2 * Copyright (C) 2001 Momchil Velikov
3 * Portions Copyright (C) 2001 Christoph Hellwig
4 * Copyright (C) 2005 SGI, Christoph Lameter
5 * Copyright (C) 2006 Nick Piggin
6 * Copyright (C) 2012 Konstantin Khlebnikov
7 * Copyright (C) 2016 Intel, Matthew Wilcox
8 * Copyright (C) 2016 Intel, Ross Zwisler
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public License as
12 * published by the Free Software Foundation; either version 2, or (at
13 * your option) any later version.
15 * This program is distributed in the hope that it will be useful, but
16 * WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 #include <linux/errno.h>
26 #include <linux/init.h>
27 #include <linux/kernel.h>
28 #include <linux/export.h>
29 #include <linux/radix-tree.h>
30 #include <linux/percpu.h>
31 #include <linux/slab.h>
32 #include <linux/kmemleak.h>
33 #include <linux/notifier.h>
34 #include <linux/cpu.h>
35 #include <linux/string.h>
36 #include <linux/bitops.h>
37 #include <linux/rcupdate.h>
38 #include <linux/preempt.h> /* in_interrupt() */
42 * The height_to_maxindex array needs to be one deeper than the maximum
43 * path as height 0 holds only 1 entry.
45 static unsigned long height_to_maxindex[RADIX_TREE_MAX_PATH + 1] __read_mostly;
48 * Radix tree node cache.
50 static struct kmem_cache *radix_tree_node_cachep;
53 * The radix tree is variable-height, so an insert operation not only has
54 * to build the branch to its corresponding item, it also has to build the
55 * branch to existing items if the size has to be increased (by
58 * The worst case is a zero height tree with just a single item at index 0,
59 * and then inserting an item at index ULONG_MAX. This requires 2 new branches
60 * of RADIX_TREE_MAX_PATH size to be created, with only the root node shared.
63 #define RADIX_TREE_PRELOAD_SIZE (RADIX_TREE_MAX_PATH * 2 - 1)
66 * Per-cpu pool of preloaded nodes
68 struct radix_tree_preload {
70 /* nodes->private_data points to next preallocated node */
71 struct radix_tree_node *nodes;
73 static DEFINE_PER_CPU(struct radix_tree_preload, radix_tree_preloads) = { 0, };
75 static inline void *ptr_to_indirect(void *ptr)
77 return (void *)((unsigned long)ptr | RADIX_TREE_INDIRECT_PTR);
80 #define RADIX_TREE_RETRY ptr_to_indirect(NULL)
82 #ifdef CONFIG_RADIX_TREE_MULTIORDER
83 /* Sibling slots point directly to another slot in the same node */
84 static inline bool is_sibling_entry(struct radix_tree_node *parent, void *node)
87 return (parent->slots <= ptr) &&
88 (ptr < parent->slots + RADIX_TREE_MAP_SIZE);
91 static inline bool is_sibling_entry(struct radix_tree_node *parent, void *node)
97 static inline unsigned long get_slot_offset(struct radix_tree_node *parent,
100 return slot - parent->slots;
103 static unsigned radix_tree_descend(struct radix_tree_node *parent,
104 struct radix_tree_node **nodep, unsigned offset)
106 void **entry = rcu_dereference_raw(parent->slots[offset]);
108 #ifdef CONFIG_RADIX_TREE_MULTIORDER
109 if (radix_tree_is_indirect_ptr(entry)) {
110 unsigned long siboff = get_slot_offset(parent, entry);
111 if (siboff < RADIX_TREE_MAP_SIZE) {
113 entry = rcu_dereference_raw(parent->slots[offset]);
118 *nodep = (void *)entry;
122 static inline gfp_t root_gfp_mask(struct radix_tree_root *root)
124 return root->gfp_mask & __GFP_BITS_MASK;
127 static inline void tag_set(struct radix_tree_node *node, unsigned int tag,
130 __set_bit(offset, node->tags[tag]);
133 static inline void tag_clear(struct radix_tree_node *node, unsigned int tag,
136 __clear_bit(offset, node->tags[tag]);
139 static inline int tag_get(struct radix_tree_node *node, unsigned int tag,
142 return test_bit(offset, node->tags[tag]);
145 static inline void root_tag_set(struct radix_tree_root *root, unsigned int tag)
147 root->gfp_mask |= (__force gfp_t)(1 << (tag + __GFP_BITS_SHIFT));
150 static inline void root_tag_clear(struct radix_tree_root *root, unsigned tag)
152 root->gfp_mask &= (__force gfp_t)~(1 << (tag + __GFP_BITS_SHIFT));
155 static inline void root_tag_clear_all(struct radix_tree_root *root)
157 root->gfp_mask &= __GFP_BITS_MASK;
160 static inline int root_tag_get(struct radix_tree_root *root, unsigned int tag)
162 return (__force int)root->gfp_mask & (1 << (tag + __GFP_BITS_SHIFT));
165 static inline unsigned root_tags_get(struct radix_tree_root *root)
167 return (__force unsigned)root->gfp_mask >> __GFP_BITS_SHIFT;
171 * Returns 1 if any slot in the node has this tag set.
172 * Otherwise returns 0.
174 static inline int any_tag_set(struct radix_tree_node *node, unsigned int tag)
177 for (idx = 0; idx < RADIX_TREE_TAG_LONGS; idx++) {
178 if (node->tags[tag][idx])
185 * radix_tree_find_next_bit - find the next set bit in a memory region
187 * @addr: The address to base the search on
188 * @size: The bitmap size in bits
189 * @offset: The bitnumber to start searching at
191 * Unrollable variant of find_next_bit() for constant size arrays.
192 * Tail bits starting from size to roundup(size, BITS_PER_LONG) must be zero.
193 * Returns next bit offset, or size if nothing found.
195 static __always_inline unsigned long
196 radix_tree_find_next_bit(const unsigned long *addr,
197 unsigned long size, unsigned long offset)
199 if (!__builtin_constant_p(size))
200 return find_next_bit(addr, size, offset);
205 addr += offset / BITS_PER_LONG;
206 tmp = *addr >> (offset % BITS_PER_LONG);
208 return __ffs(tmp) + offset;
209 offset = (offset + BITS_PER_LONG) & ~(BITS_PER_LONG - 1);
210 while (offset < size) {
213 return __ffs(tmp) + offset;
214 offset += BITS_PER_LONG;
221 static void dump_node(struct radix_tree_node *node,
222 unsigned shift, unsigned long index)
226 pr_debug("radix node: %p offset %d tags %lx %lx %lx height %d count %d parent %p\n",
228 node->tags[0][0], node->tags[1][0], node->tags[2][0],
229 node->height, node->count, node->parent);
231 for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
232 unsigned long first = index | (i << shift);
233 unsigned long last = first | ((1UL << shift) - 1);
234 void *entry = node->slots[i];
237 if (is_sibling_entry(node, entry)) {
238 pr_debug("radix sblng %p offset %ld val %p indices %ld-%ld\n",
240 *(void **)indirect_to_ptr(entry),
242 } else if (!radix_tree_is_indirect_ptr(entry)) {
243 pr_debug("radix entry %p offset %ld indices %ld-%ld\n",
244 entry, i, first, last);
246 dump_node(indirect_to_ptr(entry),
247 shift - RADIX_TREE_MAP_SHIFT, first);
253 static void radix_tree_dump(struct radix_tree_root *root)
255 pr_debug("radix root: %p height %d rnode %p tags %x\n",
256 root, root->height, root->rnode,
257 root->gfp_mask >> __GFP_BITS_SHIFT);
258 if (!radix_tree_is_indirect_ptr(root->rnode))
260 dump_node(indirect_to_ptr(root->rnode),
261 (root->height - 1) * RADIX_TREE_MAP_SHIFT, 0);
266 * This assumes that the caller has performed appropriate preallocation, and
267 * that the caller has pinned this thread of control to the current CPU.
269 static struct radix_tree_node *
270 radix_tree_node_alloc(struct radix_tree_root *root)
272 struct radix_tree_node *ret = NULL;
273 gfp_t gfp_mask = root_gfp_mask(root);
276 * Preload code isn't irq safe and it doesn't make sense to use
277 * preloading during an interrupt anyway as all the allocations have
278 * to be atomic. So just do normal allocation when in interrupt.
280 if (!gfpflags_allow_blocking(gfp_mask) && !in_interrupt()) {
281 struct radix_tree_preload *rtp;
284 * Even if the caller has preloaded, try to allocate from the
285 * cache first for the new node to get accounted.
287 ret = kmem_cache_alloc(radix_tree_node_cachep,
288 gfp_mask | __GFP_ACCOUNT | __GFP_NOWARN);
293 * Provided the caller has preloaded here, we will always
294 * succeed in getting a node here (and never reach
297 rtp = this_cpu_ptr(&radix_tree_preloads);
300 rtp->nodes = ret->private_data;
301 ret->private_data = NULL;
305 * Update the allocation stack trace as this is more useful
308 kmemleak_update_trace(ret);
311 ret = kmem_cache_alloc(radix_tree_node_cachep,
312 gfp_mask | __GFP_ACCOUNT);
314 BUG_ON(radix_tree_is_indirect_ptr(ret));
318 static void radix_tree_node_rcu_free(struct rcu_head *head)
320 struct radix_tree_node *node =
321 container_of(head, struct radix_tree_node, rcu_head);
325 * must only free zeroed nodes into the slab. radix_tree_shrink
326 * can leave us with a non-NULL entry in the first slot, so clear
327 * that here to make sure.
329 for (i = 0; i < RADIX_TREE_MAX_TAGS; i++)
330 tag_clear(node, i, 0);
332 node->slots[0] = NULL;
335 kmem_cache_free(radix_tree_node_cachep, node);
339 radix_tree_node_free(struct radix_tree_node *node)
341 call_rcu(&node->rcu_head, radix_tree_node_rcu_free);
345 * Load up this CPU's radix_tree_node buffer with sufficient objects to
346 * ensure that the addition of a single element in the tree cannot fail. On
347 * success, return zero, with preemption disabled. On error, return -ENOMEM
348 * with preemption not disabled.
350 * To make use of this facility, the radix tree must be initialised without
351 * __GFP_DIRECT_RECLAIM being passed to INIT_RADIX_TREE().
353 static int __radix_tree_preload(gfp_t gfp_mask)
355 struct radix_tree_preload *rtp;
356 struct radix_tree_node *node;
360 rtp = this_cpu_ptr(&radix_tree_preloads);
361 while (rtp->nr < RADIX_TREE_PRELOAD_SIZE) {
363 node = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
367 rtp = this_cpu_ptr(&radix_tree_preloads);
368 if (rtp->nr < RADIX_TREE_PRELOAD_SIZE) {
369 node->private_data = rtp->nodes;
373 kmem_cache_free(radix_tree_node_cachep, node);
382 * Load up this CPU's radix_tree_node buffer with sufficient objects to
383 * ensure that the addition of a single element in the tree cannot fail. On
384 * success, return zero, with preemption disabled. On error, return -ENOMEM
385 * with preemption not disabled.
387 * To make use of this facility, the radix tree must be initialised without
388 * __GFP_DIRECT_RECLAIM being passed to INIT_RADIX_TREE().
390 int radix_tree_preload(gfp_t gfp_mask)
392 /* Warn on non-sensical use... */
393 WARN_ON_ONCE(!gfpflags_allow_blocking(gfp_mask));
394 return __radix_tree_preload(gfp_mask);
396 EXPORT_SYMBOL(radix_tree_preload);
399 * The same as above function, except we don't guarantee preloading happens.
400 * We do it, if we decide it helps. On success, return zero with preemption
401 * disabled. On error, return -ENOMEM with preemption not disabled.
403 int radix_tree_maybe_preload(gfp_t gfp_mask)
405 if (gfpflags_allow_blocking(gfp_mask))
406 return __radix_tree_preload(gfp_mask);
407 /* Preloading doesn't help anything with this gfp mask, skip it */
411 EXPORT_SYMBOL(radix_tree_maybe_preload);
414 * Return the maximum key which can be store into a
415 * radix tree with height HEIGHT.
417 static inline unsigned long radix_tree_maxindex(unsigned int height)
419 return height_to_maxindex[height];
422 static inline unsigned long node_maxindex(struct radix_tree_node *node)
424 return radix_tree_maxindex(node->height);
427 static unsigned radix_tree_load_root(struct radix_tree_root *root,
428 struct radix_tree_node **nodep, unsigned long *maxindex)
430 struct radix_tree_node *node = rcu_dereference_raw(root->rnode);
434 if (likely(radix_tree_is_indirect_ptr(node))) {
435 node = indirect_to_ptr(node);
436 *maxindex = node_maxindex(node);
437 return node->height * RADIX_TREE_MAP_SHIFT;
445 * Extend a radix tree so it can store key @index.
447 static int radix_tree_extend(struct radix_tree_root *root,
450 struct radix_tree_node *slot;
454 /* Figure out what the height should be. */
455 height = root->height + 1;
456 while (index > radix_tree_maxindex(height))
459 if (root->rnode == NULL) {
460 root->height = height;
465 unsigned int newheight;
466 struct radix_tree_node *node = radix_tree_node_alloc(root);
471 /* Propagate the aggregated tag info into the new root */
472 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
473 if (root_tag_get(root, tag))
474 tag_set(node, tag, 0);
477 /* Increase the height. */
478 newheight = root->height + 1;
479 BUG_ON(newheight > BITS_PER_LONG);
480 node->height = newheight;
485 if (radix_tree_is_indirect_ptr(slot)) {
486 slot = indirect_to_ptr(slot);
488 slot = ptr_to_indirect(slot);
490 node->slots[0] = slot;
491 node = ptr_to_indirect(node);
492 rcu_assign_pointer(root->rnode, node);
493 root->height = newheight;
494 } while (height > root->height);
496 return height * RADIX_TREE_MAP_SHIFT;
500 * __radix_tree_create - create a slot in a radix tree
501 * @root: radix tree root
503 * @order: index occupies 2^order aligned slots
504 * @nodep: returns node
505 * @slotp: returns slot
507 * Create, if necessary, and return the node and slot for an item
508 * at position @index in the radix tree @root.
510 * Until there is more than one item in the tree, no nodes are
511 * allocated and @root->rnode is used as a direct slot instead of
512 * pointing to a node, in which case *@nodep will be NULL.
514 * Returns -ENOMEM, or 0 for success.
516 int __radix_tree_create(struct radix_tree_root *root, unsigned long index,
517 unsigned order, struct radix_tree_node **nodep,
520 struct radix_tree_node *node = NULL, *slot;
521 unsigned long maxindex;
522 unsigned int height, shift, offset;
523 unsigned long max = index | ((1UL << order) - 1);
525 shift = radix_tree_load_root(root, &slot, &maxindex);
527 /* Make sure the tree is high enough. */
528 if (max > maxindex) {
529 int error = radix_tree_extend(root, max);
534 if (order == shift) {
535 shift += RADIX_TREE_MAP_SHIFT;
540 height = root->height;
542 offset = 0; /* uninitialised var warning */
543 while (shift > order) {
545 /* Have to add a child node. */
546 slot = radix_tree_node_alloc(root);
549 slot->height = height;
550 slot->offset = offset;
553 rcu_assign_pointer(node->slots[offset],
554 ptr_to_indirect(slot));
557 rcu_assign_pointer(root->rnode,
558 ptr_to_indirect(slot));
559 } else if (!radix_tree_is_indirect_ptr(slot))
562 /* Go a level down */
564 shift -= RADIX_TREE_MAP_SHIFT;
565 node = indirect_to_ptr(slot);
566 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
567 offset = radix_tree_descend(node, &slot, offset);
570 #ifdef CONFIG_RADIX_TREE_MULTIORDER
571 /* Insert pointers to the canonical entry */
573 int i, n = 1 << (order - shift);
574 offset = offset & ~(n - 1);
575 slot = ptr_to_indirect(&node->slots[offset]);
576 for (i = 0; i < n; i++) {
577 if (node->slots[offset + i])
581 for (i = 1; i < n; i++) {
582 rcu_assign_pointer(node->slots[offset + i], slot);
591 *slotp = node ? node->slots + offset : (void **)&root->rnode;
596 * __radix_tree_insert - insert into a radix tree
597 * @root: radix tree root
599 * @order: key covers the 2^order indices around index
600 * @item: item to insert
602 * Insert an item into the radix tree at position @index.
604 int __radix_tree_insert(struct radix_tree_root *root, unsigned long index,
605 unsigned order, void *item)
607 struct radix_tree_node *node;
611 BUG_ON(radix_tree_is_indirect_ptr(item));
613 error = __radix_tree_create(root, index, order, &node, &slot);
618 rcu_assign_pointer(*slot, item);
621 unsigned offset = get_slot_offset(node, slot);
623 BUG_ON(tag_get(node, 0, offset));
624 BUG_ON(tag_get(node, 1, offset));
625 BUG_ON(tag_get(node, 2, offset));
627 BUG_ON(root_tags_get(root));
632 EXPORT_SYMBOL(__radix_tree_insert);
635 * __radix_tree_lookup - lookup an item in a radix tree
636 * @root: radix tree root
638 * @nodep: returns node
639 * @slotp: returns slot
641 * Lookup and return the item at position @index in the radix
644 * Until there is more than one item in the tree, no nodes are
645 * allocated and @root->rnode is used as a direct slot instead of
646 * pointing to a node, in which case *@nodep will be NULL.
648 void *__radix_tree_lookup(struct radix_tree_root *root, unsigned long index,
649 struct radix_tree_node **nodep, void ***slotp)
651 struct radix_tree_node *node, *parent;
652 unsigned long maxindex;
658 slot = (void **)&root->rnode;
659 shift = radix_tree_load_root(root, &node, &maxindex);
660 if (index > maxindex)
663 while (radix_tree_is_indirect_ptr(node)) {
666 if (node == RADIX_TREE_RETRY)
668 parent = indirect_to_ptr(node);
669 shift -= RADIX_TREE_MAP_SHIFT;
670 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
671 offset = radix_tree_descend(parent, &node, offset);
672 slot = parent->slots + offset;
683 * radix_tree_lookup_slot - lookup a slot in a radix tree
684 * @root: radix tree root
687 * Returns: the slot corresponding to the position @index in the
688 * radix tree @root. This is useful for update-if-exists operations.
690 * This function can be called under rcu_read_lock iff the slot is not
691 * modified by radix_tree_replace_slot, otherwise it must be called
692 * exclusive from other writers. Any dereference of the slot must be done
693 * using radix_tree_deref_slot.
695 void **radix_tree_lookup_slot(struct radix_tree_root *root, unsigned long index)
699 if (!__radix_tree_lookup(root, index, NULL, &slot))
703 EXPORT_SYMBOL(radix_tree_lookup_slot);
706 * radix_tree_lookup - perform lookup operation on a radix tree
707 * @root: radix tree root
710 * Lookup the item at the position @index in the radix tree @root.
712 * This function can be called under rcu_read_lock, however the caller
713 * must manage lifetimes of leaf nodes (eg. RCU may also be used to free
714 * them safely). No RCU barriers are required to access or modify the
715 * returned item, however.
717 void *radix_tree_lookup(struct radix_tree_root *root, unsigned long index)
719 return __radix_tree_lookup(root, index, NULL, NULL);
721 EXPORT_SYMBOL(radix_tree_lookup);
724 * radix_tree_tag_set - set a tag on a radix tree node
725 * @root: radix tree root
729 * Set the search tag (which must be < RADIX_TREE_MAX_TAGS)
730 * corresponding to @index in the radix tree. From
731 * the root all the way down to the leaf node.
733 * Returns the address of the tagged item. Setting a tag on a not-present
736 void *radix_tree_tag_set(struct radix_tree_root *root,
737 unsigned long index, unsigned int tag)
739 struct radix_tree_node *node, *parent;
740 unsigned long maxindex;
743 shift = radix_tree_load_root(root, &node, &maxindex);
744 BUG_ON(index > maxindex);
746 while (radix_tree_is_indirect_ptr(node)) {
749 shift -= RADIX_TREE_MAP_SHIFT;
750 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
752 parent = indirect_to_ptr(node);
753 offset = radix_tree_descend(parent, &node, offset);
756 if (!tag_get(parent, tag, offset))
757 tag_set(parent, tag, offset);
760 /* set the root's tag bit */
761 if (!root_tag_get(root, tag))
762 root_tag_set(root, tag);
766 EXPORT_SYMBOL(radix_tree_tag_set);
769 * radix_tree_tag_clear - clear a tag on a radix tree node
770 * @root: radix tree root
774 * Clear the search tag (which must be < RADIX_TREE_MAX_TAGS)
775 * corresponding to @index in the radix tree. If this causes
776 * the leaf node to have no tags set then clear the tag in the
777 * next-to-leaf node, etc.
779 * Returns the address of the tagged item on success, else NULL. ie:
780 * has the same return value and semantics as radix_tree_lookup().
782 void *radix_tree_tag_clear(struct radix_tree_root *root,
783 unsigned long index, unsigned int tag)
785 struct radix_tree_node *node, *parent;
786 unsigned long maxindex;
788 int uninitialized_var(offset);
790 shift = radix_tree_load_root(root, &node, &maxindex);
791 if (index > maxindex)
796 while (radix_tree_is_indirect_ptr(node)) {
797 shift -= RADIX_TREE_MAP_SHIFT;
798 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
800 parent = indirect_to_ptr(node);
801 offset = radix_tree_descend(parent, &node, offset);
810 if (!tag_get(parent, tag, offset))
812 tag_clear(parent, tag, offset);
813 if (any_tag_set(parent, tag))
816 index >>= RADIX_TREE_MAP_SHIFT;
817 offset = index & RADIX_TREE_MAP_MASK;
818 parent = parent->parent;
821 /* clear the root's tag bit */
822 if (root_tag_get(root, tag))
823 root_tag_clear(root, tag);
828 EXPORT_SYMBOL(radix_tree_tag_clear);
831 * radix_tree_tag_get - get a tag on a radix tree node
832 * @root: radix tree root
834 * @tag: tag index (< RADIX_TREE_MAX_TAGS)
838 * 0: tag not present or not set
841 * Note that the return value of this function may not be relied on, even if
842 * the RCU lock is held, unless tag modification and node deletion are excluded
845 int radix_tree_tag_get(struct radix_tree_root *root,
846 unsigned long index, unsigned int tag)
848 struct radix_tree_node *node, *parent;
849 unsigned long maxindex;
852 if (!root_tag_get(root, tag))
855 shift = radix_tree_load_root(root, &node, &maxindex);
856 if (index > maxindex)
861 while (radix_tree_is_indirect_ptr(node)) {
864 shift -= RADIX_TREE_MAP_SHIFT;
865 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
867 parent = indirect_to_ptr(node);
868 offset = radix_tree_descend(parent, &node, offset);
872 if (!tag_get(parent, tag, offset))
874 if (node == RADIX_TREE_RETRY)
880 EXPORT_SYMBOL(radix_tree_tag_get);
882 static inline void __set_iter_shift(struct radix_tree_iter *iter,
885 #ifdef CONFIG_RADIX_TREE_MULTIORDER
891 * radix_tree_next_chunk - find next chunk of slots for iteration
893 * @root: radix tree root
894 * @iter: iterator state
895 * @flags: RADIX_TREE_ITER_* flags and tag index
896 * Returns: pointer to chunk first slot, or NULL if iteration is over
898 void **radix_tree_next_chunk(struct radix_tree_root *root,
899 struct radix_tree_iter *iter, unsigned flags)
901 unsigned shift, tag = flags & RADIX_TREE_ITER_TAG_MASK;
902 struct radix_tree_node *rnode, *node;
903 unsigned long index, offset, maxindex;
905 if ((flags & RADIX_TREE_ITER_TAGGED) && !root_tag_get(root, tag))
909 * Catch next_index overflow after ~0UL. iter->index never overflows
910 * during iterating; it can be zero only at the beginning.
911 * And we cannot overflow iter->next_index in a single step,
912 * because RADIX_TREE_MAP_SHIFT < BITS_PER_LONG.
914 * This condition also used by radix_tree_next_slot() to stop
915 * contiguous iterating, and forbid swithing to the next chunk.
917 index = iter->next_index;
918 if (!index && iter->index)
922 shift = radix_tree_load_root(root, &rnode, &maxindex);
923 if (index > maxindex)
926 if (radix_tree_is_indirect_ptr(rnode)) {
927 rnode = indirect_to_ptr(rnode);
929 /* Single-slot tree */
931 iter->next_index = maxindex + 1;
933 __set_iter_shift(iter, shift);
934 return (void **)&root->rnode;
938 shift -= RADIX_TREE_MAP_SHIFT;
939 offset = index >> shift;
943 struct radix_tree_node *slot;
944 unsigned new_off = radix_tree_descend(node, &slot, offset);
946 if (new_off < offset) {
948 index &= ~((RADIX_TREE_MAP_SIZE << shift) - 1);
949 index |= offset << shift;
952 if ((flags & RADIX_TREE_ITER_TAGGED) ?
953 !tag_get(node, tag, offset) : !slot) {
955 if (flags & RADIX_TREE_ITER_CONTIG)
958 if (flags & RADIX_TREE_ITER_TAGGED)
959 offset = radix_tree_find_next_bit(
964 while (++offset < RADIX_TREE_MAP_SIZE) {
965 void *slot = node->slots[offset];
966 if (is_sibling_entry(node, slot))
971 index &= ~((RADIX_TREE_MAP_SIZE << shift) - 1);
972 index += offset << shift;
973 /* Overflow after ~0UL */
976 if (offset == RADIX_TREE_MAP_SIZE)
978 slot = rcu_dereference_raw(node->slots[offset]);
981 if ((slot == NULL) || (slot == RADIX_TREE_RETRY))
983 if (!radix_tree_is_indirect_ptr(slot))
986 node = indirect_to_ptr(slot);
987 shift -= RADIX_TREE_MAP_SHIFT;
988 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
991 /* Update the iterator state */
992 iter->index = index & ~((1 << shift) - 1);
993 iter->next_index = (index | ((RADIX_TREE_MAP_SIZE << shift) - 1)) + 1;
994 __set_iter_shift(iter, shift);
996 /* Construct iter->tags bit-mask from node->tags[tag] array */
997 if (flags & RADIX_TREE_ITER_TAGGED) {
998 unsigned tag_long, tag_bit;
1000 tag_long = offset / BITS_PER_LONG;
1001 tag_bit = offset % BITS_PER_LONG;
1002 iter->tags = node->tags[tag][tag_long] >> tag_bit;
1003 /* This never happens if RADIX_TREE_TAG_LONGS == 1 */
1004 if (tag_long < RADIX_TREE_TAG_LONGS - 1) {
1005 /* Pick tags from next element */
1007 iter->tags |= node->tags[tag][tag_long + 1] <<
1008 (BITS_PER_LONG - tag_bit);
1009 /* Clip chunk size, here only BITS_PER_LONG tags */
1010 iter->next_index = index + BITS_PER_LONG;
1014 return node->slots + offset;
1016 EXPORT_SYMBOL(radix_tree_next_chunk);
1019 * radix_tree_range_tag_if_tagged - for each item in given range set given
1020 * tag if item has another tag set
1021 * @root: radix tree root
1022 * @first_indexp: pointer to a starting index of a range to scan
1023 * @last_index: last index of a range to scan
1024 * @nr_to_tag: maximum number items to tag
1025 * @iftag: tag index to test
1026 * @settag: tag index to set if tested tag is set
1028 * This function scans range of radix tree from first_index to last_index
1029 * (inclusive). For each item in the range if iftag is set, the function sets
1030 * also settag. The function stops either after tagging nr_to_tag items or
1031 * after reaching last_index.
1033 * The tags must be set from the leaf level only and propagated back up the
1034 * path to the root. We must do this so that we resolve the full path before
1035 * setting any tags on intermediate nodes. If we set tags as we descend, then
1036 * we can get to the leaf node and find that the index that has the iftag
1037 * set is outside the range we are scanning. This reults in dangling tags and
1038 * can lead to problems with later tag operations (e.g. livelocks on lookups).
1040 * The function returns the number of leaves where the tag was set and sets
1041 * *first_indexp to the first unscanned index.
1042 * WARNING! *first_indexp can wrap if last_index is ULONG_MAX. Caller must
1043 * be prepared to handle that.
1045 unsigned long radix_tree_range_tag_if_tagged(struct radix_tree_root *root,
1046 unsigned long *first_indexp, unsigned long last_index,
1047 unsigned long nr_to_tag,
1048 unsigned int iftag, unsigned int settag)
1050 struct radix_tree_node *slot, *node = NULL;
1051 unsigned long maxindex;
1052 unsigned int shift = radix_tree_load_root(root, &slot, &maxindex);
1053 unsigned long tagged = 0;
1054 unsigned long index = *first_indexp;
1056 last_index = min(last_index, maxindex);
1057 if (index > last_index)
1061 if (!root_tag_get(root, iftag)) {
1062 *first_indexp = last_index + 1;
1065 if (!radix_tree_is_indirect_ptr(slot)) {
1066 *first_indexp = last_index + 1;
1067 root_tag_set(root, settag);
1071 node = indirect_to_ptr(slot);
1072 shift -= RADIX_TREE_MAP_SHIFT;
1075 unsigned long upindex;
1078 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
1079 offset = radix_tree_descend(node, &slot, offset);
1082 if (!tag_get(node, iftag, offset))
1084 /* Sibling slots never have tags set on them */
1085 if (radix_tree_is_indirect_ptr(slot)) {
1086 node = indirect_to_ptr(slot);
1087 shift -= RADIX_TREE_MAP_SHIFT;
1093 tag_set(node, settag, offset);
1095 slot = node->parent;
1096 /* walk back up the path tagging interior nodes */
1097 upindex = index >> shift;
1099 upindex >>= RADIX_TREE_MAP_SHIFT;
1100 offset = upindex & RADIX_TREE_MAP_MASK;
1102 /* stop if we find a node with the tag already set */
1103 if (tag_get(slot, settag, offset))
1105 tag_set(slot, settag, offset);
1106 slot = slot->parent;
1110 /* Go to next item at level determined by 'shift' */
1111 index = ((index >> shift) + 1) << shift;
1112 /* Overflow can happen when last_index is ~0UL... */
1113 if (index > last_index || !index)
1115 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
1116 while (offset == 0) {
1118 * We've fully scanned this node. Go up. Because
1119 * last_index is guaranteed to be in the tree, what
1120 * we do below cannot wander astray.
1122 node = node->parent;
1123 shift += RADIX_TREE_MAP_SHIFT;
1124 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
1126 if (is_sibling_entry(node, node->slots[offset]))
1128 if (tagged >= nr_to_tag)
1132 * We need not to tag the root tag if there is no tag which is set with
1133 * settag within the range from *first_indexp to last_index.
1136 root_tag_set(root, settag);
1137 *first_indexp = index;
1141 EXPORT_SYMBOL(radix_tree_range_tag_if_tagged);
1144 * radix_tree_gang_lookup - perform multiple lookup on a radix tree
1145 * @root: radix tree root
1146 * @results: where the results of the lookup are placed
1147 * @first_index: start the lookup from this key
1148 * @max_items: place up to this many items at *results
1150 * Performs an index-ascending scan of the tree for present items. Places
1151 * them at *@results and returns the number of items which were placed at
1154 * The implementation is naive.
1156 * Like radix_tree_lookup, radix_tree_gang_lookup may be called under
1157 * rcu_read_lock. In this case, rather than the returned results being
1158 * an atomic snapshot of the tree at a single point in time, the
1159 * semantics of an RCU protected gang lookup are as though multiple
1160 * radix_tree_lookups have been issued in individual locks, and results
1161 * stored in 'results'.
1164 radix_tree_gang_lookup(struct radix_tree_root *root, void **results,
1165 unsigned long first_index, unsigned int max_items)
1167 struct radix_tree_iter iter;
1169 unsigned int ret = 0;
1171 if (unlikely(!max_items))
1174 radix_tree_for_each_slot(slot, root, &iter, first_index) {
1175 results[ret] = rcu_dereference_raw(*slot);
1178 if (radix_tree_is_indirect_ptr(results[ret])) {
1179 slot = radix_tree_iter_retry(&iter);
1182 if (++ret == max_items)
1188 EXPORT_SYMBOL(radix_tree_gang_lookup);
1191 * radix_tree_gang_lookup_slot - perform multiple slot lookup on radix tree
1192 * @root: radix tree root
1193 * @results: where the results of the lookup are placed
1194 * @indices: where their indices should be placed (but usually NULL)
1195 * @first_index: start the lookup from this key
1196 * @max_items: place up to this many items at *results
1198 * Performs an index-ascending scan of the tree for present items. Places
1199 * their slots at *@results and returns the number of items which were
1200 * placed at *@results.
1202 * The implementation is naive.
1204 * Like radix_tree_gang_lookup as far as RCU and locking goes. Slots must
1205 * be dereferenced with radix_tree_deref_slot, and if using only RCU
1206 * protection, radix_tree_deref_slot may fail requiring a retry.
1209 radix_tree_gang_lookup_slot(struct radix_tree_root *root,
1210 void ***results, unsigned long *indices,
1211 unsigned long first_index, unsigned int max_items)
1213 struct radix_tree_iter iter;
1215 unsigned int ret = 0;
1217 if (unlikely(!max_items))
1220 radix_tree_for_each_slot(slot, root, &iter, first_index) {
1221 results[ret] = slot;
1223 indices[ret] = iter.index;
1224 if (++ret == max_items)
1230 EXPORT_SYMBOL(radix_tree_gang_lookup_slot);
1233 * radix_tree_gang_lookup_tag - perform multiple lookup on a radix tree
1235 * @root: radix tree root
1236 * @results: where the results of the lookup are placed
1237 * @first_index: start the lookup from this key
1238 * @max_items: place up to this many items at *results
1239 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1241 * Performs an index-ascending scan of the tree for present items which
1242 * have the tag indexed by @tag set. Places the items at *@results and
1243 * returns the number of items which were placed at *@results.
1246 radix_tree_gang_lookup_tag(struct radix_tree_root *root, void **results,
1247 unsigned long first_index, unsigned int max_items,
1250 struct radix_tree_iter iter;
1252 unsigned int ret = 0;
1254 if (unlikely(!max_items))
1257 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1258 results[ret] = rcu_dereference_raw(*slot);
1261 if (radix_tree_is_indirect_ptr(results[ret])) {
1262 slot = radix_tree_iter_retry(&iter);
1265 if (++ret == max_items)
1271 EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
1274 * radix_tree_gang_lookup_tag_slot - perform multiple slot lookup on a
1275 * radix tree based on a tag
1276 * @root: radix tree root
1277 * @results: where the results of the lookup are placed
1278 * @first_index: start the lookup from this key
1279 * @max_items: place up to this many items at *results
1280 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1282 * Performs an index-ascending scan of the tree for present items which
1283 * have the tag indexed by @tag set. Places the slots at *@results and
1284 * returns the number of slots which were placed at *@results.
1287 radix_tree_gang_lookup_tag_slot(struct radix_tree_root *root, void ***results,
1288 unsigned long first_index, unsigned int max_items,
1291 struct radix_tree_iter iter;
1293 unsigned int ret = 0;
1295 if (unlikely(!max_items))
1298 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1299 results[ret] = slot;
1300 if (++ret == max_items)
1306 EXPORT_SYMBOL(radix_tree_gang_lookup_tag_slot);
1308 #if defined(CONFIG_SHMEM) && defined(CONFIG_SWAP)
1309 #include <linux/sched.h> /* for cond_resched() */
1311 struct locate_info {
1312 unsigned long found_index;
1317 * This linear search is at present only useful to shmem_unuse_inode().
1319 static unsigned long __locate(struct radix_tree_node *slot, void *item,
1320 unsigned long index, struct locate_info *info)
1325 shift = slot->height * RADIX_TREE_MAP_SHIFT;
1328 shift -= RADIX_TREE_MAP_SHIFT;
1330 for (i = (index >> shift) & RADIX_TREE_MAP_MASK;
1331 i < RADIX_TREE_MAP_SIZE;
1332 i++, index += (1UL << shift)) {
1333 struct radix_tree_node *node =
1334 rcu_dereference_raw(slot->slots[i]);
1335 if (node == RADIX_TREE_RETRY)
1337 if (!radix_tree_is_indirect_ptr(node)) {
1339 info->found_index = index;
1345 node = indirect_to_ptr(node);
1346 if (is_sibling_entry(slot, node))
1351 if (i == RADIX_TREE_MAP_SIZE)
1356 if ((index == 0) && (i == RADIX_TREE_MAP_SIZE))
1362 * radix_tree_locate_item - search through radix tree for item
1363 * @root: radix tree root
1364 * @item: item to be found
1366 * Returns index where item was found, or -1 if not found.
1367 * Caller must hold no lock (since this time-consuming function needs
1368 * to be preemptible), and must check afterwards if item is still there.
1370 unsigned long radix_tree_locate_item(struct radix_tree_root *root, void *item)
1372 struct radix_tree_node *node;
1373 unsigned long max_index;
1374 unsigned long cur_index = 0;
1375 struct locate_info info = {
1382 node = rcu_dereference_raw(root->rnode);
1383 if (!radix_tree_is_indirect_ptr(node)) {
1386 info.found_index = 0;
1390 node = indirect_to_ptr(node);
1392 max_index = node_maxindex(node);
1393 if (cur_index > max_index) {
1398 cur_index = __locate(node, item, cur_index, &info);
1401 } while (!info.stop && cur_index <= max_index);
1403 return info.found_index;
1406 unsigned long radix_tree_locate_item(struct radix_tree_root *root, void *item)
1410 #endif /* CONFIG_SHMEM && CONFIG_SWAP */
1413 * radix_tree_shrink - shrink height of a radix tree to minimal
1414 * @root radix tree root
1416 static inline void radix_tree_shrink(struct radix_tree_root *root)
1418 /* try to shrink tree height */
1419 while (root->height > 0) {
1420 struct radix_tree_node *to_free = root->rnode;
1421 struct radix_tree_node *slot;
1423 BUG_ON(!radix_tree_is_indirect_ptr(to_free));
1424 to_free = indirect_to_ptr(to_free);
1427 * The candidate node has more than one child, or its child
1428 * is not at the leftmost slot, or it is a multiorder entry,
1431 if (to_free->count != 1)
1433 slot = to_free->slots[0];
1436 if (!radix_tree_is_indirect_ptr(slot) && (root->height > 1))
1439 if (radix_tree_is_indirect_ptr(slot)) {
1440 slot = indirect_to_ptr(slot);
1441 slot->parent = NULL;
1442 slot = ptr_to_indirect(slot);
1446 * We don't need rcu_assign_pointer(), since we are simply
1447 * moving the node from one part of the tree to another: if it
1448 * was safe to dereference the old pointer to it
1449 * (to_free->slots[0]), it will be safe to dereference the new
1450 * one (root->rnode) as far as dependent read barriers go.
1456 * We have a dilemma here. The node's slot[0] must not be
1457 * NULLed in case there are concurrent lookups expecting to
1458 * find the item. However if this was a bottom-level node,
1459 * then it may be subject to the slot pointer being visible
1460 * to callers dereferencing it. If item corresponding to
1461 * slot[0] is subsequently deleted, these callers would expect
1462 * their slot to become empty sooner or later.
1464 * For example, lockless pagecache will look up a slot, deref
1465 * the page pointer, and if the page has 0 refcount it means it
1466 * was concurrently deleted from pagecache so try the deref
1467 * again. Fortunately there is already a requirement for logic
1468 * to retry the entire slot lookup -- the indirect pointer
1469 * problem (replacing direct root node with an indirect pointer
1470 * also results in a stale slot). So tag the slot as indirect
1471 * to force callers to retry.
1473 if (!radix_tree_is_indirect_ptr(slot))
1474 to_free->slots[0] = RADIX_TREE_RETRY;
1476 radix_tree_node_free(to_free);
1481 * __radix_tree_delete_node - try to free node after clearing a slot
1482 * @root: radix tree root
1483 * @node: node containing @index
1485 * After clearing the slot at @index in @node from radix tree
1486 * rooted at @root, call this function to attempt freeing the
1487 * node and shrinking the tree.
1489 * Returns %true if @node was freed, %false otherwise.
1491 bool __radix_tree_delete_node(struct radix_tree_root *root,
1492 struct radix_tree_node *node)
1494 bool deleted = false;
1497 struct radix_tree_node *parent;
1500 if (node == indirect_to_ptr(root->rnode)) {
1501 radix_tree_shrink(root);
1502 if (root->height == 0)
1508 parent = node->parent;
1510 parent->slots[node->offset] = NULL;
1513 root_tag_clear_all(root);
1518 radix_tree_node_free(node);
1527 static inline void delete_sibling_entries(struct radix_tree_node *node,
1528 void *ptr, unsigned offset)
1530 #ifdef CONFIG_RADIX_TREE_MULTIORDER
1532 for (i = 1; offset + i < RADIX_TREE_MAP_SIZE; i++) {
1533 if (node->slots[offset + i] != ptr)
1535 node->slots[offset + i] = NULL;
1542 * radix_tree_delete_item - delete an item from a radix tree
1543 * @root: radix tree root
1545 * @item: expected item
1547 * Remove @item at @index from the radix tree rooted at @root.
1549 * Returns the address of the deleted item, or NULL if it was not present
1550 * or the entry at the given @index was not @item.
1552 void *radix_tree_delete_item(struct radix_tree_root *root,
1553 unsigned long index, void *item)
1555 struct radix_tree_node *node;
1556 unsigned int offset;
1561 entry = __radix_tree_lookup(root, index, &node, &slot);
1565 if (item && entry != item)
1569 root_tag_clear_all(root);
1574 offset = get_slot_offset(node, slot);
1577 * Clear all tags associated with the item to be deleted.
1578 * This way of doing it would be inefficient, but seldom is any set.
1580 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
1581 if (tag_get(node, tag, offset))
1582 radix_tree_tag_clear(root, index, tag);
1585 delete_sibling_entries(node, ptr_to_indirect(slot), offset);
1586 node->slots[offset] = NULL;
1589 __radix_tree_delete_node(root, node);
1593 EXPORT_SYMBOL(radix_tree_delete_item);
1596 * radix_tree_delete - delete an item from a radix tree
1597 * @root: radix tree root
1600 * Remove the item at @index from the radix tree rooted at @root.
1602 * Returns the address of the deleted item, or NULL if it was not present.
1604 void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
1606 return radix_tree_delete_item(root, index, NULL);
1608 EXPORT_SYMBOL(radix_tree_delete);
1611 * radix_tree_tagged - test whether any items in the tree are tagged
1612 * @root: radix tree root
1615 int radix_tree_tagged(struct radix_tree_root *root, unsigned int tag)
1617 return root_tag_get(root, tag);
1619 EXPORT_SYMBOL(radix_tree_tagged);
1622 radix_tree_node_ctor(void *arg)
1624 struct radix_tree_node *node = arg;
1626 memset(node, 0, sizeof(*node));
1627 INIT_LIST_HEAD(&node->private_list);
1630 static __init unsigned long __maxindex(unsigned int height)
1632 unsigned int width = height * RADIX_TREE_MAP_SHIFT;
1633 int shift = RADIX_TREE_INDEX_BITS - width;
1637 if (shift >= BITS_PER_LONG)
1639 return ~0UL >> shift;
1642 static __init void radix_tree_init_maxindex(void)
1646 for (i = 0; i < ARRAY_SIZE(height_to_maxindex); i++)
1647 height_to_maxindex[i] = __maxindex(i);
1650 static int radix_tree_callback(struct notifier_block *nfb,
1651 unsigned long action, void *hcpu)
1653 int cpu = (long)hcpu;
1654 struct radix_tree_preload *rtp;
1655 struct radix_tree_node *node;
1657 /* Free per-cpu pool of preloaded nodes */
1658 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
1659 rtp = &per_cpu(radix_tree_preloads, cpu);
1662 rtp->nodes = node->private_data;
1663 kmem_cache_free(radix_tree_node_cachep, node);
1670 void __init radix_tree_init(void)
1672 radix_tree_node_cachep = kmem_cache_create("radix_tree_node",
1673 sizeof(struct radix_tree_node), 0,
1674 SLAB_PANIC | SLAB_RECLAIM_ACCOUNT,
1675 radix_tree_node_ctor);
1676 radix_tree_init_maxindex();
1677 hotcpu_notifier(radix_tree_callback, 0);