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. All advertising materials mentioning features or use of this software
  14  *    must display the following acknowledgement:
  15  *      This product includes software developed by the University of
  16  *      California, Berkeley and its contributors.
  17  * 4. Neither the name of the University nor the names of its contributors
  18  *    may be used to endorse or promote products derived from this software
  19  *    without specific prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  31  * SUCH DAMAGE.
  32  *
  33  *      @(#)radix.c     8.4 (Berkeley) 11/2/94
  34  * $FreeBSD: src/sys/net/radix.c,v 1.20.2.3 2002/04/28 05:40:25 suz Exp $
  35  * $DragonFly: src/sys/net/radix.c,v 1.14 2006/12/22 23:44:54 swildner Exp $
  36  */
  37
  38 /*
  39  * Routines to build and maintain radix trees for routing lookups.
  40  */
  41 #include <sys/param.h>
  42 #ifdef  _KERNEL
  43 #include <sys/systm.h>
  44 #include <sys/malloc.h>
  45 #include <sys/domain.h>
  46 #include <sys/globaldata.h>
  47 #include <sys/thread.h>
  48 #else
  49 #include <stdlib.h>
  50 #endif
  51 #include <sys/syslog.h>
  52 #include <net/radix.h>
  53
  54 /*
  55  * The arguments to the radix functions are really counted byte arrays with
  56  * the length in the first byte.  struct sockaddr's fit this type structurally.
  57  */
  58 #define clen(c) (*(u_char *)(c))
  59
  60 static int rn_walktree_from(struct radix_node_head *h, char *a, char *m,
  61                             walktree_f_t *f, void *w);
  62 static int rn_walktree(struct radix_node_head *, walktree_f_t *, void *);
  63
  64 static struct radix_node
  65     *rn_insert(char *, struct radix_node_head *, boolean_t *,
  66                struct radix_node [2]),
  67     *rn_newpair(char *, int, struct radix_node[2]),
  68     *rn_search(const char *, struct radix_node *),
  69     *rn_search_m(const char *, struct radix_node *, const char *);
  70
  71 static struct radix_mask *rn_mkfreelist;
  72 static struct radix_node_head *mask_rnheads[MAXCPU];
  73
  74 static int max_keylen;
  75 static char *rn_zeros, *rn_ones;
  76
  77 static int rn_lexobetter(char *m, char *n);
  78 static struct radix_mask *
  79     rn_new_radix_mask(struct radix_node *tt, struct radix_mask *nextmask);
  80 static boolean_t
  81     rn_satisfies_leaf(char *trial, struct radix_node *leaf, int skip);
  82
  83 static __inline struct radix_mask *
  84 MKGet(struct radix_mask **l)
  85 {
  86         struct radix_mask *m;
  87
  88         if (*l != NULL) {
  89                 m = *l;
  90                 *l = m->rm_next;
  91         } else {
  92                 R_Malloc(m, struct radix_mask *, sizeof *m);
  93         }
  94         return m;
  95 }
  96
  97 static __inline void
  98 MKFree(struct radix_mask **l, struct radix_mask *m)
  99 {
 100         m->rm_next = *l;
 101         *l = m;
 102 }
 103
 104 /*
 105  * The data structure for the keys is a radix tree with one way
 106  * branching removed.  The index rn_bit at an internal node n represents a bit
 107  * position to be tested.  The tree is arranged so that all descendants
 108  * of a node n have keys whose bits all agree up to position rn_bit - 1.
 109  * (We say the index of n is rn_bit.)
 110  *
 111  * There is at least one descendant which has a one bit at position rn_bit,
 112  * and at least one with a zero there.
 113  *
 114  * A route is determined by a pair of key and mask.  We require that the
 115  * bit-wise logical and of the key and mask to be the key.
 116  * We define the index of a route to associated with the mask to be
 117  * the first bit number in the mask where 0 occurs (with bit number 0
 118  * representing the highest order bit).
 119  *
 120  * We say a mask is normal if every bit is 0, past the index of the mask.
 121  * If a node n has a descendant (k, m) with index(m) == index(n) == rn_bit,
 122  * and m is a normal mask, then the route applies to every descendant of n.
 123  * If the index(m) < rn_bit, this implies the trailing last few bits of k
 124  * before bit b are all 0, (and hence consequently true of every descendant
 125  * of n), so the route applies to all descendants of the node as well.
 126  *
 127  * Similar logic shows that a non-normal mask m such that
 128  * index(m) <= index(n) could potentially apply to many children of n.
 129  * Thus, for each non-host route, we attach its mask to a list at an internal
 130  * node as high in the tree as we can go.
 131  *
 132  * The present version of the code makes use of normal routes in short-
 133  * circuiting an explict mask and compare operation when testing whether
 134  * a key satisfies a normal route, and also in remembering the unique leaf
 135  * that governs a subtree.
 136  */
 137
 138 static struct radix_node *
 139 rn_search(const char *v, struct radix_node *head)
 140 {
 141         struct radix_node *x;
 142
 143         x = head;
 144         while (x->rn_bit >= 0) {
 145                 if (x->rn_bmask & v[x->rn_offset])
 146                         x = x->rn_right;
 147                 else
 148                         x = x->rn_left;
 149         }
 150         return (x);
 151 }
 152
 153 static struct radix_node *
 154 rn_search_m(const char *v, struct radix_node *head, const char *m)
 155 {
 156         struct radix_node *x;
 157
 158         for (x = head; x->rn_bit >= 0;) {
 159                 if ((x->rn_bmask & m[x->rn_offset]) &&
 160                     (x->rn_bmask & v[x->rn_offset]))
 161                         x = x->rn_right;
 162                 else
 163                         x = x->rn_left;
 164         }
 165         return x;
 166 }
 167
 168 boolean_t
 169 rn_refines(char *m, char *n)
 170 {
 171         char *lim, *lim2;
 172         int longer = clen(n++) - clen(m++);
 173         boolean_t masks_are_equal = TRUE;
 174
 175         lim2 = lim = n + clen(n);
 176         if (longer > 0)
 177                 lim -= longer;
 178         while (n < lim) {
 179                 if (*n & ~(*m))
 180                         return FALSE;
 181                 if (*n++ != *m++)
 182                         masks_are_equal = FALSE;
 183         }
 184         while (n < lim2)
 185                 if (*n++)
 186                         return FALSE;
 187         if (masks_are_equal && (longer < 0))
 188                 for (lim2 = m - longer; m < lim2; )
 189                         if (*m++)
 190                                 return TRUE;
 191         return (!masks_are_equal);
 192 }
 193
 194 struct radix_node *
 195 rn_lookup(char *key, char *mask, struct radix_node_head *head)
 196 {
 197         struct radix_node *x;
 198         char *netmask = NULL;
 199
 200         if (mask != NULL) {
 201                 x = rn_addmask(mask, TRUE, head->rnh_treetop->rn_offset);
 202                 if (x == NULL)
 203                         return (NULL);
 204                 netmask = x->rn_key;
 205         }
 206         x = rn_match(key, head);
 207         if (x != NULL && netmask != NULL) {
 208                 while (x != NULL && x->rn_mask != netmask)
 209                         x = x->rn_dupedkey;
 210         }
 211         return x;
 212 }
 213
 214 static boolean_t
 215 rn_satisfies_leaf(char *trial, struct radix_node *leaf, int skip)
 216 {
 217         char *cp = trial, *cp2 = leaf->rn_key, *cp3 = leaf->rn_mask;
 218         char *cplim;
 219         int length = min(clen(cp), clen(cp2));
 220
 221         if (cp3 == NULL)
 222                 cp3 = rn_ones;
 223         else
 224                 length = min(length, clen(cp3));
 225         cplim = cp + length;
 226         cp3 += skip;
 227         cp2 += skip;
 228         for (cp += skip; cp < cplim; cp++, cp2++, cp3++)
 229                 if ((*cp ^ *cp2) & *cp3)
 230                         return FALSE;
 231         return TRUE;
 232 }
 233
 234 struct radix_node *
 235 rn_match(char *key, struct radix_node_head *head)
 236 {
 237         struct radix_node *t, *x;
 238         char *cp = key, *cp2;
 239         char *cplim;
 240         struct radix_node *saved_t, *top = head->rnh_treetop;
 241         int off = top->rn_offset, klen, matched_off;
 242         int test, b, rn_bit;
 243
 244         t = rn_search(key, top);
 245         /*
 246          * See if we match exactly as a host destination
 247          * or at least learn how many bits match, for normal mask finesse.
 248          *
 249          * It doesn't hurt us to limit how many bytes to check
 250          * to the length of the mask, since if it matches we had a genuine
 251          * match and the leaf we have is the most specific one anyway;
 252          * if it didn't match with a shorter length it would fail
 253          * with a long one.  This wins big for class B&C netmasks which
 254          * are probably the most common case...
 255          */
 256         if (t->rn_mask != NULL)
 257                 klen = clen(t->rn_mask);
 258         else
 259                 klen = clen(key);
 260         cp += off; cp2 = t->rn_key + off; cplim = key + klen;
 261         for (; cp < cplim; cp++, cp2++)
 262                 if (*cp != *cp2)
 263                         goto on1;
 264         /*
 265          * This extra grot is in case we are explicitly asked
 266          * to look up the default.  Ugh!
 267          *
 268          * Never return the root node itself, it seems to cause a
 269          * lot of confusion.
 270          */
 271         if (t->rn_flags & RNF_ROOT)
 272                 t = t->rn_dupedkey;
 273         return t;
 274 on1:
 275         test = (*cp ^ *cp2) & 0xff; /* find first bit that differs */
 276         for (b = 7; (test >>= 1) > 0;)
 277                 b--;
 278         matched_off = cp - key;
 279         b += matched_off << 3;
 280         rn_bit = -1 - b;
 281         /*
 282          * If there is a host route in a duped-key chain, it will be first.
 283          */
 284         if ((saved_t = t)->rn_mask == NULL)
 285                 t = t->rn_dupedkey;
 286         for (; t; t = t->rn_dupedkey) {
 287                 /*
 288                  * Even if we don't match exactly as a host,
 289                  * we may match if the leaf we wound up at is
 290                  * a route to a net.
 291                  */
 292                 if (t->rn_flags & RNF_NORMAL) {
 293                         if (rn_bit <= t->rn_bit)
 294                                 return t;
 295                 } else if (rn_satisfies_leaf(key, t, matched_off))
 296                                 return t;
 297         }
 298         t = saved_t;
 299         /* start searching up the tree */
 300         do {
 301                 struct radix_mask *m;
 302
 303                 t = t->rn_parent;
 304                 /*
 305                  * If non-contiguous masks ever become important
 306                  * we can restore the masking and open coding of
 307                  * the search and satisfaction test and put the
 308                  * calculation of "off" back before the "do".
 309                  */
 310                 m = t->rn_mklist;
 311                 while (m != NULL) {
 312                         if (m->rm_flags & RNF_NORMAL) {
 313                                 if (rn_bit <= m->rm_bit)
 314                                         return (m->rm_leaf);
 315                         } else {
 316                                 off = min(t->rn_offset, matched_off);
 317                                 x = rn_search_m(key, t, m->rm_mask);
 318                                 while (x != NULL && x->rn_mask != m->rm_mask)
 319                                         x = x->rn_dupedkey;
 320                                 if (x && rn_satisfies_leaf(key, x, off))
 321                                         return x;
 322                         }
 323                         m = m->rm_next;
 324                 }
 325         } while (t != top);
 326         return NULL;
 327 }
 328
 329 #ifdef RN_DEBUG
 330 int rn_nodenum;
 331 struct radix_node *rn_clist;
 332 int rn_saveinfo;
 333 boolean_t rn_debug =  TRUE;
 334 #endif
 335
 336 static struct radix_node *
 337 rn_newpair(char *key, int indexbit, struct radix_node nodes[2])
 338 {
 339         struct radix_node *leaf = &nodes[0], *interior = &nodes[1];
 340
 341         interior->rn_bit = indexbit;
 342         interior->rn_bmask = 0x80 >> (indexbit & 0x7);
 343         interior->rn_offset = indexbit >> 3;
 344         interior->rn_left = leaf;
 345         interior->rn_mklist = NULL;
 346
 347         leaf->rn_bit = -1;
 348         leaf->rn_key = key;
 349         leaf->rn_parent = interior;
 350         leaf->rn_flags = interior->rn_flags = RNF_ACTIVE;
 351         leaf->rn_mklist = NULL;
 352
 353 #ifdef RN_DEBUG
 354         leaf->rn_info = rn_nodenum++;
 355         interior->rn_info = rn_nodenum++;
 356         leaf->rn_twin = interior;
 357         leaf->rn_ybro = rn_clist;
 358         rn_clist = leaf;
 359 #endif
 360         return interior;
 361 }
 362
 363 static struct radix_node *
 364 rn_insert(char *key, struct radix_node_head *head, boolean_t *dupentry,
 365           struct radix_node nodes[2])
 366 {
 367         struct radix_node *top = head->rnh_treetop;
 368         int head_off = top->rn_offset, klen = clen(key);
 369         struct radix_node *t = rn_search(key, top);
 370         char *cp = key + head_off;
 371         int b;
 372         struct radix_node *tt;
 373
 374         /*
 375          * Find first bit at which the key and t->rn_key differ
 376          */
 377     {
 378         char *cp2 = t->rn_key + head_off;
 379         int cmp_res;
 380         char *cplim = key + klen;
 381
 382         while (cp < cplim)
 383                 if (*cp2++ != *cp++)
 384                         goto on1;
 385         *dupentry = TRUE;
 386         return t;
 387 on1:
 388         *dupentry = FALSE;
 389         cmp_res = (cp[-1] ^ cp2[-1]) & 0xff;
 390         for (b = (cp - key) << 3; cmp_res; b--)
 391                 cmp_res >>= 1;
 392     }
 393     {
 394         struct radix_node *p, *x = top;
 395
 396         cp = key;
 397         do {
 398                 p = x;
 399                 if (cp[x->rn_offset] & x->rn_bmask)
 400                         x = x->rn_right;
 401                 else
 402                         x = x->rn_left;
 403         } while (b > (unsigned) x->rn_bit);
 404                                 /* x->rn_bit < b && x->rn_bit >= 0 */
 405 #ifdef RN_DEBUG
 406         if (rn_debug)
 407                 log(LOG_DEBUG, "rn_insert: Going In:\n"), traverse(p);
 408 #endif
 409         t = rn_newpair(key, b, nodes);
 410         tt = t->rn_left;
 411         if ((cp[p->rn_offset] & p->rn_bmask) == 0)
 412                 p->rn_left = t;
 413         else
 414                 p->rn_right = t;
 415         x->rn_parent = t;
 416         t->rn_parent = p; /* frees x, p as temp vars below */
 417         if ((cp[t->rn_offset] & t->rn_bmask) == 0) {
 418                 t->rn_right = x;
 419         } else {
 420                 t->rn_right = tt;
 421                 t->rn_left = x;
 422         }
 423 #ifdef RN_DEBUG
 424         if (rn_debug)
 425                 log(LOG_DEBUG, "rn_insert: Coming Out:\n"), traverse(p);
 426 #endif
 427     }
 428         return (tt);
 429 }
 430
 431 struct radix_node *
 432 rn_addmask(char *netmask, boolean_t search, int skip)
 433 {
 434         struct radix_node *x, *saved_x;
 435         char *cp, *cplim;
 436         int b = 0, mlen, m0, j;
 437         boolean_t maskduplicated, isnormal;
 438         static int last_zeroed = 0;
 439         char *addmask_key;
 440         struct radix_node_head *mask_rnh = mask_rnheads[mycpuid];
 441
 442         if ((mlen = clen(netmask)) > max_keylen)
 443                 mlen = max_keylen;
 444         if (skip == 0)
 445                 skip = 1;
 446         if (mlen <= skip)
 447                 return (mask_rnh->rnh_nodes);
 448         R_Malloc(addmask_key, char *, max_keylen);
 449         if (addmask_key == NULL)
 450                 return NULL;
 451         if (skip > 1)
 452                 bcopy(rn_ones + 1, addmask_key + 1, skip - 1);
 453         if ((m0 = mlen) > skip)
 454                 bcopy(netmask + skip, addmask_key + skip, mlen - skip);
 455         /*
 456          * Trim trailing zeroes.
 457          */
 458         for (cp = addmask_key + mlen; (cp > addmask_key) && cp[-1] == 0;)
 459                 cp--;
 460         mlen = cp - addmask_key;
 461         if (mlen <= skip) {
 462                 if (m0 >= last_zeroed)
 463                         last_zeroed = mlen;
 464                 Free(addmask_key);
 465                 return (mask_rnh->rnh_nodes);
 466         }
 467         if (m0 < last_zeroed)
 468                 bzero(addmask_key + m0, last_zeroed - m0);
 469         *addmask_key = last_zeroed = mlen;
 470         x = rn_search(addmask_key, mask_rnh->rnh_treetop);
 471         if (bcmp(addmask_key, x->rn_key, mlen) != 0)
 472                 x = NULL;
 473         if (x != NULL || search)
 474                 goto out;
 475         R_Malloc(x, struct radix_node *, max_keylen + 2 * (sizeof *x));
 476         if ((saved_x = x) == NULL)
 477                 goto out;
 478         bzero(x, max_keylen + 2 * (sizeof *x));
 479         netmask = cp = (char *)(x + 2);
 480         bcopy(addmask_key, cp, mlen);
 481         x = rn_insert(cp, mask_rnh, &maskduplicated, x);
 482         if (maskduplicated) {
 483                 log(LOG_ERR, "rn_addmask: mask impossibly already in tree");
 484                 Free(saved_x);
 485                 goto out;
 486         }
 487         /*
 488          * Calculate index of mask, and check for normalcy.
 489          */
 490         isnormal = TRUE;
 491         cplim = netmask + mlen;
 492         for (cp = netmask + skip; cp < cplim && clen(cp) == 0xff;)
 493                 cp++;
 494         if (cp != cplim) {
 495                 static const char normal_chars[] = {
 496                         0, 0x80, 0xc0, 0xe0, 0xf0, 0xf8, 0xfc, 0xfe, -1
 497                 };
 498
 499                 for (j = 0x80; (j & *cp) != 0; j >>= 1)
 500                         b++;
 501                 if (*cp != normal_chars[b] || cp != (cplim - 1))
 502                         isnormal = FALSE;
 503         }
 504         b += (cp - netmask) << 3;
 505         x->rn_bit = -1 - b;
 506         if (isnormal)
 507                 x->rn_flags |= RNF_NORMAL;
 508 out:
 509         Free(addmask_key);
 510         return (x);
 511 }
 512
 513 /* XXX: arbitrary ordering for non-contiguous masks */
 514 static boolean_t
 515 rn_lexobetter(char *mp, char *np)
 516 {
 517         char *lim;
 518
 519         if ((unsigned) *mp > (unsigned) *np)
 520                 return TRUE;/* not really, but need to check longer one first */
 521         if (*mp == *np)
 522                 for (lim = mp + clen(mp); mp < lim;)
 523                         if (*mp++ > *np++)
 524                                 return TRUE;
 525         return FALSE;
 526 }
 527
 528 static struct radix_mask *
 529 rn_new_radix_mask(struct radix_node *tt, struct radix_mask *nextmask)
 530 {
 531         struct radix_mask *m;
 532
 533         m = MKGet(&rn_mkfreelist);
 534         if (m == NULL) {
 535                 log(LOG_ERR, "Mask for route not entered\n");
 536                 return (NULL);
 537         }
 538         bzero(m, sizeof *m);
 539         m->rm_bit = tt->rn_bit;
 540         m->rm_flags = tt->rn_flags;
 541         if (tt->rn_flags & RNF_NORMAL)
 542                 m->rm_leaf = tt;
 543         else
 544                 m->rm_mask = tt->rn_mask;
 545         m->rm_next = nextmask;
 546         tt->rn_mklist = m;
 547         return m;
 548 }
 549
 550 struct radix_node *
 551 rn_addroute(char *key, char *netmask, struct radix_node_head *head,
 552             struct radix_node treenodes[2])
 553 {
 554         struct radix_node *t, *x = NULL, *tt;
 555         struct radix_node *saved_tt, *top = head->rnh_treetop;
 556         short b = 0, b_leaf = 0;
 557         boolean_t keyduplicated;
 558         char *mmask;
 559         struct radix_mask *m, **mp;
 560
 561         /*
 562          * In dealing with non-contiguous masks, there may be
 563          * many different routes which have the same mask.
 564          * We will find it useful to have a unique pointer to
 565          * the mask to speed avoiding duplicate references at
 566          * nodes and possibly save time in calculating indices.
 567          */
 568         if (netmask != NULL)  {
 569                 if ((x = rn_addmask(netmask, FALSE, top->rn_offset)) == NULL)
 570                         return (NULL);
 571                 b_leaf = x->rn_bit;
 572                 b = -1 - x->rn_bit;
 573                 netmask = x->rn_key;
 574         }
 575         /*
 576          * Deal with duplicated keys: attach node to previous instance
 577          */
 578         saved_tt = tt = rn_insert(key, head, &keyduplicated, treenodes);
 579         if (keyduplicated) {
 580                 for (t = tt; tt; t = tt, tt = tt->rn_dupedkey) {
 581                         if (tt->rn_mask == netmask)
 582                                 return (NULL);
 583                         if (netmask == NULL ||
 584                             (tt->rn_mask &&
 585                              ((b_leaf < tt->rn_bit) /* index(netmask) > node */
 586                               || rn_refines(netmask, tt->rn_mask)
 587                               || rn_lexobetter(netmask, tt->rn_mask))))
 588                                 break;
 589                 }
 590                 /*
 591                  * If the mask is not duplicated, we wouldn't
 592                  * find it among possible duplicate key entries
 593                  * anyway, so the above test doesn't hurt.
 594                  *
 595                  * We sort the masks for a duplicated key the same way as
 596                  * in a masklist -- most specific to least specific.
 597                  * This may require the unfortunate nuisance of relocating
 598                  * the head of the list.
 599                  */
 600                 if (tt == saved_tt) {
 601                         struct  radix_node *xx = x;
 602                         /* link in at head of list */
 603                         (tt = treenodes)->rn_dupedkey = t;
 604                         tt->rn_flags = t->rn_flags;
 605                         tt->rn_parent = x = t->rn_parent;
 606                         t->rn_parent = tt;                      /* parent */
 607                         if (x->rn_left == t)
 608                                 x->rn_left = tt;
 609                         else
 610                                 x->rn_right = tt;
 611                         saved_tt = tt; x = xx;
 612                 } else {
 613                         (tt = treenodes)->rn_dupedkey = t->rn_dupedkey;
 614                         t->rn_dupedkey = tt;
 615                         tt->rn_parent = t;                      /* parent */
 616                         if (tt->rn_dupedkey != NULL)            /* parent */
 617                                 tt->rn_dupedkey->rn_parent = tt; /* parent */
 618                 }
 619 #ifdef RN_DEBUG
 620                 t=tt+1; tt->rn_info = rn_nodenum++; t->rn_info = rn_nodenum++;
 621                 tt->rn_twin = t; tt->rn_ybro = rn_clist; rn_clist = tt;
 622 #endif
 623                 tt->rn_key = key;
 624                 tt->rn_bit = -1;
 625                 tt->rn_flags = RNF_ACTIVE;
 626         }
 627         /*
 628          * Put mask in tree.
 629          */
 630         if (netmask != NULL) {
 631                 tt->rn_mask = netmask;
 632                 tt->rn_bit = x->rn_bit;
 633                 tt->rn_flags |= x->rn_flags & RNF_NORMAL;
 634         }
 635         t = saved_tt->rn_parent;
 636         if (keyduplicated)
 637                 goto on2;
 638         b_leaf = -1 - t->rn_bit;
 639         if (t->rn_right == saved_tt)
 640                 x = t->rn_left;
 641         else
 642                 x = t->rn_right;
 643         /* Promote general routes from below */
 644         if (x->rn_bit < 0) {
 645                 mp = &t->rn_mklist;
 646                 while (x != NULL) {
 647                         if (x->rn_mask != NULL &&
 648                             x->rn_bit >= b_leaf &&
 649                             x->rn_mklist == NULL) {
 650                                 *mp = m = rn_new_radix_mask(x, NULL);
 651                                 if (m != NULL)
 652                                         mp = &m->rm_next;
 653                         }
 654                         x = x->rn_dupedkey;
 655                 }
 656         } else if (x->rn_mklist != NULL) {
 657                 /*
 658                  * Skip over masks whose index is > that of new node
 659                  */
 660                 for (mp = &x->rn_mklist; (m = *mp); mp = &m->rm_next)
 661                         if (m->rm_bit >= b_leaf)
 662                                 break;
 663                 t->rn_mklist = m;
 664                 *mp = NULL;
 665         }
 666 on2:
 667         /* Add new route to highest possible ancestor's list */
 668         if ((netmask == NULL) || (b > t->rn_bit ))
 669                 return tt; /* can't lift at all */
 670         b_leaf = tt->rn_bit;
 671         do {
 672                 x = t;
 673                 t = t->rn_parent;
 674         } while (b <= t->rn_bit && x != top);
 675         /*
 676          * Search through routes associated with node to
 677          * insert new route according to index.
 678          * Need same criteria as when sorting dupedkeys to avoid
 679          * double loop on deletion.
 680          */
 681         for (mp = &x->rn_mklist; (m = *mp); mp = &m->rm_next) {
 682                 if (m->rm_bit < b_leaf)
 683                         continue;
 684                 if (m->rm_bit > b_leaf)
 685                         break;
 686                 if (m->rm_flags & RNF_NORMAL) {
 687                         mmask = m->rm_leaf->rn_mask;
 688                         if (tt->rn_flags & RNF_NORMAL) {
 689                             log(LOG_ERR,
 690                                 "Non-unique normal route, mask not entered\n");
 691                                 return tt;
 692                         }
 693                 } else
 694                         mmask = m->rm_mask;
 695                 if (mmask == netmask) {
 696                         m->rm_refs++;
 697                         tt->rn_mklist = m;
 698                         return tt;
 699                 }
 700                 if (rn_refines(netmask, mmask) || rn_lexobetter(netmask, mmask))
 701                         break;
 702         }
 703         *mp = rn_new_radix_mask(tt, *mp);
 704         return tt;
 705 }
 706
 707 struct radix_node *
 708 rn_delete(char *key, char *netmask, struct radix_node_head *head)
 709 {
 710         struct radix_node *t, *p, *x, *tt;
 711         struct radix_mask *m, *saved_m, **mp;
 712         struct radix_node *dupedkey, *saved_tt, *top;
 713         int b, head_off, klen;
 714
 715         x = head->rnh_treetop;
 716         tt = rn_search(key, x);
 717         head_off = x->rn_offset;
 718         klen =  clen(key);
 719         saved_tt = tt;
 720         top = x;
 721         if (tt == NULL ||
 722             bcmp(key + head_off, tt->rn_key + head_off, klen - head_off))
 723                 return (NULL);
 724         /*
 725          * Delete our route from mask lists.
 726          */
 727         if (netmask != NULL) {
 728                 if ((x = rn_addmask(netmask, TRUE, head_off)) == NULL)
 729                         return (NULL);
 730                 netmask = x->rn_key;
 731                 while (tt->rn_mask != netmask)
 732                         if ((tt = tt->rn_dupedkey) == NULL)
 733                                 return (NULL);
 734         }
 735         if (tt->rn_mask == NULL || (saved_m = m = tt->rn_mklist) == NULL)
 736                 goto on1;
 737         if (tt->rn_flags & RNF_NORMAL) {
 738                 if (m->rm_leaf != tt || m->rm_refs > 0) {
 739                         log(LOG_ERR, "rn_delete: inconsistent annotation\n");
 740                         return (NULL);  /* dangling ref could cause disaster */
 741                 }
 742         } else {
 743                 if (m->rm_mask != tt->rn_mask) {
 744                         log(LOG_ERR, "rn_delete: inconsistent annotation\n");
 745                         goto on1;
 746                 }
 747                 if (--m->rm_refs >= 0)
 748                         goto on1;
 749         }
 750         b = -1 - tt->rn_bit;
 751         t = saved_tt->rn_parent;
 752         if (b > t->rn_bit)
 753                 goto on1; /* Wasn't lifted at all */
 754         do {
 755                 x = t;
 756                 t = t->rn_parent;
 757         } while (b <= t->rn_bit && x != top);
 758         for (mp = &x->rn_mklist; (m = *mp); mp = &m->rm_next)
 759                 if (m == saved_m) {
 760                         *mp = m->rm_next;
 761                         MKFree(&rn_mkfreelist, m);
 762                         break;
 763                 }
 764         if (m == NULL) {
 765                 log(LOG_ERR, "rn_delete: couldn't find our annotation\n");
 766                 if (tt->rn_flags & RNF_NORMAL)
 767                         return (NULL); /* Dangling ref to us */
 768         }
 769 on1:
 770         /*
 771          * Eliminate us from tree
 772          */
 773         if (tt->rn_flags & RNF_ROOT)
 774                 return (NULL);
 775 #ifdef RN_DEBUG
 776         /* Get us out of the creation list */
 777         for (t = rn_clist; t && t->rn_ybro != tt; t = t->rn_ybro) {}
 778         if (t) t->rn_ybro = tt->rn_ybro;
 779 #endif
 780         t = tt->rn_parent;
 781         dupedkey = saved_tt->rn_dupedkey;
 782         if (dupedkey != NULL) {
 783                 /*
 784                  * at this point, tt is the deletion target and saved_tt
 785                  * is the head of the dupekey chain
 786                  */
 787                 if (tt == saved_tt) {
 788                         /* remove from head of chain */
 789                         x = dupedkey; x->rn_parent = t;
 790                         if (t->rn_left == tt)
 791                                 t->rn_left = x;
 792                         else
 793                                 t->rn_right = x;
 794                 } else {
 795                         /* find node in front of tt on the chain */
 796                         for (x = p = saved_tt; p && p->rn_dupedkey != tt;)
 797                                 p = p->rn_dupedkey;
 798                         if (p) {
 799                                 p->rn_dupedkey = tt->rn_dupedkey;
 800                                 if (tt->rn_dupedkey)            /* parent */
 801                                         tt->rn_dupedkey->rn_parent = p;
 802                                                                 /* parent */
 803                         } else log(LOG_ERR, "rn_delete: couldn't find us\n");
 804                 }
 805                 t = tt + 1;
 806                 if  (t->rn_flags & RNF_ACTIVE) {
 807 #ifndef RN_DEBUG
 808                         *++x = *t;
 809                         p = t->rn_parent;
 810 #else
 811                         b = t->rn_info;
 812                         *++x = *t;
 813                         t->rn_info = b;
 814                         p = t->rn_parent;
 815 #endif
 816                         if (p->rn_left == t)
 817                                 p->rn_left = x;
 818                         else
 819                                 p->rn_right = x;
 820                         x->rn_left->rn_parent = x;
 821                         x->rn_right->rn_parent = x;
 822                 }
 823                 goto out;
 824         }
 825         if (t->rn_left == tt)
 826                 x = t->rn_right;
 827         else
 828                 x = t->rn_left;
 829         p = t->rn_parent;
 830         if (p->rn_right == t)
 831                 p->rn_right = x;
 832         else
 833                 p->rn_left = x;
 834         x->rn_parent = p;
 835         /*
 836          * Demote routes attached to us.
 837          */
 838         if (t->rn_mklist != NULL) {
 839                 if (x->rn_bit >= 0) {
 840                         for (mp = &x->rn_mklist; (m = *mp);)
 841                                 mp = &m->rm_next;
 842                         *mp = t->rn_mklist;
 843                 } else {
 844                         /*
 845                          * If there are any (key, mask) pairs in a sibling
 846                          * duped-key chain, some subset will appear sorted
 847                          * in the same order attached to our mklist.
 848                          */
 849                         for (m = t->rn_mklist; m && x; x = x->rn_dupedkey)
 850                                 if (m == x->rn_mklist) {
 851                                         struct radix_mask *mm = m->rm_next;
 852
 853                                         x->rn_mklist = NULL;
 854                                         if (--(m->rm_refs) < 0)
 855                                                 MKFree(&rn_mkfreelist, m);
 856                                         m = mm;
 857                                 }
 858                         if (m)
 859                                 log(LOG_ERR,
 860                                     "rn_delete: Orphaned Mask %p at %p\n",
 861                                     (void *)m, (void *)x);
 862                 }
 863         }
 864         /*
 865          * We may be holding an active internal node in the tree.
 866          */
 867         x = tt + 1;
 868         if (t != x) {
 869 #ifndef RN_DEBUG
 870                 *t = *x;
 871 #else
 872                 b = t->rn_info;
 873                 *t = *x;
 874                 t->rn_info = b;
 875 #endif
 876                 t->rn_left->rn_parent = t;
 877                 t->rn_right->rn_parent = t;
 878                 p = x->rn_parent;
 879                 if (p->rn_left == x)
 880                         p->rn_left = t;
 881                 else
 882                         p->rn_right = t;
 883         }
 884 out:
 885         tt->rn_flags &= ~RNF_ACTIVE;
 886         tt[1].rn_flags &= ~RNF_ACTIVE;
 887         return (tt);
 888 }
 889
 890 /*
 891  * This is the same as rn_walktree() except for the parameters and the
 892  * exit.
 893  */
 894 static int
 895 rn_walktree_from(struct radix_node_head *h, char *xa, char *xm,
 896                  walktree_f_t *f, void *w)
 897 {
 898         struct radix_node *base, *next;
 899         struct radix_node *rn, *last = NULL /* shut up gcc */;
 900         boolean_t stopping = FALSE;
 901         int lastb, error;
 902
 903         /*
 904          * rn_search_m is sort-of-open-coded here.
 905          */
 906         /* kprintf("about to search\n"); */
 907         for (rn = h->rnh_treetop; rn->rn_bit >= 0; ) {
 908                 last = rn;
 909                 /* kprintf("rn_bit %d, rn_bmask %x, xm[rn_offset] %x\n",
 910                        rn->rn_bit, rn->rn_bmask, xm[rn->rn_offset]); */
 911                 if (!(rn->rn_bmask & xm[rn->rn_offset])) {
 912                         break;
 913                 }
 914                 if (rn->rn_bmask & xa[rn->rn_offset]) {
 915                         rn = rn->rn_right;
 916                 } else {
 917                         rn = rn->rn_left;
 918                 }
 919         }
 920         /* kprintf("done searching\n"); */
 921
 922         /*
 923          * Two cases: either we stepped off the end of our mask,
 924          * in which case last == rn, or we reached a leaf, in which
 925          * case we want to start from the last node we looked at.
 926          * Either way, last is the node we want to start from.
 927          */
 928         rn = last;
 929         lastb = rn->rn_bit;
 930
 931         /* kprintf("rn %p, lastb %d\n", rn, lastb);*/
 932
 933         /*
 934          * This gets complicated because we may delete the node
 935          * while applying the function f to it, so we need to calculate
 936          * the successor node in advance.
 937          */
 938         while (rn->rn_bit >= 0)
 939                 rn = rn->rn_left;
 940
 941         while (!stopping) {
 942                 /* kprintf("node %p (%d)\n", rn, rn->rn_bit); */
 943                 base = rn;
 944                 /* If at right child go back up, otherwise, go right */
 945                 while (rn->rn_parent->rn_right == rn &&
 946                     !(rn->rn_flags & RNF_ROOT)) {
 947                         rn = rn->rn_parent;
 948
 949                         /* if went up beyond last, stop */
 950                         if (rn->rn_bit < lastb) {
 951                                 stopping = TRUE;
 952                                 /* kprintf("up too far\n"); */
 953                         }
 954                 }
 955
 956                 /* Find the next *leaf* since next node might vanish, too */
 957                 for (rn = rn->rn_parent->rn_right; rn->rn_bit >= 0;)
 958                         rn = rn->rn_left;
 959                 next = rn;
 960                 /* Process leaves */
 961                 while ((rn = base) != NULL) {
 962                         base = rn->rn_dupedkey;
 963                         /* kprintf("leaf %p\n", rn); */
 964                         if (!(rn->rn_flags & RNF_ROOT) && (error = (*f)(rn, w)))
 965                                 return (error);
 966                 }
 967                 rn = next;
 968
 969                 if (rn->rn_flags & RNF_ROOT) {
 970                         /* kprintf("root, stopping"); */
 971                         stopping = TRUE;
 972                 }
 973
 974         }
 975         return 0;
 976 }
 977
 978 static int
 979 rn_walktree(struct radix_node_head *h, walktree_f_t *f, void *w)
 980 {
 981         struct radix_node *base, *next;
 982         struct radix_node *rn = h->rnh_treetop;
 983         int error;
 984
 985         /*
 986          * This gets complicated because we may delete the node
 987          * while applying the function f to it, so we need to calculate
 988          * the successor node in advance.
 989          */
 990         /* First time through node, go left */
 991         while (rn->rn_bit >= 0)
 992                 rn = rn->rn_left;
 993         for (;;) {
 994                 base = rn;
 995                 /* If at right child go back up, otherwise, go right */
 996                 while (rn->rn_parent->rn_right == rn &&
 997                     !(rn->rn_flags & RNF_ROOT))
 998                         rn = rn->rn_parent;
 999                 /* Find the next *leaf* since next node might vanish, too */
1000                 for (rn = rn->rn_parent->rn_right; rn->rn_bit >= 0;)
1001                         rn = rn->rn_left;
1002                 next = rn;
1003                 /* Process leaves */
1004                 while ((rn = base)) {
1005                         base = rn->rn_dupedkey;
1006                         if (!(rn->rn_flags & RNF_ROOT) && (error = (*f)(rn, w)))
1007                                 return (error);
1008                 }
1009                 rn = next;
1010                 if (rn->rn_flags & RNF_ROOT)
1011                         return (0);
1012         }
1013         /* NOTREACHED */
1014 }
1015
1016 int
1017 rn_inithead(void **head, int off)
1018 {
1019         struct radix_node_head *rnh;
1020         struct radix_node *root, *left, *right;
1021
1022         if (*head != NULL)      /* already initialized */
1023                 return (1);
1024
1025         R_Malloc(rnh, struct radix_node_head *, sizeof *rnh);
1026         if (rnh == NULL)
1027                 return (0);
1028         bzero(rnh, sizeof *rnh);
1029         *head = rnh;
1030
1031         root = rn_newpair(rn_zeros, off, rnh->rnh_nodes);
1032         right = &rnh->rnh_nodes[2];
1033         root->rn_parent = root;
1034         root->rn_flags = RNF_ROOT | RNF_ACTIVE;
1035         root->rn_right = right;
1036
1037         left = root->rn_left;
1038         left->rn_bit = -1 - off;
1039         left->rn_flags = RNF_ROOT | RNF_ACTIVE;
1040
1041         *right = *left;
1042         right->rn_key = rn_ones;
1043
1044         rnh->rnh_treetop = root;
1045
1046         rnh->rnh_addaddr = rn_addroute;
1047         rnh->rnh_deladdr = rn_delete;
1048         rnh->rnh_matchaddr = rn_match;
1049         rnh->rnh_lookup = rn_lookup;
1050         rnh->rnh_walktree = rn_walktree;
1051         rnh->rnh_walktree_from = rn_walktree_from;
1052
1053         return (1);
1054 }
1055
1056 void
1057 rn_init(void)
1058 {
1059         char *cp, *cplim;
1060         int cpu;
1061 #ifdef _KERNEL
1062         struct domain *dom;
1063
1064         SLIST_FOREACH(dom, &domains, dom_next)
1065                 if (dom->dom_maxrtkey > max_keylen)
1066                         max_keylen = dom->dom_maxrtkey;
1067 #endif
1068         if (max_keylen == 0) {
1069                 log(LOG_ERR,
1070                     "rn_init: radix functions require max_keylen be set\n");
1071                 return;
1072         }
1073         R_Malloc(rn_zeros, char *, 2 * max_keylen);
1074         if (rn_zeros == NULL)
1075                 panic("rn_init");
1076         bzero(rn_zeros, 2 * max_keylen);
1077         rn_ones = cp = rn_zeros + max_keylen;
1078         cplim = rn_ones + max_keylen;
1079         while (cp < cplim)
1080                 *cp++ = -1;
1081
1082         for (cpu = 0; cpu < ncpus; ++cpu) {
1083                 if (rn_inithead((void **)&mask_rnheads[cpu], 0) == 0)
1084                         panic("rn_init 2");
1085         }
1086 }