sys/vm/vm_radix.c

   1 /*
   2  * Copyright (c) 2013 EMC Corp.
   3  * Copyright (c) 2011 Jeffrey Roberson <jeff@freebsd.org>
   4  * Copyright (c) 2008 Mayur Shardul <mayur.shardul@gmail.com>
   5  * 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  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  *
  28  */
  29
  30 /*
  31  * Path-compressed radix trie implementation.
  32  * The following code is not generalized into a general purpose library
  33  * because there are way too many parameters embedded that should really
  34  * be decided by the library consumers.  At the same time, consumers
  35  * of this code must achieve highest possible performance.
  36  *
  37  * The implementation takes into account the following rationale:
  38  * - Size of the nodes should be as small as possible but still big enough
  39  *   to avoid a large maximum depth for the trie.  This is a balance
  40  *   between the necessity to not wire too much physical memory for the nodes
  41  *   and the necessity to avoid too much cache pollution during the trie
  42  *   operations.
  43  * - There is not a huge bias toward the number of lookup operations over
  44  *   the number of insert and remove operations.  This basically implies
  45  *   that optimizations supposedly helping one operation but hurting the
  46  *   other might be carefully evaluated.
  47  * - On average not many nodes are expected to be fully populated, hence
  48  *   level compression may just complicate things.
  49  */
  50
  51 #include <sys/cdefs.h>
  52 __FBSDID("$FreeBSD$");
  53
  54 #include "opt_ddb.h"
  55
  56 #include <sys/param.h>
  57 #include <sys/systm.h>
  58 #include <sys/kernel.h>
  59 #include <sys/vmmeter.h>
  60
  61 #include <vm/uma.h>
  62 #include <vm/vm.h>
  63 #include <vm/vm_param.h>
  64 #include <vm/vm_page.h>
  65 #include <vm/vm_radix.h>
  66
  67 #ifdef DDB
  68 #include <ddb/ddb.h>
  69 #endif
  70
  71 /*
  72  * These widths should allow the pointers to a node's children to fit within
  73  * a single cache line.  The extra levels from a narrow width should not be
  74  * a problem thanks to path compression.
  75  */
  76 #ifdef __LP64__
  77 #define VM_RADIX_WIDTH  4
  78 #else
  79 #define VM_RADIX_WIDTH  3
  80 #endif
  81
  82 #define VM_RADIX_COUNT  (1 << VM_RADIX_WIDTH)
  83 #define VM_RADIX_MASK   (VM_RADIX_COUNT - 1)
  84 #define VM_RADIX_LIMIT                                                  \
  85         (howmany((sizeof(vm_pindex_t) * NBBY), VM_RADIX_WIDTH) - 1)
  86
  87 /* Flag bits stored in node pointers. */
  88 #define VM_RADIX_ISLEAF 0x1
  89 #define VM_RADIX_FLAGS  0x1
  90 #define VM_RADIX_PAD    VM_RADIX_FLAGS
  91
  92 /* Returns one unit associated with specified level. */
  93 #define VM_RADIX_UNITLEVEL(lev)                                         \
  94         ((vm_pindex_t)1 << ((VM_RADIX_LIMIT - (lev)) * VM_RADIX_WIDTH))
  95
  96 struct vm_radix_node {
  97         vm_pindex_t      rn_owner;                      /* Owner of record. */
  98         uint16_t         rn_count;                      /* Valid children. */
  99         uint16_t         rn_clev;                       /* Current level. */
 100         void            *rn_child[VM_RADIX_COUNT];      /* Child nodes. */
 101 };
 102
 103 static uma_zone_t vm_radix_node_zone;
 104
 105 /*
 106  * Allocate a radix node.  Pre-allocation should ensure that the request
 107  * will always be satisfied.
 108  */
 109 static __inline struct vm_radix_node *
 110 vm_radix_node_get(vm_pindex_t owner, uint16_t count, uint16_t clevel)
 111 {
 112         struct vm_radix_node *rnode;
 113
 114         rnode = uma_zalloc(vm_radix_node_zone, M_NOWAIT);
 115
 116         /*
 117          * The required number of nodes should already be pre-allocated
 118          * by vm_radix_prealloc().  However, UMA can hold a few nodes
 119          * in per-CPU buckets, which will not be accessible by the
 120          * current CPU.  Thus, the allocation could return NULL when
 121          * the pre-allocated pool is close to exhaustion.  Anyway,
 122          * in practice this should never occur because a new node
 123          * is not always required for insert.  Thus, the pre-allocated
 124          * pool should have some extra pages that prevent this from
 125          * becoming a problem.
 126          */
 127         if (rnode == NULL)
 128                 panic("%s: uma_zalloc() returned NULL for a new node",
 129                     __func__);
 130         rnode->rn_owner = owner;
 131         rnode->rn_count = count;
 132         rnode->rn_clev = clevel;
 133         return (rnode);
 134 }
 135
 136 /*
 137  * Free radix node.
 138  */
 139 static __inline void
 140 vm_radix_node_put(struct vm_radix_node *rnode)
 141 {
 142
 143         uma_zfree(vm_radix_node_zone, rnode);
 144 }
 145
 146 /*
 147  * Return the position in the array for a given level.
 148  */
 149 static __inline int
 150 vm_radix_slot(vm_pindex_t index, uint16_t level)
 151 {
 152
 153         return ((index >> ((VM_RADIX_LIMIT - level) * VM_RADIX_WIDTH)) &
 154             VM_RADIX_MASK);
 155 }
 156
 157 /* Trims the key after the specified level. */
 158 static __inline vm_pindex_t
 159 vm_radix_trimkey(vm_pindex_t index, uint16_t level)
 160 {
 161         vm_pindex_t ret;
 162
 163         ret = index;
 164         if (level < VM_RADIX_LIMIT) {
 165                 ret >>= (VM_RADIX_LIMIT - level) * VM_RADIX_WIDTH;
 166                 ret <<= (VM_RADIX_LIMIT - level) * VM_RADIX_WIDTH;
 167         }
 168         return (ret);
 169 }
 170
 171 /*
 172  * Get the root node for a radix tree.
 173  */
 174 static __inline struct vm_radix_node *
 175 vm_radix_getroot(struct vm_radix *rtree)
 176 {
 177
 178         return ((struct vm_radix_node *)rtree->rt_root);
 179 }
 180
 181 /*
 182  * Set the root node for a radix tree.
 183  */
 184 static __inline void
 185 vm_radix_setroot(struct vm_radix *rtree, struct vm_radix_node *rnode)
 186 {
 187
 188         rtree->rt_root = (uintptr_t)rnode;
 189 }
 190
 191 /*
 192  * Returns TRUE if the specified radix node is a leaf and FALSE otherwise.
 193  */
 194 static __inline boolean_t
 195 vm_radix_isleaf(struct vm_radix_node *rnode)
 196 {
 197
 198         return (((uintptr_t)rnode & VM_RADIX_ISLEAF) != 0);
 199 }
 200
 201 /*
 202  * Returns the associated page extracted from rnode.
 203  */
 204 static __inline vm_page_t
 205 vm_radix_topage(struct vm_radix_node *rnode)
 206 {
 207
 208         return ((vm_page_t)((uintptr_t)rnode & ~VM_RADIX_FLAGS));
 209 }
 210
 211 /*
 212  * Adds the page as a child of the provided node.
 213  */
 214 static __inline void
 215 vm_radix_addpage(struct vm_radix_node *rnode, vm_pindex_t index, uint16_t clev,
 216     vm_page_t page)
 217 {
 218         int slot;
 219
 220         slot = vm_radix_slot(index, clev);
 221         rnode->rn_child[slot] = (void *)((uintptr_t)page | VM_RADIX_ISLEAF);
 222 }
 223
 224 /*
 225  * Returns the slot where two keys differ.
 226  * It cannot accept 2 equal keys.
 227  */
 228 static __inline uint16_t
 229 vm_radix_keydiff(vm_pindex_t index1, vm_pindex_t index2)
 230 {
 231         uint16_t clev;
 232
 233         KASSERT(index1 != index2, ("%s: passing the same key value %jx",
 234             __func__, (uintmax_t)index1));
 235
 236         index1 ^= index2;
 237         for (clev = 0;; clev++)
 238                 if (vm_radix_slot(index1, clev) != 0)
 239                         return (clev);
 240 }
 241
 242 /*
 243  * Returns TRUE if it can be determined that key does not belong to the
 244  * specified rnode.  Otherwise, returns FALSE.
 245  */
 246 static __inline boolean_t
 247 vm_radix_keybarr(struct vm_radix_node *rnode, vm_pindex_t idx)
 248 {
 249
 250         if (rnode->rn_clev > 0) {
 251                 idx = vm_radix_trimkey(idx, rnode->rn_clev - 1);
 252                 return (idx != rnode->rn_owner);
 253         }
 254         return (FALSE);
 255 }
 256
 257 /*
 258  * Internal helper for vm_radix_reclaim_allnodes().
 259  * This function is recursive.
 260  */
 261 static void
 262 vm_radix_reclaim_allnodes_int(struct vm_radix_node *rnode)
 263 {
 264         int slot;
 265
 266         KASSERT(rnode->rn_count <= VM_RADIX_COUNT,
 267             ("vm_radix_reclaim_allnodes_int: bad count in rnode %p", rnode));
 268         for (slot = 0; rnode->rn_count != 0; slot++) {
 269                 if (rnode->rn_child[slot] == NULL)
 270                         continue;
 271                 if (!vm_radix_isleaf(rnode->rn_child[slot]))
 272                         vm_radix_reclaim_allnodes_int(rnode->rn_child[slot]);
 273                 rnode->rn_child[slot] = NULL;
 274                 rnode->rn_count--;
 275         }
 276         vm_radix_node_put(rnode);
 277 }
 278
 279 #ifdef INVARIANTS
 280 /*
 281  * Radix node zone destructor.
 282  */
 283 static void
 284 vm_radix_node_zone_dtor(void *mem, int size __unused, void *arg __unused)
 285 {
 286         struct vm_radix_node *rnode;
 287         int slot;
 288
 289         rnode = mem;
 290         KASSERT(rnode->rn_count == 0,
 291             ("vm_radix_node_put: rnode %p has %d children", rnode,
 292             rnode->rn_count));
 293         for (slot = 0; slot < VM_RADIX_COUNT; slot++)
 294                 KASSERT(rnode->rn_child[slot] == NULL,
 295                     ("vm_radix_node_put: rnode %p has a child", rnode));
 296 }
 297 #endif
 298
 299 /*
 300  * Radix node zone initializer.
 301  */
 302 static int
 303 vm_radix_node_zone_init(void *mem, int size __unused, int flags __unused)
 304 {
 305         struct vm_radix_node *rnode;
 306
 307         rnode = mem;
 308         memset(rnode->rn_child, 0, sizeof(rnode->rn_child));
 309         return (0);
 310 }
 311
 312 /*
 313  * Pre-allocate intermediate nodes from the UMA slab zone.
 314  */
 315 static void
 316 vm_radix_prealloc(void *arg __unused)
 317 {
 318         int nodes;
 319
 320         /*
 321          * Calculate the number of reserved nodes, discounting the pages that
 322          * are needed to store them.
 323          */
 324         nodes = ((vm_paddr_t)cnt.v_page_count * PAGE_SIZE) / (PAGE_SIZE +
 325             sizeof(struct vm_radix_node));
 326         if (!uma_zone_reserve_kva(vm_radix_node_zone, nodes))
 327                 panic("%s: unable to create new zone", __func__);
 328         uma_prealloc(vm_radix_node_zone, nodes);
 329 }
 330 SYSINIT(vm_radix_prealloc, SI_SUB_KMEM, SI_ORDER_SECOND, vm_radix_prealloc,
 331     NULL);
 332
 333 /*
 334  * Initialize the UMA slab zone.
 335  * Until vm_radix_prealloc() is called, the zone will be served by the
 336  * UMA boot-time pre-allocated pool of pages.
 337  */
 338 void
 339 vm_radix_init(void)
 340 {
 341
 342         vm_radix_node_zone = uma_zcreate("RADIX NODE",
 343             sizeof(struct vm_radix_node), NULL,
 344 #ifdef INVARIANTS
 345             vm_radix_node_zone_dtor,
 346 #else
 347             NULL,
 348 #endif
 349             vm_radix_node_zone_init, NULL, VM_RADIX_PAD, UMA_ZONE_VM |
 350             UMA_ZONE_NOFREE);
 351 }
 352
 353 /*
 354  * Inserts the key-value pair into the trie.
 355  * Panics if the key already exists.
 356  */
 357 void
 358 vm_radix_insert(struct vm_radix *rtree, vm_page_t page)
 359 {
 360         vm_pindex_t index, newind;
 361         void **parentp;
 362         struct vm_radix_node *rnode, *tmp;
 363         vm_page_t m;
 364         int slot;
 365         uint16_t clev;
 366
 367         index = page->pindex;
 368
 369         /*
 370          * The owner of record for root is not really important because it
 371          * will never be used.
 372          */
 373         rnode = vm_radix_getroot(rtree);
 374         if (rnode == NULL) {
 375                 rtree->rt_root = (uintptr_t)page | VM_RADIX_ISLEAF;
 376                 return;
 377         }
 378         parentp = (void **)&rtree->rt_root;
 379         for (;;) {
 380                 if (vm_radix_isleaf(rnode)) {
 381                         m = vm_radix_topage(rnode);
 382                         if (m->pindex == index)
 383                                 panic("%s: key %jx is already present",
 384                                     __func__, (uintmax_t)index);
 385                         clev = vm_radix_keydiff(m->pindex, index);
 386                         tmp = vm_radix_node_get(vm_radix_trimkey(index,
 387                             clev - 1), 2, clev);
 388                         *parentp = tmp;
 389                         vm_radix_addpage(tmp, index, clev, page);
 390                         vm_radix_addpage(tmp, m->pindex, clev, m);
 391                         return;
 392                 } else if (vm_radix_keybarr(rnode, index))
 393                         break;
 394                 slot = vm_radix_slot(index, rnode->rn_clev);
 395                 if (rnode->rn_child[slot] == NULL) {
 396                         rnode->rn_count++;
 397                         vm_radix_addpage(rnode, index, rnode->rn_clev, page);
 398                         return;
 399                 }
 400                 parentp = &rnode->rn_child[slot];
 401                 rnode = rnode->rn_child[slot];
 402         }
 403
 404         /*
 405          * A new node is needed because the right insertion level is reached.
 406          * Setup the new intermediate node and add the 2 children: the
 407          * new object and the older edge.
 408          */
 409         newind = rnode->rn_owner;
 410         clev = vm_radix_keydiff(newind, index);
 411         tmp = vm_radix_node_get(vm_radix_trimkey(index, clev - 1), 2,
 412             clev);
 413         *parentp = tmp;
 414         vm_radix_addpage(tmp, index, clev, page);
 415         slot = vm_radix_slot(newind, clev);
 416         tmp->rn_child[slot] = rnode;
 417 }
 418
 419 /*
 420  * Returns the value stored at the index.  If the index is not present,
 421  * NULL is returned.
 422  */
 423 vm_page_t
 424 vm_radix_lookup(struct vm_radix *rtree, vm_pindex_t index)
 425 {
 426         struct vm_radix_node *rnode;
 427         vm_page_t m;
 428         int slot;
 429
 430         rnode = vm_radix_getroot(rtree);
 431         while (rnode != NULL) {
 432                 if (vm_radix_isleaf(rnode)) {
 433                         m = vm_radix_topage(rnode);
 434                         if (m->pindex == index)
 435                                 return (m);
 436                         else
 437                                 break;
 438                 } else if (vm_radix_keybarr(rnode, index))
 439                         break;
 440                 slot = vm_radix_slot(index, rnode->rn_clev);
 441                 rnode = rnode->rn_child[slot];
 442         }
 443         return (NULL);
 444 }
 445
 446 /*
 447  * Look up the nearest entry at a position bigger than or equal to index.
 448  */
 449 vm_page_t
 450 vm_radix_lookup_ge(struct vm_radix *rtree, vm_pindex_t index)
 451 {
 452         struct vm_radix_node *stack[VM_RADIX_LIMIT];
 453         vm_pindex_t inc;
 454         vm_page_t m;
 455         struct vm_radix_node *child, *rnode;
 456 #ifdef INVARIANTS
 457         int loops = 0;
 458 #endif
 459         int slot, tos;
 460
 461         rnode = vm_radix_getroot(rtree);
 462         if (rnode == NULL)
 463                 return (NULL);
 464         else if (vm_radix_isleaf(rnode)) {
 465                 m = vm_radix_topage(rnode);
 466                 if (m->pindex >= index)
 467                         return (m);
 468                 else
 469                         return (NULL);
 470         }
 471         tos = 0;
 472         for (;;) {
 473                 /*
 474                  * If the keys differ before the current bisection node,
 475                  * then the search key might rollback to the earliest
 476                  * available bisection node or to the smallest key
 477                  * in the current node (if the owner is bigger than the
 478                  * search key).
 479                  */
 480                 if (vm_radix_keybarr(rnode, index)) {
 481                         if (index > rnode->rn_owner) {
 482 ascend:
 483                                 KASSERT(++loops < 1000,
 484                                     ("vm_radix_lookup_ge: too many loops"));
 485
 486                                 /*
 487                                  * Pop nodes from the stack until either the
 488                                  * stack is empty or a node that could have a
 489                                  * matching descendant is found.
 490                                  */
 491                                 do {
 492                                         if (tos == 0)
 493                                                 return (NULL);
 494                                         rnode = stack[--tos];
 495                                 } while (vm_radix_slot(index,
 496                                     rnode->rn_clev) == (VM_RADIX_COUNT - 1));
 497
 498                                 /*
 499                                  * The following computation cannot overflow
 500                                  * because index's slot at the current level
 501                                  * is less than VM_RADIX_COUNT - 1.
 502                                  */
 503                                 index = vm_radix_trimkey(index,
 504                                     rnode->rn_clev);
 505                                 index += VM_RADIX_UNITLEVEL(rnode->rn_clev);
 506                         } else
 507                                 index = rnode->rn_owner;
 508                         KASSERT(!vm_radix_keybarr(rnode, index),
 509                             ("vm_radix_lookup_ge: keybarr failed"));
 510                 }
 511                 slot = vm_radix_slot(index, rnode->rn_clev);
 512                 child = rnode->rn_child[slot];
 513                 if (vm_radix_isleaf(child)) {
 514                         m = vm_radix_topage(child);
 515                         if (m->pindex >= index)
 516                                 return (m);
 517                 } else if (child != NULL)
 518                         goto descend;
 519
 520                 /*
 521                  * Look for an available edge or page within the current
 522                  * bisection node.
 523                  */
 524                 if (slot < (VM_RADIX_COUNT - 1)) {
 525                         inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
 526                         index = vm_radix_trimkey(index, rnode->rn_clev);
 527                         do {
 528                                 index += inc;
 529                                 slot++;
 530                                 child = rnode->rn_child[slot];
 531                                 if (vm_radix_isleaf(child)) {
 532                                         m = vm_radix_topage(child);
 533                                         if (m->pindex >= index)
 534                                                 return (m);
 535                                 } else if (child != NULL)
 536                                         goto descend;
 537                         } while (slot < (VM_RADIX_COUNT - 1));
 538                 }
 539                 KASSERT(child == NULL || vm_radix_isleaf(child),
 540                     ("vm_radix_lookup_ge: child is radix node"));
 541
 542                 /*
 543                  * If a page or edge bigger than the search slot is not found
 544                  * in the current node, ascend to the next higher-level node.
 545                  */
 546                 goto ascend;
 547 descend:
 548                 KASSERT(rnode->rn_clev < VM_RADIX_LIMIT,
 549                     ("vm_radix_lookup_ge: pushing leaf's parent"));
 550                 KASSERT(tos < VM_RADIX_LIMIT,
 551                     ("vm_radix_lookup_ge: stack overflow"));
 552                 stack[tos++] = rnode;
 553                 rnode = child;
 554         }
 555 }
 556
 557 /*
 558  * Look up the nearest entry at a position less than or equal to index.
 559  */
 560 vm_page_t
 561 vm_radix_lookup_le(struct vm_radix *rtree, vm_pindex_t index)
 562 {
 563         struct vm_radix_node *stack[VM_RADIX_LIMIT];
 564         vm_pindex_t inc;
 565         vm_page_t m;
 566         struct vm_radix_node *child, *rnode;
 567 #ifdef INVARIANTS
 568         int loops = 0;
 569 #endif
 570         int slot, tos;
 571
 572         rnode = vm_radix_getroot(rtree);
 573         if (rnode == NULL)
 574                 return (NULL);
 575         else if (vm_radix_isleaf(rnode)) {
 576                 m = vm_radix_topage(rnode);
 577                 if (m->pindex <= index)
 578                         return (m);
 579                 else
 580                         return (NULL);
 581         }
 582         tos = 0;
 583         for (;;) {
 584                 /*
 585                  * If the keys differ before the current bisection node,
 586                  * then the search key might rollback to the earliest
 587                  * available bisection node or to the largest key
 588                  * in the current node (if the owner is smaller than the
 589                  * search key).
 590                  */
 591                 if (vm_radix_keybarr(rnode, index)) {
 592                         if (index > rnode->rn_owner) {
 593                                 index = rnode->rn_owner + VM_RADIX_COUNT *
 594                                     VM_RADIX_UNITLEVEL(rnode->rn_clev);
 595                         } else {
 596 ascend:
 597                                 KASSERT(++loops < 1000,
 598                                     ("vm_radix_lookup_le: too many loops"));
 599
 600                                 /*
 601                                  * Pop nodes from the stack until either the
 602                                  * stack is empty or a node that could have a
 603                                  * matching descendant is found.
 604                                  */
 605                                 do {
 606                                         if (tos == 0)
 607                                                 return (NULL);
 608                                         rnode = stack[--tos];
 609                                 } while (vm_radix_slot(index,
 610                                     rnode->rn_clev) == 0);
 611
 612                                 /*
 613                                  * The following computation cannot overflow
 614                                  * because index's slot at the current level
 615                                  * is greater than 0.
 616                                  */
 617                                 index = vm_radix_trimkey(index,
 618                                     rnode->rn_clev);
 619                         }
 620                         index--;
 621                         KASSERT(!vm_radix_keybarr(rnode, index),
 622                             ("vm_radix_lookup_le: keybarr failed"));
 623                 }
 624                 slot = vm_radix_slot(index, rnode->rn_clev);
 625                 child = rnode->rn_child[slot];
 626                 if (vm_radix_isleaf(child)) {
 627                         m = vm_radix_topage(child);
 628                         if (m->pindex <= index)
 629                                 return (m);
 630                 } else if (child != NULL)
 631                         goto descend;
 632
 633                 /*
 634                  * Look for an available edge or page within the current
 635                  * bisection node.
 636                  */
 637                 if (slot > 0) {
 638                         inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
 639                         index |= inc - 1;
 640                         do {
 641                                 index -= inc;
 642                                 slot--;
 643                                 child = rnode->rn_child[slot];
 644                                 if (vm_radix_isleaf(child)) {
 645                                         m = vm_radix_topage(child);
 646                                         if (m->pindex <= index)
 647                                                 return (m);
 648                                 } else if (child != NULL)
 649                                         goto descend;
 650                         } while (slot > 0);
 651                 }
 652                 KASSERT(child == NULL || vm_radix_isleaf(child),
 653                     ("vm_radix_lookup_le: child is radix node"));
 654
 655                 /*
 656                  * If a page or edge smaller than the search slot is not found
 657                  * in the current node, ascend to the next higher-level node.
 658                  */
 659                 goto ascend;
 660 descend:
 661                 KASSERT(rnode->rn_clev < VM_RADIX_LIMIT,
 662                     ("vm_radix_lookup_le: pushing leaf's parent"));
 663                 KASSERT(tos < VM_RADIX_LIMIT,
 664                     ("vm_radix_lookup_le: stack overflow"));
 665                 stack[tos++] = rnode;
 666                 rnode = child;
 667         }
 668 }
 669
 670 /*
 671  * Remove the specified index from the tree.
 672  * Panics if the key is not present.
 673  */
 674 void
 675 vm_radix_remove(struct vm_radix *rtree, vm_pindex_t index)
 676 {
 677         struct vm_radix_node *rnode, *parent;
 678         vm_page_t m;
 679         int i, slot;
 680
 681         rnode = vm_radix_getroot(rtree);
 682         if (vm_radix_isleaf(rnode)) {
 683                 m = vm_radix_topage(rnode);
 684                 if (m->pindex != index)
 685                         panic("%s: invalid key found", __func__);
 686                 vm_radix_setroot(rtree, NULL);
 687                 return;
 688         }
 689         parent = NULL;
 690         for (;;) {
 691                 if (rnode == NULL)
 692                         panic("vm_radix_remove: impossible to locate the key");
 693                 slot = vm_radix_slot(index, rnode->rn_clev);
 694                 if (vm_radix_isleaf(rnode->rn_child[slot])) {
 695                         m = vm_radix_topage(rnode->rn_child[slot]);
 696                         if (m->pindex != index)
 697                                 panic("%s: invalid key found", __func__);
 698                         rnode->rn_child[slot] = NULL;
 699                         rnode->rn_count--;
 700                         if (rnode->rn_count > 1)
 701                                 break;
 702                         for (i = 0; i < VM_RADIX_COUNT; i++)
 703                                 if (rnode->rn_child[i] != NULL)
 704                                         break;
 705                         KASSERT(i != VM_RADIX_COUNT,
 706                             ("%s: invalid node configuration", __func__));
 707                         if (parent == NULL)
 708                                 vm_radix_setroot(rtree, rnode->rn_child[i]);
 709                         else {
 710                                 slot = vm_radix_slot(index, parent->rn_clev);
 711                                 KASSERT(parent->rn_child[slot] == rnode,
 712                                     ("%s: invalid child value", __func__));
 713                                 parent->rn_child[slot] = rnode->rn_child[i];
 714                         }
 715                         rnode->rn_count--;
 716                         rnode->rn_child[i] = NULL;
 717                         vm_radix_node_put(rnode);
 718                         break;
 719                 }
 720                 parent = rnode;
 721                 rnode = rnode->rn_child[slot];
 722         }
 723 }
 724
 725 /*
 726  * Remove and free all the nodes from the radix tree.
 727  * This function is recursive but there is a tight control on it as the
 728  * maximum depth of the tree is fixed.
 729  */
 730 void
 731 vm_radix_reclaim_allnodes(struct vm_radix *rtree)
 732 {
 733         struct vm_radix_node *root;
 734
 735         root = vm_radix_getroot(rtree);
 736         if (root == NULL)
 737                 return;
 738         vm_radix_setroot(rtree, NULL);
 739         if (!vm_radix_isleaf(root))
 740                 vm_radix_reclaim_allnodes_int(root);
 741 }
 742
 743 #ifdef DDB
 744 /*
 745  * Show details about the given radix node.
 746  */
 747 DB_SHOW_COMMAND(radixnode, db_show_radixnode)
 748 {
 749         struct vm_radix_node *rnode;
 750         int i;
 751
 752         if (!have_addr)
 753                 return;
 754         rnode = (struct vm_radix_node *)addr;
 755         db_printf("radixnode %p, owner %jx, children count %u, level %u:\n",
 756             (void *)rnode, (uintmax_t)rnode->rn_owner, rnode->rn_count,
 757             rnode->rn_clev);
 758         for (i = 0; i < VM_RADIX_COUNT; i++)
 759                 if (rnode->rn_child[i] != NULL)
 760                         db_printf("slot: %d, val: %p, page: %p, clev: %d\n",
 761                             i, (void *)rnode->rn_child[i],
 762                             vm_radix_isleaf(rnode->rn_child[i]) ?
 763                             vm_radix_topage(rnode->rn_child[i]) : NULL,
 764                             rnode->rn_clev);
 765 }
 766 #endif /* DDB */