xref: /freebsd/sys/vm/vm_radix.c (revision 98fc0505155575df4170c863107153df2a751a13)
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/vmmeter.h>
62 
63 #include <vm/uma.h>
64 #include <vm/vm.h>
65 #include <vm/vm_param.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_radix.h>
68 
69 #ifdef DDB
70 #include <ddb/ddb.h>
71 #endif
72 
73 /*
74  * These widths should allow the pointers to a node's children to fit within
75  * a single cache line.  The extra levels from a narrow width should not be
76  * a problem thanks to path compression.
77  */
78 #ifdef __LP64__
79 #define	VM_RADIX_WIDTH	4
80 #else
81 #define	VM_RADIX_WIDTH	3
82 #endif
83 
84 #define	VM_RADIX_COUNT	(1 << VM_RADIX_WIDTH)
85 #define	VM_RADIX_MASK	(VM_RADIX_COUNT - 1)
86 #define	VM_RADIX_LIMIT							\
87 	(howmany(sizeof(vm_pindex_t) * NBBY, VM_RADIX_WIDTH) - 1)
88 
89 /* Flag bits stored in node pointers. */
90 #define	VM_RADIX_ISLEAF	0x1
91 #define	VM_RADIX_FLAGS	0x1
92 #define	VM_RADIX_PAD	VM_RADIX_FLAGS
93 
94 /* Returns one unit associated with specified level. */
95 #define	VM_RADIX_UNITLEVEL(lev)						\
96 	((vm_pindex_t)1 << ((lev) * VM_RADIX_WIDTH))
97 
98 struct vm_radix_node {
99 	vm_pindex_t	 rn_owner;			/* Owner of record. */
100 	uint16_t	 rn_count;			/* Valid children. */
101 	uint16_t	 rn_clev;			/* Current level. */
102 	void		*rn_child[VM_RADIX_COUNT];	/* Child nodes. */
103 };
104 
105 static uma_zone_t vm_radix_node_zone;
106 
107 /*
108  * Allocate a radix node.
109  */
110 static __inline struct vm_radix_node *
111 vm_radix_node_get(vm_pindex_t owner, uint16_t count, uint16_t clevel)
112 {
113 	struct vm_radix_node *rnode;
114 
115 	rnode = uma_zalloc(vm_radix_node_zone, M_NOWAIT);
116 	if (rnode == NULL)
117 		return (NULL);
118 	rnode->rn_owner = owner;
119 	rnode->rn_count = count;
120 	rnode->rn_clev = clevel;
121 	return (rnode);
122 }
123 
124 /*
125  * Free radix node.
126  */
127 static __inline void
128 vm_radix_node_put(struct vm_radix_node *rnode)
129 {
130 
131 	uma_zfree(vm_radix_node_zone, rnode);
132 }
133 
134 /*
135  * Return the position in the array for a given level.
136  */
137 static __inline int
138 vm_radix_slot(vm_pindex_t index, uint16_t level)
139 {
140 
141 	return ((index >> (level * VM_RADIX_WIDTH)) & VM_RADIX_MASK);
142 }
143 
144 /* Trims the key after the specified level. */
145 static __inline vm_pindex_t
146 vm_radix_trimkey(vm_pindex_t index, uint16_t level)
147 {
148 	vm_pindex_t ret;
149 
150 	ret = index;
151 	if (level > 0) {
152 		ret >>= level * VM_RADIX_WIDTH;
153 		ret <<= level * VM_RADIX_WIDTH;
154 	}
155 	return (ret);
156 }
157 
158 /*
159  * Get the root node for a radix tree.
160  */
161 static __inline struct vm_radix_node *
162 vm_radix_getroot(struct vm_radix *rtree)
163 {
164 
165 	return ((struct vm_radix_node *)rtree->rt_root);
166 }
167 
168 /*
169  * Set the root node for a radix tree.
170  */
171 static __inline void
172 vm_radix_setroot(struct vm_radix *rtree, struct vm_radix_node *rnode)
173 {
174 
175 	rtree->rt_root = (uintptr_t)rnode;
176 }
177 
178 /*
179  * Returns TRUE if the specified radix node is a leaf and FALSE otherwise.
180  */
181 static __inline boolean_t
182 vm_radix_isleaf(struct vm_radix_node *rnode)
183 {
184 
185 	return (((uintptr_t)rnode & VM_RADIX_ISLEAF) != 0);
186 }
187 
188 /*
189  * Returns the associated page extracted from rnode.
190  */
191 static __inline vm_page_t
192 vm_radix_topage(struct vm_radix_node *rnode)
193 {
194 
195 	return ((vm_page_t)((uintptr_t)rnode & ~VM_RADIX_FLAGS));
196 }
197 
198 /*
199  * Adds the page as a child of the provided node.
200  */
201 static __inline void
202 vm_radix_addpage(struct vm_radix_node *rnode, vm_pindex_t index, uint16_t clev,
203     vm_page_t page)
204 {
205 	int slot;
206 
207 	slot = vm_radix_slot(index, clev);
208 	rnode->rn_child[slot] = (void *)((uintptr_t)page | VM_RADIX_ISLEAF);
209 }
210 
211 /*
212  * Returns the slot where two keys differ.
213  * It cannot accept 2 equal keys.
214  */
215 static __inline uint16_t
216 vm_radix_keydiff(vm_pindex_t index1, vm_pindex_t index2)
217 {
218 	uint16_t clev;
219 
220 	KASSERT(index1 != index2, ("%s: passing the same key value %jx",
221 	    __func__, (uintmax_t)index1));
222 
223 	index1 ^= index2;
224 	for (clev = VM_RADIX_LIMIT;; clev--)
225 		if (vm_radix_slot(index1, clev) != 0)
226 			return (clev);
227 }
228 
229 /*
230  * Returns TRUE if it can be determined that key does not belong to the
231  * specified rnode.  Otherwise, returns FALSE.
232  */
233 static __inline boolean_t
234 vm_radix_keybarr(struct vm_radix_node *rnode, vm_pindex_t idx)
235 {
236 
237 	if (rnode->rn_clev < VM_RADIX_LIMIT) {
238 		idx = vm_radix_trimkey(idx, rnode->rn_clev + 1);
239 		return (idx != rnode->rn_owner);
240 	}
241 	return (FALSE);
242 }
243 
244 /*
245  * Internal helper for vm_radix_reclaim_allnodes().
246  * This function is recursive.
247  */
248 static void
249 vm_radix_reclaim_allnodes_int(struct vm_radix_node *rnode)
250 {
251 	int slot;
252 
253 	KASSERT(rnode->rn_count <= VM_RADIX_COUNT,
254 	    ("vm_radix_reclaim_allnodes_int: bad count in rnode %p", rnode));
255 	for (slot = 0; rnode->rn_count != 0; slot++) {
256 		if (rnode->rn_child[slot] == NULL)
257 			continue;
258 		if (!vm_radix_isleaf(rnode->rn_child[slot]))
259 			vm_radix_reclaim_allnodes_int(rnode->rn_child[slot]);
260 		rnode->rn_child[slot] = NULL;
261 		rnode->rn_count--;
262 	}
263 	vm_radix_node_put(rnode);
264 }
265 
266 #ifdef INVARIANTS
267 /*
268  * Radix node zone destructor.
269  */
270 static void
271 vm_radix_node_zone_dtor(void *mem, int size __unused, void *arg __unused)
272 {
273 	struct vm_radix_node *rnode;
274 	int slot;
275 
276 	rnode = mem;
277 	KASSERT(rnode->rn_count == 0,
278 	    ("vm_radix_node_put: rnode %p has %d children", rnode,
279 	    rnode->rn_count));
280 	for (slot = 0; slot < VM_RADIX_COUNT; slot++)
281 		KASSERT(rnode->rn_child[slot] == NULL,
282 		    ("vm_radix_node_put: rnode %p has a child", rnode));
283 }
284 #endif
285 
286 static int
287 vm_radix_node_zone_init(void *mem, int size __unused, int flags __unused)
288 {
289 	struct vm_radix_node *rnode;
290 
291 	rnode = mem;
292 	bzero(rnode, sizeof(*rnode));
293 	return (0);
294 }
295 
296 #ifndef UMA_MD_SMALL_ALLOC
297 void vm_radix_reserve_kva(void);
298 /*
299  * Reserve the KVA necessary to satisfy the node allocation.
300  * This is mandatory in architectures not supporting direct
301  * mapping as they will need otherwise to carve into the kernel maps for
302  * every node allocation, resulting into deadlocks for consumers already
303  * working with kernel maps.
304  */
305 void
306 vm_radix_reserve_kva(void)
307 {
308 
309 	/*
310 	 * Calculate the number of reserved nodes, discounting the pages that
311 	 * are needed to store them.
312 	 */
313 	if (!uma_zone_reserve_kva(vm_radix_node_zone,
314 	    ((vm_paddr_t)vm_cnt.v_page_count * PAGE_SIZE) / (PAGE_SIZE +
315 	    sizeof(struct vm_radix_node))))
316 		panic("%s: unable to reserve KVA", __func__);
317 }
318 #endif
319 
320 /*
321  * Initialize the UMA slab zone.
322  */
323 void
324 vm_radix_zinit(void)
325 {
326 
327 	vm_radix_node_zone = uma_zcreate("RADIX NODE",
328 	    sizeof(struct vm_radix_node), NULL,
329 #ifdef INVARIANTS
330 	    vm_radix_node_zone_dtor,
331 #else
332 	    NULL,
333 #endif
334 	    vm_radix_node_zone_init, NULL, VM_RADIX_PAD, UMA_ZONE_VM);
335 }
336 
337 /*
338  * Inserts the key-value pair into the trie.
339  * Panics if the key already exists.
340  */
341 int
342 vm_radix_insert(struct vm_radix *rtree, vm_page_t page)
343 {
344 	vm_pindex_t index, newind;
345 	void **parentp;
346 	struct vm_radix_node *rnode, *tmp;
347 	vm_page_t m;
348 	int slot;
349 	uint16_t clev;
350 
351 	index = page->pindex;
352 
353 	/*
354 	 * The owner of record for root is not really important because it
355 	 * will never be used.
356 	 */
357 	rnode = vm_radix_getroot(rtree);
358 	if (rnode == NULL) {
359 		rtree->rt_root = (uintptr_t)page | VM_RADIX_ISLEAF;
360 		return (0);
361 	}
362 	parentp = (void **)&rtree->rt_root;
363 	for (;;) {
364 		if (vm_radix_isleaf(rnode)) {
365 			m = vm_radix_topage(rnode);
366 			if (m->pindex == index)
367 				panic("%s: key %jx is already present",
368 				    __func__, (uintmax_t)index);
369 			clev = vm_radix_keydiff(m->pindex, index);
370 			tmp = vm_radix_node_get(vm_radix_trimkey(index,
371 			    clev + 1), 2, clev);
372 			if (tmp == NULL)
373 				return (ENOMEM);
374 			*parentp = tmp;
375 			vm_radix_addpage(tmp, index, clev, page);
376 			vm_radix_addpage(tmp, m->pindex, clev, m);
377 			return (0);
378 		} else if (vm_radix_keybarr(rnode, index))
379 			break;
380 		slot = vm_radix_slot(index, rnode->rn_clev);
381 		if (rnode->rn_child[slot] == NULL) {
382 			rnode->rn_count++;
383 			vm_radix_addpage(rnode, index, rnode->rn_clev, page);
384 			return (0);
385 		}
386 		parentp = &rnode->rn_child[slot];
387 		rnode = rnode->rn_child[slot];
388 	}
389 
390 	/*
391 	 * A new node is needed because the right insertion level is reached.
392 	 * Setup the new intermediate node and add the 2 children: the
393 	 * new object and the older edge.
394 	 */
395 	newind = rnode->rn_owner;
396 	clev = vm_radix_keydiff(newind, index);
397 	tmp = vm_radix_node_get(vm_radix_trimkey(index, clev + 1), 2, clev);
398 	if (tmp == NULL)
399 		return (ENOMEM);
400 	*parentp = tmp;
401 	vm_radix_addpage(tmp, index, clev, page);
402 	slot = vm_radix_slot(newind, clev);
403 	tmp->rn_child[slot] = rnode;
404 	return (0);
405 }
406 
407 /*
408  * Returns TRUE if the specified radix tree contains a single leaf and FALSE
409  * otherwise.
410  */
411 boolean_t
412 vm_radix_is_singleton(struct vm_radix *rtree)
413 {
414 	struct vm_radix_node *rnode;
415 
416 	rnode = vm_radix_getroot(rtree);
417 	if (rnode == NULL)
418 		return (FALSE);
419 	return (vm_radix_isleaf(rnode));
420 }
421 
422 /*
423  * Returns the value stored at the index.  If the index is not present,
424  * NULL is returned.
425  */
426 vm_page_t
427 vm_radix_lookup(struct vm_radix *rtree, vm_pindex_t index)
428 {
429 	struct vm_radix_node *rnode;
430 	vm_page_t m;
431 	int slot;
432 
433 	rnode = vm_radix_getroot(rtree);
434 	while (rnode != NULL) {
435 		if (vm_radix_isleaf(rnode)) {
436 			m = vm_radix_topage(rnode);
437 			if (m->pindex == index)
438 				return (m);
439 			else
440 				break;
441 		} else if (vm_radix_keybarr(rnode, index))
442 			break;
443 		slot = vm_radix_slot(index, rnode->rn_clev);
444 		rnode = rnode->rn_child[slot];
445 	}
446 	return (NULL);
447 }
448 
449 /*
450  * Look up the nearest entry at a position bigger than or equal to index.
451  */
452 vm_page_t
453 vm_radix_lookup_ge(struct vm_radix *rtree, vm_pindex_t index)
454 {
455 	struct vm_radix_node *stack[VM_RADIX_LIMIT];
456 	vm_pindex_t inc;
457 	vm_page_t m;
458 	struct vm_radix_node *child, *rnode;
459 #ifdef INVARIANTS
460 	int loops = 0;
461 #endif
462 	int slot, tos;
463 
464 	rnode = vm_radix_getroot(rtree);
465 	if (rnode == NULL)
466 		return (NULL);
467 	else if (vm_radix_isleaf(rnode)) {
468 		m = vm_radix_topage(rnode);
469 		if (m->pindex >= index)
470 			return (m);
471 		else
472 			return (NULL);
473 	}
474 	tos = 0;
475 	for (;;) {
476 		/*
477 		 * If the keys differ before the current bisection node,
478 		 * then the search key might rollback to the earliest
479 		 * available bisection node or to the smallest key
480 		 * in the current node (if the owner is bigger than the
481 		 * search key).
482 		 */
483 		if (vm_radix_keybarr(rnode, index)) {
484 			if (index > rnode->rn_owner) {
485 ascend:
486 				KASSERT(++loops < 1000,
487 				    ("vm_radix_lookup_ge: too many loops"));
488 
489 				/*
490 				 * Pop nodes from the stack until either the
491 				 * stack is empty or a node that could have a
492 				 * matching descendant is found.
493 				 */
494 				do {
495 					if (tos == 0)
496 						return (NULL);
497 					rnode = stack[--tos];
498 				} while (vm_radix_slot(index,
499 				    rnode->rn_clev) == (VM_RADIX_COUNT - 1));
500 
501 				/*
502 				 * The following computation cannot overflow
503 				 * because index's slot at the current level
504 				 * is less than VM_RADIX_COUNT - 1.
505 				 */
506 				index = vm_radix_trimkey(index,
507 				    rnode->rn_clev);
508 				index += VM_RADIX_UNITLEVEL(rnode->rn_clev);
509 			} else
510 				index = rnode->rn_owner;
511 			KASSERT(!vm_radix_keybarr(rnode, index),
512 			    ("vm_radix_lookup_ge: keybarr failed"));
513 		}
514 		slot = vm_radix_slot(index, rnode->rn_clev);
515 		child = rnode->rn_child[slot];
516 		if (vm_radix_isleaf(child)) {
517 			m = vm_radix_topage(child);
518 			if (m->pindex >= index)
519 				return (m);
520 		} else if (child != NULL)
521 			goto descend;
522 
523 		/*
524 		 * Look for an available edge or page within the current
525 		 * bisection node.
526 		 */
527                 if (slot < (VM_RADIX_COUNT - 1)) {
528 			inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
529 			index = vm_radix_trimkey(index, rnode->rn_clev);
530 			do {
531 				index += inc;
532 				slot++;
533 				child = rnode->rn_child[slot];
534 				if (vm_radix_isleaf(child)) {
535 					m = vm_radix_topage(child);
536 					if (m->pindex >= index)
537 						return (m);
538 				} else if (child != NULL)
539 					goto descend;
540 			} while (slot < (VM_RADIX_COUNT - 1));
541 		}
542 		KASSERT(child == NULL || vm_radix_isleaf(child),
543 		    ("vm_radix_lookup_ge: child is radix node"));
544 
545 		/*
546 		 * If a page or edge bigger than the search slot is not found
547 		 * in the current node, ascend to the next higher-level node.
548 		 */
549 		goto ascend;
550 descend:
551 		KASSERT(rnode->rn_clev > 0,
552 		    ("vm_radix_lookup_ge: pushing leaf's parent"));
553 		KASSERT(tos < VM_RADIX_LIMIT,
554 		    ("vm_radix_lookup_ge: stack overflow"));
555 		stack[tos++] = rnode;
556 		rnode = child;
557 	}
558 }
559 
560 /*
561  * Look up the nearest entry at a position less than or equal to index.
562  */
563 vm_page_t
564 vm_radix_lookup_le(struct vm_radix *rtree, vm_pindex_t index)
565 {
566 	struct vm_radix_node *stack[VM_RADIX_LIMIT];
567 	vm_pindex_t inc;
568 	vm_page_t m;
569 	struct vm_radix_node *child, *rnode;
570 #ifdef INVARIANTS
571 	int loops = 0;
572 #endif
573 	int slot, tos;
574 
575 	rnode = vm_radix_getroot(rtree);
576 	if (rnode == NULL)
577 		return (NULL);
578 	else if (vm_radix_isleaf(rnode)) {
579 		m = vm_radix_topage(rnode);
580 		if (m->pindex <= index)
581 			return (m);
582 		else
583 			return (NULL);
584 	}
585 	tos = 0;
586 	for (;;) {
587 		/*
588 		 * If the keys differ before the current bisection node,
589 		 * then the search key might rollback to the earliest
590 		 * available bisection node or to the largest key
591 		 * in the current node (if the owner is smaller than the
592 		 * search key).
593 		 */
594 		if (vm_radix_keybarr(rnode, index)) {
595 			if (index > rnode->rn_owner) {
596 				index = rnode->rn_owner + VM_RADIX_COUNT *
597 				    VM_RADIX_UNITLEVEL(rnode->rn_clev);
598 			} else {
599 ascend:
600 				KASSERT(++loops < 1000,
601 				    ("vm_radix_lookup_le: too many loops"));
602 
603 				/*
604 				 * Pop nodes from the stack until either the
605 				 * stack is empty or a node that could have a
606 				 * matching descendant is found.
607 				 */
608 				do {
609 					if (tos == 0)
610 						return (NULL);
611 					rnode = stack[--tos];
612 				} while (vm_radix_slot(index,
613 				    rnode->rn_clev) == 0);
614 
615 				/*
616 				 * The following computation cannot overflow
617 				 * because index's slot at the current level
618 				 * is greater than 0.
619 				 */
620 				index = vm_radix_trimkey(index,
621 				    rnode->rn_clev);
622 			}
623 			index--;
624 			KASSERT(!vm_radix_keybarr(rnode, index),
625 			    ("vm_radix_lookup_le: keybarr failed"));
626 		}
627 		slot = vm_radix_slot(index, rnode->rn_clev);
628 		child = rnode->rn_child[slot];
629 		if (vm_radix_isleaf(child)) {
630 			m = vm_radix_topage(child);
631 			if (m->pindex <= index)
632 				return (m);
633 		} else if (child != NULL)
634 			goto descend;
635 
636 		/*
637 		 * Look for an available edge or page within the current
638 		 * bisection node.
639 		 */
640 		if (slot > 0) {
641 			inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
642 			index |= inc - 1;
643 			do {
644 				index -= inc;
645 				slot--;
646 				child = rnode->rn_child[slot];
647 				if (vm_radix_isleaf(child)) {
648 					m = vm_radix_topage(child);
649 					if (m->pindex <= index)
650 						return (m);
651 				} else if (child != NULL)
652 					goto descend;
653 			} while (slot > 0);
654 		}
655 		KASSERT(child == NULL || vm_radix_isleaf(child),
656 		    ("vm_radix_lookup_le: child is radix node"));
657 
658 		/*
659 		 * If a page or edge smaller than the search slot is not found
660 		 * in the current node, ascend to the next higher-level node.
661 		 */
662 		goto ascend;
663 descend:
664 		KASSERT(rnode->rn_clev > 0,
665 		    ("vm_radix_lookup_le: pushing leaf's parent"));
666 		KASSERT(tos < VM_RADIX_LIMIT,
667 		    ("vm_radix_lookup_le: stack overflow"));
668 		stack[tos++] = rnode;
669 		rnode = child;
670 	}
671 }
672 
673 /*
674  * Remove the specified index from the trie, and return the value stored at
675  * that index.  If the index is not present, return NULL.
676  */
677 vm_page_t
678 vm_radix_remove(struct vm_radix *rtree, vm_pindex_t index)
679 {
680 	struct vm_radix_node *rnode, *parent;
681 	vm_page_t m;
682 	int i, slot;
683 
684 	rnode = vm_radix_getroot(rtree);
685 	if (vm_radix_isleaf(rnode)) {
686 		m = vm_radix_topage(rnode);
687 		if (m->pindex != index)
688 			return (NULL);
689 		vm_radix_setroot(rtree, NULL);
690 		return (m);
691 	}
692 	parent = NULL;
693 	for (;;) {
694 		if (rnode == NULL)
695 			return (NULL);
696 		slot = vm_radix_slot(index, rnode->rn_clev);
697 		if (vm_radix_isleaf(rnode->rn_child[slot])) {
698 			m = vm_radix_topage(rnode->rn_child[slot]);
699 			if (m->pindex != index)
700 				return (NULL);
701 			rnode->rn_child[slot] = NULL;
702 			rnode->rn_count--;
703 			if (rnode->rn_count > 1)
704 				return (m);
705 			for (i = 0; i < VM_RADIX_COUNT; i++)
706 				if (rnode->rn_child[i] != NULL)
707 					break;
708 			KASSERT(i != VM_RADIX_COUNT,
709 			    ("%s: invalid node configuration", __func__));
710 			if (parent == NULL)
711 				vm_radix_setroot(rtree, rnode->rn_child[i]);
712 			else {
713 				slot = vm_radix_slot(index, parent->rn_clev);
714 				KASSERT(parent->rn_child[slot] == rnode,
715 				    ("%s: invalid child value", __func__));
716 				parent->rn_child[slot] = rnode->rn_child[i];
717 			}
718 			rnode->rn_count--;
719 			rnode->rn_child[i] = NULL;
720 			vm_radix_node_put(rnode);
721 			return (m);
722 		}
723 		parent = rnode;
724 		rnode = rnode->rn_child[slot];
725 	}
726 }
727 
728 /*
729  * Remove and free all the nodes from the radix tree.
730  * This function is recursive but there is a tight control on it as the
731  * maximum depth of the tree is fixed.
732  */
733 void
734 vm_radix_reclaim_allnodes(struct vm_radix *rtree)
735 {
736 	struct vm_radix_node *root;
737 
738 	root = vm_radix_getroot(rtree);
739 	if (root == NULL)
740 		return;
741 	vm_radix_setroot(rtree, NULL);
742 	if (!vm_radix_isleaf(root))
743 		vm_radix_reclaim_allnodes_int(root);
744 }
745 
746 /*
747  * Replace an existing page in the trie with another one.
748  * Panics if there is not an old page in the trie at the new page's index.
749  */
750 vm_page_t
751 vm_radix_replace(struct vm_radix *rtree, vm_page_t newpage)
752 {
753 	struct vm_radix_node *rnode;
754 	vm_page_t m;
755 	vm_pindex_t index;
756 	int slot;
757 
758 	index = newpage->pindex;
759 	rnode = vm_radix_getroot(rtree);
760 	if (rnode == NULL)
761 		panic("%s: replacing page on an empty trie", __func__);
762 	if (vm_radix_isleaf(rnode)) {
763 		m = vm_radix_topage(rnode);
764 		if (m->pindex != index)
765 			panic("%s: original replacing root key not found",
766 			    __func__);
767 		rtree->rt_root = (uintptr_t)newpage | VM_RADIX_ISLEAF;
768 		return (m);
769 	}
770 	for (;;) {
771 		slot = vm_radix_slot(index, rnode->rn_clev);
772 		if (vm_radix_isleaf(rnode->rn_child[slot])) {
773 			m = vm_radix_topage(rnode->rn_child[slot]);
774 			if (m->pindex == index) {
775 				rnode->rn_child[slot] =
776 				    (void *)((uintptr_t)newpage |
777 				    VM_RADIX_ISLEAF);
778 				return (m);
779 			} else
780 				break;
781 		} else if (rnode->rn_child[slot] == NULL ||
782 		    vm_radix_keybarr(rnode->rn_child[slot], index))
783 			break;
784 		rnode = rnode->rn_child[slot];
785 	}
786 	panic("%s: original replacing page not found", __func__);
787 }
788 
789 void
790 vm_radix_wait(void)
791 {
792 	uma_zwait(vm_radix_node_zone);
793 }
794 
795 #ifdef DDB
796 /*
797  * Show details about the given radix node.
798  */
799 DB_SHOW_COMMAND(radixnode, db_show_radixnode)
800 {
801 	struct vm_radix_node *rnode;
802 	int i;
803 
804         if (!have_addr)
805                 return;
806 	rnode = (struct vm_radix_node *)addr;
807 	db_printf("radixnode %p, owner %jx, children count %u, level %u:\n",
808 	    (void *)rnode, (uintmax_t)rnode->rn_owner, rnode->rn_count,
809 	    rnode->rn_clev);
810 	for (i = 0; i < VM_RADIX_COUNT; i++)
811 		if (rnode->rn_child[i] != NULL)
812 			db_printf("slot: %d, val: %p, page: %p, clev: %d\n",
813 			    i, (void *)rnode->rn_child[i],
814 			    vm_radix_isleaf(rnode->rn_child[i]) ?
815 			    vm_radix_topage(rnode->rn_child[i]) : NULL,
816 			    rnode->rn_clev);
817 }
818 #endif /* DDB */
819