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