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 #include <linux/slab.h>
  30
  31 #include <net/ipv6.h>
  32 #include <net/ndisc.h>
  33 #include <net/addrconf.h>
  34
  35 #include <net/ip6_fib.h>
  36 #include <net/ip6_route.h>
  37
  38 #define RT6_DEBUG 2
  39
  40 #if RT6_DEBUG >= 3
  41 #define RT6_TRACE(x...) printk(KERN_DEBUG x)
  42 #else
  43 #define RT6_TRACE(x...) do { ; } while (0)
  44 #endif
  45
  46 static struct kmem_cache * fib6_node_kmem __read_mostly;
  47
  48 enum fib_walk_state_t
  49 {
  50 #ifdef CONFIG_IPV6_SUBTREES
  51         FWS_S,
  52 #endif
  53         FWS_L,
  54         FWS_R,
  55         FWS_C,
  56         FWS_U
  57 };
  58
  59 struct fib6_cleaner_t
  60 {
  61         struct fib6_walker_t w;
  62         struct net *net;
  63         int (*func)(struct rt6_info *, void *arg);
  64         void *arg;
  65 };
  66
  67 static DEFINE_RWLOCK(fib6_walker_lock);
  68
  69 #ifdef CONFIG_IPV6_SUBTREES
  70 #define FWS_INIT FWS_S
  71 #else
  72 #define FWS_INIT FWS_L
  73 #endif
  74
  75 static void fib6_prune_clones(struct net *net, struct fib6_node *fn,
  76                               struct rt6_info *rt);
  77 static struct rt6_info *fib6_find_prefix(struct net *net, struct fib6_node *fn);
  78 static struct fib6_node *fib6_repair_tree(struct net *net, struct fib6_node *fn);
  79 static int fib6_walk(struct fib6_walker_t *w);
  80 static int fib6_walk_continue(struct fib6_walker_t *w);
  81
  82 /*
  83  *      A routing update causes an increase of the serial number on the
  84  *      affected subtree. This allows for cached routes to be asynchronously
  85  *      tested when modifications are made to the destination cache as a
  86  *      result of redirects, path MTU changes, etc.
  87  */
  88
  89 static __u32 rt_sernum;
  90
  91 static void fib6_gc_timer_cb(unsigned long arg);
  92
  93 static LIST_HEAD(fib6_walkers);
  94 #define FOR_WALKERS(w) list_for_each_entry(w, &fib6_walkers, lh)
  95
  96 static inline void fib6_walker_link(struct fib6_walker_t *w)
  97 {
  98         write_lock_bh(&fib6_walker_lock);
  99         list_add(&w->lh, &fib6_walkers);
 100         write_unlock_bh(&fib6_walker_lock);
 101 }
 102
 103 static inline void fib6_walker_unlink(struct fib6_walker_t *w)
 104 {
 105         write_lock_bh(&fib6_walker_lock);
 106         list_del(&w->lh);
 107         write_unlock_bh(&fib6_walker_lock);
 108 }
 109 static __inline__ u32 fib6_new_sernum(void)
 110 {
 111         u32 n = ++rt_sernum;
 112         if ((__s32)n <= 0)
 113                 rt_sernum = n = 1;
 114         return n;
 115 }
 116
 117 /*
 118  *      Auxiliary address test functions for the radix tree.
 119  *
 120  *      These assume a 32bit processor (although it will work on
 121  *      64bit processors)
 122  */
 123
 124 /*
 125  *      test bit
 126  */
 127 #if defined(__LITTLE_ENDIAN)
 128 # define BITOP_BE32_SWIZZLE     (0x1F & ~7)
 129 #else
 130 # define BITOP_BE32_SWIZZLE     0
 131 #endif
 132
 133 static __inline__ __be32 addr_bit_set(const void *token, int fn_bit)
 134 {
 135         const __be32 *addr = token;
 136         /*
 137          * Here,
 138          *      1 << ((~fn_bit ^ BITOP_BE32_SWIZZLE) & 0x1f)
 139          * is optimized version of
 140          *      htonl(1 << ((~fn_bit)&0x1F))
 141          * See include/asm-generic/bitops/le.h.
 142          */
 143         return (__force __be32)(1 << ((~fn_bit ^ BITOP_BE32_SWIZZLE) & 0x1f)) &
 144                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->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 flowi6 *fl6,
 260                                    int flags, pol_lookup_t lookup)
 261 {
 262         return (struct dst_entry *) lookup(net, net->ipv6.fib6_main_tbl, fl6, 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->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         rcu_read_lock();
 394         for (h = s_h; h < FIB6_TABLE_HASHSZ; h++, s_e = 0) {
 395                 e = 0;
 396                 head = &net->ipv6.fib_table_hash[h];
 397                 hlist_for_each_entry_rcu(tb, node, head, tb6_hlist) {
 398                         if (e < s_e)
 399                                 goto next;
 400                         res = fib6_dump_table(tb, skb, cb);
 401                         if (res != 0)
 402                                 goto out;
 403 next:
 404                         e++;
 405                 }
 406         }
 407 out:
 408         rcu_read_unlock();
 409         cb->args[1] = e;
 410         cb->args[0] = h;
 411
 412         res = res < 0 ? res : skb->len;
 413         if (res <= 0)
 414                 fib6_dump_end(cb);
 415         return res;
 416 }
 417
 418 /*
 419  *      Routing Table
 420  *
 421  *      return the appropriate node for a routing tree "add" operation
 422  *      by either creating and inserting or by returning an existing
 423  *      node.
 424  */
 425
 426 static struct fib6_node * fib6_add_1(struct fib6_node *root, void *addr,
 427                                      int addrlen, int plen,
 428                                      int offset)
 429 {
 430         struct fib6_node *fn, *in, *ln;
 431         struct fib6_node *pn = NULL;
 432         struct rt6key *key;
 433         int     bit;
 434         __be32  dir = 0;
 435         __u32   sernum = fib6_new_sernum();
 436
 437         RT6_TRACE("fib6_add_1\n");
 438
 439         /* insert node in tree */
 440
 441         fn = root;
 442
 443         do {
 444                 key = (struct rt6key *)((u8 *)fn->leaf + offset);
 445
 446                 /*
 447                  *      Prefix match
 448                  */
 449                 if (plen < fn->fn_bit ||
 450                     !ipv6_prefix_equal(&key->addr, addr, fn->fn_bit))
 451                         goto insert_above;
 452
 453                 /*
 454                  *      Exact match ?
 455                  */
 456
 457                 if (plen == fn->fn_bit) {
 458                         /* clean up an intermediate node */
 459                         if ((fn->fn_flags & RTN_RTINFO) == 0) {
 460                                 rt6_release(fn->leaf);
 461                                 fn->leaf = NULL;
 462                         }
 463
 464                         fn->fn_sernum = sernum;
 465
 466                         return fn;
 467                 }
 468
 469                 /*
 470                  *      We have more bits to go
 471                  */
 472
 473                 /* Try to walk down on tree. */
 474                 fn->fn_sernum = sernum;
 475                 dir = addr_bit_set(addr, fn->fn_bit);
 476                 pn = fn;
 477                 fn = dir ? fn->right: fn->left;
 478         } while (fn);
 479
 480         /*
 481          *      We walked to the bottom of tree.
 482          *      Create new leaf node without children.
 483          */
 484
 485         ln = node_alloc();
 486
 487         if (ln == NULL)
 488                 return NULL;
 489         ln->fn_bit = plen;
 490
 491         ln->parent = pn;
 492         ln->fn_sernum = sernum;
 493
 494         if (dir)
 495                 pn->right = ln;
 496         else
 497                 pn->left  = ln;
 498
 499         return ln;
 500
 501
 502 insert_above:
 503         /*
 504          * split since we don't have a common prefix anymore or
 505          * we have a less significant route.
 506          * we've to insert an intermediate node on the list
 507          * this new node will point to the one we need to create
 508          * and the current
 509          */
 510
 511         pn = fn->parent;
 512
 513         /* find 1st bit in difference between the 2 addrs.
 514
 515            See comment in __ipv6_addr_diff: bit may be an invalid value,
 516            but if it is >= plen, the value is ignored in any case.
 517          */
 518
 519         bit = __ipv6_addr_diff(addr, &key->addr, addrlen);
 520
 521         /*
 522          *              (intermediate)[in]
 523          *                /        \
 524          *      (new leaf node)[ln] (old node)[fn]
 525          */
 526         if (plen > bit) {
 527                 in = node_alloc();
 528                 ln = node_alloc();
 529
 530                 if (in == NULL || ln == NULL) {
 531                         if (in)
 532                                 node_free(in);
 533                         if (ln)
 534                                 node_free(ln);
 535                         return NULL;
 536                 }
 537
 538                 /*
 539                  * new intermediate node.
 540                  * RTN_RTINFO will
 541                  * be off since that an address that chooses one of
 542                  * the branches would not match less specific routes
 543                  * in the other branch
 544                  */
 545
 546                 in->fn_bit = bit;
 547
 548                 in->parent = pn;
 549                 in->leaf = fn->leaf;
 550                 atomic_inc(&in->leaf->rt6i_ref);
 551
 552                 in->fn_sernum = sernum;
 553
 554                 /* update parent pointer */
 555                 if (dir)
 556                         pn->right = in;
 557                 else
 558                         pn->left  = in;
 559
 560                 ln->fn_bit = plen;
 561
 562                 ln->parent = in;
 563                 fn->parent = in;
 564
 565                 ln->fn_sernum = sernum;
 566
 567                 if (addr_bit_set(addr, bit)) {
 568                         in->right = ln;
 569                         in->left  = fn;
 570                 } else {
 571                         in->left  = ln;
 572                         in->right = fn;
 573                 }
 574         } else { /* plen <= bit */
 575
 576                 /*
 577                  *              (new leaf node)[ln]
 578                  *                /        \
 579                  *           (old node)[fn] NULL
 580                  */
 581
 582                 ln = node_alloc();
 583
 584                 if (ln == NULL)
 585                         return NULL;
 586
 587                 ln->fn_bit = plen;
 588
 589                 ln->parent = pn;
 590
 591                 ln->fn_sernum = sernum;
 592
 593                 if (dir)
 594                         pn->right = ln;
 595                 else
 596                         pn->left  = ln;
 597
 598                 if (addr_bit_set(&key->addr, plen))
 599                         ln->right = fn;
 600                 else
 601                         ln->left  = fn;
 602
 603                 fn->parent = ln;
 604         }
 605         return ln;
 606 }
 607
 608 /*
 609  *      Insert routing information in a node.
 610  */
 611
 612 static int fib6_add_rt2node(struct fib6_node *fn, struct rt6_info *rt,
 613                             struct nl_info *info)
 614 {
 615         struct rt6_info *iter = NULL;
 616         struct rt6_info **ins;
 617
 618         ins = &fn->leaf;
 619
 620         for (iter = fn->leaf; iter; iter=iter->dst.rt6_next) {
 621                 /*
 622                  *      Search for duplicates
 623                  */
 624
 625                 if (iter->rt6i_metric == rt->rt6i_metric) {
 626                         /*
 627                          *      Same priority level
 628                          */
 629
 630                         if (iter->rt6i_dev == rt->rt6i_dev &&
 631                             iter->rt6i_idev == rt->rt6i_idev &&
 632                             ipv6_addr_equal(&iter->rt6i_gateway,
 633                                             &rt->rt6i_gateway)) {
 634                                 if (!(iter->rt6i_flags&RTF_EXPIRES))
 635                                         return -EEXIST;
 636                                 iter->rt6i_expires = rt->rt6i_expires;
 637                                 if (!(rt->rt6i_flags&RTF_EXPIRES)) {
 638                                         iter->rt6i_flags &= ~RTF_EXPIRES;
 639                                         iter->rt6i_expires = 0;
 640                                 }
 641                                 return -EEXIST;
 642                         }
 643                 }
 644
 645                 if (iter->rt6i_metric > rt->rt6i_metric)
 646                         break;
 647
 648                 ins = &iter->dst.rt6_next;
 649         }
 650
 651         /* Reset round-robin state, if necessary */
 652         if (ins == &fn->leaf)
 653                 fn->rr_ptr = NULL;
 654
 655         /*
 656          *      insert node
 657          */
 658
 659         rt->dst.rt6_next = iter;
 660         *ins = rt;
 661         rt->rt6i_node = fn;
 662         atomic_inc(&rt->rt6i_ref);
 663         inet6_rt_notify(RTM_NEWROUTE, rt, info);
 664         info->nl_net->ipv6.rt6_stats->fib_rt_entries++;
 665
 666         if ((fn->fn_flags & RTN_RTINFO) == 0) {
 667                 info->nl_net->ipv6.rt6_stats->fib_route_nodes++;
 668                 fn->fn_flags |= RTN_RTINFO;
 669         }
 670
 671         return 0;
 672 }
 673
 674 static __inline__ void fib6_start_gc(struct net *net, struct rt6_info *rt)
 675 {
 676         if (!timer_pending(&net->ipv6.ip6_fib_timer) &&
 677             (rt->rt6i_flags & (RTF_EXPIRES|RTF_CACHE)))
 678                 mod_timer(&net->ipv6.ip6_fib_timer,
 679                           jiffies + net->ipv6.sysctl.ip6_rt_gc_interval);
 680 }
 681
 682 void fib6_force_start_gc(struct net *net)
 683 {
 684         if (!timer_pending(&net->ipv6.ip6_fib_timer))
 685                 mod_timer(&net->ipv6.ip6_fib_timer,
 686                           jiffies + net->ipv6.sysctl.ip6_rt_gc_interval);
 687 }
 688
 689 /*
 690  *      Add routing information to the routing tree.
 691  *      <destination addr>/<source addr>
 692  *      with source addr info in sub-trees
 693  */
 694
 695 int fib6_add(struct fib6_node *root, struct rt6_info *rt, struct nl_info *info)
 696 {
 697         struct fib6_node *fn, *pn = NULL;
 698         int err = -ENOMEM;
 699
 700         fn = fib6_add_1(root, &rt->rt6i_dst.addr, sizeof(struct in6_addr),
 701                         rt->rt6i_dst.plen, offsetof(struct rt6_info, rt6i_dst));
 702
 703         if (fn == NULL)
 704                 goto out;
 705
 706         pn = fn;
 707
 708 #ifdef CONFIG_IPV6_SUBTREES
 709         if (rt->rt6i_src.plen) {
 710                 struct fib6_node *sn;
 711
 712                 if (fn->subtree == NULL) {
 713                         struct fib6_node *sfn;
 714
 715                         /*
 716                          * Create subtree.
 717                          *
 718                          *              fn[main tree]
 719                          *              |
 720                          *              sfn[subtree root]
 721                          *                 \
 722                          *                  sn[new leaf node]
 723                          */
 724
 725                         /* Create subtree root node */
 726                         sfn = node_alloc();
 727                         if (sfn == NULL)
 728                                 goto st_failure;
 729
 730                         sfn->leaf = info->nl_net->ipv6.ip6_null_entry;
 731                         atomic_inc(&info->nl_net->ipv6.ip6_null_entry->rt6i_ref);
 732                         sfn->fn_flags = RTN_ROOT;
 733                         sfn->fn_sernum = fib6_new_sernum();
 734
 735                         /* Now add the first leaf node to new subtree */
 736
 737                         sn = fib6_add_1(sfn, &rt->rt6i_src.addr,
 738                                         sizeof(struct in6_addr), rt->rt6i_src.plen,
 739                                         offsetof(struct rt6_info, rt6i_src));
 740
 741                         if (sn == NULL) {
 742                                 /* If it is failed, discard just allocated
 743                                    root, and then (in st_failure) stale node
 744                                    in main tree.
 745                                  */
 746                                 node_free(sfn);
 747                                 goto st_failure;
 748                         }
 749
 750                         /* Now link new subtree to main tree */
 751                         sfn->parent = fn;
 752                         fn->subtree = sfn;
 753                 } else {
 754                         sn = fib6_add_1(fn->subtree, &rt->rt6i_src.addr,
 755                                         sizeof(struct in6_addr), rt->rt6i_src.plen,
 756                                         offsetof(struct rt6_info, rt6i_src));
 757
 758                         if (sn == NULL)
 759                                 goto st_failure;
 760                 }
 761
 762                 if (fn->leaf == NULL) {
 763                         fn->leaf = rt;
 764                         atomic_inc(&rt->rt6i_ref);
 765                 }
 766                 fn = sn;
 767         }
 768 #endif
 769
 770         err = fib6_add_rt2node(fn, rt, info);
 771
 772         if (err == 0) {
 773                 fib6_start_gc(info->nl_net, rt);
 774                 if (!(rt->rt6i_flags&RTF_CACHE))
 775                         fib6_prune_clones(info->nl_net, pn, rt);
 776         }
 777
 778 out:
 779         if (err) {
 780 #ifdef CONFIG_IPV6_SUBTREES
 781                 /*
 782                  * If fib6_add_1 has cleared the old leaf pointer in the
 783                  * super-tree leaf node we have to find a new one for it.
 784                  */
 785                 if (pn != fn && pn->leaf == rt) {
 786                         pn->leaf = NULL;
 787                         atomic_dec(&rt->rt6i_ref);
 788                 }
 789                 if (pn != fn && !pn->leaf && !(pn->fn_flags & RTN_RTINFO)) {
 790                         pn->leaf = fib6_find_prefix(info->nl_net, pn);
 791 #if RT6_DEBUG >= 2
 792                         if (!pn->leaf) {
 793                                 WARN_ON(pn->leaf == NULL);
 794                                 pn->leaf = info->nl_net->ipv6.ip6_null_entry;
 795                         }
 796 #endif
 797                         atomic_inc(&pn->leaf->rt6i_ref);
 798                 }
 799 #endif
 800                 dst_free(&rt->dst);
 801         }
 802         return err;
 803
 804 #ifdef CONFIG_IPV6_SUBTREES
 805         /* Subtree creation failed, probably main tree node
 806            is orphan. If it is, shoot it.
 807          */
 808 st_failure:
 809         if (fn && !(fn->fn_flags & (RTN_RTINFO|RTN_ROOT)))
 810                 fib6_repair_tree(info->nl_net, fn);
 811         dst_free(&rt->dst);
 812         return err;
 813 #endif
 814 }
 815
 816 /*
 817  *      Routing tree lookup
 818  *
 819  */
 820
 821 struct lookup_args {
 822         int             offset;         /* key offset on rt6_info       */
 823         const struct in6_addr   *addr;          /* search key                   */
 824 };
 825
 826 static struct fib6_node * fib6_lookup_1(struct fib6_node *root,
 827                                         struct lookup_args *args)
 828 {
 829         struct fib6_node *fn;
 830         __be32 dir;
 831
 832         if (unlikely(args->offset == 0))
 833                 return NULL;
 834
 835         /*
 836          *      Descend on a tree
 837          */
 838
 839         fn = root;
 840
 841         for (;;) {
 842                 struct fib6_node *next;
 843
 844                 dir = addr_bit_set(args->addr, fn->fn_bit);
 845
 846                 next = dir ? fn->right : fn->left;
 847
 848                 if (next) {
 849                         fn = next;
 850                         continue;
 851                 }
 852
 853                 break;
 854         }
 855
 856         while(fn) {
 857                 if (FIB6_SUBTREE(fn) || fn->fn_flags & RTN_RTINFO) {
 858                         struct rt6key *key;
 859
 860                         key = (struct rt6key *) ((u8 *) fn->leaf +
 861                                                  args->offset);
 862
 863                         if (ipv6_prefix_equal(&key->addr, args->addr, key->plen)) {
 864 #ifdef CONFIG_IPV6_SUBTREES
 865                                 if (fn->subtree) {
 866                                         struct fib6_node *sfn;
 867                                         sfn = fib6_lookup_1(fn->subtree,
 868                                                             args + 1);
 869                                         if (!sfn)
 870                                                 goto backtrack;
 871                                         fn = sfn;
 872                                 }
 873 #endif
 874                                 if (fn->fn_flags & RTN_RTINFO)
 875                                         return fn;
 876                         }
 877                 }
 878 #ifdef CONFIG_IPV6_SUBTREES
 879 backtrack:
 880 #endif
 881                 if (fn->fn_flags & RTN_ROOT)
 882                         break;
 883
 884                 fn = fn->parent;
 885         }
 886
 887         return NULL;
 888 }
 889
 890 struct fib6_node * fib6_lookup(struct fib6_node *root, const struct in6_addr *daddr,
 891                                const struct in6_addr *saddr)
 892 {
 893         struct fib6_node *fn;
 894         struct lookup_args args[] = {
 895                 {
 896                         .offset = offsetof(struct rt6_info, rt6i_dst),
 897                         .addr = daddr,
 898                 },
 899 #ifdef CONFIG_IPV6_SUBTREES
 900                 {
 901                         .offset = offsetof(struct rt6_info, rt6i_src),
 902                         .addr = saddr,
 903                 },
 904 #endif
 905                 {
 906                         .offset = 0,    /* sentinel */
 907                 }
 908         };
 909
 910         fn = fib6_lookup_1(root, daddr ? args : args + 1);
 911
 912         if (fn == NULL || fn->fn_flags & RTN_TL_ROOT)
 913                 fn = root;
 914
 915         return fn;
 916 }
 917
 918 /*
 919  *      Get node with specified destination prefix (and source prefix,
 920  *      if subtrees are used)
 921  */
 922
 923
 924 static struct fib6_node * fib6_locate_1(struct fib6_node *root,
 925                                         const struct in6_addr *addr,
 926                                         int plen, int offset)
 927 {
 928         struct fib6_node *fn;
 929
 930         for (fn = root; fn ; ) {
 931                 struct rt6key *key = (struct rt6key *)((u8 *)fn->leaf + offset);
 932
 933                 /*
 934                  *      Prefix match
 935                  */
 936                 if (plen < fn->fn_bit ||
 937                     !ipv6_prefix_equal(&key->addr, addr, fn->fn_bit))
 938                         return NULL;
 939
 940                 if (plen == fn->fn_bit)
 941                         return fn;
 942
 943                 /*
 944                  *      We have more bits to go
 945                  */
 946                 if (addr_bit_set(addr, fn->fn_bit))
 947                         fn = fn->right;
 948                 else
 949                         fn = fn->left;
 950         }
 951         return NULL;
 952 }
 953
 954 struct fib6_node * fib6_locate(struct fib6_node *root,
 955                                const struct in6_addr *daddr, int dst_len,
 956                                const struct in6_addr *saddr, int src_len)
 957 {
 958         struct fib6_node *fn;
 959
 960         fn = fib6_locate_1(root, daddr, dst_len,
 961                            offsetof(struct rt6_info, rt6i_dst));
 962
 963 #ifdef CONFIG_IPV6_SUBTREES
 964         if (src_len) {
 965                 WARN_ON(saddr == NULL);
 966                 if (fn && fn->subtree)
 967                         fn = fib6_locate_1(fn->subtree, saddr, src_len,
 968                                            offsetof(struct rt6_info, rt6i_src));
 969         }
 970 #endif
 971
 972         if (fn && fn->fn_flags&RTN_RTINFO)
 973                 return fn;
 974
 975         return NULL;
 976 }
 977
 978
 979 /*
 980  *      Deletion
 981  *
 982  */
 983
 984 static struct rt6_info *fib6_find_prefix(struct net *net, struct fib6_node *fn)
 985 {
 986         if (fn->fn_flags&RTN_ROOT)
 987                 return net->ipv6.ip6_null_entry;
 988
 989         while(fn) {
 990                 if(fn->left)
 991                         return fn->left->leaf;
 992
 993                 if(fn->right)
 994                         return fn->right->leaf;
 995
 996                 fn = FIB6_SUBTREE(fn);
 997         }
 998         return NULL;
 999 }
1000
1001 /*
1002  *      Called to trim the tree of intermediate nodes when possible. "fn"
1003  *      is the node we want to try and remove.
1004  */
1005
1006 static struct fib6_node *fib6_repair_tree(struct net *net,
1007                                            struct fib6_node *fn)
1008 {
1009         int children;
1010         int nstate;
1011         struct fib6_node *child, *pn;
1012         struct fib6_walker_t *w;
1013         int iter = 0;
1014
1015         for (;;) {
1016                 RT6_TRACE("fixing tree: plen=%d iter=%d\n", fn->fn_bit, iter);
1017                 iter++;
1018
1019                 WARN_ON(fn->fn_flags & RTN_RTINFO);
1020                 WARN_ON(fn->fn_flags & RTN_TL_ROOT);
1021                 WARN_ON(fn->leaf != NULL);
1022
1023                 children = 0;
1024                 child = NULL;
1025                 if (fn->right) child = fn->right, children |= 1;
1026                 if (fn->left) child = fn->left, children |= 2;
1027
1028                 if (children == 3 || FIB6_SUBTREE(fn)
1029 #ifdef CONFIG_IPV6_SUBTREES
1030                     /* Subtree root (i.e. fn) may have one child */
1031                     || (children && fn->fn_flags&RTN_ROOT)
1032 #endif
1033                     ) {
1034                         fn->leaf = fib6_find_prefix(net, fn);
1035 #if RT6_DEBUG >= 2
1036                         if (fn->leaf==NULL) {
1037                                 WARN_ON(!fn->leaf);
1038                                 fn->leaf = net->ipv6.ip6_null_entry;
1039                         }
1040 #endif
1041                         atomic_inc(&fn->leaf->rt6i_ref);
1042                         return fn->parent;
1043                 }
1044
1045                 pn = fn->parent;
1046 #ifdef CONFIG_IPV6_SUBTREES
1047                 if (FIB6_SUBTREE(pn) == fn) {
1048                         WARN_ON(!(fn->fn_flags & RTN_ROOT));
1049                         FIB6_SUBTREE(pn) = NULL;
1050                         nstate = FWS_L;
1051                 } else {
1052                         WARN_ON(fn->fn_flags & RTN_ROOT);
1053 #endif
1054                         if (pn->right == fn) pn->right = child;
1055                         else if (pn->left == fn) pn->left = child;
1056 #if RT6_DEBUG >= 2
1057                         else
1058                                 WARN_ON(1);
1059 #endif
1060                         if (child)
1061                                 child->parent = pn;
1062                         nstate = FWS_R;
1063 #ifdef CONFIG_IPV6_SUBTREES
1064                 }
1065 #endif
1066
1067                 read_lock(&fib6_walker_lock);
1068                 FOR_WALKERS(w) {
1069                         if (child == NULL) {
1070                                 if (w->root == fn) {
1071                                         w->root = w->node = NULL;
1072                                         RT6_TRACE("W %p adjusted by delroot 1\n", w);
1073                                 } else if (w->node == fn) {
1074                                         RT6_TRACE("W %p adjusted by delnode 1, s=%d/%d\n", w, w->state, nstate);
1075                                         w->node = pn;
1076                                         w->state = nstate;
1077                                 }
1078                         } else {
1079                                 if (w->root == fn) {
1080                                         w->root = child;
1081                                         RT6_TRACE("W %p adjusted by delroot 2\n", w);
1082                                 }
1083                                 if (w->node == fn) {
1084                                         w->node = child;
1085                                         if (children&2) {
1086                                                 RT6_TRACE("W %p adjusted by delnode 2, s=%d\n", w, w->state);
1087                                                 w->state = w->state>=FWS_R ? FWS_U : FWS_INIT;
1088                                         } else {
1089                                                 RT6_TRACE("W %p adjusted by delnode 2, s=%d\n", w, w->state);
1090                                                 w->state = w->state>=FWS_C ? FWS_U : FWS_INIT;
1091                                         }
1092                                 }
1093                         }
1094                 }
1095                 read_unlock(&fib6_walker_lock);
1096
1097                 node_free(fn);
1098                 if (pn->fn_flags&RTN_RTINFO || FIB6_SUBTREE(pn))
1099                         return pn;
1100
1101                 rt6_release(pn->leaf);
1102                 pn->leaf = NULL;
1103                 fn = pn;
1104         }
1105 }
1106
1107 static void fib6_del_route(struct fib6_node *fn, struct rt6_info **rtp,
1108                            struct nl_info *info)
1109 {
1110         struct fib6_walker_t *w;
1111         struct rt6_info *rt = *rtp;
1112         struct net *net = info->nl_net;
1113
1114         RT6_TRACE("fib6_del_route\n");
1115
1116         /* Unlink it */
1117         *rtp = rt->dst.rt6_next;
1118         rt->rt6i_node = NULL;
1119         net->ipv6.rt6_stats->fib_rt_entries--;
1120         net->ipv6.rt6_stats->fib_discarded_routes++;
1121
1122         /* Reset round-robin state, if necessary */
1123         if (fn->rr_ptr == rt)
1124                 fn->rr_ptr = NULL;
1125
1126         /* Adjust walkers */
1127         read_lock(&fib6_walker_lock);
1128         FOR_WALKERS(w) {
1129                 if (w->state == FWS_C && w->leaf == rt) {
1130                         RT6_TRACE("walker %p adjusted by delroute\n", w);
1131                         w->leaf = rt->dst.rt6_next;
1132                         if (w->leaf == NULL)
1133                                 w->state = FWS_U;
1134                 }
1135         }
1136         read_unlock(&fib6_walker_lock);
1137
1138         rt->dst.rt6_next = NULL;
1139
1140         /* If it was last route, expunge its radix tree node */
1141         if (fn->leaf == NULL) {
1142                 fn->fn_flags &= ~RTN_RTINFO;
1143                 net->ipv6.rt6_stats->fib_route_nodes--;
1144                 fn = fib6_repair_tree(net, fn);
1145         }
1146
1147         if (atomic_read(&rt->rt6i_ref) != 1) {
1148                 /* This route is used as dummy address holder in some split
1149                  * nodes. It is not leaked, but it still holds other resources,
1150                  * which must be released in time. So, scan ascendant nodes
1151                  * and replace dummy references to this route with references
1152                  * to still alive ones.
1153                  */
1154                 while (fn) {
1155                         if (!(fn->fn_flags&RTN_RTINFO) && fn->leaf == rt) {
1156                                 fn->leaf = fib6_find_prefix(net, fn);
1157                                 atomic_inc(&fn->leaf->rt6i_ref);
1158                                 rt6_release(rt);
1159                         }
1160                         fn = fn->parent;
1161                 }
1162                 /* No more references are possible at this point. */
1163                 BUG_ON(atomic_read(&rt->rt6i_ref) != 1);
1164         }
1165
1166         inet6_rt_notify(RTM_DELROUTE, rt, info);
1167         rt6_release(rt);
1168 }
1169
1170 int fib6_del(struct rt6_info *rt, struct nl_info *info)
1171 {
1172         struct net *net = info->nl_net;
1173         struct fib6_node *fn = rt->rt6i_node;
1174         struct rt6_info **rtp;
1175
1176 #if RT6_DEBUG >= 2
1177         if (rt->dst.obsolete>0) {
1178                 WARN_ON(fn != NULL);
1179                 return -ENOENT;
1180         }
1181 #endif
1182         if (fn == NULL || rt == net->ipv6.ip6_null_entry)
1183                 return -ENOENT;
1184
1185         WARN_ON(!(fn->fn_flags & RTN_RTINFO));
1186
1187         if (!(rt->rt6i_flags&RTF_CACHE)) {
1188                 struct fib6_node *pn = fn;
1189 #ifdef CONFIG_IPV6_SUBTREES
1190                 /* clones of this route might be in another subtree */
1191                 if (rt->rt6i_src.plen) {
1192                         while (!(pn->fn_flags&RTN_ROOT))
1193                                 pn = pn->parent;
1194                         pn = pn->parent;
1195                 }
1196 #endif
1197                 fib6_prune_clones(info->nl_net, pn, rt);
1198         }
1199
1200         /*
1201          *      Walk the leaf entries looking for ourself
1202          */
1203
1204         for (rtp = &fn->leaf; *rtp; rtp = &(*rtp)->dst.rt6_next) {
1205                 if (*rtp == rt) {
1206                         fib6_del_route(fn, rtp, info);
1207                         return 0;
1208                 }
1209         }
1210         return -ENOENT;
1211 }
1212
1213 /*
1214  *      Tree traversal function.
1215  *
1216  *      Certainly, it is not interrupt safe.
1217  *      However, it is internally reenterable wrt itself and fib6_add/fib6_del.
1218  *      It means, that we can modify tree during walking
1219  *      and use this function for garbage collection, clone pruning,
1220  *      cleaning tree when a device goes down etc. etc.
1221  *
1222  *      It guarantees that every node will be traversed,
1223  *      and that it will be traversed only once.
1224  *
1225  *      Callback function w->func may return:
1226  *      0 -> continue walking.
1227  *      positive value -> walking is suspended (used by tree dumps,
1228  *      and probably by gc, if it will be split to several slices)
1229  *      negative value -> terminate walking.
1230  *
1231  *      The function itself returns:
1232  *      0   -> walk is complete.
1233  *      >0  -> walk is incomplete (i.e. suspended)
1234  *      <0  -> walk is terminated by an error.
1235  */
1236
1237 static int fib6_walk_continue(struct fib6_walker_t *w)
1238 {
1239         struct fib6_node *fn, *pn;
1240
1241         for (;;) {
1242                 fn = w->node;
1243                 if (fn == NULL)
1244                         return 0;
1245
1246                 if (w->prune && fn != w->root &&
1247                     fn->fn_flags&RTN_RTINFO && w->state < FWS_C) {
1248                         w->state = FWS_C;
1249                         w->leaf = fn->leaf;
1250                 }
1251                 switch (w->state) {
1252 #ifdef CONFIG_IPV6_SUBTREES
1253                 case FWS_S:
1254                         if (FIB6_SUBTREE(fn)) {
1255                                 w->node = FIB6_SUBTREE(fn);
1256                                 continue;
1257                         }
1258                         w->state = FWS_L;
1259 #endif
1260                 case FWS_L:
1261                         if (fn->left) {
1262                                 w->node = fn->left;
1263                                 w->state = FWS_INIT;
1264                                 continue;
1265                         }
1266                         w->state = FWS_R;
1267                 case FWS_R:
1268                         if (fn->right) {
1269                                 w->node = fn->right;
1270                                 w->state = FWS_INIT;
1271                                 continue;
1272                         }
1273                         w->state = FWS_C;
1274                         w->leaf = fn->leaf;
1275                 case FWS_C:
1276                         if (w->leaf && fn->fn_flags&RTN_RTINFO) {
1277                                 int err;
1278
1279                                 if (w->count < w->skip) {
1280                                         w->count++;
1281                                         continue;
1282                                 }
1283
1284                                 err = w->func(w);
1285                                 if (err)
1286                                         return err;
1287
1288                                 w->count++;
1289                                 continue;
1290                         }
1291                         w->state = FWS_U;
1292                 case FWS_U:
1293                         if (fn == w->root)
1294                                 return 0;
1295                         pn = fn->parent;
1296                         w->node = pn;
1297 #ifdef CONFIG_IPV6_SUBTREES
1298                         if (FIB6_SUBTREE(pn) == fn) {
1299                                 WARN_ON(!(fn->fn_flags & RTN_ROOT));
1300                                 w->state = FWS_L;
1301                                 continue;
1302                         }
1303 #endif
1304                         if (pn->left == fn) {
1305                                 w->state = FWS_R;
1306                                 continue;
1307                         }
1308                         if (pn->right == fn) {
1309                                 w->state = FWS_C;
1310                                 w->leaf = w->node->leaf;
1311                                 continue;
1312                         }
1313 #if RT6_DEBUG >= 2
1314                         WARN_ON(1);
1315 #endif
1316                 }
1317         }
1318 }
1319
1320 static int fib6_walk(struct fib6_walker_t *w)
1321 {
1322         int res;
1323
1324         w->state = FWS_INIT;
1325         w->node = w->root;
1326
1327         fib6_walker_link(w);
1328         res = fib6_walk_continue(w);
1329         if (res <= 0)
1330                 fib6_walker_unlink(w);
1331         return res;
1332 }
1333
1334 static int fib6_clean_node(struct fib6_walker_t *w)
1335 {
1336         int res;
1337         struct rt6_info *rt;
1338         struct fib6_cleaner_t *c = container_of(w, struct fib6_cleaner_t, w);
1339         struct nl_info info = {
1340                 .nl_net = c->net,
1341         };
1342
1343         for (rt = w->leaf; rt; rt = rt->dst.rt6_next) {
1344                 res = c->func(rt, c->arg);
1345                 if (res < 0) {
1346                         w->leaf = rt;
1347                         res = fib6_del(rt, &info);
1348                         if (res) {
1349 #if RT6_DEBUG >= 2
1350                                 printk(KERN_DEBUG "fib6_clean_node: del failed: rt=%p@%p err=%d\n", rt, rt->rt6i_node, res);
1351 #endif
1352                                 continue;
1353                         }
1354                         return 0;
1355                 }
1356                 WARN_ON(res != 0);
1357         }
1358         w->leaf = rt;
1359         return 0;
1360 }
1361
1362 /*
1363  *      Convenient frontend to tree walker.
1364  *
1365  *      func is called on each route.
1366  *              It may return -1 -> delete this route.
1367  *                            0  -> continue walking
1368  *
1369  *      prune==1 -> only immediate children of node (certainly,
1370  *      ignoring pure split nodes) will be scanned.
1371  */
1372
1373 static void fib6_clean_tree(struct net *net, struct fib6_node *root,
1374                             int (*func)(struct rt6_info *, void *arg),
1375                             int prune, void *arg)
1376 {
1377         struct fib6_cleaner_t c;
1378
1379         c.w.root = root;
1380         c.w.func = fib6_clean_node;
1381         c.w.prune = prune;
1382         c.w.count = 0;
1383         c.w.skip = 0;
1384         c.func = func;
1385         c.arg = arg;
1386         c.net = net;
1387
1388         fib6_walk(&c.w);
1389 }
1390
1391 void fib6_clean_all(struct net *net, int (*func)(struct rt6_info *, void *arg),
1392                     int prune, void *arg)
1393 {
1394         struct fib6_table *table;
1395         struct hlist_node *node;
1396         struct hlist_head *head;
1397         unsigned int h;
1398
1399         rcu_read_lock();
1400         for (h = 0; h < FIB6_TABLE_HASHSZ; h++) {
1401                 head = &net->ipv6.fib_table_hash[h];
1402                 hlist_for_each_entry_rcu(table, node, head, tb6_hlist) {
1403                         write_lock_bh(&table->tb6_lock);
1404                         fib6_clean_tree(net, &table->tb6_root,
1405                                         func, prune, arg);
1406                         write_unlock_bh(&table->tb6_lock);
1407                 }
1408         }
1409         rcu_read_unlock();
1410 }
1411
1412 static int fib6_prune_clone(struct rt6_info *rt, void *arg)
1413 {
1414         if (rt->rt6i_flags & RTF_CACHE) {
1415                 RT6_TRACE("pruning clone %p\n", rt);
1416                 return -1;
1417         }
1418
1419         return 0;
1420 }
1421
1422 static void fib6_prune_clones(struct net *net, struct fib6_node *fn,
1423                               struct rt6_info *rt)
1424 {
1425         fib6_clean_tree(net, fn, fib6_prune_clone, 1, rt);
1426 }
1427
1428 /*
1429  *      Garbage collection
1430  */
1431
1432 static struct fib6_gc_args
1433 {
1434         int                     timeout;
1435         int                     more;
1436 } gc_args;
1437
1438 static int fib6_age(struct rt6_info *rt, void *arg)
1439 {
1440         unsigned long now = jiffies;
1441
1442         /*
1443          *      check addrconf expiration here.
1444          *      Routes are expired even if they are in use.
1445          *
1446          *      Also age clones. Note, that clones are aged out
1447          *      only if they are not in use now.
1448          */
1449
1450         if (rt->rt6i_flags&RTF_EXPIRES && rt->rt6i_expires) {
1451                 if (time_after(now, rt->rt6i_expires)) {
1452                         RT6_TRACE("expiring %p\n", rt);
1453                         return -1;
1454                 }
1455                 gc_args.more++;
1456         } else if (rt->rt6i_flags & RTF_CACHE) {
1457                 if (atomic_read(&rt->dst.__refcnt) == 0 &&
1458                     time_after_eq(now, rt->dst.lastuse + gc_args.timeout)) {
1459                         RT6_TRACE("aging clone %p\n", rt);
1460                         return -1;
1461                 } else if ((rt->rt6i_flags & RTF_GATEWAY) &&
1462                            (!(dst_get_neighbour_raw(&rt->dst)->flags & NTF_ROUTER))) {
1463                         RT6_TRACE("purging route %p via non-router but gateway\n",
1464                                   rt);
1465                         return -1;
1466                 }
1467                 gc_args.more++;
1468         }
1469
1470         return 0;
1471 }
1472
1473 static DEFINE_SPINLOCK(fib6_gc_lock);
1474
1475 void fib6_run_gc(unsigned long expires, struct net *net)
1476 {
1477         if (expires != ~0UL) {
1478                 spin_lock_bh(&fib6_gc_lock);
1479                 gc_args.timeout = expires ? (int)expires :
1480                         net->ipv6.sysctl.ip6_rt_gc_interval;
1481         } else {
1482                 if (!spin_trylock_bh(&fib6_gc_lock)) {
1483                         mod_timer(&net->ipv6.ip6_fib_timer, jiffies + HZ);
1484                         return;
1485                 }
1486                 gc_args.timeout = net->ipv6.sysctl.ip6_rt_gc_interval;
1487         }
1488
1489         gc_args.more = icmp6_dst_gc();
1490
1491         fib6_clean_all(net, fib6_age, 0, NULL);
1492
1493         if (gc_args.more)
1494                 mod_timer(&net->ipv6.ip6_fib_timer,
1495                           round_jiffies(jiffies
1496                                         + net->ipv6.sysctl.ip6_rt_gc_interval));
1497         else
1498                 del_timer(&net->ipv6.ip6_fib_timer);
1499         spin_unlock_bh(&fib6_gc_lock);
1500 }
1501
1502 static void fib6_gc_timer_cb(unsigned long arg)
1503 {
1504         fib6_run_gc(0, (struct net *)arg);
1505 }
1506
1507 static int __net_init fib6_net_init(struct net *net)
1508 {
1509         size_t size = sizeof(struct hlist_head) * FIB6_TABLE_HASHSZ;
1510
1511         setup_timer(&net->ipv6.ip6_fib_timer, fib6_gc_timer_cb, (unsigned long)net);
1512
1513         net->ipv6.rt6_stats = kzalloc(sizeof(*net->ipv6.rt6_stats), GFP_KERNEL);
1514         if (!net->ipv6.rt6_stats)
1515                 goto out_timer;
1516
1517         /* Avoid false sharing : Use at least a full cache line */
1518         size = max_t(size_t, size, L1_CACHE_BYTES);
1519
1520         net->ipv6.fib_table_hash = kzalloc(size, GFP_KERNEL);
1521         if (!net->ipv6.fib_table_hash)
1522                 goto out_rt6_stats;
1523
1524         net->ipv6.fib6_main_tbl = kzalloc(sizeof(*net->ipv6.fib6_main_tbl),
1525                                           GFP_KERNEL);
1526         if (!net->ipv6.fib6_main_tbl)
1527                 goto out_fib_table_hash;
1528
1529         net->ipv6.fib6_main_tbl->tb6_id = RT6_TABLE_MAIN;
1530         net->ipv6.fib6_main_tbl->tb6_root.leaf = net->ipv6.ip6_null_entry;
1531         net->ipv6.fib6_main_tbl->tb6_root.fn_flags =
1532                 RTN_ROOT | RTN_TL_ROOT | RTN_RTINFO;
1533
1534 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
1535         net->ipv6.fib6_local_tbl = kzalloc(sizeof(*net->ipv6.fib6_local_tbl),
1536                                            GFP_KERNEL);
1537         if (!net->ipv6.fib6_local_tbl)
1538                 goto out_fib6_main_tbl;
1539         net->ipv6.fib6_local_tbl->tb6_id = RT6_TABLE_LOCAL;
1540         net->ipv6.fib6_local_tbl->tb6_root.leaf = net->ipv6.ip6_null_entry;
1541         net->ipv6.fib6_local_tbl->tb6_root.fn_flags =
1542                 RTN_ROOT | RTN_TL_ROOT | RTN_RTINFO;
1543 #endif
1544         fib6_tables_init(net);
1545
1546         return 0;
1547
1548 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
1549 out_fib6_main_tbl:
1550         kfree(net->ipv6.fib6_main_tbl);
1551 #endif
1552 out_fib_table_hash:
1553         kfree(net->ipv6.fib_table_hash);
1554 out_rt6_stats:
1555         kfree(net->ipv6.rt6_stats);
1556 out_timer:
1557         return -ENOMEM;
1558  }
1559
1560 static void fib6_net_exit(struct net *net)
1561 {
1562         rt6_ifdown(net, NULL);
1563         del_timer_sync(&net->ipv6.ip6_fib_timer);
1564
1565 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
1566         kfree(net->ipv6.fib6_local_tbl);
1567 #endif
1568         kfree(net->ipv6.fib6_main_tbl);
1569         kfree(net->ipv6.fib_table_hash);
1570         kfree(net->ipv6.rt6_stats);
1571 }
1572
1573 static struct pernet_operations fib6_net_ops = {
1574         .init = fib6_net_init,
1575         .exit = fib6_net_exit,
1576 };
1577
1578 int __init fib6_init(void)
1579 {
1580         int ret = -ENOMEM;
1581
1582         fib6_node_kmem = kmem_cache_create("fib6_nodes",
1583                                            sizeof(struct fib6_node),
1584                                            0, SLAB_HWCACHE_ALIGN,
1585                                            NULL);
1586         if (!fib6_node_kmem)
1587                 goto out;
1588
1589         ret = register_pernet_subsys(&fib6_net_ops);
1590         if (ret)
1591                 goto out_kmem_cache_create;
1592
1593         ret = __rtnl_register(PF_INET6, RTM_GETROUTE, NULL, inet6_dump_fib,
1594                               NULL);
1595         if (ret)
1596                 goto out_unregister_subsys;
1597 out:
1598         return ret;
1599
1600 out_unregister_subsys:
1601         unregister_pernet_subsys(&fib6_net_ops);
1602 out_kmem_cache_create:
1603         kmem_cache_destroy(fib6_node_kmem);
1604         goto out;
1605 }
1606
1607 void fib6_gc_cleanup(void)
1608 {
1609         unregister_pernet_subsys(&fib6_net_ops);
1610         kmem_cache_destroy(fib6_node_kmem);
1611 }