sys/net/radix.c

   1 /*
   2  * Copyright (c) 1988, 1989, 1993
   3  *      The Regents of the University of California.  All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  * 3. Neither the name of the University nor the names of its contributors
  14  *    may be used to endorse or promote products derived from this software
  15  *    without specific prior written permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  *
  29  *      @(#)radix.c     8.4 (Berkeley) 11/2/94
  30  * $FreeBSD: src/sys/net/radix.c,v 1.20.2.3 2002/04/28 05:40:25 suz Exp $
  31  */
  32
  33 /*
  34  * Routines to build and maintain radix trees for routing lookups.
  35  */
  36 #include <sys/param.h>
  37 #ifdef  _KERNEL
  38 #include <sys/systm.h>
  39 #include <sys/malloc.h>
  40 #include <sys/domain.h>
  41 #include <sys/globaldata.h>
  42 #include <sys/thread.h>
  43 #else
  44 #include <stdlib.h>
  45 #endif
  46 #include <sys/syslog.h>
  47 #include <net/radix.h>
  48
  49 /*
  50  * The arguments to the radix functions are really counted byte arrays with
  51  * the length in the first byte.  struct sockaddr's fit this type structurally.
  52  */
  53 #define clen(c) (*(u_char *)(c))
  54
  55 static int rn_walktree_from(struct radix_node_head *h, char *a, char *m,
  56                             walktree_f_t *f, void *w);
  57 static int rn_walktree(struct radix_node_head *, walktree_f_t *, void *);
  58 static int rn_walktree_at(struct radix_node_head *h, const char *a,
  59                             const char *m, walktree_f_t *f, void *w);
  60
  61 static struct radix_node
  62     *rn_insert(char *, struct radix_node_head *, boolean_t *,
  63                struct radix_node [2]),
  64     *rn_newpair(char *, int, struct radix_node[2]),
  65     *rn_search(const char *, struct radix_node *),
  66     *rn_search_m(const char *, struct radix_node *, const char *);
  67
  68 static struct radix_mask *rn_mkfreelist[MAXCPU];
  69 static struct radix_node_head *mask_rnheads[MAXCPU];
  70
  71 static char rn_zeros[RN_MAXKEYLEN];
  72 static char rn_ones[RN_MAXKEYLEN] = RN_MAXKEYONES;
  73
  74 static boolean_t rn_lexobetter(char *m, char *n);
  75 static struct radix_mask *
  76     rn_new_radix_mask(struct radix_node *tt, struct radix_mask *nextmask);
  77 static boolean_t
  78     rn_satisfies_leaf(char *trial, struct radix_node *leaf, int skip);
  79
  80 static __inline struct radix_mask *
  81 MKGet(struct radix_mask **l)
  82 {
  83         struct radix_mask *m;
  84
  85         if (*l != NULL) {
  86                 m = *l;
  87                 *l = m->rm_next;
  88         } else {
  89                 R_Malloc(m, struct radix_mask *, sizeof *m);
  90         }
  91         return m;
  92 }
  93
  94 static __inline void
  95 MKFree(struct radix_mask **l, struct radix_mask *m)
  96 {
  97         m->rm_next = *l;
  98         *l = m;
  99 }
 100
 101 /*
 102  * The data structure for the keys is a radix tree with one way
 103  * branching removed.  The index rn_bit at an internal node n represents a bit
 104  * position to be tested.  The tree is arranged so that all descendants
 105  * of a node n have keys whose bits all agree up to position rn_bit - 1.
 106  * (We say the index of n is rn_bit.)
 107  *
 108  * There is at least one descendant which has a one bit at position rn_bit,
 109  * and at least one with a zero there.
 110  *
 111  * A route is determined by a pair of key and mask.  We require that the
 112  * bit-wise logical and of the key and mask to be the key.
 113  * We define the index of a route to associated with the mask to be
 114  * the first bit number in the mask where 0 occurs (with bit number 0
 115  * representing the highest order bit).
 116  *
 117  * We say a mask is normal if every bit is 0, past the index of the mask.
 118  * If a node n has a descendant (k, m) with index(m) == index(n) == rn_bit,
 119  * and m is a normal mask, then the route applies to every descendant of n.
 120  * If the index(m) < rn_bit, this implies the trailing last few bits of k
 121  * before bit b are all 0, (and hence consequently true of every descendant
 122  * of n), so the route applies to all descendants of the node as well.
 123  *
 124  * Similar logic shows that a non-normal mask m such that
 125  * index(m) <= index(n) could potentially apply to many children of n.
 126  * Thus, for each non-host route, we attach its mask to a list at an internal
 127  * node as high in the tree as we can go.
 128  *
 129  * The present version of the code makes use of normal routes in short-
 130  * circuiting an explict mask and compare operation when testing whether
 131  * a key satisfies a normal route, and also in remembering the unique leaf
 132  * that governs a subtree.
 133  */
 134
 135 static struct radix_node *
 136 rn_search(const char *v, struct radix_node *head)
 137 {
 138         struct radix_node *x;
 139
 140         x = head;
 141         while (x->rn_bit >= 0) {
 142                 if (x->rn_bmask & v[x->rn_offset])
 143                         x = x->rn_right;
 144                 else
 145                         x = x->rn_left;
 146         }
 147         return (x);
 148 }
 149
 150 static struct radix_node *
 151 rn_search_m(const char *v, struct radix_node *head, const char *m)
 152 {
 153         struct radix_node *x;
 154
 155         for (x = head; x->rn_bit >= 0;) {
 156                 if ((x->rn_bmask & m[x->rn_offset]) &&
 157                     (x->rn_bmask & v[x->rn_offset]))
 158                         x = x->rn_right;
 159                 else
 160                         x = x->rn_left;
 161         }
 162         return x;
 163 }
 164
 165 boolean_t
 166 rn_refines(char *m, char *n)
 167 {
 168         char *lim, *lim2;
 169         int longer = clen(n++) - clen(m++);
 170         boolean_t masks_are_equal = TRUE;
 171
 172         lim2 = lim = n + clen(n);
 173         if (longer > 0)
 174                 lim -= longer;
 175         while (n < lim) {
 176                 if (*n & ~(*m))
 177                         return FALSE;
 178                 if (*n++ != *m++)
 179                         masks_are_equal = FALSE;
 180         }
 181         while (n < lim2)
 182                 if (*n++)
 183                         return FALSE;
 184         if (masks_are_equal && (longer < 0))
 185                 for (lim2 = m - longer; m < lim2; )
 186                         if (*m++)
 187                                 return TRUE;
 188         return (!masks_are_equal);
 189 }
 190
 191 struct radix_node *
 192 rn_lookup(char *key, char *mask, struct radix_node_head *head)
 193 {
 194         struct radix_node *x;
 195         char *netmask = NULL;
 196
 197         if (mask != NULL) {
 198                 x = rn_addmask(mask, TRUE, head->rnh_treetop->rn_offset,
 199                                head->rnh_maskhead);
 200                 if (x == NULL)
 201                         return (NULL);
 202                 netmask = x->rn_key;
 203         }
 204         x = rn_match(key, head);
 205         if (x != NULL && netmask != NULL) {
 206                 while (x != NULL && x->rn_mask != netmask)
 207                         x = x->rn_dupedkey;
 208         }
 209         return x;
 210 }
 211
 212 static boolean_t
 213 rn_satisfies_leaf(char *trial, struct radix_node *leaf, int skip)
 214 {
 215         char *cp = trial, *cp2 = leaf->rn_key, *cp3 = leaf->rn_mask;
 216         char *cplim;
 217         int length = min(clen(cp), clen(cp2));
 218
 219         if (cp3 == NULL)
 220                 cp3 = rn_ones;
 221         else
 222                 length = min(length, clen(cp3));
 223         cplim = cp + length;
 224         cp3 += skip;
 225         cp2 += skip;
 226         for (cp += skip; cp < cplim; cp++, cp2++, cp3++)
 227                 if ((*cp ^ *cp2) & *cp3)
 228                         return FALSE;
 229         return TRUE;
 230 }
 231
 232 struct radix_node *
 233 rn_match(char *key, struct radix_node_head *head)
 234 {
 235         struct radix_node *t, *x;
 236         char *cp = key, *cp2;
 237         char *cplim;
 238         struct radix_node *saved_t, *top = head->rnh_treetop;
 239         int off = top->rn_offset, klen, matched_off;
 240         int test, b, rn_bit;
 241
 242         t = rn_search(key, top);
 243         /*
 244          * See if we match exactly as a host destination
 245          * or at least learn how many bits match, for normal mask finesse.
 246          *
 247          * It doesn't hurt us to limit how many bytes to check
 248          * to the length of the mask, since if it matches we had a genuine
 249          * match and the leaf we have is the most specific one anyway;
 250          * if it didn't match with a shorter length it would fail
 251          * with a long one.  This wins big for class B&C netmasks which
 252          * are probably the most common case...
 253          */
 254         if (t->rn_mask != NULL)
 255                 klen = clen(t->rn_mask);
 256         else
 257                 klen = clen(key);
 258         cp += off; cp2 = t->rn_key + off; cplim = key + klen;
 259         for (; cp < cplim; cp++, cp2++)
 260                 if (*cp != *cp2)
 261                         goto on1;
 262         /*
 263          * This extra grot is in case we are explicitly asked
 264          * to look up the default.  Ugh!
 265          *
 266          * Never return the root node itself, it seems to cause a
 267          * lot of confusion.
 268          */
 269         if (t->rn_flags & RNF_ROOT)
 270                 t = t->rn_dupedkey;
 271         return t;
 272 on1:
 273         test = (*cp ^ *cp2) & 0xff; /* find first bit that differs */
 274         for (b = 7; (test >>= 1) > 0;)
 275                 b--;
 276         matched_off = cp - key;
 277         b += matched_off << 3;
 278         rn_bit = -1 - b;
 279         /*
 280          * If there is a host route in a duped-key chain, it will be first.
 281          */
 282         if ((saved_t = t)->rn_mask == NULL)
 283                 t = t->rn_dupedkey;
 284         for (; t; t = t->rn_dupedkey) {
 285                 /*
 286                  * Even if we don't match exactly as a host,
 287                  * we may match if the leaf we wound up at is
 288                  * a route to a net.
 289                  */
 290                 if (t->rn_flags & RNF_NORMAL) {
 291                         if (rn_bit <= t->rn_bit)
 292                                 return t;
 293                 } else if (rn_satisfies_leaf(key, t, matched_off))
 294                                 return t;
 295         }
 296         t = saved_t;
 297         /* start searching up the tree */
 298         do {
 299                 struct radix_mask *m;
 300
 301                 t = t->rn_parent;
 302                 /*
 303                  * If non-contiguous masks ever become important
 304                  * we can restore the masking and open coding of
 305                  * the search and satisfaction test and put the
 306                  * calculation of "off" back before the "do".
 307                  */
 308                 m = t->rn_mklist;
 309                 while (m != NULL) {
 310                         if (m->rm_flags & RNF_NORMAL) {
 311                                 if (rn_bit <= m->rm_bit)
 312                                         return (m->rm_leaf);
 313                         } else {
 314                                 off = min(t->rn_offset, matched_off);
 315                                 x = rn_search_m(key, t, m->rm_mask);
 316                                 while (x != NULL && x->rn_mask != m->rm_mask)
 317                                         x = x->rn_dupedkey;
 318                                 if (x && rn_satisfies_leaf(key, x, off))
 319                                         return x;
 320                         }
 321                         m = m->rm_next;
 322                 }
 323         } while (t != top);
 324         return NULL;
 325 }
 326
 327 #ifdef RN_DEBUG
 328 int rn_nodenum;
 329 struct radix_node *rn_clist;
 330 int rn_saveinfo;
 331 boolean_t rn_debug =  TRUE;
 332 #endif
 333
 334 static struct radix_node *
 335 rn_newpair(char *key, int indexbit, struct radix_node nodes[2])
 336 {
 337         struct radix_node *leaf = &nodes[0], *interior = &nodes[1];
 338
 339         interior->rn_bit = indexbit;
 340         interior->rn_bmask = 0x80 >> (indexbit & 0x7);
 341         interior->rn_offset = indexbit >> 3;
 342         interior->rn_left = leaf;
 343         interior->rn_mklist = NULL;
 344
 345         leaf->rn_bit = -1;
 346         leaf->rn_key = key;
 347         leaf->rn_parent = interior;
 348         leaf->rn_flags = interior->rn_flags = RNF_ACTIVE;
 349         leaf->rn_mklist = NULL;
 350
 351 #ifdef RN_DEBUG
 352         leaf->rn_info = rn_nodenum++;
 353         interior->rn_info = rn_nodenum++;
 354         leaf->rn_twin = interior;
 355         leaf->rn_ybro = rn_clist;
 356         rn_clist = leaf;
 357 #endif
 358         return interior;
 359 }
 360
 361 static struct radix_node *
 362 rn_insert(char *key, struct radix_node_head *head, boolean_t *dupentry,
 363           struct radix_node nodes[2])
 364 {
 365         struct radix_node *top = head->rnh_treetop;
 366         int head_off = top->rn_offset, klen = clen(key);
 367         struct radix_node *t = rn_search(key, top);
 368         char *cp = key + head_off;
 369         int b;
 370         struct radix_node *tt;
 371
 372         /*
 373          * Find first bit at which the key and t->rn_key differ
 374          */
 375     {
 376         char *cp2 = t->rn_key + head_off;
 377         int cmp_res;
 378         char *cplim = key + klen;
 379
 380         while (cp < cplim)
 381                 if (*cp2++ != *cp++)
 382                         goto on1;
 383         *dupentry = TRUE;
 384         return t;
 385 on1:
 386         *dupentry = FALSE;
 387         cmp_res = (cp[-1] ^ cp2[-1]) & 0xff;
 388         for (b = (cp - key) << 3; cmp_res; b--)
 389                 cmp_res >>= 1;
 390     }
 391     {
 392         struct radix_node *p, *x = top;
 393
 394         cp = key;
 395         do {
 396                 p = x;
 397                 if (cp[x->rn_offset] & x->rn_bmask)
 398                         x = x->rn_right;
 399                 else
 400                         x = x->rn_left;
 401         } while (b > (unsigned) x->rn_bit);
 402                                 /* x->rn_bit < b && x->rn_bit >= 0 */
 403 #ifdef RN_DEBUG
 404         if (rn_debug)
 405                 log(LOG_DEBUG, "rn_insert: Going In:\n"), traverse(p);
 406 #endif
 407         t = rn_newpair(key, b, nodes);
 408         tt = t->rn_left;
 409         if ((cp[p->rn_offset] & p->rn_bmask) == 0)
 410                 p->rn_left = t;
 411         else
 412                 p->rn_right = t;
 413         x->rn_parent = t;
 414         t->rn_parent = p; /* frees x, p as temp vars below */
 415         if ((cp[t->rn_offset] & t->rn_bmask) == 0) {
 416                 t->rn_right = x;
 417         } else {
 418                 t->rn_right = tt;
 419                 t->rn_left = x;
 420         }
 421 #ifdef RN_DEBUG
 422         if (rn_debug)
 423                 log(LOG_DEBUG, "rn_insert: Coming Out:\n"), traverse(p);
 424 #endif
 425     }
 426         return (tt);
 427 }
 428
 429 struct radix_node *
 430 rn_addmask(char *netmask, boolean_t search, int skip,
 431            struct radix_node_head *mask_rnh)
 432 {
 433         struct radix_node *x, *saved_x;
 434         char *cp, *cplim;
 435         int b = 0, mlen, m0, j;
 436         boolean_t maskduplicated, isnormal;
 437         char *addmask_key;
 438
 439         if ((mlen = clen(netmask)) > RN_MAXKEYLEN)
 440                 mlen = RN_MAXKEYLEN;
 441         if (skip == 0)
 442                 skip = 1;
 443         if (mlen <= skip)
 444                 return (mask_rnh->rnh_nodes);
 445         R_Malloc(addmask_key, char *, RN_MAXKEYLEN);
 446         if (addmask_key == NULL)
 447                 return NULL;
 448         if (skip > 1)
 449                 bcopy(rn_ones + 1, addmask_key + 1, skip - 1);
 450         if ((m0 = mlen) > skip)
 451                 bcopy(netmask + skip, addmask_key + skip, mlen - skip);
 452         /*
 453          * Trim trailing zeroes.
 454          */
 455         for (cp = addmask_key + mlen; (cp > addmask_key) && cp[-1] == 0;)
 456                 cp--;
 457         mlen = cp - addmask_key;
 458         if (mlen <= skip) {
 459                 if (m0 >= mask_rnh->rnh_last_zeroed)
 460                         mask_rnh->rnh_last_zeroed = mlen;
 461                 Free(addmask_key);
 462                 return (mask_rnh->rnh_nodes);
 463         }
 464         if (m0 < mask_rnh->rnh_last_zeroed)
 465                 bzero(addmask_key + m0, mask_rnh->rnh_last_zeroed - m0);
 466         *addmask_key = mask_rnh->rnh_last_zeroed = mlen;
 467         x = rn_search(addmask_key, mask_rnh->rnh_treetop);
 468         if (x->rn_key == NULL) {
 469                 kprintf("WARNING: radix_node->rn_key is NULL rn=%p\n", x);
 470                 print_backtrace(-1);
 471                 x = NULL;
 472         } else if (bcmp(addmask_key, x->rn_key, mlen) != 0) {
 473                 x = NULL;
 474         }
 475         if (x != NULL || search)
 476                 goto out;
 477         R_Malloc(x, struct radix_node *, RN_MAXKEYLEN + 2 * (sizeof *x));
 478         if ((saved_x = x) == NULL)
 479                 goto out;
 480         bzero(x, RN_MAXKEYLEN + 2 * (sizeof *x));
 481         netmask = cp = (char *)(x + 2);
 482         bcopy(addmask_key, cp, mlen);
 483         x = rn_insert(cp, mask_rnh, &maskduplicated, x);
 484         if (maskduplicated) {
 485                 log(LOG_ERR, "rn_addmask: mask impossibly already in tree");
 486                 Free(saved_x);
 487                 goto out;
 488         }
 489         /*
 490          * Calculate index of mask, and check for normalcy.
 491          */
 492         isnormal = TRUE;
 493         cplim = netmask + mlen;
 494         for (cp = netmask + skip; cp < cplim && clen(cp) == 0xff;)
 495                 cp++;
 496         if (cp != cplim) {
 497                 static const char normal_chars[] = {
 498                         0, 0x80, 0xc0, 0xe0, 0xf0, 0xf8, 0xfc, 0xfe, -1
 499                 };
 500
 501                 for (j = 0x80; (j & *cp) != 0; j >>= 1)
 502                         b++;
 503                 if (*cp != normal_chars[b] || cp != (cplim - 1))
 504                         isnormal = FALSE;
 505         }
 506         b += (cp - netmask) << 3;
 507         x->rn_bit = -1 - b;
 508         if (isnormal)
 509                 x->rn_flags |= RNF_NORMAL;
 510 out:
 511         Free(addmask_key);
 512         return (x);
 513 }
 514
 515 /* XXX: arbitrary ordering for non-contiguous masks */
 516 static boolean_t
 517 rn_lexobetter(char *mp, char *np)
 518 {
 519         char *lim;
 520
 521         if ((unsigned) *mp > (unsigned) *np)
 522                 return TRUE;/* not really, but need to check longer one first */
 523         if (*mp == *np)
 524                 for (lim = mp + clen(mp); mp < lim;)
 525                         if (*mp++ > *np++)
 526                                 return TRUE;
 527         return FALSE;
 528 }
 529
 530 static struct radix_mask *
 531 rn_new_radix_mask(struct radix_node *tt, struct radix_mask *nextmask)
 532 {
 533         struct radix_mask *m;
 534
 535         m = MKGet(&rn_mkfreelist[mycpuid]);
 536         if (m == NULL) {
 537                 log(LOG_ERR, "Mask for route not entered\n");
 538                 return (NULL);
 539         }
 540         bzero(m, sizeof *m);
 541         m->rm_bit = tt->rn_bit;
 542         m->rm_flags = tt->rn_flags;
 543         if (tt->rn_flags & RNF_NORMAL)
 544                 m->rm_leaf = tt;
 545         else
 546                 m->rm_mask = tt->rn_mask;
 547         m->rm_next = nextmask;
 548         tt->rn_mklist = m;
 549         return m;
 550 }
 551
 552 struct radix_node *
 553 rn_addroute(char *key, char *netmask, struct radix_node_head *head,
 554             struct radix_node treenodes[2])
 555 {
 556         struct radix_node *t, *x = NULL, *tt;
 557         struct radix_node *saved_tt, *top = head->rnh_treetop;
 558         short b = 0, b_leaf = 0;
 559         boolean_t keyduplicated;
 560         char *mmask;
 561         struct radix_mask *m, **mp;
 562
 563         /*
 564          * In dealing with non-contiguous masks, there may be
 565          * many different routes which have the same mask.
 566          * We will find it useful to have a unique pointer to
 567          * the mask to speed avoiding duplicate references at
 568          * nodes and possibly save time in calculating indices.
 569          */
 570         if (netmask != NULL)  {
 571                 if ((x = rn_addmask(netmask, FALSE, top->rn_offset,
 572                                     head->rnh_maskhead)) == NULL)
 573                         return (NULL);
 574                 b_leaf = x->rn_bit;
 575                 b = -1 - x->rn_bit;
 576                 netmask = x->rn_key;
 577         }
 578         /*
 579          * Deal with duplicated keys: attach node to previous instance
 580          */
 581         saved_tt = tt = rn_insert(key, head, &keyduplicated, treenodes);
 582         if (keyduplicated) {
 583                 for (t = tt; tt; t = tt, tt = tt->rn_dupedkey) {
 584                         if (tt->rn_mask == netmask)
 585                                 return (NULL);
 586                         if (netmask == NULL ||
 587                             (tt->rn_mask &&
 588                              ((b_leaf < tt->rn_bit) /* index(netmask) > node */
 589                               || rn_refines(netmask, tt->rn_mask)
 590                               || rn_lexobetter(netmask, tt->rn_mask))))
 591                                 break;
 592                 }
 593                 /*
 594                  * If the mask is not duplicated, we wouldn't
 595                  * find it among possible duplicate key entries
 596                  * anyway, so the above test doesn't hurt.
 597                  *
 598                  * We sort the masks for a duplicated key the same way as
 599                  * in a masklist -- most specific to least specific.
 600                  * This may require the unfortunate nuisance of relocating
 601                  * the head of the list.
 602                  */
 603                 if (tt == saved_tt) {
 604                         struct  radix_node *xx = x;
 605                         /* link in at head of list */
 606                         (tt = treenodes)->rn_dupedkey = t;
 607                         tt->rn_flags = t->rn_flags;
 608                         tt->rn_parent = x = t->rn_parent;
 609                         t->rn_parent = tt;                      /* parent */
 610                         if (x->rn_left == t)
 611                                 x->rn_left = tt;
 612                         else
 613                                 x->rn_right = tt;
 614                         saved_tt = tt; x = xx;
 615                 } else {
 616                         (tt = treenodes)->rn_dupedkey = t->rn_dupedkey;
 617                         t->rn_dupedkey = tt;
 618                         tt->rn_parent = t;                      /* parent */
 619                         if (tt->rn_dupedkey != NULL)            /* parent */
 620                                 tt->rn_dupedkey->rn_parent = tt; /* parent */
 621                 }
 622 #ifdef RN_DEBUG
 623                 t=tt+1; tt->rn_info = rn_nodenum++; t->rn_info = rn_nodenum++;
 624                 tt->rn_twin = t; tt->rn_ybro = rn_clist; rn_clist = tt;
 625 #endif
 626                 tt->rn_key = key;
 627                 tt->rn_bit = -1;
 628                 tt->rn_flags = RNF_ACTIVE;
 629         }
 630         /*
 631          * Put mask in tree.
 632          */
 633         if (netmask != NULL) {
 634                 tt->rn_mask = netmask;
 635                 tt->rn_bit = x->rn_bit;
 636                 tt->rn_flags |= x->rn_flags & RNF_NORMAL;
 637         }
 638         t = saved_tt->rn_parent;
 639         if (keyduplicated)
 640                 goto on2;
 641         b_leaf = -1 - t->rn_bit;
 642         if (t->rn_right == saved_tt)
 643                 x = t->rn_left;
 644         else
 645                 x = t->rn_right;
 646         /* Promote general routes from below */
 647         if (x->rn_bit < 0) {
 648                 mp = &t->rn_mklist;
 649                 while (x != NULL) {
 650                         if (x->rn_mask != NULL &&
 651                             x->rn_bit >= b_leaf &&
 652                             x->rn_mklist == NULL) {
 653                                 *mp = m = rn_new_radix_mask(x, NULL);
 654                                 if (m != NULL)
 655                                         mp = &m->rm_next;
 656                         }
 657                         x = x->rn_dupedkey;
 658                 }
 659         } else if (x->rn_mklist != NULL) {
 660                 /*
 661                  * Skip over masks whose index is > that of new node
 662                  */
 663                 for (mp = &x->rn_mklist; (m = *mp); mp = &m->rm_next)
 664                         if (m->rm_bit >= b_leaf)
 665                                 break;
 666                 t->rn_mklist = m;
 667                 *mp = NULL;
 668         }
 669 on2:
 670         /* Add new route to highest possible ancestor's list */
 671         if ((netmask == NULL) || (b > t->rn_bit ))
 672                 return tt; /* can't lift at all */
 673         b_leaf = tt->rn_bit;
 674         do {
 675                 x = t;
 676                 t = t->rn_parent;
 677         } while (b <= t->rn_bit && x != top);
 678         /*
 679          * Search through routes associated with node to
 680          * insert new route according to index.
 681          * Need same criteria as when sorting dupedkeys to avoid
 682          * double loop on deletion.
 683          */
 684         for (mp = &x->rn_mklist; (m = *mp); mp = &m->rm_next) {
 685                 if (m->rm_bit < b_leaf)
 686                         continue;
 687                 if (m->rm_bit > b_leaf)
 688                         break;
 689                 if (m->rm_flags & RNF_NORMAL) {
 690                         mmask = m->rm_leaf->rn_mask;
 691                         if (tt->rn_flags & RNF_NORMAL) {
 692                             log(LOG_ERR,
 693                                 "Non-unique normal route, mask not entered\n");
 694                                 return tt;
 695                         }
 696                 } else
 697                         mmask = m->rm_mask;
 698                 if (mmask == netmask) {
 699                         m->rm_refs++;
 700                         tt->rn_mklist = m;
 701                         return tt;
 702                 }
 703                 if (rn_refines(netmask, mmask) || rn_lexobetter(netmask, mmask))
 704                         break;
 705         }
 706         *mp = rn_new_radix_mask(tt, *mp);
 707         return tt;
 708 }
 709
 710 struct radix_node *
 711 rn_delete(char *key, char *netmask, struct radix_node_head *head)
 712 {
 713         struct radix_node *t, *p, *x, *tt;
 714         struct radix_mask *m, *saved_m, **mp;
 715         struct radix_node *dupedkey, *saved_tt, *top;
 716         int b, head_off, klen;
 717         int cpu = mycpuid;
 718
 719         x = head->rnh_treetop;
 720         tt = rn_search(key, x);
 721         head_off = x->rn_offset;
 722         klen =  clen(key);
 723         saved_tt = tt;
 724         top = x;
 725         if (tt == NULL ||
 726             bcmp(key + head_off, tt->rn_key + head_off, klen - head_off))
 727                 return (NULL);
 728         /*
 729          * Delete our route from mask lists.
 730          */
 731         if (netmask != NULL) {
 732                 if ((x = rn_addmask(netmask, TRUE, head_off,
 733                                     head->rnh_maskhead)) == NULL)
 734                         return (NULL);
 735                 netmask = x->rn_key;
 736                 while (tt->rn_mask != netmask)
 737                         if ((tt = tt->rn_dupedkey) == NULL)
 738                                 return (NULL);
 739         }
 740         if (tt->rn_mask == NULL || (saved_m = m = tt->rn_mklist) == NULL)
 741                 goto on1;
 742         if (tt->rn_flags & RNF_NORMAL) {
 743                 if (m->rm_leaf != tt || m->rm_refs > 0) {
 744                         log(LOG_ERR, "rn_delete: inconsistent annotation\n");
 745                         return (NULL);  /* dangling ref could cause disaster */
 746                 }
 747         } else {
 748                 if (m->rm_mask != tt->rn_mask) {
 749                         log(LOG_ERR, "rn_delete: inconsistent annotation\n");
 750                         goto on1;
 751                 }
 752                 if (--m->rm_refs >= 0)
 753                         goto on1;
 754         }
 755         b = -1 - tt->rn_bit;
 756         t = saved_tt->rn_parent;
 757         if (b > t->rn_bit)
 758                 goto on1; /* Wasn't lifted at all */
 759         do {
 760                 x = t;
 761                 t = t->rn_parent;
 762         } while (b <= t->rn_bit && x != top);
 763         for (mp = &x->rn_mklist; (m = *mp); mp = &m->rm_next)
 764                 if (m == saved_m) {
 765                         *mp = m->rm_next;
 766                         MKFree(&rn_mkfreelist[cpu], m);
 767                         break;
 768                 }
 769         if (m == NULL) {
 770                 log(LOG_ERR, "rn_delete: couldn't find our annotation\n");
 771                 if (tt->rn_flags & RNF_NORMAL)
 772                         return (NULL); /* Dangling ref to us */
 773         }
 774 on1:
 775         /*
 776          * Eliminate us from tree
 777          */
 778         if (tt->rn_flags & RNF_ROOT)
 779                 return (NULL);
 780 #ifdef RN_DEBUG
 781         /* Get us out of the creation list */
 782         for (t = rn_clist; t && t->rn_ybro != tt; t = t->rn_ybro) {}
 783         if (t) t->rn_ybro = tt->rn_ybro;
 784 #endif
 785         t = tt->rn_parent;
 786         dupedkey = saved_tt->rn_dupedkey;
 787         if (dupedkey != NULL) {
 788                 /*
 789                  * at this point, tt is the deletion target and saved_tt
 790                  * is the head of the dupekey chain
 791                  */
 792                 if (tt == saved_tt) {
 793                         /* remove from head of chain */
 794                         x = dupedkey; x->rn_parent = t;
 795                         if (t->rn_left == tt)
 796                                 t->rn_left = x;
 797                         else
 798                                 t->rn_right = x;
 799                 } else {
 800                         /* find node in front of tt on the chain */
 801                         for (x = p = saved_tt; p && p->rn_dupedkey != tt;)
 802                                 p = p->rn_dupedkey;
 803                         if (p) {
 804                                 p->rn_dupedkey = tt->rn_dupedkey;
 805                                 if (tt->rn_dupedkey)            /* parent */
 806                                         tt->rn_dupedkey->rn_parent = p;
 807                                                                 /* parent */
 808                         } else log(LOG_ERR, "rn_delete: couldn't find us\n");
 809                 }
 810                 t = tt + 1;
 811                 if  (t->rn_flags & RNF_ACTIVE) {
 812 #ifndef RN_DEBUG
 813                         *++x = *t;
 814                         p = t->rn_parent;
 815 #else
 816                         b = t->rn_info;
 817                         *++x = *t;
 818                         t->rn_info = b;
 819                         p = t->rn_parent;
 820 #endif
 821                         if (p->rn_left == t)
 822                                 p->rn_left = x;
 823                         else
 824                                 p->rn_right = x;
 825                         x->rn_left->rn_parent = x;
 826                         x->rn_right->rn_parent = x;
 827                 }
 828                 goto out;
 829         }
 830         if (t->rn_left == tt)
 831                 x = t->rn_right;
 832         else
 833                 x = t->rn_left;
 834         p = t->rn_parent;
 835         if (p->rn_right == t)
 836                 p->rn_right = x;
 837         else
 838                 p->rn_left = x;
 839         x->rn_parent = p;
 840         /*
 841          * Demote routes attached to us.
 842          */
 843         if (t->rn_mklist != NULL) {
 844                 if (x->rn_bit >= 0) {
 845                         for (mp = &x->rn_mklist; (m = *mp);)
 846                                 mp = &m->rm_next;
 847                         *mp = t->rn_mklist;
 848                 } else {
 849                         /*
 850                          * If there are any (key, mask) pairs in a sibling
 851                          * duped-key chain, some subset will appear sorted
 852                          * in the same order attached to our mklist.
 853                          */
 854                         for (m = t->rn_mklist; m && x; x = x->rn_dupedkey)
 855                                 if (m == x->rn_mklist) {
 856                                         struct radix_mask *mm = m->rm_next;
 857
 858                                         x->rn_mklist = NULL;
 859                                         if (--(m->rm_refs) < 0)
 860                                                 MKFree(&rn_mkfreelist[cpu], m);
 861                                         m = mm;
 862                                 }
 863                         if (m)
 864                                 log(LOG_ERR,
 865                                     "rn_delete: Orphaned Mask %p at %p\n",
 866                                     (void *)m, (void *)x);
 867                 }
 868         }
 869         /*
 870          * We may be holding an active internal node in the tree.
 871          */
 872         x = tt + 1;
 873         if (t != x) {
 874 #ifndef RN_DEBUG
 875                 *t = *x;
 876 #else
 877                 b = t->rn_info;
 878                 *t = *x;
 879                 t->rn_info = b;
 880 #endif
 881                 t->rn_left->rn_parent = t;
 882                 t->rn_right->rn_parent = t;
 883                 p = x->rn_parent;
 884                 if (p->rn_left == x)
 885                         p->rn_left = t;
 886                 else
 887                         p->rn_right = t;
 888         }
 889 out:
 890         tt->rn_flags &= ~RNF_ACTIVE;
 891         tt[1].rn_flags &= ~RNF_ACTIVE;
 892         return (tt);
 893 }
 894
 895 /*
 896  * This is the same as rn_walktree() except for the parameters and the
 897  * exit.
 898  */
 899 static int
 900 rn_walktree_from(struct radix_node_head *h, char *xa, char *xm,
 901                  walktree_f_t *f, void *w)
 902 {
 903         struct radix_node *base, *next;
 904         struct radix_node *rn, *last = NULL /* shut up gcc */;
 905         boolean_t stopping = FALSE;
 906         int lastb, error;
 907
 908         /*
 909          * rn_search_m is sort-of-open-coded here.
 910          */
 911         /* kprintf("about to search\n"); */
 912         for (rn = h->rnh_treetop; rn->rn_bit >= 0; ) {
 913                 last = rn;
 914                 /* kprintf("rn_bit %d, rn_bmask %x, xm[rn_offset] %x\n",
 915                        rn->rn_bit, rn->rn_bmask, xm[rn->rn_offset]); */
 916                 if (!(rn->rn_bmask & xm[rn->rn_offset])) {
 917                         break;
 918                 }
 919                 if (rn->rn_bmask & xa[rn->rn_offset]) {
 920                         rn = rn->rn_right;
 921                 } else {
 922                         rn = rn->rn_left;
 923                 }
 924         }
 925         /* kprintf("done searching\n"); */
 926
 927         /*
 928          * Two cases: either we stepped off the end of our mask,
 929          * in which case last == rn, or we reached a leaf, in which
 930          * case we want to start from the last node we looked at.
 931          * Either way, last is the node we want to start from.
 932          */
 933         rn = last;
 934         lastb = rn->rn_bit;
 935
 936         /* kprintf("rn %p, lastb %d\n", rn, lastb);*/
 937
 938         /*
 939          * This gets complicated because we may delete the node
 940          * while applying the function f to it, so we need to calculate
 941          * the successor node in advance.
 942          */
 943         while (rn->rn_bit >= 0)
 944                 rn = rn->rn_left;
 945
 946         while (!stopping) {
 947                 /* kprintf("node %p (%d)\n", rn, rn->rn_bit); */
 948                 base = rn;
 949                 /* If at right child go back up, otherwise, go right */
 950                 while (rn->rn_parent->rn_right == rn &&
 951                     !(rn->rn_flags & RNF_ROOT)) {
 952                         rn = rn->rn_parent;
 953
 954                         /* if went up beyond last, stop */
 955                         if (rn->rn_bit < lastb) {
 956                                 stopping = TRUE;
 957                                 /* kprintf("up too far\n"); */
 958                         }
 959                 }
 960
 961                 /* Find the next *leaf* since next node might vanish, too */
 962                 for (rn = rn->rn_parent->rn_right; rn->rn_bit >= 0;)
 963                         rn = rn->rn_left;
 964                 next = rn;
 965                 /* Process leaves */
 966                 while ((rn = base) != NULL) {
 967                         base = rn->rn_dupedkey;
 968                         /* kprintf("leaf %p\n", rn); */
 969                         if (!(rn->rn_flags & RNF_ROOT) && (error = (*f)(rn, w)))
 970                                 return (error);
 971                 }
 972                 rn = next;
 973
 974                 if (rn->rn_flags & RNF_ROOT) {
 975                         /* kprintf("root, stopping"); */
 976                         stopping = TRUE;
 977                 }
 978
 979         }
 980         return 0;
 981 }
 982
 983 static int
 984 rn_walktree_at(struct radix_node_head *h, const char *a, const char *m,
 985     walktree_f_t *f, void *w)
 986 {
 987         struct radix_node *base, *next;
 988         struct radix_node *rn = h->rnh_treetop;
 989         int error;
 990
 991         /*
 992          * This gets complicated because we may delete the node
 993          * while applying the function f to it, so we need to calculate
 994          * the successor node in advance.
 995          */
 996         if (a == NULL) {
 997                 /* First time through node, go left */
 998                 while (rn->rn_bit >= 0)
 999                         rn = rn->rn_left;
1000         } else {
1001                 if (m != NULL)
1002                         rn = rn_search_m(a, rn, m);
1003                 else
1004                         rn = rn_search(a, rn);
1005         }
1006         for (;;) {
1007                 base = rn;
1008                 /* If at right child go back up, otherwise, go right */
1009                 while (rn->rn_parent->rn_right == rn &&
1010                     !(rn->rn_flags & RNF_ROOT))
1011                         rn = rn->rn_parent;
1012                 /* Find the next *leaf* since next node might vanish, too */
1013                 for (rn = rn->rn_parent->rn_right; rn->rn_bit >= 0;)
1014                         rn = rn->rn_left;
1015                 next = rn;
1016                 /* Process leaves */
1017                 while ((rn = base)) {
1018                         base = rn->rn_dupedkey;
1019                         if (!(rn->rn_flags & RNF_ROOT) && (error = (*f)(rn, w)))
1020                                 return (error);
1021                 }
1022                 rn = next;
1023                 if (rn->rn_flags & RNF_ROOT)
1024                         return (0);
1025         }
1026         /* NOTREACHED */
1027 }
1028
1029 static int
1030 rn_walktree(struct radix_node_head *h, walktree_f_t *f, void *w)
1031 {
1032         return rn_walktree_at(h, NULL, NULL, f, w);
1033 }
1034
1035 int
1036 rn_inithead(void **head, struct radix_node_head *maskhead, int off)
1037 {
1038         struct radix_node_head *rnh;
1039         struct radix_node *root, *left, *right;
1040
1041         if (*head != NULL)      /* already initialized */
1042                 return (1);
1043
1044         R_Malloc(rnh, struct radix_node_head *, sizeof *rnh);
1045         if (rnh == NULL)
1046                 return (0);
1047         bzero(rnh, sizeof *rnh);
1048         *head = rnh;
1049
1050         root = rn_newpair(rn_zeros, off, rnh->rnh_nodes);
1051         right = &rnh->rnh_nodes[2];
1052         root->rn_parent = root;
1053         root->rn_flags = RNF_ROOT | RNF_ACTIVE;
1054         root->rn_right = right;
1055
1056         left = root->rn_left;
1057         left->rn_bit = -1 - off;
1058         left->rn_flags = RNF_ROOT | RNF_ACTIVE;
1059
1060         *right = *left;
1061         right->rn_key = rn_ones;
1062
1063         rnh->rnh_treetop = root;
1064         rnh->rnh_maskhead = maskhead;
1065
1066         rnh->rnh_addaddr = rn_addroute;
1067         rnh->rnh_deladdr = rn_delete;
1068         rnh->rnh_matchaddr = rn_match;
1069         rnh->rnh_lookup = rn_lookup;
1070         rnh->rnh_walktree = rn_walktree;
1071         rnh->rnh_walktree_from = rn_walktree_from;
1072         rnh->rnh_walktree_at = rn_walktree_at;
1073
1074         return (1);
1075 }
1076
1077 void
1078 rn_init(void)
1079 {
1080         int cpu;
1081 #ifdef _KERNEL
1082         struct domain *dom;
1083
1084         SLIST_FOREACH(dom, &domains, dom_next) {
1085                 if (dom->dom_maxrtkey > RN_MAXKEYLEN) {
1086                         panic("domain %s maxkey too big %d/%d",
1087                               dom->dom_name, dom->dom_maxrtkey, RN_MAXKEYLEN);
1088                 }
1089         }
1090 #endif
1091         for (cpu = 0; cpu < ncpus; ++cpu) {
1092                 if (rn_inithead((void **)&mask_rnheads[cpu], NULL, 0) == 0)
1093                         panic("rn_init 2");
1094         }
1095 }
1096
1097 struct radix_node_head *
1098 rn_cpumaskhead(int cpu)
1099 {
1100         KKASSERT(mask_rnheads[cpu] != NULL);
1101         return mask_rnheads[cpu];
1102 }