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 <net/fib_notifier.h>
85 #include <trace/events/fib.h>
86 #include "fib_lookup.h"
88 static int call_fib_entry_notifier(struct notifier_block *nb, struct net *net,
89 enum fib_event_type event_type, u32 dst,
90 int dst_len, struct fib_info *fi,
91 u8 tos, u8 type, u32 tb_id)
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_info *fi,
107 u8 tos, u8 type, u32 tb_id)
109 struct fib_entry_notifier_info info = {
117 return call_fib4_notifiers(net, event_type, &info.info);
120 #define MAX_STAT_DEPTH 32
122 #define KEYLENGTH (8*sizeof(t_key))
123 #define KEY_MAX ((t_key)~0)
125 typedef unsigned int t_key;
127 #define IS_TRIE(n) ((n)->pos >= KEYLENGTH)
128 #define IS_TNODE(n) ((n)->bits)
129 #define IS_LEAF(n) (!(n)->bits)
133 unsigned char pos; /* 2log(KEYLENGTH) bits needed */
134 unsigned char bits; /* 2log(KEYLENGTH) bits needed */
137 /* This list pointer if valid if (pos | bits) == 0 (LEAF) */
138 struct hlist_head leaf;
139 /* This array is valid if (pos | bits) > 0 (TNODE) */
140 struct key_vector __rcu *tnode[0];
146 t_key empty_children; /* KEYLENGTH bits needed */
147 t_key full_children; /* KEYLENGTH bits needed */
148 struct key_vector __rcu *parent;
149 struct key_vector kv[1];
150 #define tn_bits kv[0].bits
153 #define TNODE_SIZE(n) offsetof(struct tnode, kv[0].tnode[n])
154 #define LEAF_SIZE TNODE_SIZE(1)
156 #ifdef CONFIG_IP_FIB_TRIE_STATS
157 struct trie_use_stats {
159 unsigned int backtrack;
160 unsigned int semantic_match_passed;
161 unsigned int semantic_match_miss;
162 unsigned int null_node_hit;
163 unsigned int resize_node_skipped;
168 unsigned int totdepth;
169 unsigned int maxdepth;
172 unsigned int nullpointers;
173 unsigned int prefixes;
174 unsigned int nodesizes[MAX_STAT_DEPTH];
178 struct key_vector kv[1];
179 #ifdef CONFIG_IP_FIB_TRIE_STATS
180 struct trie_use_stats __percpu *stats;
184 static struct key_vector *resize(struct trie *t, struct key_vector *tn);
185 static size_t tnode_free_size;
188 * synchronize_rcu after call_rcu for that many pages; it should be especially
189 * useful before resizing the root node with PREEMPT_NONE configs; the value was
190 * obtained experimentally, aiming to avoid visible slowdown.
192 static const int sync_pages = 128;
194 static struct kmem_cache *fn_alias_kmem __read_mostly;
195 static struct kmem_cache *trie_leaf_kmem __read_mostly;
197 static inline struct tnode *tn_info(struct key_vector *kv)
199 return container_of(kv, struct tnode, kv[0]);
202 /* caller must hold RTNL */
203 #define node_parent(tn) rtnl_dereference(tn_info(tn)->parent)
204 #define get_child(tn, i) rtnl_dereference((tn)->tnode[i])
206 /* caller must hold RCU read lock or RTNL */
207 #define node_parent_rcu(tn) rcu_dereference_rtnl(tn_info(tn)->parent)
208 #define get_child_rcu(tn, i) rcu_dereference_rtnl((tn)->tnode[i])
210 /* wrapper for rcu_assign_pointer */
211 static inline void node_set_parent(struct key_vector *n, struct key_vector *tp)
214 rcu_assign_pointer(tn_info(n)->parent, tp);
217 #define NODE_INIT_PARENT(n, p) RCU_INIT_POINTER(tn_info(n)->parent, p)
219 /* This provides us with the number of children in this node, in the case of a
220 * leaf this will return 0 meaning none of the children are accessible.
222 static inline unsigned long child_length(const struct key_vector *tn)
224 return (1ul << tn->bits) & ~(1ul);
227 #define get_cindex(key, kv) (((key) ^ (kv)->key) >> (kv)->pos)
229 static inline unsigned long get_index(t_key key, struct key_vector *kv)
231 unsigned long index = key ^ kv->key;
233 if ((BITS_PER_LONG <= KEYLENGTH) && (KEYLENGTH == kv->pos))
236 return index >> kv->pos;
239 /* To understand this stuff, an understanding of keys and all their bits is
240 * necessary. Every node in the trie has a key associated with it, but not
241 * all of the bits in that key are significant.
243 * Consider a node 'n' and its parent 'tp'.
245 * If n is a leaf, every bit in its key is significant. Its presence is
246 * necessitated by path compression, since during a tree traversal (when
247 * searching for a leaf - unless we are doing an insertion) we will completely
248 * ignore all skipped bits we encounter. Thus we need to verify, at the end of
249 * a potentially successful search, that we have indeed been walking the
252 * Note that we can never "miss" the correct key in the tree if present by
253 * following the wrong path. Path compression ensures that segments of the key
254 * that are the same for all keys with a given prefix are skipped, but the
255 * skipped part *is* identical for each node in the subtrie below the skipped
256 * bit! trie_insert() in this implementation takes care of that.
258 * if n is an internal node - a 'tnode' here, the various parts of its key
259 * have many different meanings.
262 * _________________________________________________________________
263 * | i | i | i | i | i | i | i | N | N | N | S | S | S | S | S | C |
264 * -----------------------------------------------------------------
265 * 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16
267 * _________________________________________________________________
268 * | C | C | C | u | u | u | u | u | u | u | u | u | u | u | u | u |
269 * -----------------------------------------------------------------
270 * 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
277 * First, let's just ignore the bits that come before the parent tp, that is
278 * the bits from (tp->pos + tp->bits) to 31. They are *known* but at this
279 * point we do not use them for anything.
281 * The bits from (tp->pos) to (tp->pos + tp->bits - 1) - "N", above - are the
282 * index into the parent's child array. That is, they will be used to find
283 * 'n' among tp's children.
285 * The bits from (n->pos + n->bits) to (tp->pos - 1) - "S" - are skipped bits
288 * All the bits we have seen so far are significant to the node n. The rest
289 * of the bits are really not needed or indeed known in n->key.
291 * The bits from (n->pos) to (n->pos + n->bits - 1) - "C" - are the index into
292 * n's child array, and will of course be different for each child.
294 * The rest of the bits, from 0 to (n->pos -1) - "u" - are completely unknown
298 static const int halve_threshold = 25;
299 static const int inflate_threshold = 50;
300 static const int halve_threshold_root = 15;
301 static const int inflate_threshold_root = 30;
303 static void __alias_free_mem(struct rcu_head *head)
305 struct fib_alias *fa = container_of(head, struct fib_alias, rcu);
306 kmem_cache_free(fn_alias_kmem, fa);
309 static inline void alias_free_mem_rcu(struct fib_alias *fa)
311 call_rcu(&fa->rcu, __alias_free_mem);
314 #define TNODE_KMALLOC_MAX \
315 ilog2((PAGE_SIZE - TNODE_SIZE(0)) / sizeof(struct key_vector *))
316 #define TNODE_VMALLOC_MAX \
317 ilog2((SIZE_MAX - TNODE_SIZE(0)) / sizeof(struct key_vector *))
319 static void __node_free_rcu(struct rcu_head *head)
321 struct tnode *n = container_of(head, struct tnode, rcu);
324 kmem_cache_free(trie_leaf_kmem, n);
329 #define node_free(n) call_rcu(&tn_info(n)->rcu, __node_free_rcu)
331 static struct tnode *tnode_alloc(int bits)
335 /* verify bits is within bounds */
336 if (bits > TNODE_VMALLOC_MAX)
339 /* determine size and verify it is non-zero and didn't overflow */
340 size = TNODE_SIZE(1ul << bits);
342 if (size <= PAGE_SIZE)
343 return kzalloc(size, GFP_KERNEL);
345 return vzalloc(size);
348 static inline void empty_child_inc(struct key_vector *n)
350 ++tn_info(n)->empty_children ? : ++tn_info(n)->full_children;
353 static inline void empty_child_dec(struct key_vector *n)
355 tn_info(n)->empty_children-- ? : tn_info(n)->full_children--;
358 static struct key_vector *leaf_new(t_key key, struct fib_alias *fa)
360 struct key_vector *l;
363 kv = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
367 /* initialize key vector */
372 l->slen = fa->fa_slen;
374 /* link leaf to fib alias */
375 INIT_HLIST_HEAD(&l->leaf);
376 hlist_add_head(&fa->fa_list, &l->leaf);
381 static struct key_vector *tnode_new(t_key key, int pos, int bits)
383 unsigned int shift = pos + bits;
384 struct key_vector *tn;
387 /* verify bits and pos their msb bits clear and values are valid */
388 BUG_ON(!bits || (shift > KEYLENGTH));
390 tnode = tnode_alloc(bits);
394 pr_debug("AT %p s=%zu %zu\n", tnode, TNODE_SIZE(0),
395 sizeof(struct key_vector *) << bits);
397 if (bits == KEYLENGTH)
398 tnode->full_children = 1;
400 tnode->empty_children = 1ul << bits;
403 tn->key = (shift < KEYLENGTH) ? (key >> shift) << shift : 0;
411 /* Check whether a tnode 'n' is "full", i.e. it is an internal node
412 * and no bits are skipped. See discussion in dyntree paper p. 6
414 static inline int tnode_full(struct key_vector *tn, struct key_vector *n)
416 return n && ((n->pos + n->bits) == tn->pos) && IS_TNODE(n);
419 /* Add a child at position i overwriting the old value.
420 * Update the value of full_children and empty_children.
422 static void put_child(struct key_vector *tn, unsigned long i,
423 struct key_vector *n)
425 struct key_vector *chi = get_child(tn, i);
428 BUG_ON(i >= child_length(tn));
430 /* update emptyChildren, overflow into fullChildren */
436 /* update fullChildren */
437 wasfull = tnode_full(tn, chi);
438 isfull = tnode_full(tn, n);
440 if (wasfull && !isfull)
441 tn_info(tn)->full_children--;
442 else if (!wasfull && isfull)
443 tn_info(tn)->full_children++;
445 if (n && (tn->slen < n->slen))
448 rcu_assign_pointer(tn->tnode[i], n);
451 static void update_children(struct key_vector *tn)
455 /* update all of the child parent pointers */
456 for (i = child_length(tn); i;) {
457 struct key_vector *inode = get_child(tn, --i);
462 /* Either update the children of a tnode that
463 * already belongs to us or update the child
464 * to point to ourselves.
466 if (node_parent(inode) == tn)
467 update_children(inode);
469 node_set_parent(inode, tn);
473 static inline void put_child_root(struct key_vector *tp, t_key key,
474 struct key_vector *n)
477 rcu_assign_pointer(tp->tnode[0], n);
479 put_child(tp, get_index(key, tp), n);
482 static inline void tnode_free_init(struct key_vector *tn)
484 tn_info(tn)->rcu.next = NULL;
487 static inline void tnode_free_append(struct key_vector *tn,
488 struct key_vector *n)
490 tn_info(n)->rcu.next = tn_info(tn)->rcu.next;
491 tn_info(tn)->rcu.next = &tn_info(n)->rcu;
494 static void tnode_free(struct key_vector *tn)
496 struct callback_head *head = &tn_info(tn)->rcu;
500 tnode_free_size += TNODE_SIZE(1ul << tn->bits);
503 tn = container_of(head, struct tnode, rcu)->kv;
506 if (tnode_free_size >= PAGE_SIZE * sync_pages) {
512 static struct key_vector *replace(struct trie *t,
513 struct key_vector *oldtnode,
514 struct key_vector *tn)
516 struct key_vector *tp = node_parent(oldtnode);
519 /* setup the parent pointer out of and back into this node */
520 NODE_INIT_PARENT(tn, tp);
521 put_child_root(tp, tn->key, tn);
523 /* update all of the child parent pointers */
526 /* all pointers should be clean so we are done */
527 tnode_free(oldtnode);
529 /* resize children now that oldtnode is freed */
530 for (i = child_length(tn); i;) {
531 struct key_vector *inode = get_child(tn, --i);
533 /* resize child node */
534 if (tnode_full(tn, inode))
535 tn = resize(t, inode);
541 static struct key_vector *inflate(struct trie *t,
542 struct key_vector *oldtnode)
544 struct key_vector *tn;
548 pr_debug("In inflate\n");
550 tn = tnode_new(oldtnode->key, oldtnode->pos - 1, oldtnode->bits + 1);
554 /* prepare oldtnode to be freed */
555 tnode_free_init(oldtnode);
557 /* Assemble all of the pointers in our cluster, in this case that
558 * represents all of the pointers out of our allocated nodes that
559 * point to existing tnodes and the links between our allocated
562 for (i = child_length(oldtnode), m = 1u << tn->pos; i;) {
563 struct key_vector *inode = get_child(oldtnode, --i);
564 struct key_vector *node0, *node1;
571 /* A leaf or an internal node with skipped bits */
572 if (!tnode_full(oldtnode, inode)) {
573 put_child(tn, get_index(inode->key, tn), inode);
577 /* drop the node in the old tnode free list */
578 tnode_free_append(oldtnode, inode);
580 /* An internal node with two children */
581 if (inode->bits == 1) {
582 put_child(tn, 2 * i + 1, get_child(inode, 1));
583 put_child(tn, 2 * i, get_child(inode, 0));
587 /* We will replace this node 'inode' with two new
588 * ones, 'node0' and 'node1', each with half of the
589 * original children. The two new nodes will have
590 * a position one bit further down the key and this
591 * means that the "significant" part of their keys
592 * (see the discussion near the top of this file)
593 * will differ by one bit, which will be "0" in
594 * node0's key and "1" in node1's key. Since we are
595 * moving the key position by one step, the bit that
596 * we are moving away from - the bit at position
597 * (tn->pos) - is the one that will differ between
598 * node0 and node1. So... we synthesize that bit in the
601 node1 = tnode_new(inode->key | m, inode->pos, inode->bits - 1);
604 node0 = tnode_new(inode->key, inode->pos, inode->bits - 1);
606 tnode_free_append(tn, node1);
609 tnode_free_append(tn, node0);
611 /* populate child pointers in new nodes */
612 for (k = child_length(inode), j = k / 2; j;) {
613 put_child(node1, --j, get_child(inode, --k));
614 put_child(node0, j, get_child(inode, j));
615 put_child(node1, --j, get_child(inode, --k));
616 put_child(node0, j, get_child(inode, j));
619 /* link new nodes to parent */
620 NODE_INIT_PARENT(node1, tn);
621 NODE_INIT_PARENT(node0, tn);
623 /* link parent to nodes */
624 put_child(tn, 2 * i + 1, node1);
625 put_child(tn, 2 * i, node0);
628 /* setup the parent pointers into and out of this node */
629 return replace(t, oldtnode, tn);
631 /* all pointers should be clean so we are done */
637 static struct key_vector *halve(struct trie *t,
638 struct key_vector *oldtnode)
640 struct key_vector *tn;
643 pr_debug("In halve\n");
645 tn = tnode_new(oldtnode->key, oldtnode->pos + 1, oldtnode->bits - 1);
649 /* prepare oldtnode to be freed */
650 tnode_free_init(oldtnode);
652 /* Assemble all of the pointers in our cluster, in this case that
653 * represents all of the pointers out of our allocated nodes that
654 * point to existing tnodes and the links between our allocated
657 for (i = child_length(oldtnode); i;) {
658 struct key_vector *node1 = get_child(oldtnode, --i);
659 struct key_vector *node0 = get_child(oldtnode, --i);
660 struct key_vector *inode;
662 /* At least one of the children is empty */
663 if (!node1 || !node0) {
664 put_child(tn, i / 2, node1 ? : node0);
668 /* Two nonempty children */
669 inode = tnode_new(node0->key, oldtnode->pos, 1);
672 tnode_free_append(tn, inode);
674 /* initialize pointers out of node */
675 put_child(inode, 1, node1);
676 put_child(inode, 0, node0);
677 NODE_INIT_PARENT(inode, tn);
679 /* link parent to node */
680 put_child(tn, i / 2, inode);
683 /* setup the parent pointers into and out of this node */
684 return replace(t, oldtnode, tn);
686 /* all pointers should be clean so we are done */
692 static struct key_vector *collapse(struct trie *t,
693 struct key_vector *oldtnode)
695 struct key_vector *n, *tp;
698 /* scan the tnode looking for that one child that might still exist */
699 for (n = NULL, i = child_length(oldtnode); !n && i;)
700 n = get_child(oldtnode, --i);
702 /* compress one level */
703 tp = node_parent(oldtnode);
704 put_child_root(tp, oldtnode->key, n);
705 node_set_parent(n, tp);
713 static unsigned char update_suffix(struct key_vector *tn)
715 unsigned char slen = tn->pos;
716 unsigned long stride, i;
717 unsigned char slen_max;
719 /* only vector 0 can have a suffix length greater than or equal to
720 * tn->pos + tn->bits, the second highest node will have a suffix
721 * length at most of tn->pos + tn->bits - 1
723 slen_max = min_t(unsigned char, tn->pos + tn->bits - 1, tn->slen);
725 /* search though the list of children looking for nodes that might
726 * have a suffix greater than the one we currently have. This is
727 * why we start with a stride of 2 since a stride of 1 would
728 * represent the nodes with suffix length equal to tn->pos
730 for (i = 0, stride = 0x2ul ; i < child_length(tn); i += stride) {
731 struct key_vector *n = get_child(tn, i);
733 if (!n || (n->slen <= slen))
736 /* update stride and slen based on new value */
737 stride <<= (n->slen - slen);
741 /* stop searching if we have hit the maximum possible value */
742 if (slen >= slen_max)
751 /* From "Implementing a dynamic compressed trie" by Stefan Nilsson of
752 * the Helsinki University of Technology and Matti Tikkanen of Nokia
753 * Telecommunications, page 6:
754 * "A node is doubled if the ratio of non-empty children to all
755 * children in the *doubled* node is at least 'high'."
757 * 'high' in this instance is the variable 'inflate_threshold'. It
758 * is expressed as a percentage, so we multiply it with
759 * child_length() and instead of multiplying by 2 (since the
760 * child array will be doubled by inflate()) and multiplying
761 * the left-hand side by 100 (to handle the percentage thing) we
762 * multiply the left-hand side by 50.
764 * The left-hand side may look a bit weird: child_length(tn)
765 * - tn->empty_children is of course the number of non-null children
766 * in the current node. tn->full_children is the number of "full"
767 * children, that is non-null tnodes with a skip value of 0.
768 * All of those will be doubled in the resulting inflated tnode, so
769 * we just count them one extra time here.
771 * A clearer way to write this would be:
773 * to_be_doubled = tn->full_children;
774 * not_to_be_doubled = child_length(tn) - tn->empty_children -
777 * new_child_length = child_length(tn) * 2;
779 * new_fill_factor = 100 * (not_to_be_doubled + 2*to_be_doubled) /
781 * if (new_fill_factor >= inflate_threshold)
783 * ...and so on, tho it would mess up the while () loop.
786 * 100 * (not_to_be_doubled + 2*to_be_doubled) / new_child_length >=
790 * 100 * (not_to_be_doubled + 2*to_be_doubled) >=
791 * inflate_threshold * new_child_length
793 * expand not_to_be_doubled and to_be_doubled, and shorten:
794 * 100 * (child_length(tn) - tn->empty_children +
795 * tn->full_children) >= inflate_threshold * new_child_length
797 * expand new_child_length:
798 * 100 * (child_length(tn) - tn->empty_children +
799 * tn->full_children) >=
800 * inflate_threshold * child_length(tn) * 2
803 * 50 * (tn->full_children + child_length(tn) -
804 * tn->empty_children) >= inflate_threshold *
808 static inline bool should_inflate(struct key_vector *tp, struct key_vector *tn)
810 unsigned long used = child_length(tn);
811 unsigned long threshold = used;
813 /* Keep root node larger */
814 threshold *= IS_TRIE(tp) ? inflate_threshold_root : inflate_threshold;
815 used -= tn_info(tn)->empty_children;
816 used += tn_info(tn)->full_children;
818 /* if bits == KEYLENGTH then pos = 0, and will fail below */
820 return (used > 1) && tn->pos && ((50 * used) >= threshold);
823 static inline bool should_halve(struct key_vector *tp, struct key_vector *tn)
825 unsigned long used = child_length(tn);
826 unsigned long threshold = used;
828 /* Keep root node larger */
829 threshold *= IS_TRIE(tp) ? halve_threshold_root : halve_threshold;
830 used -= tn_info(tn)->empty_children;
832 /* if bits == KEYLENGTH then used = 100% on wrap, and will fail below */
834 return (used > 1) && (tn->bits > 1) && ((100 * used) < threshold);
837 static inline bool should_collapse(struct key_vector *tn)
839 unsigned long used = child_length(tn);
841 used -= tn_info(tn)->empty_children;
843 /* account for bits == KEYLENGTH case */
844 if ((tn->bits == KEYLENGTH) && tn_info(tn)->full_children)
847 /* One child or none, time to drop us from the trie */
852 static struct key_vector *resize(struct trie *t, struct key_vector *tn)
854 #ifdef CONFIG_IP_FIB_TRIE_STATS
855 struct trie_use_stats __percpu *stats = t->stats;
857 struct key_vector *tp = node_parent(tn);
858 unsigned long cindex = get_index(tn->key, tp);
859 int max_work = MAX_WORK;
861 pr_debug("In tnode_resize %p inflate_threshold=%d threshold=%d\n",
862 tn, inflate_threshold, halve_threshold);
864 /* track the tnode via the pointer from the parent instead of
865 * doing it ourselves. This way we can let RCU fully do its
866 * thing without us interfering
868 BUG_ON(tn != get_child(tp, cindex));
870 /* Double as long as the resulting node has a number of
871 * nonempty nodes that are above the threshold.
873 while (should_inflate(tp, tn) && max_work) {
876 #ifdef CONFIG_IP_FIB_TRIE_STATS
877 this_cpu_inc(stats->resize_node_skipped);
883 tn = get_child(tp, cindex);
886 /* update parent in case inflate failed */
887 tp = node_parent(tn);
889 /* Return if at least one inflate is run */
890 if (max_work != MAX_WORK)
893 /* Halve as long as the number of empty children in this
894 * node is above threshold.
896 while (should_halve(tp, tn) && max_work) {
899 #ifdef CONFIG_IP_FIB_TRIE_STATS
900 this_cpu_inc(stats->resize_node_skipped);
906 tn = get_child(tp, cindex);
909 /* Only one child remains */
910 if (should_collapse(tn))
911 return collapse(t, tn);
913 /* update parent in case halve failed */
914 return node_parent(tn);
917 static void node_pull_suffix(struct key_vector *tn, unsigned char slen)
919 unsigned char node_slen = tn->slen;
921 while ((node_slen > tn->pos) && (node_slen > slen)) {
922 slen = update_suffix(tn);
923 if (node_slen == slen)
926 tn = node_parent(tn);
927 node_slen = tn->slen;
931 static void node_push_suffix(struct key_vector *tn, unsigned char slen)
933 while (tn->slen < slen) {
935 tn = node_parent(tn);
939 /* rcu_read_lock needs to be hold by caller from readside */
940 static struct key_vector *fib_find_node(struct trie *t,
941 struct key_vector **tp, u32 key)
943 struct key_vector *pn, *n = t->kv;
944 unsigned long index = 0;
948 n = get_child_rcu(n, index);
953 index = get_cindex(key, n);
955 /* This bit of code is a bit tricky but it combines multiple
956 * checks into a single check. The prefix consists of the
957 * prefix plus zeros for the bits in the cindex. The index
958 * is the difference between the key and this value. From
959 * this we can actually derive several pieces of data.
960 * if (index >= (1ul << bits))
961 * we have a mismatch in skip bits and failed
963 * we know the value is cindex
965 * This check is safe even if bits == KEYLENGTH due to the
966 * fact that we can only allocate a node with 32 bits if a
967 * long is greater than 32 bits.
969 if (index >= (1ul << n->bits)) {
974 /* keep searching until we find a perfect match leaf or NULL */
975 } while (IS_TNODE(n));
982 /* Return the first fib alias matching TOS with
983 * priority less than or equal to PRIO.
985 static struct fib_alias *fib_find_alias(struct hlist_head *fah, u8 slen,
986 u8 tos, u32 prio, u32 tb_id)
988 struct fib_alias *fa;
993 hlist_for_each_entry(fa, fah, fa_list) {
994 if (fa->fa_slen < slen)
996 if (fa->fa_slen != slen)
998 if (fa->tb_id > tb_id)
1000 if (fa->tb_id != tb_id)
1002 if (fa->fa_tos > tos)
1004 if (fa->fa_info->fib_priority >= prio || fa->fa_tos < tos)
1011 static void trie_rebalance(struct trie *t, struct key_vector *tn)
1013 while (!IS_TRIE(tn))
1017 static int fib_insert_node(struct trie *t, struct key_vector *tp,
1018 struct fib_alias *new, t_key key)
1020 struct key_vector *n, *l;
1022 l = leaf_new(key, new);
1026 /* retrieve child from parent node */
1027 n = get_child(tp, get_index(key, tp));
1029 /* Case 2: n is a LEAF or a TNODE and the key doesn't match.
1031 * Add a new tnode here
1032 * first tnode need some special handling
1033 * leaves us in position for handling as case 3
1036 struct key_vector *tn;
1038 tn = tnode_new(key, __fls(key ^ n->key), 1);
1042 /* initialize routes out of node */
1043 NODE_INIT_PARENT(tn, tp);
1044 put_child(tn, get_index(key, tn) ^ 1, n);
1046 /* start adding routes into the node */
1047 put_child_root(tp, key, tn);
1048 node_set_parent(n, tn);
1050 /* parent now has a NULL spot where the leaf can go */
1054 /* Case 3: n is NULL, and will just insert a new leaf */
1055 node_push_suffix(tp, new->fa_slen);
1056 NODE_INIT_PARENT(l, tp);
1057 put_child_root(tp, key, l);
1058 trie_rebalance(t, tp);
1067 static int fib_insert_alias(struct trie *t, struct key_vector *tp,
1068 struct key_vector *l, struct fib_alias *new,
1069 struct fib_alias *fa, t_key key)
1072 return fib_insert_node(t, tp, new, key);
1075 hlist_add_before_rcu(&new->fa_list, &fa->fa_list);
1077 struct fib_alias *last;
1079 hlist_for_each_entry(last, &l->leaf, fa_list) {
1080 if (new->fa_slen < last->fa_slen)
1082 if ((new->fa_slen == last->fa_slen) &&
1083 (new->tb_id > last->tb_id))
1089 hlist_add_behind_rcu(&new->fa_list, &fa->fa_list);
1091 hlist_add_head_rcu(&new->fa_list, &l->leaf);
1094 /* if we added to the tail node then we need to update slen */
1095 if (l->slen < new->fa_slen) {
1096 l->slen = new->fa_slen;
1097 node_push_suffix(tp, new->fa_slen);
1103 static bool fib_valid_key_len(u32 key, u8 plen, struct netlink_ext_ack *extack)
1105 if (plen > KEYLENGTH) {
1106 NL_SET_ERR_MSG(extack, "Invalid prefix length");
1110 if ((plen < KEYLENGTH) && (key << plen)) {
1111 NL_SET_ERR_MSG(extack,
1112 "Invalid prefix for given prefix length");
1119 /* Caller must hold RTNL. */
1120 int fib_table_insert(struct net *net, struct fib_table *tb,
1121 struct fib_config *cfg, struct netlink_ext_ack *extack)
1123 enum fib_event_type event = FIB_EVENT_ENTRY_ADD;
1124 struct trie *t = (struct trie *)tb->tb_data;
1125 struct fib_alias *fa, *new_fa;
1126 struct key_vector *l, *tp;
1127 u16 nlflags = NLM_F_EXCL;
1128 struct fib_info *fi;
1129 u8 plen = cfg->fc_dst_len;
1130 u8 slen = KEYLENGTH - plen;
1131 u8 tos = cfg->fc_tos;
1135 key = ntohl(cfg->fc_dst);
1137 if (!fib_valid_key_len(key, plen, extack))
1140 pr_debug("Insert table=%u %08x/%d\n", tb->tb_id, key, plen);
1142 fi = fib_create_info(cfg, extack);
1148 l = fib_find_node(t, &tp, key);
1149 fa = l ? fib_find_alias(&l->leaf, slen, tos, fi->fib_priority,
1152 /* Now fa, if non-NULL, points to the first fib alias
1153 * with the same keys [prefix,tos,priority], if such key already
1154 * exists or to the node before which we will insert new one.
1156 * If fa is NULL, we will need to allocate a new one and
1157 * insert to the tail of the section matching the suffix length
1161 if (fa && fa->fa_tos == tos &&
1162 fa->fa_info->fib_priority == fi->fib_priority) {
1163 struct fib_alias *fa_first, *fa_match;
1166 if (cfg->fc_nlflags & NLM_F_EXCL)
1169 nlflags &= ~NLM_F_EXCL;
1172 * 1. Find exact match for type, scope, fib_info to avoid
1174 * 2. Find next 'fa' (or head), NLM_F_APPEND inserts before it
1178 hlist_for_each_entry_from(fa, fa_list) {
1179 if ((fa->fa_slen != slen) ||
1180 (fa->tb_id != tb->tb_id) ||
1181 (fa->fa_tos != tos))
1183 if (fa->fa_info->fib_priority != fi->fib_priority)
1185 if (fa->fa_type == cfg->fc_type &&
1186 fa->fa_info == fi) {
1192 if (cfg->fc_nlflags & NLM_F_REPLACE) {
1193 struct fib_info *fi_drop;
1196 nlflags |= NLM_F_REPLACE;
1204 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1208 fi_drop = fa->fa_info;
1209 new_fa->fa_tos = fa->fa_tos;
1210 new_fa->fa_info = fi;
1211 new_fa->fa_type = cfg->fc_type;
1212 state = fa->fa_state;
1213 new_fa->fa_state = state & ~FA_S_ACCESSED;
1214 new_fa->fa_slen = fa->fa_slen;
1215 new_fa->tb_id = tb->tb_id;
1216 new_fa->fa_default = -1;
1218 call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_REPLACE,
1220 new_fa->fa_tos, cfg->fc_type,
1222 rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen,
1223 tb->tb_id, &cfg->fc_nlinfo, nlflags);
1225 hlist_replace_rcu(&fa->fa_list, &new_fa->fa_list);
1227 alias_free_mem_rcu(fa);
1229 fib_release_info(fi_drop);
1230 if (state & FA_S_ACCESSED)
1231 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1235 /* Error if we find a perfect match which
1236 * uses the same scope, type, and nexthop
1242 if (cfg->fc_nlflags & NLM_F_APPEND) {
1243 event = FIB_EVENT_ENTRY_APPEND;
1244 nlflags |= NLM_F_APPEND;
1250 if (!(cfg->fc_nlflags & NLM_F_CREATE))
1253 nlflags |= NLM_F_CREATE;
1255 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1259 new_fa->fa_info = fi;
1260 new_fa->fa_tos = tos;
1261 new_fa->fa_type = cfg->fc_type;
1262 new_fa->fa_state = 0;
1263 new_fa->fa_slen = slen;
1264 new_fa->tb_id = tb->tb_id;
1265 new_fa->fa_default = -1;
1267 /* Insert new entry to the list. */
1268 err = fib_insert_alias(t, tp, l, new_fa, fa, key);
1270 goto out_free_new_fa;
1273 tb->tb_num_default++;
1275 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1276 call_fib_entry_notifiers(net, event, key, plen, fi, tos, cfg->fc_type,
1278 rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen, new_fa->tb_id,
1279 &cfg->fc_nlinfo, nlflags);
1284 kmem_cache_free(fn_alias_kmem, new_fa);
1286 fib_release_info(fi);
1291 static inline t_key prefix_mismatch(t_key key, struct key_vector *n)
1293 t_key prefix = n->key;
1295 return (key ^ prefix) & (prefix | -prefix);
1298 /* should be called with rcu_read_lock */
1299 int fib_table_lookup(struct fib_table *tb, const struct flowi4 *flp,
1300 struct fib_result *res, int fib_flags)
1302 struct trie *t = (struct trie *) tb->tb_data;
1303 #ifdef CONFIG_IP_FIB_TRIE_STATS
1304 struct trie_use_stats __percpu *stats = t->stats;
1306 const t_key key = ntohl(flp->daddr);
1307 struct key_vector *n, *pn;
1308 struct fib_alias *fa;
1309 unsigned long index;
1312 trace_fib_table_lookup(tb->tb_id, flp);
1317 n = get_child_rcu(pn, cindex);
1321 #ifdef CONFIG_IP_FIB_TRIE_STATS
1322 this_cpu_inc(stats->gets);
1325 /* Step 1: Travel to the longest prefix match in the trie */
1327 index = get_cindex(key, n);
1329 /* This bit of code is a bit tricky but it combines multiple
1330 * checks into a single check. The prefix consists of the
1331 * prefix plus zeros for the "bits" in the prefix. The index
1332 * is the difference between the key and this value. From
1333 * this we can actually derive several pieces of data.
1334 * if (index >= (1ul << bits))
1335 * we have a mismatch in skip bits and failed
1337 * we know the value is cindex
1339 * This check is safe even if bits == KEYLENGTH due to the
1340 * fact that we can only allocate a node with 32 bits if a
1341 * long is greater than 32 bits.
1343 if (index >= (1ul << n->bits))
1346 /* we have found a leaf. Prefixes have already been compared */
1350 /* only record pn and cindex if we are going to be chopping
1351 * bits later. Otherwise we are just wasting cycles.
1353 if (n->slen > n->pos) {
1358 n = get_child_rcu(n, index);
1363 /* Step 2: Sort out leaves and begin backtracing for longest prefix */
1365 /* record the pointer where our next node pointer is stored */
1366 struct key_vector __rcu **cptr = n->tnode;
1368 /* This test verifies that none of the bits that differ
1369 * between the key and the prefix exist in the region of
1370 * the lsb and higher in the prefix.
1372 if (unlikely(prefix_mismatch(key, n)) || (n->slen == n->pos))
1375 /* exit out and process leaf */
1376 if (unlikely(IS_LEAF(n)))
1379 /* Don't bother recording parent info. Since we are in
1380 * prefix match mode we will have to come back to wherever
1381 * we started this traversal anyway
1384 while ((n = rcu_dereference(*cptr)) == NULL) {
1386 #ifdef CONFIG_IP_FIB_TRIE_STATS
1388 this_cpu_inc(stats->null_node_hit);
1390 /* If we are at cindex 0 there are no more bits for
1391 * us to strip at this level so we must ascend back
1392 * up one level to see if there are any more bits to
1393 * be stripped there.
1396 t_key pkey = pn->key;
1398 /* If we don't have a parent then there is
1399 * nothing for us to do as we do not have any
1400 * further nodes to parse.
1404 #ifdef CONFIG_IP_FIB_TRIE_STATS
1405 this_cpu_inc(stats->backtrack);
1407 /* Get Child's index */
1408 pn = node_parent_rcu(pn);
1409 cindex = get_index(pkey, pn);
1412 /* strip the least significant bit from the cindex */
1413 cindex &= cindex - 1;
1415 /* grab pointer for next child node */
1416 cptr = &pn->tnode[cindex];
1421 /* this line carries forward the xor from earlier in the function */
1422 index = key ^ n->key;
1424 /* Step 3: Process the leaf, if that fails fall back to backtracing */
1425 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list) {
1426 struct fib_info *fi = fa->fa_info;
1429 if ((BITS_PER_LONG > KEYLENGTH) || (fa->fa_slen < KEYLENGTH)) {
1430 if (index >= (1ul << fa->fa_slen))
1433 if (fa->fa_tos && fa->fa_tos != flp->flowi4_tos)
1437 if (fa->fa_info->fib_scope < flp->flowi4_scope)
1439 fib_alias_accessed(fa);
1440 err = fib_props[fa->fa_type].error;
1441 if (unlikely(err < 0)) {
1442 #ifdef CONFIG_IP_FIB_TRIE_STATS
1443 this_cpu_inc(stats->semantic_match_passed);
1447 if (fi->fib_flags & RTNH_F_DEAD)
1449 for (nhsel = 0; nhsel < fi->fib_nhs; nhsel++) {
1450 const struct fib_nh *nh = &fi->fib_nh[nhsel];
1451 struct in_device *in_dev = __in_dev_get_rcu(nh->nh_dev);
1453 if (nh->nh_flags & RTNH_F_DEAD)
1456 IN_DEV_IGNORE_ROUTES_WITH_LINKDOWN(in_dev) &&
1457 nh->nh_flags & RTNH_F_LINKDOWN &&
1458 !(fib_flags & FIB_LOOKUP_IGNORE_LINKSTATE))
1460 if (!(flp->flowi4_flags & FLOWI_FLAG_SKIP_NH_OIF)) {
1461 if (flp->flowi4_oif &&
1462 flp->flowi4_oif != nh->nh_oif)
1466 if (!(fib_flags & FIB_LOOKUP_NOREF))
1467 refcount_inc(&fi->fib_clntref);
1469 res->prefix = htonl(n->key);
1470 res->prefixlen = KEYLENGTH - fa->fa_slen;
1471 res->nh_sel = nhsel;
1472 res->type = fa->fa_type;
1473 res->scope = fi->fib_scope;
1476 res->fa_head = &n->leaf;
1477 #ifdef CONFIG_IP_FIB_TRIE_STATS
1478 this_cpu_inc(stats->semantic_match_passed);
1480 trace_fib_table_lookup_nh(nh);
1485 #ifdef CONFIG_IP_FIB_TRIE_STATS
1486 this_cpu_inc(stats->semantic_match_miss);
1490 EXPORT_SYMBOL_GPL(fib_table_lookup);
1492 static void fib_remove_alias(struct trie *t, struct key_vector *tp,
1493 struct key_vector *l, struct fib_alias *old)
1495 /* record the location of the previous list_info entry */
1496 struct hlist_node **pprev = old->fa_list.pprev;
1497 struct fib_alias *fa = hlist_entry(pprev, typeof(*fa), fa_list.next);
1499 /* remove the fib_alias from the list */
1500 hlist_del_rcu(&old->fa_list);
1502 /* if we emptied the list this leaf will be freed and we can sort
1503 * out parent suffix lengths as a part of trie_rebalance
1505 if (hlist_empty(&l->leaf)) {
1506 if (tp->slen == l->slen)
1507 node_pull_suffix(tp, tp->pos);
1508 put_child_root(tp, l->key, NULL);
1510 trie_rebalance(t, tp);
1514 /* only access fa if it is pointing at the last valid hlist_node */
1518 /* update the trie with the latest suffix length */
1519 l->slen = fa->fa_slen;
1520 node_pull_suffix(tp, fa->fa_slen);
1523 /* Caller must hold RTNL. */
1524 int fib_table_delete(struct net *net, struct fib_table *tb,
1525 struct fib_config *cfg, struct netlink_ext_ack *extack)
1527 struct trie *t = (struct trie *) tb->tb_data;
1528 struct fib_alias *fa, *fa_to_delete;
1529 struct key_vector *l, *tp;
1530 u8 plen = cfg->fc_dst_len;
1531 u8 slen = KEYLENGTH - plen;
1532 u8 tos = cfg->fc_tos;
1535 key = ntohl(cfg->fc_dst);
1537 if (!fib_valid_key_len(key, plen, extack))
1540 l = fib_find_node(t, &tp, key);
1544 fa = fib_find_alias(&l->leaf, slen, tos, 0, tb->tb_id);
1548 pr_debug("Deleting %08x/%d tos=%d t=%p\n", key, plen, tos, t);
1550 fa_to_delete = NULL;
1551 hlist_for_each_entry_from(fa, fa_list) {
1552 struct fib_info *fi = fa->fa_info;
1554 if ((fa->fa_slen != slen) ||
1555 (fa->tb_id != tb->tb_id) ||
1556 (fa->fa_tos != tos))
1559 if ((!cfg->fc_type || fa->fa_type == cfg->fc_type) &&
1560 (cfg->fc_scope == RT_SCOPE_NOWHERE ||
1561 fa->fa_info->fib_scope == cfg->fc_scope) &&
1562 (!cfg->fc_prefsrc ||
1563 fi->fib_prefsrc == cfg->fc_prefsrc) &&
1564 (!cfg->fc_protocol ||
1565 fi->fib_protocol == cfg->fc_protocol) &&
1566 fib_nh_match(cfg, fi, extack) == 0 &&
1567 fib_metrics_match(cfg, fi)) {
1576 call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, key, plen,
1577 fa_to_delete->fa_info, tos,
1578 fa_to_delete->fa_type, tb->tb_id);
1579 rtmsg_fib(RTM_DELROUTE, htonl(key), fa_to_delete, plen, tb->tb_id,
1580 &cfg->fc_nlinfo, 0);
1583 tb->tb_num_default--;
1585 fib_remove_alias(t, tp, l, fa_to_delete);
1587 if (fa_to_delete->fa_state & FA_S_ACCESSED)
1588 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1590 fib_release_info(fa_to_delete->fa_info);
1591 alias_free_mem_rcu(fa_to_delete);
1595 /* Scan for the next leaf starting at the provided key value */
1596 static struct key_vector *leaf_walk_rcu(struct key_vector **tn, t_key key)
1598 struct key_vector *pn, *n = *tn;
1599 unsigned long cindex;
1601 /* this loop is meant to try and find the key in the trie */
1603 /* record parent and next child index */
1605 cindex = (key > pn->key) ? get_index(key, pn) : 0;
1607 if (cindex >> pn->bits)
1610 /* descend into the next child */
1611 n = get_child_rcu(pn, cindex++);
1615 /* guarantee forward progress on the keys */
1616 if (IS_LEAF(n) && (n->key >= key))
1618 } while (IS_TNODE(n));
1620 /* this loop will search for the next leaf with a greater key */
1621 while (!IS_TRIE(pn)) {
1622 /* if we exhausted the parent node we will need to climb */
1623 if (cindex >= (1ul << pn->bits)) {
1624 t_key pkey = pn->key;
1626 pn = node_parent_rcu(pn);
1627 cindex = get_index(pkey, pn) + 1;
1631 /* grab the next available node */
1632 n = get_child_rcu(pn, cindex++);
1636 /* no need to compare keys since we bumped the index */
1640 /* Rescan start scanning in new node */
1646 return NULL; /* Root of trie */
1648 /* if we are at the limit for keys just return NULL for the tnode */
1653 static void fib_trie_free(struct fib_table *tb)
1655 struct trie *t = (struct trie *)tb->tb_data;
1656 struct key_vector *pn = t->kv;
1657 unsigned long cindex = 1;
1658 struct hlist_node *tmp;
1659 struct fib_alias *fa;
1661 /* walk trie in reverse order and free everything */
1663 struct key_vector *n;
1666 t_key pkey = pn->key;
1672 pn = node_parent(pn);
1674 /* drop emptied tnode */
1675 put_child_root(pn, n->key, NULL);
1678 cindex = get_index(pkey, pn);
1683 /* grab the next available node */
1684 n = get_child(pn, cindex);
1689 /* record pn and cindex for leaf walking */
1691 cindex = 1ul << n->bits;
1696 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1697 hlist_del_rcu(&fa->fa_list);
1698 alias_free_mem_rcu(fa);
1701 put_child_root(pn, n->key, NULL);
1705 #ifdef CONFIG_IP_FIB_TRIE_STATS
1706 free_percpu(t->stats);
1711 struct fib_table *fib_trie_unmerge(struct fib_table *oldtb)
1713 struct trie *ot = (struct trie *)oldtb->tb_data;
1714 struct key_vector *l, *tp = ot->kv;
1715 struct fib_table *local_tb;
1716 struct fib_alias *fa;
1720 if (oldtb->tb_data == oldtb->__data)
1723 local_tb = fib_trie_table(RT_TABLE_LOCAL, NULL);
1727 lt = (struct trie *)local_tb->tb_data;
1729 while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
1730 struct key_vector *local_l = NULL, *local_tp;
1732 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
1733 struct fib_alias *new_fa;
1735 if (local_tb->tb_id != fa->tb_id)
1738 /* clone fa for new local table */
1739 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1743 memcpy(new_fa, fa, sizeof(*fa));
1745 /* insert clone into table */
1747 local_l = fib_find_node(lt, &local_tp, l->key);
1749 if (fib_insert_alias(lt, local_tp, local_l, new_fa,
1751 kmem_cache_free(fn_alias_kmem, new_fa);
1756 /* stop loop if key wrapped back to 0 */
1764 fib_trie_free(local_tb);
1769 /* Caller must hold RTNL */
1770 void fib_table_flush_external(struct fib_table *tb)
1772 struct trie *t = (struct trie *)tb->tb_data;
1773 struct key_vector *pn = t->kv;
1774 unsigned long cindex = 1;
1775 struct hlist_node *tmp;
1776 struct fib_alias *fa;
1778 /* walk trie in reverse order */
1780 unsigned char slen = 0;
1781 struct key_vector *n;
1784 t_key pkey = pn->key;
1786 /* cannot resize the trie vector */
1790 /* update the suffix to address pulled leaves */
1791 if (pn->slen > pn->pos)
1794 /* resize completed node */
1796 cindex = get_index(pkey, pn);
1801 /* grab the next available node */
1802 n = get_child(pn, cindex);
1807 /* record pn and cindex for leaf walking */
1809 cindex = 1ul << n->bits;
1814 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1815 /* if alias was cloned to local then we just
1816 * need to remove the local copy from main
1818 if (tb->tb_id != fa->tb_id) {
1819 hlist_del_rcu(&fa->fa_list);
1820 alias_free_mem_rcu(fa);
1824 /* record local slen */
1828 /* update leaf slen */
1831 if (hlist_empty(&n->leaf)) {
1832 put_child_root(pn, n->key, NULL);
1838 /* Caller must hold RTNL. */
1839 int fib_table_flush(struct net *net, struct fib_table *tb)
1841 struct trie *t = (struct trie *)tb->tb_data;
1842 struct key_vector *pn = t->kv;
1843 unsigned long cindex = 1;
1844 struct hlist_node *tmp;
1845 struct fib_alias *fa;
1848 /* walk trie in reverse order */
1850 unsigned char slen = 0;
1851 struct key_vector *n;
1854 t_key pkey = pn->key;
1856 /* cannot resize the trie vector */
1860 /* update the suffix to address pulled leaves */
1861 if (pn->slen > pn->pos)
1864 /* resize completed node */
1866 cindex = get_index(pkey, pn);
1871 /* grab the next available node */
1872 n = get_child(pn, cindex);
1877 /* record pn and cindex for leaf walking */
1879 cindex = 1ul << n->bits;
1884 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1885 struct fib_info *fi = fa->fa_info;
1887 if (!fi || !(fi->fib_flags & RTNH_F_DEAD) ||
1888 tb->tb_id != fa->tb_id) {
1893 call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_DEL,
1895 KEYLENGTH - fa->fa_slen,
1896 fi, fa->fa_tos, fa->fa_type,
1898 hlist_del_rcu(&fa->fa_list);
1899 fib_release_info(fa->fa_info);
1900 alias_free_mem_rcu(fa);
1904 /* update leaf slen */
1907 if (hlist_empty(&n->leaf)) {
1908 put_child_root(pn, n->key, NULL);
1913 pr_debug("trie_flush found=%d\n", found);
1917 static void fib_leaf_notify(struct net *net, struct key_vector *l,
1918 struct fib_table *tb, struct notifier_block *nb)
1920 struct fib_alias *fa;
1922 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
1923 struct fib_info *fi = fa->fa_info;
1928 /* local and main table can share the same trie,
1929 * so don't notify twice for the same entry.
1931 if (tb->tb_id != fa->tb_id)
1934 call_fib_entry_notifier(nb, net, FIB_EVENT_ENTRY_ADD, l->key,
1935 KEYLENGTH - fa->fa_slen, fi, fa->fa_tos,
1936 fa->fa_type, fa->tb_id);
1940 static void fib_table_notify(struct net *net, struct fib_table *tb,
1941 struct notifier_block *nb)
1943 struct trie *t = (struct trie *)tb->tb_data;
1944 struct key_vector *l, *tp = t->kv;
1947 while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
1948 fib_leaf_notify(net, l, tb, nb);
1951 /* stop in case of wrap around */
1957 void fib_notify(struct net *net, struct notifier_block *nb)
1961 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
1962 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
1963 struct fib_table *tb;
1965 hlist_for_each_entry_rcu(tb, head, tb_hlist)
1966 fib_table_notify(net, tb, nb);
1970 static void __trie_free_rcu(struct rcu_head *head)
1972 struct fib_table *tb = container_of(head, struct fib_table, rcu);
1973 #ifdef CONFIG_IP_FIB_TRIE_STATS
1974 struct trie *t = (struct trie *)tb->tb_data;
1976 if (tb->tb_data == tb->__data)
1977 free_percpu(t->stats);
1978 #endif /* CONFIG_IP_FIB_TRIE_STATS */
1982 void fib_free_table(struct fib_table *tb)
1984 call_rcu(&tb->rcu, __trie_free_rcu);
1987 static int fn_trie_dump_leaf(struct key_vector *l, struct fib_table *tb,
1988 struct sk_buff *skb, struct netlink_callback *cb)
1990 __be32 xkey = htonl(l->key);
1991 struct fib_alias *fa;
1997 /* rcu_read_lock is hold by caller */
1998 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
2006 if (tb->tb_id != fa->tb_id) {
2011 err = fib_dump_info(skb, NETLINK_CB(cb->skb).portid,
2012 cb->nlh->nlmsg_seq, RTM_NEWROUTE,
2013 tb->tb_id, fa->fa_type,
2014 xkey, KEYLENGTH - fa->fa_slen,
2015 fa->fa_tos, fa->fa_info, NLM_F_MULTI);
2027 /* rcu_read_lock needs to be hold by caller from readside */
2028 int fib_table_dump(struct fib_table *tb, struct sk_buff *skb,
2029 struct netlink_callback *cb)
2031 struct trie *t = (struct trie *)tb->tb_data;
2032 struct key_vector *l, *tp = t->kv;
2033 /* Dump starting at last key.
2034 * Note: 0.0.0.0/0 (ie default) is first key.
2036 int count = cb->args[2];
2037 t_key key = cb->args[3];
2039 while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
2042 err = fn_trie_dump_leaf(l, tb, skb, cb);
2045 cb->args[2] = count;
2052 memset(&cb->args[4], 0,
2053 sizeof(cb->args) - 4*sizeof(cb->args[0]));
2055 /* stop loop if key wrapped back to 0 */
2061 cb->args[2] = count;
2066 void __init fib_trie_init(void)
2068 fn_alias_kmem = kmem_cache_create("ip_fib_alias",
2069 sizeof(struct fib_alias),
2070 0, SLAB_PANIC, NULL);
2072 trie_leaf_kmem = kmem_cache_create("ip_fib_trie",
2074 0, SLAB_PANIC, NULL);
2077 struct fib_table *fib_trie_table(u32 id, struct fib_table *alias)
2079 struct fib_table *tb;
2081 size_t sz = sizeof(*tb);
2084 sz += sizeof(struct trie);
2086 tb = kzalloc(sz, GFP_KERNEL);
2091 tb->tb_num_default = 0;
2092 tb->tb_data = (alias ? alias->__data : tb->__data);
2097 t = (struct trie *) tb->tb_data;
2098 t->kv[0].pos = KEYLENGTH;
2099 t->kv[0].slen = KEYLENGTH;
2100 #ifdef CONFIG_IP_FIB_TRIE_STATS
2101 t->stats = alloc_percpu(struct trie_use_stats);
2111 #ifdef CONFIG_PROC_FS
2112 /* Depth first Trie walk iterator */
2113 struct fib_trie_iter {
2114 struct seq_net_private p;
2115 struct fib_table *tb;
2116 struct key_vector *tnode;
2121 static struct key_vector *fib_trie_get_next(struct fib_trie_iter *iter)
2123 unsigned long cindex = iter->index;
2124 struct key_vector *pn = iter->tnode;
2127 pr_debug("get_next iter={node=%p index=%d depth=%d}\n",
2128 iter->tnode, iter->index, iter->depth);
2130 while (!IS_TRIE(pn)) {
2131 while (cindex < child_length(pn)) {
2132 struct key_vector *n = get_child_rcu(pn, cindex++);
2139 iter->index = cindex;
2141 /* push down one level */
2150 /* Current node exhausted, pop back up */
2152 pn = node_parent_rcu(pn);
2153 cindex = get_index(pkey, pn) + 1;
2157 /* record root node so further searches know we are done */
2164 static struct key_vector *fib_trie_get_first(struct fib_trie_iter *iter,
2167 struct key_vector *n, *pn;
2173 n = rcu_dereference(pn->tnode[0]);
2190 static void trie_collect_stats(struct trie *t, struct trie_stat *s)
2192 struct key_vector *n;
2193 struct fib_trie_iter iter;
2195 memset(s, 0, sizeof(*s));
2198 for (n = fib_trie_get_first(&iter, t); n; n = fib_trie_get_next(&iter)) {
2200 struct fib_alias *fa;
2203 s->totdepth += iter.depth;
2204 if (iter.depth > s->maxdepth)
2205 s->maxdepth = iter.depth;
2207 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list)
2211 if (n->bits < MAX_STAT_DEPTH)
2212 s->nodesizes[n->bits]++;
2213 s->nullpointers += tn_info(n)->empty_children;
2220 * This outputs /proc/net/fib_triestats
2222 static void trie_show_stats(struct seq_file *seq, struct trie_stat *stat)
2224 unsigned int i, max, pointers, bytes, avdepth;
2227 avdepth = stat->totdepth*100 / stat->leaves;
2231 seq_printf(seq, "\tAver depth: %u.%02d\n",
2232 avdepth / 100, avdepth % 100);
2233 seq_printf(seq, "\tMax depth: %u\n", stat->maxdepth);
2235 seq_printf(seq, "\tLeaves: %u\n", stat->leaves);
2236 bytes = LEAF_SIZE * stat->leaves;
2238 seq_printf(seq, "\tPrefixes: %u\n", stat->prefixes);
2239 bytes += sizeof(struct fib_alias) * stat->prefixes;
2241 seq_printf(seq, "\tInternal nodes: %u\n\t", stat->tnodes);
2242 bytes += TNODE_SIZE(0) * stat->tnodes;
2244 max = MAX_STAT_DEPTH;
2245 while (max > 0 && stat->nodesizes[max-1] == 0)
2249 for (i = 1; i < max; i++)
2250 if (stat->nodesizes[i] != 0) {
2251 seq_printf(seq, " %u: %u", i, stat->nodesizes[i]);
2252 pointers += (1<<i) * stat->nodesizes[i];
2254 seq_putc(seq, '\n');
2255 seq_printf(seq, "\tPointers: %u\n", pointers);
2257 bytes += sizeof(struct key_vector *) * pointers;
2258 seq_printf(seq, "Null ptrs: %u\n", stat->nullpointers);
2259 seq_printf(seq, "Total size: %u kB\n", (bytes + 1023) / 1024);
2262 #ifdef CONFIG_IP_FIB_TRIE_STATS
2263 static void trie_show_usage(struct seq_file *seq,
2264 const struct trie_use_stats __percpu *stats)
2266 struct trie_use_stats s = { 0 };
2269 /* loop through all of the CPUs and gather up the stats */
2270 for_each_possible_cpu(cpu) {
2271 const struct trie_use_stats *pcpu = per_cpu_ptr(stats, cpu);
2273 s.gets += pcpu->gets;
2274 s.backtrack += pcpu->backtrack;
2275 s.semantic_match_passed += pcpu->semantic_match_passed;
2276 s.semantic_match_miss += pcpu->semantic_match_miss;
2277 s.null_node_hit += pcpu->null_node_hit;
2278 s.resize_node_skipped += pcpu->resize_node_skipped;
2281 seq_printf(seq, "\nCounters:\n---------\n");
2282 seq_printf(seq, "gets = %u\n", s.gets);
2283 seq_printf(seq, "backtracks = %u\n", s.backtrack);
2284 seq_printf(seq, "semantic match passed = %u\n",
2285 s.semantic_match_passed);
2286 seq_printf(seq, "semantic match miss = %u\n", s.semantic_match_miss);
2287 seq_printf(seq, "null node hit= %u\n", s.null_node_hit);
2288 seq_printf(seq, "skipped node resize = %u\n\n", s.resize_node_skipped);
2290 #endif /* CONFIG_IP_FIB_TRIE_STATS */
2292 static void fib_table_print(struct seq_file *seq, struct fib_table *tb)
2294 if (tb->tb_id == RT_TABLE_LOCAL)
2295 seq_puts(seq, "Local:\n");
2296 else if (tb->tb_id == RT_TABLE_MAIN)
2297 seq_puts(seq, "Main:\n");
2299 seq_printf(seq, "Id %d:\n", tb->tb_id);
2303 static int fib_triestat_seq_show(struct seq_file *seq, void *v)
2305 struct net *net = (struct net *)seq->private;
2309 "Basic info: size of leaf:"
2310 " %zd bytes, size of tnode: %zd bytes.\n",
2311 LEAF_SIZE, TNODE_SIZE(0));
2313 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2314 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2315 struct fib_table *tb;
2317 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2318 struct trie *t = (struct trie *) tb->tb_data;
2319 struct trie_stat stat;
2324 fib_table_print(seq, tb);
2326 trie_collect_stats(t, &stat);
2327 trie_show_stats(seq, &stat);
2328 #ifdef CONFIG_IP_FIB_TRIE_STATS
2329 trie_show_usage(seq, t->stats);
2337 static int fib_triestat_seq_open(struct inode *inode, struct file *file)
2339 return single_open_net(inode, file, fib_triestat_seq_show);
2342 static const struct file_operations fib_triestat_fops = {
2343 .owner = THIS_MODULE,
2344 .open = fib_triestat_seq_open,
2346 .llseek = seq_lseek,
2347 .release = single_release_net,
2350 static struct key_vector *fib_trie_get_idx(struct seq_file *seq, loff_t pos)
2352 struct fib_trie_iter *iter = seq->private;
2353 struct net *net = seq_file_net(seq);
2357 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2358 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2359 struct fib_table *tb;
2361 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2362 struct key_vector *n;
2364 for (n = fib_trie_get_first(iter,
2365 (struct trie *) tb->tb_data);
2366 n; n = fib_trie_get_next(iter))
2377 static void *fib_trie_seq_start(struct seq_file *seq, loff_t *pos)
2381 return fib_trie_get_idx(seq, *pos);
2384 static void *fib_trie_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2386 struct fib_trie_iter *iter = seq->private;
2387 struct net *net = seq_file_net(seq);
2388 struct fib_table *tb = iter->tb;
2389 struct hlist_node *tb_node;
2391 struct key_vector *n;
2394 /* next node in same table */
2395 n = fib_trie_get_next(iter);
2399 /* walk rest of this hash chain */
2400 h = tb->tb_id & (FIB_TABLE_HASHSZ - 1);
2401 while ((tb_node = rcu_dereference(hlist_next_rcu(&tb->tb_hlist)))) {
2402 tb = hlist_entry(tb_node, struct fib_table, tb_hlist);
2403 n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2408 /* new hash chain */
2409 while (++h < FIB_TABLE_HASHSZ) {
2410 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2411 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2412 n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2424 static void fib_trie_seq_stop(struct seq_file *seq, void *v)
2430 static void seq_indent(struct seq_file *seq, int n)
2436 static inline const char *rtn_scope(char *buf, size_t len, enum rt_scope_t s)
2439 case RT_SCOPE_UNIVERSE: return "universe";
2440 case RT_SCOPE_SITE: return "site";
2441 case RT_SCOPE_LINK: return "link";
2442 case RT_SCOPE_HOST: return "host";
2443 case RT_SCOPE_NOWHERE: return "nowhere";
2445 snprintf(buf, len, "scope=%d", s);
2450 static const char *const rtn_type_names[__RTN_MAX] = {
2451 [RTN_UNSPEC] = "UNSPEC",
2452 [RTN_UNICAST] = "UNICAST",
2453 [RTN_LOCAL] = "LOCAL",
2454 [RTN_BROADCAST] = "BROADCAST",
2455 [RTN_ANYCAST] = "ANYCAST",
2456 [RTN_MULTICAST] = "MULTICAST",
2457 [RTN_BLACKHOLE] = "BLACKHOLE",
2458 [RTN_UNREACHABLE] = "UNREACHABLE",
2459 [RTN_PROHIBIT] = "PROHIBIT",
2460 [RTN_THROW] = "THROW",
2462 [RTN_XRESOLVE] = "XRESOLVE",
2465 static inline const char *rtn_type(char *buf, size_t len, unsigned int t)
2467 if (t < __RTN_MAX && rtn_type_names[t])
2468 return rtn_type_names[t];
2469 snprintf(buf, len, "type %u", t);
2473 /* Pretty print the trie */
2474 static int fib_trie_seq_show(struct seq_file *seq, void *v)
2476 const struct fib_trie_iter *iter = seq->private;
2477 struct key_vector *n = v;
2479 if (IS_TRIE(node_parent_rcu(n)))
2480 fib_table_print(seq, iter->tb);
2483 __be32 prf = htonl(n->key);
2485 seq_indent(seq, iter->depth-1);
2486 seq_printf(seq, " +-- %pI4/%zu %u %u %u\n",
2487 &prf, KEYLENGTH - n->pos - n->bits, n->bits,
2488 tn_info(n)->full_children,
2489 tn_info(n)->empty_children);
2491 __be32 val = htonl(n->key);
2492 struct fib_alias *fa;
2494 seq_indent(seq, iter->depth);
2495 seq_printf(seq, " |-- %pI4\n", &val);
2497 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list) {
2498 char buf1[32], buf2[32];
2500 seq_indent(seq, iter->depth + 1);
2501 seq_printf(seq, " /%zu %s %s",
2502 KEYLENGTH - fa->fa_slen,
2503 rtn_scope(buf1, sizeof(buf1),
2504 fa->fa_info->fib_scope),
2505 rtn_type(buf2, sizeof(buf2),
2508 seq_printf(seq, " tos=%d", fa->fa_tos);
2509 seq_putc(seq, '\n');
2516 static const struct seq_operations fib_trie_seq_ops = {
2517 .start = fib_trie_seq_start,
2518 .next = fib_trie_seq_next,
2519 .stop = fib_trie_seq_stop,
2520 .show = fib_trie_seq_show,
2523 static int fib_trie_seq_open(struct inode *inode, struct file *file)
2525 return seq_open_net(inode, file, &fib_trie_seq_ops,
2526 sizeof(struct fib_trie_iter));
2529 static const struct file_operations fib_trie_fops = {
2530 .owner = THIS_MODULE,
2531 .open = fib_trie_seq_open,
2533 .llseek = seq_lseek,
2534 .release = seq_release_net,
2537 struct fib_route_iter {
2538 struct seq_net_private p;
2539 struct fib_table *main_tb;
2540 struct key_vector *tnode;
2545 static struct key_vector *fib_route_get_idx(struct fib_route_iter *iter,
2548 struct key_vector *l, **tp = &iter->tnode;
2551 /* use cached location of previously found key */
2552 if (iter->pos > 0 && pos >= iter->pos) {
2561 while ((l = leaf_walk_rcu(tp, key)) && (pos-- > 0)) {
2566 /* handle unlikely case of a key wrap */
2572 iter->key = l->key; /* remember it */
2574 iter->pos = 0; /* forget it */
2579 static void *fib_route_seq_start(struct seq_file *seq, loff_t *pos)
2582 struct fib_route_iter *iter = seq->private;
2583 struct fib_table *tb;
2588 tb = fib_get_table(seq_file_net(seq), RT_TABLE_MAIN);
2593 t = (struct trie *)tb->tb_data;
2594 iter->tnode = t->kv;
2597 return fib_route_get_idx(iter, *pos);
2600 iter->key = KEY_MAX;
2602 return SEQ_START_TOKEN;
2605 static void *fib_route_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2607 struct fib_route_iter *iter = seq->private;
2608 struct key_vector *l = NULL;
2609 t_key key = iter->key + 1;
2613 /* only allow key of 0 for start of sequence */
2614 if ((v == SEQ_START_TOKEN) || key)
2615 l = leaf_walk_rcu(&iter->tnode, key);
2627 static void fib_route_seq_stop(struct seq_file *seq, void *v)
2633 static unsigned int fib_flag_trans(int type, __be32 mask, const struct fib_info *fi)
2635 unsigned int flags = 0;
2637 if (type == RTN_UNREACHABLE || type == RTN_PROHIBIT)
2639 if (fi && fi->fib_nh->nh_gw)
2640 flags |= RTF_GATEWAY;
2641 if (mask == htonl(0xFFFFFFFF))
2648 * This outputs /proc/net/route.
2649 * The format of the file is not supposed to be changed
2650 * and needs to be same as fib_hash output to avoid breaking
2653 static int fib_route_seq_show(struct seq_file *seq, void *v)
2655 struct fib_route_iter *iter = seq->private;
2656 struct fib_table *tb = iter->main_tb;
2657 struct fib_alias *fa;
2658 struct key_vector *l = v;
2661 if (v == SEQ_START_TOKEN) {
2662 seq_printf(seq, "%-127s\n", "Iface\tDestination\tGateway "
2663 "\tFlags\tRefCnt\tUse\tMetric\tMask\t\tMTU"
2668 prefix = htonl(l->key);
2670 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
2671 const struct fib_info *fi = fa->fa_info;
2672 __be32 mask = inet_make_mask(KEYLENGTH - fa->fa_slen);
2673 unsigned int flags = fib_flag_trans(fa->fa_type, mask, fi);
2675 if ((fa->fa_type == RTN_BROADCAST) ||
2676 (fa->fa_type == RTN_MULTICAST))
2679 if (fa->tb_id != tb->tb_id)
2682 seq_setwidth(seq, 127);
2686 "%s\t%08X\t%08X\t%04X\t%d\t%u\t"
2687 "%d\t%08X\t%d\t%u\t%u",
2688 fi->fib_dev ? fi->fib_dev->name : "*",
2690 fi->fib_nh->nh_gw, flags, 0, 0,
2694 fi->fib_advmss + 40 : 0),
2699 "*\t%08X\t%08X\t%04X\t%d\t%u\t"
2700 "%d\t%08X\t%d\t%u\t%u",
2701 prefix, 0, flags, 0, 0, 0,
2710 static const struct seq_operations fib_route_seq_ops = {
2711 .start = fib_route_seq_start,
2712 .next = fib_route_seq_next,
2713 .stop = fib_route_seq_stop,
2714 .show = fib_route_seq_show,
2717 static int fib_route_seq_open(struct inode *inode, struct file *file)
2719 return seq_open_net(inode, file, &fib_route_seq_ops,
2720 sizeof(struct fib_route_iter));
2723 static const struct file_operations fib_route_fops = {
2724 .owner = THIS_MODULE,
2725 .open = fib_route_seq_open,
2727 .llseek = seq_lseek,
2728 .release = seq_release_net,
2731 int __net_init fib_proc_init(struct net *net)
2733 if (!proc_create("fib_trie", S_IRUGO, net->proc_net, &fib_trie_fops))
2736 if (!proc_create("fib_triestat", S_IRUGO, net->proc_net,
2737 &fib_triestat_fops))
2740 if (!proc_create("route", S_IRUGO, net->proc_net, &fib_route_fops))
2746 remove_proc_entry("fib_triestat", net->proc_net);
2748 remove_proc_entry("fib_trie", net->proc_net);
2753 void __net_exit fib_proc_exit(struct net *net)
2755 remove_proc_entry("fib_trie", net->proc_net);
2756 remove_proc_entry("fib_triestat", net->proc_net);
2757 remove_proc_entry("route", net->proc_net);
2760 #endif /* CONFIG_PROC_FS */