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/export.h>
75 #include <net/net_namespace.h>
77 #include <net/protocol.h>
78 #include <net/route.h>
81 #include <net/ip_fib.h>
82 #include "fib_lookup.h"
84 #define MAX_STAT_DEPTH 32
86 #define KEYLENGTH (8*sizeof(t_key))
87 #define KEY_MAX ((t_key)~0)
89 typedef unsigned int t_key
;
91 #define IS_TNODE(n) ((n)->bits)
92 #define IS_LEAF(n) (!(n)->bits)
94 #define get_index(_key, _kv) (((_key) ^ (_kv)->key) >> (_kv)->pos)
98 unsigned char bits
; /* 2log(KEYLENGTH) bits needed */
99 unsigned char pos
; /* 2log(KEYLENGTH) bits needed */
101 struct tnode __rcu
*parent
;
104 /* The fields in this struct are valid if bits > 0 (TNODE) */
106 t_key empty_children
; /* KEYLENGTH bits needed */
107 t_key full_children
; /* KEYLENGTH bits needed */
108 struct tnode __rcu
*child
[0];
110 /* This list pointer if valid if bits == 0 (LEAF) */
111 struct hlist_head list
;
116 struct hlist_node hlist
;
118 u32 mask_plen
; /* ntohl(inet_make_mask(plen)) */
119 struct hlist_head falh
;
123 #ifdef CONFIG_IP_FIB_TRIE_STATS
124 struct trie_use_stats
{
126 unsigned int backtrack
;
127 unsigned int semantic_match_passed
;
128 unsigned int semantic_match_miss
;
129 unsigned int null_node_hit
;
130 unsigned int resize_node_skipped
;
135 unsigned int totdepth
;
136 unsigned int maxdepth
;
139 unsigned int nullpointers
;
140 unsigned int prefixes
;
141 unsigned int nodesizes
[MAX_STAT_DEPTH
];
145 struct tnode __rcu
*trie
;
146 #ifdef CONFIG_IP_FIB_TRIE_STATS
147 struct trie_use_stats __percpu
*stats
;
151 static void resize(struct trie
*t
, struct tnode
*tn
);
152 static size_t tnode_free_size
;
155 * synchronize_rcu after call_rcu for that many pages; it should be especially
156 * useful before resizing the root node with PREEMPT_NONE configs; the value was
157 * obtained experimentally, aiming to avoid visible slowdown.
159 static const int sync_pages
= 128;
161 static struct kmem_cache
*fn_alias_kmem __read_mostly
;
162 static struct kmem_cache
*trie_leaf_kmem __read_mostly
;
164 /* caller must hold RTNL */
165 #define node_parent(n) rtnl_dereference((n)->parent)
167 /* caller must hold RCU read lock or RTNL */
168 #define node_parent_rcu(n) rcu_dereference_rtnl((n)->parent)
170 /* wrapper for rcu_assign_pointer */
171 static inline void node_set_parent(struct tnode
*n
, struct tnode
*tp
)
174 rcu_assign_pointer(n
->parent
, tp
);
177 #define NODE_INIT_PARENT(n, p) RCU_INIT_POINTER((n)->parent, p)
179 /* This provides us with the number of children in this node, in the case of a
180 * leaf this will return 0 meaning none of the children are accessible.
182 static inline unsigned long tnode_child_length(const struct tnode
*tn
)
184 return (1ul << tn
->bits
) & ~(1ul);
187 /* caller must hold RTNL */
188 static inline struct tnode
*tnode_get_child(const struct tnode
*tn
,
191 return rtnl_dereference(tn
->child
[i
]);
194 /* caller must hold RCU read lock or RTNL */
195 static inline struct tnode
*tnode_get_child_rcu(const struct tnode
*tn
,
198 return rcu_dereference_rtnl(tn
->child
[i
]);
201 /* To understand this stuff, an understanding of keys and all their bits is
202 * necessary. Every node in the trie has a key associated with it, but not
203 * all of the bits in that key are significant.
205 * Consider a node 'n' and its parent 'tp'.
207 * If n is a leaf, every bit in its key is significant. Its presence is
208 * necessitated by path compression, since during a tree traversal (when
209 * searching for a leaf - unless we are doing an insertion) we will completely
210 * ignore all skipped bits we encounter. Thus we need to verify, at the end of
211 * a potentially successful search, that we have indeed been walking the
214 * Note that we can never "miss" the correct key in the tree if present by
215 * following the wrong path. Path compression ensures that segments of the key
216 * that are the same for all keys with a given prefix are skipped, but the
217 * skipped part *is* identical for each node in the subtrie below the skipped
218 * bit! trie_insert() in this implementation takes care of that.
220 * if n is an internal node - a 'tnode' here, the various parts of its key
221 * have many different meanings.
224 * _________________________________________________________________
225 * | i | i | i | i | i | i | i | N | N | N | S | S | S | S | S | C |
226 * -----------------------------------------------------------------
227 * 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16
229 * _________________________________________________________________
230 * | C | C | C | u | u | u | u | u | u | u | u | u | u | u | u | u |
231 * -----------------------------------------------------------------
232 * 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
239 * First, let's just ignore the bits that come before the parent tp, that is
240 * the bits from (tp->pos + tp->bits) to 31. They are *known* but at this
241 * point we do not use them for anything.
243 * The bits from (tp->pos) to (tp->pos + tp->bits - 1) - "N", above - are the
244 * index into the parent's child array. That is, they will be used to find
245 * 'n' among tp's children.
247 * The bits from (n->pos + n->bits) to (tn->pos - 1) - "S" - are skipped bits
250 * All the bits we have seen so far are significant to the node n. The rest
251 * of the bits are really not needed or indeed known in n->key.
253 * The bits from (n->pos) to (n->pos + n->bits - 1) - "C" - are the index into
254 * n's child array, and will of course be different for each child.
256 * The rest of the bits, from 0 to (n->pos + n->bits), are completely unknown
260 static const int halve_threshold
= 25;
261 static const int inflate_threshold
= 50;
262 static const int halve_threshold_root
= 15;
263 static const int inflate_threshold_root
= 30;
265 static void __alias_free_mem(struct rcu_head
*head
)
267 struct fib_alias
*fa
= container_of(head
, struct fib_alias
, rcu
);
268 kmem_cache_free(fn_alias_kmem
, fa
);
271 static inline void alias_free_mem_rcu(struct fib_alias
*fa
)
273 call_rcu(&fa
->rcu
, __alias_free_mem
);
276 #define TNODE_KMALLOC_MAX \
277 ilog2((PAGE_SIZE - sizeof(struct tnode)) / sizeof(struct tnode *))
279 static void __node_free_rcu(struct rcu_head
*head
)
281 struct tnode
*n
= container_of(head
, struct tnode
, rcu
);
284 kmem_cache_free(trie_leaf_kmem
, n
);
285 else if (n
->bits
<= TNODE_KMALLOC_MAX
)
291 #define node_free(n) call_rcu(&n->rcu, __node_free_rcu)
293 static inline void free_leaf_info(struct leaf_info
*leaf
)
295 kfree_rcu(leaf
, rcu
);
298 static struct tnode
*tnode_alloc(size_t size
)
300 if (size
<= PAGE_SIZE
)
301 return kzalloc(size
, GFP_KERNEL
);
303 return vzalloc(size
);
306 static inline void empty_child_inc(struct tnode
*n
)
308 ++n
->empty_children
? : ++n
->full_children
;
311 static inline void empty_child_dec(struct tnode
*n
)
313 n
->empty_children
-- ? : n
->full_children
--;
316 static struct tnode
*leaf_new(t_key key
)
318 struct tnode
*l
= kmem_cache_alloc(trie_leaf_kmem
, GFP_KERNEL
);
321 /* set key and pos to reflect full key value
322 * any trailing zeros in the key should be ignored
323 * as the nodes are searched
328 /* set bits to 0 indicating we are not a tnode */
331 INIT_HLIST_HEAD(&l
->list
);
336 static struct leaf_info
*leaf_info_new(int plen
)
338 struct leaf_info
*li
= kmalloc(sizeof(struct leaf_info
), GFP_KERNEL
);
341 li
->mask_plen
= ntohl(inet_make_mask(plen
));
342 INIT_HLIST_HEAD(&li
->falh
);
347 static struct tnode
*tnode_new(t_key key
, int pos
, int bits
)
349 size_t sz
= offsetof(struct tnode
, child
[1ul << bits
]);
350 struct tnode
*tn
= tnode_alloc(sz
);
351 unsigned int shift
= pos
+ bits
;
353 /* verify bits and pos their msb bits clear and values are valid */
354 BUG_ON(!bits
|| (shift
> KEYLENGTH
));
361 tn
->key
= (shift
< KEYLENGTH
) ? (key
>> shift
) << shift
: 0;
362 if (bits
== KEYLENGTH
)
363 tn
->full_children
= 1;
365 tn
->empty_children
= 1ul << bits
;
368 pr_debug("AT %p s=%zu %zu\n", tn
, sizeof(struct tnode
),
369 sizeof(struct tnode
*) << bits
);
373 /* Check whether a tnode 'n' is "full", i.e. it is an internal node
374 * and no bits are skipped. See discussion in dyntree paper p. 6
376 static inline int tnode_full(const struct tnode
*tn
, const struct tnode
*n
)
378 return n
&& ((n
->pos
+ n
->bits
) == tn
->pos
) && IS_TNODE(n
);
381 /* Add a child at position i overwriting the old value.
382 * Update the value of full_children and empty_children.
384 static void put_child(struct tnode
*tn
, unsigned long i
, struct tnode
*n
)
386 struct tnode
*chi
= tnode_get_child(tn
, i
);
389 BUG_ON(i
>= tnode_child_length(tn
));
391 /* update emptyChildren, overflow into fullChildren */
392 if (n
== NULL
&& chi
!= NULL
)
394 if (n
!= NULL
&& chi
== NULL
)
397 /* update fullChildren */
398 wasfull
= tnode_full(tn
, chi
);
399 isfull
= tnode_full(tn
, n
);
401 if (wasfull
&& !isfull
)
403 else if (!wasfull
&& isfull
)
406 if (n
&& (tn
->slen
< n
->slen
))
409 rcu_assign_pointer(tn
->child
[i
], n
);
412 static void update_children(struct tnode
*tn
)
416 /* update all of the child parent pointers */
417 for (i
= tnode_child_length(tn
); i
;) {
418 struct tnode
*inode
= tnode_get_child(tn
, --i
);
423 /* Either update the children of a tnode that
424 * already belongs to us or update the child
425 * to point to ourselves.
427 if (node_parent(inode
) == tn
)
428 update_children(inode
);
430 node_set_parent(inode
, tn
);
434 static inline void put_child_root(struct tnode
*tp
, struct trie
*t
,
435 t_key key
, struct tnode
*n
)
438 put_child(tp
, get_index(key
, tp
), n
);
440 rcu_assign_pointer(t
->trie
, n
);
443 static inline void tnode_free_init(struct tnode
*tn
)
448 static inline void tnode_free_append(struct tnode
*tn
, struct tnode
*n
)
450 n
->rcu
.next
= tn
->rcu
.next
;
451 tn
->rcu
.next
= &n
->rcu
;
454 static void tnode_free(struct tnode
*tn
)
456 struct callback_head
*head
= &tn
->rcu
;
460 tnode_free_size
+= offsetof(struct tnode
, child
[1 << tn
->bits
]);
463 tn
= container_of(head
, struct tnode
, rcu
);
466 if (tnode_free_size
>= PAGE_SIZE
* sync_pages
) {
472 static void replace(struct trie
*t
, struct tnode
*oldtnode
, struct tnode
*tn
)
474 struct tnode
*tp
= node_parent(oldtnode
);
477 /* setup the parent pointer out of and back into this node */
478 NODE_INIT_PARENT(tn
, tp
);
479 put_child_root(tp
, t
, tn
->key
, tn
);
481 /* update all of the child parent pointers */
484 /* all pointers should be clean so we are done */
485 tnode_free(oldtnode
);
487 /* resize children now that oldtnode is freed */
488 for (i
= tnode_child_length(tn
); i
;) {
489 struct tnode
*inode
= tnode_get_child(tn
, --i
);
491 /* resize child node */
492 if (tnode_full(tn
, inode
))
497 static int inflate(struct trie
*t
, struct tnode
*oldtnode
)
503 pr_debug("In inflate\n");
505 tn
= tnode_new(oldtnode
->key
, oldtnode
->pos
- 1, oldtnode
->bits
+ 1);
509 /* prepare oldtnode to be freed */
510 tnode_free_init(oldtnode
);
512 /* Assemble all of the pointers in our cluster, in this case that
513 * represents all of the pointers out of our allocated nodes that
514 * point to existing tnodes and the links between our allocated
517 for (i
= tnode_child_length(oldtnode
), m
= 1u << tn
->pos
; i
;) {
518 struct tnode
*inode
= tnode_get_child(oldtnode
, --i
);
519 struct tnode
*node0
, *node1
;
526 /* A leaf or an internal node with skipped bits */
527 if (!tnode_full(oldtnode
, inode
)) {
528 put_child(tn
, get_index(inode
->key
, tn
), inode
);
532 /* drop the node in the old tnode free list */
533 tnode_free_append(oldtnode
, inode
);
535 /* An internal node with two children */
536 if (inode
->bits
== 1) {
537 put_child(tn
, 2 * i
+ 1, tnode_get_child(inode
, 1));
538 put_child(tn
, 2 * i
, tnode_get_child(inode
, 0));
542 /* We will replace this node 'inode' with two new
543 * ones, 'node0' and 'node1', each with half of the
544 * original children. The two new nodes will have
545 * a position one bit further down the key and this
546 * means that the "significant" part of their keys
547 * (see the discussion near the top of this file)
548 * will differ by one bit, which will be "0" in
549 * node0's key and "1" in node1's key. Since we are
550 * moving the key position by one step, the bit that
551 * we are moving away from - the bit at position
552 * (tn->pos) - is the one that will differ between
553 * node0 and node1. So... we synthesize that bit in the
556 node1
= tnode_new(inode
->key
| m
, inode
->pos
, inode
->bits
- 1);
559 node0
= tnode_new(inode
->key
, inode
->pos
, inode
->bits
- 1);
561 tnode_free_append(tn
, node1
);
564 tnode_free_append(tn
, node0
);
566 /* populate child pointers in new nodes */
567 for (k
= tnode_child_length(inode
), j
= k
/ 2; j
;) {
568 put_child(node1
, --j
, tnode_get_child(inode
, --k
));
569 put_child(node0
, j
, tnode_get_child(inode
, j
));
570 put_child(node1
, --j
, tnode_get_child(inode
, --k
));
571 put_child(node0
, j
, tnode_get_child(inode
, j
));
574 /* link new nodes to parent */
575 NODE_INIT_PARENT(node1
, tn
);
576 NODE_INIT_PARENT(node0
, tn
);
578 /* link parent to nodes */
579 put_child(tn
, 2 * i
+ 1, node1
);
580 put_child(tn
, 2 * i
, node0
);
583 /* setup the parent pointers into and out of this node */
584 replace(t
, oldtnode
, tn
);
588 /* all pointers should be clean so we are done */
593 static int halve(struct trie
*t
, struct tnode
*oldtnode
)
598 pr_debug("In halve\n");
600 tn
= tnode_new(oldtnode
->key
, oldtnode
->pos
+ 1, oldtnode
->bits
- 1);
604 /* prepare oldtnode to be freed */
605 tnode_free_init(oldtnode
);
607 /* Assemble all of the pointers in our cluster, in this case that
608 * represents all of the pointers out of our allocated nodes that
609 * point to existing tnodes and the links between our allocated
612 for (i
= tnode_child_length(oldtnode
); i
;) {
613 struct tnode
*node1
= tnode_get_child(oldtnode
, --i
);
614 struct tnode
*node0
= tnode_get_child(oldtnode
, --i
);
617 /* At least one of the children is empty */
618 if (!node1
|| !node0
) {
619 put_child(tn
, i
/ 2, node1
? : node0
);
623 /* Two nonempty children */
624 inode
= tnode_new(node0
->key
, oldtnode
->pos
, 1);
629 tnode_free_append(tn
, inode
);
631 /* initialize pointers out of node */
632 put_child(inode
, 1, node1
);
633 put_child(inode
, 0, node0
);
634 NODE_INIT_PARENT(inode
, tn
);
636 /* link parent to node */
637 put_child(tn
, i
/ 2, inode
);
640 /* setup the parent pointers into and out of this node */
641 replace(t
, oldtnode
, tn
);
646 static void collapse(struct trie
*t
, struct tnode
*oldtnode
)
648 struct tnode
*n
, *tp
;
651 /* scan the tnode looking for that one child that might still exist */
652 for (n
= NULL
, i
= tnode_child_length(oldtnode
); !n
&& i
;)
653 n
= tnode_get_child(oldtnode
, --i
);
655 /* compress one level */
656 tp
= node_parent(oldtnode
);
657 put_child_root(tp
, t
, oldtnode
->key
, n
);
658 node_set_parent(n
, tp
);
664 static unsigned char update_suffix(struct tnode
*tn
)
666 unsigned char slen
= tn
->pos
;
667 unsigned long stride
, i
;
669 /* search though the list of children looking for nodes that might
670 * have a suffix greater than the one we currently have. This is
671 * why we start with a stride of 2 since a stride of 1 would
672 * represent the nodes with suffix length equal to tn->pos
674 for (i
= 0, stride
= 0x2ul
; i
< tnode_child_length(tn
); i
+= stride
) {
675 struct tnode
*n
= tnode_get_child(tn
, i
);
677 if (!n
|| (n
->slen
<= slen
))
680 /* update stride and slen based on new value */
681 stride
<<= (n
->slen
- slen
);
685 /* if slen covers all but the last bit we can stop here
686 * there will be nothing longer than that since only node
687 * 0 and 1 << (bits - 1) could have that as their suffix
690 if ((slen
+ 1) >= (tn
->pos
+ tn
->bits
))
699 /* From "Implementing a dynamic compressed trie" by Stefan Nilsson of
700 * the Helsinki University of Technology and Matti Tikkanen of Nokia
701 * Telecommunications, page 6:
702 * "A node is doubled if the ratio of non-empty children to all
703 * children in the *doubled* node is at least 'high'."
705 * 'high' in this instance is the variable 'inflate_threshold'. It
706 * is expressed as a percentage, so we multiply it with
707 * tnode_child_length() and instead of multiplying by 2 (since the
708 * child array will be doubled by inflate()) and multiplying
709 * the left-hand side by 100 (to handle the percentage thing) we
710 * multiply the left-hand side by 50.
712 * The left-hand side may look a bit weird: tnode_child_length(tn)
713 * - tn->empty_children is of course the number of non-null children
714 * in the current node. tn->full_children is the number of "full"
715 * children, that is non-null tnodes with a skip value of 0.
716 * All of those will be doubled in the resulting inflated tnode, so
717 * we just count them one extra time here.
719 * A clearer way to write this would be:
721 * to_be_doubled = tn->full_children;
722 * not_to_be_doubled = tnode_child_length(tn) - tn->empty_children -
725 * new_child_length = tnode_child_length(tn) * 2;
727 * new_fill_factor = 100 * (not_to_be_doubled + 2*to_be_doubled) /
729 * if (new_fill_factor >= inflate_threshold)
731 * ...and so on, tho it would mess up the while () loop.
734 * 100 * (not_to_be_doubled + 2*to_be_doubled) / new_child_length >=
738 * 100 * (not_to_be_doubled + 2*to_be_doubled) >=
739 * inflate_threshold * new_child_length
741 * expand not_to_be_doubled and to_be_doubled, and shorten:
742 * 100 * (tnode_child_length(tn) - tn->empty_children +
743 * tn->full_children) >= inflate_threshold * new_child_length
745 * expand new_child_length:
746 * 100 * (tnode_child_length(tn) - tn->empty_children +
747 * tn->full_children) >=
748 * inflate_threshold * tnode_child_length(tn) * 2
751 * 50 * (tn->full_children + tnode_child_length(tn) -
752 * tn->empty_children) >= inflate_threshold *
753 * tnode_child_length(tn)
756 static bool should_inflate(const struct tnode
*tp
, const struct tnode
*tn
)
758 unsigned long used
= tnode_child_length(tn
);
759 unsigned long threshold
= used
;
761 /* Keep root node larger */
762 threshold
*= tp
? inflate_threshold
: inflate_threshold_root
;
763 used
-= tn
->empty_children
;
764 used
+= tn
->full_children
;
766 /* if bits == KEYLENGTH then pos = 0, and will fail below */
768 return (used
> 1) && tn
->pos
&& ((50 * used
) >= threshold
);
771 static bool should_halve(const struct tnode
*tp
, const struct tnode
*tn
)
773 unsigned long used
= tnode_child_length(tn
);
774 unsigned long threshold
= used
;
776 /* Keep root node larger */
777 threshold
*= tp
? halve_threshold
: halve_threshold_root
;
778 used
-= tn
->empty_children
;
780 /* if bits == KEYLENGTH then used = 100% on wrap, and will fail below */
782 return (used
> 1) && (tn
->bits
> 1) && ((100 * used
) < threshold
);
785 static bool should_collapse(const struct tnode
*tn
)
787 unsigned long used
= tnode_child_length(tn
);
789 used
-= tn
->empty_children
;
791 /* account for bits == KEYLENGTH case */
792 if ((tn
->bits
== KEYLENGTH
) && tn
->full_children
)
795 /* One child or none, time to drop us from the trie */
800 static void resize(struct trie
*t
, struct tnode
*tn
)
802 struct tnode
*tp
= node_parent(tn
);
803 struct tnode __rcu
**cptr
;
804 int max_work
= MAX_WORK
;
806 pr_debug("In tnode_resize %p inflate_threshold=%d threshold=%d\n",
807 tn
, inflate_threshold
, halve_threshold
);
809 /* track the tnode via the pointer from the parent instead of
810 * doing it ourselves. This way we can let RCU fully do its
811 * thing without us interfering
813 cptr
= tp
? &tp
->child
[get_index(tn
->key
, tp
)] : &t
->trie
;
814 BUG_ON(tn
!= rtnl_dereference(*cptr
));
816 /* Double as long as the resulting node has a number of
817 * nonempty nodes that are above the threshold.
819 while (should_inflate(tp
, tn
) && max_work
) {
820 if (inflate(t
, tn
)) {
821 #ifdef CONFIG_IP_FIB_TRIE_STATS
822 this_cpu_inc(t
->stats
->resize_node_skipped
);
828 tn
= rtnl_dereference(*cptr
);
831 /* Return if at least one inflate is run */
832 if (max_work
!= MAX_WORK
)
835 /* Halve as long as the number of empty children in this
836 * node is above threshold.
838 while (should_halve(tp
, tn
) && max_work
) {
840 #ifdef CONFIG_IP_FIB_TRIE_STATS
841 this_cpu_inc(t
->stats
->resize_node_skipped
);
847 tn
= rtnl_dereference(*cptr
);
850 /* Only one child remains */
851 if (should_collapse(tn
)) {
856 /* Return if at least one deflate was run */
857 if (max_work
!= MAX_WORK
)
860 /* push the suffix length to the parent node */
861 if (tn
->slen
> tn
->pos
) {
862 unsigned char slen
= update_suffix(tn
);
864 if (tp
&& (slen
> tp
->slen
))
869 /* readside must use rcu_read_lock currently dump routines
870 via get_fa_head and dump */
872 static struct leaf_info
*find_leaf_info(struct tnode
*l
, int plen
)
874 struct hlist_head
*head
= &l
->list
;
875 struct leaf_info
*li
;
877 hlist_for_each_entry_rcu(li
, head
, hlist
)
878 if (li
->plen
== plen
)
884 static inline struct hlist_head
*get_fa_head(struct tnode
*l
, int plen
)
886 struct leaf_info
*li
= find_leaf_info(l
, plen
);
894 static void leaf_pull_suffix(struct tnode
*l
)
896 struct tnode
*tp
= node_parent(l
);
898 while (tp
&& (tp
->slen
> tp
->pos
) && (tp
->slen
> l
->slen
)) {
899 if (update_suffix(tp
) > l
->slen
)
901 tp
= node_parent(tp
);
905 static void leaf_push_suffix(struct tnode
*l
)
907 struct tnode
*tn
= node_parent(l
);
909 /* if this is a new leaf then tn will be NULL and we can sort
910 * out parent suffix lengths as a part of trie_rebalance
912 while (tn
&& (tn
->slen
< l
->slen
)) {
914 tn
= node_parent(tn
);
918 static void remove_leaf_info(struct tnode
*l
, struct leaf_info
*old
)
920 /* record the location of the previous list_info entry */
921 struct hlist_node
**pprev
= old
->hlist
.pprev
;
922 struct leaf_info
*li
= hlist_entry(pprev
, typeof(*li
), hlist
.next
);
924 /* remove the leaf info from the list */
925 hlist_del_rcu(&old
->hlist
);
927 /* only access li if it is pointing at the last valid hlist_node */
928 if (hlist_empty(&l
->list
) || (*pprev
))
931 /* update the trie with the latest suffix length */
932 l
->slen
= KEYLENGTH
- li
->plen
;
936 static void insert_leaf_info(struct tnode
*l
, struct leaf_info
*new)
938 struct hlist_head
*head
= &l
->list
;
939 struct leaf_info
*li
= NULL
, *last
= NULL
;
941 if (hlist_empty(head
)) {
942 hlist_add_head_rcu(&new->hlist
, head
);
944 hlist_for_each_entry(li
, head
, hlist
) {
945 if (new->plen
> li
->plen
)
951 hlist_add_behind_rcu(&new->hlist
, &last
->hlist
);
953 hlist_add_before_rcu(&new->hlist
, &li
->hlist
);
956 /* if we added to the tail node then we need to update slen */
957 if (l
->slen
< (KEYLENGTH
- new->plen
)) {
958 l
->slen
= KEYLENGTH
- new->plen
;
963 /* rcu_read_lock needs to be hold by caller from readside */
964 static struct tnode
*fib_find_node(struct trie
*t
, u32 key
)
966 struct tnode
*n
= rcu_dereference_rtnl(t
->trie
);
969 unsigned long index
= get_index(key
, n
);
971 /* This bit of code is a bit tricky but it combines multiple
972 * checks into a single check. The prefix consists of the
973 * prefix plus zeros for the bits in the cindex. The index
974 * is the difference between the key and this value. From
975 * this we can actually derive several pieces of data.
976 * if (index & (~0ul << bits))
977 * we have a mismatch in skip bits and failed
979 * we know the value is cindex
981 if (index
& (~0ul << n
->bits
))
984 /* we have found a leaf. Prefixes have already been compared */
988 n
= tnode_get_child_rcu(n
, index
);
994 /* Return the first fib alias matching TOS with
995 * priority less than or equal to PRIO.
997 static struct fib_alias
*fib_find_alias(struct hlist_head
*fah
, u8 tos
,
1000 struct fib_alias
*fa
;
1005 hlist_for_each_entry(fa
, fah
, fa_list
) {
1006 if (fa
->fa_tos
> tos
)
1008 if (fa
->fa_info
->fib_priority
>= prio
|| fa
->fa_tos
< tos
)
1015 static void trie_rebalance(struct trie
*t
, struct tnode
*tn
)
1019 while ((tp
= node_parent(tn
)) != NULL
) {
1024 /* Handle last (top) tnode */
1029 /* only used from updater-side */
1031 static struct hlist_head
*fib_insert_node(struct trie
*t
, u32 key
, int plen
)
1033 struct hlist_head
*fa_head
= NULL
;
1034 struct tnode
*l
, *n
, *tp
= NULL
;
1035 struct leaf_info
*li
;
1037 li
= leaf_info_new(plen
);
1040 fa_head
= &li
->falh
;
1042 n
= rtnl_dereference(t
->trie
);
1044 /* If we point to NULL, stop. Either the tree is empty and we should
1045 * just put a new leaf in if, or we have reached an empty child slot,
1046 * and we should just put our new leaf in that.
1048 * If we hit a node with a key that does't match then we should stop
1049 * and create a new tnode to replace that node and insert ourselves
1050 * and the other node into the new tnode.
1053 unsigned long index
= get_index(key
, n
);
1055 /* This bit of code is a bit tricky but it combines multiple
1056 * checks into a single check. The prefix consists of the
1057 * prefix plus zeros for the "bits" in the prefix. The index
1058 * is the difference between the key and this value. From
1059 * this we can actually derive several pieces of data.
1060 * if !(index >> bits)
1061 * we know the value is child index
1063 * we have a mismatch in skip bits and failed
1065 if (index
>> n
->bits
)
1068 /* we have found a leaf. Prefixes have already been compared */
1070 /* Case 1: n is a leaf, and prefixes match*/
1071 insert_leaf_info(n
, li
);
1076 n
= tnode_get_child_rcu(n
, index
);
1085 insert_leaf_info(l
, li
);
1087 /* Case 2: n is a LEAF or a TNODE and the key doesn't match.
1089 * Add a new tnode here
1090 * first tnode need some special handling
1091 * leaves us in position for handling as case 3
1096 tn
= tnode_new(key
, __fls(key
^ n
->key
), 1);
1103 /* initialize routes out of node */
1104 NODE_INIT_PARENT(tn
, tp
);
1105 put_child(tn
, get_index(key
, tn
) ^ 1, n
);
1107 /* start adding routes into the node */
1108 put_child_root(tp
, t
, key
, tn
);
1109 node_set_parent(n
, tn
);
1111 /* parent now has a NULL spot where the leaf can go */
1115 /* Case 3: n is NULL, and will just insert a new leaf */
1117 NODE_INIT_PARENT(l
, tp
);
1118 put_child(tp
, get_index(key
, tp
), l
);
1119 trie_rebalance(t
, tp
);
1121 rcu_assign_pointer(t
->trie
, l
);
1128 * Caller must hold RTNL.
1130 int fib_table_insert(struct fib_table
*tb
, struct fib_config
*cfg
)
1132 struct trie
*t
= (struct trie
*) tb
->tb_data
;
1133 struct fib_alias
*fa
, *new_fa
;
1134 struct hlist_head
*fa_head
= NULL
;
1135 struct fib_info
*fi
;
1136 int plen
= cfg
->fc_dst_len
;
1137 u8 tos
= cfg
->fc_tos
;
1145 key
= ntohl(cfg
->fc_dst
);
1147 pr_debug("Insert table=%u %08x/%d\n", tb
->tb_id
, key
, plen
);
1149 mask
= ntohl(inet_make_mask(plen
));
1156 fi
= fib_create_info(cfg
);
1162 l
= fib_find_node(t
, key
);
1166 fa_head
= get_fa_head(l
, plen
);
1167 fa
= fib_find_alias(fa_head
, tos
, fi
->fib_priority
);
1170 /* Now fa, if non-NULL, points to the first fib alias
1171 * with the same keys [prefix,tos,priority], if such key already
1172 * exists or to the node before which we will insert new one.
1174 * If fa is NULL, we will need to allocate a new one and
1175 * insert to the tail of the section matching the suffix length
1179 if (fa
&& fa
->fa_tos
== tos
&&
1180 fa
->fa_info
->fib_priority
== fi
->fib_priority
) {
1181 struct fib_alias
*fa_first
, *fa_match
;
1184 if (cfg
->fc_nlflags
& NLM_F_EXCL
)
1188 * 1. Find exact match for type, scope, fib_info to avoid
1190 * 2. Find next 'fa' (or head), NLM_F_APPEND inserts before it
1194 hlist_for_each_entry_from(fa
, fa_list
) {
1195 if (fa
->fa_tos
!= tos
)
1197 if (fa
->fa_info
->fib_priority
!= fi
->fib_priority
)
1199 if (fa
->fa_type
== cfg
->fc_type
&&
1200 fa
->fa_info
== fi
) {
1206 if (cfg
->fc_nlflags
& NLM_F_REPLACE
) {
1207 struct fib_info
*fi_drop
;
1217 new_fa
= kmem_cache_alloc(fn_alias_kmem
, GFP_KERNEL
);
1221 fi_drop
= fa
->fa_info
;
1222 new_fa
->fa_tos
= fa
->fa_tos
;
1223 new_fa
->fa_info
= fi
;
1224 new_fa
->fa_type
= cfg
->fc_type
;
1225 state
= fa
->fa_state
;
1226 new_fa
->fa_state
= state
& ~FA_S_ACCESSED
;
1228 hlist_replace_rcu(&fa
->fa_list
, &new_fa
->fa_list
);
1229 alias_free_mem_rcu(fa
);
1231 fib_release_info(fi_drop
);
1232 if (state
& FA_S_ACCESSED
)
1233 rt_cache_flush(cfg
->fc_nlinfo
.nl_net
);
1234 rtmsg_fib(RTM_NEWROUTE
, htonl(key
), new_fa
, plen
,
1235 tb
->tb_id
, &cfg
->fc_nlinfo
, NLM_F_REPLACE
);
1239 /* Error if we find a perfect match which
1240 * uses the same scope, type, and nexthop
1246 if (!(cfg
->fc_nlflags
& NLM_F_APPEND
))
1250 if (!(cfg
->fc_nlflags
& NLM_F_CREATE
))
1254 new_fa
= kmem_cache_alloc(fn_alias_kmem
, GFP_KERNEL
);
1258 new_fa
->fa_info
= fi
;
1259 new_fa
->fa_tos
= tos
;
1260 new_fa
->fa_type
= cfg
->fc_type
;
1261 new_fa
->fa_state
= 0;
1263 * Insert new entry to the list.
1267 fa_head
= fib_insert_node(t
, key
, plen
);
1268 if (unlikely(!fa_head
)) {
1270 goto out_free_new_fa
;
1275 tb
->tb_num_default
++;
1278 hlist_add_before_rcu(&new_fa
->fa_list
, &fa
->fa_list
);
1280 struct fib_alias
*last
;
1282 hlist_for_each_entry(last
, fa_head
, fa_list
)
1286 hlist_add_behind_rcu(&new_fa
->fa_list
, &fa
->fa_list
);
1288 hlist_add_head_rcu(&new_fa
->fa_list
, fa_head
);
1291 rt_cache_flush(cfg
->fc_nlinfo
.nl_net
);
1292 rtmsg_fib(RTM_NEWROUTE
, htonl(key
), new_fa
, plen
, tb
->tb_id
,
1293 &cfg
->fc_nlinfo
, 0);
1298 kmem_cache_free(fn_alias_kmem
, new_fa
);
1300 fib_release_info(fi
);
1305 static inline t_key
prefix_mismatch(t_key key
, struct tnode
*n
)
1307 t_key prefix
= n
->key
;
1309 return (key
^ prefix
) & (prefix
| -prefix
);
1312 /* should be called with rcu_read_lock */
1313 int fib_table_lookup(struct fib_table
*tb
, const struct flowi4
*flp
,
1314 struct fib_result
*res
, int fib_flags
)
1316 struct trie
*t
= (struct trie
*)tb
->tb_data
;
1317 #ifdef CONFIG_IP_FIB_TRIE_STATS
1318 struct trie_use_stats __percpu
*stats
= t
->stats
;
1320 const t_key key
= ntohl(flp
->daddr
);
1321 struct tnode
*n
, *pn
;
1322 struct leaf_info
*li
;
1325 n
= rcu_dereference(t
->trie
);
1329 #ifdef CONFIG_IP_FIB_TRIE_STATS
1330 this_cpu_inc(stats
->gets
);
1336 /* Step 1: Travel to the longest prefix match in the trie */
1338 unsigned long index
= get_index(key
, n
);
1340 /* This bit of code is a bit tricky but it combines multiple
1341 * checks into a single check. The prefix consists of the
1342 * prefix plus zeros for the "bits" in the prefix. The index
1343 * is the difference between the key and this value. From
1344 * this we can actually derive several pieces of data.
1345 * if (index & (~0ul << bits))
1346 * we have a mismatch in skip bits and failed
1348 * we know the value is cindex
1350 if (index
& (~0ul << n
->bits
))
1353 /* we have found a leaf. Prefixes have already been compared */
1357 /* only record pn and cindex if we are going to be chopping
1358 * bits later. Otherwise we are just wasting cycles.
1360 if (n
->slen
> n
->pos
) {
1365 n
= tnode_get_child_rcu(n
, index
);
1370 /* Step 2: Sort out leaves and begin backtracing for longest prefix */
1372 /* record the pointer where our next node pointer is stored */
1373 struct tnode __rcu
**cptr
= n
->child
;
1375 /* This test verifies that none of the bits that differ
1376 * between the key and the prefix exist in the region of
1377 * the lsb and higher in the prefix.
1379 if (unlikely(prefix_mismatch(key
, n
)) || (n
->slen
== n
->pos
))
1382 /* exit out and process leaf */
1383 if (unlikely(IS_LEAF(n
)))
1386 /* Don't bother recording parent info. Since we are in
1387 * prefix match mode we will have to come back to wherever
1388 * we started this traversal anyway
1391 while ((n
= rcu_dereference(*cptr
)) == NULL
) {
1393 #ifdef CONFIG_IP_FIB_TRIE_STATS
1395 this_cpu_inc(stats
->null_node_hit
);
1397 /* If we are at cindex 0 there are no more bits for
1398 * us to strip at this level so we must ascend back
1399 * up one level to see if there are any more bits to
1400 * be stripped there.
1403 t_key pkey
= pn
->key
;
1405 pn
= node_parent_rcu(pn
);
1408 #ifdef CONFIG_IP_FIB_TRIE_STATS
1409 this_cpu_inc(stats
->backtrack
);
1411 /* Get Child's index */
1412 cindex
= get_index(pkey
, pn
);
1415 /* strip the least significant bit from the cindex */
1416 cindex
&= cindex
- 1;
1418 /* grab pointer for next child node */
1419 cptr
= &pn
->child
[cindex
];
1424 /* Step 3: Process the leaf, if that fails fall back to backtracing */
1425 hlist_for_each_entry_rcu(li
, &n
->list
, hlist
) {
1426 struct fib_alias
*fa
;
1428 if ((key
^ n
->key
) & li
->mask_plen
)
1431 hlist_for_each_entry_rcu(fa
, &li
->falh
, fa_list
) {
1432 struct fib_info
*fi
= fa
->fa_info
;
1435 if (fa
->fa_tos
&& fa
->fa_tos
!= flp
->flowi4_tos
)
1439 if (fa
->fa_info
->fib_scope
< flp
->flowi4_scope
)
1441 fib_alias_accessed(fa
);
1442 err
= fib_props
[fa
->fa_type
].error
;
1443 if (unlikely(err
< 0)) {
1444 #ifdef CONFIG_IP_FIB_TRIE_STATS
1445 this_cpu_inc(stats
->semantic_match_passed
);
1449 if (fi
->fib_flags
& RTNH_F_DEAD
)
1451 for (nhsel
= 0; nhsel
< fi
->fib_nhs
; nhsel
++) {
1452 const struct fib_nh
*nh
= &fi
->fib_nh
[nhsel
];
1454 if (nh
->nh_flags
& RTNH_F_DEAD
)
1456 if (flp
->flowi4_oif
&& flp
->flowi4_oif
!= nh
->nh_oif
)
1459 if (!(fib_flags
& FIB_LOOKUP_NOREF
))
1460 atomic_inc(&fi
->fib_clntref
);
1462 res
->prefixlen
= li
->plen
;
1463 res
->nh_sel
= nhsel
;
1464 res
->type
= fa
->fa_type
;
1465 res
->scope
= fi
->fib_scope
;
1468 res
->fa_head
= &li
->falh
;
1469 #ifdef CONFIG_IP_FIB_TRIE_STATS
1470 this_cpu_inc(stats
->semantic_match_passed
);
1476 #ifdef CONFIG_IP_FIB_TRIE_STATS
1477 this_cpu_inc(stats
->semantic_match_miss
);
1482 EXPORT_SYMBOL_GPL(fib_table_lookup
);
1485 * Remove the leaf and return parent.
1487 static void trie_leaf_remove(struct trie
*t
, struct tnode
*l
)
1489 struct tnode
*tp
= node_parent(l
);
1491 pr_debug("entering trie_leaf_remove(%p)\n", l
);
1494 put_child(tp
, get_index(l
->key
, tp
), NULL
);
1495 trie_rebalance(t
, tp
);
1497 RCU_INIT_POINTER(t
->trie
, NULL
);
1504 * Caller must hold RTNL.
1506 int fib_table_delete(struct fib_table
*tb
, struct fib_config
*cfg
)
1508 struct trie
*t
= (struct trie
*) tb
->tb_data
;
1510 int plen
= cfg
->fc_dst_len
;
1511 u8 tos
= cfg
->fc_tos
;
1512 struct fib_alias
*fa
, *fa_to_delete
;
1513 struct hlist_head
*fa_head
;
1515 struct leaf_info
*li
;
1520 key
= ntohl(cfg
->fc_dst
);
1521 mask
= ntohl(inet_make_mask(plen
));
1527 l
= fib_find_node(t
, key
);
1532 li
= find_leaf_info(l
, plen
);
1537 fa_head
= &li
->falh
;
1538 fa
= fib_find_alias(fa_head
, tos
, 0);
1543 pr_debug("Deleting %08x/%d tos=%d t=%p\n", key
, plen
, tos
, t
);
1545 fa_to_delete
= NULL
;
1546 hlist_for_each_entry_from(fa
, fa_list
) {
1547 struct fib_info
*fi
= fa
->fa_info
;
1549 if (fa
->fa_tos
!= tos
)
1552 if ((!cfg
->fc_type
|| fa
->fa_type
== cfg
->fc_type
) &&
1553 (cfg
->fc_scope
== RT_SCOPE_NOWHERE
||
1554 fa
->fa_info
->fib_scope
== cfg
->fc_scope
) &&
1555 (!cfg
->fc_prefsrc
||
1556 fi
->fib_prefsrc
== cfg
->fc_prefsrc
) &&
1557 (!cfg
->fc_protocol
||
1558 fi
->fib_protocol
== cfg
->fc_protocol
) &&
1559 fib_nh_match(cfg
, fi
) == 0) {
1569 rtmsg_fib(RTM_DELROUTE
, htonl(key
), fa
, plen
, tb
->tb_id
,
1570 &cfg
->fc_nlinfo
, 0);
1572 hlist_del_rcu(&fa
->fa_list
);
1575 tb
->tb_num_default
--;
1577 if (hlist_empty(fa_head
)) {
1578 remove_leaf_info(l
, li
);
1582 if (hlist_empty(&l
->list
))
1583 trie_leaf_remove(t
, l
);
1585 if (fa
->fa_state
& FA_S_ACCESSED
)
1586 rt_cache_flush(cfg
->fc_nlinfo
.nl_net
);
1588 fib_release_info(fa
->fa_info
);
1589 alias_free_mem_rcu(fa
);
1593 static int trie_flush_list(struct hlist_head
*head
)
1595 struct hlist_node
*tmp
;
1596 struct fib_alias
*fa
;
1599 hlist_for_each_entry_safe(fa
, tmp
, head
, fa_list
) {
1600 struct fib_info
*fi
= fa
->fa_info
;
1602 if (fi
&& (fi
->fib_flags
& RTNH_F_DEAD
)) {
1603 hlist_del_rcu(&fa
->fa_list
);
1604 fib_release_info(fa
->fa_info
);
1605 alias_free_mem_rcu(fa
);
1612 static int trie_flush_leaf(struct tnode
*l
)
1615 struct hlist_node
*tmp
;
1616 struct leaf_info
*li
;
1617 unsigned char plen
= KEYLENGTH
;
1619 hlist_for_each_entry_safe(li
, tmp
, &l
->list
, hlist
) {
1620 found
+= trie_flush_list(&li
->falh
);
1622 if (hlist_empty(&li
->falh
)) {
1623 hlist_del_rcu(&li
->hlist
);
1631 l
->slen
= KEYLENGTH
- plen
;
1637 * Scan for the next right leaf starting at node p->child[idx]
1638 * Since we have back pointer, no recursion necessary.
1640 static struct tnode
*leaf_walk_rcu(struct tnode
*p
, struct tnode
*c
)
1643 unsigned long idx
= c
? idx
= get_index(c
->key
, p
) + 1 : 0;
1645 while (idx
< tnode_child_length(p
)) {
1646 c
= tnode_get_child_rcu(p
, idx
++);
1653 /* Rescan start scanning in new node */
1658 /* Node empty, walk back up to parent */
1660 } while ((p
= node_parent_rcu(c
)) != NULL
);
1662 return NULL
; /* Root of trie */
1665 static struct tnode
*trie_firstleaf(struct trie
*t
)
1667 struct tnode
*n
= rcu_dereference_rtnl(t
->trie
);
1672 if (IS_LEAF(n
)) /* trie is just a leaf */
1675 return leaf_walk_rcu(n
, NULL
);
1678 static struct tnode
*trie_nextleaf(struct tnode
*l
)
1680 struct tnode
*p
= node_parent_rcu(l
);
1683 return NULL
; /* trie with just one leaf */
1685 return leaf_walk_rcu(p
, l
);
1688 static struct tnode
*trie_leafindex(struct trie
*t
, int index
)
1690 struct tnode
*l
= trie_firstleaf(t
);
1692 while (l
&& index
-- > 0)
1693 l
= trie_nextleaf(l
);
1700 * Caller must hold RTNL.
1702 int fib_table_flush(struct fib_table
*tb
)
1704 struct trie
*t
= (struct trie
*) tb
->tb_data
;
1705 struct tnode
*l
, *ll
= NULL
;
1708 for (l
= trie_firstleaf(t
); l
; l
= trie_nextleaf(l
)) {
1709 found
+= trie_flush_leaf(l
);
1712 if (hlist_empty(&ll
->list
))
1713 trie_leaf_remove(t
, ll
);
1715 leaf_pull_suffix(ll
);
1722 if (hlist_empty(&ll
->list
))
1723 trie_leaf_remove(t
, ll
);
1725 leaf_pull_suffix(ll
);
1728 pr_debug("trie_flush found=%d\n", found
);
1732 void fib_free_table(struct fib_table
*tb
)
1734 #ifdef CONFIG_IP_FIB_TRIE_STATS
1735 struct trie
*t
= (struct trie
*)tb
->tb_data
;
1737 free_percpu(t
->stats
);
1738 #endif /* CONFIG_IP_FIB_TRIE_STATS */
1742 static int fn_trie_dump_fa(t_key key
, int plen
, struct hlist_head
*fah
,
1743 struct fib_table
*tb
,
1744 struct sk_buff
*skb
, struct netlink_callback
*cb
)
1747 struct fib_alias
*fa
;
1748 __be32 xkey
= htonl(key
);
1753 /* rcu_read_lock is hold by caller */
1755 hlist_for_each_entry_rcu(fa
, fah
, fa_list
) {
1761 if (fib_dump_info(skb
, NETLINK_CB(cb
->skb
).portid
,
1769 fa
->fa_info
, NLM_F_MULTI
) < 0) {
1779 static int fn_trie_dump_leaf(struct tnode
*l
, struct fib_table
*tb
,
1780 struct sk_buff
*skb
, struct netlink_callback
*cb
)
1782 struct leaf_info
*li
;
1788 /* rcu_read_lock is hold by caller */
1789 hlist_for_each_entry_rcu(li
, &l
->list
, hlist
) {
1798 if (hlist_empty(&li
->falh
))
1801 if (fn_trie_dump_fa(l
->key
, li
->plen
, &li
->falh
, tb
, skb
, cb
) < 0) {
1812 int fib_table_dump(struct fib_table
*tb
, struct sk_buff
*skb
,
1813 struct netlink_callback
*cb
)
1816 struct trie
*t
= (struct trie
*) tb
->tb_data
;
1817 t_key key
= cb
->args
[2];
1818 int count
= cb
->args
[3];
1821 /* Dump starting at last key.
1822 * Note: 0.0.0.0/0 (ie default) is first key.
1825 l
= trie_firstleaf(t
);
1827 /* Normally, continue from last key, but if that is missing
1828 * fallback to using slow rescan
1830 l
= fib_find_node(t
, key
);
1832 l
= trie_leafindex(t
, count
);
1836 cb
->args
[2] = l
->key
;
1837 if (fn_trie_dump_leaf(l
, tb
, skb
, cb
) < 0) {
1838 cb
->args
[3] = count
;
1844 l
= trie_nextleaf(l
);
1845 memset(&cb
->args
[4], 0,
1846 sizeof(cb
->args
) - 4*sizeof(cb
->args
[0]));
1848 cb
->args
[3] = count
;
1854 void __init
fib_trie_init(void)
1856 fn_alias_kmem
= kmem_cache_create("ip_fib_alias",
1857 sizeof(struct fib_alias
),
1858 0, SLAB_PANIC
, NULL
);
1860 trie_leaf_kmem
= kmem_cache_create("ip_fib_trie",
1861 max(sizeof(struct tnode
),
1862 sizeof(struct leaf_info
)),
1863 0, SLAB_PANIC
, NULL
);
1867 struct fib_table
*fib_trie_table(u32 id
)
1869 struct fib_table
*tb
;
1872 tb
= kmalloc(sizeof(struct fib_table
) + sizeof(struct trie
),
1878 tb
->tb_default
= -1;
1879 tb
->tb_num_default
= 0;
1881 t
= (struct trie
*) tb
->tb_data
;
1882 RCU_INIT_POINTER(t
->trie
, NULL
);
1883 #ifdef CONFIG_IP_FIB_TRIE_STATS
1884 t
->stats
= alloc_percpu(struct trie_use_stats
);
1894 #ifdef CONFIG_PROC_FS
1895 /* Depth first Trie walk iterator */
1896 struct fib_trie_iter
{
1897 struct seq_net_private p
;
1898 struct fib_table
*tb
;
1899 struct tnode
*tnode
;
1904 static struct tnode
*fib_trie_get_next(struct fib_trie_iter
*iter
)
1906 unsigned long cindex
= iter
->index
;
1907 struct tnode
*tn
= iter
->tnode
;
1910 /* A single entry routing table */
1914 pr_debug("get_next iter={node=%p index=%d depth=%d}\n",
1915 iter
->tnode
, iter
->index
, iter
->depth
);
1917 while (cindex
< tnode_child_length(tn
)) {
1918 struct tnode
*n
= tnode_get_child_rcu(tn
, cindex
);
1923 iter
->index
= cindex
+ 1;
1925 /* push down one level */
1936 /* Current node exhausted, pop back up */
1937 p
= node_parent_rcu(tn
);
1939 cindex
= get_index(tn
->key
, p
) + 1;
1949 static struct tnode
*fib_trie_get_first(struct fib_trie_iter
*iter
,
1957 n
= rcu_dereference(t
->trie
);
1974 static void trie_collect_stats(struct trie
*t
, struct trie_stat
*s
)
1977 struct fib_trie_iter iter
;
1979 memset(s
, 0, sizeof(*s
));
1982 for (n
= fib_trie_get_first(&iter
, t
); n
; n
= fib_trie_get_next(&iter
)) {
1984 struct leaf_info
*li
;
1987 s
->totdepth
+= iter
.depth
;
1988 if (iter
.depth
> s
->maxdepth
)
1989 s
->maxdepth
= iter
.depth
;
1991 hlist_for_each_entry_rcu(li
, &n
->list
, hlist
)
1995 if (n
->bits
< MAX_STAT_DEPTH
)
1996 s
->nodesizes
[n
->bits
]++;
1997 s
->nullpointers
+= n
->empty_children
;
2004 * This outputs /proc/net/fib_triestats
2006 static void trie_show_stats(struct seq_file
*seq
, struct trie_stat
*stat
)
2008 unsigned int i
, max
, pointers
, bytes
, avdepth
;
2011 avdepth
= stat
->totdepth
*100 / stat
->leaves
;
2015 seq_printf(seq
, "\tAver depth: %u.%02d\n",
2016 avdepth
/ 100, avdepth
% 100);
2017 seq_printf(seq
, "\tMax depth: %u\n", stat
->maxdepth
);
2019 seq_printf(seq
, "\tLeaves: %u\n", stat
->leaves
);
2020 bytes
= sizeof(struct tnode
) * stat
->leaves
;
2022 seq_printf(seq
, "\tPrefixes: %u\n", stat
->prefixes
);
2023 bytes
+= sizeof(struct leaf_info
) * stat
->prefixes
;
2025 seq_printf(seq
, "\tInternal nodes: %u\n\t", stat
->tnodes
);
2026 bytes
+= sizeof(struct tnode
) * stat
->tnodes
;
2028 max
= MAX_STAT_DEPTH
;
2029 while (max
> 0 && stat
->nodesizes
[max
-1] == 0)
2033 for (i
= 1; i
< max
; i
++)
2034 if (stat
->nodesizes
[i
] != 0) {
2035 seq_printf(seq
, " %u: %u", i
, stat
->nodesizes
[i
]);
2036 pointers
+= (1<<i
) * stat
->nodesizes
[i
];
2038 seq_putc(seq
, '\n');
2039 seq_printf(seq
, "\tPointers: %u\n", pointers
);
2041 bytes
+= sizeof(struct tnode
*) * pointers
;
2042 seq_printf(seq
, "Null ptrs: %u\n", stat
->nullpointers
);
2043 seq_printf(seq
, "Total size: %u kB\n", (bytes
+ 1023) / 1024);
2046 #ifdef CONFIG_IP_FIB_TRIE_STATS
2047 static void trie_show_usage(struct seq_file
*seq
,
2048 const struct trie_use_stats __percpu
*stats
)
2050 struct trie_use_stats s
= { 0 };
2053 /* loop through all of the CPUs and gather up the stats */
2054 for_each_possible_cpu(cpu
) {
2055 const struct trie_use_stats
*pcpu
= per_cpu_ptr(stats
, cpu
);
2057 s
.gets
+= pcpu
->gets
;
2058 s
.backtrack
+= pcpu
->backtrack
;
2059 s
.semantic_match_passed
+= pcpu
->semantic_match_passed
;
2060 s
.semantic_match_miss
+= pcpu
->semantic_match_miss
;
2061 s
.null_node_hit
+= pcpu
->null_node_hit
;
2062 s
.resize_node_skipped
+= pcpu
->resize_node_skipped
;
2065 seq_printf(seq
, "\nCounters:\n---------\n");
2066 seq_printf(seq
, "gets = %u\n", s
.gets
);
2067 seq_printf(seq
, "backtracks = %u\n", s
.backtrack
);
2068 seq_printf(seq
, "semantic match passed = %u\n",
2069 s
.semantic_match_passed
);
2070 seq_printf(seq
, "semantic match miss = %u\n", s
.semantic_match_miss
);
2071 seq_printf(seq
, "null node hit= %u\n", s
.null_node_hit
);
2072 seq_printf(seq
, "skipped node resize = %u\n\n", s
.resize_node_skipped
);
2074 #endif /* CONFIG_IP_FIB_TRIE_STATS */
2076 static void fib_table_print(struct seq_file
*seq
, struct fib_table
*tb
)
2078 if (tb
->tb_id
== RT_TABLE_LOCAL
)
2079 seq_puts(seq
, "Local:\n");
2080 else if (tb
->tb_id
== RT_TABLE_MAIN
)
2081 seq_puts(seq
, "Main:\n");
2083 seq_printf(seq
, "Id %d:\n", tb
->tb_id
);
2087 static int fib_triestat_seq_show(struct seq_file
*seq
, void *v
)
2089 struct net
*net
= (struct net
*)seq
->private;
2093 "Basic info: size of leaf:"
2094 " %Zd bytes, size of tnode: %Zd bytes.\n",
2095 sizeof(struct tnode
), sizeof(struct tnode
));
2097 for (h
= 0; h
< FIB_TABLE_HASHSZ
; h
++) {
2098 struct hlist_head
*head
= &net
->ipv4
.fib_table_hash
[h
];
2099 struct fib_table
*tb
;
2101 hlist_for_each_entry_rcu(tb
, head
, tb_hlist
) {
2102 struct trie
*t
= (struct trie
*) tb
->tb_data
;
2103 struct trie_stat stat
;
2108 fib_table_print(seq
, tb
);
2110 trie_collect_stats(t
, &stat
);
2111 trie_show_stats(seq
, &stat
);
2112 #ifdef CONFIG_IP_FIB_TRIE_STATS
2113 trie_show_usage(seq
, t
->stats
);
2121 static int fib_triestat_seq_open(struct inode
*inode
, struct file
*file
)
2123 return single_open_net(inode
, file
, fib_triestat_seq_show
);
2126 static const struct file_operations fib_triestat_fops
= {
2127 .owner
= THIS_MODULE
,
2128 .open
= fib_triestat_seq_open
,
2130 .llseek
= seq_lseek
,
2131 .release
= single_release_net
,
2134 static struct tnode
*fib_trie_get_idx(struct seq_file
*seq
, loff_t pos
)
2136 struct fib_trie_iter
*iter
= seq
->private;
2137 struct net
*net
= seq_file_net(seq
);
2141 for (h
= 0; h
< FIB_TABLE_HASHSZ
; h
++) {
2142 struct hlist_head
*head
= &net
->ipv4
.fib_table_hash
[h
];
2143 struct fib_table
*tb
;
2145 hlist_for_each_entry_rcu(tb
, head
, tb_hlist
) {
2148 for (n
= fib_trie_get_first(iter
,
2149 (struct trie
*) tb
->tb_data
);
2150 n
; n
= fib_trie_get_next(iter
))
2161 static void *fib_trie_seq_start(struct seq_file
*seq
, loff_t
*pos
)
2165 return fib_trie_get_idx(seq
, *pos
);
2168 static void *fib_trie_seq_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2170 struct fib_trie_iter
*iter
= seq
->private;
2171 struct net
*net
= seq_file_net(seq
);
2172 struct fib_table
*tb
= iter
->tb
;
2173 struct hlist_node
*tb_node
;
2178 /* next node in same table */
2179 n
= fib_trie_get_next(iter
);
2183 /* walk rest of this hash chain */
2184 h
= tb
->tb_id
& (FIB_TABLE_HASHSZ
- 1);
2185 while ((tb_node
= rcu_dereference(hlist_next_rcu(&tb
->tb_hlist
)))) {
2186 tb
= hlist_entry(tb_node
, struct fib_table
, tb_hlist
);
2187 n
= fib_trie_get_first(iter
, (struct trie
*) tb
->tb_data
);
2192 /* new hash chain */
2193 while (++h
< FIB_TABLE_HASHSZ
) {
2194 struct hlist_head
*head
= &net
->ipv4
.fib_table_hash
[h
];
2195 hlist_for_each_entry_rcu(tb
, head
, tb_hlist
) {
2196 n
= fib_trie_get_first(iter
, (struct trie
*) tb
->tb_data
);
2208 static void fib_trie_seq_stop(struct seq_file
*seq
, void *v
)
2214 static void seq_indent(struct seq_file
*seq
, int n
)
2220 static inline const char *rtn_scope(char *buf
, size_t len
, enum rt_scope_t s
)
2223 case RT_SCOPE_UNIVERSE
: return "universe";
2224 case RT_SCOPE_SITE
: return "site";
2225 case RT_SCOPE_LINK
: return "link";
2226 case RT_SCOPE_HOST
: return "host";
2227 case RT_SCOPE_NOWHERE
: return "nowhere";
2229 snprintf(buf
, len
, "scope=%d", s
);
2234 static const char *const rtn_type_names
[__RTN_MAX
] = {
2235 [RTN_UNSPEC
] = "UNSPEC",
2236 [RTN_UNICAST
] = "UNICAST",
2237 [RTN_LOCAL
] = "LOCAL",
2238 [RTN_BROADCAST
] = "BROADCAST",
2239 [RTN_ANYCAST
] = "ANYCAST",
2240 [RTN_MULTICAST
] = "MULTICAST",
2241 [RTN_BLACKHOLE
] = "BLACKHOLE",
2242 [RTN_UNREACHABLE
] = "UNREACHABLE",
2243 [RTN_PROHIBIT
] = "PROHIBIT",
2244 [RTN_THROW
] = "THROW",
2246 [RTN_XRESOLVE
] = "XRESOLVE",
2249 static inline const char *rtn_type(char *buf
, size_t len
, unsigned int t
)
2251 if (t
< __RTN_MAX
&& rtn_type_names
[t
])
2252 return rtn_type_names
[t
];
2253 snprintf(buf
, len
, "type %u", t
);
2257 /* Pretty print the trie */
2258 static int fib_trie_seq_show(struct seq_file
*seq
, void *v
)
2260 const struct fib_trie_iter
*iter
= seq
->private;
2261 struct tnode
*n
= v
;
2263 if (!node_parent_rcu(n
))
2264 fib_table_print(seq
, iter
->tb
);
2267 __be32 prf
= htonl(n
->key
);
2269 seq_indent(seq
, iter
->depth
-1);
2270 seq_printf(seq
, " +-- %pI4/%zu %u %u %u\n",
2271 &prf
, KEYLENGTH
- n
->pos
- n
->bits
, n
->bits
,
2272 n
->full_children
, n
->empty_children
);
2274 struct leaf_info
*li
;
2275 __be32 val
= htonl(n
->key
);
2277 seq_indent(seq
, iter
->depth
);
2278 seq_printf(seq
, " |-- %pI4\n", &val
);
2280 hlist_for_each_entry_rcu(li
, &n
->list
, hlist
) {
2281 struct fib_alias
*fa
;
2283 hlist_for_each_entry_rcu(fa
, &li
->falh
, fa_list
) {
2284 char buf1
[32], buf2
[32];
2286 seq_indent(seq
, iter
->depth
+1);
2287 seq_printf(seq
, " /%d %s %s", li
->plen
,
2288 rtn_scope(buf1
, sizeof(buf1
),
2289 fa
->fa_info
->fib_scope
),
2290 rtn_type(buf2
, sizeof(buf2
),
2293 seq_printf(seq
, " tos=%d", fa
->fa_tos
);
2294 seq_putc(seq
, '\n');
2302 static const struct seq_operations fib_trie_seq_ops
= {
2303 .start
= fib_trie_seq_start
,
2304 .next
= fib_trie_seq_next
,
2305 .stop
= fib_trie_seq_stop
,
2306 .show
= fib_trie_seq_show
,
2309 static int fib_trie_seq_open(struct inode
*inode
, struct file
*file
)
2311 return seq_open_net(inode
, file
, &fib_trie_seq_ops
,
2312 sizeof(struct fib_trie_iter
));
2315 static const struct file_operations fib_trie_fops
= {
2316 .owner
= THIS_MODULE
,
2317 .open
= fib_trie_seq_open
,
2319 .llseek
= seq_lseek
,
2320 .release
= seq_release_net
,
2323 struct fib_route_iter
{
2324 struct seq_net_private p
;
2325 struct trie
*main_trie
;
2330 static struct tnode
*fib_route_get_idx(struct fib_route_iter
*iter
, loff_t pos
)
2332 struct tnode
*l
= NULL
;
2333 struct trie
*t
= iter
->main_trie
;
2335 /* use cache location of last found key */
2336 if (iter
->pos
> 0 && pos
>= iter
->pos
&& (l
= fib_find_node(t
, iter
->key
)))
2340 l
= trie_firstleaf(t
);
2343 while (l
&& pos
-- > 0) {
2345 l
= trie_nextleaf(l
);
2349 iter
->key
= pos
; /* remember it */
2351 iter
->pos
= 0; /* forget it */
2356 static void *fib_route_seq_start(struct seq_file
*seq
, loff_t
*pos
)
2359 struct fib_route_iter
*iter
= seq
->private;
2360 struct fib_table
*tb
;
2363 tb
= fib_get_table(seq_file_net(seq
), RT_TABLE_MAIN
);
2367 iter
->main_trie
= (struct trie
*) tb
->tb_data
;
2369 return SEQ_START_TOKEN
;
2371 return fib_route_get_idx(iter
, *pos
- 1);
2374 static void *fib_route_seq_next(struct seq_file
*seq
, void *v
, loff_t
*pos
)
2376 struct fib_route_iter
*iter
= seq
->private;
2377 struct tnode
*l
= v
;
2380 if (v
== SEQ_START_TOKEN
) {
2382 l
= trie_firstleaf(iter
->main_trie
);
2385 l
= trie_nextleaf(l
);
2395 static void fib_route_seq_stop(struct seq_file
*seq
, void *v
)
2401 static unsigned int fib_flag_trans(int type
, __be32 mask
, const struct fib_info
*fi
)
2403 unsigned int flags
= 0;
2405 if (type
== RTN_UNREACHABLE
|| type
== RTN_PROHIBIT
)
2407 if (fi
&& fi
->fib_nh
->nh_gw
)
2408 flags
|= RTF_GATEWAY
;
2409 if (mask
== htonl(0xFFFFFFFF))
2416 * This outputs /proc/net/route.
2417 * The format of the file is not supposed to be changed
2418 * and needs to be same as fib_hash output to avoid breaking
2421 static int fib_route_seq_show(struct seq_file
*seq
, void *v
)
2423 struct tnode
*l
= v
;
2424 struct leaf_info
*li
;
2426 if (v
== SEQ_START_TOKEN
) {
2427 seq_printf(seq
, "%-127s\n", "Iface\tDestination\tGateway "
2428 "\tFlags\tRefCnt\tUse\tMetric\tMask\t\tMTU"
2433 hlist_for_each_entry_rcu(li
, &l
->list
, hlist
) {
2434 struct fib_alias
*fa
;
2435 __be32 mask
, prefix
;
2437 mask
= inet_make_mask(li
->plen
);
2438 prefix
= htonl(l
->key
);
2440 hlist_for_each_entry_rcu(fa
, &li
->falh
, fa_list
) {
2441 const struct fib_info
*fi
= fa
->fa_info
;
2442 unsigned int flags
= fib_flag_trans(fa
->fa_type
, mask
, fi
);
2444 if (fa
->fa_type
== RTN_BROADCAST
2445 || fa
->fa_type
== RTN_MULTICAST
)
2448 seq_setwidth(seq
, 127);
2452 "%s\t%08X\t%08X\t%04X\t%d\t%u\t"
2453 "%d\t%08X\t%d\t%u\t%u",
2454 fi
->fib_dev
? fi
->fib_dev
->name
: "*",
2456 fi
->fib_nh
->nh_gw
, flags
, 0, 0,
2460 fi
->fib_advmss
+ 40 : 0),
2465 "*\t%08X\t%08X\t%04X\t%d\t%u\t"
2466 "%d\t%08X\t%d\t%u\t%u",
2467 prefix
, 0, flags
, 0, 0, 0,
2477 static const struct seq_operations fib_route_seq_ops
= {
2478 .start
= fib_route_seq_start
,
2479 .next
= fib_route_seq_next
,
2480 .stop
= fib_route_seq_stop
,
2481 .show
= fib_route_seq_show
,
2484 static int fib_route_seq_open(struct inode
*inode
, struct file
*file
)
2486 return seq_open_net(inode
, file
, &fib_route_seq_ops
,
2487 sizeof(struct fib_route_iter
));
2490 static const struct file_operations fib_route_fops
= {
2491 .owner
= THIS_MODULE
,
2492 .open
= fib_route_seq_open
,
2494 .llseek
= seq_lseek
,
2495 .release
= seq_release_net
,
2498 int __net_init
fib_proc_init(struct net
*net
)
2500 if (!proc_create("fib_trie", S_IRUGO
, net
->proc_net
, &fib_trie_fops
))
2503 if (!proc_create("fib_triestat", S_IRUGO
, net
->proc_net
,
2504 &fib_triestat_fops
))
2507 if (!proc_create("route", S_IRUGO
, net
->proc_net
, &fib_route_fops
))
2513 remove_proc_entry("fib_triestat", net
->proc_net
);
2515 remove_proc_entry("fib_trie", net
->proc_net
);
2520 void __net_exit
fib_proc_exit(struct net
*net
)
2522 remove_proc_entry("fib_trie", net
->proc_net
);
2523 remove_proc_entry("fib_triestat", net
->proc_net
);
2524 remove_proc_entry("route", net
->proc_net
);
2527 #endif /* CONFIG_PROC_FS */