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/cache.h>
54 #include <linux/uaccess.h>
55 #include <linux/bitops.h>
56 #include <linux/types.h>
57 #include <linux/kernel.h>
59 #include <linux/string.h>
60 #include <linux/socket.h>
61 #include <linux/sockios.h>
62 #include <linux/errno.h>
64 #include <linux/inet.h>
65 #include <linux/inetdevice.h>
66 #include <linux/netdevice.h>
67 #include <linux/if_arp.h>
68 #include <linux/proc_fs.h>
69 #include <linux/rcupdate.h>
70 #include <linux/skbuff.h>
71 #include <linux/netlink.h>
72 #include <linux/init.h>
73 #include <linux/list.h>
74 #include <linux/slab.h>
75 #include <linux/export.h>
76 #include <linux/vmalloc.h>
77 #include <linux/notifier.h>
78 #include <net/net_namespace.h>
80 #include <net/protocol.h>
81 #include <net/route.h>
84 #include <net/ip_fib.h>
85 #include <net/fib_notifier.h>
86 #include <trace/events/fib.h>
87 #include "fib_lookup.h"
89 static int call_fib_entry_notifier(struct notifier_block *nb, struct net *net,
90 enum fib_event_type event_type, u32 dst,
91 int dst_len, struct fib_alias *fa)
93 struct fib_entry_notifier_info info = {
101 return call_fib4_notifier(nb, net, event_type, &info.info);
104 static int call_fib_entry_notifiers(struct net *net,
105 enum fib_event_type event_type, u32 dst,
106 int dst_len, struct fib_alias *fa,
107 struct netlink_ext_ack *extack)
109 struct fib_entry_notifier_info info = {
110 .info.extack = extack,
118 return call_fib4_notifiers(net, event_type, &info.info);
121 #define MAX_STAT_DEPTH 32
123 #define KEYLENGTH (8*sizeof(t_key))
124 #define KEY_MAX ((t_key)~0)
126 typedef unsigned int t_key;
128 #define IS_TRIE(n) ((n)->pos >= KEYLENGTH)
129 #define IS_TNODE(n) ((n)->bits)
130 #define IS_LEAF(n) (!(n)->bits)
134 unsigned char pos; /* 2log(KEYLENGTH) bits needed */
135 unsigned char bits; /* 2log(KEYLENGTH) bits needed */
138 /* This list pointer if valid if (pos | bits) == 0 (LEAF) */
139 struct hlist_head leaf;
140 /* This array is valid if (pos | bits) > 0 (TNODE) */
141 struct key_vector __rcu *tnode[0];
147 t_key empty_children; /* KEYLENGTH bits needed */
148 t_key full_children; /* KEYLENGTH bits needed */
149 struct key_vector __rcu *parent;
150 struct key_vector kv[1];
151 #define tn_bits kv[0].bits
154 #define TNODE_SIZE(n) offsetof(struct tnode, kv[0].tnode[n])
155 #define LEAF_SIZE TNODE_SIZE(1)
157 #ifdef CONFIG_IP_FIB_TRIE_STATS
158 struct trie_use_stats {
160 unsigned int backtrack;
161 unsigned int semantic_match_passed;
162 unsigned int semantic_match_miss;
163 unsigned int null_node_hit;
164 unsigned int resize_node_skipped;
169 unsigned int totdepth;
170 unsigned int maxdepth;
173 unsigned int nullpointers;
174 unsigned int prefixes;
175 unsigned int nodesizes[MAX_STAT_DEPTH];
179 struct key_vector kv[1];
180 #ifdef CONFIG_IP_FIB_TRIE_STATS
181 struct trie_use_stats __percpu *stats;
185 static struct key_vector *resize(struct trie *t, struct key_vector *tn);
186 static size_t tnode_free_size;
189 * synchronize_rcu after call_rcu for that many pages; it should be especially
190 * useful before resizing the root node with PREEMPT_NONE configs; the value was
191 * obtained experimentally, aiming to avoid visible slowdown.
193 static const int sync_pages = 128;
195 static struct kmem_cache *fn_alias_kmem __ro_after_init;
196 static struct kmem_cache *trie_leaf_kmem __ro_after_init;
198 static inline struct tnode *tn_info(struct key_vector *kv)
200 return container_of(kv, struct tnode, kv[0]);
203 /* caller must hold RTNL */
204 #define node_parent(tn) rtnl_dereference(tn_info(tn)->parent)
205 #define get_child(tn, i) rtnl_dereference((tn)->tnode[i])
207 /* caller must hold RCU read lock or RTNL */
208 #define node_parent_rcu(tn) rcu_dereference_rtnl(tn_info(tn)->parent)
209 #define get_child_rcu(tn, i) rcu_dereference_rtnl((tn)->tnode[i])
211 /* wrapper for rcu_assign_pointer */
212 static inline void node_set_parent(struct key_vector *n, struct key_vector *tp)
215 rcu_assign_pointer(tn_info(n)->parent, tp);
218 #define NODE_INIT_PARENT(n, p) RCU_INIT_POINTER(tn_info(n)->parent, p)
220 /* This provides us with the number of children in this node, in the case of a
221 * leaf this will return 0 meaning none of the children are accessible.
223 static inline unsigned long child_length(const struct key_vector *tn)
225 return (1ul << tn->bits) & ~(1ul);
228 #define get_cindex(key, kv) (((key) ^ (kv)->key) >> (kv)->pos)
230 static inline unsigned long get_index(t_key key, struct key_vector *kv)
232 unsigned long index = key ^ kv->key;
234 if ((BITS_PER_LONG <= KEYLENGTH) && (KEYLENGTH == kv->pos))
237 return index >> kv->pos;
240 /* To understand this stuff, an understanding of keys and all their bits is
241 * necessary. Every node in the trie has a key associated with it, but not
242 * all of the bits in that key are significant.
244 * Consider a node 'n' and its parent 'tp'.
246 * If n is a leaf, every bit in its key is significant. Its presence is
247 * necessitated by path compression, since during a tree traversal (when
248 * searching for a leaf - unless we are doing an insertion) we will completely
249 * ignore all skipped bits we encounter. Thus we need to verify, at the end of
250 * a potentially successful search, that we have indeed been walking the
253 * Note that we can never "miss" the correct key in the tree if present by
254 * following the wrong path. Path compression ensures that segments of the key
255 * that are the same for all keys with a given prefix are skipped, but the
256 * skipped part *is* identical for each node in the subtrie below the skipped
257 * bit! trie_insert() in this implementation takes care of that.
259 * if n is an internal node - a 'tnode' here, the various parts of its key
260 * have many different meanings.
263 * _________________________________________________________________
264 * | i | i | i | i | i | i | i | N | N | N | S | S | S | S | S | C |
265 * -----------------------------------------------------------------
266 * 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16
268 * _________________________________________________________________
269 * | C | C | C | u | u | u | u | u | u | u | u | u | u | u | u | u |
270 * -----------------------------------------------------------------
271 * 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
278 * First, let's just ignore the bits that come before the parent tp, that is
279 * the bits from (tp->pos + tp->bits) to 31. They are *known* but at this
280 * point we do not use them for anything.
282 * The bits from (tp->pos) to (tp->pos + tp->bits - 1) - "N", above - are the
283 * index into the parent's child array. That is, they will be used to find
284 * 'n' among tp's children.
286 * The bits from (n->pos + n->bits) to (tp->pos - 1) - "S" - are skipped bits
289 * All the bits we have seen so far are significant to the node n. The rest
290 * of the bits are really not needed or indeed known in n->key.
292 * The bits from (n->pos) to (n->pos + n->bits - 1) - "C" - are the index into
293 * n's child array, and will of course be different for each child.
295 * The rest of the bits, from 0 to (n->pos -1) - "u" - are completely unknown
299 static const int halve_threshold = 25;
300 static const int inflate_threshold = 50;
301 static const int halve_threshold_root = 15;
302 static const int inflate_threshold_root = 30;
304 static void __alias_free_mem(struct rcu_head *head)
306 struct fib_alias *fa = container_of(head, struct fib_alias, rcu);
307 kmem_cache_free(fn_alias_kmem, fa);
310 static inline void alias_free_mem_rcu(struct fib_alias *fa)
312 call_rcu(&fa->rcu, __alias_free_mem);
315 #define TNODE_KMALLOC_MAX \
316 ilog2((PAGE_SIZE - TNODE_SIZE(0)) / sizeof(struct key_vector *))
317 #define TNODE_VMALLOC_MAX \
318 ilog2((SIZE_MAX - TNODE_SIZE(0)) / sizeof(struct key_vector *))
320 static void __node_free_rcu(struct rcu_head *head)
322 struct tnode *n = container_of(head, struct tnode, rcu);
325 kmem_cache_free(trie_leaf_kmem, n);
330 #define node_free(n) call_rcu(&tn_info(n)->rcu, __node_free_rcu)
332 static struct tnode *tnode_alloc(int bits)
336 /* verify bits is within bounds */
337 if (bits > TNODE_VMALLOC_MAX)
340 /* determine size and verify it is non-zero and didn't overflow */
341 size = TNODE_SIZE(1ul << bits);
343 if (size <= PAGE_SIZE)
344 return kzalloc(size, GFP_KERNEL);
346 return vzalloc(size);
349 static inline void empty_child_inc(struct key_vector *n)
351 ++tn_info(n)->empty_children ? : ++tn_info(n)->full_children;
354 static inline void empty_child_dec(struct key_vector *n)
356 tn_info(n)->empty_children-- ? : tn_info(n)->full_children--;
359 static struct key_vector *leaf_new(t_key key, struct fib_alias *fa)
361 struct key_vector *l;
364 kv = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
368 /* initialize key vector */
373 l->slen = fa->fa_slen;
375 /* link leaf to fib alias */
376 INIT_HLIST_HEAD(&l->leaf);
377 hlist_add_head(&fa->fa_list, &l->leaf);
382 static struct key_vector *tnode_new(t_key key, int pos, int bits)
384 unsigned int shift = pos + bits;
385 struct key_vector *tn;
388 /* verify bits and pos their msb bits clear and values are valid */
389 BUG_ON(!bits || (shift > KEYLENGTH));
391 tnode = tnode_alloc(bits);
395 pr_debug("AT %p s=%zu %zu\n", tnode, TNODE_SIZE(0),
396 sizeof(struct key_vector *) << bits);
398 if (bits == KEYLENGTH)
399 tnode->full_children = 1;
401 tnode->empty_children = 1ul << bits;
404 tn->key = (shift < KEYLENGTH) ? (key >> shift) << shift : 0;
412 /* Check whether a tnode 'n' is "full", i.e. it is an internal node
413 * and no bits are skipped. See discussion in dyntree paper p. 6
415 static inline int tnode_full(struct key_vector *tn, struct key_vector *n)
417 return n && ((n->pos + n->bits) == tn->pos) && IS_TNODE(n);
420 /* Add a child at position i overwriting the old value.
421 * Update the value of full_children and empty_children.
423 static void put_child(struct key_vector *tn, unsigned long i,
424 struct key_vector *n)
426 struct key_vector *chi = get_child(tn, i);
429 BUG_ON(i >= child_length(tn));
431 /* update emptyChildren, overflow into fullChildren */
437 /* update fullChildren */
438 wasfull = tnode_full(tn, chi);
439 isfull = tnode_full(tn, n);
441 if (wasfull && !isfull)
442 tn_info(tn)->full_children--;
443 else if (!wasfull && isfull)
444 tn_info(tn)->full_children++;
446 if (n && (tn->slen < n->slen))
449 rcu_assign_pointer(tn->tnode[i], n);
452 static void update_children(struct key_vector *tn)
456 /* update all of the child parent pointers */
457 for (i = child_length(tn); i;) {
458 struct key_vector *inode = get_child(tn, --i);
463 /* Either update the children of a tnode that
464 * already belongs to us or update the child
465 * to point to ourselves.
467 if (node_parent(inode) == tn)
468 update_children(inode);
470 node_set_parent(inode, tn);
474 static inline void put_child_root(struct key_vector *tp, t_key key,
475 struct key_vector *n)
478 rcu_assign_pointer(tp->tnode[0], n);
480 put_child(tp, get_index(key, tp), n);
483 static inline void tnode_free_init(struct key_vector *tn)
485 tn_info(tn)->rcu.next = NULL;
488 static inline void tnode_free_append(struct key_vector *tn,
489 struct key_vector *n)
491 tn_info(n)->rcu.next = tn_info(tn)->rcu.next;
492 tn_info(tn)->rcu.next = &tn_info(n)->rcu;
495 static void tnode_free(struct key_vector *tn)
497 struct callback_head *head = &tn_info(tn)->rcu;
501 tnode_free_size += TNODE_SIZE(1ul << tn->bits);
504 tn = container_of(head, struct tnode, rcu)->kv;
507 if (tnode_free_size >= PAGE_SIZE * sync_pages) {
513 static struct key_vector *replace(struct trie *t,
514 struct key_vector *oldtnode,
515 struct key_vector *tn)
517 struct key_vector *tp = node_parent(oldtnode);
520 /* setup the parent pointer out of and back into this node */
521 NODE_INIT_PARENT(tn, tp);
522 put_child_root(tp, tn->key, tn);
524 /* update all of the child parent pointers */
527 /* all pointers should be clean so we are done */
528 tnode_free(oldtnode);
530 /* resize children now that oldtnode is freed */
531 for (i = child_length(tn); i;) {
532 struct key_vector *inode = get_child(tn, --i);
534 /* resize child node */
535 if (tnode_full(tn, inode))
536 tn = resize(t, inode);
542 static struct key_vector *inflate(struct trie *t,
543 struct key_vector *oldtnode)
545 struct key_vector *tn;
549 pr_debug("In inflate\n");
551 tn = tnode_new(oldtnode->key, oldtnode->pos - 1, oldtnode->bits + 1);
555 /* prepare oldtnode to be freed */
556 tnode_free_init(oldtnode);
558 /* Assemble all of the pointers in our cluster, in this case that
559 * represents all of the pointers out of our allocated nodes that
560 * point to existing tnodes and the links between our allocated
563 for (i = child_length(oldtnode), m = 1u << tn->pos; i;) {
564 struct key_vector *inode = get_child(oldtnode, --i);
565 struct key_vector *node0, *node1;
572 /* A leaf or an internal node with skipped bits */
573 if (!tnode_full(oldtnode, inode)) {
574 put_child(tn, get_index(inode->key, tn), inode);
578 /* drop the node in the old tnode free list */
579 tnode_free_append(oldtnode, inode);
581 /* An internal node with two children */
582 if (inode->bits == 1) {
583 put_child(tn, 2 * i + 1, get_child(inode, 1));
584 put_child(tn, 2 * i, get_child(inode, 0));
588 /* We will replace this node 'inode' with two new
589 * ones, 'node0' and 'node1', each with half of the
590 * original children. The two new nodes will have
591 * a position one bit further down the key and this
592 * means that the "significant" part of their keys
593 * (see the discussion near the top of this file)
594 * will differ by one bit, which will be "0" in
595 * node0's key and "1" in node1's key. Since we are
596 * moving the key position by one step, the bit that
597 * we are moving away from - the bit at position
598 * (tn->pos) - is the one that will differ between
599 * node0 and node1. So... we synthesize that bit in the
602 node1 = tnode_new(inode->key | m, inode->pos, inode->bits - 1);
605 node0 = tnode_new(inode->key, inode->pos, inode->bits - 1);
607 tnode_free_append(tn, node1);
610 tnode_free_append(tn, node0);
612 /* populate child pointers in new nodes */
613 for (k = child_length(inode), j = k / 2; j;) {
614 put_child(node1, --j, get_child(inode, --k));
615 put_child(node0, j, get_child(inode, j));
616 put_child(node1, --j, get_child(inode, --k));
617 put_child(node0, j, get_child(inode, j));
620 /* link new nodes to parent */
621 NODE_INIT_PARENT(node1, tn);
622 NODE_INIT_PARENT(node0, tn);
624 /* link parent to nodes */
625 put_child(tn, 2 * i + 1, node1);
626 put_child(tn, 2 * i, node0);
629 /* setup the parent pointers into and out of this node */
630 return replace(t, oldtnode, tn);
632 /* all pointers should be clean so we are done */
638 static struct key_vector *halve(struct trie *t,
639 struct key_vector *oldtnode)
641 struct key_vector *tn;
644 pr_debug("In halve\n");
646 tn = tnode_new(oldtnode->key, oldtnode->pos + 1, oldtnode->bits - 1);
650 /* prepare oldtnode to be freed */
651 tnode_free_init(oldtnode);
653 /* Assemble all of the pointers in our cluster, in this case that
654 * represents all of the pointers out of our allocated nodes that
655 * point to existing tnodes and the links between our allocated
658 for (i = child_length(oldtnode); i;) {
659 struct key_vector *node1 = get_child(oldtnode, --i);
660 struct key_vector *node0 = get_child(oldtnode, --i);
661 struct key_vector *inode;
663 /* At least one of the children is empty */
664 if (!node1 || !node0) {
665 put_child(tn, i / 2, node1 ? : node0);
669 /* Two nonempty children */
670 inode = tnode_new(node0->key, oldtnode->pos, 1);
673 tnode_free_append(tn, inode);
675 /* initialize pointers out of node */
676 put_child(inode, 1, node1);
677 put_child(inode, 0, node0);
678 NODE_INIT_PARENT(inode, tn);
680 /* link parent to node */
681 put_child(tn, i / 2, inode);
684 /* setup the parent pointers into and out of this node */
685 return replace(t, oldtnode, tn);
687 /* all pointers should be clean so we are done */
693 static struct key_vector *collapse(struct trie *t,
694 struct key_vector *oldtnode)
696 struct key_vector *n, *tp;
699 /* scan the tnode looking for that one child that might still exist */
700 for (n = NULL, i = child_length(oldtnode); !n && i;)
701 n = get_child(oldtnode, --i);
703 /* compress one level */
704 tp = node_parent(oldtnode);
705 put_child_root(tp, oldtnode->key, n);
706 node_set_parent(n, tp);
714 static unsigned char update_suffix(struct key_vector *tn)
716 unsigned char slen = tn->pos;
717 unsigned long stride, i;
718 unsigned char slen_max;
720 /* only vector 0 can have a suffix length greater than or equal to
721 * tn->pos + tn->bits, the second highest node will have a suffix
722 * length at most of tn->pos + tn->bits - 1
724 slen_max = min_t(unsigned char, tn->pos + tn->bits - 1, tn->slen);
726 /* search though the list of children looking for nodes that might
727 * have a suffix greater than the one we currently have. This is
728 * why we start with a stride of 2 since a stride of 1 would
729 * represent the nodes with suffix length equal to tn->pos
731 for (i = 0, stride = 0x2ul ; i < child_length(tn); i += stride) {
732 struct key_vector *n = get_child(tn, i);
734 if (!n || (n->slen <= slen))
737 /* update stride and slen based on new value */
738 stride <<= (n->slen - slen);
742 /* stop searching if we have hit the maximum possible value */
743 if (slen >= slen_max)
752 /* From "Implementing a dynamic compressed trie" by Stefan Nilsson of
753 * the Helsinki University of Technology and Matti Tikkanen of Nokia
754 * Telecommunications, page 6:
755 * "A node is doubled if the ratio of non-empty children to all
756 * children in the *doubled* node is at least 'high'."
758 * 'high' in this instance is the variable 'inflate_threshold'. It
759 * is expressed as a percentage, so we multiply it with
760 * child_length() and instead of multiplying by 2 (since the
761 * child array will be doubled by inflate()) and multiplying
762 * the left-hand side by 100 (to handle the percentage thing) we
763 * multiply the left-hand side by 50.
765 * The left-hand side may look a bit weird: child_length(tn)
766 * - tn->empty_children is of course the number of non-null children
767 * in the current node. tn->full_children is the number of "full"
768 * children, that is non-null tnodes with a skip value of 0.
769 * All of those will be doubled in the resulting inflated tnode, so
770 * we just count them one extra time here.
772 * A clearer way to write this would be:
774 * to_be_doubled = tn->full_children;
775 * not_to_be_doubled = child_length(tn) - tn->empty_children -
778 * new_child_length = child_length(tn) * 2;
780 * new_fill_factor = 100 * (not_to_be_doubled + 2*to_be_doubled) /
782 * if (new_fill_factor >= inflate_threshold)
784 * ...and so on, tho it would mess up the while () loop.
787 * 100 * (not_to_be_doubled + 2*to_be_doubled) / new_child_length >=
791 * 100 * (not_to_be_doubled + 2*to_be_doubled) >=
792 * inflate_threshold * new_child_length
794 * expand not_to_be_doubled and to_be_doubled, and shorten:
795 * 100 * (child_length(tn) - tn->empty_children +
796 * tn->full_children) >= inflate_threshold * new_child_length
798 * expand new_child_length:
799 * 100 * (child_length(tn) - tn->empty_children +
800 * tn->full_children) >=
801 * inflate_threshold * child_length(tn) * 2
804 * 50 * (tn->full_children + child_length(tn) -
805 * tn->empty_children) >= inflate_threshold *
809 static inline bool should_inflate(struct key_vector *tp, struct key_vector *tn)
811 unsigned long used = child_length(tn);
812 unsigned long threshold = used;
814 /* Keep root node larger */
815 threshold *= IS_TRIE(tp) ? inflate_threshold_root : inflate_threshold;
816 used -= tn_info(tn)->empty_children;
817 used += tn_info(tn)->full_children;
819 /* if bits == KEYLENGTH then pos = 0, and will fail below */
821 return (used > 1) && tn->pos && ((50 * used) >= threshold);
824 static inline bool should_halve(struct key_vector *tp, struct key_vector *tn)
826 unsigned long used = child_length(tn);
827 unsigned long threshold = used;
829 /* Keep root node larger */
830 threshold *= IS_TRIE(tp) ? halve_threshold_root : halve_threshold;
831 used -= tn_info(tn)->empty_children;
833 /* if bits == KEYLENGTH then used = 100% on wrap, and will fail below */
835 return (used > 1) && (tn->bits > 1) && ((100 * used) < threshold);
838 static inline bool should_collapse(struct key_vector *tn)
840 unsigned long used = child_length(tn);
842 used -= tn_info(tn)->empty_children;
844 /* account for bits == KEYLENGTH case */
845 if ((tn->bits == KEYLENGTH) && tn_info(tn)->full_children)
848 /* One child or none, time to drop us from the trie */
853 static struct key_vector *resize(struct trie *t, struct key_vector *tn)
855 #ifdef CONFIG_IP_FIB_TRIE_STATS
856 struct trie_use_stats __percpu *stats = t->stats;
858 struct key_vector *tp = node_parent(tn);
859 unsigned long cindex = get_index(tn->key, tp);
860 int max_work = MAX_WORK;
862 pr_debug("In tnode_resize %p inflate_threshold=%d threshold=%d\n",
863 tn, inflate_threshold, halve_threshold);
865 /* track the tnode via the pointer from the parent instead of
866 * doing it ourselves. This way we can let RCU fully do its
867 * thing without us interfering
869 BUG_ON(tn != get_child(tp, cindex));
871 /* Double as long as the resulting node has a number of
872 * nonempty nodes that are above the threshold.
874 while (should_inflate(tp, tn) && max_work) {
877 #ifdef CONFIG_IP_FIB_TRIE_STATS
878 this_cpu_inc(stats->resize_node_skipped);
884 tn = get_child(tp, cindex);
887 /* update parent in case inflate failed */
888 tp = node_parent(tn);
890 /* Return if at least one inflate is run */
891 if (max_work != MAX_WORK)
894 /* Halve as long as the number of empty children in this
895 * node is above threshold.
897 while (should_halve(tp, tn) && max_work) {
900 #ifdef CONFIG_IP_FIB_TRIE_STATS
901 this_cpu_inc(stats->resize_node_skipped);
907 tn = get_child(tp, cindex);
910 /* Only one child remains */
911 if (should_collapse(tn))
912 return collapse(t, tn);
914 /* update parent in case halve failed */
915 return node_parent(tn);
918 static void node_pull_suffix(struct key_vector *tn, unsigned char slen)
920 unsigned char node_slen = tn->slen;
922 while ((node_slen > tn->pos) && (node_slen > slen)) {
923 slen = update_suffix(tn);
924 if (node_slen == slen)
927 tn = node_parent(tn);
928 node_slen = tn->slen;
932 static void node_push_suffix(struct key_vector *tn, unsigned char slen)
934 while (tn->slen < slen) {
936 tn = node_parent(tn);
940 /* rcu_read_lock needs to be hold by caller from readside */
941 static struct key_vector *fib_find_node(struct trie *t,
942 struct key_vector **tp, u32 key)
944 struct key_vector *pn, *n = t->kv;
945 unsigned long index = 0;
949 n = get_child_rcu(n, index);
954 index = get_cindex(key, n);
956 /* This bit of code is a bit tricky but it combines multiple
957 * checks into a single check. The prefix consists of the
958 * prefix plus zeros for the bits in the cindex. The index
959 * is the difference between the key and this value. From
960 * this we can actually derive several pieces of data.
961 * if (index >= (1ul << bits))
962 * we have a mismatch in skip bits and failed
964 * we know the value is cindex
966 * This check is safe even if bits == KEYLENGTH due to the
967 * fact that we can only allocate a node with 32 bits if a
968 * long is greater than 32 bits.
970 if (index >= (1ul << n->bits)) {
975 /* keep searching until we find a perfect match leaf or NULL */
976 } while (IS_TNODE(n));
983 /* Return the first fib alias matching TOS with
984 * priority less than or equal to PRIO.
986 static struct fib_alias *fib_find_alias(struct hlist_head *fah, u8 slen,
987 u8 tos, u32 prio, u32 tb_id)
989 struct fib_alias *fa;
994 hlist_for_each_entry(fa, fah, fa_list) {
995 if (fa->fa_slen < slen)
997 if (fa->fa_slen != slen)
999 if (fa->tb_id > tb_id)
1001 if (fa->tb_id != tb_id)
1003 if (fa->fa_tos > tos)
1005 if (fa->fa_info->fib_priority >= prio || fa->fa_tos < tos)
1012 static void trie_rebalance(struct trie *t, struct key_vector *tn)
1014 while (!IS_TRIE(tn))
1018 static int fib_insert_node(struct trie *t, struct key_vector *tp,
1019 struct fib_alias *new, t_key key)
1021 struct key_vector *n, *l;
1023 l = leaf_new(key, new);
1027 /* retrieve child from parent node */
1028 n = get_child(tp, get_index(key, tp));
1030 /* Case 2: n is a LEAF or a TNODE and the key doesn't match.
1032 * Add a new tnode here
1033 * first tnode need some special handling
1034 * leaves us in position for handling as case 3
1037 struct key_vector *tn;
1039 tn = tnode_new(key, __fls(key ^ n->key), 1);
1043 /* initialize routes out of node */
1044 NODE_INIT_PARENT(tn, tp);
1045 put_child(tn, get_index(key, tn) ^ 1, n);
1047 /* start adding routes into the node */
1048 put_child_root(tp, key, tn);
1049 node_set_parent(n, tn);
1051 /* parent now has a NULL spot where the leaf can go */
1055 /* Case 3: n is NULL, and will just insert a new leaf */
1056 node_push_suffix(tp, new->fa_slen);
1057 NODE_INIT_PARENT(l, tp);
1058 put_child_root(tp, key, l);
1059 trie_rebalance(t, tp);
1068 /* fib notifier for ADD is sent before calling fib_insert_alias with
1069 * the expectation that the only possible failure ENOMEM
1071 static int fib_insert_alias(struct trie *t, struct key_vector *tp,
1072 struct key_vector *l, struct fib_alias *new,
1073 struct fib_alias *fa, t_key key)
1076 return fib_insert_node(t, tp, new, key);
1079 hlist_add_before_rcu(&new->fa_list, &fa->fa_list);
1081 struct fib_alias *last;
1083 hlist_for_each_entry(last, &l->leaf, fa_list) {
1084 if (new->fa_slen < last->fa_slen)
1086 if ((new->fa_slen == last->fa_slen) &&
1087 (new->tb_id > last->tb_id))
1093 hlist_add_behind_rcu(&new->fa_list, &fa->fa_list);
1095 hlist_add_head_rcu(&new->fa_list, &l->leaf);
1098 /* if we added to the tail node then we need to update slen */
1099 if (l->slen < new->fa_slen) {
1100 l->slen = new->fa_slen;
1101 node_push_suffix(tp, new->fa_slen);
1107 static bool fib_valid_key_len(u32 key, u8 plen, struct netlink_ext_ack *extack)
1109 if (plen > KEYLENGTH) {
1110 NL_SET_ERR_MSG(extack, "Invalid prefix length");
1114 if ((plen < KEYLENGTH) && (key << plen)) {
1115 NL_SET_ERR_MSG(extack,
1116 "Invalid prefix for given prefix length");
1123 /* Caller must hold RTNL. */
1124 int fib_table_insert(struct net *net, struct fib_table *tb,
1125 struct fib_config *cfg, struct netlink_ext_ack *extack)
1127 enum fib_event_type event = FIB_EVENT_ENTRY_ADD;
1128 struct trie *t = (struct trie *)tb->tb_data;
1129 struct fib_alias *fa, *new_fa;
1130 struct key_vector *l, *tp;
1131 u16 nlflags = NLM_F_EXCL;
1132 struct fib_info *fi;
1133 u8 plen = cfg->fc_dst_len;
1134 u8 slen = KEYLENGTH - plen;
1135 u8 tos = cfg->fc_tos;
1139 key = ntohl(cfg->fc_dst);
1141 if (!fib_valid_key_len(key, plen, extack))
1144 pr_debug("Insert table=%u %08x/%d\n", tb->tb_id, key, plen);
1146 fi = fib_create_info(cfg, extack);
1152 l = fib_find_node(t, &tp, key);
1153 fa = l ? fib_find_alias(&l->leaf, slen, tos, fi->fib_priority,
1156 /* Now fa, if non-NULL, points to the first fib alias
1157 * with the same keys [prefix,tos,priority], if such key already
1158 * exists or to the node before which we will insert new one.
1160 * If fa is NULL, we will need to allocate a new one and
1161 * insert to the tail of the section matching the suffix length
1165 if (fa && fa->fa_tos == tos &&
1166 fa->fa_info->fib_priority == fi->fib_priority) {
1167 struct fib_alias *fa_first, *fa_match;
1170 if (cfg->fc_nlflags & NLM_F_EXCL)
1173 nlflags &= ~NLM_F_EXCL;
1176 * 1. Find exact match for type, scope, fib_info to avoid
1178 * 2. Find next 'fa' (or head), NLM_F_APPEND inserts before it
1182 hlist_for_each_entry_from(fa, fa_list) {
1183 if ((fa->fa_slen != slen) ||
1184 (fa->tb_id != tb->tb_id) ||
1185 (fa->fa_tos != tos))
1187 if (fa->fa_info->fib_priority != fi->fib_priority)
1189 if (fa->fa_type == cfg->fc_type &&
1190 fa->fa_info == fi) {
1196 if (cfg->fc_nlflags & NLM_F_REPLACE) {
1197 struct fib_info *fi_drop;
1200 nlflags |= NLM_F_REPLACE;
1208 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1212 fi_drop = fa->fa_info;
1213 new_fa->fa_tos = fa->fa_tos;
1214 new_fa->fa_info = fi;
1215 new_fa->fa_type = cfg->fc_type;
1216 state = fa->fa_state;
1217 new_fa->fa_state = state & ~FA_S_ACCESSED;
1218 new_fa->fa_slen = fa->fa_slen;
1219 new_fa->tb_id = tb->tb_id;
1220 new_fa->fa_default = -1;
1222 err = call_fib_entry_notifiers(net,
1223 FIB_EVENT_ENTRY_REPLACE,
1227 goto out_free_new_fa;
1229 rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen,
1230 tb->tb_id, &cfg->fc_nlinfo, nlflags);
1232 hlist_replace_rcu(&fa->fa_list, &new_fa->fa_list);
1234 alias_free_mem_rcu(fa);
1236 fib_release_info(fi_drop);
1237 if (state & FA_S_ACCESSED)
1238 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1242 /* Error if we find a perfect match which
1243 * uses the same scope, type, and nexthop
1249 if (cfg->fc_nlflags & NLM_F_APPEND) {
1250 event = FIB_EVENT_ENTRY_APPEND;
1251 nlflags |= NLM_F_APPEND;
1257 if (!(cfg->fc_nlflags & NLM_F_CREATE))
1260 nlflags |= NLM_F_CREATE;
1262 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1266 new_fa->fa_info = fi;
1267 new_fa->fa_tos = tos;
1268 new_fa->fa_type = cfg->fc_type;
1269 new_fa->fa_state = 0;
1270 new_fa->fa_slen = slen;
1271 new_fa->tb_id = tb->tb_id;
1272 new_fa->fa_default = -1;
1274 err = call_fib_entry_notifiers(net, event, key, plen, new_fa, extack);
1276 goto out_free_new_fa;
1278 /* Insert new entry to the list. */
1279 err = fib_insert_alias(t, tp, l, new_fa, fa, key);
1284 tb->tb_num_default++;
1286 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1287 rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen, new_fa->tb_id,
1288 &cfg->fc_nlinfo, nlflags);
1293 /* notifier was sent that entry would be added to trie, but
1294 * the add failed and need to recover. Only failure for
1295 * fib_insert_alias is ENOMEM.
1297 NL_SET_ERR_MSG(extack, "Failed to insert route into trie");
1298 call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, key,
1299 plen, new_fa, NULL);
1301 kmem_cache_free(fn_alias_kmem, new_fa);
1303 fib_release_info(fi);
1308 static inline t_key prefix_mismatch(t_key key, struct key_vector *n)
1310 t_key prefix = n->key;
1312 return (key ^ prefix) & (prefix | -prefix);
1315 /* should be called with rcu_read_lock */
1316 int fib_table_lookup(struct fib_table *tb, const struct flowi4 *flp,
1317 struct fib_result *res, int fib_flags)
1319 struct trie *t = (struct trie *) tb->tb_data;
1320 #ifdef CONFIG_IP_FIB_TRIE_STATS
1321 struct trie_use_stats __percpu *stats = t->stats;
1323 const t_key key = ntohl(flp->daddr);
1324 struct key_vector *n, *pn;
1325 struct fib_alias *fa;
1326 unsigned long index;
1332 n = get_child_rcu(pn, cindex);
1334 trace_fib_table_lookup(tb->tb_id, flp, NULL, -EAGAIN);
1338 #ifdef CONFIG_IP_FIB_TRIE_STATS
1339 this_cpu_inc(stats->gets);
1342 /* Step 1: Travel to the longest prefix match in the trie */
1344 index = get_cindex(key, n);
1346 /* This bit of code is a bit tricky but it combines multiple
1347 * checks into a single check. The prefix consists of the
1348 * prefix plus zeros for the "bits" in the prefix. The index
1349 * is the difference between the key and this value. From
1350 * this we can actually derive several pieces of data.
1351 * if (index >= (1ul << bits))
1352 * we have a mismatch in skip bits and failed
1354 * we know the value is cindex
1356 * This check is safe even if bits == KEYLENGTH due to the
1357 * fact that we can only allocate a node with 32 bits if a
1358 * long is greater than 32 bits.
1360 if (index >= (1ul << n->bits))
1363 /* we have found a leaf. Prefixes have already been compared */
1367 /* only record pn and cindex if we are going to be chopping
1368 * bits later. Otherwise we are just wasting cycles.
1370 if (n->slen > n->pos) {
1375 n = get_child_rcu(n, index);
1380 /* Step 2: Sort out leaves and begin backtracing for longest prefix */
1382 /* record the pointer where our next node pointer is stored */
1383 struct key_vector __rcu **cptr = n->tnode;
1385 /* This test verifies that none of the bits that differ
1386 * between the key and the prefix exist in the region of
1387 * the lsb and higher in the prefix.
1389 if (unlikely(prefix_mismatch(key, n)) || (n->slen == n->pos))
1392 /* exit out and process leaf */
1393 if (unlikely(IS_LEAF(n)))
1396 /* Don't bother recording parent info. Since we are in
1397 * prefix match mode we will have to come back to wherever
1398 * we started this traversal anyway
1401 while ((n = rcu_dereference(*cptr)) == NULL) {
1403 #ifdef CONFIG_IP_FIB_TRIE_STATS
1405 this_cpu_inc(stats->null_node_hit);
1407 /* If we are at cindex 0 there are no more bits for
1408 * us to strip at this level so we must ascend back
1409 * up one level to see if there are any more bits to
1410 * be stripped there.
1413 t_key pkey = pn->key;
1415 /* If we don't have a parent then there is
1416 * nothing for us to do as we do not have any
1417 * further nodes to parse.
1420 trace_fib_table_lookup(tb->tb_id, flp,
1424 #ifdef CONFIG_IP_FIB_TRIE_STATS
1425 this_cpu_inc(stats->backtrack);
1427 /* Get Child's index */
1428 pn = node_parent_rcu(pn);
1429 cindex = get_index(pkey, pn);
1432 /* strip the least significant bit from the cindex */
1433 cindex &= cindex - 1;
1435 /* grab pointer for next child node */
1436 cptr = &pn->tnode[cindex];
1441 /* this line carries forward the xor from earlier in the function */
1442 index = key ^ n->key;
1444 /* Step 3: Process the leaf, if that fails fall back to backtracing */
1445 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list) {
1446 struct fib_info *fi = fa->fa_info;
1449 if ((BITS_PER_LONG > KEYLENGTH) || (fa->fa_slen < KEYLENGTH)) {
1450 if (index >= (1ul << fa->fa_slen))
1453 if (fa->fa_tos && fa->fa_tos != flp->flowi4_tos)
1457 if (fa->fa_info->fib_scope < flp->flowi4_scope)
1459 fib_alias_accessed(fa);
1460 err = fib_props[fa->fa_type].error;
1461 if (unlikely(err < 0)) {
1462 #ifdef CONFIG_IP_FIB_TRIE_STATS
1463 this_cpu_inc(stats->semantic_match_passed);
1465 trace_fib_table_lookup(tb->tb_id, flp, NULL, err);
1468 if (fi->fib_flags & RTNH_F_DEAD)
1470 for (nhsel = 0; nhsel < fi->fib_nhs; nhsel++) {
1471 const struct fib_nh *nh = &fi->fib_nh[nhsel];
1472 struct in_device *in_dev = __in_dev_get_rcu(nh->nh_dev);
1474 if (nh->nh_flags & RTNH_F_DEAD)
1477 IN_DEV_IGNORE_ROUTES_WITH_LINKDOWN(in_dev) &&
1478 nh->nh_flags & RTNH_F_LINKDOWN &&
1479 !(fib_flags & FIB_LOOKUP_IGNORE_LINKSTATE))
1481 if (!(flp->flowi4_flags & FLOWI_FLAG_SKIP_NH_OIF)) {
1482 if (flp->flowi4_oif &&
1483 flp->flowi4_oif != nh->nh_oif)
1487 if (!(fib_flags & FIB_LOOKUP_NOREF))
1488 refcount_inc(&fi->fib_clntref);
1490 res->prefix = htonl(n->key);
1491 res->prefixlen = KEYLENGTH - fa->fa_slen;
1492 res->nh_sel = nhsel;
1493 res->type = fa->fa_type;
1494 res->scope = fi->fib_scope;
1497 res->fa_head = &n->leaf;
1498 #ifdef CONFIG_IP_FIB_TRIE_STATS
1499 this_cpu_inc(stats->semantic_match_passed);
1501 trace_fib_table_lookup(tb->tb_id, flp, nh, err);
1506 #ifdef CONFIG_IP_FIB_TRIE_STATS
1507 this_cpu_inc(stats->semantic_match_miss);
1511 EXPORT_SYMBOL_GPL(fib_table_lookup);
1513 static void fib_remove_alias(struct trie *t, struct key_vector *tp,
1514 struct key_vector *l, struct fib_alias *old)
1516 /* record the location of the previous list_info entry */
1517 struct hlist_node **pprev = old->fa_list.pprev;
1518 struct fib_alias *fa = hlist_entry(pprev, typeof(*fa), fa_list.next);
1520 /* remove the fib_alias from the list */
1521 hlist_del_rcu(&old->fa_list);
1523 /* if we emptied the list this leaf will be freed and we can sort
1524 * out parent suffix lengths as a part of trie_rebalance
1526 if (hlist_empty(&l->leaf)) {
1527 if (tp->slen == l->slen)
1528 node_pull_suffix(tp, tp->pos);
1529 put_child_root(tp, l->key, NULL);
1531 trie_rebalance(t, tp);
1535 /* only access fa if it is pointing at the last valid hlist_node */
1539 /* update the trie with the latest suffix length */
1540 l->slen = fa->fa_slen;
1541 node_pull_suffix(tp, fa->fa_slen);
1544 /* Caller must hold RTNL. */
1545 int fib_table_delete(struct net *net, struct fib_table *tb,
1546 struct fib_config *cfg, struct netlink_ext_ack *extack)
1548 struct trie *t = (struct trie *) tb->tb_data;
1549 struct fib_alias *fa, *fa_to_delete;
1550 struct key_vector *l, *tp;
1551 u8 plen = cfg->fc_dst_len;
1552 u8 slen = KEYLENGTH - plen;
1553 u8 tos = cfg->fc_tos;
1556 key = ntohl(cfg->fc_dst);
1558 if (!fib_valid_key_len(key, plen, extack))
1561 l = fib_find_node(t, &tp, key);
1565 fa = fib_find_alias(&l->leaf, slen, tos, 0, tb->tb_id);
1569 pr_debug("Deleting %08x/%d tos=%d t=%p\n", key, plen, tos, t);
1571 fa_to_delete = NULL;
1572 hlist_for_each_entry_from(fa, fa_list) {
1573 struct fib_info *fi = fa->fa_info;
1575 if ((fa->fa_slen != slen) ||
1576 (fa->tb_id != tb->tb_id) ||
1577 (fa->fa_tos != tos))
1580 if ((!cfg->fc_type || fa->fa_type == cfg->fc_type) &&
1581 (cfg->fc_scope == RT_SCOPE_NOWHERE ||
1582 fa->fa_info->fib_scope == cfg->fc_scope) &&
1583 (!cfg->fc_prefsrc ||
1584 fi->fib_prefsrc == cfg->fc_prefsrc) &&
1585 (!cfg->fc_protocol ||
1586 fi->fib_protocol == cfg->fc_protocol) &&
1587 fib_nh_match(cfg, fi, extack) == 0 &&
1588 fib_metrics_match(cfg, fi)) {
1597 call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, key, plen,
1598 fa_to_delete, extack);
1599 rtmsg_fib(RTM_DELROUTE, htonl(key), fa_to_delete, plen, tb->tb_id,
1600 &cfg->fc_nlinfo, 0);
1603 tb->tb_num_default--;
1605 fib_remove_alias(t, tp, l, fa_to_delete);
1607 if (fa_to_delete->fa_state & FA_S_ACCESSED)
1608 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1610 fib_release_info(fa_to_delete->fa_info);
1611 alias_free_mem_rcu(fa_to_delete);
1615 /* Scan for the next leaf starting at the provided key value */
1616 static struct key_vector *leaf_walk_rcu(struct key_vector **tn, t_key key)
1618 struct key_vector *pn, *n = *tn;
1619 unsigned long cindex;
1621 /* this loop is meant to try and find the key in the trie */
1623 /* record parent and next child index */
1625 cindex = (key > pn->key) ? get_index(key, pn) : 0;
1627 if (cindex >> pn->bits)
1630 /* descend into the next child */
1631 n = get_child_rcu(pn, cindex++);
1635 /* guarantee forward progress on the keys */
1636 if (IS_LEAF(n) && (n->key >= key))
1638 } while (IS_TNODE(n));
1640 /* this loop will search for the next leaf with a greater key */
1641 while (!IS_TRIE(pn)) {
1642 /* if we exhausted the parent node we will need to climb */
1643 if (cindex >= (1ul << pn->bits)) {
1644 t_key pkey = pn->key;
1646 pn = node_parent_rcu(pn);
1647 cindex = get_index(pkey, pn) + 1;
1651 /* grab the next available node */
1652 n = get_child_rcu(pn, cindex++);
1656 /* no need to compare keys since we bumped the index */
1660 /* Rescan start scanning in new node */
1666 return NULL; /* Root of trie */
1668 /* if we are at the limit for keys just return NULL for the tnode */
1673 static void fib_trie_free(struct fib_table *tb)
1675 struct trie *t = (struct trie *)tb->tb_data;
1676 struct key_vector *pn = t->kv;
1677 unsigned long cindex = 1;
1678 struct hlist_node *tmp;
1679 struct fib_alias *fa;
1681 /* walk trie in reverse order and free everything */
1683 struct key_vector *n;
1686 t_key pkey = pn->key;
1692 pn = node_parent(pn);
1694 /* drop emptied tnode */
1695 put_child_root(pn, n->key, NULL);
1698 cindex = get_index(pkey, pn);
1703 /* grab the next available node */
1704 n = get_child(pn, cindex);
1709 /* record pn and cindex for leaf walking */
1711 cindex = 1ul << n->bits;
1716 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1717 hlist_del_rcu(&fa->fa_list);
1718 alias_free_mem_rcu(fa);
1721 put_child_root(pn, n->key, NULL);
1725 #ifdef CONFIG_IP_FIB_TRIE_STATS
1726 free_percpu(t->stats);
1731 struct fib_table *fib_trie_unmerge(struct fib_table *oldtb)
1733 struct trie *ot = (struct trie *)oldtb->tb_data;
1734 struct key_vector *l, *tp = ot->kv;
1735 struct fib_table *local_tb;
1736 struct fib_alias *fa;
1740 if (oldtb->tb_data == oldtb->__data)
1743 local_tb = fib_trie_table(RT_TABLE_LOCAL, NULL);
1747 lt = (struct trie *)local_tb->tb_data;
1749 while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
1750 struct key_vector *local_l = NULL, *local_tp;
1752 hlist_for_each_entry(fa, &l->leaf, fa_list) {
1753 struct fib_alias *new_fa;
1755 if (local_tb->tb_id != fa->tb_id)
1758 /* clone fa for new local table */
1759 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1763 memcpy(new_fa, fa, sizeof(*fa));
1765 /* insert clone into table */
1767 local_l = fib_find_node(lt, &local_tp, l->key);
1769 if (fib_insert_alias(lt, local_tp, local_l, new_fa,
1771 kmem_cache_free(fn_alias_kmem, new_fa);
1776 /* stop loop if key wrapped back to 0 */
1784 fib_trie_free(local_tb);
1789 /* Caller must hold RTNL */
1790 void fib_table_flush_external(struct fib_table *tb)
1792 struct trie *t = (struct trie *)tb->tb_data;
1793 struct key_vector *pn = t->kv;
1794 unsigned long cindex = 1;
1795 struct hlist_node *tmp;
1796 struct fib_alias *fa;
1798 /* walk trie in reverse order */
1800 unsigned char slen = 0;
1801 struct key_vector *n;
1804 t_key pkey = pn->key;
1806 /* cannot resize the trie vector */
1810 /* update the suffix to address pulled leaves */
1811 if (pn->slen > pn->pos)
1814 /* resize completed node */
1816 cindex = get_index(pkey, pn);
1821 /* grab the next available node */
1822 n = get_child(pn, cindex);
1827 /* record pn and cindex for leaf walking */
1829 cindex = 1ul << n->bits;
1834 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1835 /* if alias was cloned to local then we just
1836 * need to remove the local copy from main
1838 if (tb->tb_id != fa->tb_id) {
1839 hlist_del_rcu(&fa->fa_list);
1840 alias_free_mem_rcu(fa);
1844 /* record local slen */
1848 /* update leaf slen */
1851 if (hlist_empty(&n->leaf)) {
1852 put_child_root(pn, n->key, NULL);
1858 /* Caller must hold RTNL. */
1859 int fib_table_flush(struct net *net, struct fib_table *tb, bool flush_all)
1861 struct trie *t = (struct trie *)tb->tb_data;
1862 struct key_vector *pn = t->kv;
1863 unsigned long cindex = 1;
1864 struct hlist_node *tmp;
1865 struct fib_alias *fa;
1868 /* walk trie in reverse order */
1870 unsigned char slen = 0;
1871 struct key_vector *n;
1874 t_key pkey = pn->key;
1876 /* cannot resize the trie vector */
1880 /* update the suffix to address pulled leaves */
1881 if (pn->slen > pn->pos)
1884 /* resize completed node */
1886 cindex = get_index(pkey, pn);
1891 /* grab the next available node */
1892 n = get_child(pn, cindex);
1897 /* record pn and cindex for leaf walking */
1899 cindex = 1ul << n->bits;
1904 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1905 struct fib_info *fi = fa->fa_info;
1907 if (!fi || tb->tb_id != fa->tb_id ||
1908 (!(fi->fib_flags & RTNH_F_DEAD) &&
1909 !fib_props[fa->fa_type].error)) {
1914 /* Do not flush error routes if network namespace is
1915 * not being dismantled
1917 if (!flush_all && fib_props[fa->fa_type].error) {
1922 call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_DEL,
1924 KEYLENGTH - fa->fa_slen, fa,
1926 hlist_del_rcu(&fa->fa_list);
1927 fib_release_info(fa->fa_info);
1928 alias_free_mem_rcu(fa);
1932 /* update leaf slen */
1935 if (hlist_empty(&n->leaf)) {
1936 put_child_root(pn, n->key, NULL);
1941 pr_debug("trie_flush found=%d\n", found);
1945 static void fib_leaf_notify(struct net *net, struct key_vector *l,
1946 struct fib_table *tb, struct notifier_block *nb)
1948 struct fib_alias *fa;
1950 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
1951 struct fib_info *fi = fa->fa_info;
1956 /* local and main table can share the same trie,
1957 * so don't notify twice for the same entry.
1959 if (tb->tb_id != fa->tb_id)
1962 call_fib_entry_notifier(nb, net, FIB_EVENT_ENTRY_ADD, l->key,
1963 KEYLENGTH - fa->fa_slen, fa);
1967 static void fib_table_notify(struct net *net, struct fib_table *tb,
1968 struct notifier_block *nb)
1970 struct trie *t = (struct trie *)tb->tb_data;
1971 struct key_vector *l, *tp = t->kv;
1974 while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
1975 fib_leaf_notify(net, l, tb, nb);
1978 /* stop in case of wrap around */
1984 void fib_notify(struct net *net, struct notifier_block *nb)
1988 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
1989 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
1990 struct fib_table *tb;
1992 hlist_for_each_entry_rcu(tb, head, tb_hlist)
1993 fib_table_notify(net, tb, nb);
1997 static void __trie_free_rcu(struct rcu_head *head)
1999 struct fib_table *tb = container_of(head, struct fib_table, rcu);
2000 #ifdef CONFIG_IP_FIB_TRIE_STATS
2001 struct trie *t = (struct trie *)tb->tb_data;
2003 if (tb->tb_data == tb->__data)
2004 free_percpu(t->stats);
2005 #endif /* CONFIG_IP_FIB_TRIE_STATS */
2009 void fib_free_table(struct fib_table *tb)
2011 call_rcu(&tb->rcu, __trie_free_rcu);
2014 static int fn_trie_dump_leaf(struct key_vector *l, struct fib_table *tb,
2015 struct sk_buff *skb, struct netlink_callback *cb)
2017 __be32 xkey = htonl(l->key);
2018 struct fib_alias *fa;
2024 /* rcu_read_lock is hold by caller */
2025 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
2033 if (tb->tb_id != fa->tb_id) {
2038 err = fib_dump_info(skb, NETLINK_CB(cb->skb).portid,
2039 cb->nlh->nlmsg_seq, RTM_NEWROUTE,
2040 tb->tb_id, fa->fa_type,
2041 xkey, KEYLENGTH - fa->fa_slen,
2042 fa->fa_tos, fa->fa_info, NLM_F_MULTI);
2054 /* rcu_read_lock needs to be hold by caller from readside */
2055 int fib_table_dump(struct fib_table *tb, struct sk_buff *skb,
2056 struct netlink_callback *cb)
2058 struct trie *t = (struct trie *)tb->tb_data;
2059 struct key_vector *l, *tp = t->kv;
2060 /* Dump starting at last key.
2061 * Note: 0.0.0.0/0 (ie default) is first key.
2063 int count = cb->args[2];
2064 t_key key = cb->args[3];
2066 while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
2069 err = fn_trie_dump_leaf(l, tb, skb, cb);
2072 cb->args[2] = count;
2079 memset(&cb->args[4], 0,
2080 sizeof(cb->args) - 4*sizeof(cb->args[0]));
2082 /* stop loop if key wrapped back to 0 */
2088 cb->args[2] = count;
2093 void __init fib_trie_init(void)
2095 fn_alias_kmem = kmem_cache_create("ip_fib_alias",
2096 sizeof(struct fib_alias),
2097 0, SLAB_PANIC, NULL);
2099 trie_leaf_kmem = kmem_cache_create("ip_fib_trie",
2101 0, SLAB_PANIC, NULL);
2104 struct fib_table *fib_trie_table(u32 id, struct fib_table *alias)
2106 struct fib_table *tb;
2108 size_t sz = sizeof(*tb);
2111 sz += sizeof(struct trie);
2113 tb = kzalloc(sz, GFP_KERNEL);
2118 tb->tb_num_default = 0;
2119 tb->tb_data = (alias ? alias->__data : tb->__data);
2124 t = (struct trie *) tb->tb_data;
2125 t->kv[0].pos = KEYLENGTH;
2126 t->kv[0].slen = KEYLENGTH;
2127 #ifdef CONFIG_IP_FIB_TRIE_STATS
2128 t->stats = alloc_percpu(struct trie_use_stats);
2138 #ifdef CONFIG_PROC_FS
2139 /* Depth first Trie walk iterator */
2140 struct fib_trie_iter {
2141 struct seq_net_private p;
2142 struct fib_table *tb;
2143 struct key_vector *tnode;
2148 static struct key_vector *fib_trie_get_next(struct fib_trie_iter *iter)
2150 unsigned long cindex = iter->index;
2151 struct key_vector *pn = iter->tnode;
2154 pr_debug("get_next iter={node=%p index=%d depth=%d}\n",
2155 iter->tnode, iter->index, iter->depth);
2157 while (!IS_TRIE(pn)) {
2158 while (cindex < child_length(pn)) {
2159 struct key_vector *n = get_child_rcu(pn, cindex++);
2166 iter->index = cindex;
2168 /* push down one level */
2177 /* Current node exhausted, pop back up */
2179 pn = node_parent_rcu(pn);
2180 cindex = get_index(pkey, pn) + 1;
2184 /* record root node so further searches know we are done */
2191 static struct key_vector *fib_trie_get_first(struct fib_trie_iter *iter,
2194 struct key_vector *n, *pn;
2200 n = rcu_dereference(pn->tnode[0]);
2217 static void trie_collect_stats(struct trie *t, struct trie_stat *s)
2219 struct key_vector *n;
2220 struct fib_trie_iter iter;
2222 memset(s, 0, sizeof(*s));
2225 for (n = fib_trie_get_first(&iter, t); n; n = fib_trie_get_next(&iter)) {
2227 struct fib_alias *fa;
2230 s->totdepth += iter.depth;
2231 if (iter.depth > s->maxdepth)
2232 s->maxdepth = iter.depth;
2234 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list)
2238 if (n->bits < MAX_STAT_DEPTH)
2239 s->nodesizes[n->bits]++;
2240 s->nullpointers += tn_info(n)->empty_children;
2247 * This outputs /proc/net/fib_triestats
2249 static void trie_show_stats(struct seq_file *seq, struct trie_stat *stat)
2251 unsigned int i, max, pointers, bytes, avdepth;
2254 avdepth = stat->totdepth*100 / stat->leaves;
2258 seq_printf(seq, "\tAver depth: %u.%02d\n",
2259 avdepth / 100, avdepth % 100);
2260 seq_printf(seq, "\tMax depth: %u\n", stat->maxdepth);
2262 seq_printf(seq, "\tLeaves: %u\n", stat->leaves);
2263 bytes = LEAF_SIZE * stat->leaves;
2265 seq_printf(seq, "\tPrefixes: %u\n", stat->prefixes);
2266 bytes += sizeof(struct fib_alias) * stat->prefixes;
2268 seq_printf(seq, "\tInternal nodes: %u\n\t", stat->tnodes);
2269 bytes += TNODE_SIZE(0) * stat->tnodes;
2271 max = MAX_STAT_DEPTH;
2272 while (max > 0 && stat->nodesizes[max-1] == 0)
2276 for (i = 1; i < max; i++)
2277 if (stat->nodesizes[i] != 0) {
2278 seq_printf(seq, " %u: %u", i, stat->nodesizes[i]);
2279 pointers += (1<<i) * stat->nodesizes[i];
2281 seq_putc(seq, '\n');
2282 seq_printf(seq, "\tPointers: %u\n", pointers);
2284 bytes += sizeof(struct key_vector *) * pointers;
2285 seq_printf(seq, "Null ptrs: %u\n", stat->nullpointers);
2286 seq_printf(seq, "Total size: %u kB\n", (bytes + 1023) / 1024);
2289 #ifdef CONFIG_IP_FIB_TRIE_STATS
2290 static void trie_show_usage(struct seq_file *seq,
2291 const struct trie_use_stats __percpu *stats)
2293 struct trie_use_stats s = { 0 };
2296 /* loop through all of the CPUs and gather up the stats */
2297 for_each_possible_cpu(cpu) {
2298 const struct trie_use_stats *pcpu = per_cpu_ptr(stats, cpu);
2300 s.gets += pcpu->gets;
2301 s.backtrack += pcpu->backtrack;
2302 s.semantic_match_passed += pcpu->semantic_match_passed;
2303 s.semantic_match_miss += pcpu->semantic_match_miss;
2304 s.null_node_hit += pcpu->null_node_hit;
2305 s.resize_node_skipped += pcpu->resize_node_skipped;
2308 seq_printf(seq, "\nCounters:\n---------\n");
2309 seq_printf(seq, "gets = %u\n", s.gets);
2310 seq_printf(seq, "backtracks = %u\n", s.backtrack);
2311 seq_printf(seq, "semantic match passed = %u\n",
2312 s.semantic_match_passed);
2313 seq_printf(seq, "semantic match miss = %u\n", s.semantic_match_miss);
2314 seq_printf(seq, "null node hit= %u\n", s.null_node_hit);
2315 seq_printf(seq, "skipped node resize = %u\n\n", s.resize_node_skipped);
2317 #endif /* CONFIG_IP_FIB_TRIE_STATS */
2319 static void fib_table_print(struct seq_file *seq, struct fib_table *tb)
2321 if (tb->tb_id == RT_TABLE_LOCAL)
2322 seq_puts(seq, "Local:\n");
2323 else if (tb->tb_id == RT_TABLE_MAIN)
2324 seq_puts(seq, "Main:\n");
2326 seq_printf(seq, "Id %d:\n", tb->tb_id);
2330 static int fib_triestat_seq_show(struct seq_file *seq, void *v)
2332 struct net *net = (struct net *)seq->private;
2336 "Basic info: size of leaf:"
2337 " %zd bytes, size of tnode: %zd bytes.\n",
2338 LEAF_SIZE, TNODE_SIZE(0));
2341 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2342 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2343 struct fib_table *tb;
2345 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2346 struct trie *t = (struct trie *) tb->tb_data;
2347 struct trie_stat stat;
2352 fib_table_print(seq, tb);
2354 trie_collect_stats(t, &stat);
2355 trie_show_stats(seq, &stat);
2356 #ifdef CONFIG_IP_FIB_TRIE_STATS
2357 trie_show_usage(seq, t->stats);
2367 static struct key_vector *fib_trie_get_idx(struct seq_file *seq, loff_t pos)
2369 struct fib_trie_iter *iter = seq->private;
2370 struct net *net = seq_file_net(seq);
2374 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2375 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2376 struct fib_table *tb;
2378 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2379 struct key_vector *n;
2381 for (n = fib_trie_get_first(iter,
2382 (struct trie *) tb->tb_data);
2383 n; n = fib_trie_get_next(iter))
2394 static void *fib_trie_seq_start(struct seq_file *seq, loff_t *pos)
2398 return fib_trie_get_idx(seq, *pos);
2401 static void *fib_trie_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2403 struct fib_trie_iter *iter = seq->private;
2404 struct net *net = seq_file_net(seq);
2405 struct fib_table *tb = iter->tb;
2406 struct hlist_node *tb_node;
2408 struct key_vector *n;
2411 /* next node in same table */
2412 n = fib_trie_get_next(iter);
2416 /* walk rest of this hash chain */
2417 h = tb->tb_id & (FIB_TABLE_HASHSZ - 1);
2418 while ((tb_node = rcu_dereference(hlist_next_rcu(&tb->tb_hlist)))) {
2419 tb = hlist_entry(tb_node, struct fib_table, tb_hlist);
2420 n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2425 /* new hash chain */
2426 while (++h < FIB_TABLE_HASHSZ) {
2427 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2428 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2429 n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2441 static void fib_trie_seq_stop(struct seq_file *seq, void *v)
2447 static void seq_indent(struct seq_file *seq, int n)
2453 static inline const char *rtn_scope(char *buf, size_t len, enum rt_scope_t s)
2456 case RT_SCOPE_UNIVERSE: return "universe";
2457 case RT_SCOPE_SITE: return "site";
2458 case RT_SCOPE_LINK: return "link";
2459 case RT_SCOPE_HOST: return "host";
2460 case RT_SCOPE_NOWHERE: return "nowhere";
2462 snprintf(buf, len, "scope=%d", s);
2467 static const char *const rtn_type_names[__RTN_MAX] = {
2468 [RTN_UNSPEC] = "UNSPEC",
2469 [RTN_UNICAST] = "UNICAST",
2470 [RTN_LOCAL] = "LOCAL",
2471 [RTN_BROADCAST] = "BROADCAST",
2472 [RTN_ANYCAST] = "ANYCAST",
2473 [RTN_MULTICAST] = "MULTICAST",
2474 [RTN_BLACKHOLE] = "BLACKHOLE",
2475 [RTN_UNREACHABLE] = "UNREACHABLE",
2476 [RTN_PROHIBIT] = "PROHIBIT",
2477 [RTN_THROW] = "THROW",
2479 [RTN_XRESOLVE] = "XRESOLVE",
2482 static inline const char *rtn_type(char *buf, size_t len, unsigned int t)
2484 if (t < __RTN_MAX && rtn_type_names[t])
2485 return rtn_type_names[t];
2486 snprintf(buf, len, "type %u", t);
2490 /* Pretty print the trie */
2491 static int fib_trie_seq_show(struct seq_file *seq, void *v)
2493 const struct fib_trie_iter *iter = seq->private;
2494 struct key_vector *n = v;
2496 if (IS_TRIE(node_parent_rcu(n)))
2497 fib_table_print(seq, iter->tb);
2500 __be32 prf = htonl(n->key);
2502 seq_indent(seq, iter->depth-1);
2503 seq_printf(seq, " +-- %pI4/%zu %u %u %u\n",
2504 &prf, KEYLENGTH - n->pos - n->bits, n->bits,
2505 tn_info(n)->full_children,
2506 tn_info(n)->empty_children);
2508 __be32 val = htonl(n->key);
2509 struct fib_alias *fa;
2511 seq_indent(seq, iter->depth);
2512 seq_printf(seq, " |-- %pI4\n", &val);
2514 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list) {
2515 char buf1[32], buf2[32];
2517 seq_indent(seq, iter->depth + 1);
2518 seq_printf(seq, " /%zu %s %s",
2519 KEYLENGTH - fa->fa_slen,
2520 rtn_scope(buf1, sizeof(buf1),
2521 fa->fa_info->fib_scope),
2522 rtn_type(buf2, sizeof(buf2),
2525 seq_printf(seq, " tos=%d", fa->fa_tos);
2526 seq_putc(seq, '\n');
2533 static const struct seq_operations fib_trie_seq_ops = {
2534 .start = fib_trie_seq_start,
2535 .next = fib_trie_seq_next,
2536 .stop = fib_trie_seq_stop,
2537 .show = fib_trie_seq_show,
2540 struct fib_route_iter {
2541 struct seq_net_private p;
2542 struct fib_table *main_tb;
2543 struct key_vector *tnode;
2548 static struct key_vector *fib_route_get_idx(struct fib_route_iter *iter,
2551 struct key_vector *l, **tp = &iter->tnode;
2554 /* use cached location of previously found key */
2555 if (iter->pos > 0 && pos >= iter->pos) {
2564 while ((l = leaf_walk_rcu(tp, key)) && (pos-- > 0)) {
2569 /* handle unlikely case of a key wrap */
2575 iter->key = l->key; /* remember it */
2577 iter->pos = 0; /* forget it */
2582 static void *fib_route_seq_start(struct seq_file *seq, loff_t *pos)
2585 struct fib_route_iter *iter = seq->private;
2586 struct fib_table *tb;
2591 tb = fib_get_table(seq_file_net(seq), RT_TABLE_MAIN);
2596 t = (struct trie *)tb->tb_data;
2597 iter->tnode = t->kv;
2600 return fib_route_get_idx(iter, *pos);
2603 iter->key = KEY_MAX;
2605 return SEQ_START_TOKEN;
2608 static void *fib_route_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2610 struct fib_route_iter *iter = seq->private;
2611 struct key_vector *l = NULL;
2612 t_key key = iter->key + 1;
2616 /* only allow key of 0 for start of sequence */
2617 if ((v == SEQ_START_TOKEN) || key)
2618 l = leaf_walk_rcu(&iter->tnode, key);
2630 static void fib_route_seq_stop(struct seq_file *seq, void *v)
2636 static unsigned int fib_flag_trans(int type, __be32 mask, const struct fib_info *fi)
2638 unsigned int flags = 0;
2640 if (type == RTN_UNREACHABLE || type == RTN_PROHIBIT)
2642 if (fi && fi->fib_nh->nh_gw)
2643 flags |= RTF_GATEWAY;
2644 if (mask == htonl(0xFFFFFFFF))
2651 * This outputs /proc/net/route.
2652 * The format of the file is not supposed to be changed
2653 * and needs to be same as fib_hash output to avoid breaking
2656 static int fib_route_seq_show(struct seq_file *seq, void *v)
2658 struct fib_route_iter *iter = seq->private;
2659 struct fib_table *tb = iter->main_tb;
2660 struct fib_alias *fa;
2661 struct key_vector *l = v;
2664 if (v == SEQ_START_TOKEN) {
2665 seq_printf(seq, "%-127s\n", "Iface\tDestination\tGateway "
2666 "\tFlags\tRefCnt\tUse\tMetric\tMask\t\tMTU"
2671 prefix = htonl(l->key);
2673 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
2674 const struct fib_info *fi = fa->fa_info;
2675 __be32 mask = inet_make_mask(KEYLENGTH - fa->fa_slen);
2676 unsigned int flags = fib_flag_trans(fa->fa_type, mask, fi);
2678 if ((fa->fa_type == RTN_BROADCAST) ||
2679 (fa->fa_type == RTN_MULTICAST))
2682 if (fa->tb_id != tb->tb_id)
2685 seq_setwidth(seq, 127);
2689 "%s\t%08X\t%08X\t%04X\t%d\t%u\t"
2690 "%d\t%08X\t%d\t%u\t%u",
2691 fi->fib_dev ? fi->fib_dev->name : "*",
2693 fi->fib_nh->nh_gw, flags, 0, 0,
2697 fi->fib_advmss + 40 : 0),
2702 "*\t%08X\t%08X\t%04X\t%d\t%u\t"
2703 "%d\t%08X\t%d\t%u\t%u",
2704 prefix, 0, flags, 0, 0, 0,
2713 static const struct seq_operations fib_route_seq_ops = {
2714 .start = fib_route_seq_start,
2715 .next = fib_route_seq_next,
2716 .stop = fib_route_seq_stop,
2717 .show = fib_route_seq_show,
2720 int __net_init fib_proc_init(struct net *net)
2722 if (!proc_create_net("fib_trie", 0444, net->proc_net, &fib_trie_seq_ops,
2723 sizeof(struct fib_trie_iter)))
2726 if (!proc_create_net_single("fib_triestat", 0444, net->proc_net,
2727 fib_triestat_seq_show, NULL))
2730 if (!proc_create_net("route", 0444, net->proc_net, &fib_route_seq_ops,
2731 sizeof(struct fib_route_iter)))
2737 remove_proc_entry("fib_triestat", net->proc_net);
2739 remove_proc_entry("fib_trie", net->proc_net);
2744 void __net_exit fib_proc_exit(struct net *net)
2746 remove_proc_entry("fib_trie", net->proc_net);
2747 remove_proc_entry("fib_triestat", net->proc_net);
2748 remove_proc_entry("route", net->proc_net);
2751 #endif /* CONFIG_PROC_FS */