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 void node_free_immediate(struct fib6_node *fn)
 157 {
 158         kmem_cache_free(fib6_node_kmem, fn);
 159 }
 160
 161 static void node_free_rcu(struct rcu_head *head)
 162 {
 163         struct fib6_node *fn = container_of(head, struct fib6_node, rcu);
 164
 165         kmem_cache_free(fib6_node_kmem, fn);
 166 }
 167
 168 static void node_free(struct fib6_node *fn)
 169 {
 170         call_rcu(&fn->rcu, node_free_rcu);
 171 }
 172
 173 static __inline__ void rt6_release(struct rt6_info *rt)
 174 {
 175         if (atomic_dec_and_test(&rt->rt6i_ref))
 176                 dst_free(&rt->dst);
 177 }
 178
 179 static void fib6_link_table(struct net *net, struct fib6_table *tb)
 180 {
 181         unsigned int h;
 182
 183         /*
 184          * Initialize table lock at a single place to give lockdep a key,
 185          * tables aren't visible prior to being linked to the list.
 186          */
 187         rwlock_init(&tb->tb6_lock);
 188
 189         h = tb->tb6_id & (FIB6_TABLE_HASHSZ - 1);
 190
 191         /*
 192          * No protection necessary, this is the only list mutatation
 193          * operation, tables never disappear once they exist.
 194          */
 195         hlist_add_head_rcu(&tb->tb6_hlist, &net->ipv6.fib_table_hash[h]);
 196 }
 197
 198 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
 199
 200 static struct fib6_table *fib6_alloc_table(struct net *net, u32 id)
 201 {
 202         struct fib6_table *table;
 203
 204         table = kzalloc(sizeof(*table), GFP_ATOMIC);
 205         if (table != NULL) {
 206                 table->tb6_id = id;
 207                 table->tb6_root.leaf = net->ipv6.ip6_null_entry;
 208                 table->tb6_root.fn_flags = RTN_ROOT | RTN_TL_ROOT | RTN_RTINFO;
 209         }
 210
 211         return table;
 212 }
 213
 214 struct fib6_table *fib6_new_table(struct net *net, u32 id)
 215 {
 216         struct fib6_table *tb;
 217
 218         if (id == 0)
 219                 id = RT6_TABLE_MAIN;
 220         tb = fib6_get_table(net, id);
 221         if (tb)
 222                 return tb;
 223
 224         tb = fib6_alloc_table(net, id);
 225         if (tb != NULL)
 226                 fib6_link_table(net, tb);
 227
 228         return tb;
 229 }
 230
 231 struct fib6_table *fib6_get_table(struct net *net, u32 id)
 232 {
 233         struct fib6_table *tb;
 234         struct hlist_head *head;
 235         struct hlist_node *node;
 236         unsigned int h;
 237
 238         if (id == 0)
 239                 id = RT6_TABLE_MAIN;
 240         h = id & (FIB6_TABLE_HASHSZ - 1);
 241         rcu_read_lock();
 242         head = &net->ipv6.fib_table_hash[h];
 243         hlist_for_each_entry_rcu(tb, node, head, tb6_hlist) {
 244                 if (tb->tb6_id == id) {
 245                         rcu_read_unlock();
 246                         return tb;
 247                 }
 248         }
 249         rcu_read_unlock();
 250
 251         return NULL;
 252 }
 253
 254 static void __net_init fib6_tables_init(struct net *net)
 255 {
 256         fib6_link_table(net, net->ipv6.fib6_main_tbl);
 257         fib6_link_table(net, net->ipv6.fib6_local_tbl);
 258 }
 259 #else
 260
 261 struct fib6_table *fib6_new_table(struct net *net, u32 id)
 262 {
 263         return fib6_get_table(net, id);
 264 }
 265
 266 struct fib6_table *fib6_get_table(struct net *net, u32 id)
 267 {
 268           return net->ipv6.fib6_main_tbl;
 269 }
 270
 271 struct dst_entry *fib6_rule_lookup(struct net *net, struct flowi6 *fl6,
 272                                    int flags, pol_lookup_t lookup)
 273 {
 274         return (struct dst_entry *) lookup(net, net->ipv6.fib6_main_tbl, fl6, flags);
 275 }
 276
 277 static void __net_init fib6_tables_init(struct net *net)
 278 {
 279         fib6_link_table(net, net->ipv6.fib6_main_tbl);
 280 }
 281
 282 #endif
 283
 284 static int fib6_dump_node(struct fib6_walker_t *w)
 285 {
 286         int res;
 287         struct rt6_info *rt;
 288
 289         for (rt = w->leaf; rt; rt = rt->dst.rt6_next) {
 290                 res = rt6_dump_route(rt, w->args);
 291                 if (res < 0) {
 292                         /* Frame is full, suspend walking */
 293                         w->leaf = rt;
 294                         return 1;
 295                 }
 296                 WARN_ON(res == 0);
 297         }
 298         w->leaf = NULL;
 299         return 0;
 300 }
 301
 302 static void fib6_dump_end(struct netlink_callback *cb)
 303 {
 304         struct fib6_walker_t *w = (void*)cb->args[2];
 305
 306         if (w) {
 307                 if (cb->args[4]) {
 308                         cb->args[4] = 0;
 309                         fib6_walker_unlink(w);
 310                 }
 311                 cb->args[2] = 0;
 312                 kfree(w);
 313         }
 314         cb->done = (void*)cb->args[3];
 315         cb->args[1] = 3;
 316 }
 317
 318 static int fib6_dump_done(struct netlink_callback *cb)
 319 {
 320         fib6_dump_end(cb);
 321         return cb->done ? cb->done(cb) : 0;
 322 }
 323
 324 static int fib6_dump_table(struct fib6_table *table, struct sk_buff *skb,
 325                            struct netlink_callback *cb)
 326 {
 327         struct fib6_walker_t *w;
 328         int res;
 329
 330         w = (void *)cb->args[2];
 331         w->root = &table->tb6_root;
 332
 333         if (cb->args[4] == 0) {
 334                 w->count = 0;
 335                 w->skip = 0;
 336
 337                 read_lock_bh(&table->tb6_lock);
 338                 res = fib6_walk(w);
 339                 read_unlock_bh(&table->tb6_lock);
 340                 if (res > 0) {
 341                         cb->args[4] = 1;
 342                         cb->args[5] = w->root->fn_sernum;
 343                 }
 344         } else {
 345                 if (cb->args[5] != w->root->fn_sernum) {
 346                         /* Begin at the root if the tree changed */
 347                         cb->args[5] = w->root->fn_sernum;
 348                         w->state = FWS_INIT;
 349                         w->node = w->root;
 350                         w->skip = w->count;
 351                 } else
 352                         w->skip = 0;
 353
 354                 read_lock_bh(&table->tb6_lock);
 355                 res = fib6_walk_continue(w);
 356                 read_unlock_bh(&table->tb6_lock);
 357                 if (res <= 0) {
 358                         fib6_walker_unlink(w);
 359                         cb->args[4] = 0;
 360                 }
 361         }
 362
 363         return res;
 364 }
 365
 366 static int inet6_dump_fib(struct sk_buff *skb, struct netlink_callback *cb)
 367 {
 368         struct net *net = sock_net(skb->sk);
 369         unsigned int h, s_h;
 370         unsigned int e = 0, s_e;
 371         struct rt6_rtnl_dump_arg arg;
 372         struct fib6_walker_t *w;
 373         struct fib6_table *tb;
 374         struct hlist_node *node;
 375         struct hlist_head *head;
 376         int res = 0;
 377
 378         s_h = cb->args[0];
 379         s_e = cb->args[1];
 380
 381         w = (void *)cb->args[2];
 382         if (w == NULL) {
 383                 /* New dump:
 384                  *
 385                  * 1. hook callback destructor.
 386                  */
 387                 cb->args[3] = (long)cb->done;
 388                 cb->done = fib6_dump_done;
 389
 390                 /*
 391                  * 2. allocate and initialize walker.
 392                  */
 393                 w = kzalloc(sizeof(*w), GFP_ATOMIC);
 394                 if (w == NULL)
 395                         return -ENOMEM;
 396                 w->func = fib6_dump_node;
 397                 cb->args[2] = (long)w;
 398         }
 399
 400         arg.skb = skb;
 401         arg.cb = cb;
 402         arg.net = net;
 403         w->args = &arg;
 404
 405         rcu_read_lock();
 406         for (h = s_h; h < FIB6_TABLE_HASHSZ; h++, s_e = 0) {
 407                 e = 0;
 408                 head = &net->ipv6.fib_table_hash[h];
 409                 hlist_for_each_entry_rcu(tb, node, head, tb6_hlist) {
 410                         if (e < s_e)
 411                                 goto next;
 412                         res = fib6_dump_table(tb, skb, cb);
 413                         if (res != 0)
 414                                 goto out;
 415 next:
 416                         e++;
 417                 }
 418         }
 419 out:
 420         rcu_read_unlock();
 421         cb->args[1] = e;
 422         cb->args[0] = h;
 423
 424         res = res < 0 ? res : skb->len;
 425         if (res <= 0)
 426                 fib6_dump_end(cb);
 427         return res;
 428 }
 429
 430 /*
 431  *      Routing Table
 432  *
 433  *      return the appropriate node for a routing tree "add" operation
 434  *      by either creating and inserting or by returning an existing
 435  *      node.
 436  */
 437
 438 static struct fib6_node * fib6_add_1(struct fib6_node *root, void *addr,
 439                                      int addrlen, int plen,
 440                                      int offset)
 441 {
 442         struct fib6_node *fn, *in, *ln;
 443         struct fib6_node *pn = NULL;
 444         struct rt6key *key;
 445         int     bit;
 446         __be32  dir = 0;
 447         __u32   sernum = fib6_new_sernum();
 448
 449         RT6_TRACE("fib6_add_1\n");
 450
 451         /* insert node in tree */
 452
 453         fn = root;
 454
 455         do {
 456                 key = (struct rt6key *)((u8 *)fn->leaf + offset);
 457
 458                 /*
 459                  *      Prefix match
 460                  */
 461                 if (plen < fn->fn_bit ||
 462                     !ipv6_prefix_equal(&key->addr, addr, fn->fn_bit))
 463                         goto insert_above;
 464
 465                 /*
 466                  *      Exact match ?
 467                  */
 468
 469                 if (plen == fn->fn_bit) {
 470                         /* clean up an intermediate node */
 471                         if ((fn->fn_flags & RTN_RTINFO) == 0) {
 472                                 rt6_release(fn->leaf);
 473                                 fn->leaf = NULL;
 474                         }
 475
 476                         fn->fn_sernum = sernum;
 477
 478                         return fn;
 479                 }
 480
 481                 /*
 482                  *      We have more bits to go
 483                  */
 484
 485                 /* Try to walk down on tree. */
 486                 fn->fn_sernum = sernum;
 487                 dir = addr_bit_set(addr, fn->fn_bit);
 488                 pn = fn;
 489                 fn = dir ? fn->right: fn->left;
 490         } while (fn);
 491
 492         /*
 493          *      We walked to the bottom of tree.
 494          *      Create new leaf node without children.
 495          */
 496
 497         ln = node_alloc();
 498
 499         if (ln == NULL)
 500                 return NULL;
 501         ln->fn_bit = plen;
 502
 503         ln->parent = pn;
 504         ln->fn_sernum = sernum;
 505
 506         if (dir)
 507                 pn->right = ln;
 508         else
 509                 pn->left  = ln;
 510
 511         return ln;
 512
 513
 514 insert_above:
 515         /*
 516          * split since we don't have a common prefix anymore or
 517          * we have a less significant route.
 518          * we've to insert an intermediate node on the list
 519          * this new node will point to the one we need to create
 520          * and the current
 521          */
 522
 523         pn = fn->parent;
 524
 525         /* find 1st bit in difference between the 2 addrs.
 526
 527            See comment in __ipv6_addr_diff: bit may be an invalid value,
 528            but if it is >= plen, the value is ignored in any case.
 529          */
 530
 531         bit = __ipv6_addr_diff(addr, &key->addr, addrlen);
 532
 533         /*
 534          *              (intermediate)[in]
 535          *                /        \
 536          *      (new leaf node)[ln] (old node)[fn]
 537          */
 538         if (plen > bit) {
 539                 in = node_alloc();
 540                 ln = node_alloc();
 541
 542                 if (in == NULL || ln == NULL) {
 543                         if (in)
 544                                 node_free_immediate(in);
 545                         if (ln)
 546                                 node_free_immediate(ln);
 547                         return NULL;
 548                 }
 549
 550                 /*
 551                  * new intermediate node.
 552                  * RTN_RTINFO will
 553                  * be off since that an address that chooses one of
 554                  * the branches would not match less specific routes
 555                  * in the other branch
 556                  */
 557
 558                 in->fn_bit = bit;
 559
 560                 in->parent = pn;
 561                 in->leaf = fn->leaf;
 562                 atomic_inc(&in->leaf->rt6i_ref);
 563
 564                 in->fn_sernum = sernum;
 565
 566                 /* update parent pointer */
 567                 if (dir)
 568                         pn->right = in;
 569                 else
 570                         pn->left  = in;
 571
 572                 ln->fn_bit = plen;
 573
 574                 ln->parent = in;
 575                 fn->parent = in;
 576
 577                 ln->fn_sernum = sernum;
 578
 579                 if (addr_bit_set(addr, bit)) {
 580                         in->right = ln;
 581                         in->left  = fn;
 582                 } else {
 583                         in->left  = ln;
 584                         in->right = fn;
 585                 }
 586         } else { /* plen <= bit */
 587
 588                 /*
 589                  *              (new leaf node)[ln]
 590                  *                /        \
 591                  *           (old node)[fn] NULL
 592                  */
 593
 594                 ln = node_alloc();
 595
 596                 if (ln == NULL)
 597                         return NULL;
 598
 599                 ln->fn_bit = plen;
 600
 601                 ln->parent = pn;
 602
 603                 ln->fn_sernum = sernum;
 604
 605                 if (dir)
 606                         pn->right = ln;
 607                 else
 608                         pn->left  = ln;
 609
 610                 if (addr_bit_set(&key->addr, plen))
 611                         ln->right = fn;
 612                 else
 613                         ln->left  = fn;
 614
 615                 fn->parent = ln;
 616         }
 617         return ln;
 618 }
 619
 620 /*
 621  *      Insert routing information in a node.
 622  */
 623
 624 static int fib6_add_rt2node(struct fib6_node *fn, struct rt6_info *rt,
 625                             struct nl_info *info)
 626 {
 627         struct rt6_info *iter = NULL;
 628         struct rt6_info **ins;
 629
 630         ins = &fn->leaf;
 631
 632         for (iter = fn->leaf; iter; iter=iter->dst.rt6_next) {
 633                 /*
 634                  *      Search for duplicates
 635                  */
 636
 637                 if (iter->rt6i_metric == rt->rt6i_metric) {
 638                         /*
 639                          *      Same priority level
 640                          */
 641
 642                         if (iter->rt6i_dev == rt->rt6i_dev &&
 643                             iter->rt6i_idev == rt->rt6i_idev &&
 644                             ipv6_addr_equal(&iter->rt6i_gateway,
 645                                             &rt->rt6i_gateway)) {
 646                                 if (!(iter->rt6i_flags&RTF_EXPIRES))
 647                                         return -EEXIST;
 648                                 iter->rt6i_expires = rt->rt6i_expires;
 649                                 if (!(rt->rt6i_flags&RTF_EXPIRES)) {
 650                                         iter->rt6i_flags &= ~RTF_EXPIRES;
 651                                         iter->rt6i_expires = 0;
 652                                 }
 653                                 return -EEXIST;
 654                         }
 655                 }
 656
 657                 if (iter->rt6i_metric > rt->rt6i_metric)
 658                         break;
 659
 660                 ins = &iter->dst.rt6_next;
 661         }
 662
 663         /* Reset round-robin state, if necessary */
 664         if (ins == &fn->leaf)
 665                 fn->rr_ptr = NULL;
 666
 667         /*
 668          *      insert node
 669          */
 670
 671         rt->dst.rt6_next = iter;
 672         *ins = rt;
 673         rcu_assign_pointer(rt->rt6i_node, fn);
 674         atomic_inc(&rt->rt6i_ref);
 675         inet6_rt_notify(RTM_NEWROUTE, rt, info);
 676         info->nl_net->ipv6.rt6_stats->fib_rt_entries++;
 677
 678         if ((fn->fn_flags & RTN_RTINFO) == 0) {
 679                 info->nl_net->ipv6.rt6_stats->fib_route_nodes++;
 680                 fn->fn_flags |= RTN_RTINFO;
 681         }
 682
 683         return 0;
 684 }
 685
 686 static __inline__ void fib6_start_gc(struct net *net, struct rt6_info *rt)
 687 {
 688         if (!timer_pending(&net->ipv6.ip6_fib_timer) &&
 689             (rt->rt6i_flags & (RTF_EXPIRES|RTF_CACHE)))
 690                 mod_timer(&net->ipv6.ip6_fib_timer,
 691                           jiffies + net->ipv6.sysctl.ip6_rt_gc_interval);
 692 }
 693
 694 void fib6_force_start_gc(struct net *net)
 695 {
 696         if (!timer_pending(&net->ipv6.ip6_fib_timer))
 697                 mod_timer(&net->ipv6.ip6_fib_timer,
 698                           jiffies + net->ipv6.sysctl.ip6_rt_gc_interval);
 699 }
 700
 701 /*
 702  *      Add routing information to the routing tree.
 703  *      <destination addr>/<source addr>
 704  *      with source addr info in sub-trees
 705  */
 706
 707 int fib6_add(struct fib6_node *root, struct rt6_info *rt, struct nl_info *info)
 708 {
 709         struct fib6_node *fn, *pn = NULL;
 710         int err = -ENOMEM;
 711
 712         fn = fib6_add_1(root, &rt->rt6i_dst.addr, sizeof(struct in6_addr),
 713                         rt->rt6i_dst.plen, offsetof(struct rt6_info, rt6i_dst));
 714
 715         if (fn == NULL)
 716                 goto out;
 717
 718         pn = fn;
 719
 720 #ifdef CONFIG_IPV6_SUBTREES
 721         if (rt->rt6i_src.plen) {
 722                 struct fib6_node *sn;
 723
 724                 if (fn->subtree == NULL) {
 725                         struct fib6_node *sfn;
 726
 727                         /*
 728                          * Create subtree.
 729                          *
 730                          *              fn[main tree]
 731                          *              |
 732                          *              sfn[subtree root]
 733                          *                 \
 734                          *                  sn[new leaf node]
 735                          */
 736
 737                         /* Create subtree root node */
 738                         sfn = node_alloc();
 739                         if (sfn == NULL)
 740                                 goto st_failure;
 741
 742                         sfn->leaf = info->nl_net->ipv6.ip6_null_entry;
 743                         atomic_inc(&info->nl_net->ipv6.ip6_null_entry->rt6i_ref);
 744                         sfn->fn_flags = RTN_ROOT;
 745                         sfn->fn_sernum = fib6_new_sernum();
 746
 747                         /* Now add the first leaf node to new subtree */
 748
 749                         sn = fib6_add_1(sfn, &rt->rt6i_src.addr,
 750                                         sizeof(struct in6_addr), rt->rt6i_src.plen,
 751                                         offsetof(struct rt6_info, rt6i_src));
 752
 753                         if (sn == NULL) {
 754                                 /* If it is failed, discard just allocated
 755                                    root, and then (in st_failure) stale node
 756                                    in main tree.
 757                                  */
 758                                 node_free_immediate(sfn);
 759                                 goto st_failure;
 760                         }
 761
 762                         /* Now link new subtree to main tree */
 763                         sfn->parent = fn;
 764                         fn->subtree = sfn;
 765                 } else {
 766                         sn = fib6_add_1(fn->subtree, &rt->rt6i_src.addr,
 767                                         sizeof(struct in6_addr), rt->rt6i_src.plen,
 768                                         offsetof(struct rt6_info, rt6i_src));
 769
 770                         if (sn == NULL)
 771                                 goto st_failure;
 772                 }
 773
 774                 if (fn->leaf == NULL) {
 775                         fn->leaf = rt;
 776                         atomic_inc(&rt->rt6i_ref);
 777                 }
 778                 fn = sn;
 779         }
 780 #endif
 781
 782         err = fib6_add_rt2node(fn, rt, info);
 783
 784         if (err == 0) {
 785                 fib6_start_gc(info->nl_net, rt);
 786                 if (!(rt->rt6i_flags&RTF_CACHE))
 787                         fib6_prune_clones(info->nl_net, pn, rt);
 788         }
 789
 790 out:
 791         if (err) {
 792 #ifdef CONFIG_IPV6_SUBTREES
 793                 /*
 794                  * If fib6_add_1 has cleared the old leaf pointer in the
 795                  * super-tree leaf node we have to find a new one for it.
 796                  */
 797                 if (pn != fn && pn->leaf == rt) {
 798                         pn->leaf = NULL;
 799                         atomic_dec(&rt->rt6i_ref);
 800                 }
 801                 if (pn != fn && !pn->leaf && !(pn->fn_flags & RTN_RTINFO)) {
 802                         pn->leaf = fib6_find_prefix(info->nl_net, pn);
 803 #if RT6_DEBUG >= 2
 804                         if (!pn->leaf) {
 805                                 WARN_ON(pn->leaf == NULL);
 806                                 pn->leaf = info->nl_net->ipv6.ip6_null_entry;
 807                         }
 808 #endif
 809                         atomic_inc(&pn->leaf->rt6i_ref);
 810                 }
 811 #endif
 812                 dst_free(&rt->dst);
 813         }
 814         return err;
 815
 816 #ifdef CONFIG_IPV6_SUBTREES
 817         /* Subtree creation failed, probably main tree node
 818            is orphan. If it is, shoot it.
 819          */
 820 st_failure:
 821         if (fn && !(fn->fn_flags & (RTN_RTINFO|RTN_ROOT)))
 822                 fib6_repair_tree(info->nl_net, fn);
 823         dst_free(&rt->dst);
 824         return err;
 825 #endif
 826 }
 827
 828 /*
 829  *      Routing tree lookup
 830  *
 831  */
 832
 833 struct lookup_args {
 834         int             offset;         /* key offset on rt6_info       */
 835         const struct in6_addr   *addr;          /* search key                   */
 836 };
 837
 838 static struct fib6_node * fib6_lookup_1(struct fib6_node *root,
 839                                         struct lookup_args *args)
 840 {
 841         struct fib6_node *fn;
 842         __be32 dir;
 843
 844         if (unlikely(args->offset == 0))
 845                 return NULL;
 846
 847         /*
 848          *      Descend on a tree
 849          */
 850
 851         fn = root;
 852
 853         for (;;) {
 854                 struct fib6_node *next;
 855
 856                 dir = addr_bit_set(args->addr, fn->fn_bit);
 857
 858                 next = dir ? fn->right : fn->left;
 859
 860                 if (next) {
 861                         fn = next;
 862                         continue;
 863                 }
 864
 865                 break;
 866         }
 867
 868         while(fn) {
 869                 if (FIB6_SUBTREE(fn) || fn->fn_flags & RTN_RTINFO) {
 870                         struct rt6key *key;
 871
 872                         key = (struct rt6key *) ((u8 *) fn->leaf +
 873                                                  args->offset);
 874
 875                         if (ipv6_prefix_equal(&key->addr, args->addr, key->plen)) {
 876 #ifdef CONFIG_IPV6_SUBTREES
 877                                 if (fn->subtree) {
 878                                         struct fib6_node *sfn;
 879                                         sfn = fib6_lookup_1(fn->subtree,
 880                                                             args + 1);
 881                                         if (!sfn)
 882                                                 goto backtrack;
 883                                         fn = sfn;
 884                                 }
 885 #endif
 886                                 if (fn->fn_flags & RTN_RTINFO)
 887                                         return fn;
 888                         }
 889                 }
 890 #ifdef CONFIG_IPV6_SUBTREES
 891 backtrack:
 892 #endif
 893                 if (fn->fn_flags & RTN_ROOT)
 894                         break;
 895
 896                 fn = fn->parent;
 897         }
 898
 899         return NULL;
 900 }
 901
 902 struct fib6_node * fib6_lookup(struct fib6_node *root, const struct in6_addr *daddr,
 903                                const struct in6_addr *saddr)
 904 {
 905         struct fib6_node *fn;
 906         struct lookup_args args[] = {
 907                 {
 908                         .offset = offsetof(struct rt6_info, rt6i_dst),
 909                         .addr = daddr,
 910                 },
 911 #ifdef CONFIG_IPV6_SUBTREES
 912                 {
 913                         .offset = offsetof(struct rt6_info, rt6i_src),
 914                         .addr = saddr,
 915                 },
 916 #endif
 917                 {
 918                         .offset = 0,    /* sentinel */
 919                 }
 920         };
 921
 922         fn = fib6_lookup_1(root, daddr ? args : args + 1);
 923
 924         if (fn == NULL || fn->fn_flags & RTN_TL_ROOT)
 925                 fn = root;
 926
 927         return fn;
 928 }
 929
 930 /*
 931  *      Get node with specified destination prefix (and source prefix,
 932  *      if subtrees are used)
 933  */
 934
 935
 936 static struct fib6_node * fib6_locate_1(struct fib6_node *root,
 937                                         const struct in6_addr *addr,
 938                                         int plen, int offset)
 939 {
 940         struct fib6_node *fn;
 941
 942         for (fn = root; fn ; ) {
 943                 struct rt6key *key = (struct rt6key *)((u8 *)fn->leaf + offset);
 944
 945                 /*
 946                  *      Prefix match
 947                  */
 948                 if (plen < fn->fn_bit ||
 949                     !ipv6_prefix_equal(&key->addr, addr, fn->fn_bit))
 950                         return NULL;
 951
 952                 if (plen == fn->fn_bit)
 953                         return fn;
 954
 955                 /*
 956                  *      We have more bits to go
 957                  */
 958                 if (addr_bit_set(addr, fn->fn_bit))
 959                         fn = fn->right;
 960                 else
 961                         fn = fn->left;
 962         }
 963         return NULL;
 964 }
 965
 966 struct fib6_node * fib6_locate(struct fib6_node *root,
 967                                const struct in6_addr *daddr, int dst_len,
 968                                const struct in6_addr *saddr, int src_len)
 969 {
 970         struct fib6_node *fn;
 971
 972         fn = fib6_locate_1(root, daddr, dst_len,
 973                            offsetof(struct rt6_info, rt6i_dst));
 974
 975 #ifdef CONFIG_IPV6_SUBTREES
 976         if (src_len) {
 977                 WARN_ON(saddr == NULL);
 978                 if (fn && fn->subtree)
 979                         fn = fib6_locate_1(fn->subtree, saddr, src_len,
 980                                            offsetof(struct rt6_info, rt6i_src));
 981         }
 982 #endif
 983
 984         if (fn && fn->fn_flags&RTN_RTINFO)
 985                 return fn;
 986
 987         return NULL;
 988 }
 989
 990
 991 /*
 992  *      Deletion
 993  *
 994  */
 995
 996 static struct rt6_info *fib6_find_prefix(struct net *net, struct fib6_node *fn)
 997 {
 998         if (fn->fn_flags&RTN_ROOT)
 999                 return net->ipv6.ip6_null_entry;
1000
1001         while(fn) {
1002                 if(fn->left)
1003                         return fn->left->leaf;
1004
1005                 if(fn->right)
1006                         return fn->right->leaf;
1007
1008                 fn = FIB6_SUBTREE(fn);
1009         }
1010         return NULL;
1011 }
1012
1013 /*
1014  *      Called to trim the tree of intermediate nodes when possible. "fn"
1015  *      is the node we want to try and remove.
1016  */
1017
1018 static struct fib6_node *fib6_repair_tree(struct net *net,
1019                                            struct fib6_node *fn)
1020 {
1021         int children;
1022         int nstate;
1023         struct fib6_node *child, *pn;
1024         struct fib6_walker_t *w;
1025         int iter = 0;
1026
1027         for (;;) {
1028                 RT6_TRACE("fixing tree: plen=%d iter=%d\n", fn->fn_bit, iter);
1029                 iter++;
1030
1031                 WARN_ON(fn->fn_flags & RTN_RTINFO);
1032                 WARN_ON(fn->fn_flags & RTN_TL_ROOT);
1033                 WARN_ON(fn->leaf != NULL);
1034
1035                 children = 0;
1036                 child = NULL;
1037                 if (fn->right) child = fn->right, children |= 1;
1038                 if (fn->left) child = fn->left, children |= 2;
1039
1040                 if (children == 3 || FIB6_SUBTREE(fn)
1041 #ifdef CONFIG_IPV6_SUBTREES
1042                     /* Subtree root (i.e. fn) may have one child */
1043                     || (children && fn->fn_flags&RTN_ROOT)
1044 #endif
1045                     ) {
1046                         fn->leaf = fib6_find_prefix(net, fn);
1047 #if RT6_DEBUG >= 2
1048                         if (fn->leaf==NULL) {
1049                                 WARN_ON(!fn->leaf);
1050                                 fn->leaf = net->ipv6.ip6_null_entry;
1051                         }
1052 #endif
1053                         atomic_inc(&fn->leaf->rt6i_ref);
1054                         return fn->parent;
1055                 }
1056
1057                 pn = fn->parent;
1058 #ifdef CONFIG_IPV6_SUBTREES
1059                 if (FIB6_SUBTREE(pn) == fn) {
1060                         WARN_ON(!(fn->fn_flags & RTN_ROOT));
1061                         FIB6_SUBTREE(pn) = NULL;
1062                         nstate = FWS_L;
1063                 } else {
1064                         WARN_ON(fn->fn_flags & RTN_ROOT);
1065 #endif
1066                         if (pn->right == fn) pn->right = child;
1067                         else if (pn->left == fn) pn->left = child;
1068 #if RT6_DEBUG >= 2
1069                         else
1070                                 WARN_ON(1);
1071 #endif
1072                         if (child)
1073                                 child->parent = pn;
1074                         nstate = FWS_R;
1075 #ifdef CONFIG_IPV6_SUBTREES
1076                 }
1077 #endif
1078
1079                 read_lock(&fib6_walker_lock);
1080                 FOR_WALKERS(w) {
1081                         if (child == NULL) {
1082                                 if (w->root == fn) {
1083                                         w->root = w->node = NULL;
1084                                         RT6_TRACE("W %p adjusted by delroot 1\n", w);
1085                                 } else if (w->node == fn) {
1086                                         RT6_TRACE("W %p adjusted by delnode 1, s=%d/%d\n", w, w->state, nstate);
1087                                         w->node = pn;
1088                                         w->state = nstate;
1089                                 }
1090                         } else {
1091                                 if (w->root == fn) {
1092                                         w->root = child;
1093                                         RT6_TRACE("W %p adjusted by delroot 2\n", w);
1094                                 }
1095                                 if (w->node == fn) {
1096                                         w->node = child;
1097                                         if (children&2) {
1098                                                 RT6_TRACE("W %p adjusted by delnode 2, s=%d\n", w, w->state);
1099                                                 w->state = w->state>=FWS_R ? FWS_U : FWS_INIT;
1100                                         } else {
1101                                                 RT6_TRACE("W %p adjusted by delnode 2, s=%d\n", w, w->state);
1102                                                 w->state = w->state>=FWS_C ? FWS_U : FWS_INIT;
1103                                         }
1104                                 }
1105                         }
1106                 }
1107                 read_unlock(&fib6_walker_lock);
1108
1109                 node_free(fn);
1110                 if (pn->fn_flags&RTN_RTINFO || FIB6_SUBTREE(pn))
1111                         return pn;
1112
1113                 rt6_release(pn->leaf);
1114                 pn->leaf = NULL;
1115                 fn = pn;
1116         }
1117 }
1118
1119 static void fib6_del_route(struct fib6_node *fn, struct rt6_info **rtp,
1120                            struct nl_info *info)
1121 {
1122         struct fib6_walker_t *w;
1123         struct rt6_info *rt = *rtp;
1124         struct net *net = info->nl_net;
1125
1126         RT6_TRACE("fib6_del_route\n");
1127
1128         /* Unlink it */
1129         *rtp = rt->dst.rt6_next;
1130         rt->rt6i_node = NULL;
1131         net->ipv6.rt6_stats->fib_rt_entries--;
1132         net->ipv6.rt6_stats->fib_discarded_routes++;
1133
1134         /* Reset round-robin state, if necessary */
1135         if (fn->rr_ptr == rt)
1136                 fn->rr_ptr = NULL;
1137
1138         /* Adjust walkers */
1139         read_lock(&fib6_walker_lock);
1140         FOR_WALKERS(w) {
1141                 if (w->state == FWS_C && w->leaf == rt) {
1142                         RT6_TRACE("walker %p adjusted by delroute\n", w);
1143                         w->leaf = rt->dst.rt6_next;
1144                         if (w->leaf == NULL)
1145                                 w->state = FWS_U;
1146                 }
1147         }
1148         read_unlock(&fib6_walker_lock);
1149
1150         rt->dst.rt6_next = NULL;
1151
1152         /* If it was last route, expunge its radix tree node */
1153         if (fn->leaf == NULL) {
1154                 fn->fn_flags &= ~RTN_RTINFO;
1155                 net->ipv6.rt6_stats->fib_route_nodes--;
1156                 fn = fib6_repair_tree(net, fn);
1157         }
1158
1159         if (atomic_read(&rt->rt6i_ref) != 1) {
1160                 /* This route is used as dummy address holder in some split
1161                  * nodes. It is not leaked, but it still holds other resources,
1162                  * which must be released in time. So, scan ascendant nodes
1163                  * and replace dummy references to this route with references
1164                  * to still alive ones.
1165                  */
1166                 while (fn) {
1167                         if (!(fn->fn_flags&RTN_RTINFO) && fn->leaf == rt) {
1168                                 fn->leaf = fib6_find_prefix(net, fn);
1169                                 atomic_inc(&fn->leaf->rt6i_ref);
1170                                 rt6_release(rt);
1171                         }
1172                         fn = fn->parent;
1173                 }
1174                 /* No more references are possible at this point. */
1175                 BUG_ON(atomic_read(&rt->rt6i_ref) != 1);
1176         }
1177
1178         inet6_rt_notify(RTM_DELROUTE, rt, info);
1179         rt6_release(rt);
1180 }
1181
1182 int fib6_del(struct rt6_info *rt, struct nl_info *info)
1183 {
1184         struct fib6_node *fn = rcu_dereference_protected(rt->rt6i_node,
1185                                     lockdep_is_held(&rt->rt6i_table->tb6_lock));
1186         struct net *net = info->nl_net;
1187         struct rt6_info **rtp;
1188
1189 #if RT6_DEBUG >= 2
1190         if (rt->dst.obsolete>0) {
1191                 WARN_ON(fn != NULL);
1192                 return -ENOENT;
1193         }
1194 #endif
1195         if (fn == NULL || rt == net->ipv6.ip6_null_entry)
1196                 return -ENOENT;
1197
1198         WARN_ON(!(fn->fn_flags & RTN_RTINFO));
1199
1200         if (!(rt->rt6i_flags&RTF_CACHE)) {
1201                 struct fib6_node *pn = fn;
1202 #ifdef CONFIG_IPV6_SUBTREES
1203                 /* clones of this route might be in another subtree */
1204                 if (rt->rt6i_src.plen) {
1205                         while (!(pn->fn_flags&RTN_ROOT))
1206                                 pn = pn->parent;
1207                         pn = pn->parent;
1208                 }
1209 #endif
1210                 fib6_prune_clones(info->nl_net, pn, rt);
1211         }
1212
1213         /*
1214          *      Walk the leaf entries looking for ourself
1215          */
1216
1217         for (rtp = &fn->leaf; *rtp; rtp = &(*rtp)->dst.rt6_next) {
1218                 if (*rtp == rt) {
1219                         fib6_del_route(fn, rtp, info);
1220                         return 0;
1221                 }
1222         }
1223         return -ENOENT;
1224 }
1225
1226 /*
1227  *      Tree traversal function.
1228  *
1229  *      Certainly, it is not interrupt safe.
1230  *      However, it is internally reenterable wrt itself and fib6_add/fib6_del.
1231  *      It means, that we can modify tree during walking
1232  *      and use this function for garbage collection, clone pruning,
1233  *      cleaning tree when a device goes down etc. etc.
1234  *
1235  *      It guarantees that every node will be traversed,
1236  *      and that it will be traversed only once.
1237  *
1238  *      Callback function w->func may return:
1239  *      0 -> continue walking.
1240  *      positive value -> walking is suspended (used by tree dumps,
1241  *      and probably by gc, if it will be split to several slices)
1242  *      negative value -> terminate walking.
1243  *
1244  *      The function itself returns:
1245  *      0   -> walk is complete.
1246  *      >0  -> walk is incomplete (i.e. suspended)
1247  *      <0  -> walk is terminated by an error.
1248  */
1249
1250 static int fib6_walk_continue(struct fib6_walker_t *w)
1251 {
1252         struct fib6_node *fn, *pn;
1253
1254         for (;;) {
1255                 fn = w->node;
1256                 if (fn == NULL)
1257                         return 0;
1258
1259                 if (w->prune && fn != w->root &&
1260                     fn->fn_flags&RTN_RTINFO && w->state < FWS_C) {
1261                         w->state = FWS_C;
1262                         w->leaf = fn->leaf;
1263                 }
1264                 switch (w->state) {
1265 #ifdef CONFIG_IPV6_SUBTREES
1266                 case FWS_S:
1267                         if (FIB6_SUBTREE(fn)) {
1268                                 w->node = FIB6_SUBTREE(fn);
1269                                 continue;
1270                         }
1271                         w->state = FWS_L;
1272 #endif
1273                 case FWS_L:
1274                         if (fn->left) {
1275                                 w->node = fn->left;
1276                                 w->state = FWS_INIT;
1277                                 continue;
1278                         }
1279                         w->state = FWS_R;
1280                 case FWS_R:
1281                         if (fn->right) {
1282                                 w->node = fn->right;
1283                                 w->state = FWS_INIT;
1284                                 continue;
1285                         }
1286                         w->state = FWS_C;
1287                         w->leaf = fn->leaf;
1288                 case FWS_C:
1289                         if (w->leaf && fn->fn_flags&RTN_RTINFO) {
1290                                 int err;
1291
1292                                 if (w->skip) {
1293                                         w->skip--;
1294                                         goto skip;
1295                                 }
1296
1297                                 err = w->func(w);
1298                                 if (err)
1299                                         return err;
1300
1301                                 w->count++;
1302                                 continue;
1303                         }
1304 skip:
1305                         w->state = FWS_U;
1306                 case FWS_U:
1307                         if (fn == w->root)
1308                                 return 0;
1309                         pn = fn->parent;
1310                         w->node = pn;
1311 #ifdef CONFIG_IPV6_SUBTREES
1312                         if (FIB6_SUBTREE(pn) == fn) {
1313                                 WARN_ON(!(fn->fn_flags & RTN_ROOT));
1314                                 w->state = FWS_L;
1315                                 continue;
1316                         }
1317 #endif
1318                         if (pn->left == fn) {
1319                                 w->state = FWS_R;
1320                                 continue;
1321                         }
1322                         if (pn->right == fn) {
1323                                 w->state = FWS_C;
1324                                 w->leaf = w->node->leaf;
1325                                 continue;
1326                         }
1327 #if RT6_DEBUG >= 2
1328                         WARN_ON(1);
1329 #endif
1330                 }
1331         }
1332 }
1333
1334 static int fib6_walk(struct fib6_walker_t *w)
1335 {
1336         int res;
1337
1338         w->state = FWS_INIT;
1339         w->node = w->root;
1340
1341         fib6_walker_link(w);
1342         res = fib6_walk_continue(w);
1343         if (res <= 0)
1344                 fib6_walker_unlink(w);
1345         return res;
1346 }
1347
1348 static int fib6_clean_node(struct fib6_walker_t *w)
1349 {
1350         int res;
1351         struct rt6_info *rt;
1352         struct fib6_cleaner_t *c = container_of(w, struct fib6_cleaner_t, w);
1353         struct nl_info info = {
1354                 .nl_net = c->net,
1355         };
1356
1357         for (rt = w->leaf; rt; rt = rt->dst.rt6_next) {
1358                 res = c->func(rt, c->arg);
1359                 if (res < 0) {
1360                         w->leaf = rt;
1361                         res = fib6_del(rt, &info);
1362                         if (res) {
1363 #if RT6_DEBUG >= 2
1364                                 pr_debug("%s: del failed: rt=%p@%p err=%d\n",
1365                                          __func__, rt,
1366                                          rcu_access_pointer(rt->rt6i_node),
1367                                          res);
1368 #endif
1369                                 continue;
1370                         }
1371                         return 0;
1372                 }
1373                 WARN_ON(res != 0);
1374         }
1375         w->leaf = rt;
1376         return 0;
1377 }
1378
1379 /*
1380  *      Convenient frontend to tree walker.
1381  *
1382  *      func is called on each route.
1383  *              It may return -1 -> delete this route.
1384  *                            0  -> continue walking
1385  *
1386  *      prune==1 -> only immediate children of node (certainly,
1387  *      ignoring pure split nodes) will be scanned.
1388  */
1389
1390 static void fib6_clean_tree(struct net *net, struct fib6_node *root,
1391                             int (*func)(struct rt6_info *, void *arg),
1392                             int prune, void *arg)
1393 {
1394         struct fib6_cleaner_t c;
1395
1396         c.w.root = root;
1397         c.w.func = fib6_clean_node;
1398         c.w.prune = prune;
1399         c.w.count = 0;
1400         c.w.skip = 0;
1401         c.func = func;
1402         c.arg = arg;
1403         c.net = net;
1404
1405         fib6_walk(&c.w);
1406 }
1407
1408 void fib6_clean_all(struct net *net, int (*func)(struct rt6_info *, void *arg),
1409                     int prune, void *arg)
1410 {
1411         struct fib6_table *table;
1412         struct hlist_node *node;
1413         struct hlist_head *head;
1414         unsigned int h;
1415
1416         rcu_read_lock();
1417         for (h = 0; h < FIB6_TABLE_HASHSZ; h++) {
1418                 head = &net->ipv6.fib_table_hash[h];
1419                 hlist_for_each_entry_rcu(table, node, head, tb6_hlist) {
1420                         write_lock_bh(&table->tb6_lock);
1421                         fib6_clean_tree(net, &table->tb6_root,
1422                                         func, prune, arg);
1423                         write_unlock_bh(&table->tb6_lock);
1424                 }
1425         }
1426         rcu_read_unlock();
1427 }
1428
1429 static int fib6_prune_clone(struct rt6_info *rt, void *arg)
1430 {
1431         if (rt->rt6i_flags & RTF_CACHE) {
1432                 RT6_TRACE("pruning clone %p\n", rt);
1433                 return -1;
1434         }
1435
1436         return 0;
1437 }
1438
1439 static void fib6_prune_clones(struct net *net, struct fib6_node *fn,
1440                               struct rt6_info *rt)
1441 {
1442         fib6_clean_tree(net, fn, fib6_prune_clone, 1, rt);
1443 }
1444
1445 /*
1446  *      Garbage collection
1447  */
1448
1449 static struct fib6_gc_args
1450 {
1451         int                     timeout;
1452         int                     more;
1453 } gc_args;
1454
1455 static int fib6_age(struct rt6_info *rt, void *arg)
1456 {
1457         unsigned long now = jiffies;
1458
1459         /*
1460          *      check addrconf expiration here.
1461          *      Routes are expired even if they are in use.
1462          *
1463          *      Also age clones. Note, that clones are aged out
1464          *      only if they are not in use now.
1465          */
1466
1467         if (rt->rt6i_flags&RTF_EXPIRES && rt->rt6i_expires) {
1468                 if (time_after(now, rt->rt6i_expires)) {
1469                         RT6_TRACE("expiring %p\n", rt);
1470                         return -1;
1471                 }
1472                 gc_args.more++;
1473         } else if (rt->rt6i_flags & RTF_CACHE) {
1474                 if (atomic_read(&rt->dst.__refcnt) == 0 &&
1475                     time_after_eq(now, rt->dst.lastuse + gc_args.timeout)) {
1476                         RT6_TRACE("aging clone %p\n", rt);
1477                         return -1;
1478                 } else if ((rt->rt6i_flags & RTF_GATEWAY) &&
1479                            (!(dst_get_neighbour_raw(&rt->dst)->flags & NTF_ROUTER))) {
1480                         RT6_TRACE("purging route %p via non-router but gateway\n",
1481                                   rt);
1482                         return -1;
1483                 }
1484                 gc_args.more++;
1485         }
1486
1487         return 0;
1488 }
1489
1490 static DEFINE_SPINLOCK(fib6_gc_lock);
1491
1492 void fib6_run_gc(unsigned long expires, struct net *net, bool force)
1493 {
1494         unsigned long now;
1495
1496         if (force) {
1497                 spin_lock_bh(&fib6_gc_lock);
1498         } else if (!spin_trylock_bh(&fib6_gc_lock)) {
1499                 mod_timer(&net->ipv6.ip6_fib_timer, jiffies + HZ);
1500                 return;
1501         }
1502         gc_args.timeout = expires ? (int)expires :
1503                           net->ipv6.sysctl.ip6_rt_gc_interval;
1504
1505         gc_args.more = icmp6_dst_gc();
1506
1507         fib6_clean_all(net, fib6_age, 0, NULL);
1508         now = jiffies;
1509         net->ipv6.ip6_rt_last_gc = now;
1510
1511         if (gc_args.more)
1512                 mod_timer(&net->ipv6.ip6_fib_timer,
1513                           round_jiffies(now
1514                                         + net->ipv6.sysctl.ip6_rt_gc_interval));
1515         else
1516                 del_timer(&net->ipv6.ip6_fib_timer);
1517         spin_unlock_bh(&fib6_gc_lock);
1518 }
1519
1520 static void fib6_gc_timer_cb(unsigned long arg)
1521 {
1522         fib6_run_gc(0, (struct net *)arg, true);
1523 }
1524
1525 static int __net_init fib6_net_init(struct net *net)
1526 {
1527         size_t size = sizeof(struct hlist_head) * FIB6_TABLE_HASHSZ;
1528
1529         setup_timer(&net->ipv6.ip6_fib_timer, fib6_gc_timer_cb, (unsigned long)net);
1530
1531         net->ipv6.rt6_stats = kzalloc(sizeof(*net->ipv6.rt6_stats), GFP_KERNEL);
1532         if (!net->ipv6.rt6_stats)
1533                 goto out_timer;
1534
1535         /* Avoid false sharing : Use at least a full cache line */
1536         size = max_t(size_t, size, L1_CACHE_BYTES);
1537
1538         net->ipv6.fib_table_hash = kzalloc(size, GFP_KERNEL);
1539         if (!net->ipv6.fib_table_hash)
1540                 goto out_rt6_stats;
1541
1542         net->ipv6.fib6_main_tbl = kzalloc(sizeof(*net->ipv6.fib6_main_tbl),
1543                                           GFP_KERNEL);
1544         if (!net->ipv6.fib6_main_tbl)
1545                 goto out_fib_table_hash;
1546
1547         net->ipv6.fib6_main_tbl->tb6_id = RT6_TABLE_MAIN;
1548         net->ipv6.fib6_main_tbl->tb6_root.leaf = net->ipv6.ip6_null_entry;
1549         net->ipv6.fib6_main_tbl->tb6_root.fn_flags =
1550                 RTN_ROOT | RTN_TL_ROOT | RTN_RTINFO;
1551
1552 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
1553         net->ipv6.fib6_local_tbl = kzalloc(sizeof(*net->ipv6.fib6_local_tbl),
1554                                            GFP_KERNEL);
1555         if (!net->ipv6.fib6_local_tbl)
1556                 goto out_fib6_main_tbl;
1557         net->ipv6.fib6_local_tbl->tb6_id = RT6_TABLE_LOCAL;
1558         net->ipv6.fib6_local_tbl->tb6_root.leaf = net->ipv6.ip6_null_entry;
1559         net->ipv6.fib6_local_tbl->tb6_root.fn_flags =
1560                 RTN_ROOT | RTN_TL_ROOT | RTN_RTINFO;
1561 #endif
1562         fib6_tables_init(net);
1563
1564         return 0;
1565
1566 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
1567 out_fib6_main_tbl:
1568         kfree(net->ipv6.fib6_main_tbl);
1569 #endif
1570 out_fib_table_hash:
1571         kfree(net->ipv6.fib_table_hash);
1572 out_rt6_stats:
1573         kfree(net->ipv6.rt6_stats);
1574 out_timer:
1575         return -ENOMEM;
1576  }
1577
1578 static void fib6_net_exit(struct net *net)
1579 {
1580         rt6_ifdown(net, NULL);
1581         del_timer_sync(&net->ipv6.ip6_fib_timer);
1582
1583 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
1584         kfree(net->ipv6.fib6_local_tbl);
1585 #endif
1586         kfree(net->ipv6.fib6_main_tbl);
1587         kfree(net->ipv6.fib_table_hash);
1588         kfree(net->ipv6.rt6_stats);
1589 }
1590
1591 static struct pernet_operations fib6_net_ops = {
1592         .init = fib6_net_init,
1593         .exit = fib6_net_exit,
1594 };
1595
1596 int __init fib6_init(void)
1597 {
1598         int ret = -ENOMEM;
1599
1600         fib6_node_kmem = kmem_cache_create("fib6_nodes",
1601                                            sizeof(struct fib6_node),
1602                                            0, SLAB_HWCACHE_ALIGN,
1603                                            NULL);
1604         if (!fib6_node_kmem)
1605                 goto out;
1606
1607         ret = register_pernet_subsys(&fib6_net_ops);
1608         if (ret)
1609                 goto out_kmem_cache_create;
1610
1611         ret = __rtnl_register(PF_INET6, RTM_GETROUTE, NULL, inet6_dump_fib,
1612                               NULL);
1613         if (ret)
1614                 goto out_unregister_subsys;
1615 out:
1616         return ret;
1617
1618 out_unregister_subsys:
1619         unregister_pernet_subsys(&fib6_net_ops);
1620 out_kmem_cache_create:
1621         kmem_cache_destroy(fib6_node_kmem);
1622         goto out;
1623 }
1624
1625 void fib6_gc_cleanup(void)
1626 {
1627         unregister_pernet_subsys(&fib6_net_ops);
1628         kmem_cache_destroy(fib6_node_kmem);
1629 }