xref: /freebsd/sys/vm/vm_radix.c (revision 5ffd83dbcc34f10e07f6d3e968ae6365869615f4)
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 
221 	switch (access) {
222 	case UNSERIALIZED:
223 		smr_unserialized_store(p, v, true);
224 		break;
225 	case LOCKED:
226 		smr_serialized_store(p, v, true);
227 		break;
228 	case SMR:
229 		panic("vm_radix_node_store: Not supported in smr section.");
230 	}
231 }
232 
233 /*
234  * Get the root node for a radix tree.
235  */
236 static __inline struct vm_radix_node *
237 vm_radix_root_load(struct vm_radix *rtree, enum vm_radix_access access)
238 {
239 
240 	return (vm_radix_node_load((smrnode_t *)&rtree->rt_root, access));
241 }
242 
243 /*
244  * Set the root node for a radix tree.
245  */
246 static __inline void
247 vm_radix_root_store(struct vm_radix *rtree, struct vm_radix_node *rnode,
248     enum vm_radix_access access)
249 {
250 
251 	vm_radix_node_store((smrnode_t *)&rtree->rt_root, rnode, access);
252 }
253 
254 /*
255  * Returns TRUE if the specified radix node is a leaf and FALSE otherwise.
256  */
257 static __inline boolean_t
258 vm_radix_isleaf(struct vm_radix_node *rnode)
259 {
260 
261 	return (((uintptr_t)rnode & VM_RADIX_ISLEAF) != 0);
262 }
263 
264 /*
265  * Returns the associated page extracted from rnode.
266  */
267 static __inline vm_page_t
268 vm_radix_topage(struct vm_radix_node *rnode)
269 {
270 
271 	return ((vm_page_t)((uintptr_t)rnode & ~VM_RADIX_FLAGS));
272 }
273 
274 /*
275  * Adds the page as a child of the provided node.
276  */
277 static __inline void
278 vm_radix_addpage(struct vm_radix_node *rnode, vm_pindex_t index, uint16_t clev,
279     vm_page_t page, enum vm_radix_access access)
280 {
281 	int slot;
282 
283 	slot = vm_radix_slot(index, clev);
284 	vm_radix_node_store(&rnode->rn_child[slot],
285 	    (struct vm_radix_node *)((uintptr_t)page | VM_RADIX_ISLEAF), access);
286 }
287 
288 /*
289  * Returns the slot where two keys differ.
290  * It cannot accept 2 equal keys.
291  */
292 static __inline uint16_t
293 vm_radix_keydiff(vm_pindex_t index1, vm_pindex_t index2)
294 {
295 	uint16_t clev;
296 
297 	KASSERT(index1 != index2, ("%s: passing the same key value %jx",
298 	    __func__, (uintmax_t)index1));
299 
300 	index1 ^= index2;
301 	for (clev = VM_RADIX_LIMIT;; clev--)
302 		if (vm_radix_slot(index1, clev) != 0)
303 			return (clev);
304 }
305 
306 /*
307  * Returns TRUE if it can be determined that key does not belong to the
308  * specified rnode.  Otherwise, returns FALSE.
309  */
310 static __inline boolean_t
311 vm_radix_keybarr(struct vm_radix_node *rnode, vm_pindex_t idx)
312 {
313 
314 	if (rnode->rn_clev < VM_RADIX_LIMIT) {
315 		idx = vm_radix_trimkey(idx, rnode->rn_clev + 1);
316 		return (idx != rnode->rn_owner);
317 	}
318 	return (FALSE);
319 }
320 
321 /*
322  * Internal helper for vm_radix_reclaim_allnodes().
323  * This function is recursive.
324  */
325 static void
326 vm_radix_reclaim_allnodes_int(struct vm_radix_node *rnode)
327 {
328 	struct vm_radix_node *child;
329 	int slot;
330 
331 	KASSERT(rnode->rn_count <= VM_RADIX_COUNT,
332 	    ("vm_radix_reclaim_allnodes_int: bad count in rnode %p", rnode));
333 	for (slot = 0; rnode->rn_count != 0; slot++) {
334 		child = vm_radix_node_load(&rnode->rn_child[slot], UNSERIALIZED);
335 		if (child == NULL)
336 			continue;
337 		if (!vm_radix_isleaf(child))
338 			vm_radix_reclaim_allnodes_int(child);
339 		vm_radix_node_store(&rnode->rn_child[slot], NULL, UNSERIALIZED);
340 		rnode->rn_count--;
341 	}
342 	vm_radix_node_put(rnode, -1);
343 }
344 
345 #ifndef UMA_MD_SMALL_ALLOC
346 void vm_radix_reserve_kva(void);
347 /*
348  * Reserve the KVA necessary to satisfy the node allocation.
349  * This is mandatory in architectures not supporting direct
350  * mapping as they will need otherwise to carve into the kernel maps for
351  * every node allocation, resulting into deadlocks for consumers already
352  * working with kernel maps.
353  */
354 void
355 vm_radix_reserve_kva(void)
356 {
357 
358 	/*
359 	 * Calculate the number of reserved nodes, discounting the pages that
360 	 * are needed to store them.
361 	 */
362 	if (!uma_zone_reserve_kva(vm_radix_node_zone,
363 	    ((vm_paddr_t)vm_cnt.v_page_count * PAGE_SIZE) / (PAGE_SIZE +
364 	    sizeof(struct vm_radix_node))))
365 		panic("%s: unable to reserve KVA", __func__);
366 }
367 #endif
368 
369 /*
370  * Initialize the UMA slab zone.
371  */
372 void
373 vm_radix_zinit(void)
374 {
375 
376 	vm_radix_node_zone = uma_zcreate("RADIX NODE",
377 	    sizeof(struct vm_radix_node), NULL, NULL, NULL, NULL,
378 	    VM_RADIX_PAD, UMA_ZONE_VM | UMA_ZONE_SMR | UMA_ZONE_ZINIT);
379 	vm_radix_smr = uma_zone_get_smr(vm_radix_node_zone);
380 }
381 
382 /*
383  * Inserts the key-value pair into the trie.
384  * Panics if the key already exists.
385  */
386 int
387 vm_radix_insert(struct vm_radix *rtree, vm_page_t page)
388 {
389 	vm_pindex_t index, newind;
390 	struct vm_radix_node *rnode, *tmp;
391 	smrnode_t *parentp;
392 	vm_page_t m;
393 	int slot;
394 	uint16_t clev;
395 
396 	index = page->pindex;
397 
398 	/*
399 	 * The owner of record for root is not really important because it
400 	 * will never be used.
401 	 */
402 	rnode = vm_radix_root_load(rtree, LOCKED);
403 	if (rnode == NULL) {
404 		rtree->rt_root = (uintptr_t)page | VM_RADIX_ISLEAF;
405 		return (0);
406 	}
407 	parentp = (smrnode_t *)&rtree->rt_root;
408 	for (;;) {
409 		if (vm_radix_isleaf(rnode)) {
410 			m = vm_radix_topage(rnode);
411 			if (m->pindex == index)
412 				panic("%s: key %jx is already present",
413 				    __func__, (uintmax_t)index);
414 			clev = vm_radix_keydiff(m->pindex, index);
415 			tmp = vm_radix_node_get(vm_radix_trimkey(index,
416 			    clev + 1), 2, clev);
417 			if (tmp == NULL)
418 				return (ENOMEM);
419 			/* These writes are not yet visible due to ordering. */
420 			vm_radix_addpage(tmp, index, clev, page, UNSERIALIZED);
421 			vm_radix_addpage(tmp, m->pindex, clev, m, UNSERIALIZED);
422 			/* Synchronize to make leaf visible. */
423 			vm_radix_node_store(parentp, tmp, LOCKED);
424 			return (0);
425 		} else if (vm_radix_keybarr(rnode, index))
426 			break;
427 		slot = vm_radix_slot(index, rnode->rn_clev);
428 		parentp = &rnode->rn_child[slot];
429 		tmp = vm_radix_node_load(parentp, LOCKED);
430 		if (tmp == NULL) {
431 			rnode->rn_count++;
432 			vm_radix_addpage(rnode, index, rnode->rn_clev, page,
433 			    LOCKED);
434 			return (0);
435 		}
436 		rnode = tmp;
437 	}
438 
439 	/*
440 	 * A new node is needed because the right insertion level is reached.
441 	 * Setup the new intermediate node and add the 2 children: the
442 	 * new object and the older edge.
443 	 */
444 	newind = rnode->rn_owner;
445 	clev = vm_radix_keydiff(newind, index);
446 	tmp = vm_radix_node_get(vm_radix_trimkey(index, clev + 1), 2, clev);
447 	if (tmp == NULL)
448 		return (ENOMEM);
449 	slot = vm_radix_slot(newind, clev);
450 	/* These writes are not yet visible due to ordering. */
451 	vm_radix_addpage(tmp, index, clev, page, UNSERIALIZED);
452 	vm_radix_node_store(&tmp->rn_child[slot], rnode, UNSERIALIZED);
453 	/* Serializing write to make the above visible. */
454 	vm_radix_node_store(parentp, tmp, LOCKED);
455 
456 	return (0);
457 }
458 
459 /*
460  * Returns TRUE if the specified radix tree contains a single leaf and FALSE
461  * otherwise.
462  */
463 boolean_t
464 vm_radix_is_singleton(struct vm_radix *rtree)
465 {
466 	struct vm_radix_node *rnode;
467 
468 	rnode = vm_radix_root_load(rtree, LOCKED);
469 	if (rnode == NULL)
470 		return (FALSE);
471 	return (vm_radix_isleaf(rnode));
472 }
473 
474 /*
475  * Returns the value stored at the index.  If the index is not present,
476  * NULL is returned.
477  */
478 static __always_inline vm_page_t
479 _vm_radix_lookup(struct vm_radix *rtree, vm_pindex_t index,
480     enum vm_radix_access access)
481 {
482 	struct vm_radix_node *rnode;
483 	vm_page_t m;
484 	int slot;
485 
486 	rnode = vm_radix_root_load(rtree, access);
487 	while (rnode != NULL) {
488 		if (vm_radix_isleaf(rnode)) {
489 			m = vm_radix_topage(rnode);
490 			if (m->pindex == index)
491 				return (m);
492 			break;
493 		}
494 		if (vm_radix_keybarr(rnode, index))
495 			break;
496 		slot = vm_radix_slot(index, rnode->rn_clev);
497 		rnode = vm_radix_node_load(&rnode->rn_child[slot], access);
498 	}
499 	return (NULL);
500 }
501 
502 /*
503  * Returns the value stored at the index assuming there is an external lock.
504  *
505  * If the index is not present, NULL is returned.
506  */
507 vm_page_t
508 vm_radix_lookup(struct vm_radix *rtree, vm_pindex_t index)
509 {
510 
511 	return _vm_radix_lookup(rtree, index, LOCKED);
512 }
513 
514 /*
515  * Returns the value stored at the index without requiring an external lock.
516  *
517  * If the index is not present, NULL is returned.
518  */
519 vm_page_t
520 vm_radix_lookup_unlocked(struct vm_radix *rtree, vm_pindex_t index)
521 {
522 	vm_page_t m;
523 
524 	smr_enter(vm_radix_smr);
525 	m = _vm_radix_lookup(rtree, index, SMR);
526 	smr_exit(vm_radix_smr);
527 
528 	return (m);
529 }
530 
531 /*
532  * Look up the nearest entry at a position greater than or equal to index.
533  */
534 vm_page_t
535 vm_radix_lookup_ge(struct vm_radix *rtree, vm_pindex_t index)
536 {
537 	struct vm_radix_node *stack[VM_RADIX_LIMIT];
538 	vm_pindex_t inc;
539 	vm_page_t m;
540 	struct vm_radix_node *child, *rnode;
541 #ifdef INVARIANTS
542 	int loops = 0;
543 #endif
544 	int slot, tos;
545 
546 	rnode = vm_radix_root_load(rtree, LOCKED);
547 	if (rnode == NULL)
548 		return (NULL);
549 	else if (vm_radix_isleaf(rnode)) {
550 		m = vm_radix_topage(rnode);
551 		if (m->pindex >= index)
552 			return (m);
553 		else
554 			return (NULL);
555 	}
556 	tos = 0;
557 	for (;;) {
558 		/*
559 		 * If the keys differ before the current bisection node,
560 		 * then the search key might rollback to the earliest
561 		 * available bisection node or to the smallest key
562 		 * in the current node (if the owner is greater than the
563 		 * search key).
564 		 */
565 		if (vm_radix_keybarr(rnode, index)) {
566 			if (index > rnode->rn_owner) {
567 ascend:
568 				KASSERT(++loops < 1000,
569 				    ("vm_radix_lookup_ge: too many loops"));
570 
571 				/*
572 				 * Pop nodes from the stack until either the
573 				 * stack is empty or a node that could have a
574 				 * matching descendant is found.
575 				 */
576 				do {
577 					if (tos == 0)
578 						return (NULL);
579 					rnode = stack[--tos];
580 				} while (vm_radix_slot(index,
581 				    rnode->rn_clev) == (VM_RADIX_COUNT - 1));
582 
583 				/*
584 				 * The following computation cannot overflow
585 				 * because index's slot at the current level
586 				 * is less than VM_RADIX_COUNT - 1.
587 				 */
588 				index = vm_radix_trimkey(index,
589 				    rnode->rn_clev);
590 				index += VM_RADIX_UNITLEVEL(rnode->rn_clev);
591 			} else
592 				index = rnode->rn_owner;
593 			KASSERT(!vm_radix_keybarr(rnode, index),
594 			    ("vm_radix_lookup_ge: keybarr failed"));
595 		}
596 		slot = vm_radix_slot(index, rnode->rn_clev);
597 		child = vm_radix_node_load(&rnode->rn_child[slot], LOCKED);
598 		if (vm_radix_isleaf(child)) {
599 			m = vm_radix_topage(child);
600 			if (m->pindex >= index)
601 				return (m);
602 		} else if (child != NULL)
603 			goto descend;
604 
605 		/*
606 		 * Look for an available edge or page within the current
607 		 * bisection node.
608 		 */
609                 if (slot < (VM_RADIX_COUNT - 1)) {
610 			inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
611 			index = vm_radix_trimkey(index, rnode->rn_clev);
612 			do {
613 				index += inc;
614 				slot++;
615 				child = vm_radix_node_load(&rnode->rn_child[slot],
616 				    LOCKED);
617 				if (vm_radix_isleaf(child)) {
618 					m = vm_radix_topage(child);
619 					if (m->pindex >= index)
620 						return (m);
621 				} else if (child != NULL)
622 					goto descend;
623 			} while (slot < (VM_RADIX_COUNT - 1));
624 		}
625 		KASSERT(child == NULL || vm_radix_isleaf(child),
626 		    ("vm_radix_lookup_ge: child is radix node"));
627 
628 		/*
629 		 * If a page or edge greater than the search slot is not found
630 		 * in the current node, ascend to the next higher-level node.
631 		 */
632 		goto ascend;
633 descend:
634 		KASSERT(rnode->rn_clev > 0,
635 		    ("vm_radix_lookup_ge: pushing leaf's parent"));
636 		KASSERT(tos < VM_RADIX_LIMIT,
637 		    ("vm_radix_lookup_ge: stack overflow"));
638 		stack[tos++] = rnode;
639 		rnode = child;
640 	}
641 }
642 
643 /*
644  * Look up the nearest entry at a position less than or equal to index.
645  */
646 vm_page_t
647 vm_radix_lookup_le(struct vm_radix *rtree, vm_pindex_t index)
648 {
649 	struct vm_radix_node *stack[VM_RADIX_LIMIT];
650 	vm_pindex_t inc;
651 	vm_page_t m;
652 	struct vm_radix_node *child, *rnode;
653 #ifdef INVARIANTS
654 	int loops = 0;
655 #endif
656 	int slot, tos;
657 
658 	rnode = vm_radix_root_load(rtree, LOCKED);
659 	if (rnode == NULL)
660 		return (NULL);
661 	else if (vm_radix_isleaf(rnode)) {
662 		m = vm_radix_topage(rnode);
663 		if (m->pindex <= index)
664 			return (m);
665 		else
666 			return (NULL);
667 	}
668 	tos = 0;
669 	for (;;) {
670 		/*
671 		 * If the keys differ before the current bisection node,
672 		 * then the search key might rollback to the earliest
673 		 * available bisection node or to the largest key
674 		 * in the current node (if the owner is smaller than the
675 		 * search key).
676 		 */
677 		if (vm_radix_keybarr(rnode, index)) {
678 			if (index > rnode->rn_owner) {
679 				index = rnode->rn_owner + VM_RADIX_COUNT *
680 				    VM_RADIX_UNITLEVEL(rnode->rn_clev);
681 			} else {
682 ascend:
683 				KASSERT(++loops < 1000,
684 				    ("vm_radix_lookup_le: too many loops"));
685 
686 				/*
687 				 * Pop nodes from the stack until either the
688 				 * stack is empty or a node that could have a
689 				 * matching descendant is found.
690 				 */
691 				do {
692 					if (tos == 0)
693 						return (NULL);
694 					rnode = stack[--tos];
695 				} while (vm_radix_slot(index,
696 				    rnode->rn_clev) == 0);
697 
698 				/*
699 				 * The following computation cannot overflow
700 				 * because index's slot at the current level
701 				 * is greater than 0.
702 				 */
703 				index = vm_radix_trimkey(index,
704 				    rnode->rn_clev);
705 			}
706 			index--;
707 			KASSERT(!vm_radix_keybarr(rnode, index),
708 			    ("vm_radix_lookup_le: keybarr failed"));
709 		}
710 		slot = vm_radix_slot(index, rnode->rn_clev);
711 		child = vm_radix_node_load(&rnode->rn_child[slot], LOCKED);
712 		if (vm_radix_isleaf(child)) {
713 			m = vm_radix_topage(child);
714 			if (m->pindex <= index)
715 				return (m);
716 		} else if (child != NULL)
717 			goto descend;
718 
719 		/*
720 		 * Look for an available edge or page within the current
721 		 * bisection node.
722 		 */
723 		if (slot > 0) {
724 			inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
725 			index |= inc - 1;
726 			do {
727 				index -= inc;
728 				slot--;
729 				child = vm_radix_node_load(&rnode->rn_child[slot],
730 				    LOCKED);
731 				if (vm_radix_isleaf(child)) {
732 					m = vm_radix_topage(child);
733 					if (m->pindex <= index)
734 						return (m);
735 				} else if (child != NULL)
736 					goto descend;
737 			} while (slot > 0);
738 		}
739 		KASSERT(child == NULL || vm_radix_isleaf(child),
740 		    ("vm_radix_lookup_le: child is radix node"));
741 
742 		/*
743 		 * If a page or edge smaller than the search slot is not found
744 		 * in the current node, ascend to the next higher-level node.
745 		 */
746 		goto ascend;
747 descend:
748 		KASSERT(rnode->rn_clev > 0,
749 		    ("vm_radix_lookup_le: pushing leaf's parent"));
750 		KASSERT(tos < VM_RADIX_LIMIT,
751 		    ("vm_radix_lookup_le: stack overflow"));
752 		stack[tos++] = rnode;
753 		rnode = child;
754 	}
755 }
756 
757 /*
758  * Remove the specified index from the trie, and return the value stored at
759  * that index.  If the index is not present, return NULL.
760  */
761 vm_page_t
762 vm_radix_remove(struct vm_radix *rtree, vm_pindex_t index)
763 {
764 	struct vm_radix_node *rnode, *parent, *tmp;
765 	vm_page_t m;
766 	int i, slot;
767 
768 	rnode = vm_radix_root_load(rtree, LOCKED);
769 	if (vm_radix_isleaf(rnode)) {
770 		m = vm_radix_topage(rnode);
771 		if (m->pindex != index)
772 			return (NULL);
773 		vm_radix_root_store(rtree, NULL, LOCKED);
774 		return (m);
775 	}
776 	parent = NULL;
777 	for (;;) {
778 		if (rnode == NULL)
779 			return (NULL);
780 		slot = vm_radix_slot(index, rnode->rn_clev);
781 		tmp = vm_radix_node_load(&rnode->rn_child[slot], LOCKED);
782 		if (vm_radix_isleaf(tmp)) {
783 			m = vm_radix_topage(tmp);
784 			if (m->pindex != index)
785 				return (NULL);
786 			vm_radix_node_store(&rnode->rn_child[slot], NULL, LOCKED);
787 			rnode->rn_count--;
788 			if (rnode->rn_count > 1)
789 				return (m);
790 			for (i = 0; i < VM_RADIX_COUNT; i++)
791 				if (vm_radix_node_load(&rnode->rn_child[i],
792 				    LOCKED) != NULL)
793 					break;
794 			KASSERT(i != VM_RADIX_COUNT,
795 			    ("%s: invalid node configuration", __func__));
796 			tmp = vm_radix_node_load(&rnode->rn_child[i], LOCKED);
797 			if (parent == NULL)
798 				vm_radix_root_store(rtree, tmp, LOCKED);
799 			else {
800 				slot = vm_radix_slot(index, parent->rn_clev);
801 				KASSERT(vm_radix_node_load(
802 				    &parent->rn_child[slot], LOCKED) == rnode,
803 				    ("%s: invalid child value", __func__));
804 				vm_radix_node_store(&parent->rn_child[slot],
805 				    tmp, LOCKED);
806 			}
807 			/*
808 			 * The child is still valid and we can not zero the
809 			 * pointer until all smr references are gone.
810 			 */
811 			rnode->rn_count--;
812 			vm_radix_node_put(rnode, i);
813 			return (m);
814 		}
815 		parent = rnode;
816 		rnode = tmp;
817 	}
818 }
819 
820 /*
821  * Remove and free all the nodes from the radix tree.
822  * This function is recursive but there is a tight control on it as the
823  * maximum depth of the tree is fixed.
824  */
825 void
826 vm_radix_reclaim_allnodes(struct vm_radix *rtree)
827 {
828 	struct vm_radix_node *root;
829 
830 	root = vm_radix_root_load(rtree, LOCKED);
831 	if (root == NULL)
832 		return;
833 	vm_radix_root_store(rtree, NULL, UNSERIALIZED);
834 	if (!vm_radix_isleaf(root))
835 		vm_radix_reclaim_allnodes_int(root);
836 }
837 
838 /*
839  * Replace an existing page in the trie with another one.
840  * Panics if there is not an old page in the trie at the new page's index.
841  */
842 vm_page_t
843 vm_radix_replace(struct vm_radix *rtree, vm_page_t newpage)
844 {
845 	struct vm_radix_node *rnode, *tmp;
846 	vm_page_t m;
847 	vm_pindex_t index;
848 	int slot;
849 
850 	index = newpage->pindex;
851 	rnode = vm_radix_root_load(rtree, LOCKED);
852 	if (rnode == NULL)
853 		panic("%s: replacing page on an empty trie", __func__);
854 	if (vm_radix_isleaf(rnode)) {
855 		m = vm_radix_topage(rnode);
856 		if (m->pindex != index)
857 			panic("%s: original replacing root key not found",
858 			    __func__);
859 		rtree->rt_root = (uintptr_t)newpage | VM_RADIX_ISLEAF;
860 		return (m);
861 	}
862 	for (;;) {
863 		slot = vm_radix_slot(index, rnode->rn_clev);
864 		tmp = vm_radix_node_load(&rnode->rn_child[slot], LOCKED);
865 		if (vm_radix_isleaf(tmp)) {
866 			m = vm_radix_topage(tmp);
867 			if (m->pindex == index) {
868 				vm_radix_node_store(&rnode->rn_child[slot],
869 				    (struct vm_radix_node *)((uintptr_t)newpage |
870 				    VM_RADIX_ISLEAF), LOCKED);
871 				return (m);
872 			} else
873 				break;
874 		} else if (tmp == NULL || vm_radix_keybarr(tmp, index))
875 			break;
876 		rnode = tmp;
877 	}
878 	panic("%s: original replacing page not found", __func__);
879 }
880 
881 void
882 vm_radix_wait(void)
883 {
884 	uma_zwait(vm_radix_node_zone);
885 }
886 
887 #ifdef DDB
888 /*
889  * Show details about the given radix node.
890  */
891 DB_SHOW_COMMAND(radixnode, db_show_radixnode)
892 {
893 	struct vm_radix_node *rnode, *tmp;
894 	int i;
895 
896         if (!have_addr)
897                 return;
898 	rnode = (struct vm_radix_node *)addr;
899 	db_printf("radixnode %p, owner %jx, children count %u, level %u:\n",
900 	    (void *)rnode, (uintmax_t)rnode->rn_owner, rnode->rn_count,
901 	    rnode->rn_clev);
902 	for (i = 0; i < VM_RADIX_COUNT; i++) {
903 		tmp = vm_radix_node_load(&rnode->rn_child[i], UNSERIALIZED);
904 		if (tmp != NULL)
905 			db_printf("slot: %d, val: %p, page: %p, clev: %d\n",
906 			    i, (void *)tmp,
907 			    vm_radix_isleaf(tmp) ?  vm_radix_topage(tmp) : NULL,
908 			    rnode->rn_clev);
909 	}
910 }
911 #endif /* DDB */
912