xref: /freebsd/sys/net/radix.c (revision e307eb94ae520d98dc1d346a0c53667a41beab5d)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1988, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	@(#)radix.c	8.5 (Berkeley) 5/19/95
32  * $FreeBSD$
33  */
34 
35 /*
36  * Routines to build and maintain radix trees for routing lookups.
37  */
38 #include <sys/param.h>
39 #ifdef	_KERNEL
40 #include <sys/lock.h>
41 #include <sys/mutex.h>
42 #include <sys/rmlock.h>
43 #include <sys/systm.h>
44 #include <sys/malloc.h>
45 #include <sys/syslog.h>
46 #include <net/radix.h>
47 #include "opt_mpath.h"
48 #ifdef RADIX_MPATH
49 #include <net/radix_mpath.h>
50 #endif
51 #else /* !_KERNEL */
52 #include <stdio.h>
53 #include <strings.h>
54 #include <stdlib.h>
55 #define log(x, arg...)	fprintf(stderr, ## arg)
56 #define panic(x)	fprintf(stderr, "PANIC: %s", x), exit(1)
57 #define min(a, b) ((a) < (b) ? (a) : (b) )
58 #include <net/radix.h>
59 #endif /* !_KERNEL */
60 
61 static struct radix_node
62 	 *rn_insert(void *, struct radix_head *, int *,
63 	     struct radix_node [2]),
64 	 *rn_newpair(void *, int, struct radix_node[2]),
65 	 *rn_search(void *, struct radix_node *),
66 	 *rn_search_m(void *, struct radix_node *, void *);
67 static struct radix_node *rn_addmask(void *, struct radix_mask_head *, int,int);
68 
69 static void rn_detachhead_internal(struct radix_head *);
70 
71 #define	RADIX_MAX_KEY_LEN	32
72 
73 static char rn_zeros[RADIX_MAX_KEY_LEN];
74 static char rn_ones[RADIX_MAX_KEY_LEN] = {
75 	-1, -1, -1, -1, -1, -1, -1, -1,
76 	-1, -1, -1, -1, -1, -1, -1, -1,
77 	-1, -1, -1, -1, -1, -1, -1, -1,
78 	-1, -1, -1, -1, -1, -1, -1, -1,
79 };
80 
81 static int	rn_lexobetter(void *m_arg, void *n_arg);
82 static struct radix_mask *
83 		rn_new_radix_mask(struct radix_node *tt,
84 		    struct radix_mask *next);
85 static int	rn_satisfies_leaf(char *trial, struct radix_node *leaf,
86 		    int skip);
87 
88 /*
89  * The data structure for the keys is a radix tree with one way
90  * branching removed.  The index rn_bit at an internal node n represents a bit
91  * position to be tested.  The tree is arranged so that all descendants
92  * of a node n have keys whose bits all agree up to position rn_bit - 1.
93  * (We say the index of n is rn_bit.)
94  *
95  * There is at least one descendant which has a one bit at position rn_bit,
96  * and at least one with a zero there.
97  *
98  * A route is determined by a pair of key and mask.  We require that the
99  * bit-wise logical and of the key and mask to be the key.
100  * We define the index of a route to associated with the mask to be
101  * the first bit number in the mask where 0 occurs (with bit number 0
102  * representing the highest order bit).
103  *
104  * We say a mask is normal if every bit is 0, past the index of the mask.
105  * If a node n has a descendant (k, m) with index(m) == index(n) == rn_bit,
106  * and m is a normal mask, then the route applies to every descendant of n.
107  * If the index(m) < rn_bit, this implies the trailing last few bits of k
108  * before bit b are all 0, (and hence consequently true of every descendant
109  * of n), so the route applies to all descendants of the node as well.
110  *
111  * Similar logic shows that a non-normal mask m such that
112  * index(m) <= index(n) could potentially apply to many children of n.
113  * Thus, for each non-host route, we attach its mask to a list at an internal
114  * node as high in the tree as we can go.
115  *
116  * The present version of the code makes use of normal routes in short-
117  * circuiting an explict mask and compare operation when testing whether
118  * a key satisfies a normal route, and also in remembering the unique leaf
119  * that governs a subtree.
120  */
121 
122 /*
123  * Most of the functions in this code assume that the key/mask arguments
124  * are sockaddr-like structures, where the first byte is an u_char
125  * indicating the size of the entire structure.
126  *
127  * To make the assumption more explicit, we use the LEN() macro to access
128  * this field. It is safe to pass an expression with side effects
129  * to LEN() as the argument is evaluated only once.
130  * We cast the result to int as this is the dominant usage.
131  */
132 #define LEN(x) ( (int) (*(const u_char *)(x)) )
133 
134 /*
135  * XXX THIS NEEDS TO BE FIXED
136  * In the code, pointers to keys and masks are passed as either
137  * 'void *' (because callers use to pass pointers of various kinds), or
138  * 'caddr_t' (which is fine for pointer arithmetics, but not very
139  * clean when you dereference it to access data). Furthermore, caddr_t
140  * is really 'char *', while the natural type to operate on keys and
141  * masks would be 'u_char'. This mismatch require a lot of casts and
142  * intermediate variables to adapt types that clutter the code.
143  */
144 
145 /*
146  * Search a node in the tree matching the key.
147  */
148 static struct radix_node *
149 rn_search(void *v_arg, struct radix_node *head)
150 {
151 	struct radix_node *x;
152 	caddr_t v;
153 
154 	for (x = head, v = v_arg; x->rn_bit >= 0;) {
155 		if (x->rn_bmask & v[x->rn_offset])
156 			x = x->rn_right;
157 		else
158 			x = x->rn_left;
159 	}
160 	return (x);
161 }
162 
163 /*
164  * Same as above, but with an additional mask.
165  * XXX note this function is used only once.
166  */
167 static struct radix_node *
168 rn_search_m(void *v_arg, struct radix_node *head, void *m_arg)
169 {
170 	struct radix_node *x;
171 	caddr_t v = v_arg, m = m_arg;
172 
173 	for (x = head; x->rn_bit >= 0;) {
174 		if ((x->rn_bmask & m[x->rn_offset]) &&
175 		    (x->rn_bmask & v[x->rn_offset]))
176 			x = x->rn_right;
177 		else
178 			x = x->rn_left;
179 	}
180 	return (x);
181 }
182 
183 int
184 rn_refines(void *m_arg, void *n_arg)
185 {
186 	caddr_t m = m_arg, n = n_arg;
187 	caddr_t lim, lim2 = lim = n + LEN(n);
188 	int longer = LEN(n++) - LEN(m++);
189 	int masks_are_equal = 1;
190 
191 	if (longer > 0)
192 		lim -= longer;
193 	while (n < lim) {
194 		if (*n & ~(*m))
195 			return (0);
196 		if (*n++ != *m++)
197 			masks_are_equal = 0;
198 	}
199 	while (n < lim2)
200 		if (*n++)
201 			return (0);
202 	if (masks_are_equal && (longer < 0))
203 		for (lim2 = m - longer; m < lim2; )
204 			if (*m++)
205 				return (1);
206 	return (!masks_are_equal);
207 }
208 
209 /*
210  * Search for exact match in given @head.
211  * Assume host bits are cleared in @v_arg if @m_arg is not NULL
212  * Note that prefixes with /32 or /128 masks are treated differently
213  * from host routes.
214  */
215 struct radix_node *
216 rn_lookup(void *v_arg, void *m_arg, struct radix_head *head)
217 {
218 	struct radix_node *x;
219 	caddr_t netmask;
220 
221 	if (m_arg != NULL) {
222 		/*
223 		 * Most common case: search exact prefix/mask
224 		 */
225 		x = rn_addmask(m_arg, head->rnh_masks, 1,
226 		    head->rnh_treetop->rn_offset);
227 		if (x == NULL)
228 			return (NULL);
229 		netmask = x->rn_key;
230 
231 		x = rn_match(v_arg, head);
232 
233 		while (x != NULL && x->rn_mask != netmask)
234 			x = x->rn_dupedkey;
235 
236 		return (x);
237 	}
238 
239 	/*
240 	 * Search for host address.
241 	 */
242 	if ((x = rn_match(v_arg, head)) == NULL)
243 		return (NULL);
244 
245 	/* Check if found key is the same */
246 	if (LEN(x->rn_key) != LEN(v_arg) || bcmp(x->rn_key, v_arg, LEN(v_arg)))
247 		return (NULL);
248 
249 	/* Check if this is not host route */
250 	if (x->rn_mask != NULL)
251 		return (NULL);
252 
253 	return (x);
254 }
255 
256 static int
257 rn_satisfies_leaf(char *trial, struct radix_node *leaf, int skip)
258 {
259 	char *cp = trial, *cp2 = leaf->rn_key, *cp3 = leaf->rn_mask;
260 	char *cplim;
261 	int length = min(LEN(cp), LEN(cp2));
262 
263 	if (cp3 == NULL)
264 		cp3 = rn_ones;
265 	else
266 		length = min(length, LEN(cp3));
267 	cplim = cp + length; cp3 += skip; cp2 += skip;
268 	for (cp += skip; cp < cplim; cp++, cp2++, cp3++)
269 		if ((*cp ^ *cp2) & *cp3)
270 			return (0);
271 	return (1);
272 }
273 
274 /*
275  * Search for longest-prefix match in given @head
276  */
277 struct radix_node *
278 rn_match(void *v_arg, struct radix_head *head)
279 {
280 	caddr_t v = v_arg;
281 	struct radix_node *t = head->rnh_treetop, *x;
282 	caddr_t cp = v, cp2;
283 	caddr_t cplim;
284 	struct radix_node *saved_t, *top = t;
285 	int off = t->rn_offset, vlen = LEN(cp), matched_off;
286 	int test, b, rn_bit;
287 
288 	/*
289 	 * Open code rn_search(v, top) to avoid overhead of extra
290 	 * subroutine call.
291 	 */
292 	for (; t->rn_bit >= 0; ) {
293 		if (t->rn_bmask & cp[t->rn_offset])
294 			t = t->rn_right;
295 		else
296 			t = t->rn_left;
297 	}
298 	/*
299 	 * See if we match exactly as a host destination
300 	 * or at least learn how many bits match, for normal mask finesse.
301 	 *
302 	 * It doesn't hurt us to limit how many bytes to check
303 	 * to the length of the mask, since if it matches we had a genuine
304 	 * match and the leaf we have is the most specific one anyway;
305 	 * if it didn't match with a shorter length it would fail
306 	 * with a long one.  This wins big for class B&C netmasks which
307 	 * are probably the most common case...
308 	 */
309 	if (t->rn_mask)
310 		vlen = *(u_char *)t->rn_mask;
311 	cp += off; cp2 = t->rn_key + off; cplim = v + vlen;
312 	for (; cp < cplim; cp++, cp2++)
313 		if (*cp != *cp2)
314 			goto on1;
315 	/*
316 	 * This extra grot is in case we are explicitly asked
317 	 * to look up the default.  Ugh!
318 	 *
319 	 * Never return the root node itself, it seems to cause a
320 	 * lot of confusion.
321 	 */
322 	if (t->rn_flags & RNF_ROOT)
323 		t = t->rn_dupedkey;
324 	return (t);
325 on1:
326 	test = (*cp ^ *cp2) & 0xff; /* find first bit that differs */
327 	for (b = 7; (test >>= 1) > 0;)
328 		b--;
329 	matched_off = cp - v;
330 	b += matched_off << 3;
331 	rn_bit = -1 - b;
332 	/*
333 	 * If there is a host route in a duped-key chain, it will be first.
334 	 */
335 	if ((saved_t = t)->rn_mask == 0)
336 		t = t->rn_dupedkey;
337 	for (; t; t = t->rn_dupedkey)
338 		/*
339 		 * Even if we don't match exactly as a host,
340 		 * we may match if the leaf we wound up at is
341 		 * a route to a net.
342 		 */
343 		if (t->rn_flags & RNF_NORMAL) {
344 			if (rn_bit <= t->rn_bit)
345 				return (t);
346 		} else if (rn_satisfies_leaf(v, t, matched_off))
347 				return (t);
348 	t = saved_t;
349 	/* start searching up the tree */
350 	do {
351 		struct radix_mask *m;
352 		t = t->rn_parent;
353 		m = t->rn_mklist;
354 		/*
355 		 * If non-contiguous masks ever become important
356 		 * we can restore the masking and open coding of
357 		 * the search and satisfaction test and put the
358 		 * calculation of "off" back before the "do".
359 		 */
360 		while (m) {
361 			if (m->rm_flags & RNF_NORMAL) {
362 				if (rn_bit <= m->rm_bit)
363 					return (m->rm_leaf);
364 			} else {
365 				off = min(t->rn_offset, matched_off);
366 				x = rn_search_m(v, t, m->rm_mask);
367 				while (x && x->rn_mask != m->rm_mask)
368 					x = x->rn_dupedkey;
369 				if (x && rn_satisfies_leaf(v, x, off))
370 					return (x);
371 			}
372 			m = m->rm_mklist;
373 		}
374 	} while (t != top);
375 	return (0);
376 }
377 
378 #ifdef RN_DEBUG
379 int	rn_nodenum;
380 struct	radix_node *rn_clist;
381 int	rn_saveinfo;
382 int	rn_debug =  1;
383 #endif
384 
385 /*
386  * Whenever we add a new leaf to the tree, we also add a parent node,
387  * so we allocate them as an array of two elements: the first one must be
388  * the leaf (see RNTORT() in route.c), the second one is the parent.
389  * This routine initializes the relevant fields of the nodes, so that
390  * the leaf is the left child of the parent node, and both nodes have
391  * (almost) all all fields filled as appropriate.
392  * (XXX some fields are left unset, see the '#if 0' section).
393  * The function returns a pointer to the parent node.
394  */
395 
396 static struct radix_node *
397 rn_newpair(void *v, int b, struct radix_node nodes[2])
398 {
399 	struct radix_node *tt = nodes, *t = tt + 1;
400 	t->rn_bit = b;
401 	t->rn_bmask = 0x80 >> (b & 7);
402 	t->rn_left = tt;
403 	t->rn_offset = b >> 3;
404 
405 #if 0  /* XXX perhaps we should fill these fields as well. */
406 	t->rn_parent = t->rn_right = NULL;
407 
408 	tt->rn_mask = NULL;
409 	tt->rn_dupedkey = NULL;
410 	tt->rn_bmask = 0;
411 #endif
412 	tt->rn_bit = -1;
413 	tt->rn_key = (caddr_t)v;
414 	tt->rn_parent = t;
415 	tt->rn_flags = t->rn_flags = RNF_ACTIVE;
416 	tt->rn_mklist = t->rn_mklist = 0;
417 #ifdef RN_DEBUG
418 	tt->rn_info = rn_nodenum++; t->rn_info = rn_nodenum++;
419 	tt->rn_twin = t;
420 	tt->rn_ybro = rn_clist;
421 	rn_clist = tt;
422 #endif
423 	return (t);
424 }
425 
426 static struct radix_node *
427 rn_insert(void *v_arg, struct radix_head *head, int *dupentry,
428     struct radix_node nodes[2])
429 {
430 	caddr_t v = v_arg;
431 	struct radix_node *top = head->rnh_treetop;
432 	int head_off = top->rn_offset, vlen = LEN(v);
433 	struct radix_node *t = rn_search(v_arg, top);
434 	caddr_t cp = v + head_off;
435 	int b;
436 	struct radix_node *p, *tt, *x;
437     	/*
438 	 * Find first bit at which v and t->rn_key differ
439 	 */
440 	caddr_t cp2 = t->rn_key + head_off;
441 	int cmp_res;
442 	caddr_t cplim = v + vlen;
443 
444 	while (cp < cplim)
445 		if (*cp2++ != *cp++)
446 			goto on1;
447 	*dupentry = 1;
448 	return (t);
449 on1:
450 	*dupentry = 0;
451 	cmp_res = (cp[-1] ^ cp2[-1]) & 0xff;
452 	for (b = (cp - v) << 3; cmp_res; b--)
453 		cmp_res >>= 1;
454 
455 	x = top;
456 	cp = v;
457 	do {
458 		p = x;
459 		if (cp[x->rn_offset] & x->rn_bmask)
460 			x = x->rn_right;
461 		else
462 			x = x->rn_left;
463 	} while (b > (unsigned) x->rn_bit);
464 				/* x->rn_bit < b && x->rn_bit >= 0 */
465 #ifdef RN_DEBUG
466 	if (rn_debug)
467 		log(LOG_DEBUG, "rn_insert: Going In:\n"), traverse(p);
468 #endif
469 	t = rn_newpair(v_arg, b, nodes);
470 	tt = t->rn_left;
471 	if ((cp[p->rn_offset] & p->rn_bmask) == 0)
472 		p->rn_left = t;
473 	else
474 		p->rn_right = t;
475 	x->rn_parent = t;
476 	t->rn_parent = p; /* frees x, p as temp vars below */
477 	if ((cp[t->rn_offset] & t->rn_bmask) == 0) {
478 		t->rn_right = x;
479 	} else {
480 		t->rn_right = tt;
481 		t->rn_left = x;
482 	}
483 #ifdef RN_DEBUG
484 	if (rn_debug)
485 		log(LOG_DEBUG, "rn_insert: Coming Out:\n"), traverse(p);
486 #endif
487 	return (tt);
488 }
489 
490 struct radix_node *
491 rn_addmask(void *n_arg, struct radix_mask_head *maskhead, int search, int skip)
492 {
493 	unsigned char *netmask = n_arg;
494 	unsigned char *cp, *cplim;
495 	struct radix_node *x;
496 	int b = 0, mlen, j;
497 	int maskduplicated, isnormal;
498 	struct radix_node *saved_x;
499 	unsigned char addmask_key[RADIX_MAX_KEY_LEN];
500 
501 	if ((mlen = LEN(netmask)) > RADIX_MAX_KEY_LEN)
502 		mlen = RADIX_MAX_KEY_LEN;
503 	if (skip == 0)
504 		skip = 1;
505 	if (mlen <= skip)
506 		return (maskhead->mask_nodes);
507 
508 	bzero(addmask_key, RADIX_MAX_KEY_LEN);
509 	if (skip > 1)
510 		bcopy(rn_ones + 1, addmask_key + 1, skip - 1);
511 	bcopy(netmask + skip, addmask_key + skip, mlen - skip);
512 	/*
513 	 * Trim trailing zeroes.
514 	 */
515 	for (cp = addmask_key + mlen; (cp > addmask_key) && cp[-1] == 0;)
516 		cp--;
517 	mlen = cp - addmask_key;
518 	if (mlen <= skip)
519 		return (maskhead->mask_nodes);
520 	*addmask_key = mlen;
521 	x = rn_search(addmask_key, maskhead->head.rnh_treetop);
522 	if (bcmp(addmask_key, x->rn_key, mlen) != 0)
523 		x = NULL;
524 	if (x || search)
525 		return (x);
526 	R_Zalloc(x, struct radix_node *, RADIX_MAX_KEY_LEN + 2 * sizeof (*x));
527 	if ((saved_x = x) == NULL)
528 		return (0);
529 	netmask = cp = (unsigned char *)(x + 2);
530 	bcopy(addmask_key, cp, mlen);
531 	x = rn_insert(cp, &maskhead->head, &maskduplicated, x);
532 	if (maskduplicated) {
533 		log(LOG_ERR, "rn_addmask: mask impossibly already in tree");
534 		R_Free(saved_x);
535 		return (x);
536 	}
537 	/*
538 	 * Calculate index of mask, and check for normalcy.
539 	 * First find the first byte with a 0 bit, then if there are
540 	 * more bits left (remember we already trimmed the trailing 0's),
541 	 * the bits should be contiguous, otherwise we have got
542 	 * a non-contiguous mask.
543 	 */
544 #define	CONTIG(_c)	(((~(_c) + 1) & (_c)) == (unsigned char)(~(_c) + 1))
545 	cplim = netmask + mlen;
546 	isnormal = 1;
547 	for (cp = netmask + skip; (cp < cplim) && *(u_char *)cp == 0xff;)
548 		cp++;
549 	if (cp != cplim) {
550 		for (j = 0x80; (j & *cp) != 0; j >>= 1)
551 			b++;
552 		if (!CONTIG(*cp) || cp != (cplim - 1))
553 			isnormal = 0;
554 	}
555 	b += (cp - netmask) << 3;
556 	x->rn_bit = -1 - b;
557 	if (isnormal)
558 		x->rn_flags |= RNF_NORMAL;
559 	return (x);
560 }
561 
562 static int	/* XXX: arbitrary ordering for non-contiguous masks */
563 rn_lexobetter(void *m_arg, void *n_arg)
564 {
565 	u_char *mp = m_arg, *np = n_arg, *lim;
566 
567 	if (LEN(mp) > LEN(np))
568 		return (1);  /* not really, but need to check longer one first */
569 	if (LEN(mp) == LEN(np))
570 		for (lim = mp + LEN(mp); mp < lim;)
571 			if (*mp++ > *np++)
572 				return (1);
573 	return (0);
574 }
575 
576 static struct radix_mask *
577 rn_new_radix_mask(struct radix_node *tt, struct radix_mask *next)
578 {
579 	struct radix_mask *m;
580 
581 	R_Malloc(m, struct radix_mask *, sizeof (struct radix_mask));
582 	if (m == NULL) {
583 		log(LOG_ERR, "Failed to allocate route mask\n");
584 		return (0);
585 	}
586 	bzero(m, sizeof(*m));
587 	m->rm_bit = tt->rn_bit;
588 	m->rm_flags = tt->rn_flags;
589 	if (tt->rn_flags & RNF_NORMAL)
590 		m->rm_leaf = tt;
591 	else
592 		m->rm_mask = tt->rn_mask;
593 	m->rm_mklist = next;
594 	tt->rn_mklist = m;
595 	return (m);
596 }
597 
598 struct radix_node *
599 rn_addroute(void *v_arg, void *n_arg, struct radix_head *head,
600     struct radix_node treenodes[2])
601 {
602 	caddr_t v = (caddr_t)v_arg, netmask = (caddr_t)n_arg;
603 	struct radix_node *t, *x = NULL, *tt;
604 	struct radix_node *saved_tt, *top = head->rnh_treetop;
605 	short b = 0, b_leaf = 0;
606 	int keyduplicated;
607 	caddr_t mmask;
608 	struct radix_mask *m, **mp;
609 
610 	/*
611 	 * In dealing with non-contiguous masks, there may be
612 	 * many different routes which have the same mask.
613 	 * We will find it useful to have a unique pointer to
614 	 * the mask to speed avoiding duplicate references at
615 	 * nodes and possibly save time in calculating indices.
616 	 */
617 	if (netmask)  {
618 		x = rn_addmask(netmask, head->rnh_masks, 0, top->rn_offset);
619 		if (x == NULL)
620 			return (0);
621 		b_leaf = x->rn_bit;
622 		b = -1 - x->rn_bit;
623 		netmask = x->rn_key;
624 	}
625 	/*
626 	 * Deal with duplicated keys: attach node to previous instance
627 	 */
628 	saved_tt = tt = rn_insert(v, head, &keyduplicated, treenodes);
629 	if (keyduplicated) {
630 		for (t = tt; tt; t = tt, tt = tt->rn_dupedkey) {
631 #ifdef RADIX_MPATH
632 			/* permit multipath, if enabled for the family */
633 			if (rn_mpath_capable(head) && netmask == tt->rn_mask) {
634 				/*
635 				 * go down to the end of multipaths, so that
636 				 * new entry goes into the end of rn_dupedkey
637 				 * chain.
638 				 */
639 				do {
640 					t = tt;
641 					tt = tt->rn_dupedkey;
642 				} while (tt && t->rn_mask == tt->rn_mask);
643 				break;
644 			}
645 #endif
646 			if (tt->rn_mask == netmask)
647 				return (0);
648 			if (netmask == 0 ||
649 			    (tt->rn_mask &&
650 			     ((b_leaf < tt->rn_bit) /* index(netmask) > node */
651 			      || rn_refines(netmask, tt->rn_mask)
652 			      || rn_lexobetter(netmask, tt->rn_mask))))
653 				break;
654 		}
655 		/*
656 		 * If the mask is not duplicated, we wouldn't
657 		 * find it among possible duplicate key entries
658 		 * anyway, so the above test doesn't hurt.
659 		 *
660 		 * We sort the masks for a duplicated key the same way as
661 		 * in a masklist -- most specific to least specific.
662 		 * This may require the unfortunate nuisance of relocating
663 		 * the head of the list.
664 		 *
665 		 * We also reverse, or doubly link the list through the
666 		 * parent pointer.
667 		 */
668 		if (tt == saved_tt) {
669 			struct	radix_node *xx = x;
670 			/* link in at head of list */
671 			(tt = treenodes)->rn_dupedkey = t;
672 			tt->rn_flags = t->rn_flags;
673 			tt->rn_parent = x = t->rn_parent;
674 			t->rn_parent = tt;	 		/* parent */
675 			if (x->rn_left == t)
676 				x->rn_left = tt;
677 			else
678 				x->rn_right = tt;
679 			saved_tt = tt; x = xx;
680 		} else {
681 			(tt = treenodes)->rn_dupedkey = t->rn_dupedkey;
682 			t->rn_dupedkey = tt;
683 			tt->rn_parent = t;			/* parent */
684 			if (tt->rn_dupedkey)			/* parent */
685 				tt->rn_dupedkey->rn_parent = tt; /* parent */
686 		}
687 #ifdef RN_DEBUG
688 		t=tt+1; tt->rn_info = rn_nodenum++; t->rn_info = rn_nodenum++;
689 		tt->rn_twin = t; tt->rn_ybro = rn_clist; rn_clist = tt;
690 #endif
691 		tt->rn_key = (caddr_t) v;
692 		tt->rn_bit = -1;
693 		tt->rn_flags = RNF_ACTIVE;
694 	}
695 	/*
696 	 * Put mask in tree.
697 	 */
698 	if (netmask) {
699 		tt->rn_mask = netmask;
700 		tt->rn_bit = x->rn_bit;
701 		tt->rn_flags |= x->rn_flags & RNF_NORMAL;
702 	}
703 	t = saved_tt->rn_parent;
704 	if (keyduplicated)
705 		goto on2;
706 	b_leaf = -1 - t->rn_bit;
707 	if (t->rn_right == saved_tt)
708 		x = t->rn_left;
709 	else
710 		x = t->rn_right;
711 	/* Promote general routes from below */
712 	if (x->rn_bit < 0) {
713 	    for (mp = &t->rn_mklist; x; x = x->rn_dupedkey)
714 		if (x->rn_mask && (x->rn_bit >= b_leaf) && x->rn_mklist == 0) {
715 			*mp = m = rn_new_radix_mask(x, 0);
716 			if (m)
717 				mp = &m->rm_mklist;
718 		}
719 	} else if (x->rn_mklist) {
720 		/*
721 		 * Skip over masks whose index is > that of new node
722 		 */
723 		for (mp = &x->rn_mklist; (m = *mp); mp = &m->rm_mklist)
724 			if (m->rm_bit >= b_leaf)
725 				break;
726 		t->rn_mklist = m; *mp = NULL;
727 	}
728 on2:
729 	/* Add new route to highest possible ancestor's list */
730 	if ((netmask == 0) || (b > t->rn_bit ))
731 		return (tt); /* can't lift at all */
732 	b_leaf = tt->rn_bit;
733 	do {
734 		x = t;
735 		t = t->rn_parent;
736 	} while (b <= t->rn_bit && x != top);
737 	/*
738 	 * Search through routes associated with node to
739 	 * insert new route according to index.
740 	 * Need same criteria as when sorting dupedkeys to avoid
741 	 * double loop on deletion.
742 	 */
743 	for (mp = &x->rn_mklist; (m = *mp); mp = &m->rm_mklist) {
744 		if (m->rm_bit < b_leaf)
745 			continue;
746 		if (m->rm_bit > b_leaf)
747 			break;
748 		if (m->rm_flags & RNF_NORMAL) {
749 			mmask = m->rm_leaf->rn_mask;
750 			if (tt->rn_flags & RNF_NORMAL) {
751 #if !defined(RADIX_MPATH)
752 			    log(LOG_ERR,
753 			        "Non-unique normal route, mask not entered\n");
754 #endif
755 				return (tt);
756 			}
757 		} else
758 			mmask = m->rm_mask;
759 		if (mmask == netmask) {
760 			m->rm_refs++;
761 			tt->rn_mklist = m;
762 			return (tt);
763 		}
764 		if (rn_refines(netmask, mmask)
765 		    || rn_lexobetter(netmask, mmask))
766 			break;
767 	}
768 	*mp = rn_new_radix_mask(tt, *mp);
769 	return (tt);
770 }
771 
772 struct radix_node *
773 rn_delete(void *v_arg, void *netmask_arg, struct radix_head *head)
774 {
775 	struct radix_node *t, *p, *x, *tt;
776 	struct radix_mask *m, *saved_m, **mp;
777 	struct radix_node *dupedkey, *saved_tt, *top;
778 	caddr_t v, netmask;
779 	int b, head_off, vlen;
780 
781 	v = v_arg;
782 	netmask = netmask_arg;
783 	x = head->rnh_treetop;
784 	tt = rn_search(v, x);
785 	head_off = x->rn_offset;
786 	vlen =  LEN(v);
787 	saved_tt = tt;
788 	top = x;
789 	if (tt == NULL ||
790 	    bcmp(v + head_off, tt->rn_key + head_off, vlen - head_off))
791 		return (0);
792 	/*
793 	 * Delete our route from mask lists.
794 	 */
795 	if (netmask) {
796 		x = rn_addmask(netmask, head->rnh_masks, 1, head_off);
797 		if (x == NULL)
798 			return (0);
799 		netmask = x->rn_key;
800 		while (tt->rn_mask != netmask)
801 			if ((tt = tt->rn_dupedkey) == NULL)
802 				return (0);
803 	}
804 	if (tt->rn_mask == 0 || (saved_m = m = tt->rn_mklist) == NULL)
805 		goto on1;
806 	if (tt->rn_flags & RNF_NORMAL) {
807 		if (m->rm_leaf != tt || m->rm_refs > 0) {
808 			log(LOG_ERR, "rn_delete: inconsistent annotation\n");
809 			return (0);  /* dangling ref could cause disaster */
810 		}
811 	} else {
812 		if (m->rm_mask != tt->rn_mask) {
813 			log(LOG_ERR, "rn_delete: inconsistent annotation\n");
814 			goto on1;
815 		}
816 		if (--m->rm_refs >= 0)
817 			goto on1;
818 	}
819 	b = -1 - tt->rn_bit;
820 	t = saved_tt->rn_parent;
821 	if (b > t->rn_bit)
822 		goto on1; /* Wasn't lifted at all */
823 	do {
824 		x = t;
825 		t = t->rn_parent;
826 	} while (b <= t->rn_bit && x != top);
827 	for (mp = &x->rn_mklist; (m = *mp); mp = &m->rm_mklist)
828 		if (m == saved_m) {
829 			*mp = m->rm_mklist;
830 			R_Free(m);
831 			break;
832 		}
833 	if (m == NULL) {
834 		log(LOG_ERR, "rn_delete: couldn't find our annotation\n");
835 		if (tt->rn_flags & RNF_NORMAL)
836 			return (0); /* Dangling ref to us */
837 	}
838 on1:
839 	/*
840 	 * Eliminate us from tree
841 	 */
842 	if (tt->rn_flags & RNF_ROOT)
843 		return (0);
844 #ifdef RN_DEBUG
845 	/* Get us out of the creation list */
846 	for (t = rn_clist; t && t->rn_ybro != tt; t = t->rn_ybro) {}
847 	if (t) t->rn_ybro = tt->rn_ybro;
848 #endif
849 	t = tt->rn_parent;
850 	dupedkey = saved_tt->rn_dupedkey;
851 	if (dupedkey) {
852 		/*
853 		 * Here, tt is the deletion target and
854 		 * saved_tt is the head of the dupekey chain.
855 		 */
856 		if (tt == saved_tt) {
857 			/* remove from head of chain */
858 			x = dupedkey; x->rn_parent = t;
859 			if (t->rn_left == tt)
860 				t->rn_left = x;
861 			else
862 				t->rn_right = x;
863 		} else {
864 			/* find node in front of tt on the chain */
865 			for (x = p = saved_tt; p && p->rn_dupedkey != tt;)
866 				p = p->rn_dupedkey;
867 			if (p) {
868 				p->rn_dupedkey = tt->rn_dupedkey;
869 				if (tt->rn_dupedkey)		/* parent */
870 					tt->rn_dupedkey->rn_parent = p;
871 								/* parent */
872 			} else log(LOG_ERR, "rn_delete: couldn't find us\n");
873 		}
874 		t = tt + 1;
875 		if  (t->rn_flags & RNF_ACTIVE) {
876 #ifndef RN_DEBUG
877 			*++x = *t;
878 			p = t->rn_parent;
879 #else
880 			b = t->rn_info;
881 			*++x = *t;
882 			t->rn_info = b;
883 			p = t->rn_parent;
884 #endif
885 			if (p->rn_left == t)
886 				p->rn_left = x;
887 			else
888 				p->rn_right = x;
889 			x->rn_left->rn_parent = x;
890 			x->rn_right->rn_parent = x;
891 		}
892 		goto out;
893 	}
894 	if (t->rn_left == tt)
895 		x = t->rn_right;
896 	else
897 		x = t->rn_left;
898 	p = t->rn_parent;
899 	if (p->rn_right == t)
900 		p->rn_right = x;
901 	else
902 		p->rn_left = x;
903 	x->rn_parent = p;
904 	/*
905 	 * Demote routes attached to us.
906 	 */
907 	if (t->rn_mklist) {
908 		if (x->rn_bit >= 0) {
909 			for (mp = &x->rn_mklist; (m = *mp);)
910 				mp = &m->rm_mklist;
911 			*mp = t->rn_mklist;
912 		} else {
913 			/* If there are any key,mask pairs in a sibling
914 			   duped-key chain, some subset will appear sorted
915 			   in the same order attached to our mklist */
916 			for (m = t->rn_mklist; m && x; x = x->rn_dupedkey)
917 				if (m == x->rn_mklist) {
918 					struct radix_mask *mm = m->rm_mklist;
919 					x->rn_mklist = 0;
920 					if (--(m->rm_refs) < 0)
921 						R_Free(m);
922 					m = mm;
923 				}
924 			if (m)
925 				log(LOG_ERR,
926 				    "rn_delete: Orphaned Mask %p at %p\n",
927 				    m, x);
928 		}
929 	}
930 	/*
931 	 * We may be holding an active internal node in the tree.
932 	 */
933 	x = tt + 1;
934 	if (t != x) {
935 #ifndef RN_DEBUG
936 		*t = *x;
937 #else
938 		b = t->rn_info;
939 		*t = *x;
940 		t->rn_info = b;
941 #endif
942 		t->rn_left->rn_parent = t;
943 		t->rn_right->rn_parent = t;
944 		p = x->rn_parent;
945 		if (p->rn_left == x)
946 			p->rn_left = t;
947 		else
948 			p->rn_right = t;
949 	}
950 out:
951 	tt->rn_flags &= ~RNF_ACTIVE;
952 	tt[1].rn_flags &= ~RNF_ACTIVE;
953 	return (tt);
954 }
955 
956 /*
957  * This is the same as rn_walktree() except for the parameters and the
958  * exit.
959  */
960 int
961 rn_walktree_from(struct radix_head *h, void *a, void *m,
962     walktree_f_t *f, void *w)
963 {
964 	int error;
965 	struct radix_node *base, *next;
966 	u_char *xa = (u_char *)a;
967 	u_char *xm = (u_char *)m;
968 	struct radix_node *rn, *last = NULL; /* shut up gcc */
969 	int stopping = 0;
970 	int lastb;
971 
972 	KASSERT(m != NULL, ("%s: mask needs to be specified", __func__));
973 
974 	/*
975 	 * rn_search_m is sort-of-open-coded here. We cannot use the
976 	 * function because we need to keep track of the last node seen.
977 	 */
978 	/* printf("about to search\n"); */
979 	for (rn = h->rnh_treetop; rn->rn_bit >= 0; ) {
980 		last = rn;
981 		/* printf("rn_bit %d, rn_bmask %x, xm[rn_offset] %x\n",
982 		       rn->rn_bit, rn->rn_bmask, xm[rn->rn_offset]); */
983 		if (!(rn->rn_bmask & xm[rn->rn_offset])) {
984 			break;
985 		}
986 		if (rn->rn_bmask & xa[rn->rn_offset]) {
987 			rn = rn->rn_right;
988 		} else {
989 			rn = rn->rn_left;
990 		}
991 	}
992 	/* printf("done searching\n"); */
993 
994 	/*
995 	 * Two cases: either we stepped off the end of our mask,
996 	 * in which case last == rn, or we reached a leaf, in which
997 	 * case we want to start from the leaf.
998 	 */
999 	if (rn->rn_bit >= 0)
1000 		rn = last;
1001 	lastb = last->rn_bit;
1002 
1003 	/* printf("rn %p, lastb %d\n", rn, lastb);*/
1004 
1005 	/*
1006 	 * This gets complicated because we may delete the node
1007 	 * while applying the function f to it, so we need to calculate
1008 	 * the successor node in advance.
1009 	 */
1010 	while (rn->rn_bit >= 0)
1011 		rn = rn->rn_left;
1012 
1013 	while (!stopping) {
1014 		/* printf("node %p (%d)\n", rn, rn->rn_bit); */
1015 		base = rn;
1016 		/* If at right child go back up, otherwise, go right */
1017 		while (rn->rn_parent->rn_right == rn
1018 		       && !(rn->rn_flags & RNF_ROOT)) {
1019 			rn = rn->rn_parent;
1020 
1021 			/* if went up beyond last, stop */
1022 			if (rn->rn_bit <= lastb) {
1023 				stopping = 1;
1024 				/* printf("up too far\n"); */
1025 				/*
1026 				 * XXX we should jump to the 'Process leaves'
1027 				 * part, because the values of 'rn' and 'next'
1028 				 * we compute will not be used. Not a big deal
1029 				 * because this loop will terminate, but it is
1030 				 * inefficient and hard to understand!
1031 				 */
1032 			}
1033 		}
1034 
1035 		/*
1036 		 * At the top of the tree, no need to traverse the right
1037 		 * half, prevent the traversal of the entire tree in the
1038 		 * case of default route.
1039 		 */
1040 		if (rn->rn_parent->rn_flags & RNF_ROOT)
1041 			stopping = 1;
1042 
1043 		/* Find the next *leaf* since next node might vanish, too */
1044 		for (rn = rn->rn_parent->rn_right; rn->rn_bit >= 0;)
1045 			rn = rn->rn_left;
1046 		next = rn;
1047 		/* Process leaves */
1048 		while ((rn = base) != NULL) {
1049 			base = rn->rn_dupedkey;
1050 			/* printf("leaf %p\n", rn); */
1051 			if (!(rn->rn_flags & RNF_ROOT)
1052 			    && (error = (*f)(rn, w)))
1053 				return (error);
1054 		}
1055 		rn = next;
1056 
1057 		if (rn->rn_flags & RNF_ROOT) {
1058 			/* printf("root, stopping"); */
1059 			stopping = 1;
1060 		}
1061 	}
1062 	return (0);
1063 }
1064 
1065 int
1066 rn_walktree(struct radix_head *h, walktree_f_t *f, void *w)
1067 {
1068 	int error;
1069 	struct radix_node *base, *next;
1070 	struct radix_node *rn = h->rnh_treetop;
1071 	/*
1072 	 * This gets complicated because we may delete the node
1073 	 * while applying the function f to it, so we need to calculate
1074 	 * the successor node in advance.
1075 	 */
1076 
1077 	/* First time through node, go left */
1078 	while (rn->rn_bit >= 0)
1079 		rn = rn->rn_left;
1080 	for (;;) {
1081 		base = rn;
1082 		/* If at right child go back up, otherwise, go right */
1083 		while (rn->rn_parent->rn_right == rn
1084 		       && (rn->rn_flags & RNF_ROOT) == 0)
1085 			rn = rn->rn_parent;
1086 		/* Find the next *leaf* since next node might vanish, too */
1087 		for (rn = rn->rn_parent->rn_right; rn->rn_bit >= 0;)
1088 			rn = rn->rn_left;
1089 		next = rn;
1090 		/* Process leaves */
1091 		while ((rn = base)) {
1092 			base = rn->rn_dupedkey;
1093 			if (!(rn->rn_flags & RNF_ROOT)
1094 			    && (error = (*f)(rn, w)))
1095 				return (error);
1096 		}
1097 		rn = next;
1098 		if (rn->rn_flags & RNF_ROOT)
1099 			return (0);
1100 	}
1101 	/* NOTREACHED */
1102 }
1103 
1104 /*
1105  * Initialize an empty tree. This has 3 nodes, which are passed
1106  * via base_nodes (in the order <left,root,right>) and are
1107  * marked RNF_ROOT so they cannot be freed.
1108  * The leaves have all-zero and all-one keys, with significant
1109  * bits starting at 'off'.
1110  */
1111 void
1112 rn_inithead_internal(struct radix_head *rh, struct radix_node *base_nodes, int off)
1113 {
1114 	struct radix_node *t, *tt, *ttt;
1115 
1116 	t = rn_newpair(rn_zeros, off, base_nodes);
1117 	ttt = base_nodes + 2;
1118 	t->rn_right = ttt;
1119 	t->rn_parent = t;
1120 	tt = t->rn_left;	/* ... which in turn is base_nodes */
1121 	tt->rn_flags = t->rn_flags = RNF_ROOT | RNF_ACTIVE;
1122 	tt->rn_bit = -1 - off;
1123 	*ttt = *tt;
1124 	ttt->rn_key = rn_ones;
1125 
1126 	rh->rnh_treetop = t;
1127 }
1128 
1129 static void
1130 rn_detachhead_internal(struct radix_head *head)
1131 {
1132 
1133 	KASSERT((head != NULL),
1134 	    ("%s: head already freed", __func__));
1135 
1136 	/* Free <left,root,right> nodes. */
1137 	R_Free(head);
1138 }
1139 
1140 /* Functions used by 'struct radix_node_head' users */
1141 
1142 int
1143 rn_inithead(void **head, int off)
1144 {
1145 	struct radix_node_head *rnh;
1146 	struct radix_mask_head *rmh;
1147 
1148 	rnh = *head;
1149 	rmh = NULL;
1150 
1151 	if (*head != NULL)
1152 		return (1);
1153 
1154 	R_Zalloc(rnh, struct radix_node_head *, sizeof (*rnh));
1155 	R_Zalloc(rmh, struct radix_mask_head *, sizeof (*rmh));
1156 	if (rnh == NULL || rmh == NULL) {
1157 		if (rnh != NULL)
1158 			R_Free(rnh);
1159 		if (rmh != NULL)
1160 			R_Free(rmh);
1161 		return (0);
1162 	}
1163 
1164 	/* Init trees */
1165 	rn_inithead_internal(&rnh->rh, rnh->rnh_nodes, off);
1166 	rn_inithead_internal(&rmh->head, rmh->mask_nodes, 0);
1167 	*head = rnh;
1168 	rnh->rh.rnh_masks = rmh;
1169 
1170 	/* Finally, set base callbacks */
1171 	rnh->rnh_addaddr = rn_addroute;
1172 	rnh->rnh_deladdr = rn_delete;
1173 	rnh->rnh_matchaddr = rn_match;
1174 	rnh->rnh_lookup = rn_lookup;
1175 	rnh->rnh_walktree = rn_walktree;
1176 	rnh->rnh_walktree_from = rn_walktree_from;
1177 
1178 	return (1);
1179 }
1180 
1181 static int
1182 rn_freeentry(struct radix_node *rn, void *arg)
1183 {
1184 	struct radix_head * const rnh = arg;
1185 	struct radix_node *x;
1186 
1187 	x = (struct radix_node *)rn_delete(rn + 2, NULL, rnh);
1188 	if (x != NULL)
1189 		R_Free(x);
1190 	return (0);
1191 }
1192 
1193 int
1194 rn_detachhead(void **head)
1195 {
1196 	struct radix_node_head *rnh;
1197 
1198 	KASSERT((head != NULL && *head != NULL),
1199 	    ("%s: head already freed", __func__));
1200 
1201 	rnh = (struct radix_node_head *)(*head);
1202 
1203 	rn_walktree(&rnh->rh.rnh_masks->head, rn_freeentry, rnh->rh.rnh_masks);
1204 	rn_detachhead_internal(&rnh->rh.rnh_masks->head);
1205 	rn_detachhead_internal(&rnh->rh);
1206 
1207 	*head = NULL;
1208 
1209 	return (1);
1210 }
1211