (rule->iifname[0]) {
 999                if (nla_put_string(skb, FRA_IIFNAME, rule->iifname))
1000                        goto nla_put_failure;
1001                if (rule->iifindex == -1)
1002                        frh->flags |= FIB_RULE_IIF_DETACHED;
1003        }
1004
1005        if (rule->oifname[0]) {
1006                if (nla_put_string(skb, FRA_OIFNAME, rule->oifname))
1007                        goto nla_put_failure;
1008                if (rule->oifindex == -1)
1009                        frh->flags |= FIB_RULE_OIF_DETACHED;
1010        }
1011
1012        if ((rule->pref &&
1013             nla_put_u32(skb, FRA_PRIORITY, rule->pref)) ||
1014            (rule->mark &&
1015             nla_put_u32(skb, FRA_FWMARK, rule->mark)) ||
1016            ((rule->mark_mask || rule->mark) &&
1017             nla_put_u32(skb, FRA_FWMASK, rule->mark_mask)) ||
1018            (rule->target &&
1019             nla_put_u32(skb, FRA_GOTO, rule->target)) ||
1020            (rule->tun_id &&
1021             nla_put_be64(skb, FRA_TUN_ID, rule->tun_id, FRA_PAD)) ||
1022            (rule->l3mdev &&
1023             nla_put_u8(skb, FRA_L3MDEV, rule->l3mdev)) ||
1024            (uid_range_set(&rule->uid_range) &&
1025             nla_put_uid_range(skb, &rule->uid_range)) ||
1026            (fib_rule_port_range_set(&rule->sport_range) &&
1027             nla_put_port_range(skb, FRA_SPORT_RANGE, &rule->sport_range)) ||
1028            (fib_rule_port_range_set(&rule->dport_range) &&
1029             nla_put_port_range(skb, FRA_DPORT_RANGE, &rule->dport_range)) ||
1030            (rule->ip_proto && nla_put_u8(skb, FRA_IP_PROTO, rule->ip_proto)))
1031                goto nla_put_failure;
1032
1033        if (rule->suppress_ifgroup != -1) {
1034                if (nla_put_u32(skb, FRA_SUPPRESS_IFGROUP, rule->suppress_ifgroup))
1035                        goto nla_put_failure;
1036        }
1037
1038        if (ops->fill(rule, skb, frh) < 0)
1039                goto nla_put_failure;
1040
1041        nlmsg_end(skb, nlh);
1042        return 0;
1043
1044nla_put_failure:
1045        nlmsg_cancel(skb, nlh);
1046        return -EMSGSIZE;
1047}
1048
1049static int dump_rules(struct sk_buff *skb, struct netlink_callback *cb,
1050                      struct fib_rules_ops *ops)
1051{
1052        int idx = 0;
1053        struct fib_rule *rule;
1054        int err = 0;
1055
1056        rcu_read_lock();
1057        list_for_each_entry_rcu(rule, &ops->rules_list, list) {
1058                if (idx < cb->args[1])
1059                        goto skip;
1060
1061                err = fib_nl_fill_rule(skb, rule, NETLINK_CB(cb->skb).portid,
1062                                       cb->nlh->nlmsg_seq, RTM_NEWRULE,
1063                                       NLM_F_MULTI, ops);
1064                if (err)
1065                        break;
1066skip:
1067                idx++;
1068        }
1069        rcu_read_unlock();
1070        cb->args[1] = idx;
1071        rules_ops_put(ops);
1072
1073        return err;
1074}
1075
1076static int fib_valid_dumprule_req(const struct nlmsghdr *nlh,
1077                                   struct netlink_ext_ack *extack)
1078{
1079        struct fib_rule_hdr *frh;
1080
1081        if (nlh->nlmsg_len < nlmsg_msg_size(sizeof(*frh))) {
1082                NL_SET_ERR_MSG(extack, "Invalid header for fib rule dump request");
1083                return -EINVAL;
1084        }
1085
1086        frh = nlmsg_data(nlh);
1087        if (frh->dst_len || frh->src_len || frh->tos || frh->table ||
1088            frh->res1 || frh->res2 || frh->action || frh->flags) {
1089                NL_SET_ERR_MSG(extack,
1090                               "Invalid values in header for fib rule dump request");
1091                return -EINVAL;
1092        }
1093
1094        if (nlmsg_attrlen(nlh, sizeof(*frh))) {
1095                NL_SET_ERR_MSG(extack, "Invalid data after header in fib rule dump request");
1096                return -EINVAL;
1097        }
1098
1099        return 0;
1100}
1101
1102static int fib_nl_dumprule(struct sk_buff *skb, struct netlink_callback *cb)
1103{
1104        const struct nlmsghdr *nlh = cb->nlh;
1105        struct net *net = sock_net(skb->sk);
1106        struct fib_rules_ops *ops;
1107        int idx = 0, family;
1108
1109        if (cb->strict_check) {
1110                int err = fib_valid_dumprule_req(nlh, cb->extack);
1111
1112                if (err < 0)
1113                        return err;
1114        }
1115
1116        family = rtnl_msg_family(nlh);
1117        if (family != AF_UNSPEC) {
1118                /* Protocol specific dump request */
1119                ops = lookup_rules_ops(net, family);
1120                if (ops == NULL)
1121                        return -EAFNOSUPPORT;
1122
1123                dump_rules(skb, cb, ops);
1124
1125                return skb->len;
1126        }
1127
1128        rcu_read_lock();
1129        list_for_each_entry_rcu(ops, &net->rules_ops, list) {
1130                if (idx < cb->args[0] || !try_module_get(ops->owner))
1131                        goto skip;
1132
1133                if (dump_rules(skb, cb, ops) < 0)
1134                        break;
1135
1136                cb->args[1] = 0;
1137skip:
1138                idx++;
1139        }
1140        rcu_read_unlock();
1141        cb->args[0] = idx;
1142
1143        return skb->len;
1144}
1145
1146static void notify_rule_change(int event, struct fib_rule *rule,
1147                               struct fib_rules_ops *ops, struct nlmsghdr *nlh,
1148                               u32 pid)
1149{
1150        struct net *net;
1151        struct sk_buff *skb;
1152        int err = -ENOBUFS;
1153
1154        net = ops->fro_net;
1155        skb = nlmsg_new(fib_rule_nlmsg_size(ops, rule), GFP_KERNEL);
1156        if (skb == NULL)
1157                goto errout;
1158
1159        err = fib_nl_fill_rule(skb, rule, pid, nlh->nlmsg_seq, event, 0, ops);
1160        if (err < 0) {
1161                /* -EMSGSIZE implies BUG in fib_rule_nlmsg_size() */
1162                WARN_ON(err == -EMSGSIZE);
1163                kfree_skb(skb);
1164                goto errout;
1165        }
1166
1167        rtnl_notify(skb, net, pid, ops->nlgroup, nlh, GFP_KERNEL);
1168        return;
1169errout:
1170        if (err < 0)
1171                rtnl_set_sk_err(net, ops->nlgroup, err);
1172}
1173
1174static void attach_rules(struct list_head *rules, struct net_device *dev)
1175{
1176        struct fib_rule *rule;
1177
1178        list_for_each_entry(rule, rules, list) {
1179                if (rule->iifindex == -1 &&
1180                    strcmp(dev->name, rule->iifname) == 0)
1181                        rule->iifindex = dev->ifindex;
1182                if (rule->oifindex == -1 &&
1183                    strcmp(dev->name, rule->oifname) == 0)
1184                        rule->oifindex = dev->ifindex;
1185        }
1186}
1187
1188static void detach_rules(struct list_head *rules, struct net_device *dev)
1189{
1190        struct fib_rule *rule;
1191
1192        list_for_each_entry(rule, rules, list) {
1193                if (rule->iifindex == dev->ifindex)
1194                        rule->iifindex = -1;
1195                if (rule->oifindex == dev->ifindex)
1196                        rule->oifindex = -1;
1197        }
1198}
1199
1200
1201static int fib_rules_event(struct notifier_block *this, unsigned long event,
1202                           void *ptr)
1203{
1204        struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1205        struct net *net = dev_net(dev);
1206        struct fib_rules_ops *ops;
1207
1208        ASSERT_RTNL();
1209
1210        switch (event) {
1211        case NETDEV_REGISTER:
1212                list_for_each_entry(ops, &net->rules_ops, list)
1213                        attach_rules(&ops->rules_list, dev);
1214                break;
1215
1216        case NETDEV_CHANGENAME:
1217                list_for_each_entry(ops, &net->rules_ops, list) {
1218                        detach_rules(&ops->rules_list, dev);
1219                        attach_rules(&ops->rules_list, dev);
1220                }
1221                break;
1222
1223        case NETDEV_UNREGISTER:
1224                list_for_each_entry(ops, &net->rules_ops, list)
1225                        detach_rules(&ops->rules_list, dev);
1226                break;
1227        }
1228
1229        return NOTIFY_DONE;
1230}
1231
1232static struct notifier_block fib_rules_notifier = {
1233        .notifier_call = fib_rules_event,
1234};
1235
1236static int __net_init fib_rules_net_init(struct net *net)
1237{
1238        INIT_LIST_HEAD(&net->rules_ops);
1239        spin_lock_init(&net->rules_mod_lock);
1240        return 0;
1241}
1242
1243static void __net_exit fib_rules_net_exit(struct net *net)
1244{
1245        WARN_ON_ONCE(!list_empty(&net->rules_ops));
1246}
1247
1248static struct pernet_operations fib_rules_net_ops = {
1249        .init = fib_rules_net_init,
1250        .exit = fib_rules_net_exit,
1251};
1252
1253static int __init fib_rules_init(void)
1254{
1255        int err;
1256        rtnl_register(PF_UNSPEC, RTM_NEWRULE, fib_nl_newrule, NULL, 0);
1257        rtnl_register(PF_UNSPEC, RTM_DELRULE, fib_nl_delrule, NULL, 0);
1258        rtnl_register(PF_UNSPEC, RTM_GETRULE, NULL, fib_nl_dumprule, 0);
1259
1260        err = register_pernet_subsys(&fib_rules_net_ops);
1261        if (err < 0)
1262                goto fail;
1263
1264        err = register_netdevice_notifier(&fib_rules_notifier);
1265        if (err < 0)
1266                goto fail_unregister;
1267
1268        return 0;
1269
1270fail_unregister:
1271        unregister_pernet_subsys(&fib_rules_net_ops);
1272fail:
1273        rtnl_unregister(PF_UNSPEC, RTM_NEWRULE);
1274        rtnl_unregister(PF_UNSPEC, RTM_DELRULE);
1275        rtnl_unregister(PF_UNSPEC, RTM_GETRULE);
1276        return err;
1277}
1278
1279subsys_initcall(fib_rules_init);
1280