2 * This program is free software; you can redistribute it and/or
3 * modify it under the terms of the GNU General Public License
4 * as published by the Free Software Foundation; either version
5 * 2 of the License, or (at your option) any later version.
7 * Robert Olsson <robert.olsson@its.uu.se> Uppsala Universitet
8 * & Swedish University of Agricultural Sciences.
10 * Jens Laas <jens.laas@data.slu.se> Swedish University of
11 * Agricultural Sciences.
13 * Hans Liss <hans.liss@its.uu.se> Uppsala Universitet
15 * This work is based on the LPC-trie which is originally described in:
17 * An experimental study of compression methods for dynamic tries
18 * Stefan Nilsson and Matti Tikkanen. Algorithmica, 33(1):19-33, 2002.
19 * http://www.csc.kth.se/~snilsson/software/dyntrie2/
22 * IP-address lookup using LC-tries. Stefan Nilsson and Gunnar Karlsson
23 * IEEE Journal on Selected Areas in Communications, 17(6):1083-1092, June 1999
26 * Code from fib_hash has been reused which includes the following header:
29 * INET An implementation of the TCP/IP protocol suite for the LINUX
30 * operating system. INET is implemented using the BSD Socket
31 * interface as the means of communication with the user level.
33 * IPv4 FIB: lookup engine and maintenance routines.
36 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
38 * This program is free software; you can redistribute it and/or
39 * modify it under the terms of the GNU General Public License
40 * as published by the Free Software Foundation; either version
41 * 2 of the License, or (at your option) any later version.
43 * Substantial contributions to this work comes from:
45 * David S. Miller, <davem@davemloft.net>
46 * Stephen Hemminger <shemminger@osdl.org>
47 * Paul E. McKenney <paulmck@us.ibm.com>
48 * Patrick McHardy <kaber@trash.net>
51 #define VERSION "0.409"
53 #include <asm/uaccess.h>
54 #include <linux/bitops.h>
55 #include <linux/types.h>
56 #include <linux/kernel.h>
58 #include <linux/string.h>
59 #include <linux/socket.h>
60 #include <linux/sockios.h>
61 #include <linux/errno.h>
63 #include <linux/inet.h>
64 #include <linux/inetdevice.h>
65 #include <linux/netdevice.h>
66 #include <linux/if_arp.h>
67 #include <linux/proc_fs.h>
68 #include <linux/rcupdate.h>
69 #include <linux/skbuff.h>
70 #include <linux/netlink.h>
71 #include <linux/init.h>
72 #include <linux/list.h>
73 #include <linux/slab.h>
74 #include <linux/export.h>
75 #include <linux/vmalloc.h>
76 #include <net/net_namespace.h>
78 #include <net/protocol.h>
79 #include <net/route.h>
82 #include <net/ip_fib.h>
83 #include <net/switchdev.h>
84 #include "fib_lookup.h"
86 #define MAX_STAT_DEPTH 32
88 #define KEYLENGTH (8*sizeof(t_key))
89 #define KEY_MAX ((t_key)~0)
91 typedef unsigned int t_key;
93 #define IS_TRIE(n) ((n)->pos >= KEYLENGTH)
94 #define IS_TNODE(n) ((n)->bits)
95 #define IS_LEAF(n) (!(n)->bits)
99 unsigned char pos; /* 2log(KEYLENGTH) bits needed */
100 unsigned char bits; /* 2log(KEYLENGTH) bits needed */
103 /* This list pointer if valid if (pos | bits) == 0 (LEAF) */
104 struct hlist_head leaf;
105 /* This array is valid if (pos | bits) > 0 (TNODE) */
106 struct key_vector __rcu *tnode[0];
112 t_key empty_children; /* KEYLENGTH bits needed */
113 t_key full_children; /* KEYLENGTH bits needed */
114 struct key_vector __rcu *parent;
115 struct key_vector kv[1];
116 #define tn_bits kv[0].bits
119 #define TNODE_SIZE(n) offsetof(struct tnode, kv[0].tnode[n])
120 #define LEAF_SIZE TNODE_SIZE(1)
122 #ifdef CONFIG_IP_FIB_TRIE_STATS
123 struct trie_use_stats {
125 unsigned int backtrack;
126 unsigned int semantic_match_passed;
127 unsigned int semantic_match_miss;
128 unsigned int null_node_hit;
129 unsigned int resize_node_skipped;
134 unsigned int totdepth;
135 unsigned int maxdepth;
138 unsigned int nullpointers;
139 unsigned int prefixes;
140 unsigned int nodesizes[MAX_STAT_DEPTH];
144 struct key_vector kv[1];
145 #ifdef CONFIG_IP_FIB_TRIE_STATS
146 struct trie_use_stats __percpu *stats;
150 static struct key_vector *resize(struct trie *t, struct key_vector *tn);
151 static size_t tnode_free_size;
154 * synchronize_rcu after call_rcu for that many pages; it should be especially
155 * useful before resizing the root node with PREEMPT_NONE configs; the value was
156 * obtained experimentally, aiming to avoid visible slowdown.
158 static const int sync_pages = 128;
160 static struct kmem_cache *fn_alias_kmem __read_mostly;
161 static struct kmem_cache *trie_leaf_kmem __read_mostly;
163 static inline struct tnode *tn_info(struct key_vector *kv)
165 return container_of(kv, struct tnode, kv[0]);
168 /* caller must hold RTNL */
169 #define node_parent(tn) rtnl_dereference(tn_info(tn)->parent)
170 #define get_child(tn, i) rtnl_dereference((tn)->tnode[i])
172 /* caller must hold RCU read lock or RTNL */
173 #define node_parent_rcu(tn) rcu_dereference_rtnl(tn_info(tn)->parent)
174 #define get_child_rcu(tn, i) rcu_dereference_rtnl((tn)->tnode[i])
176 /* wrapper for rcu_assign_pointer */
177 static inline void node_set_parent(struct key_vector *n, struct key_vector *tp)
180 rcu_assign_pointer(tn_info(n)->parent, tp);
183 #define NODE_INIT_PARENT(n, p) RCU_INIT_POINTER(tn_info(n)->parent, p)
185 /* This provides us with the number of children in this node, in the case of a
186 * leaf this will return 0 meaning none of the children are accessible.
188 static inline unsigned long child_length(const struct key_vector *tn)
190 return (1ul << tn->bits) & ~(1ul);
193 #define get_cindex(key, kv) (((key) ^ (kv)->key) >> (kv)->pos)
195 static inline unsigned long get_index(t_key key, struct key_vector *kv)
197 unsigned long index = key ^ kv->key;
199 if ((BITS_PER_LONG <= KEYLENGTH) && (KEYLENGTH == kv->pos))
202 return index >> kv->pos;
205 /* To understand this stuff, an understanding of keys and all their bits is
206 * necessary. Every node in the trie has a key associated with it, but not
207 * all of the bits in that key are significant.
209 * Consider a node 'n' and its parent 'tp'.
211 * If n is a leaf, every bit in its key is significant. Its presence is
212 * necessitated by path compression, since during a tree traversal (when
213 * searching for a leaf - unless we are doing an insertion) we will completely
214 * ignore all skipped bits we encounter. Thus we need to verify, at the end of
215 * a potentially successful search, that we have indeed been walking the
218 * Note that we can never "miss" the correct key in the tree if present by
219 * following the wrong path. Path compression ensures that segments of the key
220 * that are the same for all keys with a given prefix are skipped, but the
221 * skipped part *is* identical for each node in the subtrie below the skipped
222 * bit! trie_insert() in this implementation takes care of that.
224 * if n is an internal node - a 'tnode' here, the various parts of its key
225 * have many different meanings.
228 * _________________________________________________________________
229 * | i | i | i | i | i | i | i | N | N | N | S | S | S | S | S | C |
230 * -----------------------------------------------------------------
231 * 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16
233 * _________________________________________________________________
234 * | C | C | C | u | u | u | u | u | u | u | u | u | u | u | u | u |
235 * -----------------------------------------------------------------
236 * 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
243 * First, let's just ignore the bits that come before the parent tp, that is
244 * the bits from (tp->pos + tp->bits) to 31. They are *known* but at this
245 * point we do not use them for anything.
247 * The bits from (tp->pos) to (tp->pos + tp->bits - 1) - "N", above - are the
248 * index into the parent's child array. That is, they will be used to find
249 * 'n' among tp's children.
251 * The bits from (n->pos + n->bits) to (tn->pos - 1) - "S" - are skipped bits
254 * All the bits we have seen so far are significant to the node n. The rest
255 * of the bits are really not needed or indeed known in n->key.
257 * The bits from (n->pos) to (n->pos + n->bits - 1) - "C" - are the index into
258 * n's child array, and will of course be different for each child.
260 * The rest of the bits, from 0 to (n->pos + n->bits), are completely unknown
264 static const int halve_threshold = 25;
265 static const int inflate_threshold = 50;
266 static const int halve_threshold_root = 15;
267 static const int inflate_threshold_root = 30;
269 static void __alias_free_mem(struct rcu_head *head)
271 struct fib_alias *fa = container_of(head, struct fib_alias, rcu);
272 kmem_cache_free(fn_alias_kmem, fa);
275 static inline void alias_free_mem_rcu(struct fib_alias *fa)
277 call_rcu(&fa->rcu, __alias_free_mem);
280 #define TNODE_KMALLOC_MAX \
281 ilog2((PAGE_SIZE - TNODE_SIZE(0)) / sizeof(struct key_vector *))
282 #define TNODE_VMALLOC_MAX \
283 ilog2((SIZE_MAX - TNODE_SIZE(0)) / sizeof(struct key_vector *))
285 static void __node_free_rcu(struct rcu_head *head)
287 struct tnode *n = container_of(head, struct tnode, rcu);
290 kmem_cache_free(trie_leaf_kmem, n);
291 else if (n->tn_bits <= TNODE_KMALLOC_MAX)
297 #define node_free(n) call_rcu(&tn_info(n)->rcu, __node_free_rcu)
299 static struct tnode *tnode_alloc(int bits)
303 /* verify bits is within bounds */
304 if (bits > TNODE_VMALLOC_MAX)
307 /* determine size and verify it is non-zero and didn't overflow */
308 size = TNODE_SIZE(1ul << bits);
310 if (size <= PAGE_SIZE)
311 return kzalloc(size, GFP_KERNEL);
313 return vzalloc(size);
316 static inline void empty_child_inc(struct key_vector *n)
318 ++tn_info(n)->empty_children ? : ++tn_info(n)->full_children;
321 static inline void empty_child_dec(struct key_vector *n)
323 tn_info(n)->empty_children-- ? : tn_info(n)->full_children--;
326 static struct key_vector *leaf_new(t_key key, struct fib_alias *fa)
328 struct key_vector *l;
331 kv = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
335 /* initialize key vector */
340 l->slen = fa->fa_slen;
342 /* link leaf to fib alias */
343 INIT_HLIST_HEAD(&l->leaf);
344 hlist_add_head(&fa->fa_list, &l->leaf);
349 static struct key_vector *tnode_new(t_key key, int pos, int bits)
351 unsigned int shift = pos + bits;
352 struct key_vector *tn;
355 /* verify bits and pos their msb bits clear and values are valid */
356 BUG_ON(!bits || (shift > KEYLENGTH));
358 tnode = tnode_alloc(bits);
362 pr_debug("AT %p s=%zu %zu\n", tnode, TNODE_SIZE(0),
363 sizeof(struct key_vector *) << bits);
365 if (bits == KEYLENGTH)
366 tnode->full_children = 1;
368 tnode->empty_children = 1ul << bits;
371 tn->key = (shift < KEYLENGTH) ? (key >> shift) << shift : 0;
379 /* Check whether a tnode 'n' is "full", i.e. it is an internal node
380 * and no bits are skipped. See discussion in dyntree paper p. 6
382 static inline int tnode_full(struct key_vector *tn, struct key_vector *n)
384 return n && ((n->pos + n->bits) == tn->pos) && IS_TNODE(n);
387 /* Add a child at position i overwriting the old value.
388 * Update the value of full_children and empty_children.
390 static void put_child(struct key_vector *tn, unsigned long i,
391 struct key_vector *n)
393 struct key_vector *chi = get_child(tn, i);
396 BUG_ON(i >= child_length(tn));
398 /* update emptyChildren, overflow into fullChildren */
404 /* update fullChildren */
405 wasfull = tnode_full(tn, chi);
406 isfull = tnode_full(tn, n);
408 if (wasfull && !isfull)
409 tn_info(tn)->full_children--;
410 else if (!wasfull && isfull)
411 tn_info(tn)->full_children++;
413 if (n && (tn->slen < n->slen))
416 rcu_assign_pointer(tn->tnode[i], n);
419 static void update_children(struct key_vector *tn)
423 /* update all of the child parent pointers */
424 for (i = child_length(tn); i;) {
425 struct key_vector *inode = get_child(tn, --i);
430 /* Either update the children of a tnode that
431 * already belongs to us or update the child
432 * to point to ourselves.
434 if (node_parent(inode) == tn)
435 update_children(inode);
437 node_set_parent(inode, tn);
441 static inline void put_child_root(struct key_vector *tp, t_key key,
442 struct key_vector *n)
445 rcu_assign_pointer(tp->tnode[0], n);
447 put_child(tp, get_index(key, tp), n);
450 static inline void tnode_free_init(struct key_vector *tn)
452 tn_info(tn)->rcu.next = NULL;
455 static inline void tnode_free_append(struct key_vector *tn,
456 struct key_vector *n)
458 tn_info(n)->rcu.next = tn_info(tn)->rcu.next;
459 tn_info(tn)->rcu.next = &tn_info(n)->rcu;
462 static void tnode_free(struct key_vector *tn)
464 struct callback_head *head = &tn_info(tn)->rcu;
468 tnode_free_size += TNODE_SIZE(1ul << tn->bits);
471 tn = container_of(head, struct tnode, rcu)->kv;
474 if (tnode_free_size >= PAGE_SIZE * sync_pages) {
480 static struct key_vector *replace(struct trie *t,
481 struct key_vector *oldtnode,
482 struct key_vector *tn)
484 struct key_vector *tp = node_parent(oldtnode);
487 /* setup the parent pointer out of and back into this node */
488 NODE_INIT_PARENT(tn, tp);
489 put_child_root(tp, tn->key, tn);
491 /* update all of the child parent pointers */
494 /* all pointers should be clean so we are done */
495 tnode_free(oldtnode);
497 /* resize children now that oldtnode is freed */
498 for (i = child_length(tn); i;) {
499 struct key_vector *inode = get_child(tn, --i);
501 /* resize child node */
502 if (tnode_full(tn, inode))
503 tn = resize(t, inode);
509 static struct key_vector *inflate(struct trie *t,
510 struct key_vector *oldtnode)
512 struct key_vector *tn;
516 pr_debug("In inflate\n");
518 tn = tnode_new(oldtnode->key, oldtnode->pos - 1, oldtnode->bits + 1);
522 /* prepare oldtnode to be freed */
523 tnode_free_init(oldtnode);
525 /* Assemble all of the pointers in our cluster, in this case that
526 * represents all of the pointers out of our allocated nodes that
527 * point to existing tnodes and the links between our allocated
530 for (i = child_length(oldtnode), m = 1u << tn->pos; i;) {
531 struct key_vector *inode = get_child(oldtnode, --i);
532 struct key_vector *node0, *node1;
539 /* A leaf or an internal node with skipped bits */
540 if (!tnode_full(oldtnode, inode)) {
541 put_child(tn, get_index(inode->key, tn), inode);
545 /* drop the node in the old tnode free list */
546 tnode_free_append(oldtnode, inode);
548 /* An internal node with two children */
549 if (inode->bits == 1) {
550 put_child(tn, 2 * i + 1, get_child(inode, 1));
551 put_child(tn, 2 * i, get_child(inode, 0));
555 /* We will replace this node 'inode' with two new
556 * ones, 'node0' and 'node1', each with half of the
557 * original children. The two new nodes will have
558 * a position one bit further down the key and this
559 * means that the "significant" part of their keys
560 * (see the discussion near the top of this file)
561 * will differ by one bit, which will be "0" in
562 * node0's key and "1" in node1's key. Since we are
563 * moving the key position by one step, the bit that
564 * we are moving away from - the bit at position
565 * (tn->pos) - is the one that will differ between
566 * node0 and node1. So... we synthesize that bit in the
569 node1 = tnode_new(inode->key | m, inode->pos, inode->bits - 1);
572 node0 = tnode_new(inode->key, inode->pos, inode->bits - 1);
574 tnode_free_append(tn, node1);
577 tnode_free_append(tn, node0);
579 /* populate child pointers in new nodes */
580 for (k = child_length(inode), j = k / 2; j;) {
581 put_child(node1, --j, get_child(inode, --k));
582 put_child(node0, j, get_child(inode, j));
583 put_child(node1, --j, get_child(inode, --k));
584 put_child(node0, j, get_child(inode, j));
587 /* link new nodes to parent */
588 NODE_INIT_PARENT(node1, tn);
589 NODE_INIT_PARENT(node0, tn);
591 /* link parent to nodes */
592 put_child(tn, 2 * i + 1, node1);
593 put_child(tn, 2 * i, node0);
596 /* setup the parent pointers into and out of this node */
597 return replace(t, oldtnode, tn);
599 /* all pointers should be clean so we are done */
605 static struct key_vector *halve(struct trie *t,
606 struct key_vector *oldtnode)
608 struct key_vector *tn;
611 pr_debug("In halve\n");
613 tn = tnode_new(oldtnode->key, oldtnode->pos + 1, oldtnode->bits - 1);
617 /* prepare oldtnode to be freed */
618 tnode_free_init(oldtnode);
620 /* Assemble all of the pointers in our cluster, in this case that
621 * represents all of the pointers out of our allocated nodes that
622 * point to existing tnodes and the links between our allocated
625 for (i = child_length(oldtnode); i;) {
626 struct key_vector *node1 = get_child(oldtnode, --i);
627 struct key_vector *node0 = get_child(oldtnode, --i);
628 struct key_vector *inode;
630 /* At least one of the children is empty */
631 if (!node1 || !node0) {
632 put_child(tn, i / 2, node1 ? : node0);
636 /* Two nonempty children */
637 inode = tnode_new(node0->key, oldtnode->pos, 1);
640 tnode_free_append(tn, inode);
642 /* initialize pointers out of node */
643 put_child(inode, 1, node1);
644 put_child(inode, 0, node0);
645 NODE_INIT_PARENT(inode, tn);
647 /* link parent to node */
648 put_child(tn, i / 2, inode);
651 /* setup the parent pointers into and out of this node */
652 return replace(t, oldtnode, tn);
654 /* all pointers should be clean so we are done */
660 static struct key_vector *collapse(struct trie *t,
661 struct key_vector *oldtnode)
663 struct key_vector *n, *tp;
666 /* scan the tnode looking for that one child that might still exist */
667 for (n = NULL, i = child_length(oldtnode); !n && i;)
668 n = get_child(oldtnode, --i);
670 /* compress one level */
671 tp = node_parent(oldtnode);
672 put_child_root(tp, oldtnode->key, n);
673 node_set_parent(n, tp);
681 static unsigned char update_suffix(struct key_vector *tn)
683 unsigned char slen = tn->pos;
684 unsigned long stride, i;
686 /* search though the list of children looking for nodes that might
687 * have a suffix greater than the one we currently have. This is
688 * why we start with a stride of 2 since a stride of 1 would
689 * represent the nodes with suffix length equal to tn->pos
691 for (i = 0, stride = 0x2ul ; i < child_length(tn); i += stride) {
692 struct key_vector *n = get_child(tn, i);
694 if (!n || (n->slen <= slen))
697 /* update stride and slen based on new value */
698 stride <<= (n->slen - slen);
702 /* if slen covers all but the last bit we can stop here
703 * there will be nothing longer than that since only node
704 * 0 and 1 << (bits - 1) could have that as their suffix
707 if ((slen + 1) >= (tn->pos + tn->bits))
716 /* From "Implementing a dynamic compressed trie" by Stefan Nilsson of
717 * the Helsinki University of Technology and Matti Tikkanen of Nokia
718 * Telecommunications, page 6:
719 * "A node is doubled if the ratio of non-empty children to all
720 * children in the *doubled* node is at least 'high'."
722 * 'high' in this instance is the variable 'inflate_threshold'. It
723 * is expressed as a percentage, so we multiply it with
724 * child_length() and instead of multiplying by 2 (since the
725 * child array will be doubled by inflate()) and multiplying
726 * the left-hand side by 100 (to handle the percentage thing) we
727 * multiply the left-hand side by 50.
729 * The left-hand side may look a bit weird: child_length(tn)
730 * - tn->empty_children is of course the number of non-null children
731 * in the current node. tn->full_children is the number of "full"
732 * children, that is non-null tnodes with a skip value of 0.
733 * All of those will be doubled in the resulting inflated tnode, so
734 * we just count them one extra time here.
736 * A clearer way to write this would be:
738 * to_be_doubled = tn->full_children;
739 * not_to_be_doubled = child_length(tn) - tn->empty_children -
742 * new_child_length = child_length(tn) * 2;
744 * new_fill_factor = 100 * (not_to_be_doubled + 2*to_be_doubled) /
746 * if (new_fill_factor >= inflate_threshold)
748 * ...and so on, tho it would mess up the while () loop.
751 * 100 * (not_to_be_doubled + 2*to_be_doubled) / new_child_length >=
755 * 100 * (not_to_be_doubled + 2*to_be_doubled) >=
756 * inflate_threshold * new_child_length
758 * expand not_to_be_doubled and to_be_doubled, and shorten:
759 * 100 * (child_length(tn) - tn->empty_children +
760 * tn->full_children) >= inflate_threshold * new_child_length
762 * expand new_child_length:
763 * 100 * (child_length(tn) - tn->empty_children +
764 * tn->full_children) >=
765 * inflate_threshold * child_length(tn) * 2
768 * 50 * (tn->full_children + child_length(tn) -
769 * tn->empty_children) >= inflate_threshold *
773 static inline bool should_inflate(struct key_vector *tp, struct key_vector *tn)
775 unsigned long used = child_length(tn);
776 unsigned long threshold = used;
778 /* Keep root node larger */
779 threshold *= IS_TRIE(tp) ? inflate_threshold_root : inflate_threshold;
780 used -= tn_info(tn)->empty_children;
781 used += tn_info(tn)->full_children;
783 /* if bits == KEYLENGTH then pos = 0, and will fail below */
785 return (used > 1) && tn->pos && ((50 * used) >= threshold);
788 static inline bool should_halve(struct key_vector *tp, struct key_vector *tn)
790 unsigned long used = child_length(tn);
791 unsigned long threshold = used;
793 /* Keep root node larger */
794 threshold *= IS_TRIE(tp) ? halve_threshold_root : halve_threshold;
795 used -= tn_info(tn)->empty_children;
797 /* if bits == KEYLENGTH then used = 100% on wrap, and will fail below */
799 return (used > 1) && (tn->bits > 1) && ((100 * used) < threshold);
802 static inline bool should_collapse(struct key_vector *tn)
804 unsigned long used = child_length(tn);
806 used -= tn_info(tn)->empty_children;
808 /* account for bits == KEYLENGTH case */
809 if ((tn->bits == KEYLENGTH) && tn_info(tn)->full_children)
812 /* One child or none, time to drop us from the trie */
817 static struct key_vector *resize(struct trie *t, struct key_vector *tn)
819 #ifdef CONFIG_IP_FIB_TRIE_STATS
820 struct trie_use_stats __percpu *stats = t->stats;
822 struct key_vector *tp = node_parent(tn);
823 unsigned long cindex = get_index(tn->key, tp);
824 int max_work = MAX_WORK;
826 pr_debug("In tnode_resize %p inflate_threshold=%d threshold=%d\n",
827 tn, inflate_threshold, halve_threshold);
829 /* track the tnode via the pointer from the parent instead of
830 * doing it ourselves. This way we can let RCU fully do its
831 * thing without us interfering
833 BUG_ON(tn != get_child(tp, cindex));
835 /* Double as long as the resulting node has a number of
836 * nonempty nodes that are above the threshold.
838 while (should_inflate(tp, tn) && max_work) {
841 #ifdef CONFIG_IP_FIB_TRIE_STATS
842 this_cpu_inc(stats->resize_node_skipped);
848 tn = get_child(tp, cindex);
851 /* update parent in case inflate failed */
852 tp = node_parent(tn);
854 /* Return if at least one inflate is run */
855 if (max_work != MAX_WORK)
858 /* Halve as long as the number of empty children in this
859 * node is above threshold.
861 while (should_halve(tp, tn) && max_work) {
864 #ifdef CONFIG_IP_FIB_TRIE_STATS
865 this_cpu_inc(stats->resize_node_skipped);
871 tn = get_child(tp, cindex);
874 /* Only one child remains */
875 if (should_collapse(tn))
876 return collapse(t, tn);
878 /* update parent in case halve failed */
879 tp = node_parent(tn);
881 /* Return if at least one deflate was run */
882 if (max_work != MAX_WORK)
885 /* push the suffix length to the parent node */
886 if (tn->slen > tn->pos) {
887 unsigned char slen = update_suffix(tn);
896 static void leaf_pull_suffix(struct key_vector *tp, struct key_vector *l)
898 while ((tp->slen > tp->pos) && (tp->slen > l->slen)) {
899 if (update_suffix(tp) > l->slen)
901 tp = node_parent(tp);
905 static void leaf_push_suffix(struct key_vector *tn, struct key_vector *l)
907 /* if this is a new leaf then tn will be NULL and we can sort
908 * out parent suffix lengths as a part of trie_rebalance
910 while (tn->slen < l->slen) {
912 tn = node_parent(tn);
916 /* rcu_read_lock needs to be hold by caller from readside */
917 static struct key_vector *fib_find_node(struct trie *t,
918 struct key_vector **tp, u32 key)
920 struct key_vector *pn, *n = t->kv;
921 unsigned long index = 0;
925 n = get_child_rcu(n, index);
930 index = get_cindex(key, n);
932 /* This bit of code is a bit tricky but it combines multiple
933 * checks into a single check. The prefix consists of the
934 * prefix plus zeros for the bits in the cindex. The index
935 * is the difference between the key and this value. From
936 * this we can actually derive several pieces of data.
937 * if (index >= (1ul << bits))
938 * we have a mismatch in skip bits and failed
940 * we know the value is cindex
942 * This check is safe even if bits == KEYLENGTH due to the
943 * fact that we can only allocate a node with 32 bits if a
944 * long is greater than 32 bits.
946 if (index >= (1ul << n->bits)) {
951 /* keep searching until we find a perfect match leaf or NULL */
952 } while (IS_TNODE(n));
959 /* Return the first fib alias matching TOS with
960 * priority less than or equal to PRIO.
962 static struct fib_alias *fib_find_alias(struct hlist_head *fah, u8 slen,
963 u8 tos, u32 prio, u32 tb_id)
965 struct fib_alias *fa;
970 hlist_for_each_entry(fa, fah, fa_list) {
971 if (fa->fa_slen < slen)
973 if (fa->fa_slen != slen)
975 if (fa->tb_id > tb_id)
977 if (fa->tb_id != tb_id)
979 if (fa->fa_tos > tos)
981 if (fa->fa_info->fib_priority >= prio || fa->fa_tos < tos)
988 static void trie_rebalance(struct trie *t, struct key_vector *tn)
994 static int fib_insert_node(struct trie *t, struct key_vector *tp,
995 struct fib_alias *new, t_key key)
997 struct key_vector *n, *l;
999 l = leaf_new(key, new);
1003 /* retrieve child from parent node */
1004 n = get_child(tp, get_index(key, tp));
1006 /* Case 2: n is a LEAF or a TNODE and the key doesn't match.
1008 * Add a new tnode here
1009 * first tnode need some special handling
1010 * leaves us in position for handling as case 3
1013 struct key_vector *tn;
1015 tn = tnode_new(key, __fls(key ^ n->key), 1);
1019 /* initialize routes out of node */
1020 NODE_INIT_PARENT(tn, tp);
1021 put_child(tn, get_index(key, tn) ^ 1, n);
1023 /* start adding routes into the node */
1024 put_child_root(tp, key, tn);
1025 node_set_parent(n, tn);
1027 /* parent now has a NULL spot where the leaf can go */
1031 /* Case 3: n is NULL, and will just insert a new leaf */
1032 NODE_INIT_PARENT(l, tp);
1033 put_child_root(tp, key, l);
1034 trie_rebalance(t, tp);
1043 static int fib_insert_alias(struct trie *t, struct key_vector *tp,
1044 struct key_vector *l, struct fib_alias *new,
1045 struct fib_alias *fa, t_key key)
1048 return fib_insert_node(t, tp, new, key);
1051 hlist_add_before_rcu(&new->fa_list, &fa->fa_list);
1053 struct fib_alias *last;
1055 hlist_for_each_entry(last, &l->leaf, fa_list) {
1056 if (new->fa_slen < last->fa_slen)
1058 if ((new->fa_slen == last->fa_slen) &&
1059 (new->tb_id > last->tb_id))
1065 hlist_add_behind_rcu(&new->fa_list, &fa->fa_list);
1067 hlist_add_head_rcu(&new->fa_list, &l->leaf);
1070 /* if we added to the tail node then we need to update slen */
1071 if (l->slen < new->fa_slen) {
1072 l->slen = new->fa_slen;
1073 leaf_push_suffix(tp, l);
1079 /* Caller must hold RTNL. */
1080 int fib_table_insert(struct fib_table *tb, struct fib_config *cfg)
1082 struct trie *t = (struct trie *)tb->tb_data;
1083 struct fib_alias *fa, *new_fa;
1084 struct key_vector *l, *tp;
1085 unsigned int nlflags = 0;
1086 struct fib_info *fi;
1087 u8 plen = cfg->fc_dst_len;
1088 u8 slen = KEYLENGTH - plen;
1089 u8 tos = cfg->fc_tos;
1093 if (plen > KEYLENGTH)
1096 key = ntohl(cfg->fc_dst);
1098 pr_debug("Insert table=%u %08x/%d\n", tb->tb_id, key, plen);
1100 if ((plen < KEYLENGTH) && (key << plen))
1103 fi = fib_create_info(cfg);
1109 l = fib_find_node(t, &tp, key);
1110 fa = l ? fib_find_alias(&l->leaf, slen, tos, fi->fib_priority,
1113 /* Now fa, if non-NULL, points to the first fib alias
1114 * with the same keys [prefix,tos,priority], if such key already
1115 * exists or to the node before which we will insert new one.
1117 * If fa is NULL, we will need to allocate a new one and
1118 * insert to the tail of the section matching the suffix length
1122 if (fa && fa->fa_tos == tos &&
1123 fa->fa_info->fib_priority == fi->fib_priority) {
1124 struct fib_alias *fa_first, *fa_match;
1127 if (cfg->fc_nlflags & NLM_F_EXCL)
1131 * 1. Find exact match for type, scope, fib_info to avoid
1133 * 2. Find next 'fa' (or head), NLM_F_APPEND inserts before it
1137 hlist_for_each_entry_from(fa, fa_list) {
1138 if ((fa->fa_slen != slen) ||
1139 (fa->tb_id != tb->tb_id) ||
1140 (fa->fa_tos != tos))
1142 if (fa->fa_info->fib_priority != fi->fib_priority)
1144 if (fa->fa_type == cfg->fc_type &&
1145 fa->fa_info == fi) {
1151 if (cfg->fc_nlflags & NLM_F_REPLACE) {
1152 struct fib_info *fi_drop;
1162 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1166 fi_drop = fa->fa_info;
1167 new_fa->fa_tos = fa->fa_tos;
1168 new_fa->fa_info = fi;
1169 new_fa->fa_type = cfg->fc_type;
1170 state = fa->fa_state;
1171 new_fa->fa_state = state & ~FA_S_ACCESSED;
1172 new_fa->fa_slen = fa->fa_slen;
1173 new_fa->tb_id = tb->tb_id;
1174 new_fa->fa_default = -1;
1176 err = switchdev_fib_ipv4_add(key, plen, fi,
1182 switchdev_fib_ipv4_abort(fi);
1183 kmem_cache_free(fn_alias_kmem, new_fa);
1187 hlist_replace_rcu(&fa->fa_list, &new_fa->fa_list);
1189 alias_free_mem_rcu(fa);
1191 fib_release_info(fi_drop);
1192 if (state & FA_S_ACCESSED)
1193 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1194 rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen,
1195 tb->tb_id, &cfg->fc_nlinfo, NLM_F_REPLACE);
1199 /* Error if we find a perfect match which
1200 * uses the same scope, type, and nexthop
1206 if (cfg->fc_nlflags & NLM_F_APPEND)
1207 nlflags = NLM_F_APPEND;
1212 if (!(cfg->fc_nlflags & NLM_F_CREATE))
1216 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1220 new_fa->fa_info = fi;
1221 new_fa->fa_tos = tos;
1222 new_fa->fa_type = cfg->fc_type;
1223 new_fa->fa_state = 0;
1224 new_fa->fa_slen = slen;
1225 new_fa->tb_id = tb->tb_id;
1226 new_fa->fa_default = -1;
1228 /* (Optionally) offload fib entry to switch hardware. */
1229 err = switchdev_fib_ipv4_add(key, plen, fi, tos, cfg->fc_type,
1230 cfg->fc_nlflags, tb->tb_id);
1232 switchdev_fib_ipv4_abort(fi);
1233 goto out_free_new_fa;
1236 /* Insert new entry to the list. */
1237 err = fib_insert_alias(t, tp, l, new_fa, fa, key);
1239 goto out_sw_fib_del;
1242 tb->tb_num_default++;
1244 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1245 rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen, new_fa->tb_id,
1246 &cfg->fc_nlinfo, nlflags);
1251 switchdev_fib_ipv4_del(key, plen, fi, tos, cfg->fc_type, tb->tb_id);
1253 kmem_cache_free(fn_alias_kmem, new_fa);
1255 fib_release_info(fi);
1260 static inline t_key prefix_mismatch(t_key key, struct key_vector *n)
1262 t_key prefix = n->key;
1264 return (key ^ prefix) & (prefix | -prefix);
1267 /* should be called with rcu_read_lock */
1268 int fib_table_lookup(struct fib_table *tb, const struct flowi4 *flp,
1269 struct fib_result *res, int fib_flags)
1271 struct trie *t = (struct trie *) tb->tb_data;
1272 #ifdef CONFIG_IP_FIB_TRIE_STATS
1273 struct trie_use_stats __percpu *stats = t->stats;
1275 const t_key key = ntohl(flp->daddr);
1276 struct key_vector *n, *pn;
1277 struct fib_alias *fa;
1278 unsigned long index;
1284 n = get_child_rcu(pn, cindex);
1288 #ifdef CONFIG_IP_FIB_TRIE_STATS
1289 this_cpu_inc(stats->gets);
1292 /* Step 1: Travel to the longest prefix match in the trie */
1294 index = get_cindex(key, n);
1296 /* This bit of code is a bit tricky but it combines multiple
1297 * checks into a single check. The prefix consists of the
1298 * prefix plus zeros for the "bits" in the prefix. The index
1299 * is the difference between the key and this value. From
1300 * this we can actually derive several pieces of data.
1301 * if (index >= (1ul << bits))
1302 * we have a mismatch in skip bits and failed
1304 * we know the value is cindex
1306 * This check is safe even if bits == KEYLENGTH due to the
1307 * fact that we can only allocate a node with 32 bits if a
1308 * long is greater than 32 bits.
1310 if (index >= (1ul << n->bits))
1313 /* we have found a leaf. Prefixes have already been compared */
1317 /* only record pn and cindex if we are going to be chopping
1318 * bits later. Otherwise we are just wasting cycles.
1320 if (n->slen > n->pos) {
1325 n = get_child_rcu(n, index);
1330 /* Step 2: Sort out leaves and begin backtracing for longest prefix */
1332 /* record the pointer where our next node pointer is stored */
1333 struct key_vector __rcu **cptr = n->tnode;
1335 /* This test verifies that none of the bits that differ
1336 * between the key and the prefix exist in the region of
1337 * the lsb and higher in the prefix.
1339 if (unlikely(prefix_mismatch(key, n)) || (n->slen == n->pos))
1342 /* exit out and process leaf */
1343 if (unlikely(IS_LEAF(n)))
1346 /* Don't bother recording parent info. Since we are in
1347 * prefix match mode we will have to come back to wherever
1348 * we started this traversal anyway
1351 while ((n = rcu_dereference(*cptr)) == NULL) {
1353 #ifdef CONFIG_IP_FIB_TRIE_STATS
1355 this_cpu_inc(stats->null_node_hit);
1357 /* If we are at cindex 0 there are no more bits for
1358 * us to strip at this level so we must ascend back
1359 * up one level to see if there are any more bits to
1360 * be stripped there.
1363 t_key pkey = pn->key;
1365 /* If we don't have a parent then there is
1366 * nothing for us to do as we do not have any
1367 * further nodes to parse.
1371 #ifdef CONFIG_IP_FIB_TRIE_STATS
1372 this_cpu_inc(stats->backtrack);
1374 /* Get Child's index */
1375 pn = node_parent_rcu(pn);
1376 cindex = get_index(pkey, pn);
1379 /* strip the least significant bit from the cindex */
1380 cindex &= cindex - 1;
1382 /* grab pointer for next child node */
1383 cptr = &pn->tnode[cindex];
1388 /* this line carries forward the xor from earlier in the function */
1389 index = key ^ n->key;
1391 /* Step 3: Process the leaf, if that fails fall back to backtracing */
1392 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list) {
1393 struct fib_info *fi = fa->fa_info;
1396 if ((index >= (1ul << fa->fa_slen)) &&
1397 ((BITS_PER_LONG > KEYLENGTH) || (fa->fa_slen != KEYLENGTH)))
1399 if (fa->fa_tos && fa->fa_tos != flp->flowi4_tos)
1403 if (fa->fa_info->fib_scope < flp->flowi4_scope)
1405 fib_alias_accessed(fa);
1406 err = fib_props[fa->fa_type].error;
1407 if (unlikely(err < 0)) {
1408 #ifdef CONFIG_IP_FIB_TRIE_STATS
1409 this_cpu_inc(stats->semantic_match_passed);
1413 if (fi->fib_flags & RTNH_F_DEAD)
1415 for (nhsel = 0; nhsel < fi->fib_nhs; nhsel++) {
1416 const struct fib_nh *nh = &fi->fib_nh[nhsel];
1417 struct in_device *in_dev = __in_dev_get_rcu(nh->nh_dev);
1419 if (nh->nh_flags & RTNH_F_DEAD)
1422 IN_DEV_IGNORE_ROUTES_WITH_LINKDOWN(in_dev) &&
1423 nh->nh_flags & RTNH_F_LINKDOWN &&
1424 !(fib_flags & FIB_LOOKUP_IGNORE_LINKSTATE))
1426 if (flp->flowi4_oif && flp->flowi4_oif != nh->nh_oif)
1429 if (!(fib_flags & FIB_LOOKUP_NOREF))
1430 atomic_inc(&fi->fib_clntref);
1432 res->prefixlen = KEYLENGTH - fa->fa_slen;
1433 res->nh_sel = nhsel;
1434 res->type = fa->fa_type;
1435 res->scope = fi->fib_scope;
1438 res->fa_head = &n->leaf;
1439 #ifdef CONFIG_IP_FIB_TRIE_STATS
1440 this_cpu_inc(stats->semantic_match_passed);
1445 #ifdef CONFIG_IP_FIB_TRIE_STATS
1446 this_cpu_inc(stats->semantic_match_miss);
1450 EXPORT_SYMBOL_GPL(fib_table_lookup);
1452 static void fib_remove_alias(struct trie *t, struct key_vector *tp,
1453 struct key_vector *l, struct fib_alias *old)
1455 /* record the location of the previous list_info entry */
1456 struct hlist_node **pprev = old->fa_list.pprev;
1457 struct fib_alias *fa = hlist_entry(pprev, typeof(*fa), fa_list.next);
1459 /* remove the fib_alias from the list */
1460 hlist_del_rcu(&old->fa_list);
1462 /* if we emptied the list this leaf will be freed and we can sort
1463 * out parent suffix lengths as a part of trie_rebalance
1465 if (hlist_empty(&l->leaf)) {
1466 put_child_root(tp, l->key, NULL);
1468 trie_rebalance(t, tp);
1472 /* only access fa if it is pointing at the last valid hlist_node */
1476 /* update the trie with the latest suffix length */
1477 l->slen = fa->fa_slen;
1478 leaf_pull_suffix(tp, l);
1481 /* Caller must hold RTNL. */
1482 int fib_table_delete(struct fib_table *tb, struct fib_config *cfg)
1484 struct trie *t = (struct trie *) tb->tb_data;
1485 struct fib_alias *fa, *fa_to_delete;
1486 struct key_vector *l, *tp;
1487 u8 plen = cfg->fc_dst_len;
1488 u8 slen = KEYLENGTH - plen;
1489 u8 tos = cfg->fc_tos;
1492 if (plen > KEYLENGTH)
1495 key = ntohl(cfg->fc_dst);
1497 if ((plen < KEYLENGTH) && (key << plen))
1500 l = fib_find_node(t, &tp, key);
1504 fa = fib_find_alias(&l->leaf, slen, tos, 0, tb->tb_id);
1508 pr_debug("Deleting %08x/%d tos=%d t=%p\n", key, plen, tos, t);
1510 fa_to_delete = NULL;
1511 hlist_for_each_entry_from(fa, fa_list) {
1512 struct fib_info *fi = fa->fa_info;
1514 if ((fa->fa_slen != slen) ||
1515 (fa->tb_id != tb->tb_id) ||
1516 (fa->fa_tos != tos))
1519 if ((!cfg->fc_type || fa->fa_type == cfg->fc_type) &&
1520 (cfg->fc_scope == RT_SCOPE_NOWHERE ||
1521 fa->fa_info->fib_scope == cfg->fc_scope) &&
1522 (!cfg->fc_prefsrc ||
1523 fi->fib_prefsrc == cfg->fc_prefsrc) &&
1524 (!cfg->fc_protocol ||
1525 fi->fib_protocol == cfg->fc_protocol) &&
1526 fib_nh_match(cfg, fi) == 0) {
1535 switchdev_fib_ipv4_del(key, plen, fa_to_delete->fa_info, tos,
1536 cfg->fc_type, tb->tb_id);
1538 rtmsg_fib(RTM_DELROUTE, htonl(key), fa_to_delete, plen, tb->tb_id,
1539 &cfg->fc_nlinfo, 0);
1542 tb->tb_num_default--;
1544 fib_remove_alias(t, tp, l, fa_to_delete);
1546 if (fa_to_delete->fa_state & FA_S_ACCESSED)
1547 rt_cache_flush(cfg->fc_nlinfo.nl_net);
1549 fib_release_info(fa_to_delete->fa_info);
1550 alias_free_mem_rcu(fa_to_delete);
1554 /* Scan for the next leaf starting at the provided key value */
1555 static struct key_vector *leaf_walk_rcu(struct key_vector **tn, t_key key)
1557 struct key_vector *pn, *n = *tn;
1558 unsigned long cindex;
1560 /* this loop is meant to try and find the key in the trie */
1562 /* record parent and next child index */
1564 cindex = key ? get_index(key, pn) : 0;
1566 if (cindex >> pn->bits)
1569 /* descend into the next child */
1570 n = get_child_rcu(pn, cindex++);
1574 /* guarantee forward progress on the keys */
1575 if (IS_LEAF(n) && (n->key >= key))
1577 } while (IS_TNODE(n));
1579 /* this loop will search for the next leaf with a greater key */
1580 while (!IS_TRIE(pn)) {
1581 /* if we exhausted the parent node we will need to climb */
1582 if (cindex >= (1ul << pn->bits)) {
1583 t_key pkey = pn->key;
1585 pn = node_parent_rcu(pn);
1586 cindex = get_index(pkey, pn) + 1;
1590 /* grab the next available node */
1591 n = get_child_rcu(pn, cindex++);
1595 /* no need to compare keys since we bumped the index */
1599 /* Rescan start scanning in new node */
1605 return NULL; /* Root of trie */
1607 /* if we are at the limit for keys just return NULL for the tnode */
1612 static void fib_trie_free(struct fib_table *tb)
1614 struct trie *t = (struct trie *)tb->tb_data;
1615 struct key_vector *pn = t->kv;
1616 unsigned long cindex = 1;
1617 struct hlist_node *tmp;
1618 struct fib_alias *fa;
1620 /* walk trie in reverse order and free everything */
1622 struct key_vector *n;
1625 t_key pkey = pn->key;
1631 pn = node_parent(pn);
1633 /* drop emptied tnode */
1634 put_child_root(pn, n->key, NULL);
1637 cindex = get_index(pkey, pn);
1642 /* grab the next available node */
1643 n = get_child(pn, cindex);
1648 /* record pn and cindex for leaf walking */
1650 cindex = 1ul << n->bits;
1655 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1656 hlist_del_rcu(&fa->fa_list);
1657 alias_free_mem_rcu(fa);
1660 put_child_root(pn, n->key, NULL);
1664 #ifdef CONFIG_IP_FIB_TRIE_STATS
1665 free_percpu(t->stats);
1670 struct fib_table *fib_trie_unmerge(struct fib_table *oldtb)
1672 struct trie *ot = (struct trie *)oldtb->tb_data;
1673 struct key_vector *l, *tp = ot->kv;
1674 struct fib_table *local_tb;
1675 struct fib_alias *fa;
1679 if (oldtb->tb_data == oldtb->__data)
1682 local_tb = fib_trie_table(RT_TABLE_LOCAL, NULL);
1686 lt = (struct trie *)local_tb->tb_data;
1688 while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
1689 struct key_vector *local_l = NULL, *local_tp;
1691 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
1692 struct fib_alias *new_fa;
1694 if (local_tb->tb_id != fa->tb_id)
1697 /* clone fa for new local table */
1698 new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1702 memcpy(new_fa, fa, sizeof(*fa));
1704 /* insert clone into table */
1706 local_l = fib_find_node(lt, &local_tp, l->key);
1708 if (fib_insert_alias(lt, local_tp, local_l, new_fa,
1713 /* stop loop if key wrapped back to 0 */
1721 fib_trie_free(local_tb);
1726 /* Caller must hold RTNL */
1727 void fib_table_flush_external(struct fib_table *tb)
1729 struct trie *t = (struct trie *)tb->tb_data;
1730 struct key_vector *pn = t->kv;
1731 unsigned long cindex = 1;
1732 struct hlist_node *tmp;
1733 struct fib_alias *fa;
1735 /* walk trie in reverse order */
1737 unsigned char slen = 0;
1738 struct key_vector *n;
1741 t_key pkey = pn->key;
1743 /* cannot resize the trie vector */
1747 /* resize completed node */
1749 cindex = get_index(pkey, pn);
1754 /* grab the next available node */
1755 n = get_child(pn, cindex);
1760 /* record pn and cindex for leaf walking */
1762 cindex = 1ul << n->bits;
1767 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1768 struct fib_info *fi = fa->fa_info;
1770 /* if alias was cloned to local then we just
1771 * need to remove the local copy from main
1773 if (tb->tb_id != fa->tb_id) {
1774 hlist_del_rcu(&fa->fa_list);
1775 alias_free_mem_rcu(fa);
1779 /* record local slen */
1782 if (!fi || !(fi->fib_flags & RTNH_F_OFFLOAD))
1785 switchdev_fib_ipv4_del(n->key, KEYLENGTH - fa->fa_slen,
1786 fi, fa->fa_tos, fa->fa_type,
1790 /* update leaf slen */
1793 if (hlist_empty(&n->leaf)) {
1794 put_child_root(pn, n->key, NULL);
1800 /* Caller must hold RTNL. */
1801 int fib_table_flush(struct fib_table *tb)
1803 struct trie *t = (struct trie *)tb->tb_data;
1804 struct key_vector *pn = t->kv;
1805 unsigned long cindex = 1;
1806 struct hlist_node *tmp;
1807 struct fib_alias *fa;
1810 /* walk trie in reverse order */
1812 unsigned char slen = 0;
1813 struct key_vector *n;
1816 t_key pkey = pn->key;
1818 /* cannot resize the trie vector */
1822 /* resize completed node */
1824 cindex = get_index(pkey, pn);
1829 /* grab the next available node */
1830 n = get_child(pn, cindex);
1835 /* record pn and cindex for leaf walking */
1837 cindex = 1ul << n->bits;
1842 hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1843 struct fib_info *fi = fa->fa_info;
1845 if (!fi || !(fi->fib_flags & RTNH_F_DEAD)) {
1850 switchdev_fib_ipv4_del(n->key, KEYLENGTH - fa->fa_slen,
1851 fi, fa->fa_tos, fa->fa_type,
1853 hlist_del_rcu(&fa->fa_list);
1854 fib_release_info(fa->fa_info);
1855 alias_free_mem_rcu(fa);
1859 /* update leaf slen */
1862 if (hlist_empty(&n->leaf)) {
1863 put_child_root(pn, n->key, NULL);
1868 pr_debug("trie_flush found=%d\n", found);
1872 static void __trie_free_rcu(struct rcu_head *head)
1874 struct fib_table *tb = container_of(head, struct fib_table, rcu);
1875 #ifdef CONFIG_IP_FIB_TRIE_STATS
1876 struct trie *t = (struct trie *)tb->tb_data;
1878 if (tb->tb_data == tb->__data)
1879 free_percpu(t->stats);
1880 #endif /* CONFIG_IP_FIB_TRIE_STATS */
1884 void fib_free_table(struct fib_table *tb)
1886 call_rcu(&tb->rcu, __trie_free_rcu);
1889 static int fn_trie_dump_leaf(struct key_vector *l, struct fib_table *tb,
1890 struct sk_buff *skb, struct netlink_callback *cb)
1892 __be32 xkey = htonl(l->key);
1893 struct fib_alias *fa;
1899 /* rcu_read_lock is hold by caller */
1900 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
1906 if (tb->tb_id != fa->tb_id) {
1911 if (fib_dump_info(skb, NETLINK_CB(cb->skb).portid,
1917 KEYLENGTH - fa->fa_slen,
1919 fa->fa_info, NLM_F_MULTI) < 0) {
1930 /* rcu_read_lock needs to be hold by caller from readside */
1931 int fib_table_dump(struct fib_table *tb, struct sk_buff *skb,
1932 struct netlink_callback *cb)
1934 struct trie *t = (struct trie *)tb->tb_data;
1935 struct key_vector *l, *tp = t->kv;
1936 /* Dump starting at last key.
1937 * Note: 0.0.0.0/0 (ie default) is first key.
1939 int count = cb->args[2];
1940 t_key key = cb->args[3];
1942 while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
1943 if (fn_trie_dump_leaf(l, tb, skb, cb) < 0) {
1945 cb->args[2] = count;
1952 memset(&cb->args[4], 0,
1953 sizeof(cb->args) - 4*sizeof(cb->args[0]));
1955 /* stop loop if key wrapped back to 0 */
1961 cb->args[2] = count;
1966 void __init fib_trie_init(void)
1968 fn_alias_kmem = kmem_cache_create("ip_fib_alias",
1969 sizeof(struct fib_alias),
1970 0, SLAB_PANIC, NULL);
1972 trie_leaf_kmem = kmem_cache_create("ip_fib_trie",
1974 0, SLAB_PANIC, NULL);
1977 struct fib_table *fib_trie_table(u32 id, struct fib_table *alias)
1979 struct fib_table *tb;
1981 size_t sz = sizeof(*tb);
1984 sz += sizeof(struct trie);
1986 tb = kzalloc(sz, GFP_KERNEL);
1991 tb->tb_num_default = 0;
1992 tb->tb_data = (alias ? alias->__data : tb->__data);
1997 t = (struct trie *) tb->tb_data;
1998 t->kv[0].pos = KEYLENGTH;
1999 t->kv[0].slen = KEYLENGTH;
2000 #ifdef CONFIG_IP_FIB_TRIE_STATS
2001 t->stats = alloc_percpu(struct trie_use_stats);
2011 #ifdef CONFIG_PROC_FS
2012 /* Depth first Trie walk iterator */
2013 struct fib_trie_iter {
2014 struct seq_net_private p;
2015 struct fib_table *tb;
2016 struct key_vector *tnode;
2021 static struct key_vector *fib_trie_get_next(struct fib_trie_iter *iter)
2023 unsigned long cindex = iter->index;
2024 struct key_vector *pn = iter->tnode;
2027 pr_debug("get_next iter={node=%p index=%d depth=%d}\n",
2028 iter->tnode, iter->index, iter->depth);
2030 while (!IS_TRIE(pn)) {
2031 while (cindex < child_length(pn)) {
2032 struct key_vector *n = get_child_rcu(pn, cindex++);
2039 iter->index = cindex;
2041 /* push down one level */
2050 /* Current node exhausted, pop back up */
2052 pn = node_parent_rcu(pn);
2053 cindex = get_index(pkey, pn) + 1;
2057 /* record root node so further searches know we are done */
2064 static struct key_vector *fib_trie_get_first(struct fib_trie_iter *iter,
2067 struct key_vector *n, *pn;
2073 n = rcu_dereference(pn->tnode[0]);
2090 static void trie_collect_stats(struct trie *t, struct trie_stat *s)
2092 struct key_vector *n;
2093 struct fib_trie_iter iter;
2095 memset(s, 0, sizeof(*s));
2098 for (n = fib_trie_get_first(&iter, t); n; n = fib_trie_get_next(&iter)) {
2100 struct fib_alias *fa;
2103 s->totdepth += iter.depth;
2104 if (iter.depth > s->maxdepth)
2105 s->maxdepth = iter.depth;
2107 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list)
2111 if (n->bits < MAX_STAT_DEPTH)
2112 s->nodesizes[n->bits]++;
2113 s->nullpointers += tn_info(n)->empty_children;
2120 * This outputs /proc/net/fib_triestats
2122 static void trie_show_stats(struct seq_file *seq, struct trie_stat *stat)
2124 unsigned int i, max, pointers, bytes, avdepth;
2127 avdepth = stat->totdepth*100 / stat->leaves;
2131 seq_printf(seq, "\tAver depth: %u.%02d\n",
2132 avdepth / 100, avdepth % 100);
2133 seq_printf(seq, "\tMax depth: %u\n", stat->maxdepth);
2135 seq_printf(seq, "\tLeaves: %u\n", stat->leaves);
2136 bytes = LEAF_SIZE * stat->leaves;
2138 seq_printf(seq, "\tPrefixes: %u\n", stat->prefixes);
2139 bytes += sizeof(struct fib_alias) * stat->prefixes;
2141 seq_printf(seq, "\tInternal nodes: %u\n\t", stat->tnodes);
2142 bytes += TNODE_SIZE(0) * stat->tnodes;
2144 max = MAX_STAT_DEPTH;
2145 while (max > 0 && stat->nodesizes[max-1] == 0)
2149 for (i = 1; i < max; i++)
2150 if (stat->nodesizes[i] != 0) {
2151 seq_printf(seq, " %u: %u", i, stat->nodesizes[i]);
2152 pointers += (1<<i) * stat->nodesizes[i];
2154 seq_putc(seq, '\n');
2155 seq_printf(seq, "\tPointers: %u\n", pointers);
2157 bytes += sizeof(struct key_vector *) * pointers;
2158 seq_printf(seq, "Null ptrs: %u\n", stat->nullpointers);
2159 seq_printf(seq, "Total size: %u kB\n", (bytes + 1023) / 1024);
2162 #ifdef CONFIG_IP_FIB_TRIE_STATS
2163 static void trie_show_usage(struct seq_file *seq,
2164 const struct trie_use_stats __percpu *stats)
2166 struct trie_use_stats s = { 0 };
2169 /* loop through all of the CPUs and gather up the stats */
2170 for_each_possible_cpu(cpu) {
2171 const struct trie_use_stats *pcpu = per_cpu_ptr(stats, cpu);
2173 s.gets += pcpu->gets;
2174 s.backtrack += pcpu->backtrack;
2175 s.semantic_match_passed += pcpu->semantic_match_passed;
2176 s.semantic_match_miss += pcpu->semantic_match_miss;
2177 s.null_node_hit += pcpu->null_node_hit;
2178 s.resize_node_skipped += pcpu->resize_node_skipped;
2181 seq_printf(seq, "\nCounters:\n---------\n");
2182 seq_printf(seq, "gets = %u\n", s.gets);
2183 seq_printf(seq, "backtracks = %u\n", s.backtrack);
2184 seq_printf(seq, "semantic match passed = %u\n",
2185 s.semantic_match_passed);
2186 seq_printf(seq, "semantic match miss = %u\n", s.semantic_match_miss);
2187 seq_printf(seq, "null node hit= %u\n", s.null_node_hit);
2188 seq_printf(seq, "skipped node resize = %u\n\n", s.resize_node_skipped);
2190 #endif /* CONFIG_IP_FIB_TRIE_STATS */
2192 static void fib_table_print(struct seq_file *seq, struct fib_table *tb)
2194 if (tb->tb_id == RT_TABLE_LOCAL)
2195 seq_puts(seq, "Local:\n");
2196 else if (tb->tb_id == RT_TABLE_MAIN)
2197 seq_puts(seq, "Main:\n");
2199 seq_printf(seq, "Id %d:\n", tb->tb_id);
2203 static int fib_triestat_seq_show(struct seq_file *seq, void *v)
2205 struct net *net = (struct net *)seq->private;
2209 "Basic info: size of leaf:"
2210 " %Zd bytes, size of tnode: %Zd bytes.\n",
2211 LEAF_SIZE, TNODE_SIZE(0));
2213 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2214 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2215 struct fib_table *tb;
2217 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2218 struct trie *t = (struct trie *) tb->tb_data;
2219 struct trie_stat stat;
2224 fib_table_print(seq, tb);
2226 trie_collect_stats(t, &stat);
2227 trie_show_stats(seq, &stat);
2228 #ifdef CONFIG_IP_FIB_TRIE_STATS
2229 trie_show_usage(seq, t->stats);
2237 static int fib_triestat_seq_open(struct inode *inode, struct file *file)
2239 return single_open_net(inode, file, fib_triestat_seq_show);
2242 static const struct file_operations fib_triestat_fops = {
2243 .owner = THIS_MODULE,
2244 .open = fib_triestat_seq_open,
2246 .llseek = seq_lseek,
2247 .release = single_release_net,
2250 static struct key_vector *fib_trie_get_idx(struct seq_file *seq, loff_t pos)
2252 struct fib_trie_iter *iter = seq->private;
2253 struct net *net = seq_file_net(seq);
2257 for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2258 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2259 struct fib_table *tb;
2261 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2262 struct key_vector *n;
2264 for (n = fib_trie_get_first(iter,
2265 (struct trie *) tb->tb_data);
2266 n; n = fib_trie_get_next(iter))
2277 static void *fib_trie_seq_start(struct seq_file *seq, loff_t *pos)
2281 return fib_trie_get_idx(seq, *pos);
2284 static void *fib_trie_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2286 struct fib_trie_iter *iter = seq->private;
2287 struct net *net = seq_file_net(seq);
2288 struct fib_table *tb = iter->tb;
2289 struct hlist_node *tb_node;
2291 struct key_vector *n;
2294 /* next node in same table */
2295 n = fib_trie_get_next(iter);
2299 /* walk rest of this hash chain */
2300 h = tb->tb_id & (FIB_TABLE_HASHSZ - 1);
2301 while ((tb_node = rcu_dereference(hlist_next_rcu(&tb->tb_hlist)))) {
2302 tb = hlist_entry(tb_node, struct fib_table, tb_hlist);
2303 n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2308 /* new hash chain */
2309 while (++h < FIB_TABLE_HASHSZ) {
2310 struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2311 hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2312 n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2324 static void fib_trie_seq_stop(struct seq_file *seq, void *v)
2330 static void seq_indent(struct seq_file *seq, int n)
2336 static inline const char *rtn_scope(char *buf, size_t len, enum rt_scope_t s)
2339 case RT_SCOPE_UNIVERSE: return "universe";
2340 case RT_SCOPE_SITE: return "site";
2341 case RT_SCOPE_LINK: return "link";
2342 case RT_SCOPE_HOST: return "host";
2343 case RT_SCOPE_NOWHERE: return "nowhere";
2345 snprintf(buf, len, "scope=%d", s);
2350 static const char *const rtn_type_names[__RTN_MAX] = {
2351 [RTN_UNSPEC] = "UNSPEC",
2352 [RTN_UNICAST] = "UNICAST",
2353 [RTN_LOCAL] = "LOCAL",
2354 [RTN_BROADCAST] = "BROADCAST",
2355 [RTN_ANYCAST] = "ANYCAST",
2356 [RTN_MULTICAST] = "MULTICAST",
2357 [RTN_BLACKHOLE] = "BLACKHOLE",
2358 [RTN_UNREACHABLE] = "UNREACHABLE",
2359 [RTN_PROHIBIT] = "PROHIBIT",
2360 [RTN_THROW] = "THROW",
2362 [RTN_XRESOLVE] = "XRESOLVE",
2365 static inline const char *rtn_type(char *buf, size_t len, unsigned int t)
2367 if (t < __RTN_MAX && rtn_type_names[t])
2368 return rtn_type_names[t];
2369 snprintf(buf, len, "type %u", t);
2373 /* Pretty print the trie */
2374 static int fib_trie_seq_show(struct seq_file *seq, void *v)
2376 const struct fib_trie_iter *iter = seq->private;
2377 struct key_vector *n = v;
2379 if (IS_TRIE(node_parent_rcu(n)))
2380 fib_table_print(seq, iter->tb);
2383 __be32 prf = htonl(n->key);
2385 seq_indent(seq, iter->depth-1);
2386 seq_printf(seq, " +-- %pI4/%zu %u %u %u\n",
2387 &prf, KEYLENGTH - n->pos - n->bits, n->bits,
2388 tn_info(n)->full_children,
2389 tn_info(n)->empty_children);
2391 __be32 val = htonl(n->key);
2392 struct fib_alias *fa;
2394 seq_indent(seq, iter->depth);
2395 seq_printf(seq, " |-- %pI4\n", &val);
2397 hlist_for_each_entry_rcu(fa, &n->leaf, fa_list) {
2398 char buf1[32], buf2[32];
2400 seq_indent(seq, iter->depth + 1);
2401 seq_printf(seq, " /%zu %s %s",
2402 KEYLENGTH - fa->fa_slen,
2403 rtn_scope(buf1, sizeof(buf1),
2404 fa->fa_info->fib_scope),
2405 rtn_type(buf2, sizeof(buf2),
2408 seq_printf(seq, " tos=%d", fa->fa_tos);
2409 seq_putc(seq, '\n');
2416 static const struct seq_operations fib_trie_seq_ops = {
2417 .start = fib_trie_seq_start,
2418 .next = fib_trie_seq_next,
2419 .stop = fib_trie_seq_stop,
2420 .show = fib_trie_seq_show,
2423 static int fib_trie_seq_open(struct inode *inode, struct file *file)
2425 return seq_open_net(inode, file, &fib_trie_seq_ops,
2426 sizeof(struct fib_trie_iter));
2429 static const struct file_operations fib_trie_fops = {
2430 .owner = THIS_MODULE,
2431 .open = fib_trie_seq_open,
2433 .llseek = seq_lseek,
2434 .release = seq_release_net,
2437 struct fib_route_iter {
2438 struct seq_net_private p;
2439 struct fib_table *main_tb;
2440 struct key_vector *tnode;
2445 static struct key_vector *fib_route_get_idx(struct fib_route_iter *iter,
2448 struct fib_table *tb = iter->main_tb;
2449 struct key_vector *l, **tp = &iter->tnode;
2453 /* use cache location of next-to-find key */
2454 if (iter->pos > 0 && pos >= iter->pos) {
2458 t = (struct trie *)tb->tb_data;
2459 iter->tnode = t->kv;
2464 while ((l = leaf_walk_rcu(tp, key)) != NULL) {
2473 /* handle unlikely case of a key wrap */
2479 iter->key = key; /* remember it */
2481 iter->pos = 0; /* forget it */
2486 static void *fib_route_seq_start(struct seq_file *seq, loff_t *pos)
2489 struct fib_route_iter *iter = seq->private;
2490 struct fib_table *tb;
2495 tb = fib_get_table(seq_file_net(seq), RT_TABLE_MAIN);
2502 return fib_route_get_idx(iter, *pos);
2504 t = (struct trie *)tb->tb_data;
2505 iter->tnode = t->kv;
2509 return SEQ_START_TOKEN;
2512 static void *fib_route_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2514 struct fib_route_iter *iter = seq->private;
2515 struct key_vector *l = NULL;
2516 t_key key = iter->key;
2520 /* only allow key of 0 for start of sequence */
2521 if ((v == SEQ_START_TOKEN) || key)
2522 l = leaf_walk_rcu(&iter->tnode, key);
2525 iter->key = l->key + 1;
2534 static void fib_route_seq_stop(struct seq_file *seq, void *v)
2540 static unsigned int fib_flag_trans(int type, __be32 mask, const struct fib_info *fi)
2542 unsigned int flags = 0;
2544 if (type == RTN_UNREACHABLE || type == RTN_PROHIBIT)
2546 if (fi && fi->fib_nh->nh_gw)
2547 flags |= RTF_GATEWAY;
2548 if (mask == htonl(0xFFFFFFFF))
2555 * This outputs /proc/net/route.
2556 * The format of the file is not supposed to be changed
2557 * and needs to be same as fib_hash output to avoid breaking
2560 static int fib_route_seq_show(struct seq_file *seq, void *v)
2562 struct fib_route_iter *iter = seq->private;
2563 struct fib_table *tb = iter->main_tb;
2564 struct fib_alias *fa;
2565 struct key_vector *l = v;
2568 if (v == SEQ_START_TOKEN) {
2569 seq_printf(seq, "%-127s\n", "Iface\tDestination\tGateway "
2570 "\tFlags\tRefCnt\tUse\tMetric\tMask\t\tMTU"
2575 prefix = htonl(l->key);
2577 hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
2578 const struct fib_info *fi = fa->fa_info;
2579 __be32 mask = inet_make_mask(KEYLENGTH - fa->fa_slen);
2580 unsigned int flags = fib_flag_trans(fa->fa_type, mask, fi);
2582 if ((fa->fa_type == RTN_BROADCAST) ||
2583 (fa->fa_type == RTN_MULTICAST))
2586 if (fa->tb_id != tb->tb_id)
2589 seq_setwidth(seq, 127);
2593 "%s\t%08X\t%08X\t%04X\t%d\t%u\t"
2594 "%d\t%08X\t%d\t%u\t%u",
2595 fi->fib_dev ? fi->fib_dev->name : "*",
2597 fi->fib_nh->nh_gw, flags, 0, 0,
2601 fi->fib_advmss + 40 : 0),
2606 "*\t%08X\t%08X\t%04X\t%d\t%u\t"
2607 "%d\t%08X\t%d\t%u\t%u",
2608 prefix, 0, flags, 0, 0, 0,
2617 static const struct seq_operations fib_route_seq_ops = {
2618 .start = fib_route_seq_start,
2619 .next = fib_route_seq_next,
2620 .stop = fib_route_seq_stop,
2621 .show = fib_route_seq_show,
2624 static int fib_route_seq_open(struct inode *inode, struct file *file)
2626 return seq_open_net(inode, file, &fib_route_seq_ops,
2627 sizeof(struct fib_route_iter));
2630 static const struct file_operations fib_route_fops = {
2631 .owner = THIS_MODULE,
2632 .open = fib_route_seq_open,
2634 .llseek = seq_lseek,
2635 .release = seq_release_net,
2638 int __net_init fib_proc_init(struct net *net)
2640 if (!proc_create("fib_trie", S_IRUGO, net->proc_net, &fib_trie_fops))
2643 if (!proc_create("fib_triestat", S_IRUGO, net->proc_net,
2644 &fib_triestat_fops))
2647 if (!proc_create("route", S_IRUGO, net->proc_net, &fib_route_fops))
2653 remove_proc_entry("fib_triestat", net->proc_net);
2655 remove_proc_entry("fib_trie", net->proc_net);
2660 void __net_exit fib_proc_exit(struct net *net)
2662 remove_proc_entry("fib_trie", net->proc_net);
2663 remove_proc_entry("fib_triestat", net->proc_net);
2664 remove_proc_entry("route", net->proc_net);
2667 #endif /* CONFIG_PROC_FS */