SAFE public projects git trees. - safe/jmp/linux-2.6/blob - net/ipv6/ip6_fib.c

   1 /*
   2  *      Linux INET6 implementation
   3  *      Forwarding Information Database
   4  *
   5  *      Authors:
   6  *      Pedro Roque             <roque@di.fc.ul.pt>
   7  *
   8  *      This program is free software; you can redistribute it and/or
   9  *      modify it under the terms of the GNU General Public License
  10  *      as published by the Free Software Foundation; either version
  11  *      2 of the License, or (at your option) any later version.
  12  */
  13
  14 /*
  15  *      Changes:
  16  *      Yuji SEKIYA @USAGI:     Support default route on router node;
  17  *                              remove ip6_null_entry from the top of
  18  *                              routing table.
  19  *      Ville Nuorvala:         Fixed routing subtrees.
  20  */
  21 #include <linux/errno.h>
  22 #include <linux/types.h>
  23 #include <linux/net.h>
  24 #include <linux/route.h>
  25 #include <linux/netdevice.h>
  26 #include <linux/in6.h>
  27 #include <linux/init.h>
  28 #include <linux/list.h>
  29
  30 #ifdef  CONFIG_PROC_FS
  31 #include <linux/proc_fs.h>
  32 #endif
  33
  34 #include <net/ipv6.h>
  35 #include <net/ndisc.h>
  36 #include <net/addrconf.h>
  37
  38 #include <net/ip6_fib.h>
  39 #include <net/ip6_route.h>
  40
  41 #define RT6_DEBUG 2
  42
  43 #if RT6_DEBUG >= 3
  44 #define RT6_TRACE(x...) printk(KERN_DEBUG x)
  45 #else
  46 #define RT6_TRACE(x...) do { ; } while (0)
  47 #endif
  48
  49 static struct kmem_cache * fib6_node_kmem __read_mostly;
  50
  51 enum fib_walk_state_t
  52 {
  53 #ifdef CONFIG_IPV6_SUBTREES
  54         FWS_S,
  55 #endif
  56         FWS_L,
  57         FWS_R,
  58         FWS_C,
  59         FWS_U
  60 };
  61
  62 struct fib6_cleaner_t
  63 {
  64         struct fib6_walker_t w;
  65         struct net *net;
  66         int (*func)(struct rt6_info *, void *arg);
  67         void *arg;
  68 };
  69
  70 static DEFINE_RWLOCK(fib6_walker_lock);
  71
  72 #ifdef CONFIG_IPV6_SUBTREES
  73 #define FWS_INIT FWS_S
  74 #else
  75 #define FWS_INIT FWS_L
  76 #endif
  77
  78 static void fib6_prune_clones(struct net *net, struct fib6_node *fn,
  79                               struct rt6_info *rt);
  80 static struct rt6_info *fib6_find_prefix(struct net *net, struct fib6_node *fn);
  81 static struct fib6_node *fib6_repair_tree(struct net *net, struct fib6_node *fn);
  82 static int fib6_walk(struct fib6_walker_t *w);
  83 static int fib6_walk_continue(struct fib6_walker_t *w);
  84
  85 /*
  86  *      A routing update causes an increase of the serial number on the
  87  *      affected subtree. This allows for cached routes to be asynchronously
  88  *      tested when modifications are made to the destination cache as a
  89  *      result of redirects, path MTU changes, etc.
  90  */
  91
  92 static __u32 rt_sernum;
  93
  94 static void fib6_gc_timer_cb(unsigned long arg);
  95
  96 static struct fib6_walker_t fib6_walker_list = {
  97         .prev   = &fib6_walker_list,
  98         .next   = &fib6_walker_list,
  99 };
 100
 101 #define FOR_WALKERS(w) for ((w)=fib6_walker_list.next; (w) != &fib6_walker_list; (w)=(w)->next)
 102
 103 static inline void fib6_walker_link(struct fib6_walker_t *w)
 104 {
 105         write_lock_bh(&fib6_walker_lock);
 106         w->next = fib6_walker_list.next;
 107         w->prev = &fib6_walker_list;
 108         w->next->prev = w;
 109         w->prev->next = w;
 110         write_unlock_bh(&fib6_walker_lock);
 111 }
 112
 113 static inline void fib6_walker_unlink(struct fib6_walker_t *w)
 114 {
 115         write_lock_bh(&fib6_walker_lock);
 116         w->next->prev = w->prev;
 117         w->prev->next = w->next;
 118         w->prev = w->next = w;
 119         write_unlock_bh(&fib6_walker_lock);
 120 }
 121 static __inline__ u32 fib6_new_sernum(void)
 122 {
 123         u32 n = ++rt_sernum;
 124         if ((__s32)n <= 0)
 125                 rt_sernum = n = 1;
 126         return n;
 127 }
 128
 129 /*
 130  *      Auxiliary address test functions for the radix tree.
 131  *
 132  *      These assume a 32bit processor (although it will work on
 133  *      64bit processors)
 134  */
 135
 136 /*
 137  *      test bit
 138  */
 139
 140 static __inline__ __be32 addr_bit_set(void *token, int fn_bit)
 141 {
 142         __be32 *addr = token;
 143
 144         return htonl(1 << ((~fn_bit)&0x1F)) & addr[fn_bit>>5];
 145 }
 146
 147 static __inline__ struct fib6_node * node_alloc(void)
 148 {
 149         struct fib6_node *fn;
 150
 151         fn = kmem_cache_zalloc(fib6_node_kmem, GFP_ATOMIC);
 152
 153         return fn;
 154 }
 155
 156 static __inline__ void node_free(struct fib6_node * fn)
 157 {
 158         kmem_cache_free(fib6_node_kmem, fn);
 159 }
 160
 161 static __inline__ void rt6_release(struct rt6_info *rt)
 162 {
 163         if (atomic_dec_and_test(&rt->rt6i_ref))
 164                 dst_free(&rt->u.dst);
 165 }
 166
 167 static void fib6_link_table(struct net *net, struct fib6_table *tb)
 168 {
 169         unsigned int h;
 170
 171         /*
 172          * Initialize table lock at a single place to give lockdep a key,
 173          * tables aren't visible prior to being linked to the list.
 174          */
 175         rwlock_init(&tb->tb6_lock);
 176
 177         h = tb->tb6_id & (FIB6_TABLE_HASHSZ - 1);
 178
 179         /*
 180          * No protection necessary, this is the only list mutatation
 181          * operation, tables never disappear once they exist.
 182          */
 183         hlist_add_head_rcu(&tb->tb6_hlist, &net->ipv6.fib_table_hash[h]);
 184 }
 185
 186 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
 187
 188 static struct fib6_table *fib6_alloc_table(struct net *net, u32 id)
 189 {
 190         struct fib6_table *table;
 191
 192         table = kzalloc(sizeof(*table), GFP_ATOMIC);
 193         if (table != NULL) {
 194                 table->tb6_id = id;
 195                 table->tb6_root.leaf = net->ipv6.ip6_null_entry;
 196                 table->tb6_root.fn_flags = RTN_ROOT | RTN_TL_ROOT | RTN_RTINFO;
 197         }
 198
 199         return table;
 200 }
 201
 202 struct fib6_table *fib6_new_table(struct net *net, u32 id)
 203 {
 204         struct fib6_table *tb;
 205
 206         if (id == 0)
 207                 id = RT6_TABLE_MAIN;
 208         tb = fib6_get_table(net, id);
 209         if (tb)
 210                 return tb;
 211
 212         tb = fib6_alloc_table(net, id);
 213         if (tb != NULL)
 214                 fib6_link_table(net, tb);
 215
 216         return tb;
 217 }
 218
 219 struct fib6_table *fib6_get_table(struct net *net, u32 id)
 220 {
 221         struct fib6_table *tb;
 222         struct hlist_head *head;
 223         struct hlist_node *node;
 224         unsigned int h;
 225
 226         if (id == 0)
 227                 id = RT6_TABLE_MAIN;
 228         h = id & (FIB6_TABLE_HASHSZ - 1);
 229         rcu_read_lock();
 230         head = &net->ipv6.fib_table_hash[h];
 231         hlist_for_each_entry_rcu(tb, node, head, tb6_hlist) {
 232                 if (tb->tb6_id == id) {
 233                         rcu_read_unlock();
 234                         return tb;
 235                 }
 236         }
 237         rcu_read_unlock();
 238
 239         return NULL;
 240 }
 241
 242 static void __net_init fib6_tables_init(struct net *net)
 243 {
 244         fib6_link_table(net, net->ipv6.fib6_main_tbl);
 245         fib6_link_table(net, net->ipv6.fib6_local_tbl);
 246 }
 247 #else
 248
 249 struct fib6_table *fib6_new_table(struct net *net, u32 id)
 250 {
 251         return fib6_get_table(net, id);
 252 }
 253
 254 struct fib6_table *fib6_get_table(struct net *net, u32 id)
 255 {
 256           return net->ipv6.fib6_main_tbl;
 257 }
 258
 259 struct dst_entry *fib6_rule_lookup(struct net *net, struct flowi *fl,
 260                                    int flags, pol_lookup_t lookup)
 261 {
 262         return (struct dst_entry *) lookup(net, net->ipv6.fib6_main_tbl, fl, flags);
 263 }
 264
 265 static void __net_init fib6_tables_init(struct net *net)
 266 {
 267         fib6_link_table(net, net->ipv6.fib6_main_tbl);
 268 }
 269
 270 #endif
 271
 272 static int fib6_dump_node(struct fib6_walker_t *w)
 273 {
 274         int res;
 275         struct rt6_info *rt;
 276
 277         for (rt = w->leaf; rt; rt = rt->u.dst.rt6_next) {
 278                 res = rt6_dump_route(rt, w->args);
 279                 if (res < 0) {
 280                         /* Frame is full, suspend walking */
 281                         w->leaf = rt;
 282                         return 1;
 283                 }
 284                 WARN_ON(res == 0);
 285         }
 286         w->leaf = NULL;
 287         return 0;
 288 }
 289
 290 static void fib6_dump_end(struct netlink_callback *cb)
 291 {
 292         struct fib6_walker_t *w = (void*)cb->args[2];
 293
 294         if (w) {
 295                 if (cb->args[4]) {
 296                         cb->args[4] = 0;
 297                         fib6_walker_unlink(w);
 298                 }
 299                 cb->args[2] = 0;
 300                 kfree(w);
 301         }
 302         cb->done = (void*)cb->args[3];
 303         cb->args[1] = 3;
 304 }
 305
 306 static int fib6_dump_done(struct netlink_callback *cb)
 307 {
 308         fib6_dump_end(cb);
 309         return cb->done ? cb->done(cb) : 0;
 310 }
 311
 312 static int fib6_dump_table(struct fib6_table *table, struct sk_buff *skb,
 313                            struct netlink_callback *cb)
 314 {
 315         struct fib6_walker_t *w;
 316         int res;
 317
 318         w = (void *)cb->args[2];
 319         w->root = &table->tb6_root;
 320
 321         if (cb->args[4] == 0) {
 322                 w->count = 0;
 323                 w->skip = 0;
 324
 325                 read_lock_bh(&table->tb6_lock);
 326                 res = fib6_walk(w);
 327                 read_unlock_bh(&table->tb6_lock);
 328                 if (res > 0) {
 329                         cb->args[4] = 1;
 330                         cb->args[5] = w->root->fn_sernum;
 331                 }
 332         } else {
 333                 if (cb->args[5] != w->root->fn_sernum) {
 334                         /* Begin at the root if the tree changed */
 335                         cb->args[5] = w->root->fn_sernum;
 336                         w->state = FWS_INIT;
 337                         w->node = w->root;
 338                         w->skip = w->count;
 339                 } else
 340                         w->skip = 0;
 341
 342                 read_lock_bh(&table->tb6_lock);
 343                 res = fib6_walk_continue(w);
 344                 read_unlock_bh(&table->tb6_lock);
 345                 if (res <= 0) {
 346                         fib6_walker_unlink(w);
 347                         cb->args[4] = 0;
 348                 }
 349         }
 350
 351         return res;
 352 }
 353
 354 static int inet6_dump_fib(struct sk_buff *skb, struct netlink_callback *cb)
 355 {
 356         struct net *net = sock_net(skb->sk);
 357         unsigned int h, s_h;
 358         unsigned int e = 0, s_e;
 359         struct rt6_rtnl_dump_arg arg;
 360         struct fib6_walker_t *w;
 361         struct fib6_table *tb;
 362         struct hlist_node *node;
 363         struct hlist_head *head;
 364         int res = 0;
 365
 366         s_h = cb->args[0];
 367         s_e = cb->args[1];
 368
 369         w = (void *)cb->args[2];
 370         if (w == NULL) {
 371                 /* New dump:
 372                  *
 373                  * 1. hook callback destructor.
 374                  */
 375                 cb->args[3] = (long)cb->done;
 376                 cb->done = fib6_dump_done;
 377
 378                 /*
 379                  * 2. allocate and initialize walker.
 380                  */
 381                 w = kzalloc(sizeof(*w), GFP_ATOMIC);
 382                 if (w == NULL)
 383                         return -ENOMEM;
 384                 w->func = fib6_dump_node;
 385                 cb->args[2] = (long)w;
 386         }
 387
 388         arg.skb = skb;
 389         arg.cb = cb;
 390         arg.net = net;
 391         w->args = &arg;
 392
 393         for (h = s_h; h < FIB6_TABLE_HASHSZ; h++, s_e = 0) {
 394                 e = 0;
 395                 head = &net->ipv6.fib_table_hash[h];
 396                 hlist_for_each_entry(tb, node, head, tb6_hlist) {
 397                         if (e < s_e)
 398                                 goto next;
 399                         res = fib6_dump_table(tb, skb, cb);
 400                         if (res != 0)
 401                                 goto out;
 402 next:
 403                         e++;
 404                 }
 405         }
 406 out:
 407         cb->args[1] = e;
 408         cb->args[0] = h;
 409
 410         res = res < 0 ? res : skb->len;
 411         if (res <= 0)
 412                 fib6_dump_end(cb);
 413         return res;
 414 }
 415
 416 /*
 417  *      Routing Table
 418  *
 419  *      return the appropriate node for a routing tree "add" operation
 420  *      by either creating and inserting or by returning an existing
 421  *      node.
 422  */
 423
 424 static struct fib6_node * fib6_add_1(struct fib6_node *root, void *addr,
 425                                      int addrlen, int plen,
 426                                      int offset)
 427 {
 428         struct fib6_node *fn, *in, *ln;
 429         struct fib6_node *pn = NULL;
 430         struct rt6key *key;
 431         int     bit;
 432         __be32  dir = 0;
 433         __u32   sernum = fib6_new_sernum();
 434
 435         RT6_TRACE("fib6_add_1\n");
 436
 437         /* insert node in tree */
 438
 439         fn = root;
 440
 441         do {
 442                 key = (struct rt6key *)((u8 *)fn->leaf + offset);
 443
 444                 /*
 445                  *      Prefix match
 446                  */
 447                 if (plen < fn->fn_bit ||
 448                     !ipv6_prefix_equal(&key->addr, addr, fn->fn_bit))
 449                         goto insert_above;
 450
 451                 /*
 452                  *      Exact match ?
 453                  */
 454
 455                 if (plen == fn->fn_bit) {
 456                         /* clean up an intermediate node */
 457                         if ((fn->fn_flags & RTN_RTINFO) == 0) {
 458                                 rt6_release(fn->leaf);
 459                                 fn->leaf = NULL;
 460                         }
 461
 462                         fn->fn_sernum = sernum;
 463
 464                         return fn;
 465                 }
 466
 467                 /*
 468                  *      We have more bits to go
 469                  */
 470
 471                 /* Try to walk down on tree. */
 472                 fn->fn_sernum = sernum;
 473                 dir = addr_bit_set(addr, fn->fn_bit);
 474                 pn = fn;
 475                 fn = dir ? fn->right: fn->left;
 476         } while (fn);
 477
 478         /*
 479          *      We walked to the bottom of tree.
 480          *      Create new leaf node without children.
 481          */
 482
 483         ln = node_alloc();
 484
 485         if (ln == NULL)
 486                 return NULL;
 487         ln->fn_bit = plen;
 488
 489         ln->parent = pn;
 490         ln->fn_sernum = sernum;
 491
 492         if (dir)
 493                 pn->right = ln;
 494         else
 495                 pn->left  = ln;
 496
 497         return ln;
 498
 499
 500 insert_above:
 501         /*
 502          * split since we don't have a common prefix anymore or
 503          * we have a less significant route.
 504          * we've to insert an intermediate node on the list
 505          * this new node will point to the one we need to create
 506          * and the current
 507          */
 508
 509         pn = fn->parent;
 510
 511         /* find 1st bit in difference between the 2 addrs.
 512
 513            See comment in __ipv6_addr_diff: bit may be an invalid value,
 514            but if it is >= plen, the value is ignored in any case.
 515          */
 516
 517         bit = __ipv6_addr_diff(addr, &key->addr, addrlen);
 518
 519         /*
 520          *              (intermediate)[in]
 521          *                /        \
 522          *      (new leaf node)[ln] (old node)[fn]
 523          */
 524         if (plen > bit) {
 525                 in = node_alloc();
 526                 ln = node_alloc();
 527
 528                 if (in == NULL || ln == NULL) {
 529                         if (in)
 530                                 node_free(in);
 531                         if (ln)
 532                                 node_free(ln);
 533                         return NULL;
 534                 }
 535
 536                 /*
 537                  * new intermediate node.
 538                  * RTN_RTINFO will
 539                  * be off since that an address that chooses one of
 540                  * the branches would not match less specific routes
 541                  * in the other branch
 542                  */
 543
 544                 in->fn_bit = bit;
 545
 546                 in->parent = pn;
 547                 in->leaf = fn->leaf;
 548                 atomic_inc(&in->leaf->rt6i_ref);
 549
 550                 in->fn_sernum = sernum;
 551
 552                 /* update parent pointer */
 553                 if (dir)
 554                         pn->right = in;
 555                 else
 556                         pn->left  = in;
 557
 558                 ln->fn_bit = plen;
 559
 560                 ln->parent = in;
 561                 fn->parent = in;
 562
 563                 ln->fn_sernum = sernum;
 564
 565                 if (addr_bit_set(addr, bit)) {
 566                         in->right = ln;
 567                         in->left  = fn;
 568                 } else {
 569                         in->left  = ln;
 570                         in->right = fn;
 571                 }
 572         } else { /* plen <= bit */
 573
 574                 /*
 575                  *              (new leaf node)[ln]
 576                  *                /        \
 577                  *           (old node)[fn] NULL
 578                  */
 579
 580                 ln = node_alloc();
 581
 582                 if (ln == NULL)
 583                         return NULL;
 584
 585                 ln->fn_bit = plen;
 586
 587                 ln->parent = pn;
 588
 589                 ln->fn_sernum = sernum;
 590
 591                 if (dir)
 592                         pn->right = ln;
 593                 else
 594                         pn->left  = ln;
 595
 596                 if (addr_bit_set(&key->addr, plen))
 597                         ln->right = fn;
 598                 else
 599                         ln->left  = fn;
 600
 601                 fn->parent = ln;
 602         }
 603         return ln;
 604 }
 605
 606 /*
 607  *      Insert routing information in a node.
 608  */
 609
 610 static int fib6_add_rt2node(struct fib6_node *fn, struct rt6_info *rt,
 611                             struct nl_info *info)
 612 {
 613         struct rt6_info *iter = NULL;
 614         struct rt6_info **ins;
 615
 616         ins = &fn->leaf;
 617
 618         for (iter = fn->leaf; iter; iter=iter->u.dst.rt6_next) {
 619                 /*
 620                  *      Search for duplicates
 621                  */
 622
 623                 if (iter->rt6i_metric == rt->rt6i_metric) {
 624                         /*
 625                          *      Same priority level
 626                          */
 627
 628                         if (iter->rt6i_dev == rt->rt6i_dev &&
 629                             iter->rt6i_idev == rt->rt6i_idev &&
 630                             ipv6_addr_equal(&iter->rt6i_gateway,
 631                                             &rt->rt6i_gateway)) {
 632                                 if (!(iter->rt6i_flags&RTF_EXPIRES))
 633                                         return -EEXIST;
 634                                 iter->rt6i_expires = rt->rt6i_expires;
 635                                 if (!(rt->rt6i_flags&RTF_EXPIRES)) {
 636                                         iter->rt6i_flags &= ~RTF_EXPIRES;
 637                                         iter->rt6i_expires = 0;
 638                                 }
 639                                 return -EEXIST;
 640                         }
 641                 }
 642
 643                 if (iter->rt6i_metric > rt->rt6i_metric)
 644                         break;
 645
 646                 ins = &iter->u.dst.rt6_next;
 647         }
 648
 649         /* Reset round-robin state, if necessary */
 650         if (ins == &fn->leaf)
 651                 fn->rr_ptr = NULL;
 652
 653         /*
 654          *      insert node
 655          */
 656
 657         rt->u.dst.rt6_next = iter;
 658         *ins = rt;
 659         rt->rt6i_node = fn;
 660         atomic_inc(&rt->rt6i_ref);
 661         inet6_rt_notify(RTM_NEWROUTE, rt, info);
 662         info->nl_net->ipv6.rt6_stats->fib_rt_entries++;
 663
 664         if ((fn->fn_flags & RTN_RTINFO) == 0) {
 665                 info->nl_net->ipv6.rt6_stats->fib_route_nodes++;
 666                 fn->fn_flags |= RTN_RTINFO;
 667         }
 668
 669         return 0;
 670 }
 671
 672 static __inline__ void fib6_start_gc(struct net *net, struct rt6_info *rt)
 673 {
 674         if (!timer_pending(&net->ipv6.ip6_fib_timer) &&
 675             (rt->rt6i_flags & (RTF_EXPIRES|RTF_CACHE)))
 676                 mod_timer(&net->ipv6.ip6_fib_timer,
 677                           jiffies + net->ipv6.sysctl.ip6_rt_gc_interval);
 678 }
 679
 680 void fib6_force_start_gc(struct net *net)
 681 {
 682         if (!timer_pending(&net->ipv6.ip6_fib_timer))
 683                 mod_timer(&net->ipv6.ip6_fib_timer,
 684                           jiffies + net->ipv6.sysctl.ip6_rt_gc_interval);
 685 }
 686
 687 /*
 688  *      Add routing information to the routing tree.
 689  *      <destination addr>/<source addr>
 690  *      with source addr info in sub-trees
 691  */
 692
 693 int fib6_add(struct fib6_node *root, struct rt6_info *rt, struct nl_info *info)
 694 {
 695         struct fib6_node *fn, *pn = NULL;
 696         int err = -ENOMEM;
 697
 698         fn = fib6_add_1(root, &rt->rt6i_dst.addr, sizeof(struct in6_addr),
 699                         rt->rt6i_dst.plen, offsetof(struct rt6_info, rt6i_dst));
 700
 701         if (fn == NULL)
 702                 goto out;
 703
 704         pn = fn;
 705
 706 #ifdef CONFIG_IPV6_SUBTREES
 707         if (rt->rt6i_src.plen) {
 708                 struct fib6_node *sn;
 709
 710                 if (fn->subtree == NULL) {
 711                         struct fib6_node *sfn;
 712
 713                         /*
 714                          * Create subtree.
 715                          *
 716                          *              fn[main tree]
 717                          *              |
 718                          *              sfn[subtree root]
 719                          *                 \
 720                          *                  sn[new leaf node]
 721                          */
 722
 723                         /* Create subtree root node */
 724                         sfn = node_alloc();
 725                         if (sfn == NULL)
 726                                 goto st_failure;
 727
 728                         sfn->leaf = info->nl_net->ipv6.ip6_null_entry;
 729                         atomic_inc(&info->nl_net->ipv6.ip6_null_entry->rt6i_ref);
 730                         sfn->fn_flags = RTN_ROOT;
 731                         sfn->fn_sernum = fib6_new_sernum();
 732
 733                         /* Now add the first leaf node to new subtree */
 734
 735                         sn = fib6_add_1(sfn, &rt->rt6i_src.addr,
 736                                         sizeof(struct in6_addr), rt->rt6i_src.plen,
 737                                         offsetof(struct rt6_info, rt6i_src));
 738
 739                         if (sn == NULL) {
 740                                 /* If it is failed, discard just allocated
 741                                    root, and then (in st_failure) stale node
 742                                    in main tree.
 743                                  */
 744                                 node_free(sfn);
 745                                 goto st_failure;
 746                         }
 747
 748                         /* Now link new subtree to main tree */
 749                         sfn->parent = fn;
 750                         fn->subtree = sfn;
 751                 } else {
 752                         sn = fib6_add_1(fn->subtree, &rt->rt6i_src.addr,
 753                                         sizeof(struct in6_addr), rt->rt6i_src.plen,
 754                                         offsetof(struct rt6_info, rt6i_src));
 755
 756                         if (sn == NULL)
 757                                 goto st_failure;
 758                 }
 759
 760                 if (fn->leaf == NULL) {
 761                         fn->leaf = rt;
 762                         atomic_inc(&rt->rt6i_ref);
 763                 }
 764                 fn = sn;
 765         }
 766 #endif
 767
 768         err = fib6_add_rt2node(fn, rt, info);
 769
 770         if (err == 0) {
 771                 fib6_start_gc(info->nl_net, rt);
 772                 if (!(rt->rt6i_flags&RTF_CACHE))
 773                         fib6_prune_clones(info->nl_net, pn, rt);
 774         }
 775
 776 out:
 777         if (err) {
 778 #ifdef CONFIG_IPV6_SUBTREES
 779                 /*
 780                  * If fib6_add_1 has cleared the old leaf pointer in the
 781                  * super-tree leaf node we have to find a new one for it.
 782                  */
 783                 if (pn != fn && pn->leaf == rt) {
 784                         pn->leaf = NULL;
 785                         atomic_dec(&rt->rt6i_ref);
 786                 }
 787                 if (pn != fn && !pn->leaf && !(pn->fn_flags & RTN_RTINFO)) {
 788                         pn->leaf = fib6_find_prefix(info->nl_net, pn);
 789 #if RT6_DEBUG >= 2
 790                         if (!pn->leaf) {
 791                                 WARN_ON(pn->leaf == NULL);
 792                                 pn->leaf = info->nl_net->ipv6.ip6_null_entry;
 793                         }
 794 #endif
 795                         atomic_inc(&pn->leaf->rt6i_ref);
 796                 }
 797 #endif
 798                 dst_free(&rt->u.dst);
 799         }
 800         return err;
 801
 802 #ifdef CONFIG_IPV6_SUBTREES
 803         /* Subtree creation failed, probably main tree node
 804            is orphan. If it is, shoot it.
 805          */
 806 st_failure:
 807         if (fn && !(fn->fn_flags & (RTN_RTINFO|RTN_ROOT)))
 808                 fib6_repair_tree(info->nl_net, fn);
 809         dst_free(&rt->u.dst);
 810         return err;
 811 #endif
 812 }
 813
 814 /*
 815  *      Routing tree lookup
 816  *
 817  */
 818
 819 struct lookup_args {
 820         int             offset;         /* key offset on rt6_info       */
 821         struct in6_addr *addr;          /* search key                   */
 822 };
 823
 824 static struct fib6_node * fib6_lookup_1(struct fib6_node *root,
 825                                         struct lookup_args *args)
 826 {
 827         struct fib6_node *fn;
 828         __be32 dir;
 829
 830         if (unlikely(args->offset == 0))
 831                 return NULL;
 832
 833         /*
 834          *      Descend on a tree
 835          */
 836
 837         fn = root;
 838
 839         for (;;) {
 840                 struct fib6_node *next;
 841
 842                 dir = addr_bit_set(args->addr, fn->fn_bit);
 843
 844                 next = dir ? fn->right : fn->left;
 845
 846                 if (next) {
 847                         fn = next;
 848                         continue;
 849                 }
 850
 851                 break;
 852         }
 853
 854         while(fn) {
 855                 if (FIB6_SUBTREE(fn) || fn->fn_flags & RTN_RTINFO) {
 856                         struct rt6key *key;
 857
 858                         key = (struct rt6key *) ((u8 *) fn->leaf +
 859                                                  args->offset);
 860
 861                         if (ipv6_prefix_equal(&key->addr, args->addr, key->plen)) {
 862 #ifdef CONFIG_IPV6_SUBTREES
 863                                 if (fn->subtree)
 864                                         fn = fib6_lookup_1(fn->subtree, args + 1);
 865 #endif
 866                                 if (!fn || fn->fn_flags & RTN_RTINFO)
 867                                         return fn;
 868                         }
 869                 }
 870
 871                 if (fn->fn_flags & RTN_ROOT)
 872                         break;
 873
 874                 fn = fn->parent;
 875         }
 876
 877         return NULL;
 878 }
 879
 880 struct fib6_node * fib6_lookup(struct fib6_node *root, struct in6_addr *daddr,
 881                                struct in6_addr *saddr)
 882 {
 883         struct fib6_node *fn;
 884         struct lookup_args args[] = {
 885                 {
 886                         .offset = offsetof(struct rt6_info, rt6i_dst),
 887                         .addr = daddr,
 888                 },
 889 #ifdef CONFIG_IPV6_SUBTREES
 890                 {
 891                         .offset = offsetof(struct rt6_info, rt6i_src),
 892                         .addr = saddr,
 893                 },
 894 #endif
 895                 {
 896                         .offset = 0,    /* sentinel */
 897                 }
 898         };
 899
 900         fn = fib6_lookup_1(root, daddr ? args : args + 1);
 901
 902         if (fn == NULL || fn->fn_flags & RTN_TL_ROOT)
 903                 fn = root;
 904
 905         return fn;
 906 }
 907
 908 /*
 909  *      Get node with specified destination prefix (and source prefix,
 910  *      if subtrees are used)
 911  */
 912
 913
 914 static struct fib6_node * fib6_locate_1(struct fib6_node *root,
 915                                         struct in6_addr *addr,
 916                                         int plen, int offset)
 917 {
 918         struct fib6_node *fn;
 919
 920         for (fn = root; fn ; ) {
 921                 struct rt6key *key = (struct rt6key *)((u8 *)fn->leaf + offset);
 922
 923                 /*
 924                  *      Prefix match
 925                  */
 926                 if (plen < fn->fn_bit ||
 927                     !ipv6_prefix_equal(&key->addr, addr, fn->fn_bit))
 928                         return NULL;
 929
 930                 if (plen == fn->fn_bit)
 931                         return fn;
 932
 933                 /*
 934                  *      We have more bits to go
 935                  */
 936                 if (addr_bit_set(addr, fn->fn_bit))
 937                         fn = fn->right;
 938                 else
 939                         fn = fn->left;
 940         }
 941         return NULL;
 942 }
 943
 944 struct fib6_node * fib6_locate(struct fib6_node *root,
 945                                struct in6_addr *daddr, int dst_len,
 946                                struct in6_addr *saddr, int src_len)
 947 {
 948         struct fib6_node *fn;
 949
 950         fn = fib6_locate_1(root, daddr, dst_len,
 951                            offsetof(struct rt6_info, rt6i_dst));
 952
 953 #ifdef CONFIG_IPV6_SUBTREES
 954         if (src_len) {
 955                 WARN_ON(saddr == NULL);
 956                 if (fn && fn->subtree)
 957                         fn = fib6_locate_1(fn->subtree, saddr, src_len,
 958                                            offsetof(struct rt6_info, rt6i_src));
 959         }
 960 #endif
 961
 962         if (fn && fn->fn_flags&RTN_RTINFO)
 963                 return fn;
 964
 965         return NULL;
 966 }
 967
 968
 969 /*
 970  *      Deletion
 971  *
 972  */
 973
 974 static struct rt6_info *fib6_find_prefix(struct net *net, struct fib6_node *fn)
 975 {
 976         if (fn->fn_flags&RTN_ROOT)
 977                 return net->ipv6.ip6_null_entry;
 978
 979         while(fn) {
 980                 if(fn->left)
 981                         return fn->left->leaf;
 982
 983                 if(fn->right)
 984                         return fn->right->leaf;
 985
 986                 fn = FIB6_SUBTREE(fn);
 987         }
 988         return NULL;
 989 }
 990
 991 /*
 992  *      Called to trim the tree of intermediate nodes when possible. "fn"
 993  *      is the node we want to try and remove.
 994  */
 995
 996 static struct fib6_node *fib6_repair_tree(struct net *net,
 997                                            struct fib6_node *fn)
 998 {
 999         int children;
1000         int nstate;
1001         struct fib6_node *child, *pn;
1002         struct fib6_walker_t *w;
1003         int iter = 0;
1004
1005         for (;;) {
1006                 RT6_TRACE("fixing tree: plen=%d iter=%d\n", fn->fn_bit, iter);
1007                 iter++;
1008
1009                 WARN_ON(fn->fn_flags & RTN_RTINFO);
1010                 WARN_ON(fn->fn_flags & RTN_TL_ROOT);
1011                 WARN_ON(fn->leaf != NULL);
1012
1013                 children = 0;
1014                 child = NULL;
1015                 if (fn->right) child = fn->right, children |= 1;
1016                 if (fn->left) child = fn->left, children |= 2;
1017
1018                 if (children == 3 || FIB6_SUBTREE(fn)
1019 #ifdef CONFIG_IPV6_SUBTREES
1020                     /* Subtree root (i.e. fn) may have one child */
1021                     || (children && fn->fn_flags&RTN_ROOT)
1022 #endif
1023                     ) {
1024                         fn->leaf = fib6_find_prefix(net, fn);
1025 #if RT6_DEBUG >= 2
1026                         if (fn->leaf==NULL) {
1027                                 WARN_ON(!fn->leaf);
1028                                 fn->leaf = net->ipv6.ip6_null_entry;
1029                         }
1030 #endif
1031                         atomic_inc(&fn->leaf->rt6i_ref);
1032                         return fn->parent;
1033                 }
1034
1035                 pn = fn->parent;
1036 #ifdef CONFIG_IPV6_SUBTREES
1037                 if (FIB6_SUBTREE(pn) == fn) {
1038                         WARN_ON(!(fn->fn_flags & RTN_ROOT));
1039                         FIB6_SUBTREE(pn) = NULL;
1040                         nstate = FWS_L;
1041                 } else {
1042                         WARN_ON(fn->fn_flags & RTN_ROOT);
1043 #endif
1044                         if (pn->right == fn) pn->right = child;
1045                         else if (pn->left == fn) pn->left = child;
1046 #if RT6_DEBUG >= 2
1047                         else
1048                                 WARN_ON(1);
1049 #endif
1050                         if (child)
1051                                 child->parent = pn;
1052                         nstate = FWS_R;
1053 #ifdef CONFIG_IPV6_SUBTREES
1054                 }
1055 #endif
1056
1057                 read_lock(&fib6_walker_lock);
1058                 FOR_WALKERS(w) {
1059                         if (child == NULL) {
1060                                 if (w->root == fn) {
1061                                         w->root = w->node = NULL;
1062                                         RT6_TRACE("W %p adjusted by delroot 1\n", w);
1063                                 } else if (w->node == fn) {
1064                                         RT6_TRACE("W %p adjusted by delnode 1, s=%d/%d\n", w, w->state, nstate);
1065                                         w->node = pn;
1066                                         w->state = nstate;
1067                                 }
1068                         } else {
1069                                 if (w->root == fn) {
1070                                         w->root = child;
1071                                         RT6_TRACE("W %p adjusted by delroot 2\n", w);
1072                                 }
1073                                 if (w->node == fn) {
1074                                         w->node = child;
1075                                         if (children&2) {
1076                                                 RT6_TRACE("W %p adjusted by delnode 2, s=%d\n", w, w->state);
1077                                                 w->state = w->state>=FWS_R ? FWS_U : FWS_INIT;
1078                                         } else {
1079                                                 RT6_TRACE("W %p adjusted by delnode 2, s=%d\n", w, w->state);
1080                                                 w->state = w->state>=FWS_C ? FWS_U : FWS_INIT;
1081                                         }
1082                                 }
1083                         }
1084                 }
1085                 read_unlock(&fib6_walker_lock);
1086
1087                 node_free(fn);
1088                 if (pn->fn_flags&RTN_RTINFO || FIB6_SUBTREE(pn))
1089                         return pn;
1090
1091                 rt6_release(pn->leaf);
1092                 pn->leaf = NULL;
1093                 fn = pn;
1094         }
1095 }
1096
1097 static void fib6_del_route(struct fib6_node *fn, struct rt6_info **rtp,
1098                            struct nl_info *info)
1099 {
1100         struct fib6_walker_t *w;
1101         struct rt6_info *rt = *rtp;
1102         struct net *net = info->nl_net;
1103
1104         RT6_TRACE("fib6_del_route\n");
1105
1106         /* Unlink it */
1107         *rtp = rt->u.dst.rt6_next;
1108         rt->rt6i_node = NULL;
1109         net->ipv6.rt6_stats->fib_rt_entries--;
1110         net->ipv6.rt6_stats->fib_discarded_routes++;
1111
1112         /* Reset round-robin state, if necessary */
1113         if (fn->rr_ptr == rt)
1114                 fn->rr_ptr = NULL;
1115
1116         /* Adjust walkers */
1117         read_lock(&fib6_walker_lock);
1118         FOR_WALKERS(w) {
1119                 if (w->state == FWS_C && w->leaf == rt) {
1120                         RT6_TRACE("walker %p adjusted by delroute\n", w);
1121                         w->leaf = rt->u.dst.rt6_next;
1122                         if (w->leaf == NULL)
1123                                 w->state = FWS_U;
1124                 }
1125         }
1126         read_unlock(&fib6_walker_lock);
1127
1128         rt->u.dst.rt6_next = NULL;
1129
1130         /* If it was last route, expunge its radix tree node */
1131         if (fn->leaf == NULL) {
1132                 fn->fn_flags &= ~RTN_RTINFO;
1133                 net->ipv6.rt6_stats->fib_route_nodes--;
1134                 fn = fib6_repair_tree(net, fn);
1135         }
1136
1137         if (atomic_read(&rt->rt6i_ref) != 1) {
1138                 /* This route is used as dummy address holder in some split
1139                  * nodes. It is not leaked, but it still holds other resources,
1140                  * which must be released in time. So, scan ascendant nodes
1141                  * and replace dummy references to this route with references
1142                  * to still alive ones.
1143                  */
1144                 while (fn) {
1145                         if (!(fn->fn_flags&RTN_RTINFO) && fn->leaf == rt) {
1146                                 fn->leaf = fib6_find_prefix(net, fn);
1147                                 atomic_inc(&fn->leaf->rt6i_ref);
1148                                 rt6_release(rt);
1149                         }
1150                         fn = fn->parent;
1151                 }
1152                 /* No more references are possible at this point. */
1153                 BUG_ON(atomic_read(&rt->rt6i_ref) != 1);
1154         }
1155
1156         inet6_rt_notify(RTM_DELROUTE, rt, info);
1157         rt6_release(rt);
1158 }
1159
1160 int fib6_del(struct rt6_info *rt, struct nl_info *info)
1161 {
1162         struct net *net = info->nl_net;
1163         struct fib6_node *fn = rt->rt6i_node;
1164         struct rt6_info **rtp;
1165
1166 #if RT6_DEBUG >= 2
1167         if (rt->u.dst.obsolete>0) {
1168                 WARN_ON(fn != NULL);
1169                 return -ENOENT;
1170         }
1171 #endif
1172         if (fn == NULL || rt == net->ipv6.ip6_null_entry)
1173                 return -ENOENT;
1174
1175         WARN_ON(!(fn->fn_flags & RTN_RTINFO));
1176
1177         if (!(rt->rt6i_flags&RTF_CACHE)) {
1178                 struct fib6_node *pn = fn;
1179 #ifdef CONFIG_IPV6_SUBTREES
1180                 /* clones of this route might be in another subtree */
1181                 if (rt->rt6i_src.plen) {
1182                         while (!(pn->fn_flags&RTN_ROOT))
1183                                 pn = pn->parent;
1184                         pn = pn->parent;
1185                 }
1186 #endif
1187                 fib6_prune_clones(info->nl_net, pn, rt);
1188         }
1189
1190         /*
1191          *      Walk the leaf entries looking for ourself
1192          */
1193
1194         for (rtp = &fn->leaf; *rtp; rtp = &(*rtp)->u.dst.rt6_next) {
1195                 if (*rtp == rt) {
1196                         fib6_del_route(fn, rtp, info);
1197                         return 0;
1198                 }
1199         }
1200         return -ENOENT;
1201 }
1202
1203 /*
1204  *      Tree traversal function.
1205  *
1206  *      Certainly, it is not interrupt safe.
1207  *      However, it is internally reenterable wrt itself and fib6_add/fib6_del.
1208  *      It means, that we can modify tree during walking
1209  *      and use this function for garbage collection, clone pruning,
1210  *      cleaning tree when a device goes down etc. etc.
1211  *
1212  *      It guarantees that every node will be traversed,
1213  *      and that it will be traversed only once.
1214  *
1215  *      Callback function w->func may return:
1216  *      0 -> continue walking.
1217  *      positive value -> walking is suspended (used by tree dumps,
1218  *      and probably by gc, if it will be split to several slices)
1219  *      negative value -> terminate walking.
1220  *
1221  *      The function itself returns:
1222  *      0   -> walk is complete.
1223  *      >0  -> walk is incomplete (i.e. suspended)
1224  *      <0  -> walk is terminated by an error.
1225  */
1226
1227 static int fib6_walk_continue(struct fib6_walker_t *w)
1228 {
1229         struct fib6_node *fn, *pn;
1230
1231         for (;;) {
1232                 fn = w->node;
1233                 if (fn == NULL)
1234                         return 0;
1235
1236                 if (w->prune && fn != w->root &&
1237                     fn->fn_flags&RTN_RTINFO && w->state < FWS_C) {
1238                         w->state = FWS_C;
1239                         w->leaf = fn->leaf;
1240                 }
1241                 switch (w->state) {
1242 #ifdef CONFIG_IPV6_SUBTREES
1243                 case FWS_S:
1244                         if (FIB6_SUBTREE(fn)) {
1245                                 w->node = FIB6_SUBTREE(fn);
1246                                 continue;
1247                         }
1248                         w->state = FWS_L;
1249 #endif
1250                 case FWS_L:
1251                         if (fn->left) {
1252                                 w->node = fn->left;
1253                                 w->state = FWS_INIT;
1254                                 continue;
1255                         }
1256                         w->state = FWS_R;
1257                 case FWS_R:
1258                         if (fn->right) {
1259                                 w->node = fn->right;
1260                                 w->state = FWS_INIT;
1261                                 continue;
1262                         }
1263                         w->state = FWS_C;
1264                         w->leaf = fn->leaf;
1265                 case FWS_C:
1266                         if (w->leaf && fn->fn_flags&RTN_RTINFO) {
1267                                 int err;
1268
1269                                 if (w->count < w->skip) {
1270                                         w->count++;
1271                                         continue;
1272                                 }
1273
1274                                 err = w->func(w);
1275                                 if (err)
1276                                         return err;
1277
1278                                 w->count++;
1279                                 continue;
1280                         }
1281                         w->state = FWS_U;
1282                 case FWS_U:
1283                         if (fn == w->root)
1284                                 return 0;
1285                         pn = fn->parent;
1286                         w->node = pn;
1287 #ifdef CONFIG_IPV6_SUBTREES
1288                         if (FIB6_SUBTREE(pn) == fn) {
1289                                 WARN_ON(!(fn->fn_flags & RTN_ROOT));
1290                                 w->state = FWS_L;
1291                                 continue;
1292                         }
1293 #endif
1294                         if (pn->left == fn) {
1295                                 w->state = FWS_R;
1296                                 continue;
1297                         }
1298                         if (pn->right == fn) {
1299                                 w->state = FWS_C;
1300                                 w->leaf = w->node->leaf;
1301                                 continue;
1302                         }
1303 #if RT6_DEBUG >= 2
1304                         WARN_ON(1);
1305 #endif
1306                 }
1307         }
1308 }
1309
1310 static int fib6_walk(struct fib6_walker_t *w)
1311 {
1312         int res;
1313
1314         w->state = FWS_INIT;
1315         w->node = w->root;
1316
1317         fib6_walker_link(w);
1318         res = fib6_walk_continue(w);
1319         if (res <= 0)
1320                 fib6_walker_unlink(w);
1321         return res;
1322 }
1323
1324 static int fib6_clean_node(struct fib6_walker_t *w)
1325 {
1326         int res;
1327         struct rt6_info *rt;
1328         struct fib6_cleaner_t *c = container_of(w, struct fib6_cleaner_t, w);
1329         struct nl_info info = {
1330                 .nl_net = c->net,
1331         };
1332
1333         for (rt = w->leaf; rt; rt = rt->u.dst.rt6_next) {
1334                 res = c->func(rt, c->arg);
1335                 if (res < 0) {
1336                         w->leaf = rt;
1337                         res = fib6_del(rt, &info);
1338                         if (res) {
1339 #if RT6_DEBUG >= 2
1340                                 printk(KERN_DEBUG "fib6_clean_node: del failed: rt=%p@%p err=%d\n", rt, rt->rt6i_node, res);
1341 #endif
1342                                 continue;
1343                         }
1344                         return 0;
1345                 }
1346                 WARN_ON(res != 0);
1347         }
1348         w->leaf = rt;
1349         return 0;
1350 }
1351
1352 /*
1353  *      Convenient frontend to tree walker.
1354  *
1355  *      func is called on each route.
1356  *              It may return -1 -> delete this route.
1357  *                            0  -> continue walking
1358  *
1359  *      prune==1 -> only immediate children of node (certainly,
1360  *      ignoring pure split nodes) will be scanned.
1361  */
1362
1363 static void fib6_clean_tree(struct net *net, struct fib6_node *root,
1364                             int (*func)(struct rt6_info *, void *arg),
1365                             int prune, void *arg)
1366 {
1367         struct fib6_cleaner_t c;
1368
1369         c.w.root = root;
1370         c.w.func = fib6_clean_node;
1371         c.w.prune = prune;
1372         c.w.count = 0;
1373         c.w.skip = 0;
1374         c.func = func;
1375         c.arg = arg;
1376         c.net = net;
1377
1378         fib6_walk(&c.w);
1379 }
1380
1381 void fib6_clean_all(struct net *net, int (*func)(struct rt6_info *, void *arg),
1382                     int prune, void *arg)
1383 {
1384         struct fib6_table *table;
1385         struct hlist_node *node;
1386         struct hlist_head *head;
1387         unsigned int h;
1388
1389         rcu_read_lock();
1390         for (h = 0; h < FIB6_TABLE_HASHSZ; h++) {
1391                 head = &net->ipv6.fib_table_hash[h];
1392                 hlist_for_each_entry_rcu(table, node, head, tb6_hlist) {
1393                         write_lock_bh(&table->tb6_lock);
1394                         fib6_clean_tree(net, &table->tb6_root,
1395                                         func, prune, arg);
1396                         write_unlock_bh(&table->tb6_lock);
1397                 }
1398         }
1399         rcu_read_unlock();
1400 }
1401
1402 static int fib6_prune_clone(struct rt6_info *rt, void *arg)
1403 {
1404         if (rt->rt6i_flags & RTF_CACHE) {
1405                 RT6_TRACE("pruning clone %p\n", rt);
1406                 return -1;
1407         }
1408
1409         return 0;
1410 }
1411
1412 static void fib6_prune_clones(struct net *net, struct fib6_node *fn,
1413                               struct rt6_info *rt)
1414 {
1415         fib6_clean_tree(net, fn, fib6_prune_clone, 1, rt);
1416 }
1417
1418 /*
1419  *      Garbage collection
1420  */
1421
1422 static struct fib6_gc_args
1423 {
1424         int                     timeout;
1425         int                     more;
1426 } gc_args;
1427
1428 static int fib6_age(struct rt6_info *rt, void *arg)
1429 {
1430         unsigned long now = jiffies;
1431
1432         /*
1433          *      check addrconf expiration here.
1434          *      Routes are expired even if they are in use.
1435          *
1436          *      Also age clones. Note, that clones are aged out
1437          *      only if they are not in use now.
1438          */
1439
1440         if (rt->rt6i_flags&RTF_EXPIRES && rt->rt6i_expires) {
1441                 if (time_after(now, rt->rt6i_expires)) {
1442                         RT6_TRACE("expiring %p\n", rt);
1443                         return -1;
1444                 }
1445                 gc_args.more++;
1446         } else if (rt->rt6i_flags & RTF_CACHE) {
1447                 if (atomic_read(&rt->u.dst.__refcnt) == 0 &&
1448                     time_after_eq(now, rt->u.dst.lastuse + gc_args.timeout)) {
1449                         RT6_TRACE("aging clone %p\n", rt);
1450                         return -1;
1451                 } else if ((rt->rt6i_flags & RTF_GATEWAY) &&
1452                            (!(rt->rt6i_nexthop->flags & NTF_ROUTER))) {
1453                         RT6_TRACE("purging route %p via non-router but gateway\n",
1454                                   rt);
1455                         return -1;
1456                 }
1457                 gc_args.more++;
1458         }
1459
1460         return 0;
1461 }
1462
1463 static DEFINE_SPINLOCK(fib6_gc_lock);
1464
1465 void fib6_run_gc(unsigned long expires, struct net *net)
1466 {
1467         if (expires != ~0UL) {
1468                 spin_lock_bh(&fib6_gc_lock);
1469                 gc_args.timeout = expires ? (int)expires :
1470                         net->ipv6.sysctl.ip6_rt_gc_interval;
1471         } else {
1472                 if (!spin_trylock_bh(&fib6_gc_lock)) {
1473                         mod_timer(&net->ipv6.ip6_fib_timer, jiffies + HZ);
1474                         return;
1475                 }
1476                 gc_args.timeout = net->ipv6.sysctl.ip6_rt_gc_interval;
1477         }
1478
1479         gc_args.more = icmp6_dst_gc();
1480
1481         fib6_clean_all(net, fib6_age, 0, NULL);
1482
1483         if (gc_args.more)
1484                 mod_timer(&net->ipv6.ip6_fib_timer,
1485                           round_jiffies(jiffies
1486                                         + net->ipv6.sysctl.ip6_rt_gc_interval));
1487         else
1488                 del_timer(&net->ipv6.ip6_fib_timer);
1489         spin_unlock_bh(&fib6_gc_lock);
1490 }
1491
1492 static void fib6_gc_timer_cb(unsigned long arg)
1493 {
1494         fib6_run_gc(0, (struct net *)arg);
1495 }
1496
1497 static int __net_init fib6_net_init(struct net *net)
1498 {
1499         setup_timer(&net->ipv6.ip6_fib_timer, fib6_gc_timer_cb, (unsigned long)net);
1500
1501         net->ipv6.rt6_stats = kzalloc(sizeof(*net->ipv6.rt6_stats), GFP_KERNEL);
1502         if (!net->ipv6.rt6_stats)
1503                 goto out_timer;
1504
1505         net->ipv6.fib_table_hash = kcalloc(FIB6_TABLE_HASHSZ,
1506                                            sizeof(*net->ipv6.fib_table_hash),
1507                                            GFP_KERNEL);
1508         if (!net->ipv6.fib_table_hash)
1509                 goto out_rt6_stats;
1510
1511         net->ipv6.fib6_main_tbl = kzalloc(sizeof(*net->ipv6.fib6_main_tbl),
1512                                           GFP_KERNEL);
1513         if (!net->ipv6.fib6_main_tbl)
1514                 goto out_fib_table_hash;
1515
1516         net->ipv6.fib6_main_tbl->tb6_id = RT6_TABLE_MAIN;
1517         net->ipv6.fib6_main_tbl->tb6_root.leaf = net->ipv6.ip6_null_entry;
1518         net->ipv6.fib6_main_tbl->tb6_root.fn_flags =
1519                 RTN_ROOT | RTN_TL_ROOT | RTN_RTINFO;
1520
1521 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
1522         net->ipv6.fib6_local_tbl = kzalloc(sizeof(*net->ipv6.fib6_local_tbl),
1523                                            GFP_KERNEL);
1524         if (!net->ipv6.fib6_local_tbl)
1525                 goto out_fib6_main_tbl;
1526         net->ipv6.fib6_local_tbl->tb6_id = RT6_TABLE_LOCAL;
1527         net->ipv6.fib6_local_tbl->tb6_root.leaf = net->ipv6.ip6_null_entry;
1528         net->ipv6.fib6_local_tbl->tb6_root.fn_flags =
1529                 RTN_ROOT | RTN_TL_ROOT | RTN_RTINFO;
1530 #endif
1531         fib6_tables_init(net);
1532
1533         return 0;
1534
1535 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
1536 out_fib6_main_tbl:
1537         kfree(net->ipv6.fib6_main_tbl);
1538 #endif
1539 out_fib_table_hash:
1540         kfree(net->ipv6.fib_table_hash);
1541 out_rt6_stats:
1542         kfree(net->ipv6.rt6_stats);
1543 out_timer:
1544         return -ENOMEM;
1545  }
1546
1547 static void fib6_net_exit(struct net *net)
1548 {
1549         rt6_ifdown(net, NULL);
1550         del_timer_sync(&net->ipv6.ip6_fib_timer);
1551
1552 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
1553         kfree(net->ipv6.fib6_local_tbl);
1554 #endif
1555         kfree(net->ipv6.fib6_main_tbl);
1556         kfree(net->ipv6.fib_table_hash);
1557         kfree(net->ipv6.rt6_stats);
1558 }
1559
1560 static struct pernet_operations fib6_net_ops = {
1561         .init = fib6_net_init,
1562         .exit = fib6_net_exit,
1563 };
1564
1565 int __init fib6_init(void)
1566 {
1567         int ret = -ENOMEM;
1568
1569         fib6_node_kmem = kmem_cache_create("fib6_nodes",
1570                                            sizeof(struct fib6_node),
1571                                            0, SLAB_HWCACHE_ALIGN,
1572                                            NULL);
1573         if (!fib6_node_kmem)
1574                 goto out;
1575
1576         ret = register_pernet_subsys(&fib6_net_ops);
1577         if (ret)
1578                 goto out_kmem_cache_create;
1579
1580         ret = __rtnl_register(PF_INET6, RTM_GETROUTE, NULL, inet6_dump_fib);
1581         if (ret)
1582                 goto out_unregister_subsys;
1583 out:
1584         return ret;
1585
1586 out_unregister_subsys:
1587         unregister_pernet_subsys(&fib6_net_ops);
1588 out_kmem_cache_create:
1589         kmem_cache_destroy(fib6_node_kmem);
1590         goto out;
1591 }
1592
1593 void fib6_gc_cleanup(void)
1594 {
1595         unregister_pernet_subsys(&fib6_net_ops);
1596         kmem_cache_destroy(fib6_node_kmem);
1597 }