xref: /openbmc/linux/net/ipv4/fib_trie.c (revision fa7f32422ea1ac276b45b96a540ed5981caaa61f)
1 /*
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.
6  *
7  *   Robert Olsson <robert.olsson@its.uu.se> Uppsala Universitet
8  *     & Swedish University of Agricultural Sciences.
9  *
10  *   Jens Laas <jens.laas@data.slu.se> Swedish University of
11  *     Agricultural Sciences.
12  *
13  *   Hans Liss <hans.liss@its.uu.se>  Uppsala Universitet
14  *
15  * This work is based on the LPC-trie which is originally described in:
16  *
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/
20  *
21  *
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
24  *
25  *
26  * Code from fib_hash has been reused which includes the following header:
27  *
28  *
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.
32  *
33  *		IPv4 FIB: lookup engine and maintenance routines.
34  *
35  *
36  * Authors:	Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
37  *
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.
42  *
43  * Substantial contributions to this work comes from:
44  *
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>
49  */
50 
51 #define VERSION "0.409"
52 
53 #include <asm/uaccess.h>
54 #include <linux/bitops.h>
55 #include <linux/types.h>
56 #include <linux/kernel.h>
57 #include <linux/mm.h>
58 #include <linux/string.h>
59 #include <linux/socket.h>
60 #include <linux/sockios.h>
61 #include <linux/errno.h>
62 #include <linux/in.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>
78 #include <net/ip.h>
79 #include <net/protocol.h>
80 #include <net/route.h>
81 #include <net/tcp.h>
82 #include <net/sock.h>
83 #include <net/ip_fib.h>
84 #include <trace/events/fib.h>
85 #include "fib_lookup.h"
86 
87 static BLOCKING_NOTIFIER_HEAD(fib_chain);
88 
89 int register_fib_notifier(struct notifier_block *nb)
90 {
91 	return blocking_notifier_chain_register(&fib_chain, nb);
92 }
93 EXPORT_SYMBOL(register_fib_notifier);
94 
95 int unregister_fib_notifier(struct notifier_block *nb)
96 {
97 	return blocking_notifier_chain_unregister(&fib_chain, nb);
98 }
99 EXPORT_SYMBOL(unregister_fib_notifier);
100 
101 int call_fib_notifiers(struct net *net, enum fib_event_type event_type,
102 		       struct fib_notifier_info *info)
103 {
104 	info->net = net;
105 	return blocking_notifier_call_chain(&fib_chain, event_type, info);
106 }
107 
108 static int call_fib_entry_notifiers(struct net *net,
109 				    enum fib_event_type event_type, u32 dst,
110 				    int dst_len, struct fib_info *fi,
111 				    u8 tos, u8 type, u32 tb_id, u32 nlflags)
112 {
113 	struct fib_entry_notifier_info info = {
114 		.dst = dst,
115 		.dst_len = dst_len,
116 		.fi = fi,
117 		.tos = tos,
118 		.type = type,
119 		.tb_id = tb_id,
120 		.nlflags = nlflags,
121 	};
122 	return call_fib_notifiers(net, event_type, &info.info);
123 }
124 
125 #define MAX_STAT_DEPTH 32
126 
127 #define KEYLENGTH	(8*sizeof(t_key))
128 #define KEY_MAX		((t_key)~0)
129 
130 typedef unsigned int t_key;
131 
132 #define IS_TRIE(n)	((n)->pos >= KEYLENGTH)
133 #define IS_TNODE(n)	((n)->bits)
134 #define IS_LEAF(n)	(!(n)->bits)
135 
136 struct key_vector {
137 	t_key key;
138 	unsigned char pos;		/* 2log(KEYLENGTH) bits needed */
139 	unsigned char bits;		/* 2log(KEYLENGTH) bits needed */
140 	unsigned char slen;
141 	union {
142 		/* This list pointer if valid if (pos | bits) == 0 (LEAF) */
143 		struct hlist_head leaf;
144 		/* This array is valid if (pos | bits) > 0 (TNODE) */
145 		struct key_vector __rcu *tnode[0];
146 	};
147 };
148 
149 struct tnode {
150 	struct rcu_head rcu;
151 	t_key empty_children;		/* KEYLENGTH bits needed */
152 	t_key full_children;		/* KEYLENGTH bits needed */
153 	struct key_vector __rcu *parent;
154 	struct key_vector kv[1];
155 #define tn_bits kv[0].bits
156 };
157 
158 #define TNODE_SIZE(n)	offsetof(struct tnode, kv[0].tnode[n])
159 #define LEAF_SIZE	TNODE_SIZE(1)
160 
161 #ifdef CONFIG_IP_FIB_TRIE_STATS
162 struct trie_use_stats {
163 	unsigned int gets;
164 	unsigned int backtrack;
165 	unsigned int semantic_match_passed;
166 	unsigned int semantic_match_miss;
167 	unsigned int null_node_hit;
168 	unsigned int resize_node_skipped;
169 };
170 #endif
171 
172 struct trie_stat {
173 	unsigned int totdepth;
174 	unsigned int maxdepth;
175 	unsigned int tnodes;
176 	unsigned int leaves;
177 	unsigned int nullpointers;
178 	unsigned int prefixes;
179 	unsigned int nodesizes[MAX_STAT_DEPTH];
180 };
181 
182 struct trie {
183 	struct key_vector kv[1];
184 #ifdef CONFIG_IP_FIB_TRIE_STATS
185 	struct trie_use_stats __percpu *stats;
186 #endif
187 };
188 
189 static struct key_vector *resize(struct trie *t, struct key_vector *tn);
190 static size_t tnode_free_size;
191 
192 /*
193  * synchronize_rcu after call_rcu for that many pages; it should be especially
194  * useful before resizing the root node with PREEMPT_NONE configs; the value was
195  * obtained experimentally, aiming to avoid visible slowdown.
196  */
197 static const int sync_pages = 128;
198 
199 static struct kmem_cache *fn_alias_kmem __read_mostly;
200 static struct kmem_cache *trie_leaf_kmem __read_mostly;
201 
202 static inline struct tnode *tn_info(struct key_vector *kv)
203 {
204 	return container_of(kv, struct tnode, kv[0]);
205 }
206 
207 /* caller must hold RTNL */
208 #define node_parent(tn) rtnl_dereference(tn_info(tn)->parent)
209 #define get_child(tn, i) rtnl_dereference((tn)->tnode[i])
210 
211 /* caller must hold RCU read lock or RTNL */
212 #define node_parent_rcu(tn) rcu_dereference_rtnl(tn_info(tn)->parent)
213 #define get_child_rcu(tn, i) rcu_dereference_rtnl((tn)->tnode[i])
214 
215 /* wrapper for rcu_assign_pointer */
216 static inline void node_set_parent(struct key_vector *n, struct key_vector *tp)
217 {
218 	if (n)
219 		rcu_assign_pointer(tn_info(n)->parent, tp);
220 }
221 
222 #define NODE_INIT_PARENT(n, p) RCU_INIT_POINTER(tn_info(n)->parent, p)
223 
224 /* This provides us with the number of children in this node, in the case of a
225  * leaf this will return 0 meaning none of the children are accessible.
226  */
227 static inline unsigned long child_length(const struct key_vector *tn)
228 {
229 	return (1ul << tn->bits) & ~(1ul);
230 }
231 
232 #define get_cindex(key, kv) (((key) ^ (kv)->key) >> (kv)->pos)
233 
234 static inline unsigned long get_index(t_key key, struct key_vector *kv)
235 {
236 	unsigned long index = key ^ kv->key;
237 
238 	if ((BITS_PER_LONG <= KEYLENGTH) && (KEYLENGTH == kv->pos))
239 		return 0;
240 
241 	return index >> kv->pos;
242 }
243 
244 /* To understand this stuff, an understanding of keys and all their bits is
245  * necessary. Every node in the trie has a key associated with it, but not
246  * all of the bits in that key are significant.
247  *
248  * Consider a node 'n' and its parent 'tp'.
249  *
250  * If n is a leaf, every bit in its key is significant. Its presence is
251  * necessitated by path compression, since during a tree traversal (when
252  * searching for a leaf - unless we are doing an insertion) we will completely
253  * ignore all skipped bits we encounter. Thus we need to verify, at the end of
254  * a potentially successful search, that we have indeed been walking the
255  * correct key path.
256  *
257  * Note that we can never "miss" the correct key in the tree if present by
258  * following the wrong path. Path compression ensures that segments of the key
259  * that are the same for all keys with a given prefix are skipped, but the
260  * skipped part *is* identical for each node in the subtrie below the skipped
261  * bit! trie_insert() in this implementation takes care of that.
262  *
263  * if n is an internal node - a 'tnode' here, the various parts of its key
264  * have many different meanings.
265  *
266  * Example:
267  * _________________________________________________________________
268  * | i | i | i | i | i | i | i | N | N | N | S | S | S | S | S | C |
269  * -----------------------------------------------------------------
270  *  31  30  29  28  27  26  25  24  23  22  21  20  19  18  17  16
271  *
272  * _________________________________________________________________
273  * | C | C | C | u | u | u | u | u | u | u | u | u | u | u | u | u |
274  * -----------------------------------------------------------------
275  *  15  14  13  12  11  10   9   8   7   6   5   4   3   2   1   0
276  *
277  * tp->pos = 22
278  * tp->bits = 3
279  * n->pos = 13
280  * n->bits = 4
281  *
282  * First, let's just ignore the bits that come before the parent tp, that is
283  * the bits from (tp->pos + tp->bits) to 31. They are *known* but at this
284  * point we do not use them for anything.
285  *
286  * The bits from (tp->pos) to (tp->pos + tp->bits - 1) - "N", above - are the
287  * index into the parent's child array. That is, they will be used to find
288  * 'n' among tp's children.
289  *
290  * The bits from (n->pos + n->bits) to (tp->pos - 1) - "S" - are skipped bits
291  * for the node n.
292  *
293  * All the bits we have seen so far are significant to the node n. The rest
294  * of the bits are really not needed or indeed known in n->key.
295  *
296  * The bits from (n->pos) to (n->pos + n->bits - 1) - "C" - are the index into
297  * n's child array, and will of course be different for each child.
298  *
299  * The rest of the bits, from 0 to (n->pos -1) - "u" - are completely unknown
300  * at this point.
301  */
302 
303 static const int halve_threshold = 25;
304 static const int inflate_threshold = 50;
305 static const int halve_threshold_root = 15;
306 static const int inflate_threshold_root = 30;
307 
308 static void __alias_free_mem(struct rcu_head *head)
309 {
310 	struct fib_alias *fa = container_of(head, struct fib_alias, rcu);
311 	kmem_cache_free(fn_alias_kmem, fa);
312 }
313 
314 static inline void alias_free_mem_rcu(struct fib_alias *fa)
315 {
316 	call_rcu(&fa->rcu, __alias_free_mem);
317 }
318 
319 #define TNODE_KMALLOC_MAX \
320 	ilog2((PAGE_SIZE - TNODE_SIZE(0)) / sizeof(struct key_vector *))
321 #define TNODE_VMALLOC_MAX \
322 	ilog2((SIZE_MAX - TNODE_SIZE(0)) / sizeof(struct key_vector *))
323 
324 static void __node_free_rcu(struct rcu_head *head)
325 {
326 	struct tnode *n = container_of(head, struct tnode, rcu);
327 
328 	if (!n->tn_bits)
329 		kmem_cache_free(trie_leaf_kmem, n);
330 	else
331 		kvfree(n);
332 }
333 
334 #define node_free(n) call_rcu(&tn_info(n)->rcu, __node_free_rcu)
335 
336 static struct tnode *tnode_alloc(int bits)
337 {
338 	size_t size;
339 
340 	/* verify bits is within bounds */
341 	if (bits > TNODE_VMALLOC_MAX)
342 		return NULL;
343 
344 	/* determine size and verify it is non-zero and didn't overflow */
345 	size = TNODE_SIZE(1ul << bits);
346 
347 	if (size <= PAGE_SIZE)
348 		return kzalloc(size, GFP_KERNEL);
349 	else
350 		return vzalloc(size);
351 }
352 
353 static inline void empty_child_inc(struct key_vector *n)
354 {
355 	++tn_info(n)->empty_children ? : ++tn_info(n)->full_children;
356 }
357 
358 static inline void empty_child_dec(struct key_vector *n)
359 {
360 	tn_info(n)->empty_children-- ? : tn_info(n)->full_children--;
361 }
362 
363 static struct key_vector *leaf_new(t_key key, struct fib_alias *fa)
364 {
365 	struct key_vector *l;
366 	struct tnode *kv;
367 
368 	kv = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
369 	if (!kv)
370 		return NULL;
371 
372 	/* initialize key vector */
373 	l = kv->kv;
374 	l->key = key;
375 	l->pos = 0;
376 	l->bits = 0;
377 	l->slen = fa->fa_slen;
378 
379 	/* link leaf to fib alias */
380 	INIT_HLIST_HEAD(&l->leaf);
381 	hlist_add_head(&fa->fa_list, &l->leaf);
382 
383 	return l;
384 }
385 
386 static struct key_vector *tnode_new(t_key key, int pos, int bits)
387 {
388 	unsigned int shift = pos + bits;
389 	struct key_vector *tn;
390 	struct tnode *tnode;
391 
392 	/* verify bits and pos their msb bits clear and values are valid */
393 	BUG_ON(!bits || (shift > KEYLENGTH));
394 
395 	tnode = tnode_alloc(bits);
396 	if (!tnode)
397 		return NULL;
398 
399 	pr_debug("AT %p s=%zu %zu\n", tnode, TNODE_SIZE(0),
400 		 sizeof(struct key_vector *) << bits);
401 
402 	if (bits == KEYLENGTH)
403 		tnode->full_children = 1;
404 	else
405 		tnode->empty_children = 1ul << bits;
406 
407 	tn = tnode->kv;
408 	tn->key = (shift < KEYLENGTH) ? (key >> shift) << shift : 0;
409 	tn->pos = pos;
410 	tn->bits = bits;
411 	tn->slen = pos;
412 
413 	return tn;
414 }
415 
416 /* Check whether a tnode 'n' is "full", i.e. it is an internal node
417  * and no bits are skipped. See discussion in dyntree paper p. 6
418  */
419 static inline int tnode_full(struct key_vector *tn, struct key_vector *n)
420 {
421 	return n && ((n->pos + n->bits) == tn->pos) && IS_TNODE(n);
422 }
423 
424 /* Add a child at position i overwriting the old value.
425  * Update the value of full_children and empty_children.
426  */
427 static void put_child(struct key_vector *tn, unsigned long i,
428 		      struct key_vector *n)
429 {
430 	struct key_vector *chi = get_child(tn, i);
431 	int isfull, wasfull;
432 
433 	BUG_ON(i >= child_length(tn));
434 
435 	/* update emptyChildren, overflow into fullChildren */
436 	if (!n && chi)
437 		empty_child_inc(tn);
438 	if (n && !chi)
439 		empty_child_dec(tn);
440 
441 	/* update fullChildren */
442 	wasfull = tnode_full(tn, chi);
443 	isfull = tnode_full(tn, n);
444 
445 	if (wasfull && !isfull)
446 		tn_info(tn)->full_children--;
447 	else if (!wasfull && isfull)
448 		tn_info(tn)->full_children++;
449 
450 	if (n && (tn->slen < n->slen))
451 		tn->slen = n->slen;
452 
453 	rcu_assign_pointer(tn->tnode[i], n);
454 }
455 
456 static void update_children(struct key_vector *tn)
457 {
458 	unsigned long i;
459 
460 	/* update all of the child parent pointers */
461 	for (i = child_length(tn); i;) {
462 		struct key_vector *inode = get_child(tn, --i);
463 
464 		if (!inode)
465 			continue;
466 
467 		/* Either update the children of a tnode that
468 		 * already belongs to us or update the child
469 		 * to point to ourselves.
470 		 */
471 		if (node_parent(inode) == tn)
472 			update_children(inode);
473 		else
474 			node_set_parent(inode, tn);
475 	}
476 }
477 
478 static inline void put_child_root(struct key_vector *tp, t_key key,
479 				  struct key_vector *n)
480 {
481 	if (IS_TRIE(tp))
482 		rcu_assign_pointer(tp->tnode[0], n);
483 	else
484 		put_child(tp, get_index(key, tp), n);
485 }
486 
487 static inline void tnode_free_init(struct key_vector *tn)
488 {
489 	tn_info(tn)->rcu.next = NULL;
490 }
491 
492 static inline void tnode_free_append(struct key_vector *tn,
493 				     struct key_vector *n)
494 {
495 	tn_info(n)->rcu.next = tn_info(tn)->rcu.next;
496 	tn_info(tn)->rcu.next = &tn_info(n)->rcu;
497 }
498 
499 static void tnode_free(struct key_vector *tn)
500 {
501 	struct callback_head *head = &tn_info(tn)->rcu;
502 
503 	while (head) {
504 		head = head->next;
505 		tnode_free_size += TNODE_SIZE(1ul << tn->bits);
506 		node_free(tn);
507 
508 		tn = container_of(head, struct tnode, rcu)->kv;
509 	}
510 
511 	if (tnode_free_size >= PAGE_SIZE * sync_pages) {
512 		tnode_free_size = 0;
513 		synchronize_rcu();
514 	}
515 }
516 
517 static struct key_vector *replace(struct trie *t,
518 				  struct key_vector *oldtnode,
519 				  struct key_vector *tn)
520 {
521 	struct key_vector *tp = node_parent(oldtnode);
522 	unsigned long i;
523 
524 	/* setup the parent pointer out of and back into this node */
525 	NODE_INIT_PARENT(tn, tp);
526 	put_child_root(tp, tn->key, tn);
527 
528 	/* update all of the child parent pointers */
529 	update_children(tn);
530 
531 	/* all pointers should be clean so we are done */
532 	tnode_free(oldtnode);
533 
534 	/* resize children now that oldtnode is freed */
535 	for (i = child_length(tn); i;) {
536 		struct key_vector *inode = get_child(tn, --i);
537 
538 		/* resize child node */
539 		if (tnode_full(tn, inode))
540 			tn = resize(t, inode);
541 	}
542 
543 	return tp;
544 }
545 
546 static struct key_vector *inflate(struct trie *t,
547 				  struct key_vector *oldtnode)
548 {
549 	struct key_vector *tn;
550 	unsigned long i;
551 	t_key m;
552 
553 	pr_debug("In inflate\n");
554 
555 	tn = tnode_new(oldtnode->key, oldtnode->pos - 1, oldtnode->bits + 1);
556 	if (!tn)
557 		goto notnode;
558 
559 	/* prepare oldtnode to be freed */
560 	tnode_free_init(oldtnode);
561 
562 	/* Assemble all of the pointers in our cluster, in this case that
563 	 * represents all of the pointers out of our allocated nodes that
564 	 * point to existing tnodes and the links between our allocated
565 	 * nodes.
566 	 */
567 	for (i = child_length(oldtnode), m = 1u << tn->pos; i;) {
568 		struct key_vector *inode = get_child(oldtnode, --i);
569 		struct key_vector *node0, *node1;
570 		unsigned long j, k;
571 
572 		/* An empty child */
573 		if (!inode)
574 			continue;
575 
576 		/* A leaf or an internal node with skipped bits */
577 		if (!tnode_full(oldtnode, inode)) {
578 			put_child(tn, get_index(inode->key, tn), inode);
579 			continue;
580 		}
581 
582 		/* drop the node in the old tnode free list */
583 		tnode_free_append(oldtnode, inode);
584 
585 		/* An internal node with two children */
586 		if (inode->bits == 1) {
587 			put_child(tn, 2 * i + 1, get_child(inode, 1));
588 			put_child(tn, 2 * i, get_child(inode, 0));
589 			continue;
590 		}
591 
592 		/* We will replace this node 'inode' with two new
593 		 * ones, 'node0' and 'node1', each with half of the
594 		 * original children. The two new nodes will have
595 		 * a position one bit further down the key and this
596 		 * means that the "significant" part of their keys
597 		 * (see the discussion near the top of this file)
598 		 * will differ by one bit, which will be "0" in
599 		 * node0's key and "1" in node1's key. Since we are
600 		 * moving the key position by one step, the bit that
601 		 * we are moving away from - the bit at position
602 		 * (tn->pos) - is the one that will differ between
603 		 * node0 and node1. So... we synthesize that bit in the
604 		 * two new keys.
605 		 */
606 		node1 = tnode_new(inode->key | m, inode->pos, inode->bits - 1);
607 		if (!node1)
608 			goto nomem;
609 		node0 = tnode_new(inode->key, inode->pos, inode->bits - 1);
610 
611 		tnode_free_append(tn, node1);
612 		if (!node0)
613 			goto nomem;
614 		tnode_free_append(tn, node0);
615 
616 		/* populate child pointers in new nodes */
617 		for (k = child_length(inode), j = k / 2; j;) {
618 			put_child(node1, --j, get_child(inode, --k));
619 			put_child(node0, j, get_child(inode, j));
620 			put_child(node1, --j, get_child(inode, --k));
621 			put_child(node0, j, get_child(inode, j));
622 		}
623 
624 		/* link new nodes to parent */
625 		NODE_INIT_PARENT(node1, tn);
626 		NODE_INIT_PARENT(node0, tn);
627 
628 		/* link parent to nodes */
629 		put_child(tn, 2 * i + 1, node1);
630 		put_child(tn, 2 * i, node0);
631 	}
632 
633 	/* setup the parent pointers into and out of this node */
634 	return replace(t, oldtnode, tn);
635 nomem:
636 	/* all pointers should be clean so we are done */
637 	tnode_free(tn);
638 notnode:
639 	return NULL;
640 }
641 
642 static struct key_vector *halve(struct trie *t,
643 				struct key_vector *oldtnode)
644 {
645 	struct key_vector *tn;
646 	unsigned long i;
647 
648 	pr_debug("In halve\n");
649 
650 	tn = tnode_new(oldtnode->key, oldtnode->pos + 1, oldtnode->bits - 1);
651 	if (!tn)
652 		goto notnode;
653 
654 	/* prepare oldtnode to be freed */
655 	tnode_free_init(oldtnode);
656 
657 	/* Assemble all of the pointers in our cluster, in this case that
658 	 * represents all of the pointers out of our allocated nodes that
659 	 * point to existing tnodes and the links between our allocated
660 	 * nodes.
661 	 */
662 	for (i = child_length(oldtnode); i;) {
663 		struct key_vector *node1 = get_child(oldtnode, --i);
664 		struct key_vector *node0 = get_child(oldtnode, --i);
665 		struct key_vector *inode;
666 
667 		/* At least one of the children is empty */
668 		if (!node1 || !node0) {
669 			put_child(tn, i / 2, node1 ? : node0);
670 			continue;
671 		}
672 
673 		/* Two nonempty children */
674 		inode = tnode_new(node0->key, oldtnode->pos, 1);
675 		if (!inode)
676 			goto nomem;
677 		tnode_free_append(tn, inode);
678 
679 		/* initialize pointers out of node */
680 		put_child(inode, 1, node1);
681 		put_child(inode, 0, node0);
682 		NODE_INIT_PARENT(inode, tn);
683 
684 		/* link parent to node */
685 		put_child(tn, i / 2, inode);
686 	}
687 
688 	/* setup the parent pointers into and out of this node */
689 	return replace(t, oldtnode, tn);
690 nomem:
691 	/* all pointers should be clean so we are done */
692 	tnode_free(tn);
693 notnode:
694 	return NULL;
695 }
696 
697 static struct key_vector *collapse(struct trie *t,
698 				   struct key_vector *oldtnode)
699 {
700 	struct key_vector *n, *tp;
701 	unsigned long i;
702 
703 	/* scan the tnode looking for that one child that might still exist */
704 	for (n = NULL, i = child_length(oldtnode); !n && i;)
705 		n = get_child(oldtnode, --i);
706 
707 	/* compress one level */
708 	tp = node_parent(oldtnode);
709 	put_child_root(tp, oldtnode->key, n);
710 	node_set_parent(n, tp);
711 
712 	/* drop dead node */
713 	node_free(oldtnode);
714 
715 	return tp;
716 }
717 
718 static unsigned char update_suffix(struct key_vector *tn)
719 {
720 	unsigned char slen = tn->pos;
721 	unsigned long stride, i;
722 
723 	/* search though the list of children looking for nodes that might
724 	 * have a suffix greater than the one we currently have.  This is
725 	 * why we start with a stride of 2 since a stride of 1 would
726 	 * represent the nodes with suffix length equal to tn->pos
727 	 */
728 	for (i = 0, stride = 0x2ul ; i < child_length(tn); i += stride) {
729 		struct key_vector *n = get_child(tn, i);
730 
731 		if (!n || (n->slen <= slen))
732 			continue;
733 
734 		/* update stride and slen based on new value */
735 		stride <<= (n->slen - slen);
736 		slen = n->slen;
737 		i &= ~(stride - 1);
738 
739 		/* if slen covers all but the last bit we can stop here
740 		 * there will be nothing longer than that since only node
741 		 * 0 and 1 << (bits - 1) could have that as their suffix
742 		 * length.
743 		 */
744 		if ((slen + 1) >= (tn->pos + tn->bits))
745 			break;
746 	}
747 
748 	tn->slen = slen;
749 
750 	return slen;
751 }
752 
753 /* From "Implementing a dynamic compressed trie" by Stefan Nilsson of
754  * the Helsinki University of Technology and Matti Tikkanen of Nokia
755  * Telecommunications, page 6:
756  * "A node is doubled if the ratio of non-empty children to all
757  * children in the *doubled* node is at least 'high'."
758  *
759  * 'high' in this instance is the variable 'inflate_threshold'. It
760  * is expressed as a percentage, so we multiply it with
761  * child_length() and instead of multiplying by 2 (since the
762  * child array will be doubled by inflate()) and multiplying
763  * the left-hand side by 100 (to handle the percentage thing) we
764  * multiply the left-hand side by 50.
765  *
766  * The left-hand side may look a bit weird: child_length(tn)
767  * - tn->empty_children is of course the number of non-null children
768  * in the current node. tn->full_children is the number of "full"
769  * children, that is non-null tnodes with a skip value of 0.
770  * All of those will be doubled in the resulting inflated tnode, so
771  * we just count them one extra time here.
772  *
773  * A clearer way to write this would be:
774  *
775  * to_be_doubled = tn->full_children;
776  * not_to_be_doubled = child_length(tn) - tn->empty_children -
777  *     tn->full_children;
778  *
779  * new_child_length = child_length(tn) * 2;
780  *
781  * new_fill_factor = 100 * (not_to_be_doubled + 2*to_be_doubled) /
782  *      new_child_length;
783  * if (new_fill_factor >= inflate_threshold)
784  *
785  * ...and so on, tho it would mess up the while () loop.
786  *
787  * anyway,
788  * 100 * (not_to_be_doubled + 2*to_be_doubled) / new_child_length >=
789  *      inflate_threshold
790  *
791  * avoid a division:
792  * 100 * (not_to_be_doubled + 2*to_be_doubled) >=
793  *      inflate_threshold * new_child_length
794  *
795  * expand not_to_be_doubled and to_be_doubled, and shorten:
796  * 100 * (child_length(tn) - tn->empty_children +
797  *    tn->full_children) >= inflate_threshold * new_child_length
798  *
799  * expand new_child_length:
800  * 100 * (child_length(tn) - tn->empty_children +
801  *    tn->full_children) >=
802  *      inflate_threshold * child_length(tn) * 2
803  *
804  * shorten again:
805  * 50 * (tn->full_children + child_length(tn) -
806  *    tn->empty_children) >= inflate_threshold *
807  *    child_length(tn)
808  *
809  */
810 static inline bool should_inflate(struct key_vector *tp, struct key_vector *tn)
811 {
812 	unsigned long used = child_length(tn);
813 	unsigned long threshold = used;
814 
815 	/* Keep root node larger */
816 	threshold *= IS_TRIE(tp) ? inflate_threshold_root : inflate_threshold;
817 	used -= tn_info(tn)->empty_children;
818 	used += tn_info(tn)->full_children;
819 
820 	/* if bits == KEYLENGTH then pos = 0, and will fail below */
821 
822 	return (used > 1) && tn->pos && ((50 * used) >= threshold);
823 }
824 
825 static inline bool should_halve(struct key_vector *tp, struct key_vector *tn)
826 {
827 	unsigned long used = child_length(tn);
828 	unsigned long threshold = used;
829 
830 	/* Keep root node larger */
831 	threshold *= IS_TRIE(tp) ? halve_threshold_root : halve_threshold;
832 	used -= tn_info(tn)->empty_children;
833 
834 	/* if bits == KEYLENGTH then used = 100% on wrap, and will fail below */
835 
836 	return (used > 1) && (tn->bits > 1) && ((100 * used) < threshold);
837 }
838 
839 static inline bool should_collapse(struct key_vector *tn)
840 {
841 	unsigned long used = child_length(tn);
842 
843 	used -= tn_info(tn)->empty_children;
844 
845 	/* account for bits == KEYLENGTH case */
846 	if ((tn->bits == KEYLENGTH) && tn_info(tn)->full_children)
847 		used -= KEY_MAX;
848 
849 	/* One child or none, time to drop us from the trie */
850 	return used < 2;
851 }
852 
853 #define MAX_WORK 10
854 static struct key_vector *resize(struct trie *t, struct key_vector *tn)
855 {
856 #ifdef CONFIG_IP_FIB_TRIE_STATS
857 	struct trie_use_stats __percpu *stats = t->stats;
858 #endif
859 	struct key_vector *tp = node_parent(tn);
860 	unsigned long cindex = get_index(tn->key, tp);
861 	int max_work = MAX_WORK;
862 
863 	pr_debug("In tnode_resize %p inflate_threshold=%d threshold=%d\n",
864 		 tn, inflate_threshold, halve_threshold);
865 
866 	/* track the tnode via the pointer from the parent instead of
867 	 * doing it ourselves.  This way we can let RCU fully do its
868 	 * thing without us interfering
869 	 */
870 	BUG_ON(tn != get_child(tp, cindex));
871 
872 	/* Double as long as the resulting node has a number of
873 	 * nonempty nodes that are above the threshold.
874 	 */
875 	while (should_inflate(tp, tn) && max_work) {
876 		tp = inflate(t, tn);
877 		if (!tp) {
878 #ifdef CONFIG_IP_FIB_TRIE_STATS
879 			this_cpu_inc(stats->resize_node_skipped);
880 #endif
881 			break;
882 		}
883 
884 		max_work--;
885 		tn = get_child(tp, cindex);
886 	}
887 
888 	/* update parent in case inflate failed */
889 	tp = node_parent(tn);
890 
891 	/* Return if at least one inflate is run */
892 	if (max_work != MAX_WORK)
893 		return tp;
894 
895 	/* Halve as long as the number of empty children in this
896 	 * node is above threshold.
897 	 */
898 	while (should_halve(tp, tn) && max_work) {
899 		tp = halve(t, tn);
900 		if (!tp) {
901 #ifdef CONFIG_IP_FIB_TRIE_STATS
902 			this_cpu_inc(stats->resize_node_skipped);
903 #endif
904 			break;
905 		}
906 
907 		max_work--;
908 		tn = get_child(tp, cindex);
909 	}
910 
911 	/* Only one child remains */
912 	if (should_collapse(tn))
913 		return collapse(t, tn);
914 
915 	/* update parent in case halve failed */
916 	tp = node_parent(tn);
917 
918 	/* Return if at least one deflate was run */
919 	if (max_work != MAX_WORK)
920 		return tp;
921 
922 	/* push the suffix length to the parent node */
923 	if (tn->slen > tn->pos) {
924 		unsigned char slen = update_suffix(tn);
925 
926 		if (slen > tp->slen)
927 			tp->slen = slen;
928 	}
929 
930 	return tp;
931 }
932 
933 static void leaf_pull_suffix(struct key_vector *tp, struct key_vector *l)
934 {
935 	while ((tp->slen > tp->pos) && (tp->slen > l->slen)) {
936 		if (update_suffix(tp) > l->slen)
937 			break;
938 		tp = node_parent(tp);
939 	}
940 }
941 
942 static void leaf_push_suffix(struct key_vector *tn, struct key_vector *l)
943 {
944 	/* if this is a new leaf then tn will be NULL and we can sort
945 	 * out parent suffix lengths as a part of trie_rebalance
946 	 */
947 	while (tn->slen < l->slen) {
948 		tn->slen = l->slen;
949 		tn = node_parent(tn);
950 	}
951 }
952 
953 /* rcu_read_lock needs to be hold by caller from readside */
954 static struct key_vector *fib_find_node(struct trie *t,
955 					struct key_vector **tp, u32 key)
956 {
957 	struct key_vector *pn, *n = t->kv;
958 	unsigned long index = 0;
959 
960 	do {
961 		pn = n;
962 		n = get_child_rcu(n, index);
963 
964 		if (!n)
965 			break;
966 
967 		index = get_cindex(key, n);
968 
969 		/* This bit of code is a bit tricky but it combines multiple
970 		 * checks into a single check.  The prefix consists of the
971 		 * prefix plus zeros for the bits in the cindex. The index
972 		 * is the difference between the key and this value.  From
973 		 * this we can actually derive several pieces of data.
974 		 *   if (index >= (1ul << bits))
975 		 *     we have a mismatch in skip bits and failed
976 		 *   else
977 		 *     we know the value is cindex
978 		 *
979 		 * This check is safe even if bits == KEYLENGTH due to the
980 		 * fact that we can only allocate a node with 32 bits if a
981 		 * long is greater than 32 bits.
982 		 */
983 		if (index >= (1ul << n->bits)) {
984 			n = NULL;
985 			break;
986 		}
987 
988 		/* keep searching until we find a perfect match leaf or NULL */
989 	} while (IS_TNODE(n));
990 
991 	*tp = pn;
992 
993 	return n;
994 }
995 
996 /* Return the first fib alias matching TOS with
997  * priority less than or equal to PRIO.
998  */
999 static struct fib_alias *fib_find_alias(struct hlist_head *fah, u8 slen,
1000 					u8 tos, u32 prio, u32 tb_id)
1001 {
1002 	struct fib_alias *fa;
1003 
1004 	if (!fah)
1005 		return NULL;
1006 
1007 	hlist_for_each_entry(fa, fah, fa_list) {
1008 		if (fa->fa_slen < slen)
1009 			continue;
1010 		if (fa->fa_slen != slen)
1011 			break;
1012 		if (fa->tb_id > tb_id)
1013 			continue;
1014 		if (fa->tb_id != tb_id)
1015 			break;
1016 		if (fa->fa_tos > tos)
1017 			continue;
1018 		if (fa->fa_info->fib_priority >= prio || fa->fa_tos < tos)
1019 			return fa;
1020 	}
1021 
1022 	return NULL;
1023 }
1024 
1025 static void trie_rebalance(struct trie *t, struct key_vector *tn)
1026 {
1027 	while (!IS_TRIE(tn))
1028 		tn = resize(t, tn);
1029 }
1030 
1031 static int fib_insert_node(struct trie *t, struct key_vector *tp,
1032 			   struct fib_alias *new, t_key key)
1033 {
1034 	struct key_vector *n, *l;
1035 
1036 	l = leaf_new(key, new);
1037 	if (!l)
1038 		goto noleaf;
1039 
1040 	/* retrieve child from parent node */
1041 	n = get_child(tp, get_index(key, tp));
1042 
1043 	/* Case 2: n is a LEAF or a TNODE and the key doesn't match.
1044 	 *
1045 	 *  Add a new tnode here
1046 	 *  first tnode need some special handling
1047 	 *  leaves us in position for handling as case 3
1048 	 */
1049 	if (n) {
1050 		struct key_vector *tn;
1051 
1052 		tn = tnode_new(key, __fls(key ^ n->key), 1);
1053 		if (!tn)
1054 			goto notnode;
1055 
1056 		/* initialize routes out of node */
1057 		NODE_INIT_PARENT(tn, tp);
1058 		put_child(tn, get_index(key, tn) ^ 1, n);
1059 
1060 		/* start adding routes into the node */
1061 		put_child_root(tp, key, tn);
1062 		node_set_parent(n, tn);
1063 
1064 		/* parent now has a NULL spot where the leaf can go */
1065 		tp = tn;
1066 	}
1067 
1068 	/* Case 3: n is NULL, and will just insert a new leaf */
1069 	NODE_INIT_PARENT(l, tp);
1070 	put_child_root(tp, key, l);
1071 	trie_rebalance(t, tp);
1072 
1073 	return 0;
1074 notnode:
1075 	node_free(l);
1076 noleaf:
1077 	return -ENOMEM;
1078 }
1079 
1080 static int fib_insert_alias(struct trie *t, struct key_vector *tp,
1081 			    struct key_vector *l, struct fib_alias *new,
1082 			    struct fib_alias *fa, t_key key)
1083 {
1084 	if (!l)
1085 		return fib_insert_node(t, tp, new, key);
1086 
1087 	if (fa) {
1088 		hlist_add_before_rcu(&new->fa_list, &fa->fa_list);
1089 	} else {
1090 		struct fib_alias *last;
1091 
1092 		hlist_for_each_entry(last, &l->leaf, fa_list) {
1093 			if (new->fa_slen < last->fa_slen)
1094 				break;
1095 			if ((new->fa_slen == last->fa_slen) &&
1096 			    (new->tb_id > last->tb_id))
1097 				break;
1098 			fa = last;
1099 		}
1100 
1101 		if (fa)
1102 			hlist_add_behind_rcu(&new->fa_list, &fa->fa_list);
1103 		else
1104 			hlist_add_head_rcu(&new->fa_list, &l->leaf);
1105 	}
1106 
1107 	/* if we added to the tail node then we need to update slen */
1108 	if (l->slen < new->fa_slen) {
1109 		l->slen = new->fa_slen;
1110 		leaf_push_suffix(tp, l);
1111 	}
1112 
1113 	return 0;
1114 }
1115 
1116 /* Caller must hold RTNL. */
1117 int fib_table_insert(struct net *net, struct fib_table *tb,
1118 		     struct fib_config *cfg)
1119 {
1120 	struct trie *t = (struct trie *)tb->tb_data;
1121 	struct fib_alias *fa, *new_fa;
1122 	struct key_vector *l, *tp;
1123 	u16 nlflags = NLM_F_EXCL;
1124 	struct fib_info *fi;
1125 	u8 plen = cfg->fc_dst_len;
1126 	u8 slen = KEYLENGTH - plen;
1127 	u8 tos = cfg->fc_tos;
1128 	u32 key;
1129 	int err;
1130 
1131 	if (plen > KEYLENGTH)
1132 		return -EINVAL;
1133 
1134 	key = ntohl(cfg->fc_dst);
1135 
1136 	pr_debug("Insert table=%u %08x/%d\n", tb->tb_id, key, plen);
1137 
1138 	if ((plen < KEYLENGTH) && (key << plen))
1139 		return -EINVAL;
1140 
1141 	fi = fib_create_info(cfg);
1142 	if (IS_ERR(fi)) {
1143 		err = PTR_ERR(fi);
1144 		goto err;
1145 	}
1146 
1147 	l = fib_find_node(t, &tp, key);
1148 	fa = l ? fib_find_alias(&l->leaf, slen, tos, fi->fib_priority,
1149 				tb->tb_id) : NULL;
1150 
1151 	/* Now fa, if non-NULL, points to the first fib alias
1152 	 * with the same keys [prefix,tos,priority], if such key already
1153 	 * exists or to the node before which we will insert new one.
1154 	 *
1155 	 * If fa is NULL, we will need to allocate a new one and
1156 	 * insert to the tail of the section matching the suffix length
1157 	 * of the new alias.
1158 	 */
1159 
1160 	if (fa && fa->fa_tos == tos &&
1161 	    fa->fa_info->fib_priority == fi->fib_priority) {
1162 		struct fib_alias *fa_first, *fa_match;
1163 
1164 		err = -EEXIST;
1165 		if (cfg->fc_nlflags & NLM_F_EXCL)
1166 			goto out;
1167 
1168 		nlflags &= ~NLM_F_EXCL;
1169 
1170 		/* We have 2 goals:
1171 		 * 1. Find exact match for type, scope, fib_info to avoid
1172 		 * duplicate routes
1173 		 * 2. Find next 'fa' (or head), NLM_F_APPEND inserts before it
1174 		 */
1175 		fa_match = NULL;
1176 		fa_first = fa;
1177 		hlist_for_each_entry_from(fa, fa_list) {
1178 			if ((fa->fa_slen != slen) ||
1179 			    (fa->tb_id != tb->tb_id) ||
1180 			    (fa->fa_tos != tos))
1181 				break;
1182 			if (fa->fa_info->fib_priority != fi->fib_priority)
1183 				break;
1184 			if (fa->fa_type == cfg->fc_type &&
1185 			    fa->fa_info == fi) {
1186 				fa_match = fa;
1187 				break;
1188 			}
1189 		}
1190 
1191 		if (cfg->fc_nlflags & NLM_F_REPLACE) {
1192 			struct fib_info *fi_drop;
1193 			u8 state;
1194 
1195 			nlflags |= NLM_F_REPLACE;
1196 			fa = fa_first;
1197 			if (fa_match) {
1198 				if (fa == fa_match)
1199 					err = 0;
1200 				goto out;
1201 			}
1202 			err = -ENOBUFS;
1203 			new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1204 			if (!new_fa)
1205 				goto out;
1206 
1207 			fi_drop = fa->fa_info;
1208 			new_fa->fa_tos = fa->fa_tos;
1209 			new_fa->fa_info = fi;
1210 			new_fa->fa_type = cfg->fc_type;
1211 			state = fa->fa_state;
1212 			new_fa->fa_state = state & ~FA_S_ACCESSED;
1213 			new_fa->fa_slen = fa->fa_slen;
1214 			new_fa->tb_id = tb->tb_id;
1215 			new_fa->fa_default = -1;
1216 
1217 			hlist_replace_rcu(&fa->fa_list, &new_fa->fa_list);
1218 
1219 			alias_free_mem_rcu(fa);
1220 
1221 			fib_release_info(fi_drop);
1222 			if (state & FA_S_ACCESSED)
1223 				rt_cache_flush(cfg->fc_nlinfo.nl_net);
1224 
1225 			call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_ADD,
1226 						 key, plen, fi,
1227 						 new_fa->fa_tos, cfg->fc_type,
1228 						 tb->tb_id, cfg->fc_nlflags);
1229 			rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen,
1230 				tb->tb_id, &cfg->fc_nlinfo, nlflags);
1231 
1232 			goto succeeded;
1233 		}
1234 		/* Error if we find a perfect match which
1235 		 * uses the same scope, type, and nexthop
1236 		 * information.
1237 		 */
1238 		if (fa_match)
1239 			goto out;
1240 
1241 		if (cfg->fc_nlflags & NLM_F_APPEND)
1242 			nlflags |= NLM_F_APPEND;
1243 		else
1244 			fa = fa_first;
1245 	}
1246 	err = -ENOENT;
1247 	if (!(cfg->fc_nlflags & NLM_F_CREATE))
1248 		goto out;
1249 
1250 	nlflags |= NLM_F_CREATE;
1251 	err = -ENOBUFS;
1252 	new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1253 	if (!new_fa)
1254 		goto out;
1255 
1256 	new_fa->fa_info = fi;
1257 	new_fa->fa_tos = tos;
1258 	new_fa->fa_type = cfg->fc_type;
1259 	new_fa->fa_state = 0;
1260 	new_fa->fa_slen = slen;
1261 	new_fa->tb_id = tb->tb_id;
1262 	new_fa->fa_default = -1;
1263 
1264 	/* Insert new entry to the list. */
1265 	err = fib_insert_alias(t, tp, l, new_fa, fa, key);
1266 	if (err)
1267 		goto out_free_new_fa;
1268 
1269 	if (!plen)
1270 		tb->tb_num_default++;
1271 
1272 	rt_cache_flush(cfg->fc_nlinfo.nl_net);
1273 	call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_ADD, key, plen, fi, tos,
1274 				 cfg->fc_type, tb->tb_id, cfg->fc_nlflags);
1275 	rtmsg_fib(RTM_NEWROUTE, htonl(key), new_fa, plen, new_fa->tb_id,
1276 		  &cfg->fc_nlinfo, nlflags);
1277 succeeded:
1278 	return 0;
1279 
1280 out_free_new_fa:
1281 	kmem_cache_free(fn_alias_kmem, new_fa);
1282 out:
1283 	fib_release_info(fi);
1284 err:
1285 	return err;
1286 }
1287 
1288 static inline t_key prefix_mismatch(t_key key, struct key_vector *n)
1289 {
1290 	t_key prefix = n->key;
1291 
1292 	return (key ^ prefix) & (prefix | -prefix);
1293 }
1294 
1295 /* should be called with rcu_read_lock */
1296 int fib_table_lookup(struct fib_table *tb, const struct flowi4 *flp,
1297 		     struct fib_result *res, int fib_flags)
1298 {
1299 	struct trie *t = (struct trie *) tb->tb_data;
1300 #ifdef CONFIG_IP_FIB_TRIE_STATS
1301 	struct trie_use_stats __percpu *stats = t->stats;
1302 #endif
1303 	const t_key key = ntohl(flp->daddr);
1304 	struct key_vector *n, *pn;
1305 	struct fib_alias *fa;
1306 	unsigned long index;
1307 	t_key cindex;
1308 
1309 	trace_fib_table_lookup(tb->tb_id, flp);
1310 
1311 	pn = t->kv;
1312 	cindex = 0;
1313 
1314 	n = get_child_rcu(pn, cindex);
1315 	if (!n)
1316 		return -EAGAIN;
1317 
1318 #ifdef CONFIG_IP_FIB_TRIE_STATS
1319 	this_cpu_inc(stats->gets);
1320 #endif
1321 
1322 	/* Step 1: Travel to the longest prefix match in the trie */
1323 	for (;;) {
1324 		index = get_cindex(key, n);
1325 
1326 		/* This bit of code is a bit tricky but it combines multiple
1327 		 * checks into a single check.  The prefix consists of the
1328 		 * prefix plus zeros for the "bits" in the prefix. The index
1329 		 * is the difference between the key and this value.  From
1330 		 * this we can actually derive several pieces of data.
1331 		 *   if (index >= (1ul << bits))
1332 		 *     we have a mismatch in skip bits and failed
1333 		 *   else
1334 		 *     we know the value is cindex
1335 		 *
1336 		 * This check is safe even if bits == KEYLENGTH due to the
1337 		 * fact that we can only allocate a node with 32 bits if a
1338 		 * long is greater than 32 bits.
1339 		 */
1340 		if (index >= (1ul << n->bits))
1341 			break;
1342 
1343 		/* we have found a leaf. Prefixes have already been compared */
1344 		if (IS_LEAF(n))
1345 			goto found;
1346 
1347 		/* only record pn and cindex if we are going to be chopping
1348 		 * bits later.  Otherwise we are just wasting cycles.
1349 		 */
1350 		if (n->slen > n->pos) {
1351 			pn = n;
1352 			cindex = index;
1353 		}
1354 
1355 		n = get_child_rcu(n, index);
1356 		if (unlikely(!n))
1357 			goto backtrace;
1358 	}
1359 
1360 	/* Step 2: Sort out leaves and begin backtracing for longest prefix */
1361 	for (;;) {
1362 		/* record the pointer where our next node pointer is stored */
1363 		struct key_vector __rcu **cptr = n->tnode;
1364 
1365 		/* This test verifies that none of the bits that differ
1366 		 * between the key and the prefix exist in the region of
1367 		 * the lsb and higher in the prefix.
1368 		 */
1369 		if (unlikely(prefix_mismatch(key, n)) || (n->slen == n->pos))
1370 			goto backtrace;
1371 
1372 		/* exit out and process leaf */
1373 		if (unlikely(IS_LEAF(n)))
1374 			break;
1375 
1376 		/* Don't bother recording parent info.  Since we are in
1377 		 * prefix match mode we will have to come back to wherever
1378 		 * we started this traversal anyway
1379 		 */
1380 
1381 		while ((n = rcu_dereference(*cptr)) == NULL) {
1382 backtrace:
1383 #ifdef CONFIG_IP_FIB_TRIE_STATS
1384 			if (!n)
1385 				this_cpu_inc(stats->null_node_hit);
1386 #endif
1387 			/* If we are at cindex 0 there are no more bits for
1388 			 * us to strip at this level so we must ascend back
1389 			 * up one level to see if there are any more bits to
1390 			 * be stripped there.
1391 			 */
1392 			while (!cindex) {
1393 				t_key pkey = pn->key;
1394 
1395 				/* If we don't have a parent then there is
1396 				 * nothing for us to do as we do not have any
1397 				 * further nodes to parse.
1398 				 */
1399 				if (IS_TRIE(pn))
1400 					return -EAGAIN;
1401 #ifdef CONFIG_IP_FIB_TRIE_STATS
1402 				this_cpu_inc(stats->backtrack);
1403 #endif
1404 				/* Get Child's index */
1405 				pn = node_parent_rcu(pn);
1406 				cindex = get_index(pkey, pn);
1407 			}
1408 
1409 			/* strip the least significant bit from the cindex */
1410 			cindex &= cindex - 1;
1411 
1412 			/* grab pointer for next child node */
1413 			cptr = &pn->tnode[cindex];
1414 		}
1415 	}
1416 
1417 found:
1418 	/* this line carries forward the xor from earlier in the function */
1419 	index = key ^ n->key;
1420 
1421 	/* Step 3: Process the leaf, if that fails fall back to backtracing */
1422 	hlist_for_each_entry_rcu(fa, &n->leaf, fa_list) {
1423 		struct fib_info *fi = fa->fa_info;
1424 		int nhsel, err;
1425 
1426 		if ((BITS_PER_LONG > KEYLENGTH) || (fa->fa_slen < KEYLENGTH)) {
1427 			if (index >= (1ul << fa->fa_slen))
1428 				continue;
1429 		}
1430 		if (fa->fa_tos && fa->fa_tos != flp->flowi4_tos)
1431 			continue;
1432 		if (fi->fib_dead)
1433 			continue;
1434 		if (fa->fa_info->fib_scope < flp->flowi4_scope)
1435 			continue;
1436 		fib_alias_accessed(fa);
1437 		err = fib_props[fa->fa_type].error;
1438 		if (unlikely(err < 0)) {
1439 #ifdef CONFIG_IP_FIB_TRIE_STATS
1440 			this_cpu_inc(stats->semantic_match_passed);
1441 #endif
1442 			return err;
1443 		}
1444 		if (fi->fib_flags & RTNH_F_DEAD)
1445 			continue;
1446 		for (nhsel = 0; nhsel < fi->fib_nhs; nhsel++) {
1447 			const struct fib_nh *nh = &fi->fib_nh[nhsel];
1448 			struct in_device *in_dev = __in_dev_get_rcu(nh->nh_dev);
1449 
1450 			if (nh->nh_flags & RTNH_F_DEAD)
1451 				continue;
1452 			if (in_dev &&
1453 			    IN_DEV_IGNORE_ROUTES_WITH_LINKDOWN(in_dev) &&
1454 			    nh->nh_flags & RTNH_F_LINKDOWN &&
1455 			    !(fib_flags & FIB_LOOKUP_IGNORE_LINKSTATE))
1456 				continue;
1457 			if (!(flp->flowi4_flags & FLOWI_FLAG_SKIP_NH_OIF)) {
1458 				if (flp->flowi4_oif &&
1459 				    flp->flowi4_oif != nh->nh_oif)
1460 					continue;
1461 			}
1462 
1463 			if (!(fib_flags & FIB_LOOKUP_NOREF))
1464 				atomic_inc(&fi->fib_clntref);
1465 
1466 			res->prefixlen = KEYLENGTH - fa->fa_slen;
1467 			res->nh_sel = nhsel;
1468 			res->type = fa->fa_type;
1469 			res->scope = fi->fib_scope;
1470 			res->fi = fi;
1471 			res->table = tb;
1472 			res->fa_head = &n->leaf;
1473 #ifdef CONFIG_IP_FIB_TRIE_STATS
1474 			this_cpu_inc(stats->semantic_match_passed);
1475 #endif
1476 			trace_fib_table_lookup_nh(nh);
1477 
1478 			return err;
1479 		}
1480 	}
1481 #ifdef CONFIG_IP_FIB_TRIE_STATS
1482 	this_cpu_inc(stats->semantic_match_miss);
1483 #endif
1484 	goto backtrace;
1485 }
1486 EXPORT_SYMBOL_GPL(fib_table_lookup);
1487 
1488 static void fib_remove_alias(struct trie *t, struct key_vector *tp,
1489 			     struct key_vector *l, struct fib_alias *old)
1490 {
1491 	/* record the location of the previous list_info entry */
1492 	struct hlist_node **pprev = old->fa_list.pprev;
1493 	struct fib_alias *fa = hlist_entry(pprev, typeof(*fa), fa_list.next);
1494 
1495 	/* remove the fib_alias from the list */
1496 	hlist_del_rcu(&old->fa_list);
1497 
1498 	/* if we emptied the list this leaf will be freed and we can sort
1499 	 * out parent suffix lengths as a part of trie_rebalance
1500 	 */
1501 	if (hlist_empty(&l->leaf)) {
1502 		put_child_root(tp, l->key, NULL);
1503 		node_free(l);
1504 		trie_rebalance(t, tp);
1505 		return;
1506 	}
1507 
1508 	/* only access fa if it is pointing at the last valid hlist_node */
1509 	if (*pprev)
1510 		return;
1511 
1512 	/* update the trie with the latest suffix length */
1513 	l->slen = fa->fa_slen;
1514 	leaf_pull_suffix(tp, l);
1515 }
1516 
1517 /* Caller must hold RTNL. */
1518 int fib_table_delete(struct net *net, struct fib_table *tb,
1519 		     struct fib_config *cfg)
1520 {
1521 	struct trie *t = (struct trie *) tb->tb_data;
1522 	struct fib_alias *fa, *fa_to_delete;
1523 	struct key_vector *l, *tp;
1524 	u8 plen = cfg->fc_dst_len;
1525 	u8 slen = KEYLENGTH - plen;
1526 	u8 tos = cfg->fc_tos;
1527 	u32 key;
1528 
1529 	if (plen > KEYLENGTH)
1530 		return -EINVAL;
1531 
1532 	key = ntohl(cfg->fc_dst);
1533 
1534 	if ((plen < KEYLENGTH) && (key << plen))
1535 		return -EINVAL;
1536 
1537 	l = fib_find_node(t, &tp, key);
1538 	if (!l)
1539 		return -ESRCH;
1540 
1541 	fa = fib_find_alias(&l->leaf, slen, tos, 0, tb->tb_id);
1542 	if (!fa)
1543 		return -ESRCH;
1544 
1545 	pr_debug("Deleting %08x/%d tos=%d t=%p\n", key, plen, tos, t);
1546 
1547 	fa_to_delete = NULL;
1548 	hlist_for_each_entry_from(fa, fa_list) {
1549 		struct fib_info *fi = fa->fa_info;
1550 
1551 		if ((fa->fa_slen != slen) ||
1552 		    (fa->tb_id != tb->tb_id) ||
1553 		    (fa->fa_tos != tos))
1554 			break;
1555 
1556 		if ((!cfg->fc_type || fa->fa_type == cfg->fc_type) &&
1557 		    (cfg->fc_scope == RT_SCOPE_NOWHERE ||
1558 		     fa->fa_info->fib_scope == cfg->fc_scope) &&
1559 		    (!cfg->fc_prefsrc ||
1560 		     fi->fib_prefsrc == cfg->fc_prefsrc) &&
1561 		    (!cfg->fc_protocol ||
1562 		     fi->fib_protocol == cfg->fc_protocol) &&
1563 		    fib_nh_match(cfg, fi) == 0) {
1564 			fa_to_delete = fa;
1565 			break;
1566 		}
1567 	}
1568 
1569 	if (!fa_to_delete)
1570 		return -ESRCH;
1571 
1572 	call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, key, plen,
1573 				 fa_to_delete->fa_info, tos, cfg->fc_type,
1574 				 tb->tb_id, 0);
1575 	rtmsg_fib(RTM_DELROUTE, htonl(key), fa_to_delete, plen, tb->tb_id,
1576 		  &cfg->fc_nlinfo, 0);
1577 
1578 	if (!plen)
1579 		tb->tb_num_default--;
1580 
1581 	fib_remove_alias(t, tp, l, fa_to_delete);
1582 
1583 	if (fa_to_delete->fa_state & FA_S_ACCESSED)
1584 		rt_cache_flush(cfg->fc_nlinfo.nl_net);
1585 
1586 	fib_release_info(fa_to_delete->fa_info);
1587 	alias_free_mem_rcu(fa_to_delete);
1588 	return 0;
1589 }
1590 
1591 /* Scan for the next leaf starting at the provided key value */
1592 static struct key_vector *leaf_walk_rcu(struct key_vector **tn, t_key key)
1593 {
1594 	struct key_vector *pn, *n = *tn;
1595 	unsigned long cindex;
1596 
1597 	/* this loop is meant to try and find the key in the trie */
1598 	do {
1599 		/* record parent and next child index */
1600 		pn = n;
1601 		cindex = (key > pn->key) ? get_index(key, pn) : 0;
1602 
1603 		if (cindex >> pn->bits)
1604 			break;
1605 
1606 		/* descend into the next child */
1607 		n = get_child_rcu(pn, cindex++);
1608 		if (!n)
1609 			break;
1610 
1611 		/* guarantee forward progress on the keys */
1612 		if (IS_LEAF(n) && (n->key >= key))
1613 			goto found;
1614 	} while (IS_TNODE(n));
1615 
1616 	/* this loop will search for the next leaf with a greater key */
1617 	while (!IS_TRIE(pn)) {
1618 		/* if we exhausted the parent node we will need to climb */
1619 		if (cindex >= (1ul << pn->bits)) {
1620 			t_key pkey = pn->key;
1621 
1622 			pn = node_parent_rcu(pn);
1623 			cindex = get_index(pkey, pn) + 1;
1624 			continue;
1625 		}
1626 
1627 		/* grab the next available node */
1628 		n = get_child_rcu(pn, cindex++);
1629 		if (!n)
1630 			continue;
1631 
1632 		/* no need to compare keys since we bumped the index */
1633 		if (IS_LEAF(n))
1634 			goto found;
1635 
1636 		/* Rescan start scanning in new node */
1637 		pn = n;
1638 		cindex = 0;
1639 	}
1640 
1641 	*tn = pn;
1642 	return NULL; /* Root of trie */
1643 found:
1644 	/* if we are at the limit for keys just return NULL for the tnode */
1645 	*tn = pn;
1646 	return n;
1647 }
1648 
1649 static void fib_trie_free(struct fib_table *tb)
1650 {
1651 	struct trie *t = (struct trie *)tb->tb_data;
1652 	struct key_vector *pn = t->kv;
1653 	unsigned long cindex = 1;
1654 	struct hlist_node *tmp;
1655 	struct fib_alias *fa;
1656 
1657 	/* walk trie in reverse order and free everything */
1658 	for (;;) {
1659 		struct key_vector *n;
1660 
1661 		if (!(cindex--)) {
1662 			t_key pkey = pn->key;
1663 
1664 			if (IS_TRIE(pn))
1665 				break;
1666 
1667 			n = pn;
1668 			pn = node_parent(pn);
1669 
1670 			/* drop emptied tnode */
1671 			put_child_root(pn, n->key, NULL);
1672 			node_free(n);
1673 
1674 			cindex = get_index(pkey, pn);
1675 
1676 			continue;
1677 		}
1678 
1679 		/* grab the next available node */
1680 		n = get_child(pn, cindex);
1681 		if (!n)
1682 			continue;
1683 
1684 		if (IS_TNODE(n)) {
1685 			/* record pn and cindex for leaf walking */
1686 			pn = n;
1687 			cindex = 1ul << n->bits;
1688 
1689 			continue;
1690 		}
1691 
1692 		hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1693 			hlist_del_rcu(&fa->fa_list);
1694 			alias_free_mem_rcu(fa);
1695 		}
1696 
1697 		put_child_root(pn, n->key, NULL);
1698 		node_free(n);
1699 	}
1700 
1701 #ifdef CONFIG_IP_FIB_TRIE_STATS
1702 	free_percpu(t->stats);
1703 #endif
1704 	kfree(tb);
1705 }
1706 
1707 struct fib_table *fib_trie_unmerge(struct fib_table *oldtb)
1708 {
1709 	struct trie *ot = (struct trie *)oldtb->tb_data;
1710 	struct key_vector *l, *tp = ot->kv;
1711 	struct fib_table *local_tb;
1712 	struct fib_alias *fa;
1713 	struct trie *lt;
1714 	t_key key = 0;
1715 
1716 	if (oldtb->tb_data == oldtb->__data)
1717 		return oldtb;
1718 
1719 	local_tb = fib_trie_table(RT_TABLE_LOCAL, NULL);
1720 	if (!local_tb)
1721 		return NULL;
1722 
1723 	lt = (struct trie *)local_tb->tb_data;
1724 
1725 	while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
1726 		struct key_vector *local_l = NULL, *local_tp;
1727 
1728 		hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
1729 			struct fib_alias *new_fa;
1730 
1731 			if (local_tb->tb_id != fa->tb_id)
1732 				continue;
1733 
1734 			/* clone fa for new local table */
1735 			new_fa = kmem_cache_alloc(fn_alias_kmem, GFP_KERNEL);
1736 			if (!new_fa)
1737 				goto out;
1738 
1739 			memcpy(new_fa, fa, sizeof(*fa));
1740 
1741 			/* insert clone into table */
1742 			if (!local_l)
1743 				local_l = fib_find_node(lt, &local_tp, l->key);
1744 
1745 			if (fib_insert_alias(lt, local_tp, local_l, new_fa,
1746 					     NULL, l->key))
1747 				goto out;
1748 		}
1749 
1750 		/* stop loop if key wrapped back to 0 */
1751 		key = l->key + 1;
1752 		if (key < l->key)
1753 			break;
1754 	}
1755 
1756 	return local_tb;
1757 out:
1758 	fib_trie_free(local_tb);
1759 
1760 	return NULL;
1761 }
1762 
1763 /* Caller must hold RTNL. */
1764 int fib_table_flush(struct net *net, struct fib_table *tb)
1765 {
1766 	struct trie *t = (struct trie *)tb->tb_data;
1767 	struct key_vector *pn = t->kv;
1768 	unsigned long cindex = 1;
1769 	struct hlist_node *tmp;
1770 	struct fib_alias *fa;
1771 	int found = 0;
1772 
1773 	/* walk trie in reverse order */
1774 	for (;;) {
1775 		unsigned char slen = 0;
1776 		struct key_vector *n;
1777 
1778 		if (!(cindex--)) {
1779 			t_key pkey = pn->key;
1780 
1781 			/* cannot resize the trie vector */
1782 			if (IS_TRIE(pn))
1783 				break;
1784 
1785 			/* resize completed node */
1786 			pn = resize(t, pn);
1787 			cindex = get_index(pkey, pn);
1788 
1789 			continue;
1790 		}
1791 
1792 		/* grab the next available node */
1793 		n = get_child(pn, cindex);
1794 		if (!n)
1795 			continue;
1796 
1797 		if (IS_TNODE(n)) {
1798 			/* record pn and cindex for leaf walking */
1799 			pn = n;
1800 			cindex = 1ul << n->bits;
1801 
1802 			continue;
1803 		}
1804 
1805 		hlist_for_each_entry_safe(fa, tmp, &n->leaf, fa_list) {
1806 			struct fib_info *fi = fa->fa_info;
1807 
1808 			if (!fi || !(fi->fib_flags & RTNH_F_DEAD)) {
1809 				slen = fa->fa_slen;
1810 				continue;
1811 			}
1812 
1813 			call_fib_entry_notifiers(net, FIB_EVENT_ENTRY_DEL,
1814 						 n->key,
1815 						 KEYLENGTH - fa->fa_slen,
1816 						 fi, fa->fa_tos, fa->fa_type,
1817 						 tb->tb_id, 0);
1818 			hlist_del_rcu(&fa->fa_list);
1819 			fib_release_info(fa->fa_info);
1820 			alias_free_mem_rcu(fa);
1821 			found++;
1822 		}
1823 
1824 		/* update leaf slen */
1825 		n->slen = slen;
1826 
1827 		if (hlist_empty(&n->leaf)) {
1828 			put_child_root(pn, n->key, NULL);
1829 			node_free(n);
1830 		}
1831 	}
1832 
1833 	pr_debug("trie_flush found=%d\n", found);
1834 	return found;
1835 }
1836 
1837 static void __trie_free_rcu(struct rcu_head *head)
1838 {
1839 	struct fib_table *tb = container_of(head, struct fib_table, rcu);
1840 #ifdef CONFIG_IP_FIB_TRIE_STATS
1841 	struct trie *t = (struct trie *)tb->tb_data;
1842 
1843 	if (tb->tb_data == tb->__data)
1844 		free_percpu(t->stats);
1845 #endif /* CONFIG_IP_FIB_TRIE_STATS */
1846 	kfree(tb);
1847 }
1848 
1849 void fib_free_table(struct fib_table *tb)
1850 {
1851 	call_rcu(&tb->rcu, __trie_free_rcu);
1852 }
1853 
1854 static int fn_trie_dump_leaf(struct key_vector *l, struct fib_table *tb,
1855 			     struct sk_buff *skb, struct netlink_callback *cb)
1856 {
1857 	__be32 xkey = htonl(l->key);
1858 	struct fib_alias *fa;
1859 	int i, s_i;
1860 
1861 	s_i = cb->args[4];
1862 	i = 0;
1863 
1864 	/* rcu_read_lock is hold by caller */
1865 	hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
1866 		if (i < s_i) {
1867 			i++;
1868 			continue;
1869 		}
1870 
1871 		if (tb->tb_id != fa->tb_id) {
1872 			i++;
1873 			continue;
1874 		}
1875 
1876 		if (fib_dump_info(skb, NETLINK_CB(cb->skb).portid,
1877 				  cb->nlh->nlmsg_seq,
1878 				  RTM_NEWROUTE,
1879 				  tb->tb_id,
1880 				  fa->fa_type,
1881 				  xkey,
1882 				  KEYLENGTH - fa->fa_slen,
1883 				  fa->fa_tos,
1884 				  fa->fa_info, NLM_F_MULTI) < 0) {
1885 			cb->args[4] = i;
1886 			return -1;
1887 		}
1888 		i++;
1889 	}
1890 
1891 	cb->args[4] = i;
1892 	return skb->len;
1893 }
1894 
1895 /* rcu_read_lock needs to be hold by caller from readside */
1896 int fib_table_dump(struct fib_table *tb, struct sk_buff *skb,
1897 		   struct netlink_callback *cb)
1898 {
1899 	struct trie *t = (struct trie *)tb->tb_data;
1900 	struct key_vector *l, *tp = t->kv;
1901 	/* Dump starting at last key.
1902 	 * Note: 0.0.0.0/0 (ie default) is first key.
1903 	 */
1904 	int count = cb->args[2];
1905 	t_key key = cb->args[3];
1906 
1907 	while ((l = leaf_walk_rcu(&tp, key)) != NULL) {
1908 		if (fn_trie_dump_leaf(l, tb, skb, cb) < 0) {
1909 			cb->args[3] = key;
1910 			cb->args[2] = count;
1911 			return -1;
1912 		}
1913 
1914 		++count;
1915 		key = l->key + 1;
1916 
1917 		memset(&cb->args[4], 0,
1918 		       sizeof(cb->args) - 4*sizeof(cb->args[0]));
1919 
1920 		/* stop loop if key wrapped back to 0 */
1921 		if (key < l->key)
1922 			break;
1923 	}
1924 
1925 	cb->args[3] = key;
1926 	cb->args[2] = count;
1927 
1928 	return skb->len;
1929 }
1930 
1931 void __init fib_trie_init(void)
1932 {
1933 	fn_alias_kmem = kmem_cache_create("ip_fib_alias",
1934 					  sizeof(struct fib_alias),
1935 					  0, SLAB_PANIC, NULL);
1936 
1937 	trie_leaf_kmem = kmem_cache_create("ip_fib_trie",
1938 					   LEAF_SIZE,
1939 					   0, SLAB_PANIC, NULL);
1940 }
1941 
1942 struct fib_table *fib_trie_table(u32 id, struct fib_table *alias)
1943 {
1944 	struct fib_table *tb;
1945 	struct trie *t;
1946 	size_t sz = sizeof(*tb);
1947 
1948 	if (!alias)
1949 		sz += sizeof(struct trie);
1950 
1951 	tb = kzalloc(sz, GFP_KERNEL);
1952 	if (!tb)
1953 		return NULL;
1954 
1955 	tb->tb_id = id;
1956 	tb->tb_num_default = 0;
1957 	tb->tb_data = (alias ? alias->__data : tb->__data);
1958 
1959 	if (alias)
1960 		return tb;
1961 
1962 	t = (struct trie *) tb->tb_data;
1963 	t->kv[0].pos = KEYLENGTH;
1964 	t->kv[0].slen = KEYLENGTH;
1965 #ifdef CONFIG_IP_FIB_TRIE_STATS
1966 	t->stats = alloc_percpu(struct trie_use_stats);
1967 	if (!t->stats) {
1968 		kfree(tb);
1969 		tb = NULL;
1970 	}
1971 #endif
1972 
1973 	return tb;
1974 }
1975 
1976 #ifdef CONFIG_PROC_FS
1977 /* Depth first Trie walk iterator */
1978 struct fib_trie_iter {
1979 	struct seq_net_private p;
1980 	struct fib_table *tb;
1981 	struct key_vector *tnode;
1982 	unsigned int index;
1983 	unsigned int depth;
1984 };
1985 
1986 static struct key_vector *fib_trie_get_next(struct fib_trie_iter *iter)
1987 {
1988 	unsigned long cindex = iter->index;
1989 	struct key_vector *pn = iter->tnode;
1990 	t_key pkey;
1991 
1992 	pr_debug("get_next iter={node=%p index=%d depth=%d}\n",
1993 		 iter->tnode, iter->index, iter->depth);
1994 
1995 	while (!IS_TRIE(pn)) {
1996 		while (cindex < child_length(pn)) {
1997 			struct key_vector *n = get_child_rcu(pn, cindex++);
1998 
1999 			if (!n)
2000 				continue;
2001 
2002 			if (IS_LEAF(n)) {
2003 				iter->tnode = pn;
2004 				iter->index = cindex;
2005 			} else {
2006 				/* push down one level */
2007 				iter->tnode = n;
2008 				iter->index = 0;
2009 				++iter->depth;
2010 			}
2011 
2012 			return n;
2013 		}
2014 
2015 		/* Current node exhausted, pop back up */
2016 		pkey = pn->key;
2017 		pn = node_parent_rcu(pn);
2018 		cindex = get_index(pkey, pn) + 1;
2019 		--iter->depth;
2020 	}
2021 
2022 	/* record root node so further searches know we are done */
2023 	iter->tnode = pn;
2024 	iter->index = 0;
2025 
2026 	return NULL;
2027 }
2028 
2029 static struct key_vector *fib_trie_get_first(struct fib_trie_iter *iter,
2030 					     struct trie *t)
2031 {
2032 	struct key_vector *n, *pn;
2033 
2034 	if (!t)
2035 		return NULL;
2036 
2037 	pn = t->kv;
2038 	n = rcu_dereference(pn->tnode[0]);
2039 	if (!n)
2040 		return NULL;
2041 
2042 	if (IS_TNODE(n)) {
2043 		iter->tnode = n;
2044 		iter->index = 0;
2045 		iter->depth = 1;
2046 	} else {
2047 		iter->tnode = pn;
2048 		iter->index = 0;
2049 		iter->depth = 0;
2050 	}
2051 
2052 	return n;
2053 }
2054 
2055 static void trie_collect_stats(struct trie *t, struct trie_stat *s)
2056 {
2057 	struct key_vector *n;
2058 	struct fib_trie_iter iter;
2059 
2060 	memset(s, 0, sizeof(*s));
2061 
2062 	rcu_read_lock();
2063 	for (n = fib_trie_get_first(&iter, t); n; n = fib_trie_get_next(&iter)) {
2064 		if (IS_LEAF(n)) {
2065 			struct fib_alias *fa;
2066 
2067 			s->leaves++;
2068 			s->totdepth += iter.depth;
2069 			if (iter.depth > s->maxdepth)
2070 				s->maxdepth = iter.depth;
2071 
2072 			hlist_for_each_entry_rcu(fa, &n->leaf, fa_list)
2073 				++s->prefixes;
2074 		} else {
2075 			s->tnodes++;
2076 			if (n->bits < MAX_STAT_DEPTH)
2077 				s->nodesizes[n->bits]++;
2078 			s->nullpointers += tn_info(n)->empty_children;
2079 		}
2080 	}
2081 	rcu_read_unlock();
2082 }
2083 
2084 /*
2085  *	This outputs /proc/net/fib_triestats
2086  */
2087 static void trie_show_stats(struct seq_file *seq, struct trie_stat *stat)
2088 {
2089 	unsigned int i, max, pointers, bytes, avdepth;
2090 
2091 	if (stat->leaves)
2092 		avdepth = stat->totdepth*100 / stat->leaves;
2093 	else
2094 		avdepth = 0;
2095 
2096 	seq_printf(seq, "\tAver depth:     %u.%02d\n",
2097 		   avdepth / 100, avdepth % 100);
2098 	seq_printf(seq, "\tMax depth:      %u\n", stat->maxdepth);
2099 
2100 	seq_printf(seq, "\tLeaves:         %u\n", stat->leaves);
2101 	bytes = LEAF_SIZE * stat->leaves;
2102 
2103 	seq_printf(seq, "\tPrefixes:       %u\n", stat->prefixes);
2104 	bytes += sizeof(struct fib_alias) * stat->prefixes;
2105 
2106 	seq_printf(seq, "\tInternal nodes: %u\n\t", stat->tnodes);
2107 	bytes += TNODE_SIZE(0) * stat->tnodes;
2108 
2109 	max = MAX_STAT_DEPTH;
2110 	while (max > 0 && stat->nodesizes[max-1] == 0)
2111 		max--;
2112 
2113 	pointers = 0;
2114 	for (i = 1; i < max; i++)
2115 		if (stat->nodesizes[i] != 0) {
2116 			seq_printf(seq, "  %u: %u",  i, stat->nodesizes[i]);
2117 			pointers += (1<<i) * stat->nodesizes[i];
2118 		}
2119 	seq_putc(seq, '\n');
2120 	seq_printf(seq, "\tPointers: %u\n", pointers);
2121 
2122 	bytes += sizeof(struct key_vector *) * pointers;
2123 	seq_printf(seq, "Null ptrs: %u\n", stat->nullpointers);
2124 	seq_printf(seq, "Total size: %u  kB\n", (bytes + 1023) / 1024);
2125 }
2126 
2127 #ifdef CONFIG_IP_FIB_TRIE_STATS
2128 static void trie_show_usage(struct seq_file *seq,
2129 			    const struct trie_use_stats __percpu *stats)
2130 {
2131 	struct trie_use_stats s = { 0 };
2132 	int cpu;
2133 
2134 	/* loop through all of the CPUs and gather up the stats */
2135 	for_each_possible_cpu(cpu) {
2136 		const struct trie_use_stats *pcpu = per_cpu_ptr(stats, cpu);
2137 
2138 		s.gets += pcpu->gets;
2139 		s.backtrack += pcpu->backtrack;
2140 		s.semantic_match_passed += pcpu->semantic_match_passed;
2141 		s.semantic_match_miss += pcpu->semantic_match_miss;
2142 		s.null_node_hit += pcpu->null_node_hit;
2143 		s.resize_node_skipped += pcpu->resize_node_skipped;
2144 	}
2145 
2146 	seq_printf(seq, "\nCounters:\n---------\n");
2147 	seq_printf(seq, "gets = %u\n", s.gets);
2148 	seq_printf(seq, "backtracks = %u\n", s.backtrack);
2149 	seq_printf(seq, "semantic match passed = %u\n",
2150 		   s.semantic_match_passed);
2151 	seq_printf(seq, "semantic match miss = %u\n", s.semantic_match_miss);
2152 	seq_printf(seq, "null node hit= %u\n", s.null_node_hit);
2153 	seq_printf(seq, "skipped node resize = %u\n\n", s.resize_node_skipped);
2154 }
2155 #endif /*  CONFIG_IP_FIB_TRIE_STATS */
2156 
2157 static void fib_table_print(struct seq_file *seq, struct fib_table *tb)
2158 {
2159 	if (tb->tb_id == RT_TABLE_LOCAL)
2160 		seq_puts(seq, "Local:\n");
2161 	else if (tb->tb_id == RT_TABLE_MAIN)
2162 		seq_puts(seq, "Main:\n");
2163 	else
2164 		seq_printf(seq, "Id %d:\n", tb->tb_id);
2165 }
2166 
2167 
2168 static int fib_triestat_seq_show(struct seq_file *seq, void *v)
2169 {
2170 	struct net *net = (struct net *)seq->private;
2171 	unsigned int h;
2172 
2173 	seq_printf(seq,
2174 		   "Basic info: size of leaf:"
2175 		   " %Zd bytes, size of tnode: %Zd bytes.\n",
2176 		   LEAF_SIZE, TNODE_SIZE(0));
2177 
2178 	for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2179 		struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2180 		struct fib_table *tb;
2181 
2182 		hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2183 			struct trie *t = (struct trie *) tb->tb_data;
2184 			struct trie_stat stat;
2185 
2186 			if (!t)
2187 				continue;
2188 
2189 			fib_table_print(seq, tb);
2190 
2191 			trie_collect_stats(t, &stat);
2192 			trie_show_stats(seq, &stat);
2193 #ifdef CONFIG_IP_FIB_TRIE_STATS
2194 			trie_show_usage(seq, t->stats);
2195 #endif
2196 		}
2197 	}
2198 
2199 	return 0;
2200 }
2201 
2202 static int fib_triestat_seq_open(struct inode *inode, struct file *file)
2203 {
2204 	return single_open_net(inode, file, fib_triestat_seq_show);
2205 }
2206 
2207 static const struct file_operations fib_triestat_fops = {
2208 	.owner	= THIS_MODULE,
2209 	.open	= fib_triestat_seq_open,
2210 	.read	= seq_read,
2211 	.llseek	= seq_lseek,
2212 	.release = single_release_net,
2213 };
2214 
2215 static struct key_vector *fib_trie_get_idx(struct seq_file *seq, loff_t pos)
2216 {
2217 	struct fib_trie_iter *iter = seq->private;
2218 	struct net *net = seq_file_net(seq);
2219 	loff_t idx = 0;
2220 	unsigned int h;
2221 
2222 	for (h = 0; h < FIB_TABLE_HASHSZ; h++) {
2223 		struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2224 		struct fib_table *tb;
2225 
2226 		hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2227 			struct key_vector *n;
2228 
2229 			for (n = fib_trie_get_first(iter,
2230 						    (struct trie *) tb->tb_data);
2231 			     n; n = fib_trie_get_next(iter))
2232 				if (pos == idx++) {
2233 					iter->tb = tb;
2234 					return n;
2235 				}
2236 		}
2237 	}
2238 
2239 	return NULL;
2240 }
2241 
2242 static void *fib_trie_seq_start(struct seq_file *seq, loff_t *pos)
2243 	__acquires(RCU)
2244 {
2245 	rcu_read_lock();
2246 	return fib_trie_get_idx(seq, *pos);
2247 }
2248 
2249 static void *fib_trie_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2250 {
2251 	struct fib_trie_iter *iter = seq->private;
2252 	struct net *net = seq_file_net(seq);
2253 	struct fib_table *tb = iter->tb;
2254 	struct hlist_node *tb_node;
2255 	unsigned int h;
2256 	struct key_vector *n;
2257 
2258 	++*pos;
2259 	/* next node in same table */
2260 	n = fib_trie_get_next(iter);
2261 	if (n)
2262 		return n;
2263 
2264 	/* walk rest of this hash chain */
2265 	h = tb->tb_id & (FIB_TABLE_HASHSZ - 1);
2266 	while ((tb_node = rcu_dereference(hlist_next_rcu(&tb->tb_hlist)))) {
2267 		tb = hlist_entry(tb_node, struct fib_table, tb_hlist);
2268 		n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2269 		if (n)
2270 			goto found;
2271 	}
2272 
2273 	/* new hash chain */
2274 	while (++h < FIB_TABLE_HASHSZ) {
2275 		struct hlist_head *head = &net->ipv4.fib_table_hash[h];
2276 		hlist_for_each_entry_rcu(tb, head, tb_hlist) {
2277 			n = fib_trie_get_first(iter, (struct trie *) tb->tb_data);
2278 			if (n)
2279 				goto found;
2280 		}
2281 	}
2282 	return NULL;
2283 
2284 found:
2285 	iter->tb = tb;
2286 	return n;
2287 }
2288 
2289 static void fib_trie_seq_stop(struct seq_file *seq, void *v)
2290 	__releases(RCU)
2291 {
2292 	rcu_read_unlock();
2293 }
2294 
2295 static void seq_indent(struct seq_file *seq, int n)
2296 {
2297 	while (n-- > 0)
2298 		seq_puts(seq, "   ");
2299 }
2300 
2301 static inline const char *rtn_scope(char *buf, size_t len, enum rt_scope_t s)
2302 {
2303 	switch (s) {
2304 	case RT_SCOPE_UNIVERSE: return "universe";
2305 	case RT_SCOPE_SITE:	return "site";
2306 	case RT_SCOPE_LINK:	return "link";
2307 	case RT_SCOPE_HOST:	return "host";
2308 	case RT_SCOPE_NOWHERE:	return "nowhere";
2309 	default:
2310 		snprintf(buf, len, "scope=%d", s);
2311 		return buf;
2312 	}
2313 }
2314 
2315 static const char *const rtn_type_names[__RTN_MAX] = {
2316 	[RTN_UNSPEC] = "UNSPEC",
2317 	[RTN_UNICAST] = "UNICAST",
2318 	[RTN_LOCAL] = "LOCAL",
2319 	[RTN_BROADCAST] = "BROADCAST",
2320 	[RTN_ANYCAST] = "ANYCAST",
2321 	[RTN_MULTICAST] = "MULTICAST",
2322 	[RTN_BLACKHOLE] = "BLACKHOLE",
2323 	[RTN_UNREACHABLE] = "UNREACHABLE",
2324 	[RTN_PROHIBIT] = "PROHIBIT",
2325 	[RTN_THROW] = "THROW",
2326 	[RTN_NAT] = "NAT",
2327 	[RTN_XRESOLVE] = "XRESOLVE",
2328 };
2329 
2330 static inline const char *rtn_type(char *buf, size_t len, unsigned int t)
2331 {
2332 	if (t < __RTN_MAX && rtn_type_names[t])
2333 		return rtn_type_names[t];
2334 	snprintf(buf, len, "type %u", t);
2335 	return buf;
2336 }
2337 
2338 /* Pretty print the trie */
2339 static int fib_trie_seq_show(struct seq_file *seq, void *v)
2340 {
2341 	const struct fib_trie_iter *iter = seq->private;
2342 	struct key_vector *n = v;
2343 
2344 	if (IS_TRIE(node_parent_rcu(n)))
2345 		fib_table_print(seq, iter->tb);
2346 
2347 	if (IS_TNODE(n)) {
2348 		__be32 prf = htonl(n->key);
2349 
2350 		seq_indent(seq, iter->depth-1);
2351 		seq_printf(seq, "  +-- %pI4/%zu %u %u %u\n",
2352 			   &prf, KEYLENGTH - n->pos - n->bits, n->bits,
2353 			   tn_info(n)->full_children,
2354 			   tn_info(n)->empty_children);
2355 	} else {
2356 		__be32 val = htonl(n->key);
2357 		struct fib_alias *fa;
2358 
2359 		seq_indent(seq, iter->depth);
2360 		seq_printf(seq, "  |-- %pI4\n", &val);
2361 
2362 		hlist_for_each_entry_rcu(fa, &n->leaf, fa_list) {
2363 			char buf1[32], buf2[32];
2364 
2365 			seq_indent(seq, iter->depth + 1);
2366 			seq_printf(seq, "  /%zu %s %s",
2367 				   KEYLENGTH - fa->fa_slen,
2368 				   rtn_scope(buf1, sizeof(buf1),
2369 					     fa->fa_info->fib_scope),
2370 				   rtn_type(buf2, sizeof(buf2),
2371 					    fa->fa_type));
2372 			if (fa->fa_tos)
2373 				seq_printf(seq, " tos=%d", fa->fa_tos);
2374 			seq_putc(seq, '\n');
2375 		}
2376 	}
2377 
2378 	return 0;
2379 }
2380 
2381 static const struct seq_operations fib_trie_seq_ops = {
2382 	.start  = fib_trie_seq_start,
2383 	.next   = fib_trie_seq_next,
2384 	.stop   = fib_trie_seq_stop,
2385 	.show   = fib_trie_seq_show,
2386 };
2387 
2388 static int fib_trie_seq_open(struct inode *inode, struct file *file)
2389 {
2390 	return seq_open_net(inode, file, &fib_trie_seq_ops,
2391 			    sizeof(struct fib_trie_iter));
2392 }
2393 
2394 static const struct file_operations fib_trie_fops = {
2395 	.owner  = THIS_MODULE,
2396 	.open   = fib_trie_seq_open,
2397 	.read   = seq_read,
2398 	.llseek = seq_lseek,
2399 	.release = seq_release_net,
2400 };
2401 
2402 struct fib_route_iter {
2403 	struct seq_net_private p;
2404 	struct fib_table *main_tb;
2405 	struct key_vector *tnode;
2406 	loff_t	pos;
2407 	t_key	key;
2408 };
2409 
2410 static struct key_vector *fib_route_get_idx(struct fib_route_iter *iter,
2411 					    loff_t pos)
2412 {
2413 	struct key_vector *l, **tp = &iter->tnode;
2414 	t_key key;
2415 
2416 	/* use cache location of next-to-find key */
2417 	if (iter->pos > 0 && pos >= iter->pos) {
2418 		pos -= iter->pos;
2419 		key = iter->key;
2420 	} else {
2421 		iter->pos = 0;
2422 		key = 0;
2423 	}
2424 
2425 	while ((l = leaf_walk_rcu(tp, key)) != NULL) {
2426 		key = l->key + 1;
2427 		iter->pos++;
2428 
2429 		if (--pos <= 0)
2430 			break;
2431 
2432 		l = NULL;
2433 
2434 		/* handle unlikely case of a key wrap */
2435 		if (!key)
2436 			break;
2437 	}
2438 
2439 	if (l)
2440 		iter->key = key;	/* remember it */
2441 	else
2442 		iter->pos = 0;		/* forget it */
2443 
2444 	return l;
2445 }
2446 
2447 static void *fib_route_seq_start(struct seq_file *seq, loff_t *pos)
2448 	__acquires(RCU)
2449 {
2450 	struct fib_route_iter *iter = seq->private;
2451 	struct fib_table *tb;
2452 	struct trie *t;
2453 
2454 	rcu_read_lock();
2455 
2456 	tb = fib_get_table(seq_file_net(seq), RT_TABLE_MAIN);
2457 	if (!tb)
2458 		return NULL;
2459 
2460 	iter->main_tb = tb;
2461 	t = (struct trie *)tb->tb_data;
2462 	iter->tnode = t->kv;
2463 
2464 	if (*pos != 0)
2465 		return fib_route_get_idx(iter, *pos);
2466 
2467 	iter->pos = 0;
2468 	iter->key = 0;
2469 
2470 	return SEQ_START_TOKEN;
2471 }
2472 
2473 static void *fib_route_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2474 {
2475 	struct fib_route_iter *iter = seq->private;
2476 	struct key_vector *l = NULL;
2477 	t_key key = iter->key;
2478 
2479 	++*pos;
2480 
2481 	/* only allow key of 0 for start of sequence */
2482 	if ((v == SEQ_START_TOKEN) || key)
2483 		l = leaf_walk_rcu(&iter->tnode, key);
2484 
2485 	if (l) {
2486 		iter->key = l->key + 1;
2487 		iter->pos++;
2488 	} else {
2489 		iter->pos = 0;
2490 	}
2491 
2492 	return l;
2493 }
2494 
2495 static void fib_route_seq_stop(struct seq_file *seq, void *v)
2496 	__releases(RCU)
2497 {
2498 	rcu_read_unlock();
2499 }
2500 
2501 static unsigned int fib_flag_trans(int type, __be32 mask, const struct fib_info *fi)
2502 {
2503 	unsigned int flags = 0;
2504 
2505 	if (type == RTN_UNREACHABLE || type == RTN_PROHIBIT)
2506 		flags = RTF_REJECT;
2507 	if (fi && fi->fib_nh->nh_gw)
2508 		flags |= RTF_GATEWAY;
2509 	if (mask == htonl(0xFFFFFFFF))
2510 		flags |= RTF_HOST;
2511 	flags |= RTF_UP;
2512 	return flags;
2513 }
2514 
2515 /*
2516  *	This outputs /proc/net/route.
2517  *	The format of the file is not supposed to be changed
2518  *	and needs to be same as fib_hash output to avoid breaking
2519  *	legacy utilities
2520  */
2521 static int fib_route_seq_show(struct seq_file *seq, void *v)
2522 {
2523 	struct fib_route_iter *iter = seq->private;
2524 	struct fib_table *tb = iter->main_tb;
2525 	struct fib_alias *fa;
2526 	struct key_vector *l = v;
2527 	__be32 prefix;
2528 
2529 	if (v == SEQ_START_TOKEN) {
2530 		seq_printf(seq, "%-127s\n", "Iface\tDestination\tGateway "
2531 			   "\tFlags\tRefCnt\tUse\tMetric\tMask\t\tMTU"
2532 			   "\tWindow\tIRTT");
2533 		return 0;
2534 	}
2535 
2536 	prefix = htonl(l->key);
2537 
2538 	hlist_for_each_entry_rcu(fa, &l->leaf, fa_list) {
2539 		const struct fib_info *fi = fa->fa_info;
2540 		__be32 mask = inet_make_mask(KEYLENGTH - fa->fa_slen);
2541 		unsigned int flags = fib_flag_trans(fa->fa_type, mask, fi);
2542 
2543 		if ((fa->fa_type == RTN_BROADCAST) ||
2544 		    (fa->fa_type == RTN_MULTICAST))
2545 			continue;
2546 
2547 		if (fa->tb_id != tb->tb_id)
2548 			continue;
2549 
2550 		seq_setwidth(seq, 127);
2551 
2552 		if (fi)
2553 			seq_printf(seq,
2554 				   "%s\t%08X\t%08X\t%04X\t%d\t%u\t"
2555 				   "%d\t%08X\t%d\t%u\t%u",
2556 				   fi->fib_dev ? fi->fib_dev->name : "*",
2557 				   prefix,
2558 				   fi->fib_nh->nh_gw, flags, 0, 0,
2559 				   fi->fib_priority,
2560 				   mask,
2561 				   (fi->fib_advmss ?
2562 				    fi->fib_advmss + 40 : 0),
2563 				   fi->fib_window,
2564 				   fi->fib_rtt >> 3);
2565 		else
2566 			seq_printf(seq,
2567 				   "*\t%08X\t%08X\t%04X\t%d\t%u\t"
2568 				   "%d\t%08X\t%d\t%u\t%u",
2569 				   prefix, 0, flags, 0, 0, 0,
2570 				   mask, 0, 0, 0);
2571 
2572 		seq_pad(seq, '\n');
2573 	}
2574 
2575 	return 0;
2576 }
2577 
2578 static const struct seq_operations fib_route_seq_ops = {
2579 	.start  = fib_route_seq_start,
2580 	.next   = fib_route_seq_next,
2581 	.stop   = fib_route_seq_stop,
2582 	.show   = fib_route_seq_show,
2583 };
2584 
2585 static int fib_route_seq_open(struct inode *inode, struct file *file)
2586 {
2587 	return seq_open_net(inode, file, &fib_route_seq_ops,
2588 			    sizeof(struct fib_route_iter));
2589 }
2590 
2591 static const struct file_operations fib_route_fops = {
2592 	.owner  = THIS_MODULE,
2593 	.open   = fib_route_seq_open,
2594 	.read   = seq_read,
2595 	.llseek = seq_lseek,
2596 	.release = seq_release_net,
2597 };
2598 
2599 int __net_init fib_proc_init(struct net *net)
2600 {
2601 	if (!proc_create("fib_trie", S_IRUGO, net->proc_net, &fib_trie_fops))
2602 		goto out1;
2603 
2604 	if (!proc_create("fib_triestat", S_IRUGO, net->proc_net,
2605 			 &fib_triestat_fops))
2606 		goto out2;
2607 
2608 	if (!proc_create("route", S_IRUGO, net->proc_net, &fib_route_fops))
2609 		goto out3;
2610 
2611 	return 0;
2612 
2613 out3:
2614 	remove_proc_entry("fib_triestat", net->proc_net);
2615 out2:
2616 	remove_proc_entry("fib_trie", net->proc_net);
2617 out1:
2618 	return -ENOMEM;
2619 }
2620 
2621 void __net_exit fib_proc_exit(struct net *net)
2622 {
2623 	remove_proc_entry("fib_trie", net->proc_net);
2624 	remove_proc_entry("fib_triestat", net->proc_net);
2625 	remove_proc_entry("route", net->proc_net);
2626 }
2627 
2628 #endif /* CONFIG_PROC_FS */
2629