sys/vm/vm_radix.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2013 EMC Corp.
   5  * Copyright (c) 2011 Jeffrey Roberson <jeff@freebsd.org>
   6  * Copyright (c) 2008 Mayur Shardul <mayur.shardul@gmail.com>
   7  * All rights reserved.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  28  * SUCH DAMAGE.
  29  *
  30  */
  31
  32 /*
  33  * Path-compressed radix trie implementation.
  34  * The following code is not generalized into a general purpose library
  35  * because there are way too many parameters embedded that should really
  36  * be decided by the library consumers.  At the same time, consumers
  37  * of this code must achieve highest possible performance.
  38  *
  39  * The implementation takes into account the following rationale:
  40  * - Size of the nodes should be as small as possible but still big enough
  41  *   to avoid a large maximum depth for the trie.  This is a balance
  42  *   between the necessity to not wire too much physical memory for the nodes
  43  *   and the necessity to avoid too much cache pollution during the trie
  44  *   operations.
  45  * - There is not a huge bias toward the number of lookup operations over
  46  *   the number of insert and remove operations.  This basically implies
  47  *   that optimizations supposedly helping one operation but hurting the
  48  *   other might be carefully evaluated.
  49  * - On average not many nodes are expected to be fully populated, hence
  50  *   level compression may just complicate things.
  51  */
  52
  53 #include <sys/cdefs.h>
  54 __FBSDID("$FreeBSD$");
  55
  56 #include "opt_ddb.h"
  57
  58 #include <sys/param.h>
  59 #include <sys/systm.h>
  60 #include <sys/kernel.h>
  61 #include <sys/proc.h>
  62 #include <sys/vmmeter.h>
  63 #include <sys/smr.h>
  64 #include <sys/smr_types.h>
  65
  66 #include <vm/uma.h>
  67 #include <vm/vm.h>
  68 #include <vm/vm_param.h>
  69 #include <vm/vm_object.h>
  70 #include <vm/vm_page.h>
  71 #include <vm/vm_radix.h>
  72
  73 #ifdef DDB
  74 #include <ddb/ddb.h>
  75 #endif
  76
  77 /*
  78  * These widths should allow the pointers to a node's children to fit within
  79  * a single cache line.  The extra levels from a narrow width should not be
  80  * a problem thanks to path compression.
  81  */
  82 #ifdef __LP64__
  83 #define VM_RADIX_WIDTH  4
  84 #else
  85 #define VM_RADIX_WIDTH  3
  86 #endif
  87
  88 #define VM_RADIX_COUNT  (1 << VM_RADIX_WIDTH)
  89 #define VM_RADIX_MASK   (VM_RADIX_COUNT - 1)
  90 #define VM_RADIX_LIMIT                                                  \
  91         (howmany(sizeof(vm_pindex_t) * NBBY, VM_RADIX_WIDTH) - 1)
  92
  93 /* Flag bits stored in node pointers. */
  94 #define VM_RADIX_ISLEAF 0x1
  95 #define VM_RADIX_FLAGS  0x1
  96 #define VM_RADIX_PAD    VM_RADIX_FLAGS
  97
  98 /* Returns one unit associated with specified level. */
  99 #define VM_RADIX_UNITLEVEL(lev)                                         \
 100         ((vm_pindex_t)1 << ((lev) * VM_RADIX_WIDTH))
 101
 102 enum vm_radix_access { SMR, LOCKED, UNSERIALIZED };
 103
 104 struct vm_radix_node;
 105 typedef SMR_POINTER(struct vm_radix_node *) smrnode_t;
 106
 107 struct vm_radix_node {
 108         vm_pindex_t     rn_owner;                       /* Owner of record. */
 109         uint16_t        rn_count;                       /* Valid children. */
 110         uint8_t         rn_clev;                        /* Current level. */
 111         int8_t          rn_last;                        /* zero last ptr. */
 112         smrnode_t       rn_child[VM_RADIX_COUNT];       /* Child nodes. */
 113 };
 114
 115 static uma_zone_t vm_radix_node_zone;
 116 static smr_t vm_radix_smr;
 117
 118 static void vm_radix_node_store(smrnode_t *p, struct vm_radix_node *v,
 119     enum vm_radix_access access);
 120
 121 /*
 122  * Allocate a radix node.
 123  */
 124 static struct vm_radix_node *
 125 vm_radix_node_get(vm_pindex_t owner, uint16_t count, uint16_t clevel)
 126 {
 127         struct vm_radix_node *rnode;
 128
 129         rnode = uma_zalloc_smr(vm_radix_node_zone, M_NOWAIT);
 130         if (rnode == NULL)
 131                 return (NULL);
 132
 133         /*
 134          * We want to clear the last child pointer after the final section
 135          * has exited so lookup can not return false negatives.  It is done
 136          * here because it will be cache-cold in the dtor callback.
 137          */
 138         if (rnode->rn_last != 0) {
 139                 vm_radix_node_store(&rnode->rn_child[rnode->rn_last - 1],
 140                     NULL, UNSERIALIZED);
 141                 rnode->rn_last = 0;
 142         }
 143         rnode->rn_owner = owner;
 144         rnode->rn_count = count;
 145         rnode->rn_clev = clevel;
 146         return (rnode);
 147 }
 148
 149 /*
 150  * Free radix node.
 151  */
 152 static __inline void
 153 vm_radix_node_put(struct vm_radix_node *rnode, int8_t last)
 154 {
 155 #ifdef INVARIANTS
 156         int slot;
 157
 158         KASSERT(rnode->rn_count == 0,
 159             ("vm_radix_node_put: rnode %p has %d children", rnode,
 160             rnode->rn_count));
 161         for (slot = 0; slot < VM_RADIX_COUNT; slot++) {
 162                 if (slot == last)
 163                         continue;
 164                 KASSERT(smr_unserialized_load(&rnode->rn_child[slot], true) ==
 165                     NULL, ("vm_radix_node_put: rnode %p has a child", rnode));
 166         }
 167 #endif
 168         /* Off by one so a freshly zero'd node is not assigned to. */
 169         rnode->rn_last = last + 1;
 170         uma_zfree_smr(vm_radix_node_zone, rnode);
 171 }
 172
 173 /*
 174  * Return the position in the array for a given level.
 175  */
 176 static __inline int
 177 vm_radix_slot(vm_pindex_t index, uint16_t level)
 178 {
 179
 180         return ((index >> (level * VM_RADIX_WIDTH)) & VM_RADIX_MASK);
 181 }
 182
 183 /* Trims the key after the specified level. */
 184 static __inline vm_pindex_t
 185 vm_radix_trimkey(vm_pindex_t index, uint16_t level)
 186 {
 187         vm_pindex_t ret;
 188
 189         ret = index;
 190         if (level > 0) {
 191                 ret >>= level * VM_RADIX_WIDTH;
 192                 ret <<= level * VM_RADIX_WIDTH;
 193         }
 194         return (ret);
 195 }
 196
 197 /*
 198  * Fetch a node pointer from a slot in another node.
 199  */
 200 static __inline struct vm_radix_node *
 201 vm_radix_node_load(smrnode_t *p, enum vm_radix_access access)
 202 {
 203
 204         switch (access) {
 205         case UNSERIALIZED:
 206                 return (smr_unserialized_load(p, true));
 207         case LOCKED:
 208                 return (smr_serialized_load(p, true));
 209         case SMR:
 210                 return (smr_entered_load(p, vm_radix_smr));
 211         }
 212         __assert_unreachable();
 213 }
 214
 215 static __inline void
 216 vm_radix_node_store(smrnode_t *p, struct vm_radix_node *v,
 217     enum vm_radix_access access)
 218 {
 219
 220         switch (access) {
 221         case UNSERIALIZED:
 222                 smr_unserialized_store(p, v, true);
 223                 break;
 224         case LOCKED:
 225                 smr_serialized_store(p, v, true);
 226                 break;
 227         case SMR:
 228                 panic("vm_radix_node_store: Not supported in smr section.");
 229         }
 230 }
 231
 232 /*
 233  * Get the root node for a radix tree.
 234  */
 235 static __inline struct vm_radix_node *
 236 vm_radix_root_load(struct vm_radix *rtree, enum vm_radix_access access)
 237 {
 238
 239         return (vm_radix_node_load((smrnode_t *)&rtree->rt_root, access));
 240 }
 241
 242 /*
 243  * Set the root node for a radix tree.
 244  */
 245 static __inline void
 246 vm_radix_root_store(struct vm_radix *rtree, struct vm_radix_node *rnode,
 247     enum vm_radix_access access)
 248 {
 249
 250         vm_radix_node_store((smrnode_t *)&rtree->rt_root, rnode, access);
 251 }
 252
 253 /*
 254  * Returns TRUE if the specified radix node is a leaf and FALSE otherwise.
 255  */
 256 static __inline boolean_t
 257 vm_radix_isleaf(struct vm_radix_node *rnode)
 258 {
 259
 260         return (((uintptr_t)rnode & VM_RADIX_ISLEAF) != 0);
 261 }
 262
 263 /*
 264  * Returns the associated page extracted from rnode.
 265  */
 266 static __inline vm_page_t
 267 vm_radix_topage(struct vm_radix_node *rnode)
 268 {
 269
 270         return ((vm_page_t)((uintptr_t)rnode & ~VM_RADIX_FLAGS));
 271 }
 272
 273 /*
 274  * Adds the page as a child of the provided node.
 275  */
 276 static __inline void
 277 vm_radix_addpage(struct vm_radix_node *rnode, vm_pindex_t index, uint16_t clev,
 278     vm_page_t page, enum vm_radix_access access)
 279 {
 280         int slot;
 281
 282         slot = vm_radix_slot(index, clev);
 283         vm_radix_node_store(&rnode->rn_child[slot],
 284             (struct vm_radix_node *)((uintptr_t)page | VM_RADIX_ISLEAF), access);
 285 }
 286
 287 /*
 288  * Returns the slot where two keys differ.
 289  * It cannot accept 2 equal keys.
 290  */
 291 static __inline uint16_t
 292 vm_radix_keydiff(vm_pindex_t index1, vm_pindex_t index2)
 293 {
 294         uint16_t clev;
 295
 296         KASSERT(index1 != index2, ("%s: passing the same key value %jx",
 297             __func__, (uintmax_t)index1));
 298
 299         index1 ^= index2;
 300         for (clev = VM_RADIX_LIMIT;; clev--)
 301                 if (vm_radix_slot(index1, clev) != 0)
 302                         return (clev);
 303 }
 304
 305 /*
 306  * Returns TRUE if it can be determined that key does not belong to the
 307  * specified rnode.  Otherwise, returns FALSE.
 308  */
 309 static __inline boolean_t
 310 vm_radix_keybarr(struct vm_radix_node *rnode, vm_pindex_t idx)
 311 {
 312
 313         if (rnode->rn_clev < VM_RADIX_LIMIT) {
 314                 idx = vm_radix_trimkey(idx, rnode->rn_clev + 1);
 315                 return (idx != rnode->rn_owner);
 316         }
 317         return (FALSE);
 318 }
 319
 320 /*
 321  * Internal helper for vm_radix_reclaim_allnodes().
 322  * This function is recursive.
 323  */
 324 static void
 325 vm_radix_reclaim_allnodes_int(struct vm_radix_node *rnode)
 326 {
 327         struct vm_radix_node *child;
 328         int slot;
 329
 330         KASSERT(rnode->rn_count <= VM_RADIX_COUNT,
 331             ("vm_radix_reclaim_allnodes_int: bad count in rnode %p", rnode));
 332         for (slot = 0; rnode->rn_count != 0; slot++) {
 333                 child = vm_radix_node_load(&rnode->rn_child[slot], UNSERIALIZED);
 334                 if (child == NULL)
 335                         continue;
 336                 if (!vm_radix_isleaf(child))
 337                         vm_radix_reclaim_allnodes_int(child);
 338                 vm_radix_node_store(&rnode->rn_child[slot], NULL, UNSERIALIZED);
 339                 rnode->rn_count--;
 340         }
 341         vm_radix_node_put(rnode, -1);
 342 }
 343
 344 #ifndef UMA_MD_SMALL_ALLOC
 345 void vm_radix_reserve_kva(void);
 346 /*
 347  * Reserve the KVA necessary to satisfy the node allocation.
 348  * This is mandatory in architectures not supporting direct
 349  * mapping as they will need otherwise to carve into the kernel maps for
 350  * every node allocation, resulting into deadlocks for consumers already
 351  * working with kernel maps.
 352  */
 353 void
 354 vm_radix_reserve_kva(void)
 355 {
 356
 357         /*
 358          * Calculate the number of reserved nodes, discounting the pages that
 359          * are needed to store them.
 360          */
 361         if (!uma_zone_reserve_kva(vm_radix_node_zone,
 362             ((vm_paddr_t)vm_cnt.v_page_count * PAGE_SIZE) / (PAGE_SIZE +
 363             sizeof(struct vm_radix_node))))
 364                 panic("%s: unable to reserve KVA", __func__);
 365 }
 366 #endif
 367
 368 /*
 369  * Initialize the UMA slab zone.
 370  */
 371 void
 372 vm_radix_zinit(void)
 373 {
 374
 375         vm_radix_node_zone = uma_zcreate("RADIX NODE",
 376             sizeof(struct vm_radix_node), NULL, NULL, NULL, NULL,
 377             VM_RADIX_PAD, UMA_ZONE_VM | UMA_ZONE_SMR | UMA_ZONE_ZINIT);
 378         vm_radix_smr = uma_zone_get_smr(vm_radix_node_zone);
 379 }
 380
 381 /*
 382  * Inserts the key-value pair into the trie.
 383  * Panics if the key already exists.
 384  */
 385 int
 386 vm_radix_insert(struct vm_radix *rtree, vm_page_t page)
 387 {
 388         vm_pindex_t index, newind;
 389         struct vm_radix_node *rnode, *tmp;
 390         smrnode_t *parentp;
 391         vm_page_t m;
 392         int slot;
 393         uint16_t clev;
 394
 395         index = page->pindex;
 396
 397         /*
 398          * The owner of record for root is not really important because it
 399          * will never be used.
 400          */
 401         rnode = vm_radix_root_load(rtree, LOCKED);
 402         if (rnode == NULL) {
 403                 rtree->rt_root = (uintptr_t)page | VM_RADIX_ISLEAF;
 404                 return (0);
 405         }
 406         parentp = (smrnode_t *)&rtree->rt_root;
 407         for (;;) {
 408                 if (vm_radix_isleaf(rnode)) {
 409                         m = vm_radix_topage(rnode);
 410                         if (m->pindex == index)
 411                                 panic("%s: key %jx is already present",
 412                                     __func__, (uintmax_t)index);
 413                         clev = vm_radix_keydiff(m->pindex, index);
 414                         tmp = vm_radix_node_get(vm_radix_trimkey(index,
 415                             clev + 1), 2, clev);
 416                         if (tmp == NULL)
 417                                 return (ENOMEM);
 418                         /* These writes are not yet visible due to ordering. */
 419                         vm_radix_addpage(tmp, index, clev, page, UNSERIALIZED);
 420                         vm_radix_addpage(tmp, m->pindex, clev, m, UNSERIALIZED);
 421                         /* Synchronize to make leaf visible. */
 422                         vm_radix_node_store(parentp, tmp, LOCKED);
 423                         return (0);
 424                 } else if (vm_radix_keybarr(rnode, index))
 425                         break;
 426                 slot = vm_radix_slot(index, rnode->rn_clev);
 427                 parentp = &rnode->rn_child[slot];
 428                 tmp = vm_radix_node_load(parentp, LOCKED);
 429                 if (tmp == NULL) {
 430                         rnode->rn_count++;
 431                         vm_radix_addpage(rnode, index, rnode->rn_clev, page,
 432                             LOCKED);
 433                         return (0);
 434                 }
 435                 rnode = tmp;
 436         }
 437
 438         /*
 439          * A new node is needed because the right insertion level is reached.
 440          * Setup the new intermediate node and add the 2 children: the
 441          * new object and the older edge.
 442          */
 443         newind = rnode->rn_owner;
 444         clev = vm_radix_keydiff(newind, index);
 445         tmp = vm_radix_node_get(vm_radix_trimkey(index, clev + 1), 2, clev);
 446         if (tmp == NULL)
 447                 return (ENOMEM);
 448         slot = vm_radix_slot(newind, clev);
 449         /* These writes are not yet visible due to ordering. */
 450         vm_radix_addpage(tmp, index, clev, page, UNSERIALIZED);
 451         vm_radix_node_store(&tmp->rn_child[slot], rnode, UNSERIALIZED);
 452         /* Serializing write to make the above visible. */
 453         vm_radix_node_store(parentp, tmp, LOCKED);
 454
 455         return (0);
 456 }
 457
 458 /*
 459  * Returns TRUE if the specified radix tree contains a single leaf and FALSE
 460  * otherwise.
 461  */
 462 boolean_t
 463 vm_radix_is_singleton(struct vm_radix *rtree)
 464 {
 465         struct vm_radix_node *rnode;
 466
 467         rnode = vm_radix_root_load(rtree, LOCKED);
 468         if (rnode == NULL)
 469                 return (FALSE);
 470         return (vm_radix_isleaf(rnode));
 471 }
 472
 473 /*
 474  * Returns the value stored at the index.  If the index is not present,
 475  * NULL is returned.
 476  */
 477 static __always_inline vm_page_t
 478 _vm_radix_lookup(struct vm_radix *rtree, vm_pindex_t index,
 479     enum vm_radix_access access)
 480 {
 481         struct vm_radix_node *rnode;
 482         vm_page_t m;
 483         int slot;
 484
 485         rnode = vm_radix_root_load(rtree, access);
 486         while (rnode != NULL) {
 487                 if (vm_radix_isleaf(rnode)) {
 488                         m = vm_radix_topage(rnode);
 489                         if (m->pindex == index)
 490                                 return (m);
 491                         break;
 492                 }
 493                 if (vm_radix_keybarr(rnode, index))
 494                         break;
 495                 slot = vm_radix_slot(index, rnode->rn_clev);
 496                 rnode = vm_radix_node_load(&rnode->rn_child[slot], access);
 497         }
 498         return (NULL);
 499 }
 500
 501 /*
 502  * Returns the value stored at the index assuming there is an external lock.
 503  *
 504  * If the index is not present, NULL is returned.
 505  */
 506 vm_page_t
 507 vm_radix_lookup(struct vm_radix *rtree, vm_pindex_t index)
 508 {
 509
 510         return _vm_radix_lookup(rtree, index, LOCKED);
 511 }
 512
 513 /*
 514  * Returns the value stored at the index without requiring an external lock.
 515  *
 516  * If the index is not present, NULL is returned.
 517  */
 518 vm_page_t
 519 vm_radix_lookup_unlocked(struct vm_radix *rtree, vm_pindex_t index)
 520 {
 521         vm_page_t m;
 522
 523         smr_enter(vm_radix_smr);
 524         m = _vm_radix_lookup(rtree, index, SMR);
 525         smr_exit(vm_radix_smr);
 526
 527         return (m);
 528 }
 529
 530 /*
 531  * Look up the nearest entry at a position greater than or equal to index.
 532  */
 533 vm_page_t
 534 vm_radix_lookup_ge(struct vm_radix *rtree, vm_pindex_t index)
 535 {
 536         struct vm_radix_node *stack[VM_RADIX_LIMIT];
 537         vm_pindex_t inc;
 538         vm_page_t m;
 539         struct vm_radix_node *child, *rnode;
 540 #ifdef INVARIANTS
 541         int loops = 0;
 542 #endif
 543         int slot, tos;
 544
 545         rnode = vm_radix_root_load(rtree, LOCKED);
 546         if (rnode == NULL)
 547                 return (NULL);
 548         else if (vm_radix_isleaf(rnode)) {
 549                 m = vm_radix_topage(rnode);
 550                 if (m->pindex >= index)
 551                         return (m);
 552                 else
 553                         return (NULL);
 554         }
 555         tos = 0;
 556         for (;;) {
 557                 /*
 558                  * If the keys differ before the current bisection node,
 559                  * then the search key might rollback to the earliest
 560                  * available bisection node or to the smallest key
 561                  * in the current node (if the owner is greater than the
 562                  * search key).
 563                  */
 564                 if (vm_radix_keybarr(rnode, index)) {
 565                         if (index > rnode->rn_owner) {
 566 ascend:
 567                                 KASSERT(++loops < 1000,
 568                                     ("vm_radix_lookup_ge: too many loops"));
 569
 570                                 /*
 571                                  * Pop nodes from the stack until either the
 572                                  * stack is empty or a node that could have a
 573                                  * matching descendant is found.
 574                                  */
 575                                 do {
 576                                         if (tos == 0)
 577                                                 return (NULL);
 578                                         rnode = stack[--tos];
 579                                 } while (vm_radix_slot(index,
 580                                     rnode->rn_clev) == (VM_RADIX_COUNT - 1));
 581
 582                                 /*
 583                                  * The following computation cannot overflow
 584                                  * because index's slot at the current level
 585                                  * is less than VM_RADIX_COUNT - 1.
 586                                  */
 587                                 index = vm_radix_trimkey(index,
 588                                     rnode->rn_clev);
 589                                 index += VM_RADIX_UNITLEVEL(rnode->rn_clev);
 590                         } else
 591                                 index = rnode->rn_owner;
 592                         KASSERT(!vm_radix_keybarr(rnode, index),
 593                             ("vm_radix_lookup_ge: keybarr failed"));
 594                 }
 595                 slot = vm_radix_slot(index, rnode->rn_clev);
 596                 child = vm_radix_node_load(&rnode->rn_child[slot], LOCKED);
 597                 if (vm_radix_isleaf(child)) {
 598                         m = vm_radix_topage(child);
 599                         if (m->pindex >= index)
 600                                 return (m);
 601                 } else if (child != NULL)
 602                         goto descend;
 603
 604                 /*
 605                  * Look for an available edge or page within the current
 606                  * bisection node.
 607                  */
 608                 if (slot < (VM_RADIX_COUNT - 1)) {
 609                         inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
 610                         index = vm_radix_trimkey(index, rnode->rn_clev);
 611                         do {
 612                                 index += inc;
 613                                 slot++;
 614                                 child = vm_radix_node_load(&rnode->rn_child[slot],
 615                                     LOCKED);
 616                                 if (vm_radix_isleaf(child)) {
 617                                         m = vm_radix_topage(child);
 618                                         if (m->pindex >= index)
 619                                                 return (m);
 620                                 } else if (child != NULL)
 621                                         goto descend;
 622                         } while (slot < (VM_RADIX_COUNT - 1));
 623                 }
 624                 KASSERT(child == NULL || vm_radix_isleaf(child),
 625                     ("vm_radix_lookup_ge: child is radix node"));
 626
 627                 /*
 628                  * If a page or edge greater than the search slot is not found
 629                  * in the current node, ascend to the next higher-level node.
 630                  */
 631                 goto ascend;
 632 descend:
 633                 KASSERT(rnode->rn_clev > 0,
 634                     ("vm_radix_lookup_ge: pushing leaf's parent"));
 635                 KASSERT(tos < VM_RADIX_LIMIT,
 636                     ("vm_radix_lookup_ge: stack overflow"));
 637                 stack[tos++] = rnode;
 638                 rnode = child;
 639         }
 640 }
 641
 642 /*
 643  * Look up the nearest entry at a position less than or equal to index.
 644  */
 645 vm_page_t
 646 vm_radix_lookup_le(struct vm_radix *rtree, vm_pindex_t index)
 647 {
 648         struct vm_radix_node *stack[VM_RADIX_LIMIT];
 649         vm_pindex_t inc;
 650         vm_page_t m;
 651         struct vm_radix_node *child, *rnode;
 652 #ifdef INVARIANTS
 653         int loops = 0;
 654 #endif
 655         int slot, tos;
 656
 657         rnode = vm_radix_root_load(rtree, LOCKED);
 658         if (rnode == NULL)
 659                 return (NULL);
 660         else if (vm_radix_isleaf(rnode)) {
 661                 m = vm_radix_topage(rnode);
 662                 if (m->pindex <= index)
 663                         return (m);
 664                 else
 665                         return (NULL);
 666         }
 667         tos = 0;
 668         for (;;) {
 669                 /*
 670                  * If the keys differ before the current bisection node,
 671                  * then the search key might rollback to the earliest
 672                  * available bisection node or to the largest key
 673                  * in the current node (if the owner is smaller than the
 674                  * search key).
 675                  */
 676                 if (vm_radix_keybarr(rnode, index)) {
 677                         if (index > rnode->rn_owner) {
 678                                 index = rnode->rn_owner + VM_RADIX_COUNT *
 679                                     VM_RADIX_UNITLEVEL(rnode->rn_clev);
 680                         } else {
 681 ascend:
 682                                 KASSERT(++loops < 1000,
 683                                     ("vm_radix_lookup_le: too many loops"));
 684
 685                                 /*
 686                                  * Pop nodes from the stack until either the
 687                                  * stack is empty or a node that could have a
 688                                  * matching descendant is found.
 689                                  */
 690                                 do {
 691                                         if (tos == 0)
 692                                                 return (NULL);
 693                                         rnode = stack[--tos];
 694                                 } while (vm_radix_slot(index,
 695                                     rnode->rn_clev) == 0);
 696
 697                                 /*
 698                                  * The following computation cannot overflow
 699                                  * because index's slot at the current level
 700                                  * is greater than 0.
 701                                  */
 702                                 index = vm_radix_trimkey(index,
 703                                     rnode->rn_clev);
 704                         }
 705                         index--;
 706                         KASSERT(!vm_radix_keybarr(rnode, index),
 707                             ("vm_radix_lookup_le: keybarr failed"));
 708                 }
 709                 slot = vm_radix_slot(index, rnode->rn_clev);
 710                 child = vm_radix_node_load(&rnode->rn_child[slot], LOCKED);
 711                 if (vm_radix_isleaf(child)) {
 712                         m = vm_radix_topage(child);
 713                         if (m->pindex <= index)
 714                                 return (m);
 715                 } else if (child != NULL)
 716                         goto descend;
 717
 718                 /*
 719                  * Look for an available edge or page within the current
 720                  * bisection node.
 721                  */
 722                 if (slot > 0) {
 723                         inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
 724                         index |= inc - 1;
 725                         do {
 726                                 index -= inc;
 727                                 slot--;
 728                                 child = vm_radix_node_load(&rnode->rn_child[slot],
 729                                     LOCKED);
 730                                 if (vm_radix_isleaf(child)) {
 731                                         m = vm_radix_topage(child);
 732                                         if (m->pindex <= index)
 733                                                 return (m);
 734                                 } else if (child != NULL)
 735                                         goto descend;
 736                         } while (slot > 0);
 737                 }
 738                 KASSERT(child == NULL || vm_radix_isleaf(child),
 739                     ("vm_radix_lookup_le: child is radix node"));
 740
 741                 /*
 742                  * If a page or edge smaller than the search slot is not found
 743                  * in the current node, ascend to the next higher-level node.
 744                  */
 745                 goto ascend;
 746 descend:
 747                 KASSERT(rnode->rn_clev > 0,
 748                     ("vm_radix_lookup_le: pushing leaf's parent"));
 749                 KASSERT(tos < VM_RADIX_LIMIT,
 750                     ("vm_radix_lookup_le: stack overflow"));
 751                 stack[tos++] = rnode;
 752                 rnode = child;
 753         }
 754 }
 755
 756 /*
 757  * Remove the specified index from the trie, and return the value stored at
 758  * that index.  If the index is not present, return NULL.
 759  */
 760 vm_page_t
 761 vm_radix_remove(struct vm_radix *rtree, vm_pindex_t index)
 762 {
 763         struct vm_radix_node *rnode, *parent, *tmp;
 764         vm_page_t m;
 765         int i, slot;
 766
 767         rnode = vm_radix_root_load(rtree, LOCKED);
 768         if (vm_radix_isleaf(rnode)) {
 769                 m = vm_radix_topage(rnode);
 770                 if (m->pindex != index)
 771                         return (NULL);
 772                 vm_radix_root_store(rtree, NULL, LOCKED);
 773                 return (m);
 774         }
 775         parent = NULL;
 776         for (;;) {
 777                 if (rnode == NULL)
 778                         return (NULL);
 779                 slot = vm_radix_slot(index, rnode->rn_clev);
 780                 tmp = vm_radix_node_load(&rnode->rn_child[slot], LOCKED);
 781                 if (vm_radix_isleaf(tmp)) {
 782                         m = vm_radix_topage(tmp);
 783                         if (m->pindex != index)
 784                                 return (NULL);
 785                         vm_radix_node_store(&rnode->rn_child[slot], NULL, LOCKED);
 786                         rnode->rn_count--;
 787                         if (rnode->rn_count > 1)
 788                                 return (m);
 789                         for (i = 0; i < VM_RADIX_COUNT; i++)
 790                                 if (vm_radix_node_load(&rnode->rn_child[i],
 791                                     LOCKED) != NULL)
 792                                         break;
 793                         KASSERT(i != VM_RADIX_COUNT,
 794                             ("%s: invalid node configuration", __func__));
 795                         tmp = vm_radix_node_load(&rnode->rn_child[i], LOCKED);
 796                         if (parent == NULL)
 797                                 vm_radix_root_store(rtree, tmp, LOCKED);
 798                         else {
 799                                 slot = vm_radix_slot(index, parent->rn_clev);
 800                                 KASSERT(vm_radix_node_load(
 801                                     &parent->rn_child[slot], LOCKED) == rnode,
 802                                     ("%s: invalid child value", __func__));
 803                                 vm_radix_node_store(&parent->rn_child[slot],
 804                                     tmp, LOCKED);
 805                         }
 806                         /*
 807                          * The child is still valid and we can not zero the
 808                          * pointer until all smr references are gone.
 809                          */
 810                         rnode->rn_count--;
 811                         vm_radix_node_put(rnode, i);
 812                         return (m);
 813                 }
 814                 parent = rnode;
 815                 rnode = tmp;
 816         }
 817 }
 818
 819 /*
 820  * Remove and free all the nodes from the radix tree.
 821  * This function is recursive but there is a tight control on it as the
 822  * maximum depth of the tree is fixed.
 823  */
 824 void
 825 vm_radix_reclaim_allnodes(struct vm_radix *rtree)
 826 {
 827         struct vm_radix_node *root;
 828
 829         root = vm_radix_root_load(rtree, LOCKED);
 830         if (root == NULL)
 831                 return;
 832         vm_radix_root_store(rtree, NULL, UNSERIALIZED);
 833         if (!vm_radix_isleaf(root))
 834                 vm_radix_reclaim_allnodes_int(root);
 835 }
 836
 837 /*
 838  * Replace an existing page in the trie with another one.
 839  * Panics if there is not an old page in the trie at the new page's index.
 840  */
 841 vm_page_t
 842 vm_radix_replace(struct vm_radix *rtree, vm_page_t newpage)
 843 {
 844         struct vm_radix_node *rnode, *tmp;
 845         vm_page_t m;
 846         vm_pindex_t index;
 847         int slot;
 848
 849         index = newpage->pindex;
 850         rnode = vm_radix_root_load(rtree, LOCKED);
 851         if (rnode == NULL)
 852                 panic("%s: replacing page on an empty trie", __func__);
 853         if (vm_radix_isleaf(rnode)) {
 854                 m = vm_radix_topage(rnode);
 855                 if (m->pindex != index)
 856                         panic("%s: original replacing root key not found",
 857                             __func__);
 858                 rtree->rt_root = (uintptr_t)newpage | VM_RADIX_ISLEAF;
 859                 return (m);
 860         }
 861         for (;;) {
 862                 slot = vm_radix_slot(index, rnode->rn_clev);
 863                 tmp = vm_radix_node_load(&rnode->rn_child[slot], LOCKED);
 864                 if (vm_radix_isleaf(tmp)) {
 865                         m = vm_radix_topage(tmp);
 866                         if (m->pindex == index) {
 867                                 vm_radix_node_store(&rnode->rn_child[slot],
 868                                     (struct vm_radix_node *)((uintptr_t)newpage |
 869                                     VM_RADIX_ISLEAF), LOCKED);
 870                                 return (m);
 871                         } else
 872                                 break;
 873                 } else if (tmp == NULL || vm_radix_keybarr(tmp, index))
 874                         break;
 875                 rnode = tmp;
 876         }
 877         panic("%s: original replacing page not found", __func__);
 878 }
 879
 880 void
 881 vm_radix_wait(void)
 882 {
 883         uma_zwait(vm_radix_node_zone);
 884 }
 885
 886 #ifdef DDB
 887 /*
 888  * Show details about the given radix node.
 889  */
 890 DB_SHOW_COMMAND(radixnode, db_show_radixnode)
 891 {
 892         struct vm_radix_node *rnode, *tmp;
 893         int i;
 894
 895         if (!have_addr)
 896                 return;
 897         rnode = (struct vm_radix_node *)addr;
 898         db_printf("radixnode %p, owner %jx, children count %u, level %u:\n",
 899             (void *)rnode, (uintmax_t)rnode->rn_owner, rnode->rn_count,
 900             rnode->rn_clev);
 901         for (i = 0; i < VM_RADIX_COUNT; i++) {
 902                 tmp = vm_radix_node_load(&rnode->rn_child[i], UNSERIALIZED);
 903                 if (tmp != NULL)
 904                         db_printf("slot: %d, val: %p, page: %p, clev: %d\n",
 905                             i, (void *)tmp,
 906                             vm_radix_isleaf(tmp) ?  vm_radix_topage(tmp) : NULL,
 907                             rnode->rn_clev);
 908         }
 909 }
 910 #endif /* DDB */