2 * Copyright (c) 2013 EMC Corp.
3 * Copyright (c) 2011 Jeffrey Roberson <jeff@freebsd.org>
4 * Copyright (c) 2008 Mayur Shardul <mayur.shardul@gmail.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * Path-compressed radix trie implementation.
32 * The following code is not generalized into a general purpose library
33 * because there are way too many parameters embedded that should really
34 * be decided by the library consumers. At the same time, consumers
35 * of this code must achieve highest possible performance.
37 * The implementation takes into account the following rationale:
38 * - Size of the nodes should be as small as possible but still big enough
39 * to avoid a large maximum depth for the trie. This is a balance
40 * between the necessity to not wire too much physical memory for the nodes
41 * and the necessity to avoid too much cache pollution during the trie
43 * - There is not a huge bias toward the number of lookup operations over
44 * the number of insert and remove operations. This basically implies
45 * that optimizations supposedly helping one operation but hurting the
46 * other might be carefully evaluated.
47 * - On average not many nodes are expected to be fully populated, hence
48 * level compression may just complicate things.
51 #include <sys/cdefs.h>
52 __FBSDID("$FreeBSD$");
56 #include <sys/param.h>
57 #include <sys/systm.h>
58 #include <sys/kernel.h>
59 #include <sys/vmmeter.h>
63 #include <vm/vm_param.h>
64 #include <vm/vm_page.h>
65 #include <vm/vm_radix.h>
72 * These widths should allow the pointers to a node's children to fit within
73 * a single cache line. The extra levels from a narrow width should not be
74 * a problem thanks to path compression.
77 #define VM_RADIX_WIDTH 4
79 #define VM_RADIX_WIDTH 3
82 #define VM_RADIX_COUNT (1 << VM_RADIX_WIDTH)
83 #define VM_RADIX_MASK (VM_RADIX_COUNT - 1)
84 #define VM_RADIX_LIMIT \
85 (howmany((sizeof(vm_pindex_t) * NBBY), VM_RADIX_WIDTH) - 1)
87 /* Flag bits stored in node pointers. */
88 #define VM_RADIX_ISLEAF 0x1
89 #define VM_RADIX_FLAGS 0x1
90 #define VM_RADIX_PAD VM_RADIX_FLAGS
92 /* Returns one unit associated with specified level. */
93 #define VM_RADIX_UNITLEVEL(lev) \
94 ((vm_pindex_t)1 << ((VM_RADIX_LIMIT - (lev)) * VM_RADIX_WIDTH))
96 struct vm_radix_node {
97 vm_pindex_t rn_owner; /* Owner of record. */
98 uint16_t rn_count; /* Valid children. */
99 uint16_t rn_clev; /* Current level. */
100 void *rn_child[VM_RADIX_COUNT]; /* Child nodes. */
103 static uma_zone_t vm_radix_node_zone;
106 * Allocate a radix node. Pre-allocation should ensure that the request
107 * will always be satisfied.
109 static __inline struct vm_radix_node *
110 vm_radix_node_get(vm_pindex_t owner, uint16_t count, uint16_t clevel)
112 struct vm_radix_node *rnode;
114 rnode = uma_zalloc(vm_radix_node_zone, M_NOWAIT);
117 * The required number of nodes should already be pre-allocated
118 * by vm_radix_prealloc(). However, UMA can hold a few nodes
119 * in per-CPU buckets, which will not be accessible by the
120 * current CPU. Thus, the allocation could return NULL when
121 * the pre-allocated pool is close to exhaustion. Anyway,
122 * in practice this should never occur because a new node
123 * is not always required for insert. Thus, the pre-allocated
124 * pool should have some extra pages that prevent this from
125 * becoming a problem.
128 panic("%s: uma_zalloc() returned NULL for a new node",
130 rnode->rn_owner = owner;
131 rnode->rn_count = count;
132 rnode->rn_clev = clevel;
140 vm_radix_node_put(struct vm_radix_node *rnode)
143 uma_zfree(vm_radix_node_zone, rnode);
147 * Return the position in the array for a given level.
150 vm_radix_slot(vm_pindex_t index, uint16_t level)
153 return ((index >> ((VM_RADIX_LIMIT - level) * VM_RADIX_WIDTH)) &
157 /* Trims the key after the specified level. */
158 static __inline vm_pindex_t
159 vm_radix_trimkey(vm_pindex_t index, uint16_t level)
164 if (level < VM_RADIX_LIMIT) {
165 ret >>= (VM_RADIX_LIMIT - level) * VM_RADIX_WIDTH;
166 ret <<= (VM_RADIX_LIMIT - level) * VM_RADIX_WIDTH;
172 * Get the root node for a radix tree.
174 static __inline struct vm_radix_node *
175 vm_radix_getroot(struct vm_radix *rtree)
178 return ((struct vm_radix_node *)rtree->rt_root);
182 * Set the root node for a radix tree.
185 vm_radix_setroot(struct vm_radix *rtree, struct vm_radix_node *rnode)
188 rtree->rt_root = (uintptr_t)rnode;
192 * Returns TRUE if the specified radix node is a leaf and FALSE otherwise.
194 static __inline boolean_t
195 vm_radix_isleaf(struct vm_radix_node *rnode)
198 return (((uintptr_t)rnode & VM_RADIX_ISLEAF) != 0);
202 * Returns the associated page extracted from rnode.
204 static __inline vm_page_t
205 vm_radix_topage(struct vm_radix_node *rnode)
208 return ((vm_page_t)((uintptr_t)rnode & ~VM_RADIX_FLAGS));
212 * Adds the page as a child of the provided node.
215 vm_radix_addpage(struct vm_radix_node *rnode, vm_pindex_t index, uint16_t clev,
220 slot = vm_radix_slot(index, clev);
221 rnode->rn_child[slot] = (void *)((uintptr_t)page | VM_RADIX_ISLEAF);
225 * Returns the slot where two keys differ.
226 * It cannot accept 2 equal keys.
228 static __inline uint16_t
229 vm_radix_keydiff(vm_pindex_t index1, vm_pindex_t index2)
233 KASSERT(index1 != index2, ("%s: passing the same key value %jx",
234 __func__, (uintmax_t)index1));
237 for (clev = 0; clev <= VM_RADIX_LIMIT ; clev++)
238 if (vm_radix_slot(index1, clev))
240 panic("%s: cannot reach this point", __func__);
245 * Returns TRUE if it can be determined that key does not belong to the
246 * specified rnode. Otherwise, returns FALSE.
248 static __inline boolean_t
249 vm_radix_keybarr(struct vm_radix_node *rnode, vm_pindex_t idx)
252 if (rnode->rn_clev > 0) {
253 idx = vm_radix_trimkey(idx, rnode->rn_clev - 1);
254 return (idx != rnode->rn_owner);
260 * Adjusts the idx key to the first upper level available, based on a valid
261 * initial level and map of available levels.
262 * Returns a value bigger than 0 to signal that there are not valid levels
266 vm_radix_addlev(vm_pindex_t *idx, boolean_t *levels, uint16_t ilev)
270 for (; levels[ilev] == FALSE ||
271 vm_radix_slot(*idx, ilev) == (VM_RADIX_COUNT - 1); ilev--)
275 *idx = vm_radix_trimkey(*idx, ilev);
276 *idx += VM_RADIX_UNITLEVEL(ilev);
277 return (*idx < wrapidx);
281 * Adjusts the idx key to the first lower level available, based on a valid
282 * initial level and map of available levels.
283 * Returns a value bigger than 0 to signal that there are not valid levels
287 vm_radix_declev(vm_pindex_t *idx, boolean_t *levels, uint16_t ilev)
291 for (; levels[ilev] == FALSE ||
292 vm_radix_slot(*idx, ilev) == 0; ilev--)
296 *idx = vm_radix_trimkey(*idx, ilev);
297 *idx |= VM_RADIX_UNITLEVEL(ilev) - 1;
298 *idx -= VM_RADIX_UNITLEVEL(ilev);
299 return (*idx > wrapidx);
303 * Internal helper for vm_radix_reclaim_allnodes().
304 * This function is recursive.
307 vm_radix_reclaim_allnodes_int(struct vm_radix_node *rnode)
311 KASSERT(rnode->rn_count <= VM_RADIX_COUNT,
312 ("vm_radix_reclaim_allnodes_int: bad count in rnode %p", rnode));
313 for (slot = 0; rnode->rn_count != 0; slot++) {
314 if (rnode->rn_child[slot] == NULL)
316 if (!vm_radix_isleaf(rnode->rn_child[slot]))
317 vm_radix_reclaim_allnodes_int(rnode->rn_child[slot]);
318 rnode->rn_child[slot] = NULL;
321 vm_radix_node_put(rnode);
326 * Radix node zone destructor.
329 vm_radix_node_zone_dtor(void *mem, int size __unused, void *arg __unused)
331 struct vm_radix_node *rnode;
335 KASSERT(rnode->rn_count == 0,
336 ("vm_radix_node_put: rnode %p has %d children", rnode,
338 for (slot = 0; slot < VM_RADIX_COUNT; slot++)
339 KASSERT(rnode->rn_child[slot] == NULL,
340 ("vm_radix_node_put: rnode %p has a child", rnode));
345 * Radix node zone initializer.
348 vm_radix_node_zone_init(void *mem, int size __unused, int flags __unused)
350 struct vm_radix_node *rnode;
353 memset(rnode->rn_child, 0, sizeof(rnode->rn_child));
358 * Pre-allocate intermediate nodes from the UMA slab zone.
361 vm_radix_prealloc(void *arg __unused)
364 if (!uma_zone_reserve_kva(vm_radix_node_zone, cnt.v_page_count))
365 panic("%s: unable to create new zone", __func__);
366 uma_prealloc(vm_radix_node_zone, cnt.v_page_count);
368 SYSINIT(vm_radix_prealloc, SI_SUB_KMEM, SI_ORDER_SECOND, vm_radix_prealloc,
372 * Initialize the UMA slab zone.
373 * Until vm_radix_prealloc() is called, the zone will be served by the
374 * UMA boot-time pre-allocated pool of pages.
380 vm_radix_node_zone = uma_zcreate("RADIX NODE",
381 sizeof(struct vm_radix_node), NULL,
383 vm_radix_node_zone_dtor,
387 vm_radix_node_zone_init, NULL, VM_RADIX_PAD, UMA_ZONE_VM |
392 * Inserts the key-value pair into the trie.
393 * Panics if the key already exists.
396 vm_radix_insert(struct vm_radix *rtree, vm_page_t page)
398 vm_pindex_t index, newind;
399 struct vm_radix_node *parent, *rnode, *tmp;
404 index = page->pindex;
407 * The owner of record for root is not really important because it
408 * will never be used.
410 rnode = vm_radix_getroot(rtree);
412 rnode = vm_radix_node_get(0, 1, 0);
413 vm_radix_setroot(rtree, rnode);
414 vm_radix_addpage(rnode, index, 0, page);
418 slot = vm_radix_slot(index, rnode->rn_clev);
419 if (vm_radix_isleaf(rnode->rn_child[slot])) {
420 m = vm_radix_topage(rnode->rn_child[slot]);
421 if (m->pindex == index)
422 panic("%s: key %jx is already present",
423 __func__, (uintmax_t)index);
424 clev = vm_radix_keydiff(m->pindex, index);
425 tmp = vm_radix_node_get(vm_radix_trimkey(index,
427 rnode->rn_child[slot] = tmp;
428 vm_radix_addpage(tmp, index, clev, page);
429 vm_radix_addpage(tmp, m->pindex, clev, m);
432 if (rnode->rn_child[slot] == NULL) {
434 vm_radix_addpage(rnode, index, rnode->rn_clev, page);
438 rnode = rnode->rn_child[slot];
439 } while (!vm_radix_keybarr(rnode, index));
442 * A new node is needed because the right insertion level is reached.
443 * Setup the new intermediate node and add the 2 children: the
444 * new object and the older edge.
446 newind = rnode->rn_owner;
447 clev = vm_radix_keydiff(newind, index);
448 tmp = vm_radix_node_get(vm_radix_trimkey(index, clev - 1), 2,
450 parent->rn_child[slot] = tmp;
451 vm_radix_addpage(tmp, index, clev, page);
452 slot = vm_radix_slot(newind, clev);
453 tmp->rn_child[slot] = rnode;
457 * Returns the value stored at the index. If the index is not present,
461 vm_radix_lookup(struct vm_radix *rtree, vm_pindex_t index)
463 struct vm_radix_node *rnode;
467 rnode = vm_radix_getroot(rtree);
468 while (rnode != NULL) {
469 if (vm_radix_isleaf(rnode)) {
470 m = vm_radix_topage(rnode);
471 if (m->pindex == index)
475 } else if (vm_radix_keybarr(rnode, index))
477 slot = vm_radix_slot(index, rnode->rn_clev);
478 rnode = rnode->rn_child[slot];
484 * Look up the nearest entry at a position bigger than or equal to index.
487 vm_radix_lookup_ge(struct vm_radix *rtree, vm_pindex_t index)
491 struct vm_radix_node *child, *rnode;
494 boolean_t maplevels[VM_RADIX_LIMIT + 1];
499 rnode = vm_radix_getroot(rtree);
502 else if (vm_radix_isleaf(rnode)) {
503 m = vm_radix_topage(rnode);
504 if (m->pindex >= index)
510 KASSERT(++loops < 1000, ("%s: too many loops", __func__));
511 for (difflev = 0; difflev < (VM_RADIX_LIMIT + 1); difflev++)
512 maplevels[difflev] = FALSE;
514 maplevels[rnode->rn_clev] = TRUE;
517 * If the keys differ before the current bisection node
518 * the search key might rollback to the earliest
519 * available bisection node, or to the smaller value
520 * in the current domain (if the owner is bigger than the
522 * The maplevels array records any node has been seen
523 * at a given level. This aids the search for a valid
526 if (vm_radix_keybarr(rnode, index)) {
527 difflev = vm_radix_keydiff(index, rnode->rn_owner);
528 if (index > rnode->rn_owner) {
529 if (vm_radix_addlev(&index, maplevels,
533 index = vm_radix_trimkey(rnode->rn_owner,
535 rnode = vm_radix_getroot(rtree);
538 slot = vm_radix_slot(index, rnode->rn_clev);
539 child = rnode->rn_child[slot];
540 if (vm_radix_isleaf(child)) {
541 m = vm_radix_topage(child);
542 if (m->pindex >= index)
544 } else if (child != NULL)
548 * Look for an available edge or page within the current
551 if (slot < (VM_RADIX_COUNT - 1)) {
552 inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
553 index = vm_radix_trimkey(index, rnode->rn_clev);
557 child = rnode->rn_child[slot];
558 if (vm_radix_isleaf(child)) {
559 m = vm_radix_topage(child);
560 if (m->pindex >= index)
562 } else if (child != NULL)
564 } while (slot < (VM_RADIX_COUNT - 1));
566 KASSERT(child == NULL || vm_radix_isleaf(child),
567 ("vm_radix_lookup_ge: child is radix node"));
570 * If a valid page or edge bigger than the search slot is
571 * found in the traversal, skip to the next higher-level key.
573 if (rnode->rn_clev == 0 || vm_radix_addlev(&index, maplevels,
574 rnode->rn_clev - 1) > 0)
576 rnode = vm_radix_getroot(rtree);
585 * Look up the nearest entry at a position less than or equal to index.
588 vm_radix_lookup_le(struct vm_radix *rtree, vm_pindex_t index)
592 struct vm_radix_node *child, *rnode;
595 boolean_t maplevels[VM_RADIX_LIMIT + 1];
600 rnode = vm_radix_getroot(rtree);
603 else if (vm_radix_isleaf(rnode)) {
604 m = vm_radix_topage(rnode);
605 if (m->pindex <= index)
611 KASSERT(++loops < 1000, ("%s: too many loops", __func__));
612 for (difflev = 0; difflev < (VM_RADIX_LIMIT + 1); difflev++)
613 maplevels[difflev] = FALSE;
615 maplevels[rnode->rn_clev] = TRUE;
618 * If the keys differ before the current bisection node
619 * the search key might rollback to the earliest
620 * available bisection node, or to the higher value
621 * in the current domain (if the owner is smaller than the
623 * The maplevels array records any node has been seen
624 * at a given level. This aids the search for a valid
627 if (vm_radix_keybarr(rnode, index)) {
628 difflev = vm_radix_keydiff(index, rnode->rn_owner);
629 if (index > rnode->rn_owner) {
630 index = vm_radix_trimkey(rnode->rn_owner,
632 index |= VM_RADIX_UNITLEVEL(difflev) - 1;
633 } else if (vm_radix_declev(&index, maplevels,
636 rnode = vm_radix_getroot(rtree);
639 slot = vm_radix_slot(index, rnode->rn_clev);
640 child = rnode->rn_child[slot];
641 if (vm_radix_isleaf(child)) {
642 m = vm_radix_topage(child);
643 if (m->pindex <= index)
645 } else if (child != NULL)
649 * Look for an available edge or page within the current
653 inc = VM_RADIX_UNITLEVEL(rnode->rn_clev);
654 index = vm_radix_trimkey(index, rnode->rn_clev);
659 child = rnode->rn_child[slot];
660 if (vm_radix_isleaf(child)) {
661 m = vm_radix_topage(child);
662 if (m->pindex <= index)
664 } else if (child != NULL)
668 KASSERT(child == NULL || vm_radix_isleaf(child),
669 ("vm_radix_lookup_le: child is radix node"));
672 * If a valid page or edge smaller than the search slot is
673 * found in the traversal, skip to the next higher-level key.
675 if (rnode->rn_clev == 0 || vm_radix_declev(&index, maplevels,
676 rnode->rn_clev - 1) > 0)
678 rnode = vm_radix_getroot(rtree);
687 * Remove the specified index from the tree.
688 * Panics if the key is not present.
691 vm_radix_remove(struct vm_radix *rtree, vm_pindex_t index)
693 struct vm_radix_node *rnode, *parent;
698 rnode = vm_radix_getroot(rtree);
701 panic("vm_radix_remove: impossible to locate the key");
702 slot = vm_radix_slot(index, rnode->rn_clev);
703 if (vm_radix_isleaf(rnode->rn_child[slot])) {
704 m = vm_radix_topage(rnode->rn_child[slot]);
705 if (m->pindex != index)
706 panic("%s: invalid key found", __func__);
707 rnode->rn_child[slot] = NULL;
709 if (rnode->rn_count > 1)
711 if (parent == NULL) {
712 if (rnode->rn_count == 0) {
713 vm_radix_node_put(rnode);
714 vm_radix_setroot(rtree, NULL);
718 for (i = 0; i < VM_RADIX_COUNT; i++)
719 if (rnode->rn_child[i] != NULL)
721 KASSERT(i != VM_RADIX_COUNT,
722 ("%s: invalid node configuration", __func__));
723 slot = vm_radix_slot(index, parent->rn_clev);
724 KASSERT(parent->rn_child[slot] == rnode,
725 ("%s: invalid child value", __func__));
726 parent->rn_child[slot] = rnode->rn_child[i];
728 rnode->rn_child[i] = NULL;
729 vm_radix_node_put(rnode);
733 rnode = rnode->rn_child[slot];
738 * Remove and free all the nodes from the radix tree.
739 * This function is recursive but there is a tight control on it as the
740 * maximum depth of the tree is fixed.
743 vm_radix_reclaim_allnodes(struct vm_radix *rtree)
745 struct vm_radix_node *root;
747 root = vm_radix_getroot(rtree);
750 vm_radix_setroot(rtree, NULL);
751 vm_radix_reclaim_allnodes_int(root);
756 * Show details about the given radix node.
758 DB_SHOW_COMMAND(radixnode, db_show_radixnode)
760 struct vm_radix_node *rnode;
765 rnode = (struct vm_radix_node *)addr;
766 db_printf("radixnode %p, owner %jx, children count %u, level %u:\n",
767 (void *)rnode, (uintmax_t)rnode->rn_owner, rnode->rn_count,
769 for (i = 0; i < VM_RADIX_COUNT; i++)
770 if (rnode->rn_child[i] != NULL)
771 db_printf("slot: %d, val: %p, page: %p, clev: %d\n",
772 i, (void *)rnode->rn_child[i],
773 vm_radix_isleaf(rnode->rn_child[i]) ?
774 vm_radix_topage(rnode->rn_child[i]) : NULL,