2 * This program is free software; you can redistribute it and/or
3 * modify it under the terms of the GNU General Public License
4 * as published by the Free Software Foundation; either version
5 * 2 of the License, or (at your option) any later version.
7 * Robert Olsson <robert.olsson@its.uu.se> Uppsala Universitet
8 * & Swedish University of Agricultural Sciences.
10 * Jens Laas <jens.laas@data.slu.se> Swedish University of
11 * Agricultural Sciences.
13 * Hans Liss <hans.liss@its.uu.se> Uppsala Universitet
15 * This work is based on the LPC-trie which is originally described in:
17 * An experimental study of compression methods for dynamic tries
18 * Stefan Nilsson and Matti Tikkanen. Algorithmica, 33(1):19-33, 2002.
19 * http://www.csc.kth.se/~snilsson/software/dyntrie2/
22 * IP-address lookup using LC-tries. Stefan Nilsson and Gunnar Karlsson
23 * IEEE Journal on Selected Areas in Communications, 17(6):1083-1092, June 1999
26 * Code from fib_hash has been reused which includes the following header:
29 * INET An implementation of the TCP/IP protocol suite for the LINUX
30 * operating system. INET is implemented using the BSD Socket
31 * interface as the means of communication with the user level.
33 * IPv4 FIB: lookup engine and maintenance routines.
36 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
38 * This program is free software; you can redistribute it and/or
39 * modify it under the terms of the GNU General Public License
40 * as published by the Free Software Foundation; either version
41 * 2 of the License, or (at your option) any later version.
43 * Substantial contributions to this work comes from:
45 * David S. Miller, <davem@davemloft.net>
46 * Stephen Hemminger <shemminger@osdl.org>
47 * Paul E. McKenney <paulmck@us.ibm.com>
48 * Patrick McHardy <kaber@trash.net>
51 #define VERSION "0.409"
53 #include <asm/uaccess.h>
54 #include <linux/bitops.h>
55 #include <linux/types.h>
56 #include <linux/kernel.h>
58 #include <linux/string.h>
59 #include <linux/socket.h>
60 #include <linux/sockios.h>
61 #include <linux/errno.h>
63 #include <linux/inet.h>
64 #include <linux/inetdevice.h>
65 #include <linux/netdevice.h>
66 #include <linux/if_arp.h>
67 #include <linux/proc_fs.h>
68 #include <linux/rcupdate.h>
69 #include <linux/skbuff.h>
70 #include <linux/netlink.h>
71 #include <linux/init.h>
72 #include <linux/list.h>
73 #include <linux/slab.h>
74 #include <linux/prefetch.h>
75 #include <linux/export.h>
76 #include <net/net_namespace.h>
78 #include <net/protocol.h>
79 #include <net/route.h>
82 #include <net/ip_fib.h>
83 #include "fib_lookup.h"
85 #define MAX_STAT_DEPTH 32
87 #define KEYLENGTH (8*sizeof(t_key))
89 typedef unsigned int t_key;
93 #define NODE_TYPE_MASK 0x1UL
94 #define NODE_TYPE(node) ((node)->parent & NODE_TYPE_MASK)
96 #define IS_TNODE(n) (!(n->parent & T_LEAF))
97 #define IS_LEAF(n) (n->parent & T_LEAF)
100 unsigned long parent;
105 unsigned long parent;
107 struct hlist_head list;
112 struct hlist_node hlist;
114 u32 mask_plen; /* ntohl(inet_make_mask(plen)) */
115 struct list_head falh;
120 unsigned long parent;
122 unsigned char pos; /* 2log(KEYLENGTH) bits needed */
123 unsigned char bits; /* 2log(KEYLENGTH) bits needed */
124 unsigned int full_children; /* KEYLENGTH bits needed */
125 unsigned int empty_children; /* KEYLENGTH bits needed */
128 struct work_struct work;
129 struct tnode *tnode_free;
131 struct rt_trie_node __rcu *child[0];
134 #ifdef CONFIG_IP_FIB_TRIE_STATS
135 struct trie_use_stats {
137 unsigned int backtrack;
138 unsigned int semantic_match_passed;
139 unsigned int semantic_match_miss;
140 unsigned int null_node_hit;
141 unsigned int resize_node_skipped;
146 unsigned int totdepth;
147 unsigned int maxdepth;
150 unsigned int nullpointers;
151 unsigned int prefixes;
152 unsigned int nodesizes[MAX_STAT_DEPTH];
156 struct rt_trie_node __rcu *trie;
157 #ifdef CONFIG_IP_FIB_TRIE_STATS
158 struct trie_use_stats stats;
162 static void tnode_put_child_reorg(struct tnode *tn, int i, struct rt_trie_node *n,
164 static struct rt_trie_node *resize(struct trie *t, struct tnode *tn);
165 static struct tnode *inflate(struct trie *t, struct tnode *tn);
166 static struct tnode *halve(struct trie *t, struct tnode *tn);
167 /* tnodes to free after resize(); protected by RTNL */
168 static struct tnode *tnode_free_head;
169 static size_t tnode_free_size;
172 * synchronize_rcu after call_rcu for that many pages; it should be especially
173 * useful before resizing the root node with PREEMPT_NONE configs; the value was
174 * obtained experimentally, aiming to avoid visible slowdown.
176 static const int sync_pages = 128;
178 static struct kmem_cache *fn_alias_kmem __read_mostly;
179 static struct kmem_cache *trie_leaf_kmem __read_mostly;
182 * caller must hold RTNL
184 static inline struct tnode *node_parent(const struct rt_trie_node *node)
186 unsigned long parent;
188 parent = rcu_dereference_index_check(node->parent, lockdep_rtnl_is_held());
190 return (struct tnode *)(parent & ~NODE_TYPE_MASK);
194 * caller must hold RCU read lock or RTNL
196 static inline struct tnode *node_parent_rcu(const struct rt_trie_node *node)
198 unsigned long parent;
200 parent = rcu_dereference_index_check(node->parent, rcu_read_lock_held() ||
201 lockdep_rtnl_is_held());
203 return (struct tnode *)(parent & ~NODE_TYPE_MASK);
206 /* Same as rcu_assign_pointer
207 * but that macro() assumes that value is a pointer.
209 static inline void node_set_parent(struct rt_trie_node *node, struct tnode *ptr)
212 node->parent = (unsigned long)ptr | NODE_TYPE(node);
216 * caller must hold RTNL
218 static inline struct rt_trie_node *tnode_get_child(const struct tnode *tn, unsigned int i)
220 BUG_ON(i >= 1U << tn->bits);
222 return rtnl_dereference(tn->child[i]);
226 * caller must hold RCU read lock or RTNL
228 static inline struct rt_trie_node *tnode_get_child_rcu(const struct tnode *tn, unsigned int i)
230 BUG_ON(i >= 1U << tn->bits);
232 return rcu_dereference_rtnl(tn->child[i]);
235 static inline int tnode_child_length(const struct tnode *tn)
237 return 1 << tn->bits;
240 static inline t_key mask_pfx(t_key k, unsigned int l)
242 return (l == 0) ? 0 : k >> (KEYLENGTH-l) << (KEYLENGTH-l);
245 static inline t_key tkey_extract_bits(t_key a, unsigned int offset, unsigned int bits)
247 if (offset < KEYLENGTH)
248 return ((t_key)(a << offset)) >> (KEYLENGTH - bits);
253 static inline int tkey_equals(t_key a, t_key b)
258 static inline int tkey_sub_equals(t_key a, int offset, int bits, t_key b)
260 if (bits == 0 || offset >= KEYLENGTH)
262 bits = bits > KEYLENGTH ? KEYLENGTH : bits;
263 return ((a ^ b) << offset) >> (KEYLENGTH - bits) == 0;
266 static inline int tkey_mismatch(t_key a, int offset, t_key b)
273 while ((diff << i) >> (KEYLENGTH-1) == 0)
279 To understand this stuff, an understanding of keys and all their bits is
280 necessary. Every node in the trie has a key associated with it, but not
281 all of the bits in that key are significant.
283 Consider a node 'n' and its parent 'tp'.
285 If n is a leaf, every bit in its key is significant. Its presence is
286 necessitated by path compression, since during a tree traversal (when
287 searching for a leaf - unless we are doing an insertion) we will completely
288 ignore all skipped bits we encounter. Thus we need to verify, at the end of
289 a potentially successful search, that we have indeed been walking the
292 Note that we can never "miss" the correct key in the tree if present by
293 following the wrong path. Path compression ensures that segments of the key
294 that are the same for all keys with a given prefix are skipped, but the
295 skipped part *is* identical for each node in the subtrie below the skipped
296 bit! trie_insert() in this implementation takes care of that - note the
297 call to tkey_sub_equals() in trie_insert().
299 if n is an internal node - a 'tnode' here, the various parts of its key
300 have many different meanings.
303 _________________________________________________________________
304 | i | i | i | i | i | i | i | N | N | N | S | S | S | S | S | C |
305 -----------------------------------------------------------------
306 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
308 _________________________________________________________________
309 | C | C | C | u | u | u | u | u | u | u | u | u | u | u | u | u |
310 -----------------------------------------------------------------
311 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
318 First, let's just ignore the bits that come before the parent tp, that is
319 the bits from 0 to (tp->pos-1). They are *known* but at this point we do
320 not use them for anything.
322 The bits from (tp->pos) to (tp->pos + tp->bits - 1) - "N", above - are the
323 index into the parent's child array. That is, they will be used to find
324 'n' among tp's children.
326 The bits from (tp->pos + tp->bits) to (n->pos - 1) - "S" - are skipped bits
329 All the bits we have seen so far are significant to the node n. The rest
330 of the bits are really not needed or indeed known in n->key.
332 The bits from (n->pos) to (n->pos + n->bits - 1) - "C" - are the index into
333 n's child array, and will of course be different for each child.
336 The rest of the bits, from (n->pos + n->bits) onward, are completely unknown
341 static inline void check_tnode(const struct tnode *tn)
343 WARN_ON(tn && tn->pos+tn->bits > 32);
346 static const int halve_threshold = 25;
347 static const int inflate_threshold = 50;
348 static const int halve_threshold_root = 15;
349 static const int inflate_threshold_root = 30;
351 static void __alias_free_mem(struct rcu_head *head)
353 struct fib_alias *fa = container_of(head, struct fib_alias, rcu);
354 kmem_cache_free(fn_alias_kmem, fa);
357 static inline void alias_free_mem_rcu(struct fib_alias *fa)
359 call_rcu(&fa->rcu, __alias_free_mem);
362 static void __leaf_free_rcu(struct rcu_head *head)
364 struct leaf *l = container_of(head, struct leaf, rcu);
365 kmem_cache_free(trie_leaf_kmem, l);
368 static inline void free_leaf(struct leaf *l)
370 call_rcu_bh(&l->rcu, __leaf_free_rcu);
373 static inline void free_leaf_info(struct leaf_info *leaf)
375 kfree_rcu(leaf, rcu);
378 static struct tnode *tnode_alloc(size_t size)
380 if (size <= PAGE_SIZE)
381 return kzalloc(size, GFP_KERNEL);
383 return vzalloc(size);
386 static void __tnode_vfree(struct work_struct *arg)
388 struct tnode *tn = container_of(arg, struct tnode, work);
392 static void __tnode_free_rcu(struct rcu_head *head)
394 struct tnode *tn = container_of(head, struct tnode, rcu);
395 size_t size = sizeof(struct tnode) +
396 (sizeof(struct rt_trie_node *) << tn->bits);
398 if (size <= PAGE_SIZE)
401 INIT_WORK(&tn->work, __tnode_vfree);
402 schedule_work(&tn->work);
406 static inline void tnode_free(struct tnode *tn)
409 free_leaf((struct leaf *) tn);
411 call_rcu(&tn->rcu, __tnode_free_rcu);
414 static void tnode_free_safe(struct tnode *tn)
417 tn->tnode_free = tnode_free_head;
418 tnode_free_head = tn;
419 tnode_free_size += sizeof(struct tnode) +
420 (sizeof(struct rt_trie_node *) << tn->bits);
423 static void tnode_free_flush(void)
427 while ((tn = tnode_free_head)) {
428 tnode_free_head = tn->tnode_free;
429 tn->tnode_free = NULL;
433 if (tnode_free_size >= PAGE_SIZE * sync_pages) {
439 static struct leaf *leaf_new(void)
441 struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
444 INIT_HLIST_HEAD(&l->list);
449 static struct leaf_info *leaf_info_new(int plen)
451 struct leaf_info *li = kmalloc(sizeof(struct leaf_info), GFP_KERNEL);
454 li->mask_plen = ntohl(inet_make_mask(plen));
455 INIT_LIST_HEAD(&li->falh);
460 static struct tnode *tnode_new(t_key key, int pos, int bits)
462 size_t sz = sizeof(struct tnode) + (sizeof(struct rt_trie_node *) << bits);
463 struct tnode *tn = tnode_alloc(sz);
466 tn->parent = T_TNODE;
470 tn->full_children = 0;
471 tn->empty_children = 1<<bits;
474 pr_debug("AT %p s=%zu %zu\n", tn, sizeof(struct tnode),
475 sizeof(struct rt_trie_node *) << bits);
480 * Check whether a tnode 'n' is "full", i.e. it is an internal node
481 * and no bits are skipped. See discussion in dyntree paper p. 6
484 static inline int tnode_full(const struct tnode *tn, const struct rt_trie_node *n)
486 if (n == NULL || IS_LEAF(n))
489 return ((struct tnode *) n)->pos == tn->pos + tn->bits;
492 static inline void put_child(struct tnode *tn, int i,
493 struct rt_trie_node *n)
495 tnode_put_child_reorg(tn, i, n, -1);
499 * Add a child at position i overwriting the old value.
500 * Update the value of full_children and empty_children.
503 static void tnode_put_child_reorg(struct tnode *tn, int i, struct rt_trie_node *n,
506 struct rt_trie_node *chi = rtnl_dereference(tn->child[i]);
509 BUG_ON(i >= 1<<tn->bits);
511 /* update emptyChildren */
512 if (n == NULL && chi != NULL)
513 tn->empty_children++;
514 else if (n != NULL && chi == NULL)
515 tn->empty_children--;
517 /* update fullChildren */
519 wasfull = tnode_full(tn, chi);
521 isfull = tnode_full(tn, n);
522 if (wasfull && !isfull)
524 else if (!wasfull && isfull)
528 node_set_parent(n, tn);
530 rcu_assign_pointer(tn->child[i], n);
534 static struct rt_trie_node *resize(struct trie *t, struct tnode *tn)
537 struct tnode *old_tn;
538 int inflate_threshold_use;
539 int halve_threshold_use;
545 pr_debug("In tnode_resize %p inflate_threshold=%d threshold=%d\n",
546 tn, inflate_threshold, halve_threshold);
549 if (tn->empty_children == tnode_child_length(tn)) {
554 if (tn->empty_children == tnode_child_length(tn) - 1)
557 * Double as long as the resulting node has a number of
558 * nonempty nodes that are above the threshold.
562 * From "Implementing a dynamic compressed trie" by Stefan Nilsson of
563 * the Helsinki University of Technology and Matti Tikkanen of Nokia
564 * Telecommunications, page 6:
565 * "A node is doubled if the ratio of non-empty children to all
566 * children in the *doubled* node is at least 'high'."
568 * 'high' in this instance is the variable 'inflate_threshold'. It
569 * is expressed as a percentage, so we multiply it with
570 * tnode_child_length() and instead of multiplying by 2 (since the
571 * child array will be doubled by inflate()) and multiplying
572 * the left-hand side by 100 (to handle the percentage thing) we
573 * multiply the left-hand side by 50.
575 * The left-hand side may look a bit weird: tnode_child_length(tn)
576 * - tn->empty_children is of course the number of non-null children
577 * in the current node. tn->full_children is the number of "full"
578 * children, that is non-null tnodes with a skip value of 0.
579 * All of those will be doubled in the resulting inflated tnode, so
580 * we just count them one extra time here.
582 * A clearer way to write this would be:
584 * to_be_doubled = tn->full_children;
585 * not_to_be_doubled = tnode_child_length(tn) - tn->empty_children -
588 * new_child_length = tnode_child_length(tn) * 2;
590 * new_fill_factor = 100 * (not_to_be_doubled + 2*to_be_doubled) /
592 * if (new_fill_factor >= inflate_threshold)
594 * ...and so on, tho it would mess up the while () loop.
597 * 100 * (not_to_be_doubled + 2*to_be_doubled) / new_child_length >=
601 * 100 * (not_to_be_doubled + 2*to_be_doubled) >=
602 * inflate_threshold * new_child_length
604 * expand not_to_be_doubled and to_be_doubled, and shorten:
605 * 100 * (tnode_child_length(tn) - tn->empty_children +
606 * tn->full_children) >= inflate_threshold * new_child_length
608 * expand new_child_length:
609 * 100 * (tnode_child_length(tn) - tn->empty_children +
610 * tn->full_children) >=
611 * inflate_threshold * tnode_child_length(tn) * 2
614 * 50 * (tn->full_children + tnode_child_length(tn) -
615 * tn->empty_children) >= inflate_threshold *
616 * tnode_child_length(tn)
622 /* Keep root node larger */
624 if (!node_parent((struct rt_trie_node *)tn)) {
625 inflate_threshold_use = inflate_threshold_root;
626 halve_threshold_use = halve_threshold_root;
628 inflate_threshold_use = inflate_threshold;
629 halve_threshold_use = halve_threshold;
633 while ((tn->full_children > 0 && max_work-- &&
634 50 * (tn->full_children + tnode_child_length(tn)
635 - tn->empty_children)
636 >= inflate_threshold_use * tnode_child_length(tn))) {
643 #ifdef CONFIG_IP_FIB_TRIE_STATS
644 t->stats.resize_node_skipped++;
652 /* Return if at least one inflate is run */
653 if (max_work != MAX_WORK)
654 return (struct rt_trie_node *) tn;
657 * Halve as long as the number of empty children in this
658 * node is above threshold.
662 while (tn->bits > 1 && max_work-- &&
663 100 * (tnode_child_length(tn) - tn->empty_children) <
664 halve_threshold_use * tnode_child_length(tn)) {
670 #ifdef CONFIG_IP_FIB_TRIE_STATS
671 t->stats.resize_node_skipped++;
678 /* Only one child remains */
679 if (tn->empty_children == tnode_child_length(tn) - 1) {
681 for (i = 0; i < tnode_child_length(tn); i++) {
682 struct rt_trie_node *n;
684 n = rtnl_dereference(tn->child[i]);
688 /* compress one level */
690 node_set_parent(n, NULL);
695 return (struct rt_trie_node *) tn;
699 static void tnode_clean_free(struct tnode *tn)
702 struct tnode *tofree;
704 for (i = 0; i < tnode_child_length(tn); i++) {
705 tofree = (struct tnode *)rtnl_dereference(tn->child[i]);
712 static struct tnode *inflate(struct trie *t, struct tnode *tn)
714 struct tnode *oldtnode = tn;
715 int olen = tnode_child_length(tn);
718 pr_debug("In inflate\n");
720 tn = tnode_new(oldtnode->key, oldtnode->pos, oldtnode->bits + 1);
723 return ERR_PTR(-ENOMEM);
726 * Preallocate and store tnodes before the actual work so we
727 * don't get into an inconsistent state if memory allocation
728 * fails. In case of failure we return the oldnode and inflate
729 * of tnode is ignored.
732 for (i = 0; i < olen; i++) {
735 inode = (struct tnode *) tnode_get_child(oldtnode, i);
738 inode->pos == oldtnode->pos + oldtnode->bits &&
740 struct tnode *left, *right;
741 t_key m = ~0U << (KEYLENGTH - 1) >> inode->pos;
743 left = tnode_new(inode->key&(~m), inode->pos + 1,
748 right = tnode_new(inode->key|m, inode->pos + 1,
756 put_child(tn, 2*i, (struct rt_trie_node *) left);
757 put_child(tn, 2*i+1, (struct rt_trie_node *) right);
761 for (i = 0; i < olen; i++) {
763 struct rt_trie_node *node = tnode_get_child(oldtnode, i);
764 struct tnode *left, *right;
771 /* A leaf or an internal node with skipped bits */
773 if (IS_LEAF(node) || ((struct tnode *) node)->pos >
774 tn->pos + tn->bits - 1) {
775 if (tkey_extract_bits(node->key,
776 oldtnode->pos + oldtnode->bits,
778 put_child(tn, 2*i, node);
780 put_child(tn, 2*i+1, node);
784 /* An internal node with two children */
785 inode = (struct tnode *) node;
787 if (inode->bits == 1) {
788 put_child(tn, 2*i, rtnl_dereference(inode->child[0]));
789 put_child(tn, 2*i+1, rtnl_dereference(inode->child[1]));
791 tnode_free_safe(inode);
795 /* An internal node with more than two children */
797 /* We will replace this node 'inode' with two new
798 * ones, 'left' and 'right', each with half of the
799 * original children. The two new nodes will have
800 * a position one bit further down the key and this
801 * means that the "significant" part of their keys
802 * (see the discussion near the top of this file)
803 * will differ by one bit, which will be "0" in
804 * left's key and "1" in right's key. Since we are
805 * moving the key position by one step, the bit that
806 * we are moving away from - the bit at position
807 * (inode->pos) - is the one that will differ between
808 * left and right. So... we synthesize that bit in the
810 * The mask 'm' below will be a single "one" bit at
811 * the position (inode->pos)
814 /* Use the old key, but set the new significant
818 left = (struct tnode *) tnode_get_child(tn, 2*i);
819 put_child(tn, 2*i, NULL);
823 right = (struct tnode *) tnode_get_child(tn, 2*i+1);
824 put_child(tn, 2*i+1, NULL);
828 size = tnode_child_length(left);
829 for (j = 0; j < size; j++) {
830 put_child(left, j, rtnl_dereference(inode->child[j]));
831 put_child(right, j, rtnl_dereference(inode->child[j + size]));
833 put_child(tn, 2*i, resize(t, left));
834 put_child(tn, 2*i+1, resize(t, right));
836 tnode_free_safe(inode);
838 tnode_free_safe(oldtnode);
841 tnode_clean_free(tn);
842 return ERR_PTR(-ENOMEM);
845 static struct tnode *halve(struct trie *t, struct tnode *tn)
847 struct tnode *oldtnode = tn;
848 struct rt_trie_node *left, *right;
850 int olen = tnode_child_length(tn);
852 pr_debug("In halve\n");
854 tn = tnode_new(oldtnode->key, oldtnode->pos, oldtnode->bits - 1);
857 return ERR_PTR(-ENOMEM);
860 * Preallocate and store tnodes before the actual work so we
861 * don't get into an inconsistent state if memory allocation
862 * fails. In case of failure we return the oldnode and halve
863 * of tnode is ignored.
866 for (i = 0; i < olen; i += 2) {
867 left = tnode_get_child(oldtnode, i);
868 right = tnode_get_child(oldtnode, i+1);
870 /* Two nonempty children */
874 newn = tnode_new(left->key, tn->pos + tn->bits, 1);
879 put_child(tn, i/2, (struct rt_trie_node *)newn);
884 for (i = 0; i < olen; i += 2) {
885 struct tnode *newBinNode;
887 left = tnode_get_child(oldtnode, i);
888 right = tnode_get_child(oldtnode, i+1);
890 /* At least one of the children is empty */
892 if (right == NULL) /* Both are empty */
894 put_child(tn, i/2, right);
899 put_child(tn, i/2, left);
903 /* Two nonempty children */
904 newBinNode = (struct tnode *) tnode_get_child(tn, i/2);
905 put_child(tn, i/2, NULL);
906 put_child(newBinNode, 0, left);
907 put_child(newBinNode, 1, right);
908 put_child(tn, i/2, resize(t, newBinNode));
910 tnode_free_safe(oldtnode);
913 tnode_clean_free(tn);
914 return ERR_PTR(-ENOMEM);
917 /* readside must use rcu_read_lock currently dump routines
918 via get_fa_head and dump */
920 static struct leaf_info *find_leaf_info(struct leaf *l, int plen)
922 struct hlist_head *head = &l->list;
923 struct hlist_node *node;
924 struct leaf_info *li;
926 hlist_for_each_entry_rcu(li, node, head, hlist)
927 if (li->plen == plen)
933 static inline struct list_head *get_fa_head(struct leaf *l, int plen)
935 struct leaf_info *li = find_leaf_info(l, plen);
943 static void insert_leaf_info(struct hlist_head *head, struct leaf_info *new)
945 struct leaf_info *li = NULL, *last = NULL;
946 struct hlist_node *node;
948 if (hlist_empty(head)) {
949 hlist_add_head_rcu(&new->hlist, head);
951 hlist_for_each_entry(li, node, head, hlist) {
952 if (new->plen > li->plen)
958 hlist_add_after_rcu(&last->hlist, &new->hlist);
960 hlist_add_before_rcu(&new->hlist, &li->hlist);
964 /* rcu_read_lock needs to be hold by caller from readside */
967 fib_find_node(struct trie *t, u32 key)
971 struct rt_trie_node *n;
974 n = rcu_dereference_rtnl(t->trie);
976 while (n != NULL && NODE_TYPE(n) == T_TNODE) {
977 tn = (struct tnode *) n;
981 if (tkey_sub_equals(tn->key, pos, tn->pos-pos, key)) {
982 pos = tn->pos + tn->bits;
983 n = tnode_get_child_rcu(tn,
984 tkey_extract_bits(key,
990 /* Case we have found a leaf. Compare prefixes */
992 if (n != NULL && IS_LEAF(n) && tkey_equals(key, n->key))
993 return (struct leaf *)n;
998 static void trie_rebalance(struct trie *t, struct tnode *tn)
1006 while (tn != NULL && (tp = node_parent((struct rt_trie_node *)tn)) != NULL) {
1007 cindex = tkey_extract_bits(key, tp->pos, tp->bits);
1008 wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
1009 tn = (struct tnode *)resize(t, tn);
1011 tnode_put_child_reorg(tp, cindex,
1012 (struct rt_trie_node *)tn, wasfull);
1014 tp = node_parent((struct rt_trie_node *) tn);
1016 rcu_assign_pointer(t->trie, (struct rt_trie_node *)tn);
1024 /* Handle last (top) tnode */
1026 tn = (struct tnode *)resize(t, tn);
1028 rcu_assign_pointer(t->trie, (struct rt_trie_node *)tn);
1032 /* only used from updater-side */
1034 static struct list_head *fib_insert_node(struct trie *t, u32 key, int plen)
1037 struct tnode *tp = NULL, *tn = NULL;
1038 struct rt_trie_node *n;
1041 struct list_head *fa_head = NULL;
1042 struct leaf_info *li;
1046 n = rtnl_dereference(t->trie);
1048 /* If we point to NULL, stop. Either the tree is empty and we should
1049 * just put a new leaf in if, or we have reached an empty child slot,
1050 * and we should just put our new leaf in that.
1051 * If we point to a T_TNODE, check if it matches our key. Note that
1052 * a T_TNODE might be skipping any number of bits - its 'pos' need
1053 * not be the parent's 'pos'+'bits'!
1055 * If it does match the current key, get pos/bits from it, extract
1056 * the index from our key, push the T_TNODE and walk the tree.
1058 * If it doesn't, we have to replace it with a new T_TNODE.
1060 * If we point to a T_LEAF, it might or might not have the same key
1061 * as we do. If it does, just change the value, update the T_LEAF's
1062 * value, and return it.
1063 * If it doesn't, we need to replace it with a T_TNODE.
1066 while (n != NULL && NODE_TYPE(n) == T_TNODE) {
1067 tn = (struct tnode *) n;
1071 if (tkey_sub_equals(tn->key, pos, tn->pos-pos, key)) {
1073 pos = tn->pos + tn->bits;
1074 n = tnode_get_child(tn,
1075 tkey_extract_bits(key,
1079 BUG_ON(n && node_parent(n) != tn);
1085 * n ----> NULL, LEAF or TNODE
1087 * tp is n's (parent) ----> NULL or TNODE
1090 BUG_ON(tp && IS_LEAF(tp));
1092 /* Case 1: n is a leaf. Compare prefixes */
1094 if (n != NULL && IS_LEAF(n) && tkey_equals(key, n->key)) {
1095 l = (struct leaf *) n;
1096 li = leaf_info_new(plen);
1101 fa_head = &li->falh;
1102 insert_leaf_info(&l->list, li);
1111 li = leaf_info_new(plen);
1118 fa_head = &li->falh;
1119 insert_leaf_info(&l->list, li);
1121 if (t->trie && n == NULL) {
1122 /* Case 2: n is NULL, and will just insert a new leaf */
1124 node_set_parent((struct rt_trie_node *)l, tp);
1126 cindex = tkey_extract_bits(key, tp->pos, tp->bits);
1127 put_child(tp, cindex, (struct rt_trie_node *)l);
1129 /* Case 3: n is a LEAF or a TNODE and the key doesn't match. */
1131 * Add a new tnode here
1132 * first tnode need some special handling
1136 pos = tp->pos+tp->bits;
1141 newpos = tkey_mismatch(key, pos, n->key);
1142 tn = tnode_new(n->key, newpos, 1);
1145 tn = tnode_new(key, newpos, 1); /* First tnode */
1154 node_set_parent((struct rt_trie_node *)tn, tp);
1156 missbit = tkey_extract_bits(key, newpos, 1);
1157 put_child(tn, missbit, (struct rt_trie_node *)l);
1158 put_child(tn, 1-missbit, n);
1161 cindex = tkey_extract_bits(key, tp->pos, tp->bits);
1162 put_child(tp, cindex, (struct rt_trie_node *)tn);
1164 rcu_assign_pointer(t->trie, (struct rt_trie_node *)tn);
1169 if (tp && tp->pos + tp->bits > 32)
1170 pr_warn("fib_trie tp=%p pos=%d, bits=%d, key=%0x plen=%d\n",
1171 tp, tp->pos, tp->bits, key, plen);
1173 /* Rebalance the trie */
1175 trie_rebalance(t, tp);
1181 * Caller must hold RTNL.
1183 int fib_table_insert(struct fib_table *tb, struct fib_config *cfg)
1185 struct trie *t = (struct trie *) tb->tb_data;
1186 struct fib_alias *fa, *new_fa;
1187 struct list_head *fa_head = NULL;
1188 struct fib_info *fi;
1189 int plen = cfg->fc_dst_len;
1190 u8 tos = cfg->fc_tos;
1198 key = ntohl(cfg->fc_dst);
1200 pr_debug("Insert table=%u %08x/%d\n", tb->tb_id, key, plen);
1202 mask = ntohl(inet_make_mask(plen));
1209 fi = fib_create_info(cfg);
1215 l = fib_find_node(t, key);
1219 fa_head = get_fa_head(l, plen);
1220 fa = fib_find_alias(fa_head, tos, fi->fib_priority);
1223 /* Now fa, if non-NULL, points to the first fib alias
1224 * with the same keys [prefix,tos,priority], if such key already
1225 * exists or to the node before which we will insert new one.
1227 * If fa is NULL, we will need to allocate a new one and
1228 * insert to the head of f.
1230 * If f is NULL, no fib node matched the destination key
1231 * and we need to allocate a new one of those as well.
1234 if (fa && fa->fa_tos == tos &&
1235 fa->fa_info->fib_priority == fi->fib_priority) {
1236 struct fib_alias *fa_first, *fa_match;
1239 if (cfg->fc_nlflags & NLM_F_EXCL)
1243 * 1. Find exact match for type, scope, fib_info to avoid
1245 * 2. Find next 'fa' (or head), NLM_F_APPEND inserts before it
1249 fa = list_entry(fa->fa_list.prev, struct fib_alias, fa_list);
1250 list_for_each_entry_continue(fa, fa_head, fa_list) {
1251 if (fa->fa_tos != tos)
1253 if (fa->fa_info->fib_priority != fi->fib_priority)
1255 if (fa->fa_type == cfg->fc_type &&
1256 fa->fa_info == fi) {
1262 if (cfg->fc_nlflags & NLM_F_REPLACE) {
1263 struct fib_info *fi_drop;
1273 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1277 fi_drop = fa->fa_info;
1278 new_fa->fa_tos = fa->fa_tos;
1279 new_fa->fa_info = fi;
1280 new_fa->fa_type = cfg->fc_type;
1281 state = fa->fa_state;
1282 new_fa->fa_state = state & ~FA_S_ACCESSED;
1284 list_replace_rcu(&fa->fa_list, &new_fa->fa_list);
1285 alias_free_mem_rcu(fa);
1287 fib_release_info(fi_drop);
1288 if (state & FA_S_ACCESSED)
1289 rt_cache_flush(cfg->fc_nlinfo.nl_net, -1);
1290 rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen,
1291 tb->tb_id, &cfg->fc_nlinfo, NLM_F_REPLACE);
1295 /* Error if we find a perfect match which
1296 * uses the same scope, type, and nexthop
1302 if (!(cfg->fc_nlflags & NLM_F_APPEND))
1306 if (!(cfg->fc_nlflags & NLM_F_CREATE))
1310 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1314 new_fa->fa_info = fi;
1315 new_fa->fa_tos = tos;
1316 new_fa->fa_type = cfg->fc_type;
1317 new_fa->fa_state = 0;
1319 * Insert new entry to the list.
1323 fa_head = fib_insert_node(t, key, plen);
1324 if (unlikely(!fa_head)) {
1326 goto out_free_new_fa;
1331 tb->tb_num_default++;
1333 list_add_tail_rcu(&new_fa->fa_list,
1334 (fa ? &fa->fa_list : fa_head));
1336 rt_cache_flush(cfg->fc_nlinfo.nl_net, -1);
1337 rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen, tb->tb_id,
1338 &cfg->fc_nlinfo, 0);
1343 kmem_cache_free(fn_alias_kmem, new_fa);
1345 fib_release_info(fi);
1350 /* should be called with rcu_read_lock */
1351 static int check_leaf(struct fib_table *tb, struct trie *t, struct leaf *l,
1352 t_key key, const struct flowi4 *flp,
1353 struct fib_result *res, int fib_flags)
1355 struct leaf_info *li;
1356 struct hlist_head *hhead = &l->list;
1357 struct hlist_node *node;
1359 hlist_for_each_entry_rcu(li, node, hhead, hlist) {
1360 struct fib_alias *fa;
1362 if (l->key != (key & li->mask_plen))
1365 list_for_each_entry_rcu(fa, &li->falh, fa_list) {
1366 struct fib_info *fi = fa->fa_info;
1369 if (fa->fa_tos && fa->fa_tos != flp->flowi4_tos)
1373 if (fa->fa_info->fib_scope < flp->flowi4_scope)
1375 fib_alias_accessed(fa);
1376 err = fib_props[fa->fa_type].error;
1378 #ifdef CONFIG_IP_FIB_TRIE_STATS
1379 t->stats.semantic_match_passed++;
1383 if (fi->fib_flags & RTNH_F_DEAD)
1385 for (nhsel = 0; nhsel < fi->fib_nhs; nhsel++) {
1386 const struct fib_nh *nh = &fi->fib_nh[nhsel];
1388 if (nh->nh_flags & RTNH_F_DEAD)
1390 if (flp->flowi4_oif && flp->flowi4_oif != nh->nh_oif)
1393 #ifdef CONFIG_IP_FIB_TRIE_STATS
1394 t->stats.semantic_match_passed++;
1396 res->prefixlen = li->plen;
1397 res->nh_sel = nhsel;
1398 res->type = fa->fa_type;
1399 res->scope = fa->fa_info->fib_scope;
1402 res->fa_head = &li->falh;
1403 if (!(fib_flags & FIB_LOOKUP_NOREF))
1404 atomic_inc(&fi->fib_clntref);
1409 #ifdef CONFIG_IP_FIB_TRIE_STATS
1410 t->stats.semantic_match_miss++;
1417 int fib_table_lookup(struct fib_table *tb, const struct flowi4 *flp,
1418 struct fib_result *res, int fib_flags)
1420 struct trie *t = (struct trie *) tb->tb_data;
1422 struct rt_trie_node *n;
1424 unsigned int pos, bits;
1425 t_key key = ntohl(flp->daddr);
1426 unsigned int chopped_off;
1428 unsigned int current_prefix_length = KEYLENGTH;
1430 t_key pref_mismatch;
1434 n = rcu_dereference(t->trie);
1438 #ifdef CONFIG_IP_FIB_TRIE_STATS
1444 ret = check_leaf(tb, t, (struct leaf *)n, key, flp, res, fib_flags);
1448 pn = (struct tnode *) n;
1456 cindex = tkey_extract_bits(mask_pfx(key, current_prefix_length),
1459 n = tnode_get_child_rcu(pn, cindex);
1462 #ifdef CONFIG_IP_FIB_TRIE_STATS
1463 t->stats.null_node_hit++;
1469 ret = check_leaf(tb, t, (struct leaf *)n, key, flp, res, fib_flags);
1475 cn = (struct tnode *)n;
1478 * It's a tnode, and we can do some extra checks here if we
1479 * like, to avoid descending into a dead-end branch.
1480 * This tnode is in the parent's child array at index
1481 * key[p_pos..p_pos+p_bits] but potentially with some bits
1482 * chopped off, so in reality the index may be just a
1483 * subprefix, padded with zero at the end.
1484 * We can also take a look at any skipped bits in this
1485 * tnode - everything up to p_pos is supposed to be ok,
1486 * and the non-chopped bits of the index (se previous
1487 * paragraph) are also guaranteed ok, but the rest is
1488 * considered unknown.
1490 * The skipped bits are key[pos+bits..cn->pos].
1493 /* If current_prefix_length < pos+bits, we are already doing
1494 * actual prefix matching, which means everything from
1495 * pos+(bits-chopped_off) onward must be zero along some
1496 * branch of this subtree - otherwise there is *no* valid
1497 * prefix present. Here we can only check the skipped
1498 * bits. Remember, since we have already indexed into the
1499 * parent's child array, we know that the bits we chopped of
1503 /* NOTA BENE: Checking only skipped bits
1504 for the new node here */
1506 if (current_prefix_length < pos+bits) {
1507 if (tkey_extract_bits(cn->key, current_prefix_length,
1508 cn->pos - current_prefix_length)
1514 * If chopped_off=0, the index is fully validated and we
1515 * only need to look at the skipped bits for this, the new,
1516 * tnode. What we actually want to do is to find out if
1517 * these skipped bits match our key perfectly, or if we will
1518 * have to count on finding a matching prefix further down,
1519 * because if we do, we would like to have some way of
1520 * verifying the existence of such a prefix at this point.
1523 /* The only thing we can do at this point is to verify that
1524 * any such matching prefix can indeed be a prefix to our
1525 * key, and if the bits in the node we are inspecting that
1526 * do not match our key are not ZERO, this cannot be true.
1527 * Thus, find out where there is a mismatch (before cn->pos)
1528 * and verify that all the mismatching bits are zero in the
1533 * Note: We aren't very concerned about the piece of
1534 * the key that precede pn->pos+pn->bits, since these
1535 * have already been checked. The bits after cn->pos
1536 * aren't checked since these are by definition
1537 * "unknown" at this point. Thus, what we want to see
1538 * is if we are about to enter the "prefix matching"
1539 * state, and in that case verify that the skipped
1540 * bits that will prevail throughout this subtree are
1541 * zero, as they have to be if we are to find a
1545 pref_mismatch = mask_pfx(cn->key ^ key, cn->pos);
1548 * In short: If skipped bits in this node do not match
1549 * the search key, enter the "prefix matching"
1552 if (pref_mismatch) {
1553 int mp = KEYLENGTH - fls(pref_mismatch);
1555 if (tkey_extract_bits(cn->key, mp, cn->pos - mp) != 0)
1558 if (current_prefix_length >= cn->pos)
1559 current_prefix_length = mp;
1562 pn = (struct tnode *)n; /* Descend */
1569 /* As zero don't change the child key (cindex) */
1570 while ((chopped_off <= pn->bits)
1571 && !(cindex & (1<<(chopped_off-1))))
1574 /* Decrease current_... with bits chopped off */
1575 if (current_prefix_length > pn->pos + pn->bits - chopped_off)
1576 current_prefix_length = pn->pos + pn->bits
1580 * Either we do the actual chop off according or if we have
1581 * chopped off all bits in this tnode walk up to our parent.
1584 if (chopped_off <= pn->bits) {
1585 cindex &= ~(1 << (chopped_off-1));
1587 struct tnode *parent = node_parent_rcu((struct rt_trie_node *) pn);
1591 /* Get Child's index */
1592 cindex = tkey_extract_bits(pn->key, parent->pos, parent->bits);
1596 #ifdef CONFIG_IP_FIB_TRIE_STATS
1597 t->stats.backtrack++;
1608 EXPORT_SYMBOL_GPL(fib_table_lookup);
1611 * Remove the leaf and return parent.
1613 static void trie_leaf_remove(struct trie *t, struct leaf *l)
1615 struct tnode *tp = node_parent((struct rt_trie_node *) l);
1617 pr_debug("entering trie_leaf_remove(%p)\n", l);
1620 t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
1621 put_child(tp, cindex, NULL);
1622 trie_rebalance(t, tp);
1624 RCU_INIT_POINTER(t->trie, NULL);
1630 * Caller must hold RTNL.
1632 int fib_table_delete(struct fib_table *tb, struct fib_config *cfg)
1634 struct trie *t = (struct trie *) tb->tb_data;
1636 int plen = cfg->fc_dst_len;
1637 u8 tos = cfg->fc_tos;
1638 struct fib_alias *fa, *fa_to_delete;
1639 struct list_head *fa_head;
1641 struct leaf_info *li;
1646 key = ntohl(cfg->fc_dst);
1647 mask = ntohl(inet_make_mask(plen));
1653 l = fib_find_node(t, key);
1658 fa_head = get_fa_head(l, plen);
1659 fa = fib_find_alias(fa_head, tos, 0);
1664 pr_debug("Deleting %08x/%d tos=%d t=%p\n", key, plen, tos, t);
1666 fa_to_delete = NULL;
1667 fa = list_entry(fa->fa_list.prev, struct fib_alias, fa_list);
1668 list_for_each_entry_continue(fa, fa_head, fa_list) {
1669 struct fib_info *fi = fa->fa_info;
1671 if (fa->fa_tos != tos)
1674 if ((!cfg->fc_type || fa->fa_type == cfg->fc_type) &&
1675 (cfg->fc_scope == RT_SCOPE_NOWHERE ||
1676 fa->fa_info->fib_scope == cfg->fc_scope) &&
1677 (!cfg->fc_prefsrc ||
1678 fi->fib_prefsrc == cfg->fc_prefsrc) &&
1679 (!cfg->fc_protocol ||
1680 fi->fib_protocol == cfg->fc_protocol) &&
1681 fib_nh_match(cfg, fi) == 0) {
1691 rtmsg_fib(RTM_DELROUTE, htonl(key), fa, plen, tb->tb_id,
1692 &cfg->fc_nlinfo, 0);
1694 l = fib_find_node(t, key);
1695 li = find_leaf_info(l, plen);
1697 list_del_rcu(&fa->fa_list);
1700 tb->tb_num_default--;
1702 if (list_empty(fa_head)) {
1703 hlist_del_rcu(&li->hlist);
1707 if (hlist_empty(&l->list))
1708 trie_leaf_remove(t, l);
1710 if (fa->fa_state & FA_S_ACCESSED)
1711 rt_cache_flush(cfg->fc_nlinfo.nl_net, -1);
1713 fib_release_info(fa->fa_info);
1714 alias_free_mem_rcu(fa);
1718 static int trie_flush_list(struct list_head *head)
1720 struct fib_alias *fa, *fa_node;
1723 list_for_each_entry_safe(fa, fa_node, head, fa_list) {
1724 struct fib_info *fi = fa->fa_info;
1726 if (fi && (fi->fib_flags & RTNH_F_DEAD)) {
1727 list_del_rcu(&fa->fa_list);
1728 fib_release_info(fa->fa_info);
1729 alias_free_mem_rcu(fa);
1736 static int trie_flush_leaf(struct leaf *l)
1739 struct hlist_head *lih = &l->list;
1740 struct hlist_node *node, *tmp;
1741 struct leaf_info *li = NULL;
1743 hlist_for_each_entry_safe(li, node, tmp, lih, hlist) {
1744 found += trie_flush_list(&li->falh);
1746 if (list_empty(&li->falh)) {
1747 hlist_del_rcu(&li->hlist);
1755 * Scan for the next right leaf starting at node p->child[idx]
1756 * Since we have back pointer, no recursion necessary.
1758 static struct leaf *leaf_walk_rcu(struct tnode *p, struct rt_trie_node *c)
1764 idx = tkey_extract_bits(c->key, p->pos, p->bits) + 1;
1768 while (idx < 1u << p->bits) {
1769 c = tnode_get_child_rcu(p, idx++);
1774 prefetch(rcu_dereference_rtnl(p->child[idx]));
1775 return (struct leaf *) c;
1778 /* Rescan start scanning in new node */
1779 p = (struct tnode *) c;
1783 /* Node empty, walk back up to parent */
1784 c = (struct rt_trie_node *) p;
1785 } while ((p = node_parent_rcu(c)) != NULL);
1787 return NULL; /* Root of trie */
1790 static struct leaf *trie_firstleaf(struct trie *t)
1792 struct tnode *n = (struct tnode *)rcu_dereference_rtnl(t->trie);
1797 if (IS_LEAF(n)) /* trie is just a leaf */
1798 return (struct leaf *) n;
1800 return leaf_walk_rcu(n, NULL);
1803 static struct leaf *trie_nextleaf(struct leaf *l)
1805 struct rt_trie_node *c = (struct rt_trie_node *) l;
1806 struct tnode *p = node_parent_rcu(c);
1809 return NULL; /* trie with just one leaf */
1811 return leaf_walk_rcu(p, c);
1814 static struct leaf *trie_leafindex(struct trie *t, int index)
1816 struct leaf *l = trie_firstleaf(t);
1818 while (l && index-- > 0)
1819 l = trie_nextleaf(l);
1826 * Caller must hold RTNL.
1828 int fib_table_flush(struct fib_table *tb)
1830 struct trie *t = (struct trie *) tb->tb_data;
1831 struct leaf *l, *ll = NULL;
1834 for (l = trie_firstleaf(t); l; l = trie_nextleaf(l)) {
1835 found += trie_flush_leaf(l);
1837 if (ll && hlist_empty(&ll->list))
1838 trie_leaf_remove(t, ll);
1842 if (ll && hlist_empty(&ll->list))
1843 trie_leaf_remove(t, ll);
1845 pr_debug("trie_flush found=%d\n", found);
1849 void fib_free_table(struct fib_table *tb)
1854 static int fn_trie_dump_fa(t_key key, int plen, struct list_head *fah,
1855 struct fib_table *tb,
1856 struct sk_buff *skb, struct netlink_callback *cb)
1859 struct fib_alias *fa;
1860 __be32 xkey = htonl(key);
1865 /* rcu_read_lock is hold by caller */
1867 list_for_each_entry_rcu(fa, fah, fa_list) {
1873 if (fib_dump_info(skb, NETLINK_CB(cb->skb).pid,
1881 fa->fa_info, NLM_F_MULTI) < 0) {
1891 static int fn_trie_dump_leaf(struct leaf *l, struct fib_table *tb,
1892 struct sk_buff *skb, struct netlink_callback *cb)
1894 struct leaf_info *li;
1895 struct hlist_node *node;
1901 /* rcu_read_lock is hold by caller */
1902 hlist_for_each_entry_rcu(li, node, &l->list, hlist) {
1911 if (list_empty(&li->falh))
1914 if (fn_trie_dump_fa(l->key, li->plen, &li->falh, tb, skb, cb) < 0) {
1925 int fib_table_dump(struct fib_table *tb, struct sk_buff *skb,
1926 struct netlink_callback *cb)
1929 struct trie *t = (struct trie *) tb->tb_data;
1930 t_key key = cb->args[2];
1931 int count = cb->args[3];
1934 /* Dump starting at last key.
1935 * Note: 0.0.0.0/0 (ie default) is first key.
1938 l = trie_firstleaf(t);
1940 /* Normally, continue from last key, but if that is missing
1941 * fallback to using slow rescan
1943 l = fib_find_node(t, key);
1945 l = trie_leafindex(t, count);
1949 cb->args[2] = l->key;
1950 if (fn_trie_dump_leaf(l, tb, skb, cb) < 0) {
1951 cb->args[3] = count;
1957 l = trie_nextleaf(l);
1958 memset(&cb->args[4], 0,
1959 sizeof(cb->args) - 4*sizeof(cb->args[0]));
1961 cb->args[3] = count;
1967 void __init fib_trie_init(void)
1969 fn_alias_kmem = kmem_cache_create("ip_fib_alias",
1970 sizeof(struct fib_alias),
1971 0, SLAB_PANIC, NULL);
1973 trie_leaf_kmem = kmem_cache_create("ip_fib_trie",
1974 max(sizeof(struct leaf),
1975 sizeof(struct leaf_info)),
1976 0, SLAB_PANIC, NULL);
1980 struct fib_table *fib_trie_table(u32 id)
1982 struct fib_table *tb;
1985 tb = kmalloc(sizeof(struct fib_table) + sizeof(struct trie),
1991 tb->tb_default = -1;
1992 tb->tb_num_default = 0;
1994 t = (struct trie *) tb->tb_data;
1995 memset(t, 0, sizeof(*t));
2000 #ifdef CONFIG_PROC_FS
2001 /* Depth first Trie walk iterator */
2002 struct fib_trie_iter {
2003 struct seq_net_private p;
2004 struct fib_table *tb;
2005 struct tnode *tnode;
2010 static struct rt_trie_node *fib_trie_get_next(struct fib_trie_iter *iter)
2012 struct tnode *tn = iter->tnode;
2013 unsigned int cindex = iter->index;
2016 /* A single entry routing table */
2020 pr_debug("get_next iter={node=%p index=%d depth=%d}\n",
2021 iter->tnode, iter->index, iter->depth);
2023 while (cindex < (1<<tn->bits)) {
2024 struct rt_trie_node *n = tnode_get_child_rcu(tn, cindex);
2029 iter->index = cindex + 1;
2031 /* push down one level */
2032 iter->tnode = (struct tnode *) n;
2042 /* Current node exhausted, pop back up */
2043 p = node_parent_rcu((struct rt_trie_node *)tn);
2045 cindex = tkey_extract_bits(tn->key, p->pos, p->bits)+1;
2055 static struct rt_trie_node *fib_trie_get_first(struct fib_trie_iter *iter,
2058 struct rt_trie_node *n;
2063 n = rcu_dereference(t->trie);
2068 iter->tnode = (struct tnode *) n;
2080 static void trie_collect_stats(struct trie *t, struct trie_stat *s)
2082 struct rt_trie_node *n;
2083 struct fib_trie_iter iter;
2085 memset(s, 0, sizeof(*s));
2088 for (n = fib_trie_get_first(&iter, t); n; n = fib_trie_get_next(&iter)) {
2090 struct leaf *l = (struct leaf *)n;
2091 struct leaf_info *li;
2092 struct hlist_node *tmp;
2095 s->totdepth += iter.depth;
2096 if (iter.depth > s->maxdepth)
2097 s->maxdepth = iter.depth;
2099 hlist_for_each_entry_rcu(li, tmp, &l->list, hlist)
2102 const struct tnode *tn = (const struct tnode *) n;
2106 if (tn->bits < MAX_STAT_DEPTH)
2107 s->nodesizes[tn->bits]++;
2109 for (i = 0; i < (1<<tn->bits); i++)
2118 * This outputs /proc/net/fib_triestats
2120 static void trie_show_stats(struct seq_file *seq, struct trie_stat *stat)
2122 unsigned int i, max, pointers, bytes, avdepth;
2125 avdepth = stat->totdepth*100 / stat->leaves;
2129 seq_printf(seq, "\tAver depth: %u.%02d\n",
2130 avdepth / 100, avdepth % 100);
2131 seq_printf(seq, "\tMax depth: %u\n", stat->maxdepth);
2133 seq_printf(seq, "\tLeaves: %u\n", stat->leaves);
2134 bytes = sizeof(struct leaf) * stat->leaves;
2136 seq_printf(seq, "\tPrefixes: %u\n", stat->prefixes);
2137 bytes += sizeof(struct leaf_info) * stat->prefixes;
2139 seq_printf(seq, "\tInternal nodes: %u\n\t", stat->tnodes);
2140 bytes += sizeof(struct tnode) * stat->tnodes;
2142 max = MAX_STAT_DEPTH;
2143 while (max > 0 && stat->nodesizes[max-1] == 0)
2147 for (i = 1; i <= max; i++)
2148 if (stat->nodesizes[i] != 0) {
2149 seq_printf(seq, " %u: %u", i, stat->nodesizes[i]);
2150 pointers += (1<<i) * stat->nodesizes[i];
2152 seq_putc(seq, '\n');
2153 seq_printf(seq, "\tPointers: %u\n", pointers);
2155 bytes += sizeof(struct rt_trie_node *) * pointers;
2156 seq_printf(seq, "Null ptrs: %u\n", stat->nullpointers);
2157 seq_printf(seq, "Total size: %u kB\n", (bytes + 1023) / 1024);
2160 #ifdef CONFIG_IP_FIB_TRIE_STATS
2161 static void trie_show_usage(struct seq_file *seq,
2162 const struct trie_use_stats *stats)
2164 seq_printf(seq, "\nCounters:\n---------\n");
2165 seq_printf(seq, "gets = %u\n", stats->gets);
2166 seq_printf(seq, "backtracks = %u\n", stats->backtrack);
2167 seq_printf(seq, "semantic match passed = %u\n",
2168 stats->semantic_match_passed);
2169 seq_printf(seq, "semantic match miss = %u\n",
2170 stats->semantic_match_miss);
2171 seq_printf(seq, "null node hit= %u\n", stats->null_node_hit);
2172 seq_printf(seq, "skipped node resize = %u\n\n",
2173 stats->resize_node_skipped);
2175 #endif /* CONFIG_IP_FIB_TRIE_STATS */
2177 static void fib_table_print(struct seq_file *seq, struct fib_table *tb)
2179 if (tb->tb_id == RT_TABLE_LOCAL)
2180 seq_puts(seq, "Local:\n");
2181 else if (tb->tb_id == RT_TABLE_MAIN)
2182 seq_puts(seq, "Main:\n");
2184 seq_printf(seq, "Id %d:\n", tb->tb_id);
2188 static int fib_triestat_seq_show(struct seq_file *seq, void *v)
2190 struct net *net = (struct net *)seq->private;
2194 "Basic info: size of leaf:"
2195 " %Zd bytes, size of tnode: %Zd bytes.\n",
2196 sizeof(struct leaf), sizeof(struct tnode));
2198 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2199 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2200 struct hlist_node *node;
2201 struct fib_table *tb;
2203 hlist_for_each_entry_rcu(tb, node, head, tb_hlist) {
2204 struct trie *t = (struct trie *) tb->tb_data;
2205 struct trie_stat stat;
2210 fib_table_print(seq, tb);
2212 trie_collect_stats(t, &stat);
2213 trie_show_stats(seq, &stat);
2214 #ifdef CONFIG_IP_FIB_TRIE_STATS
2215 trie_show_usage(seq, &t->stats);
2223 static int fib_triestat_seq_open(struct inode *inode, struct file *file)
2225 return single_open_net(inode, file, fib_triestat_seq_show);
2228 static const struct file_operations fib_triestat_fops = {
2229 .owner = THIS_MODULE,
2230 .open = fib_triestat_seq_open,
2232 .llseek = seq_lseek,
2233 .release = single_release_net,
2236 static struct rt_trie_node *fib_trie_get_idx(struct seq_file *seq, loff_t pos)
2238 struct fib_trie_iter *iter = seq->private;
2239 struct net *net = seq_file_net(seq);
2243 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2244 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2245 struct hlist_node *node;
2246 struct fib_table *tb;
2248 hlist_for_each_entry_rcu(tb, node, head, tb_hlist) {
2249 struct rt_trie_node *n;
2251 for (n = fib_trie_get_first(iter,
2252 (struct trie *) tb->tb_data);
2253 n; n = fib_trie_get_next(iter))
2264 static void *fib_trie_seq_start(struct seq_file *seq, loff_t *pos)
2268 return fib_trie_get_idx(seq, *pos);
2271 static void *fib_trie_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2273 struct fib_trie_iter *iter = seq->private;
2274 struct net *net = seq_file_net(seq);
2275 struct fib_table *tb = iter->tb;
2276 struct hlist_node *tb_node;
2278 struct rt_trie_node *n;
2281 /* next node in same table */
2282 n = fib_trie_get_next(iter);
2286 /* walk rest of this hash chain */
2287 h = tb->tb_id & (FIB_TABLE_HASHSZ - 1);
2288 while ((tb_node = rcu_dereference(hlist_next_rcu(&tb->tb_hlist)))) {
2289 tb = hlist_entry(tb_node, struct fib_table, tb_hlist);
2290 n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2295 /* new hash chain */
2296 while (++h < FIB_TABLE_HASHSZ) {
2297 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2298 hlist_for_each_entry_rcu(tb, tb_node, head, tb_hlist) {
2299 n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2311 static void fib_trie_seq_stop(struct seq_file *seq, void *v)
2317 static void seq_indent(struct seq_file *seq, int n)
2323 static inline const char *rtn_scope(char *buf, size_t len, enum rt_scope_t s)
2326 case RT_SCOPE_UNIVERSE: return "universe";
2327 case RT_SCOPE_SITE: return "site";
2328 case RT_SCOPE_LINK: return "link";
2329 case RT_SCOPE_HOST: return "host";
2330 case RT_SCOPE_NOWHERE: return "nowhere";
2332 snprintf(buf, len, "scope=%d", s);
2337 static const char *const rtn_type_names[__RTN_MAX] = {
2338 [RTN_UNSPEC] = "UNSPEC",
2339 [RTN_UNICAST] = "UNICAST",
2340 [RTN_LOCAL] = "LOCAL",
2341 [RTN_BROADCAST] = "BROADCAST",
2342 [RTN_ANYCAST] = "ANYCAST",
2343 [RTN_MULTICAST] = "MULTICAST",
2344 [RTN_BLACKHOLE] = "BLACKHOLE",
2345 [RTN_UNREACHABLE] = "UNREACHABLE",
2346 [RTN_PROHIBIT] = "PROHIBIT",
2347 [RTN_THROW] = "THROW",
2349 [RTN_XRESOLVE] = "XRESOLVE",
2352 static inline const char *rtn_type(char *buf, size_t len, unsigned int t)
2354 if (t < __RTN_MAX && rtn_type_names[t])
2355 return rtn_type_names[t];
2356 snprintf(buf, len, "type %u", t);
2360 /* Pretty print the trie */
2361 static int fib_trie_seq_show(struct seq_file *seq, void *v)
2363 const struct fib_trie_iter *iter = seq->private;
2364 struct rt_trie_node *n = v;
2366 if (!node_parent_rcu(n))
2367 fib_table_print(seq, iter->tb);
2370 struct tnode *tn = (struct tnode *) n;
2371 __be32 prf = htonl(mask_pfx(tn->key, tn->pos));
2373 seq_indent(seq, iter->depth-1);
2374 seq_printf(seq, " +-- %pI4/%d %d %d %d\n",
2375 &prf, tn->pos, tn->bits, tn->full_children,
2376 tn->empty_children);
2379 struct leaf *l = (struct leaf *) n;
2380 struct leaf_info *li;
2381 struct hlist_node *node;
2382 __be32 val = htonl(l->key);
2384 seq_indent(seq, iter->depth);
2385 seq_printf(seq, " |-- %pI4\n", &val);
2387 hlist_for_each_entry_rcu(li, node, &l->list, hlist) {
2388 struct fib_alias *fa;
2390 list_for_each_entry_rcu(fa, &li->falh, fa_list) {
2391 char buf1[32], buf2[32];
2393 seq_indent(seq, iter->depth+1);
2394 seq_printf(seq, " /%d %s %s", li->plen,
2395 rtn_scope(buf1, sizeof(buf1),
2396 fa->fa_info->fib_scope),
2397 rtn_type(buf2, sizeof(buf2),
2400 seq_printf(seq, " tos=%d", fa->fa_tos);
2401 seq_putc(seq, '\n');
2409 static const struct seq_operations fib_trie_seq_ops = {
2410 .start = fib_trie_seq_start,
2411 .next = fib_trie_seq_next,
2412 .stop = fib_trie_seq_stop,
2413 .show = fib_trie_seq_show,
2416 static int fib_trie_seq_open(struct inode *inode, struct file *file)
2418 return seq_open_net(inode, file, &fib_trie_seq_ops,
2419 sizeof(struct fib_trie_iter));
2422 static const struct file_operations fib_trie_fops = {
2423 .owner = THIS_MODULE,
2424 .open = fib_trie_seq_open,
2426 .llseek = seq_lseek,
2427 .release = seq_release_net,
2430 struct fib_route_iter {
2431 struct seq_net_private p;
2432 struct trie *main_trie;
2437 static struct leaf *fib_route_get_idx(struct fib_route_iter *iter, loff_t pos)
2439 struct leaf *l = NULL;
2440 struct trie *t = iter->main_trie;
2442 /* use cache location of last found key */
2443 if (iter->pos > 0 && pos >= iter->pos && (l = fib_find_node(t, iter->key)))
2447 l = trie_firstleaf(t);
2450 while (l && pos-- > 0) {
2452 l = trie_nextleaf(l);
2456 iter->key = pos; /* remember it */
2458 iter->pos = 0; /* forget it */
2463 static void *fib_route_seq_start(struct seq_file *seq, loff_t *pos)
2466 struct fib_route_iter *iter = seq->private;
2467 struct fib_table *tb;
2470 tb = fib_get_table(seq_file_net(seq), RT_TABLE_MAIN);
2474 iter->main_trie = (struct trie *) tb->tb_data;
2476 return SEQ_START_TOKEN;
2478 return fib_route_get_idx(iter, *pos - 1);
2481 static void *fib_route_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2483 struct fib_route_iter *iter = seq->private;
2487 if (v == SEQ_START_TOKEN) {
2489 l = trie_firstleaf(iter->main_trie);
2492 l = trie_nextleaf(l);
2502 static void fib_route_seq_stop(struct seq_file *seq, void *v)
2508 static unsigned int fib_flag_trans(int type, __be32 mask, const struct fib_info *fi)
2510 unsigned int flags = 0;
2512 if (type == RTN_UNREACHABLE || type == RTN_PROHIBIT)
2514 if (fi && fi->fib_nh->nh_gw)
2515 flags |= RTF_GATEWAY;
2516 if (mask == htonl(0xFFFFFFFF))
2523 * This outputs /proc/net/route.
2524 * The format of the file is not supposed to be changed
2525 * and needs to be same as fib_hash output to avoid breaking
2528 static int fib_route_seq_show(struct seq_file *seq, void *v)
2531 struct leaf_info *li;
2532 struct hlist_node *node;
2534 if (v == SEQ_START_TOKEN) {
2535 seq_printf(seq, "%-127s\n", "Iface\tDestination\tGateway "
2536 "\tFlags\tRefCnt\tUse\tMetric\tMask\t\tMTU"
2541 hlist_for_each_entry_rcu(li, node, &l->list, hlist) {
2542 struct fib_alias *fa;
2543 __be32 mask, prefix;
2545 mask = inet_make_mask(li->plen);
2546 prefix = htonl(l->key);
2548 list_for_each_entry_rcu(fa, &li->falh, fa_list) {
2549 const struct fib_info *fi = fa->fa_info;
2550 unsigned int flags = fib_flag_trans(fa->fa_type, mask, fi);
2553 if (fa->fa_type == RTN_BROADCAST
2554 || fa->fa_type == RTN_MULTICAST)
2559 "%s\t%08X\t%08X\t%04X\t%d\t%u\t"
2560 "%d\t%08X\t%d\t%u\t%u%n",
2561 fi->fib_dev ? fi->fib_dev->name : "*",
2563 fi->fib_nh->nh_gw, flags, 0, 0,
2567 fi->fib_advmss + 40 : 0),
2569 fi->fib_rtt >> 3, &len);
2572 "*\t%08X\t%08X\t%04X\t%d\t%u\t"
2573 "%d\t%08X\t%d\t%u\t%u%n",
2574 prefix, 0, flags, 0, 0, 0,
2575 mask, 0, 0, 0, &len);
2577 seq_printf(seq, "%*s\n", 127 - len, "");
2584 static const struct seq_operations fib_route_seq_ops = {
2585 .start = fib_route_seq_start,
2586 .next = fib_route_seq_next,
2587 .stop = fib_route_seq_stop,
2588 .show = fib_route_seq_show,
2591 static int fib_route_seq_open(struct inode *inode, struct file *file)
2593 return seq_open_net(inode, file, &fib_route_seq_ops,
2594 sizeof(struct fib_route_iter));
2597 static const struct file_operations fib_route_fops = {
2598 .owner = THIS_MODULE,
2599 .open = fib_route_seq_open,
2601 .llseek = seq_lseek,
2602 .release = seq_release_net,
2605 int __net_init fib_proc_init(struct net *net)
2607 if (!proc_net_fops_create(net, "fib_trie", S_IRUGO, &fib_trie_fops))
2610 if (!proc_net_fops_create(net, "fib_triestat", S_IRUGO,
2611 &fib_triestat_fops))
2614 if (!proc_net_fops_create(net, "route", S_IRUGO, &fib_route_fops))
2620 proc_net_remove(net, "fib_triestat");
2622 proc_net_remove(net, "fib_trie");
2627 void __net_exit fib_proc_exit(struct net *net)
2629 proc_net_remove(net, "fib_trie");
2630 proc_net_remove(net, "fib_triestat");
2631 proc_net_remove(net, "route");
2634 #endif /* CONFIG_PROC_FS */