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 <linux/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 <linux/vmalloc.h>
76 #include <linux/notifier.h>
77 #include <net/net_namespace.h>
79 #include <net/protocol.h>
80 #include <net/route.h>
83 #include <net/ip_fib.h>
84 #include <trace/events/fib.h>
85 #include "fib_lookup.h"
87 static int call_fib_entry_notifier(struct notifier_block *nb, struct net *net,
88 enum fib_event_type event_type, u32 dst,
89 int dst_len, struct fib_info *fi,
90 u8 tos, u8 type, u32 tb_id)
92 struct fib_entry_notifier_info info = {
100 return call_fib_notifier(nb, net, event_type, &info.info);
103 static int call_fib_entry_notifiers(struct net *net,
104 enum fib_event_type event_type, u32 dst,
105 int dst_len, struct fib_info *fi,
106 u8 tos, u8 type, u32 tb_id)
108 struct fib_entry_notifier_info info = {
116 return call_fib_notifiers(net, event_type, &info.info);
119 #define MAX_STAT_DEPTH 32
121 #define KEYLENGTH (8*sizeof(t_key))
122 #define KEY_MAX ((t_key)~0)
124 typedef unsigned int t_key;
126 #define IS_TRIE(n) ((n)->pos >= KEYLENGTH)
127 #define IS_TNODE(n) ((n)->bits)
128 #define IS_LEAF(n) (!(n)->bits)
132 unsigned char pos; /* 2log(KEYLENGTH) bits needed */
133 unsigned char bits; /* 2log(KEYLENGTH) bits needed */
136 /* This list pointer if valid if (pos | bits) == 0 (LEAF) */
137 struct hlist_head leaf;
138 /* This array is valid if (pos | bits) > 0 (TNODE) */
139 struct key_vector __rcu *tnode[0];
145 t_key empty_children; /* KEYLENGTH bits needed */
146 t_key full_children; /* KEYLENGTH bits needed */
147 struct key_vector __rcu *parent;
148 struct key_vector kv[1];
149 #define tn_bits kv[0].bits
152 #define TNODE_SIZE(n) offsetof(struct tnode, kv[0].tnode[n])
153 #define LEAF_SIZE TNODE_SIZE(1)
155 #ifdef CONFIG_IP_FIB_TRIE_STATS
156 struct trie_use_stats {
158 unsigned int backtrack;
159 unsigned int semantic_match_passed;
160 unsigned int semantic_match_miss;
161 unsigned int null_node_hit;
162 unsigned int resize_node_skipped;
167 unsigned int totdepth;
168 unsigned int maxdepth;
171 unsigned int nullpointers;
172 unsigned int prefixes;
173 unsigned int nodesizes[MAX_STAT_DEPTH];
177 struct key_vector kv[1];
178 #ifdef CONFIG_IP_FIB_TRIE_STATS
179 struct trie_use_stats __percpu *stats;
183 static struct key_vector *resize(struct trie *t, struct key_vector *tn);
184 static size_t tnode_free_size;
187 * synchronize_rcu after call_rcu for that many pages; it should be especially
188 * useful before resizing the root node with PREEMPT_NONE configs; the value was
189 * obtained experimentally, aiming to avoid visible slowdown.
191 static const int sync_pages = 128;
193 static struct kmem_cache *fn_alias_kmem __read_mostly;
194 static struct kmem_cache *trie_leaf_kmem __read_mostly;
196 static inline struct tnode *tn_info(struct key_vector *kv)
198 return container_of(kv, struct tnode, kv[0]);
201 /* caller must hold RTNL */
202 #define node_parent(tn) rtnl_dereference(tn_info(tn)->parent)
203 #define get_child(tn, i) rtnl_dereference((tn)->tnode[i])
205 /* caller must hold RCU read lock or RTNL */
206 #define node_parent_rcu(tn) rcu_dereference_rtnl(tn_info(tn)->parent)
207 #define get_child_rcu(tn, i) rcu_dereference_rtnl((tn)->tnode[i])
209 /* wrapper for rcu_assign_pointer */
210 static inline void node_set_parent(struct key_vector *n, struct key_vector *tp)
213 rcu_assign_pointer(tn_info(n)->parent, tp);
216 #define NODE_INIT_PARENT(n, p) RCU_INIT_POINTER(tn_info(n)->parent, p)
218 /* This provides us with the number of children in this node, in the case of a
219 * leaf this will return 0 meaning none of the children are accessible.
221 static inline unsigned long child_length(const struct key_vector *tn)
223 return (1ul << tn->bits) & ~(1ul);
226 #define get_cindex(key, kv) (((key) ^ (kv)->key) >> (kv)->pos)
228 static inline unsigned long get_index(t_key key, struct key_vector *kv)
230 unsigned long index = key ^ kv->key;
232 if ((BITS_PER_LONG <= KEYLENGTH) && (KEYLENGTH == kv->pos))
235 return index >> kv->pos;
238 /* To understand this stuff, an understanding of keys and all their bits is
239 * necessary. Every node in the trie has a key associated with it, but not
240 * all of the bits in that key are significant.
242 * Consider a node 'n' and its parent 'tp'.
244 * If n is a leaf, every bit in its key is significant. Its presence is
245 * necessitated by path compression, since during a tree traversal (when
246 * searching for a leaf - unless we are doing an insertion) we will completely
247 * ignore all skipped bits we encounter. Thus we need to verify, at the end of
248 * a potentially successful search, that we have indeed been walking the
251 * Note that we can never "miss" the correct key in the tree if present by
252 * following the wrong path. Path compression ensures that segments of the key
253 * that are the same for all keys with a given prefix are skipped, but the
254 * skipped part *is* identical for each node in the subtrie below the skipped
255 * bit! trie_insert() in this implementation takes care of that.
257 * if n is an internal node - a 'tnode' here, the various parts of its key
258 * have many different meanings.
261 * _________________________________________________________________
262 * | i | i | i | i | i | i | i | N | N | N | S | S | S | S | S | C |
263 * -----------------------------------------------------------------
264 * 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16
266 * _________________________________________________________________
267 * | C | C | C | u | u | u | u | u | u | u | u | u | u | u | u | u |
268 * -----------------------------------------------------------------
269 * 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
276 * First, let's just ignore the bits that come before the parent tp, that is
277 * the bits from (tp->pos + tp->bits) to 31. They are *known* but at this
278 * point we do not use them for anything.
280 * The bits from (tp->pos) to (tp->pos + tp->bits - 1) - "N", above - are the
281 * index into the parent's child array. That is, they will be used to find
282 * 'n' among tp's children.
284 * The bits from (n->pos + n->bits) to (tp->pos - 1) - "S" - are skipped bits
287 * All the bits we have seen so far are significant to the node n. The rest
288 * of the bits are really not needed or indeed known in n->key.
290 * The bits from (n->pos) to (n->pos + n->bits - 1) - "C" - are the index into
291 * n's child array, and will of course be different for each child.
293 * The rest of the bits, from 0 to (n->pos -1) - "u" - are completely unknown
297 static const int halve_threshold = 25;
298 static const int inflate_threshold = 50;
299 static const int halve_threshold_root = 15;
300 static const int inflate_threshold_root = 30;
302 static void __alias_free_mem(struct rcu_head *head)
304 struct fib_alias *fa = container_of(head, struct fib_alias, rcu);
305 kmem_cache_free(fn_alias_kmem, fa);
308 static inline void alias_free_mem_rcu(struct fib_alias *fa)
310 call_rcu(&fa->rcu, __alias_free_mem);
313 #define TNODE_KMALLOC_MAX \
314 ilog2((PAGE_SIZE - TNODE_SIZE(0)) / sizeof(struct key_vector *))
315 #define TNODE_VMALLOC_MAX \
316 ilog2((SIZE_MAX - TNODE_SIZE(0)) / sizeof(struct key_vector *))
318 static void __node_free_rcu(struct rcu_head *head)
320 struct tnode *n = container_of(head, struct tnode, rcu);
323 kmem_cache_free(trie_leaf_kmem, n);
328 #define node_free(n) call_rcu(&tn_info(n)->rcu, __node_free_rcu)
330 static struct tnode *tnode_alloc(int bits)
334 /* verify bits is within bounds */
335 if (bits > TNODE_VMALLOC_MAX)
338 /* determine size and verify it is non-zero and didn't overflow */
339 size = TNODE_SIZE(1ul << bits);
341 if (size <= PAGE_SIZE)
342 return kzalloc(size, GFP_KERNEL);
344 return vzalloc(size);
347 static inline void empty_child_inc(struct key_vector *n)
349 ++tn_info(n)->empty_children ? : ++tn_info(n)->full_children;
352 static inline void empty_child_dec(struct key_vector *n)
354 tn_info(n)->empty_children-- ? : tn_info(n)->full_children--;
357 static struct key_vector *leaf_new(t_key key, struct fib_alias *fa)
359 struct key_vector *l;
362 kv = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
366 /* initialize key vector */
371 l->slen = fa->fa_slen;
373 /* link leaf to fib alias */
374 INIT_HLIST_HEAD(&l->leaf);
375 hlist_add_head(&fa->fa_list, &l->leaf);
380 static struct key_vector *tnode_new(t_key key, int pos, int bits)
382 unsigned int shift = pos + bits;
383 struct key_vector *tn;
386 /* verify bits and pos their msb bits clear and values are valid */
387 BUG_ON(!bits || (shift > KEYLENGTH));
389 tnode = tnode_alloc(bits);
393 pr_debug("AT %p s=%zu %zu\n", tnode, TNODE_SIZE(0),
394 sizeof(struct key_vector *) << bits);
396 if (bits == KEYLENGTH)
397 tnode->full_children = 1;
399 tnode->empty_children = 1ul << bits;
402 tn->key = (shift < KEYLENGTH) ? (key >> shift) << shift : 0;
410 /* Check whether a tnode 'n' is "full", i.e. it is an internal node
411 * and no bits are skipped. See discussion in dyntree paper p. 6
413 static inline int tnode_full(struct key_vector *tn, struct key_vector *n)
415 return n && ((n->pos + n->bits) == tn->pos) && IS_TNODE(n);
418 /* Add a child at position i overwriting the old value.
419 * Update the value of full_children and empty_children.
421 static void put_child(struct key_vector *tn, unsigned long i,
422 struct key_vector *n)
424 struct key_vector *chi = get_child(tn, i);
427 BUG_ON(i >= child_length(tn));
429 /* update emptyChildren, overflow into fullChildren */
435 /* update fullChildren */
436 wasfull = tnode_full(tn, chi);
437 isfull = tnode_full(tn, n);
439 if (wasfull && !isfull)
440 tn_info(tn)->full_children--;
441 else if (!wasfull && isfull)
442 tn_info(tn)->full_children++;
444 if (n && (tn->slen < n->slen))
447 rcu_assign_pointer(tn->tnode[i], n);
450 static void update_children(struct key_vector *tn)
454 /* update all of the child parent pointers */
455 for (i = child_length(tn); i;) {
456 struct key_vector *inode = get_child(tn, --i);
461 /* Either update the children of a tnode that
462 * already belongs to us or update the child
463 * to point to ourselves.
465 if (node_parent(inode) == tn)
466 update_children(inode);
468 node_set_parent(inode, tn);
472 static inline void put_child_root(struct key_vector *tp, t_key key,
473 struct key_vector *n)
476 rcu_assign_pointer(tp->tnode[0], n);
478 put_child(tp, get_index(key, tp), n);
481 static inline void tnode_free_init(struct key_vector *tn)
483 tn_info(tn)->rcu.next = NULL;
486 static inline void tnode_free_append(struct key_vector *tn,
487 struct key_vector *n)
489 tn_info(n)->rcu.next = tn_info(tn)->rcu.next;
490 tn_info(tn)->rcu.next = &tn_info(n)->rcu;
493 static void tnode_free(struct key_vector *tn)
495 struct callback_head *head = &tn_info(tn)->rcu;
499 tnode_free_size += TNODE_SIZE(1ul << tn->bits);
502 tn = container_of(head, struct tnode, rcu)->kv;
505 if (tnode_free_size >= PAGE_SIZE * sync_pages) {
511 static struct key_vector *replace(struct trie *t,
512 struct key_vector *oldtnode,
513 struct key_vector *tn)
515 struct key_vector *tp = node_parent(oldtnode);
518 /* setup the parent pointer out of and back into this node */
519 NODE_INIT_PARENT(tn, tp);
520 put_child_root(tp, tn->key, tn);
522 /* update all of the child parent pointers */
525 /* all pointers should be clean so we are done */
526 tnode_free(oldtnode);
528 /* resize children now that oldtnode is freed */
529 for (i = child_length(tn); i;) {
530 struct key_vector *inode = get_child(tn, --i);
532 /* resize child node */
533 if (tnode_full(tn, inode))
534 tn = resize(t, inode);
540 static struct key_vector *inflate(struct trie *t,
541 struct key_vector *oldtnode)
543 struct key_vector *tn;
547 pr_debug("In inflate\n");
549 tn = tnode_new(oldtnode->key, oldtnode->pos - 1, oldtnode->bits + 1);
553 /* prepare oldtnode to be freed */
554 tnode_free_init(oldtnode);
556 /* Assemble all of the pointers in our cluster, in this case that
557 * represents all of the pointers out of our allocated nodes that
558 * point to existing tnodes and the links between our allocated
561 for (i = child_length(oldtnode), m = 1u << tn->pos; i;) {
562 struct key_vector *inode = get_child(oldtnode, --i);
563 struct key_vector *node0, *node1;
570 /* A leaf or an internal node with skipped bits */
571 if (!tnode_full(oldtnode, inode)) {
572 put_child(tn, get_index(inode->key, tn), inode);
576 /* drop the node in the old tnode free list */
577 tnode_free_append(oldtnode, inode);
579 /* An internal node with two children */
580 if (inode->bits == 1) {
581 put_child(tn, 2 * i + 1, get_child(inode, 1));
582 put_child(tn, 2 * i, get_child(inode, 0));
586 /* We will replace this node 'inode' with two new
587 * ones, 'node0' and 'node1', each with half of the
588 * original children. The two new nodes will have
589 * a position one bit further down the key and this
590 * means that the "significant" part of their keys
591 * (see the discussion near the top of this file)
592 * will differ by one bit, which will be "0" in
593 * node0's key and "1" in node1's key. Since we are
594 * moving the key position by one step, the bit that
595 * we are moving away from - the bit at position
596 * (tn->pos) - is the one that will differ between
597 * node0 and node1. So... we synthesize that bit in the
600 node1 = tnode_new(inode->key | m, inode->pos, inode->bits - 1);
603 node0 = tnode_new(inode->key, inode->pos, inode->bits - 1);
605 tnode_free_append(tn, node1);
608 tnode_free_append(tn, node0);
610 /* populate child pointers in new nodes */
611 for (k = child_length(inode), j = k / 2; j;) {
612 put_child(node1, --j, get_child(inode, --k));
613 put_child(node0, j, get_child(inode, j));
614 put_child(node1, --j, get_child(inode, --k));
615 put_child(node0, j, get_child(inode, j));
618 /* link new nodes to parent */
619 NODE_INIT_PARENT(node1, tn);
620 NODE_INIT_PARENT(node0, tn);
622 /* link parent to nodes */
623 put_child(tn, 2 * i + 1, node1);
624 put_child(tn, 2 * i, node0);
627 /* setup the parent pointers into and out of this node */
628 return replace(t, oldtnode, tn);
630 /* all pointers should be clean so we are done */
636 static struct key_vector *halve(struct trie *t,
637 struct key_vector *oldtnode)
639 struct key_vector *tn;
642 pr_debug("In halve\n");
644 tn = tnode_new(oldtnode->key, oldtnode->pos + 1, oldtnode->bits - 1);
648 /* prepare oldtnode to be freed */
649 tnode_free_init(oldtnode);
651 /* Assemble all of the pointers in our cluster, in this case that
652 * represents all of the pointers out of our allocated nodes that
653 * point to existing tnodes and the links between our allocated
656 for (i = child_length(oldtnode); i;) {
657 struct key_vector *node1 = get_child(oldtnode, --i);
658 struct key_vector *node0 = get_child(oldtnode, --i);
659 struct key_vector *inode;
661 /* At least one of the children is empty */
662 if (!node1 || !node0) {
663 put_child(tn, i / 2, node1 ? : node0);
667 /* Two nonempty children */
668 inode = tnode_new(node0->key, oldtnode->pos, 1);
671 tnode_free_append(tn, inode);
673 /* initialize pointers out of node */
674 put_child(inode, 1, node1);
675 put_child(inode, 0, node0);
676 NODE_INIT_PARENT(inode, tn);
678 /* link parent to node */
679 put_child(tn, i / 2, inode);
682 /* setup the parent pointers into and out of this node */
683 return replace(t, oldtnode, tn);
685 /* all pointers should be clean so we are done */
691 static struct key_vector *collapse(struct trie *t,
692 struct key_vector *oldtnode)
694 struct key_vector *n, *tp;
697 /* scan the tnode looking for that one child that might still exist */
698 for (n = NULL, i = child_length(oldtnode); !n && i;)
699 n = get_child(oldtnode, --i);
701 /* compress one level */
702 tp = node_parent(oldtnode);
703 put_child_root(tp, oldtnode->key, n);
704 node_set_parent(n, tp);
712 static unsigned char update_suffix(struct key_vector *tn)
714 unsigned char slen = tn->pos;
715 unsigned long stride, i;
716 unsigned char slen_max;
718 /* only vector 0 can have a suffix length greater than or equal to
719 * tn->pos + tn->bits, the second highest node will have a suffix
720 * length at most of tn->pos + tn->bits - 1
722 slen_max = min_t(unsigned char, tn->pos + tn->bits - 1, tn->slen);
724 /* search though the list of children looking for nodes that might
725 * have a suffix greater than the one we currently have. This is
726 * why we start with a stride of 2 since a stride of 1 would
727 * represent the nodes with suffix length equal to tn->pos
729 for (i = 0, stride = 0x2ul ; i < child_length(tn); i += stride) {
730 struct key_vector *n = get_child(tn, i);
732 if (!n || (n->slen <= slen))
735 /* update stride and slen based on new value */
736 stride <<= (n->slen - slen);
740 /* stop searching if we have hit the maximum possible value */
741 if (slen >= slen_max)
750 /* From "Implementing a dynamic compressed trie" by Stefan Nilsson of
751 * the Helsinki University of Technology and Matti Tikkanen of Nokia
752 * Telecommunications, page 6:
753 * "A node is doubled if the ratio of non-empty children to all
754 * children in the *doubled* node is at least 'high'."
756 * 'high' in this instance is the variable 'inflate_threshold'. It
757 * is expressed as a percentage, so we multiply it with
758 * child_length() and instead of multiplying by 2 (since the
759 * child array will be doubled by inflate()) and multiplying
760 * the left-hand side by 100 (to handle the percentage thing) we
761 * multiply the left-hand side by 50.
763 * The left-hand side may look a bit weird: child_length(tn)
764 * - tn->empty_children is of course the number of non-null children
765 * in the current node. tn->full_children is the number of "full"
766 * children, that is non-null tnodes with a skip value of 0.
767 * All of those will be doubled in the resulting inflated tnode, so
768 * we just count them one extra time here.
770 * A clearer way to write this would be:
772 * to_be_doubled = tn->full_children;
773 * not_to_be_doubled = child_length(tn) - tn->empty_children -
776 * new_child_length = child_length(tn) * 2;
778 * new_fill_factor = 100 * (not_to_be_doubled + 2*to_be_doubled) /
780 * if (new_fill_factor >= inflate_threshold)
782 * ...and so on, tho it would mess up the while () loop.
785 * 100 * (not_to_be_doubled + 2*to_be_doubled) / new_child_length >=
789 * 100 * (not_to_be_doubled + 2*to_be_doubled) >=
790 * inflate_threshold * new_child_length
792 * expand not_to_be_doubled and to_be_doubled, and shorten:
793 * 100 * (child_length(tn) - tn->empty_children +
794 * tn->full_children) >= inflate_threshold * new_child_length
796 * expand new_child_length:
797 * 100 * (child_length(tn) - tn->empty_children +
798 * tn->full_children) >=
799 * inflate_threshold * child_length(tn) * 2
802 * 50 * (tn->full_children + child_length(tn) -
803 * tn->empty_children) >= inflate_threshold *
807 static inline bool should_inflate(struct key_vector *tp, struct key_vector *tn)
809 unsigned long used = child_length(tn);
810 unsigned long threshold = used;
812 /* Keep root node larger */
813 threshold *= IS_TRIE(tp) ? inflate_threshold_root : inflate_threshold;
814 used -= tn_info(tn)->empty_children;
815 used += tn_info(tn)->full_children;
817 /* if bits == KEYLENGTH then pos = 0, and will fail below */
819 return (used > 1) && tn->pos && ((50 * used) >= threshold);
822 static inline bool should_halve(struct key_vector *tp, struct key_vector *tn)
824 unsigned long used = child_length(tn);
825 unsigned long threshold = used;
827 /* Keep root node larger */
828 threshold *= IS_TRIE(tp) ? halve_threshold_root : halve_threshold;
829 used -= tn_info(tn)->empty_children;
831 /* if bits == KEYLENGTH then used = 100% on wrap, and will fail below */
833 return (used > 1) && (tn->bits > 1) && ((100 * used) < threshold);
836 static inline bool should_collapse(struct key_vector *tn)
838 unsigned long used = child_length(tn);
840 used -= tn_info(tn)->empty_children;
842 /* account for bits == KEYLENGTH case */
843 if ((tn->bits == KEYLENGTH) && tn_info(tn)->full_children)
846 /* One child or none, time to drop us from the trie */
851 static struct key_vector *resize(struct trie *t, struct key_vector *tn)
853 #ifdef CONFIG_IP_FIB_TRIE_STATS
854 struct trie_use_stats __percpu *stats = t->stats;
856 struct key_vector *tp = node_parent(tn);
857 unsigned long cindex = get_index(tn->key, tp);
858 int max_work = MAX_WORK;
860 pr_debug("In tnode_resize %p inflate_threshold=%d threshold=%d\n",
861 tn, inflate_threshold, halve_threshold);
863 /* track the tnode via the pointer from the parent instead of
864 * doing it ourselves. This way we can let RCU fully do its
865 * thing without us interfering
867 BUG_ON(tn != get_child(tp, cindex));
869 /* Double as long as the resulting node has a number of
870 * nonempty nodes that are above the threshold.
872 while (should_inflate(tp, tn) && max_work) {
875 #ifdef CONFIG_IP_FIB_TRIE_STATS
876 this_cpu_inc(stats->resize_node_skipped);
882 tn = get_child(tp, cindex);
885 /* update parent in case inflate failed */
886 tp = node_parent(tn);
888 /* Return if at least one inflate is run */
889 if (max_work != MAX_WORK)
892 /* Halve as long as the number of empty children in this
893 * node is above threshold.
895 while (should_halve(tp, tn) && max_work) {
898 #ifdef CONFIG_IP_FIB_TRIE_STATS
899 this_cpu_inc(stats->resize_node_skipped);
905 tn = get_child(tp, cindex);
908 /* Only one child remains */
909 if (should_collapse(tn))
910 return collapse(t, tn);
912 /* update parent in case halve failed */
913 return node_parent(tn);
916 static void node_pull_suffix(struct key_vector *tn, unsigned char slen)
918 unsigned char node_slen = tn->slen;
920 while ((node_slen > tn->pos) && (node_slen > slen)) {
921 slen = update_suffix(tn);
922 if (node_slen == slen)
925 tn = node_parent(tn);
926 node_slen = tn->slen;
930 static void node_push_suffix(struct key_vector *tn, unsigned char slen)
932 while (tn->slen < slen) {
934 tn = node_parent(tn);
938 /* rcu_read_lock needs to be hold by caller from readside */
939 static struct key_vector *fib_find_node(struct trie *t,
940 struct key_vector **tp, u32 key)
942 struct key_vector *pn, *n = t->kv;
943 unsigned long index = 0;
947 n = get_child_rcu(n, index);
952 index = get_cindex(key, n);
954 /* This bit of code is a bit tricky but it combines multiple
955 * checks into a single check. The prefix consists of the
956 * prefix plus zeros for the bits in the cindex. The index
957 * is the difference between the key and this value. From
958 * this we can actually derive several pieces of data.
959 * if (index >= (1ul << bits))
960 * we have a mismatch in skip bits and failed
962 * we know the value is cindex
964 * This check is safe even if bits == KEYLENGTH due to the
965 * fact that we can only allocate a node with 32 bits if a
966 * long is greater than 32 bits.
968 if (index >= (1ul << n->bits)) {
973 /* keep searching until we find a perfect match leaf or NULL */
974 } while (IS_TNODE(n));
981 /* Return the first fib alias matching TOS with
982 * priority less than or equal to PRIO.
984 static struct fib_alias *fib_find_alias(struct hlist_head *fah, u8 slen,
985 u8 tos, u32 prio, u32 tb_id)
987 struct fib_alias *fa;
992 hlist_for_each_entry(fa, fah, fa_list) {
993 if (fa->fa_slen < slen)
995 if (fa->fa_slen != slen)
997 if (fa->tb_id > tb_id)
999 if (fa->tb_id != tb_id)
1001 if (fa->fa_tos > tos)
1003 if (fa->fa_info->fib_priority >= prio || fa->fa_tos < tos)
1010 static void trie_rebalance(struct trie *t, struct key_vector *tn)
1012 while (!IS_TRIE(tn))
1016 static int fib_insert_node(struct trie *t, struct key_vector *tp,
1017 struct fib_alias *new, t_key key)
1019 struct key_vector *n, *l;
1021 l = leaf_new(key, new);
1025 /* retrieve child from parent node */
1026 n = get_child(tp, get_index(key, tp));
1028 /* Case 2: n is a LEAF or a TNODE and the key doesn't match.
1030 * Add a new tnode here
1031 * first tnode need some special handling
1032 * leaves us in position for handling as case 3
1035 struct key_vector *tn;
1037 tn = tnode_new(key, __fls(key ^ n->key), 1);
1041 /* initialize routes out of node */
1042 NODE_INIT_PARENT(tn, tp);
1043 put_child(tn, get_index(key, tn) ^ 1, n);
1045 /* start adding routes into the node */
1046 put_child_root(tp, key, tn);
1047 node_set_parent(n, tn);
1049 /* parent now has a NULL spot where the leaf can go */
1053 /* Case 3: n is NULL, and will just insert a new leaf */
1054 node_push_suffix(tp, new->fa_slen);
1055 NODE_INIT_PARENT(l, tp);
1056 put_child_root(tp, key, l);
1057 trie_rebalance(t, tp);
1066 static int fib_insert_alias(struct trie *t, struct key_vector *tp,
1067 struct key_vector *l, struct fib_alias *new,
1068 struct fib_alias *fa, t_key key)
1071 return fib_insert_node(t, tp, new, key);
1074 hlist_add_before_rcu(&new->fa_list, &fa->fa_list);
1076 struct fib_alias *last;
1078 hlist_for_each_entry(last, &l->leaf, fa_list) {
1079 if (new->fa_slen < last->fa_slen)
1081 if ((new->fa_slen == last->fa_slen) &&
1082 (new->tb_id > last->tb_id))
1088 hlist_add_behind_rcu(&new->fa_list, &fa->fa_list);
1090 hlist_add_head_rcu(&new->fa_list, &l->leaf);
1093 /* if we added to the tail node then we need to update slen */
1094 if (l->slen < new->fa_slen) {
1095 l->slen = new->fa_slen;
1096 node_push_suffix(tp, new->fa_slen);
1102 static bool fib_valid_key_len(u32 key, u8 plen, struct netlink_ext_ack *extack)
1104 if (plen > KEYLENGTH) {
1105 NL_SET_ERR_MSG(extack, "Invalid prefix length");
1109 if ((plen < KEYLENGTH) && (key << plen)) {
1110 NL_SET_ERR_MSG(extack,
1111 "Invalid prefix for given prefix length");
1118 /* Caller must hold RTNL. */
1119 int fib_table_insert(struct net *net, struct fib_table *tb,
1120 struct fib_config *cfg, struct netlink_ext_ack *extack)
1122 enum fib_event_type event = FIB_EVENT_ENTRY_ADD;
1123 struct trie *t = (struct trie *)tb->tb_data;
1124 struct fib_alias *fa, *new_fa;
1125 struct key_vector *l, *tp;
1126 u16 nlflags = NLM_F_EXCL;
1127 struct fib_info *fi;
1128 u8 plen = cfg->fc_dst_len;
1129 u8 slen = KEYLENGTH - plen;
1130 u8 tos = cfg->fc_tos;
1134 key = ntohl(cfg->fc_dst);
1136 if (!fib_valid_key_len(key, plen, extack))
1139 pr_debug("Insert table=%u %08x/%d\n", tb->tb_id, key, plen);
1141 fi = fib_create_info(cfg, extack);
1147 l = fib_find_node(t, &tp, key);
1148 fa = l ? fib_find_alias(&l->leaf, slen, tos, fi->fib_priority,
1151 /* Now fa, if non-NULL, points to the first fib alias
1152 * with the same keys [prefix,tos,priority], if such key already
1153 * exists or to the node before which we will insert new one.
1155 * If fa is NULL, we will need to allocate a new one and
1156 * insert to the tail of the section matching the suffix length
1160 if (fa && fa->fa_tos == tos &&
1161 fa->fa_info->fib_priority == fi->fib_priority) {
1162 struct fib_alias *fa_first, *fa_match;
1165 if (cfg->fc_nlflags & NLM_F_EXCL)
1168 nlflags &= ~NLM_F_EXCL;
1171 * 1. Find exact match for type, scope, fib_info to avoid
1173 * 2. Find next 'fa' (or head), NLM_F_APPEND inserts before it
1177 hlist_for_each_entry_from(fa, fa_list) {
1178 if ((fa->fa_slen != slen) ||
1179 (fa->tb_id != tb->tb_id) ||
1180 (fa->fa_tos != tos))
1182 if (fa->fa_info->fib_priority != fi->fib_priority)
1184 if (fa->fa_type == cfg->fc_type &&
1185 fa->fa_info == fi) {
1191 if (cfg->fc_nlflags & NLM_F_REPLACE) {
1192 struct fib_info *fi_drop;
1195 nlflags |= NLM_F_REPLACE;
1203 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1207 fi_drop = fa->fa_info;
1208 new_fa->fa_tos = fa->fa_tos;
1209 new_fa->fa_info = fi;
1210 new_fa->fa_type = cfg->fc_type;
1211 state = fa->fa_state;
1212 new_fa->fa_state = state & ~FA_S_ACCESSED;
1213 new_fa->fa_slen = fa->fa_slen;
1214 new_fa->tb_id = tb->tb_id;
1215 new_fa->fa_default = -1;
1217 call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_REPLACE,
1219 new_fa->fa_tos, cfg->fc_type,
1221 rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen,
1222 tb->tb_id, &cfg->fc_nlinfo, nlflags);
1224 hlist_replace_rcu(&fa->fa_list, &new_fa->fa_list);
1226 alias_free_mem_rcu(fa);
1228 fib_release_info(fi_drop);
1229 if (state & FA_S_ACCESSED)
1230 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1234 /* Error if we find a perfect match which
1235 * uses the same scope, type, and nexthop
1241 if (cfg->fc_nlflags & NLM_F_APPEND) {
1242 event = FIB_EVENT_ENTRY_APPEND;
1243 nlflags |= NLM_F_APPEND;
1249 if (!(cfg->fc_nlflags & NLM_F_CREATE))
1252 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;
1262 new_fa->fa_slen = slen;
1263 new_fa->tb_id = tb->tb_id;
1264 new_fa->fa_default = -1;
1266 /* Insert new entry to the list. */
1267 err = fib_insert_alias(t, tp, l, new_fa, fa, key);
1269 goto out_free_new_fa;
1272 tb->tb_num_default++;
1274 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1275 call_fib_entry_notifiers(net, event, key, plen, fi, tos, cfg->fc_type,
1277 rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen, new_fa->tb_id,
1278 &cfg->fc_nlinfo, nlflags);
1283 kmem_cache_free(fn_alias_kmem, new_fa);
1285 fib_release_info(fi);
1290 static inline t_key prefix_mismatch(t_key key, struct key_vector *n)
1292 t_key prefix = n->key;
1294 return (key ^ prefix) & (prefix | -prefix);
1297 /* should be called with rcu_read_lock */
1298 int fib_table_lookup(struct fib_table *tb, const struct flowi4 *flp,
1299 struct fib_result *res, int fib_flags)
1301 struct trie *t = (struct trie *) tb->tb_data;
1302 #ifdef CONFIG_IP_FIB_TRIE_STATS
1303 struct trie_use_stats __percpu *stats = t->stats;
1305 const t_key key = ntohl(flp->daddr);
1306 struct key_vector *n, *pn;
1307 struct fib_alias *fa;
1308 unsigned long index;
1311 trace_fib_table_lookup(tb->tb_id, flp);
1316 n = get_child_rcu(pn, cindex);
1320 #ifdef CONFIG_IP_FIB_TRIE_STATS
1321 this_cpu_inc(stats->gets);
1324 /* Step 1: Travel to the longest prefix match in the trie */
1326 index = get_cindex(key, n);
1328 /* This bit of code is a bit tricky but it combines multiple
1329 * checks into a single check. The prefix consists of the
1330 * prefix plus zeros for the "bits" in the prefix. The index
1331 * is the difference between the key and this value. From
1332 * this we can actually derive several pieces of data.
1333 * if (index >= (1ul << bits))
1334 * we have a mismatch in skip bits and failed
1336 * we know the value is cindex
1338 * This check is safe even if bits == KEYLENGTH due to the
1339 * fact that we can only allocate a node with 32 bits if a
1340 * long is greater than 32 bits.
1342 if (index >= (1ul << n->bits))
1345 /* we have found a leaf. Prefixes have already been compared */
1349 /* only record pn and cindex if we are going to be chopping
1350 * bits later. Otherwise we are just wasting cycles.
1352 if (n->slen > n->pos) {
1357 n = get_child_rcu(n, index);
1362 /* Step 2: Sort out leaves and begin backtracing for longest prefix */
1364 /* record the pointer where our next node pointer is stored */
1365 struct key_vector __rcu **cptr = n->tnode;
1367 /* This test verifies that none of the bits that differ
1368 * between the key and the prefix exist in the region of
1369 * the lsb and higher in the prefix.
1371 if (unlikely(prefix_mismatch(key, n)) || (n->slen == n->pos))
1374 /* exit out and process leaf */
1375 if (unlikely(IS_LEAF(n)))
1378 /* Don't bother recording parent info. Since we are in
1379 * prefix match mode we will have to come back to wherever
1380 * we started this traversal anyway
1383 while ((n = rcu_dereference(*cptr)) == NULL) {
1385 #ifdef CONFIG_IP_FIB_TRIE_STATS
1387 this_cpu_inc(stats->null_node_hit);
1389 /* If we are at cindex 0 there are no more bits for
1390 * us to strip at this level so we must ascend back
1391 * up one level to see if there are any more bits to
1392 * be stripped there.
1395 t_key pkey = pn->key;
1397 /* If we don't have a parent then there is
1398 * nothing for us to do as we do not have any
1399 * further nodes to parse.
1403 #ifdef CONFIG_IP_FIB_TRIE_STATS
1404 this_cpu_inc(stats->backtrack);
1406 /* Get Child's index */
1407 pn = node_parent_rcu(pn);
1408 cindex = get_index(pkey, pn);
1411 /* strip the least significant bit from the cindex */
1412 cindex &= cindex - 1;
1414 /* grab pointer for next child node */
1415 cptr = &pn->tnode[cindex];
1420 /* this line carries forward the xor from earlier in the function */
1421 index = key ^ n->key;
1423 /* Step 3: Process the leaf, if that fails fall back to backtracing */
1424 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list) {
1425 struct fib_info *fi = fa->fa_info;
1428 if ((BITS_PER_LONG > KEYLENGTH) || (fa->fa_slen < KEYLENGTH)) {
1429 if (index >= (1ul << fa->fa_slen))
1432 if (fa->fa_tos && fa->fa_tos != flp->flowi4_tos)
1436 if (fa->fa_info->fib_scope < flp->flowi4_scope)
1438 fib_alias_accessed(fa);
1439 err = fib_props[fa->fa_type].error;
1440 if (unlikely(err < 0)) {
1441 #ifdef CONFIG_IP_FIB_TRIE_STATS
1442 this_cpu_inc(stats->semantic_match_passed);
1446 if (fi->fib_flags & RTNH_F_DEAD)
1448 for (nhsel = 0; nhsel < fi->fib_nhs; nhsel++) {
1449 const struct fib_nh *nh = &fi->fib_nh[nhsel];
1450 struct in_device *in_dev = __in_dev_get_rcu(nh->nh_dev);
1452 if (nh->nh_flags & RTNH_F_DEAD)
1455 IN_DEV_IGNORE_ROUTES_WITH_LINKDOWN(in_dev) &&
1456 nh->nh_flags & RTNH_F_LINKDOWN &&
1457 !(fib_flags & FIB_LOOKUP_IGNORE_LINKSTATE))
1459 if (!(flp->flowi4_flags & FLOWI_FLAG_SKIP_NH_OIF)) {
1460 if (flp->flowi4_oif &&
1461 flp->flowi4_oif != nh->nh_oif)
1465 if (!(fib_flags & FIB_LOOKUP_NOREF))
1466 refcount_inc(&fi->fib_clntref);
1468 res->prefix = htonl(n->key);
1469 res->prefixlen = KEYLENGTH - fa->fa_slen;
1470 res->nh_sel = nhsel;
1471 res->type = fa->fa_type;
1472 res->scope = fi->fib_scope;
1475 res->fa_head = &n->leaf;
1476 #ifdef CONFIG_IP_FIB_TRIE_STATS
1477 this_cpu_inc(stats->semantic_match_passed);
1479 trace_fib_table_lookup_nh(nh);
1484 #ifdef CONFIG_IP_FIB_TRIE_STATS
1485 this_cpu_inc(stats->semantic_match_miss);
1489 EXPORT_SYMBOL_GPL(fib_table_lookup);
1491 static void fib_remove_alias(struct trie *t, struct key_vector *tp,
1492 struct key_vector *l, struct fib_alias *old)
1494 /* record the location of the previous list_info entry */
1495 struct hlist_node **pprev = old->fa_list.pprev;
1496 struct fib_alias *fa = hlist_entry(pprev, typeof(*fa), fa_list.next);
1498 /* remove the fib_alias from the list */
1499 hlist_del_rcu(&old->fa_list);
1501 /* if we emptied the list this leaf will be freed and we can sort
1502 * out parent suffix lengths as a part of trie_rebalance
1504 if (hlist_empty(&l->leaf)) {
1505 if (tp->slen == l->slen)
1506 node_pull_suffix(tp, tp->pos);
1507 put_child_root(tp, l->key, NULL);
1509 trie_rebalance(t, tp);
1513 /* only access fa if it is pointing at the last valid hlist_node */
1517 /* update the trie with the latest suffix length */
1518 l->slen = fa->fa_slen;
1519 node_pull_suffix(tp, fa->fa_slen);
1522 /* Caller must hold RTNL. */
1523 int fib_table_delete(struct net *net, struct fib_table *tb,
1524 struct fib_config *cfg, struct netlink_ext_ack *extack)
1526 struct trie *t = (struct trie *) tb->tb_data;
1527 struct fib_alias *fa, *fa_to_delete;
1528 struct key_vector *l, *tp;
1529 u8 plen = cfg->fc_dst_len;
1530 u8 slen = KEYLENGTH - plen;
1531 u8 tos = cfg->fc_tos;
1534 key = ntohl(cfg->fc_dst);
1536 if (!fib_valid_key_len(key, plen, extack))
1539 l = fib_find_node(t, &tp, key);
1543 fa = fib_find_alias(&l->leaf, slen, tos, 0, tb->tb_id);
1547 pr_debug("Deleting %08x/%d tos=%d t=%p\n", key, plen, tos, t);
1549 fa_to_delete = NULL;
1550 hlist_for_each_entry_from(fa, fa_list) {
1551 struct fib_info *fi = fa->fa_info;
1553 if ((fa->fa_slen != slen) ||
1554 (fa->tb_id != tb->tb_id) ||
1555 (fa->fa_tos != tos))
1558 if ((!cfg->fc_type || fa->fa_type == cfg->fc_type) &&
1559 (cfg->fc_scope == RT_SCOPE_NOWHERE ||
1560 fa->fa_info->fib_scope == cfg->fc_scope) &&
1561 (!cfg->fc_prefsrc ||
1562 fi->fib_prefsrc == cfg->fc_prefsrc) &&
1563 (!cfg->fc_protocol ||
1564 fi->fib_protocol == cfg->fc_protocol) &&
1565 fib_nh_match(cfg, fi, extack) == 0) {
1574 call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, key, plen,
1575 fa_to_delete->fa_info, tos,
1576 fa_to_delete->fa_type, tb->tb_id);
1577 rtmsg_fib(RTM_DELROUTE, htonl(key), fa_to_delete, plen, tb->tb_id,
1578 &cfg->fc_nlinfo, 0);
1581 tb->tb_num_default--;
1583 fib_remove_alias(t, tp, l, fa_to_delete);
1585 if (fa_to_delete->fa_state & FA_S_ACCESSED)
1586 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1588 fib_release_info(fa_to_delete->fa_info);
1589 alias_free_mem_rcu(fa_to_delete);
1593 /* Scan for the next leaf starting at the provided key value */
1594 static struct key_vector *leaf_walk_rcu(struct key_vector **tn, t_key key)
1596 struct key_vector *pn, *n = *tn;
1597 unsigned long cindex;
1599 /* this loop is meant to try and find the key in the trie */
1601 /* record parent and next child index */
1603 cindex = (key > pn->key) ? get_index(key, pn) : 0;
1605 if (cindex >> pn->bits)
1608 /* descend into the next child */
1609 n = get_child_rcu(pn, cindex++);
1613 /* guarantee forward progress on the keys */
1614 if (IS_LEAF(n) && (n->key >= key))
1616 } while (IS_TNODE(n));
1618 /* this loop will search for the next leaf with a greater key */
1619 while (!IS_TRIE(pn)) {
1620 /* if we exhausted the parent node we will need to climb */
1621 if (cindex >= (1ul << pn->bits)) {
1622 t_key pkey = pn->key;
1624 pn = node_parent_rcu(pn);
1625 cindex = get_index(pkey, pn) + 1;
1629 /* grab the next available node */
1630 n = get_child_rcu(pn, cindex++);
1634 /* no need to compare keys since we bumped the index */
1638 /* Rescan start scanning in new node */
1644 return NULL; /* Root of trie */
1646 /* if we are at the limit for keys just return NULL for the tnode */
1651 static void fib_trie_free(struct fib_table *tb)
1653 struct trie *t = (struct trie *)tb->tb_data;
1654 struct key_vector *pn = t->kv;
1655 unsigned long cindex = 1;
1656 struct hlist_node *tmp;
1657 struct fib_alias *fa;
1659 /* walk trie in reverse order and free everything */
1661 struct key_vector *n;
1664 t_key pkey = pn->key;
1670 pn = node_parent(pn);
1672 /* drop emptied tnode */
1673 put_child_root(pn, n->key, NULL);
1676 cindex = get_index(pkey, pn);
1681 /* grab the next available node */
1682 n = get_child(pn, cindex);
1687 /* record pn and cindex for leaf walking */
1689 cindex = 1ul << n->bits;
1694 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1695 hlist_del_rcu(&fa->fa_list);
1696 alias_free_mem_rcu(fa);
1699 put_child_root(pn, n->key, NULL);
1703 #ifdef CONFIG_IP_FIB_TRIE_STATS
1704 free_percpu(t->stats);
1709 struct fib_table *fib_trie_unmerge(struct fib_table *oldtb)
1711 struct trie *ot = (struct trie *)oldtb->tb_data;
1712 struct key_vector *l, *tp = ot->kv;
1713 struct fib_table *local_tb;
1714 struct fib_alias *fa;
1718 if (oldtb->tb_data == oldtb->__data)
1721 local_tb = fib_trie_table(RT_TABLE_LOCAL, NULL);
1725 lt = (struct trie *)local_tb->tb_data;
1727 while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
1728 struct key_vector *local_l = NULL, *local_tp;
1730 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
1731 struct fib_alias *new_fa;
1733 if (local_tb->tb_id != fa->tb_id)
1736 /* clone fa for new local table */
1737 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1741 memcpy(new_fa, fa, sizeof(*fa));
1743 /* insert clone into table */
1745 local_l = fib_find_node(lt, &local_tp, l->key);
1747 if (fib_insert_alias(lt, local_tp, local_l, new_fa,
1749 kmem_cache_free(fn_alias_kmem, new_fa);
1754 /* stop loop if key wrapped back to 0 */
1762 fib_trie_free(local_tb);
1767 /* Caller must hold RTNL */
1768 void fib_table_flush_external(struct fib_table *tb)
1770 struct trie *t = (struct trie *)tb->tb_data;
1771 struct key_vector *pn = t->kv;
1772 unsigned long cindex = 1;
1773 struct hlist_node *tmp;
1774 struct fib_alias *fa;
1776 /* walk trie in reverse order */
1778 unsigned char slen = 0;
1779 struct key_vector *n;
1782 t_key pkey = pn->key;
1784 /* cannot resize the trie vector */
1788 /* update the suffix to address pulled leaves */
1789 if (pn->slen > pn->pos)
1792 /* resize completed node */
1794 cindex = get_index(pkey, pn);
1799 /* grab the next available node */
1800 n = get_child(pn, cindex);
1805 /* record pn and cindex for leaf walking */
1807 cindex = 1ul << n->bits;
1812 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1813 /* if alias was cloned to local then we just
1814 * need to remove the local copy from main
1816 if (tb->tb_id != fa->tb_id) {
1817 hlist_del_rcu(&fa->fa_list);
1818 alias_free_mem_rcu(fa);
1822 /* record local slen */
1826 /* update leaf slen */
1829 if (hlist_empty(&n->leaf)) {
1830 put_child_root(pn, n->key, NULL);
1836 /* Caller must hold RTNL. */
1837 int fib_table_flush(struct net *net, struct fib_table *tb)
1839 struct trie *t = (struct trie *)tb->tb_data;
1840 struct key_vector *pn = t->kv;
1841 unsigned long cindex = 1;
1842 struct hlist_node *tmp;
1843 struct fib_alias *fa;
1846 /* walk trie in reverse order */
1848 unsigned char slen = 0;
1849 struct key_vector *n;
1852 t_key pkey = pn->key;
1854 /* cannot resize the trie vector */
1858 /* update the suffix to address pulled leaves */
1859 if (pn->slen > pn->pos)
1862 /* resize completed node */
1864 cindex = get_index(pkey, pn);
1869 /* grab the next available node */
1870 n = get_child(pn, cindex);
1875 /* record pn and cindex for leaf walking */
1877 cindex = 1ul << n->bits;
1882 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1883 struct fib_info *fi = fa->fa_info;
1885 if (!fi || !(fi->fib_flags & RTNH_F_DEAD) ||
1886 tb->tb_id != fa->tb_id) {
1891 call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_DEL,
1893 KEYLENGTH - fa->fa_slen,
1894 fi, fa->fa_tos, fa->fa_type,
1896 hlist_del_rcu(&fa->fa_list);
1897 fib_release_info(fa->fa_info);
1898 alias_free_mem_rcu(fa);
1902 /* update leaf slen */
1905 if (hlist_empty(&n->leaf)) {
1906 put_child_root(pn, n->key, NULL);
1911 pr_debug("trie_flush found=%d\n", found);
1915 static void fib_leaf_notify(struct net *net, struct key_vector *l,
1916 struct fib_table *tb, struct notifier_block *nb)
1918 struct fib_alias *fa;
1920 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
1921 struct fib_info *fi = fa->fa_info;
1926 /* local and main table can share the same trie,
1927 * so don't notify twice for the same entry.
1929 if (tb->tb_id != fa->tb_id)
1932 call_fib_entry_notifier(nb, net, FIB_EVENT_ENTRY_ADD, l->key,
1933 KEYLENGTH - fa->fa_slen, fi, fa->fa_tos,
1934 fa->fa_type, fa->tb_id);
1938 static void fib_table_notify(struct net *net, struct fib_table *tb,
1939 struct notifier_block *nb)
1941 struct trie *t = (struct trie *)tb->tb_data;
1942 struct key_vector *l, *tp = t->kv;
1945 while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
1946 fib_leaf_notify(net, l, tb, nb);
1949 /* stop in case of wrap around */
1955 void fib_notify(struct net *net, struct notifier_block *nb)
1959 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
1960 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
1961 struct fib_table *tb;
1963 hlist_for_each_entry_rcu(tb, head, tb_hlist)
1964 fib_table_notify(net, tb, nb);
1968 static void __trie_free_rcu(struct rcu_head *head)
1970 struct fib_table *tb = container_of(head, struct fib_table, rcu);
1971 #ifdef CONFIG_IP_FIB_TRIE_STATS
1972 struct trie *t = (struct trie *)tb->tb_data;
1974 if (tb->tb_data == tb->__data)
1975 free_percpu(t->stats);
1976 #endif /* CONFIG_IP_FIB_TRIE_STATS */
1980 void fib_free_table(struct fib_table *tb)
1982 call_rcu(&tb->rcu, __trie_free_rcu);
1985 static int fn_trie_dump_leaf(struct key_vector *l, struct fib_table *tb,
1986 struct sk_buff *skb, struct netlink_callback *cb)
1988 __be32 xkey = htonl(l->key);
1989 struct fib_alias *fa;
1995 /* rcu_read_lock is hold by caller */
1996 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
2004 if (tb->tb_id != fa->tb_id) {
2009 err = fib_dump_info(skb, NETLINK_CB(cb->skb).portid,
2010 cb->nlh->nlmsg_seq, RTM_NEWROUTE,
2011 tb->tb_id, fa->fa_type,
2012 xkey, KEYLENGTH - fa->fa_slen,
2013 fa->fa_tos, fa->fa_info, NLM_F_MULTI);
2025 /* rcu_read_lock needs to be hold by caller from readside */
2026 int fib_table_dump(struct fib_table *tb, struct sk_buff *skb,
2027 struct netlink_callback *cb)
2029 struct trie *t = (struct trie *)tb->tb_data;
2030 struct key_vector *l, *tp = t->kv;
2031 /* Dump starting at last key.
2032 * Note: 0.0.0.0/0 (ie default) is first key.
2034 int count = cb->args[2];
2035 t_key key = cb->args[3];
2037 while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
2040 err = fn_trie_dump_leaf(l, tb, skb, cb);
2043 cb->args[2] = count;
2050 memset(&cb->args[4], 0,
2051 sizeof(cb->args) - 4*sizeof(cb->args[0]));
2053 /* stop loop if key wrapped back to 0 */
2059 cb->args[2] = count;
2064 void __init fib_trie_init(void)
2066 fn_alias_kmem = kmem_cache_create("ip_fib_alias",
2067 sizeof(struct fib_alias),
2068 0, SLAB_PANIC, NULL);
2070 trie_leaf_kmem = kmem_cache_create("ip_fib_trie",
2072 0, SLAB_PANIC, NULL);
2075 struct fib_table *fib_trie_table(u32 id, struct fib_table *alias)
2077 struct fib_table *tb;
2079 size_t sz = sizeof(*tb);
2082 sz += sizeof(struct trie);
2084 tb = kzalloc(sz, GFP_KERNEL);
2089 tb->tb_num_default = 0;
2090 tb->tb_data = (alias ? alias->__data : tb->__data);
2095 t = (struct trie *) tb->tb_data;
2096 t->kv[0].pos = KEYLENGTH;
2097 t->kv[0].slen = KEYLENGTH;
2098 #ifdef CONFIG_IP_FIB_TRIE_STATS
2099 t->stats = alloc_percpu(struct trie_use_stats);
2109 #ifdef CONFIG_PROC_FS
2110 /* Depth first Trie walk iterator */
2111 struct fib_trie_iter {
2112 struct seq_net_private p;
2113 struct fib_table *tb;
2114 struct key_vector *tnode;
2119 static struct key_vector *fib_trie_get_next(struct fib_trie_iter *iter)
2121 unsigned long cindex = iter->index;
2122 struct key_vector *pn = iter->tnode;
2125 pr_debug("get_next iter={node=%p index=%d depth=%d}\n",
2126 iter->tnode, iter->index, iter->depth);
2128 while (!IS_TRIE(pn)) {
2129 while (cindex < child_length(pn)) {
2130 struct key_vector *n = get_child_rcu(pn, cindex++);
2137 iter->index = cindex;
2139 /* push down one level */
2148 /* Current node exhausted, pop back up */
2150 pn = node_parent_rcu(pn);
2151 cindex = get_index(pkey, pn) + 1;
2155 /* record root node so further searches know we are done */
2162 static struct key_vector *fib_trie_get_first(struct fib_trie_iter *iter,
2165 struct key_vector *n, *pn;
2171 n = rcu_dereference(pn->tnode[0]);
2188 static void trie_collect_stats(struct trie *t, struct trie_stat *s)
2190 struct key_vector *n;
2191 struct fib_trie_iter iter;
2193 memset(s, 0, sizeof(*s));
2196 for (n = fib_trie_get_first(&iter, t); n; n = fib_trie_get_next(&iter)) {
2198 struct fib_alias *fa;
2201 s->totdepth += iter.depth;
2202 if (iter.depth > s->maxdepth)
2203 s->maxdepth = iter.depth;
2205 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list)
2209 if (n->bits < MAX_STAT_DEPTH)
2210 s->nodesizes[n->bits]++;
2211 s->nullpointers += tn_info(n)->empty_children;
2218 * This outputs /proc/net/fib_triestats
2220 static void trie_show_stats(struct seq_file *seq, struct trie_stat *stat)
2222 unsigned int i, max, pointers, bytes, avdepth;
2225 avdepth = stat->totdepth*100 / stat->leaves;
2229 seq_printf(seq, "\tAver depth: %u.%02d\n",
2230 avdepth / 100, avdepth % 100);
2231 seq_printf(seq, "\tMax depth: %u\n", stat->maxdepth);
2233 seq_printf(seq, "\tLeaves: %u\n", stat->leaves);
2234 bytes = LEAF_SIZE * stat->leaves;
2236 seq_printf(seq, "\tPrefixes: %u\n", stat->prefixes);
2237 bytes += sizeof(struct fib_alias) * stat->prefixes;
2239 seq_printf(seq, "\tInternal nodes: %u\n\t", stat->tnodes);
2240 bytes += TNODE_SIZE(0) * stat->tnodes;
2242 max = MAX_STAT_DEPTH;
2243 while (max > 0 && stat->nodesizes[max-1] == 0)
2247 for (i = 1; i < max; i++)
2248 if (stat->nodesizes[i] != 0) {
2249 seq_printf(seq, " %u: %u", i, stat->nodesizes[i]);
2250 pointers += (1<<i) * stat->nodesizes[i];
2252 seq_putc(seq, '\n');
2253 seq_printf(seq, "\tPointers: %u\n", pointers);
2255 bytes += sizeof(struct key_vector *) * pointers;
2256 seq_printf(seq, "Null ptrs: %u\n", stat->nullpointers);
2257 seq_printf(seq, "Total size: %u kB\n", (bytes + 1023) / 1024);
2260 #ifdef CONFIG_IP_FIB_TRIE_STATS
2261 static void trie_show_usage(struct seq_file *seq,
2262 const struct trie_use_stats __percpu *stats)
2264 struct trie_use_stats s = { 0 };
2267 /* loop through all of the CPUs and gather up the stats */
2268 for_each_possible_cpu(cpu) {
2269 const struct trie_use_stats *pcpu = per_cpu_ptr(stats, cpu);
2271 s.gets += pcpu->gets;
2272 s.backtrack += pcpu->backtrack;
2273 s.semantic_match_passed += pcpu->semantic_match_passed;
2274 s.semantic_match_miss += pcpu->semantic_match_miss;
2275 s.null_node_hit += pcpu->null_node_hit;
2276 s.resize_node_skipped += pcpu->resize_node_skipped;
2279 seq_printf(seq, "\nCounters:\n---------\n");
2280 seq_printf(seq, "gets = %u\n", s.gets);
2281 seq_printf(seq, "backtracks = %u\n", s.backtrack);
2282 seq_printf(seq, "semantic match passed = %u\n",
2283 s.semantic_match_passed);
2284 seq_printf(seq, "semantic match miss = %u\n", s.semantic_match_miss);
2285 seq_printf(seq, "null node hit= %u\n", s.null_node_hit);
2286 seq_printf(seq, "skipped node resize = %u\n\n", s.resize_node_skipped);
2288 #endif /* CONFIG_IP_FIB_TRIE_STATS */
2290 static void fib_table_print(struct seq_file *seq, struct fib_table *tb)
2292 if (tb->tb_id == RT_TABLE_LOCAL)
2293 seq_puts(seq, "Local:\n");
2294 else if (tb->tb_id == RT_TABLE_MAIN)
2295 seq_puts(seq, "Main:\n");
2297 seq_printf(seq, "Id %d:\n", tb->tb_id);
2301 static int fib_triestat_seq_show(struct seq_file *seq, void *v)
2303 struct net *net = (struct net *)seq->private;
2307 "Basic info: size of leaf:"
2308 " %zd bytes, size of tnode: %zd bytes.\n",
2309 LEAF_SIZE, TNODE_SIZE(0));
2311 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2312 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2313 struct fib_table *tb;
2315 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2316 struct trie *t = (struct trie *) tb->tb_data;
2317 struct trie_stat stat;
2322 fib_table_print(seq, tb);
2324 trie_collect_stats(t, &stat);
2325 trie_show_stats(seq, &stat);
2326 #ifdef CONFIG_IP_FIB_TRIE_STATS
2327 trie_show_usage(seq, t->stats);
2335 static int fib_triestat_seq_open(struct inode *inode, struct file *file)
2337 return single_open_net(inode, file, fib_triestat_seq_show);
2340 static const struct file_operations fib_triestat_fops = {
2341 .owner = THIS_MODULE,
2342 .open = fib_triestat_seq_open,
2344 .llseek = seq_lseek,
2345 .release = single_release_net,
2348 static struct key_vector *fib_trie_get_idx(struct seq_file *seq, loff_t pos)
2350 struct fib_trie_iter *iter = seq->private;
2351 struct net *net = seq_file_net(seq);
2355 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2356 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2357 struct fib_table *tb;
2359 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2360 struct key_vector *n;
2362 for (n = fib_trie_get_first(iter,
2363 (struct trie *) tb->tb_data);
2364 n; n = fib_trie_get_next(iter))
2375 static void *fib_trie_seq_start(struct seq_file *seq, loff_t *pos)
2379 return fib_trie_get_idx(seq, *pos);
2382 static void *fib_trie_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2384 struct fib_trie_iter *iter = seq->private;
2385 struct net *net = seq_file_net(seq);
2386 struct fib_table *tb = iter->tb;
2387 struct hlist_node *tb_node;
2389 struct key_vector *n;
2392 /* next node in same table */
2393 n = fib_trie_get_next(iter);
2397 /* walk rest of this hash chain */
2398 h = tb->tb_id & (FIB_TABLE_HASHSZ - 1);
2399 while ((tb_node = rcu_dereference(hlist_next_rcu(&tb->tb_hlist)))) {
2400 tb = hlist_entry(tb_node, struct fib_table, tb_hlist);
2401 n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2406 /* new hash chain */
2407 while (++h < FIB_TABLE_HASHSZ) {
2408 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2409 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2410 n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2422 static void fib_trie_seq_stop(struct seq_file *seq, void *v)
2428 static void seq_indent(struct seq_file *seq, int n)
2434 static inline const char *rtn_scope(char *buf, size_t len, enum rt_scope_t s)
2437 case RT_SCOPE_UNIVERSE: return "universe";
2438 case RT_SCOPE_SITE: return "site";
2439 case RT_SCOPE_LINK: return "link";
2440 case RT_SCOPE_HOST: return "host";
2441 case RT_SCOPE_NOWHERE: return "nowhere";
2443 snprintf(buf, len, "scope=%d", s);
2448 static const char *const rtn_type_names[__RTN_MAX] = {
2449 [RTN_UNSPEC] = "UNSPEC",
2450 [RTN_UNICAST] = "UNICAST",
2451 [RTN_LOCAL] = "LOCAL",
2452 [RTN_BROADCAST] = "BROADCAST",
2453 [RTN_ANYCAST] = "ANYCAST",
2454 [RTN_MULTICAST] = "MULTICAST",
2455 [RTN_BLACKHOLE] = "BLACKHOLE",
2456 [RTN_UNREACHABLE] = "UNREACHABLE",
2457 [RTN_PROHIBIT] = "PROHIBIT",
2458 [RTN_THROW] = "THROW",
2460 [RTN_XRESOLVE] = "XRESOLVE",
2463 static inline const char *rtn_type(char *buf, size_t len, unsigned int t)
2465 if (t < __RTN_MAX && rtn_type_names[t])
2466 return rtn_type_names[t];
2467 snprintf(buf, len, "type %u", t);
2471 /* Pretty print the trie */
2472 static int fib_trie_seq_show(struct seq_file *seq, void *v)
2474 const struct fib_trie_iter *iter = seq->private;
2475 struct key_vector *n = v;
2477 if (IS_TRIE(node_parent_rcu(n)))
2478 fib_table_print(seq, iter->tb);
2481 __be32 prf = htonl(n->key);
2483 seq_indent(seq, iter->depth-1);
2484 seq_printf(seq, " +-- %pI4/%zu %u %u %u\n",
2485 &prf, KEYLENGTH - n->pos - n->bits, n->bits,
2486 tn_info(n)->full_children,
2487 tn_info(n)->empty_children);
2489 __be32 val = htonl(n->key);
2490 struct fib_alias *fa;
2492 seq_indent(seq, iter->depth);
2493 seq_printf(seq, " |-- %pI4\n", &val);
2495 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list) {
2496 char buf1[32], buf2[32];
2498 seq_indent(seq, iter->depth + 1);
2499 seq_printf(seq, " /%zu %s %s",
2500 KEYLENGTH - fa->fa_slen,
2501 rtn_scope(buf1, sizeof(buf1),
2502 fa->fa_info->fib_scope),
2503 rtn_type(buf2, sizeof(buf2),
2506 seq_printf(seq, " tos=%d", fa->fa_tos);
2507 seq_putc(seq, '\n');
2514 static const struct seq_operations fib_trie_seq_ops = {
2515 .start = fib_trie_seq_start,
2516 .next = fib_trie_seq_next,
2517 .stop = fib_trie_seq_stop,
2518 .show = fib_trie_seq_show,
2521 static int fib_trie_seq_open(struct inode *inode, struct file *file)
2523 return seq_open_net(inode, file, &fib_trie_seq_ops,
2524 sizeof(struct fib_trie_iter));
2527 static const struct file_operations fib_trie_fops = {
2528 .owner = THIS_MODULE,
2529 .open = fib_trie_seq_open,
2531 .llseek = seq_lseek,
2532 .release = seq_release_net,
2535 struct fib_route_iter {
2536 struct seq_net_private p;
2537 struct fib_table *main_tb;
2538 struct key_vector *tnode;
2543 static struct key_vector *fib_route_get_idx(struct fib_route_iter *iter,
2546 struct key_vector *l, **tp = &iter->tnode;
2549 /* use cached location of previously found key */
2550 if (iter->pos > 0 && pos >= iter->pos) {
2559 while ((l = leaf_walk_rcu(tp, key)) && (pos-- > 0)) {
2564 /* handle unlikely case of a key wrap */
2570 iter->key = l->key; /* remember it */
2572 iter->pos = 0; /* forget it */
2577 static void *fib_route_seq_start(struct seq_file *seq, loff_t *pos)
2580 struct fib_route_iter *iter = seq->private;
2581 struct fib_table *tb;
2586 tb = fib_get_table(seq_file_net(seq), RT_TABLE_MAIN);
2591 t = (struct trie *)tb->tb_data;
2592 iter->tnode = t->kv;
2595 return fib_route_get_idx(iter, *pos);
2598 iter->key = KEY_MAX;
2600 return SEQ_START_TOKEN;
2603 static void *fib_route_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2605 struct fib_route_iter *iter = seq->private;
2606 struct key_vector *l = NULL;
2607 t_key key = iter->key + 1;
2611 /* only allow key of 0 for start of sequence */
2612 if ((v == SEQ_START_TOKEN) || key)
2613 l = leaf_walk_rcu(&iter->tnode, key);
2625 static void fib_route_seq_stop(struct seq_file *seq, void *v)
2631 static unsigned int fib_flag_trans(int type, __be32 mask, const struct fib_info *fi)
2633 unsigned int flags = 0;
2635 if (type == RTN_UNREACHABLE || type == RTN_PROHIBIT)
2637 if (fi && fi->fib_nh->nh_gw)
2638 flags |= RTF_GATEWAY;
2639 if (mask == htonl(0xFFFFFFFF))
2646 * This outputs /proc/net/route.
2647 * The format of the file is not supposed to be changed
2648 * and needs to be same as fib_hash output to avoid breaking
2651 static int fib_route_seq_show(struct seq_file *seq, void *v)
2653 struct fib_route_iter *iter = seq->private;
2654 struct fib_table *tb = iter->main_tb;
2655 struct fib_alias *fa;
2656 struct key_vector *l = v;
2659 if (v == SEQ_START_TOKEN) {
2660 seq_printf(seq, "%-127s\n", "Iface\tDestination\tGateway "
2661 "\tFlags\tRefCnt\tUse\tMetric\tMask\t\tMTU"
2666 prefix = htonl(l->key);
2668 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
2669 const struct fib_info *fi = fa->fa_info;
2670 __be32 mask = inet_make_mask(KEYLENGTH - fa->fa_slen);
2671 unsigned int flags = fib_flag_trans(fa->fa_type, mask, fi);
2673 if ((fa->fa_type == RTN_BROADCAST) ||
2674 (fa->fa_type == RTN_MULTICAST))
2677 if (fa->tb_id != tb->tb_id)
2680 seq_setwidth(seq, 127);
2684 "%s\t%08X\t%08X\t%04X\t%d\t%u\t"
2685 "%d\t%08X\t%d\t%u\t%u",
2686 fi->fib_dev ? fi->fib_dev->name : "*",
2688 fi->fib_nh->nh_gw, flags, 0, 0,
2692 fi->fib_advmss + 40 : 0),
2697 "*\t%08X\t%08X\t%04X\t%d\t%u\t"
2698 "%d\t%08X\t%d\t%u\t%u",
2699 prefix, 0, flags, 0, 0, 0,
2708 static const struct seq_operations fib_route_seq_ops = {
2709 .start = fib_route_seq_start,
2710 .next = fib_route_seq_next,
2711 .stop = fib_route_seq_stop,
2712 .show = fib_route_seq_show,
2715 static int fib_route_seq_open(struct inode *inode, struct file *file)
2717 return seq_open_net(inode, file, &fib_route_seq_ops,
2718 sizeof(struct fib_route_iter));
2721 static const struct file_operations fib_route_fops = {
2722 .owner = THIS_MODULE,
2723 .open = fib_route_seq_open,
2725 .llseek = seq_lseek,
2726 .release = seq_release_net,
2729 int __net_init fib_proc_init(struct net *net)
2731 if (!proc_create("fib_trie", S_IRUGO, net->proc_net, &fib_trie_fops))
2734 if (!proc_create("fib_triestat", S_IRUGO, net->proc_net,
2735 &fib_triestat_fops))
2738 if (!proc_create("route", S_IRUGO, net->proc_net, &fib_route_fops))
2744 remove_proc_entry("fib_triestat", net->proc_net);
2746 remove_proc_entry("fib_trie", net->proc_net);
2751 void __net_exit fib_proc_exit(struct net *net)
2753 remove_proc_entry("fib_trie", net->proc_net);
2754 remove_proc_entry("fib_triestat", net->proc_net);
2755 remove_proc_entry("route", net->proc_net);
2758 #endif /* CONFIG_PROC_FS */