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 << ((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.
 107  */
 108 static __inline struct vm_radix_node *
 109 vm_radix_node_get(vm_pindex_t owner, uint16_t count, uint16_t clevel)
 110 {
 111         struct vm_radix_node *rnode;
 112
 113         rnode = uma_zalloc(vm_radix_node_zone, M_NOWAIT | M_ZERO);
 114         if (rnode == NULL)
 115                 return (NULL);
 116         rnode->rn_owner = owner;
 117         rnode->rn_count = count;
 118         rnode->rn_clev = clevel;
 119         return (rnode);
 120 }
 121
 122 /*
 123  * Free radix node.
 124  */
 125 static __inline void
 126 vm_radix_node_put(struct vm_radix_node *rnode)
 127 {
 128
 129         uma_zfree(vm_radix_node_zone, rnode);
 130 }
 131
 132 /*
 133  * Return the position in the array for a given level.
 134  */
 135 static __inline int
 136 vm_radix_slot(vm_pindex_t index, uint16_t level)
 137 {
 138
 139         return ((index >> (level * VM_RADIX_WIDTH)) & VM_RADIX_MASK);
 140 }
 141
 142 /* Trims the key after the specified level. */
 143 static __inline vm_pindex_t
 144 vm_radix_trimkey(vm_pindex_t index, uint16_t level)
 145 {
 146         vm_pindex_t ret;
 147
 148         ret = index;
 149         if (level > 0) {
 150                 ret >>= level * VM_RADIX_WIDTH;
 151                 ret <<= level * VM_RADIX_WIDTH;
 152         }
 153         return (ret);
 154 }
 155
 156 /*
 157  * Get the root node for a radix tree.
 158  */
 159 static __inline struct vm_radix_node *
 160 vm_radix_getroot(struct vm_radix *rtree)
 161 {
 162
 163         return ((struct vm_radix_node *)rtree->rt_root);
 164 }
 165
 166 /*
 167  * Set the root node for a radix tree.
 168  */
 169 static __inline void
 170 vm_radix_setroot(struct vm_radix *rtree, struct vm_radix_node *rnode)
 171 {
 172
 173         rtree->rt_root = (uintptr_t)rnode;
 174 }
 175
 176 /*
 177  * Returns TRUE if the specified radix node is a leaf and FALSE otherwise.
 178  */
 179 static __inline boolean_t
 180 vm_radix_isleaf(struct vm_radix_node *rnode)
 181 {
 182
 183         return (((uintptr_t)rnode & VM_RADIX_ISLEAF) != 0);
 184 }
 185
 186 /*
 187  * Returns the associated page extracted from rnode.
 188  */
 189 static __inline vm_page_t
 190 vm_radix_topage(struct vm_radix_node *rnode)
 191 {
 192
 193         return ((vm_page_t)((uintptr_t)rnode & ~VM_RADIX_FLAGS));
 194 }
 195
 196 /*
 197  * Adds the page as a child of the provided node.
 198  */
 199 static __inline void
 200 vm_radix_addpage(struct vm_radix_node *rnode, vm_pindex_t index, uint16_t clev,
 201     vm_page_t page)
 202 {
 203         int slot;
 204
 205         slot = vm_radix_slot(index, clev);
 206         rnode->rn_child[slot] = (void *)((uintptr_t)page | VM_RADIX_ISLEAF);
 207 }
 208
 209 /*
 210  * Returns the slot where two keys differ.
 211  * It cannot accept 2 equal keys.
 212  */
 213 static __inline uint16_t
 214 vm_radix_keydiff(vm_pindex_t index1, vm_pindex_t index2)
 215 {
 216         uint16_t clev;
 217
 218         KASSERT(index1 != index2, ("%s: passing the same key value %jx",
 219             __func__, (uintmax_t)index1));
 220
 221         index1 ^= index2;
 222         for (clev = VM_RADIX_LIMIT;; clev--)
 223                 if (vm_radix_slot(index1, clev) != 0)
 224                         return (clev);
 225 }
 226
 227 /*
 228  * Returns TRUE if it can be determined that key does not belong to the
 229  * specified rnode.  Otherwise, returns FALSE.
 230  */
 231 static __inline boolean_t
 232 vm_radix_keybarr(struct vm_radix_node *rnode, vm_pindex_t idx)
 233 {
 234
 235         if (rnode->rn_clev < VM_RADIX_LIMIT) {
 236                 idx = vm_radix_trimkey(idx, rnode->rn_clev + 1);
 237                 return (idx != rnode->rn_owner);
 238         }
 239         return (FALSE);
 240 }
 241
 242 /*
 243  * Internal helper for vm_radix_reclaim_allnodes().
 244  * This function is recursive.
 245  */
 246 static void
 247 vm_radix_reclaim_allnodes_int(struct vm_radix_node *rnode)
 248 {
 249         int slot;
 250
 251         KASSERT(rnode->rn_count <= VM_RADIX_COUNT,
 252             ("vm_radix_reclaim_allnodes_int: bad count in rnode %p", rnode));
 253         for (slot = 0; rnode->rn_count != 0; slot++) {
 254                 if (rnode->rn_child[slot] == NULL)
 255                         continue;
 256                 if (!vm_radix_isleaf(rnode->rn_child[slot]))
 257                         vm_radix_reclaim_allnodes_int(rnode->rn_child[slot]);
 258                 rnode->rn_child[slot] = NULL;
 259                 rnode->rn_count--;
 260         }
 261         vm_radix_node_put(rnode);
 262 }
 263
 264 #ifdef INVARIANTS
 265 /*
 266  * Radix node zone destructor.
 267  */
 268 static void
 269 vm_radix_node_zone_dtor(void *mem, int size __unused, void *arg __unused)
 270 {
 271         struct vm_radix_node *rnode;
 272         int slot;
 273
 274         rnode = mem;
 275         KASSERT(rnode->rn_count == 0,
 276             ("vm_radix_node_put: rnode %p has %d children", rnode,
 277             rnode->rn_count));
 278         for (slot = 0; slot < VM_RADIX_COUNT; slot++)
 279                 KASSERT(rnode->rn_child[slot] == NULL,
 280                     ("vm_radix_node_put: rnode %p has a child", rnode));
 281 }
 282 #endif
 283
 284 #ifndef UMA_MD_SMALL_ALLOC
 285 /*
 286  * Reserve the KVA necessary to satisfy the node allocation.
 287  * This is mandatory in architectures not supporting direct
 288  * mapping as they will need otherwise to carve into the kernel maps for
 289  * every node allocation, resulting into deadlocks for consumers already
 290  * working with kernel maps.
 291  */
 292 static void
 293 vm_radix_reserve_kva(void *arg __unused)
 294 {
 295
 296         /*
 297          * Calculate the number of reserved nodes, discounting the pages that
 298          * are needed to store them.
 299          */
 300         if (!uma_zone_reserve_kva(vm_radix_node_zone,
 301             ((vm_paddr_t)cnt.v_page_count * PAGE_SIZE) / (PAGE_SIZE +
 302             sizeof(struct vm_radix_node))))
 303                 panic("%s: unable to reserve KVA", __func__);
 304 }
 305 SYSINIT(vm_radix_reserve_kva, SI_SUB_KMEM, SI_ORDER_SECOND,
 306     vm_radix_reserve_kva, NULL);
 307 #endif
 308
 309 /*
 310  * Initialize the UMA slab zone.
 311  * Until vm_radix_prealloc() is called, the zone will be served by the
 312  * UMA boot-time pre-allocated pool of pages.
 313  */
 314 void
 315 vm_radix_init(void)
 316 {
 317
 318         vm_radix_node_zone = uma_zcreate("RADIX NODE",
 319             sizeof(struct vm_radix_node), NULL,
 320 #ifdef INVARIANTS
 321             vm_radix_node_zone_dtor,
 322 #else
 323             NULL,
 324 #endif
 325             NULL, NULL, VM_RADIX_PAD, UMA_ZONE_VM);
 326 }
 327
 328 /*
 329  * Inserts the key-value pair into the trie.
 330  * Panics if the key already exists.
 331  */
 332 int
 333 vm_radix_insert(struct vm_radix *rtree, vm_page_t page)
 334 {
 335         vm_pindex_t index, newind;
 336         void **parentp;
 337         struct vm_radix_node *rnode, *tmp;
 338         vm_page_t m;
 339         int slot;
 340         uint16_t clev;
 341
 342         index = page->pindex;
 343
 344 restart:
 345
 346         /*
 347          * The owner of record for root is not really important because it
 348          * will never be used.
 349          */
 350         rnode = vm_radix_getroot(rtree);
 351         if (rnode == NULL) {
 352                 rtree->rt_root = (uintptr_t)page | VM_RADIX_ISLEAF;
 353                 return (0);
 354         }
 355         parentp = (void **)&rtree->rt_root;
 356         for (;;) {
 357                 if (vm_radix_isleaf(rnode)) {
 358                         m = vm_radix_topage(rnode);
 359                         if (m->pindex == index)
 360                                 panic("%s: key %jx is already present",
 361                                     __func__, (uintmax_t)index);
 362                         clev = vm_radix_keydiff(m->pindex, index);
 363
 364                         /*
 365                          * During node allocation the trie that is being
 366                          * walked can be modified because of recursing radix
 367                          * trie operations.
 368                          * If this is the case, the recursing functions signal
 369                          * such situation and the insert operation must
 370                          * start from scratch again.
 371                          * The freed radix node will then be in the UMA
 372                          * caches very likely to avoid the same situation
 373                          * to happen.
 374                          */
 375                         rtree->rt_flags |= RT_INSERT_INPROG;
 376                         tmp = vm_radix_node_get(vm_radix_trimkey(index,
 377                             clev + 1), 2, clev);
 378                         rtree->rt_flags &= ~RT_INSERT_INPROG;
 379                         if (tmp == NULL) {
 380                                 rtree->rt_flags &= ~RT_TRIE_MODIFIED;
 381                                 return (ENOMEM);
 382                         }
 383                         if ((rtree->rt_flags & RT_TRIE_MODIFIED) != 0) {
 384                                 rtree->rt_flags &= ~RT_TRIE_MODIFIED;
 385                                 tmp->rn_count = 0;
 386                                 vm_radix_node_put(tmp);
 387                                 goto restart;
 388                         }
 389                         *parentp = tmp;
 390                         vm_radix_addpage(tmp, index, clev, page);
 391                         vm_radix_addpage(tmp, m->pindex, clev, m);
 392                         return (0);
 393                 } else if (vm_radix_keybarr(rnode, index))
 394                         break;
 395                 slot = vm_radix_slot(index, rnode->rn_clev);
 396                 if (rnode->rn_child[slot] == NULL) {
 397                         rnode->rn_count++;
 398                         vm_radix_addpage(rnode, index, rnode->rn_clev, page);
 399                         return (0);
 400                 }
 401                 parentp = &rnode->rn_child[slot];
 402                 rnode = rnode->rn_child[slot];
 403         }
 404
 405         /*
 406          * A new node is needed because the right insertion level is reached.
 407          * Setup the new intermediate node and add the 2 children: the
 408          * new object and the older edge.
 409          */
 410         newind = rnode->rn_owner;
 411         clev = vm_radix_keydiff(newind, index);
 412
 413         /* See the comments above. */
 414         rtree->rt_flags |= RT_INSERT_INPROG;
 415         tmp = vm_radix_node_get(vm_radix_trimkey(index, clev + 1), 2, clev);
 416         rtree->rt_flags &= ~RT_INSERT_INPROG;
 417         if (tmp == NULL) {
 418                 rtree->rt_flags &= ~RT_TRIE_MODIFIED;
 419                 return (ENOMEM);
 420         }
 421         if ((rtree->rt_flags & RT_TRIE_MODIFIED) != 0) {
 422                 rtree->rt_flags &= ~RT_TRIE_MODIFIED;
 423                 tmp->rn_count = 0;
 424                 vm_radix_node_put(tmp);
 425                 goto restart;
 426         }
 427         *parentp = tmp;
 428         vm_radix_addpage(tmp, index, clev, page);
 429         slot = vm_radix_slot(newind, clev);
 430         tmp->rn_child[slot] = rnode;
 431         return (0);
 432 }
 433
 434 /*
 435  * Returns TRUE if the specified radix tree contains a single leaf and FALSE
 436  * otherwise.
 437  */
 438 boolean_t
 439 vm_radix_is_singleton(struct vm_radix *rtree)
 440 {
 441         struct vm_radix_node *rnode;
 442
 443         rnode = vm_radix_getroot(rtree);
 444         if (rnode == NULL)
 445                 return (FALSE);
 446         return (vm_radix_isleaf(rnode));
 447 }
 448
 449 /*
 450  * Returns the value stored at the index.  If the index is not present,
 451  * NULL is returned.
 452  */
 453 vm_page_t
 454 vm_radix_lookup(struct vm_radix *rtree, vm_pindex_t index)
 455 {
 456         struct vm_radix_node *rnode;
 457         vm_page_t m;
 458         int slot;
 459
 460         rnode = vm_radix_getroot(rtree);
 461         while (rnode != NULL) {
 462                 if (vm_radix_isleaf(rnode)) {
 463                         m = vm_radix_topage(rnode);
 464                         if (m->pindex == index)
 465                                 return (m);
 466                         else
 467                                 break;
 468                 } else if (vm_radix_keybarr(rnode, index))
 469                         break;
 470                 slot = vm_radix_slot(index, rnode->rn_clev);
 471                 rnode = rnode->rn_child[slot];
 472         }
 473         return (NULL);
 474 }
 475
 476 /*
 477  * Look up the nearest entry at a position bigger than or equal to index.
 478  */
 479 vm_page_t
 480 vm_radix_lookup_ge(struct vm_radix *rtree, vm_pindex_t index)
 481 {
 482         struct vm_radix_node *stack[VM_RADIX_LIMIT];
 483         vm_pindex_t inc;
 484         vm_page_t m;
 485         struct vm_radix_node *child, *rnode;
 486 #ifdef INVARIANTS
 487         int loops = 0;
 488 #endif
 489         int slot, tos;
 490
 491         rnode = vm_radix_getroot(rtree);
 492         if (rnode == NULL)
 493                 return (NULL);
 494         else if (vm_radix_isleaf(rnode)) {
 495                 m = vm_radix_topage(rnode);
 496                 if (m->pindex >= index)
 497                         return (m);
 498                 else
 499                         return (NULL);
 500         }
 501         tos = 0;
 502         for (;;) {
 503                 /*
 504                  * If the keys differ before the current bisection node,
 505                  * then the search key might rollback to the earliest
 506                  * available bisection node or to the smallest key
 507                  * in the current node (if the owner is bigger than the
 508                  * search key).
 509                  */
 510                 if (vm_radix_keybarr(rnode, index)) {
 511                         if (index > rnode->rn_owner) {
 512 ascend:
 513                                 KASSERT(++loops < 1000,
 514                                     ("vm_radix_lookup_ge: too many loops"));
 515
 516                                 /*
 517                                  * Pop nodes from the stack until either the
 518                                  * stack is empty or a node that could have a
 519                                  * matching descendant is found.
 520                                  */
 521                                 do {
 522                                         if (tos == 0)
 523                                                 return (NULL);
 524                                         rnode = stack[--tos];
 525                                 } while (vm_radix_slot(index,
 526                                     rnode->rn_clev) == (VM_RADIX_COUNT - 1));
 527
 528                                 /*
 529                                  * The following computation cannot overflow
 530                                  * because index's slot at the current level
 531                                  * is less than VM_RADIX_COUNT - 1.
 532                                  */
 533                                 index = vm_radix_trimkey(index,
 534                                     rnode->rn_clev);
 535                                 index += VM_RADIX_UNITLEVEL(rnode->rn_clev);
 536                         } else
 537                                 index = rnode->rn_owner;
 538                         KASSERT(!vm_radix_keybarr(rnode, index),
 539                             ("vm_radix_lookup_ge: keybarr failed"));
 540                 }
 541                 slot = vm_radix_slot(index, rnode->rn_clev);
 542                 child = rnode->rn_child[slot];
 543                 if (vm_radix_isleaf(child)) {
 544                         m = vm_radix_topage(child);
 545                         if (m->pindex >= index)
 546                                 return (m);
 547                 } else if (child != NULL)
 548                         goto descend;
 549
 550                 /*
 551                  * Look for an available edge or page within the current
 552                  * bisection node.
 553                  */
 554                 if (slot < (VM_RADIX_COUNT - 1)) {
 555                         inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
 556                         index = vm_radix_trimkey(index, rnode->rn_clev);
 557                         do {
 558                                 index += inc;
 559                                 slot++;
 560                                 child = rnode->rn_child[slot];
 561                                 if (vm_radix_isleaf(child)) {
 562                                         m = vm_radix_topage(child);
 563                                         if (m->pindex >= index)
 564                                                 return (m);
 565                                 } else if (child != NULL)
 566                                         goto descend;
 567                         } while (slot < (VM_RADIX_COUNT - 1));
 568                 }
 569                 KASSERT(child == NULL || vm_radix_isleaf(child),
 570                     ("vm_radix_lookup_ge: child is radix node"));
 571
 572                 /*
 573                  * If a page or edge bigger than the search slot is not found
 574                  * in the current node, ascend to the next higher-level node.
 575                  */
 576                 goto ascend;
 577 descend:
 578                 KASSERT(rnode->rn_clev > 0,
 579                     ("vm_radix_lookup_ge: pushing leaf's parent"));
 580                 KASSERT(tos < VM_RADIX_LIMIT,
 581                     ("vm_radix_lookup_ge: stack overflow"));
 582                 stack[tos++] = rnode;
 583                 rnode = child;
 584         }
 585 }
 586
 587 /*
 588  * Look up the nearest entry at a position less than or equal to index.
 589  */
 590 vm_page_t
 591 vm_radix_lookup_le(struct vm_radix *rtree, vm_pindex_t index)
 592 {
 593         struct vm_radix_node *stack[VM_RADIX_LIMIT];
 594         vm_pindex_t inc;
 595         vm_page_t m;
 596         struct vm_radix_node *child, *rnode;
 597 #ifdef INVARIANTS
 598         int loops = 0;
 599 #endif
 600         int slot, tos;
 601
 602         rnode = vm_radix_getroot(rtree);
 603         if (rnode == NULL)
 604                 return (NULL);
 605         else if (vm_radix_isleaf(rnode)) {
 606                 m = vm_radix_topage(rnode);
 607                 if (m->pindex <= index)
 608                         return (m);
 609                 else
 610                         return (NULL);
 611         }
 612         tos = 0;
 613         for (;;) {
 614                 /*
 615                  * If the keys differ before the current bisection node,
 616                  * then the search key might rollback to the earliest
 617                  * available bisection node or to the largest key
 618                  * in the current node (if the owner is smaller than the
 619                  * search key).
 620                  */
 621                 if (vm_radix_keybarr(rnode, index)) {
 622                         if (index > rnode->rn_owner) {
 623                                 index = rnode->rn_owner + VM_RADIX_COUNT *
 624                                     VM_RADIX_UNITLEVEL(rnode->rn_clev);
 625                         } else {
 626 ascend:
 627                                 KASSERT(++loops < 1000,
 628                                     ("vm_radix_lookup_le: too many loops"));
 629
 630                                 /*
 631                                  * Pop nodes from the stack until either the
 632                                  * stack is empty or a node that could have a
 633                                  * matching descendant is found.
 634                                  */
 635                                 do {
 636                                         if (tos == 0)
 637                                                 return (NULL);
 638                                         rnode = stack[--tos];
 639                                 } while (vm_radix_slot(index,
 640                                     rnode->rn_clev) == 0);
 641
 642                                 /*
 643                                  * The following computation cannot overflow
 644                                  * because index's slot at the current level
 645                                  * is greater than 0.
 646                                  */
 647                                 index = vm_radix_trimkey(index,
 648                                     rnode->rn_clev);
 649                         }
 650                         index--;
 651                         KASSERT(!vm_radix_keybarr(rnode, index),
 652                             ("vm_radix_lookup_le: keybarr failed"));
 653                 }
 654                 slot = vm_radix_slot(index, rnode->rn_clev);
 655                 child = rnode->rn_child[slot];
 656                 if (vm_radix_isleaf(child)) {
 657                         m = vm_radix_topage(child);
 658                         if (m->pindex <= index)
 659                                 return (m);
 660                 } else if (child != NULL)
 661                         goto descend;
 662
 663                 /*
 664                  * Look for an available edge or page within the current
 665                  * bisection node.
 666                  */
 667                 if (slot > 0) {
 668                         inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
 669                         index |= inc - 1;
 670                         do {
 671                                 index -= inc;
 672                                 slot--;
 673                                 child = rnode->rn_child[slot];
 674                                 if (vm_radix_isleaf(child)) {
 675                                         m = vm_radix_topage(child);
 676                                         if (m->pindex <= index)
 677                                                 return (m);
 678                                 } else if (child != NULL)
 679                                         goto descend;
 680                         } while (slot > 0);
 681                 }
 682                 KASSERT(child == NULL || vm_radix_isleaf(child),
 683                     ("vm_radix_lookup_le: child is radix node"));
 684
 685                 /*
 686                  * If a page or edge smaller than the search slot is not found
 687                  * in the current node, ascend to the next higher-level node.
 688                  */
 689                 goto ascend;
 690 descend:
 691                 KASSERT(rnode->rn_clev > 0,
 692                     ("vm_radix_lookup_le: pushing leaf's parent"));
 693                 KASSERT(tos < VM_RADIX_LIMIT,
 694                     ("vm_radix_lookup_le: stack overflow"));
 695                 stack[tos++] = rnode;
 696                 rnode = child;
 697         }
 698 }
 699
 700 /*
 701  * Remove the specified index from the tree.
 702  * Panics if the key is not present.
 703  */
 704 void
 705 vm_radix_remove(struct vm_radix *rtree, vm_pindex_t index)
 706 {
 707         struct vm_radix_node *rnode, *parent;
 708         vm_page_t m;
 709         int i, slot;
 710
 711         /*
 712          * Detect if a page is going to be removed from a trie which is
 713          * already undergoing another trie operation.
 714          * Right now this is only possible for vm_radix_remove() recursing
 715          * into vm_radix_insert().
 716          * If this is the case, the caller must be notified about this
 717          * situation.  It will also takecare to update the RT_TRIE_MODIFIED
 718          * accordingly.
 719          * The RT_TRIE_MODIFIED bit is set here because the remove operation
 720          * will always succeed.
 721          */
 722         if ((rtree->rt_flags & RT_INSERT_INPROG) != 0)
 723                 rtree->rt_flags |= RT_TRIE_MODIFIED;
 724
 725         rnode = vm_radix_getroot(rtree);
 726         if (vm_radix_isleaf(rnode)) {
 727                 m = vm_radix_topage(rnode);
 728                 if (m->pindex != index)
 729                         panic("%s: invalid key found", __func__);
 730                 vm_radix_setroot(rtree, NULL);
 731                 return;
 732         }
 733         parent = NULL;
 734         for (;;) {
 735                 if (rnode == NULL)
 736                         panic("vm_radix_remove: impossible to locate the key");
 737                 slot = vm_radix_slot(index, rnode->rn_clev);
 738                 if (vm_radix_isleaf(rnode->rn_child[slot])) {
 739                         m = vm_radix_topage(rnode->rn_child[slot]);
 740                         if (m->pindex != index)
 741                                 panic("%s: invalid key found", __func__);
 742                         rnode->rn_child[slot] = NULL;
 743                         rnode->rn_count--;
 744                         if (rnode->rn_count > 1)
 745                                 break;
 746                         for (i = 0; i < VM_RADIX_COUNT; i++)
 747                                 if (rnode->rn_child[i] != NULL)
 748                                         break;
 749                         KASSERT(i != VM_RADIX_COUNT,
 750                             ("%s: invalid node configuration", __func__));
 751                         if (parent == NULL)
 752                                 vm_radix_setroot(rtree, rnode->rn_child[i]);
 753                         else {
 754                                 slot = vm_radix_slot(index, parent->rn_clev);
 755                                 KASSERT(parent->rn_child[slot] == rnode,
 756                                     ("%s: invalid child value", __func__));
 757                                 parent->rn_child[slot] = rnode->rn_child[i];
 758                         }
 759                         rnode->rn_count--;
 760                         rnode->rn_child[i] = NULL;
 761                         vm_radix_node_put(rnode);
 762                         break;
 763                 }
 764                 parent = rnode;
 765                 rnode = rnode->rn_child[slot];
 766         }
 767 }
 768
 769 /*
 770  * Remove and free all the nodes from the radix tree.
 771  * This function is recursive but there is a tight control on it as the
 772  * maximum depth of the tree is fixed.
 773  */
 774 void
 775 vm_radix_reclaim_allnodes(struct vm_radix *rtree)
 776 {
 777         struct vm_radix_node *root;
 778
 779         KASSERT((rtree->rt_flags & RT_INSERT_INPROG) == 0,
 780             ("vm_radix_reclaim_allnodes: unexpected trie recursion"));
 781
 782         root = vm_radix_getroot(rtree);
 783         if (root == NULL)
 784                 return;
 785         vm_radix_setroot(rtree, NULL);
 786         if (!vm_radix_isleaf(root))
 787                 vm_radix_reclaim_allnodes_int(root);
 788 }
 789
 790 /*
 791  * Replace an existing page into the trie with another one.
 792  * Panics if the replacing page is not present or if the new page has an
 793  * invalid key.
 794  */
 795 vm_page_t
 796 vm_radix_replace(struct vm_radix *rtree, vm_page_t newpage, vm_pindex_t index)
 797 {
 798         struct vm_radix_node *rnode;
 799         vm_page_t m;
 800         int slot;
 801
 802         KASSERT(newpage->pindex == index, ("%s: newpage index invalid",
 803             __func__));
 804
 805         rnode = vm_radix_getroot(rtree);
 806         if (rnode == NULL)
 807                 panic("%s: replacing page on an empty trie", __func__);
 808         if (vm_radix_isleaf(rnode)) {
 809                 m = vm_radix_topage(rnode);
 810                 if (m->pindex != index)
 811                         panic("%s: original replacing root key not found",
 812                             __func__);
 813                 rtree->rt_root = (uintptr_t)newpage | VM_RADIX_ISLEAF;
 814                 return (m);
 815         }
 816         for (;;) {
 817                 slot = vm_radix_slot(index, rnode->rn_clev);
 818                 if (vm_radix_isleaf(rnode->rn_child[slot])) {
 819                         m = vm_radix_topage(rnode->rn_child[slot]);
 820                         if (m->pindex == index) {
 821                                 rnode->rn_child[slot] =
 822                                     (void *)((uintptr_t)newpage |
 823                                     VM_RADIX_ISLEAF);
 824                                 return (m);
 825                         } else
 826                                 break;
 827                 } else if (rnode->rn_child[slot] == NULL ||
 828                     vm_radix_keybarr(rnode->rn_child[slot], index))
 829                         break;
 830                 rnode = rnode->rn_child[slot];
 831         }
 832         panic("%s: original replacing page not found", __func__);
 833 }
 834
 835 #ifdef DDB
 836 /*
 837  * Show details about the given radix node.
 838  */
 839 DB_SHOW_COMMAND(radixnode, db_show_radixnode)
 840 {
 841         struct vm_radix_node *rnode;
 842         int i;
 843
 844         if (!have_addr)
 845                 return;
 846         rnode = (struct vm_radix_node *)addr;
 847         db_printf("radixnode %p, owner %jx, children count %u, level %u:\n",
 848             (void *)rnode, (uintmax_t)rnode->rn_owner, rnode->rn_count,
 849             rnode->rn_clev);
 850         for (i = 0; i < VM_RADIX_COUNT; i++)
 851                 if (rnode->rn_child[i] != NULL)
 852                         db_printf("slot: %d, val: %p, page: %p, clev: %d\n",
 853                             i, (void *)rnode->rn_child[i],
 854                             vm_radix_isleaf(rnode->rn_child[i]) ?
 855                             vm_radix_topage(rnode->rn_child[i]) : NULL,
 856                             rnode->rn_clev);
 857 }
 858 #endif /* DDB */