1 /*
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 *
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public License as
10 * published by the Free Software Foundation; either version 2, or (at
11 * your option) any later version.
12 *
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 */
23 #include <linux/errno.h>
24 #include <linux/init.h>
25 #include <linux/kernel.h>
26 #include <linux/export.h>
27 #include <linux/radix-tree.h>
28 #include <linux/percpu.h>
29 #include <linux/slab.h>
30 #include <linux/notifier.h>
31 #include <linux/cpu.h>
32 #include <linux/string.h>
33 #include <linux/bitops.h>
34 #include <linux/rcupdate.h>
35 #include <linux/hardirq.h> /* in_interrupt() */
38 #ifdef __KERNEL__
39 #define RADIX_TREE_MAP_SHIFT (CONFIG_BASE_SMALL ? 4 : 6)
40 #else
41 #define RADIX_TREE_MAP_SHIFT 3 /* For more stressful testing */
42 #endif
44 #define RADIX_TREE_MAP_SIZE (1UL << RADIX_TREE_MAP_SHIFT)
45 #define RADIX_TREE_MAP_MASK (RADIX_TREE_MAP_SIZE-1)
47 #define RADIX_TREE_TAG_LONGS \
48 ((RADIX_TREE_MAP_SIZE + BITS_PER_LONG - 1) / BITS_PER_LONG)
50 struct radix_tree_node {
51 unsigned int height; /* Height from the bottom */
52 unsigned int count;
53 union {
54 struct radix_tree_node *parent; /* Used when ascending tree */
55 struct rcu_head rcu_head; /* Used when freeing node */
56 };
57 void __rcu *slots[RADIX_TREE_MAP_SIZE];
58 unsigned long tags[RADIX_TREE_MAX_TAGS][RADIX_TREE_TAG_LONGS];
59 };
61 #define RADIX_TREE_INDEX_BITS (8 /* CHAR_BIT */ * sizeof(unsigned long))
62 #define RADIX_TREE_MAX_PATH (DIV_ROUND_UP(RADIX_TREE_INDEX_BITS, \
63 RADIX_TREE_MAP_SHIFT))
65 /*
66 * The height_to_maxindex array needs to be one deeper than the maximum
67 * path as height 0 holds only 1 entry.
68 */
69 static unsigned long height_to_maxindex[RADIX_TREE_MAX_PATH + 1] __read_mostly;
71 /*
72 * Radix tree node cache.
73 */
74 static struct kmem_cache *radix_tree_node_cachep;
76 /*
77 * The radix tree is variable-height, so an insert operation not only has
78 * to build the branch to its corresponding item, it also has to build the
79 * branch to existing items if the size has to be increased (by
80 * radix_tree_extend).
81 *
82 * The worst case is a zero height tree with just a single item at index 0,
83 * and then inserting an item at index ULONG_MAX. This requires 2 new branches
84 * of RADIX_TREE_MAX_PATH size to be created, with only the root node shared.
85 * Hence:
86 */
87 #define RADIX_TREE_PRELOAD_SIZE (RADIX_TREE_MAX_PATH * 2 - 1)
89 /*
90 * Per-cpu pool of preloaded nodes
91 */
92 struct radix_tree_preload {
93 int nr;
94 struct radix_tree_node *nodes[RADIX_TREE_PRELOAD_SIZE];
95 };
96 static DEFINE_PER_CPU(struct radix_tree_preload, radix_tree_preloads) = { 0, };
98 static inline void *ptr_to_indirect(void *ptr)
99 {
100 return (void *)((unsigned long)ptr | RADIX_TREE_INDIRECT_PTR);
101 }
103 static inline void *indirect_to_ptr(void *ptr)
104 {
105 return (void *)((unsigned long)ptr & ~RADIX_TREE_INDIRECT_PTR);
106 }
108 static inline gfp_t root_gfp_mask(struct radix_tree_root *root)
109 {
110 return root->gfp_mask & __GFP_BITS_MASK;
111 }
113 static inline void tag_set(struct radix_tree_node *node, unsigned int tag,
114 int offset)
115 {
116 __set_bit(offset, node->tags[tag]);
117 }
119 static inline void tag_clear(struct radix_tree_node *node, unsigned int tag,
120 int offset)
121 {
122 __clear_bit(offset, node->tags[tag]);
123 }
125 static inline int tag_get(struct radix_tree_node *node, unsigned int tag,
126 int offset)
127 {
128 return test_bit(offset, node->tags[tag]);
129 }
131 static inline void root_tag_set(struct radix_tree_root *root, unsigned int tag)
132 {
133 root->gfp_mask |= (__force gfp_t)(1 << (tag + __GFP_BITS_SHIFT));
134 }
136 static inline void root_tag_clear(struct radix_tree_root *root, unsigned int tag)
137 {
138 root->gfp_mask &= (__force gfp_t)~(1 << (tag + __GFP_BITS_SHIFT));
139 }
141 static inline void root_tag_clear_all(struct radix_tree_root *root)
142 {
143 root->gfp_mask &= __GFP_BITS_MASK;
144 }
146 static inline int root_tag_get(struct radix_tree_root *root, unsigned int tag)
147 {
148 return (__force unsigned)root->gfp_mask & (1 << (tag + __GFP_BITS_SHIFT));
149 }
151 /*
152 * Returns 1 if any slot in the node has this tag set.
153 * Otherwise returns 0.
154 */
155 static inline int any_tag_set(struct radix_tree_node *node, unsigned int tag)
156 {
157 int idx;
158 for (idx = 0; idx < RADIX_TREE_TAG_LONGS; idx++) {
159 if (node->tags[tag][idx])
160 return 1;
161 }
162 return 0;
163 }
165 /**
166 * radix_tree_find_next_bit - find the next set bit in a memory region
167 *
168 * @addr: The address to base the search on
169 * @size: The bitmap size in bits
170 * @offset: The bitnumber to start searching at
171 *
172 * Unrollable variant of find_next_bit() for constant size arrays.
173 * Tail bits starting from size to roundup(size, BITS_PER_LONG) must be zero.
174 * Returns next bit offset, or size if nothing found.
175 */
176 static __always_inline unsigned long
177 radix_tree_find_next_bit(const unsigned long *addr,
178 unsigned long size, unsigned long offset)
179 {
180 if (!__builtin_constant_p(size))
181 return find_next_bit(addr, size, offset);
183 if (offset < size) {
184 unsigned long tmp;
186 addr += offset / BITS_PER_LONG;
187 tmp = *addr >> (offset % BITS_PER_LONG);
188 if (tmp)
189 return __ffs(tmp) + offset;
190 offset = (offset + BITS_PER_LONG) & ~(BITS_PER_LONG - 1);
191 while (offset < size) {
192 tmp = *++addr;
193 if (tmp)
194 return __ffs(tmp) + offset;
195 offset += BITS_PER_LONG;
196 }
197 }
198 return size;
199 }
201 /*
202 * This assumes that the caller has performed appropriate preallocation, and
203 * that the caller has pinned this thread of control to the current CPU.
204 */
205 static struct radix_tree_node *
206 radix_tree_node_alloc(struct radix_tree_root *root)
207 {
208 struct radix_tree_node *ret = NULL;
209 gfp_t gfp_mask = root_gfp_mask(root);
211 /*
212 * Preload code isn't irq safe and it doesn't make sence to use
213 * preloading in the interrupt anyway as all the allocations have to
214 * be atomic. So just do normal allocation when in interrupt.
215 */
216 if (!(gfp_mask & __GFP_WAIT) && !in_interrupt()) {
217 struct radix_tree_preload *rtp;
219 /*
220 * Provided the caller has preloaded here, we will always
221 * succeed in getting a node here (and never reach
222 * kmem_cache_alloc)
223 */
224 rtp = &__get_cpu_var(radix_tree_preloads);
225 if (rtp->nr) {
226 ret = rtp->nodes[rtp->nr - 1];
227 rtp->nodes[rtp->nr - 1] = NULL;
228 rtp->nr--;
229 }
230 }
231 if (ret == NULL)
232 ret = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
234 BUG_ON(radix_tree_is_indirect_ptr(ret));
235 return ret;
236 }
238 static void radix_tree_node_rcu_free(struct rcu_head *head)
239 {
240 struct radix_tree_node *node =
241 container_of(head, struct radix_tree_node, rcu_head);
242 int i;
244 /*
245 * must only free zeroed nodes into the slab. radix_tree_shrink
246 * can leave us with a non-NULL entry in the first slot, so clear
247 * that here to make sure.
248 */
249 for (i = 0; i < RADIX_TREE_MAX_TAGS; i++)
250 tag_clear(node, i, 0);
252 node->slots[0] = NULL;
253 node->count = 0;
255 kmem_cache_free(radix_tree_node_cachep, node);
256 }
258 static inline void
259 radix_tree_node_free(struct radix_tree_node *node)
260 {
261 call_rcu(&node->rcu_head, radix_tree_node_rcu_free);
262 }
264 /*
265 * Load up this CPU's radix_tree_node buffer with sufficient objects to
266 * ensure that the addition of a single element in the tree cannot fail. On
267 * success, return zero, with preemption disabled. On error, return -ENOMEM
268 * with preemption not disabled.
269 *
270 * To make use of this facility, the radix tree must be initialised without
271 * __GFP_WAIT being passed to INIT_RADIX_TREE().
272 */
273 static int __radix_tree_preload(gfp_t gfp_mask)
274 {
275 struct radix_tree_preload *rtp;
276 struct radix_tree_node *node;
277 int ret = -ENOMEM;
279 preempt_disable();
280 rtp = &__get_cpu_var(radix_tree_preloads);
281 while (rtp->nr < ARRAY_SIZE(rtp->nodes)) {
282 preempt_enable();
283 node = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
284 if (node == NULL)
285 goto out;
286 preempt_disable();
287 rtp = &__get_cpu_var(radix_tree_preloads);
288 if (rtp->nr < ARRAY_SIZE(rtp->nodes))
289 rtp->nodes[rtp->nr++] = node;
290 else
291 kmem_cache_free(radix_tree_node_cachep, node);
292 }
293 ret = 0;
294 out:
295 return ret;
296 }
298 /*
299 * Load up this CPU's radix_tree_node buffer with sufficient objects to
300 * ensure that the addition of a single element in the tree cannot fail. On
301 * success, return zero, with preemption disabled. On error, return -ENOMEM
302 * with preemption not disabled.
303 *
304 * To make use of this facility, the radix tree must be initialised without
305 * __GFP_WAIT being passed to INIT_RADIX_TREE().
306 */
307 int radix_tree_preload(gfp_t gfp_mask)
308 {
309 /* Warn on non-sensical use... */
310 WARN_ON_ONCE(!(gfp_mask & __GFP_WAIT));
311 return __radix_tree_preload(gfp_mask);
312 }
313 EXPORT_SYMBOL(radix_tree_preload);
315 /*
316 * The same as above function, except we don't guarantee preloading happens.
317 * We do it, if we decide it helps. On success, return zero with preemption
318 * disabled. On error, return -ENOMEM with preemption not disabled.
319 */
320 int radix_tree_maybe_preload(gfp_t gfp_mask)
321 {
322 if (gfp_mask & __GFP_WAIT)
323 return __radix_tree_preload(gfp_mask);
324 /* Preloading doesn't help anything with this gfp mask, skip it */
325 preempt_disable();
326 return 0;
327 }
328 EXPORT_SYMBOL(radix_tree_maybe_preload);
330 /*
331 * Return the maximum key which can be store into a
332 * radix tree with height HEIGHT.
333 */
334 static inline unsigned long radix_tree_maxindex(unsigned int height)
335 {
336 return height_to_maxindex[height];
337 }
339 /*
340 * Extend a radix tree so it can store key @index.
341 */
342 static int radix_tree_extend(struct radix_tree_root *root, unsigned long index)
343 {
344 struct radix_tree_node *node;
345 struct radix_tree_node *slot;
346 unsigned int height;
347 int tag;
349 /* Figure out what the height should be. */
350 height = root->height + 1;
351 while (index > radix_tree_maxindex(height))
352 height++;
354 if (root->rnode == NULL) {
355 root->height = height;
356 goto out;
357 }
359 do {
360 unsigned int newheight;
361 if (!(node = radix_tree_node_alloc(root)))
362 return -ENOMEM;
364 /* Propagate the aggregated tag info into the new root */
365 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
366 if (root_tag_get(root, tag))
367 tag_set(node, tag, 0);
368 }
370 /* Increase the height. */
371 newheight = root->height+1;
372 node->height = newheight;
373 node->count = 1;
374 node->parent = NULL;
375 slot = root->rnode;
376 if (newheight > 1) {
377 slot = indirect_to_ptr(slot);
378 slot->parent = node;
379 }
380 node->slots[0] = slot;
381 node = ptr_to_indirect(node);
382 rcu_assign_pointer(root->rnode, node);
383 root->height = newheight;
384 } while (height > root->height);
385 out:
386 return 0;
387 }
389 /**
390 * radix_tree_insert - insert into a radix tree
391 * @root: radix tree root
392 * @index: index key
393 * @item: item to insert
394 *
395 * Insert an item into the radix tree at position @index.
396 */
397 int radix_tree_insert(struct radix_tree_root *root,
398 unsigned long index, void *item)
399 {
400 struct radix_tree_node *node = NULL, *slot;
401 unsigned int height, shift;
402 int offset;
403 int error;
405 BUG_ON(radix_tree_is_indirect_ptr(item));
407 /* Make sure the tree is high enough. */
408 if (index > radix_tree_maxindex(root->height)) {
409 error = radix_tree_extend(root, index);
410 if (error)
411 return error;
412 }
414 slot = indirect_to_ptr(root->rnode);
416 height = root->height;
417 shift = (height-1) * RADIX_TREE_MAP_SHIFT;
419 offset = 0; /* uninitialised var warning */
420 while (height > 0) {
421 if (slot == NULL) {
422 /* Have to add a child node. */
423 if (!(slot = radix_tree_node_alloc(root)))
424 return -ENOMEM;
425 slot->height = height;
426 slot->parent = node;
427 if (node) {
428 rcu_assign_pointer(node->slots[offset], slot);
429 node->count++;
430 } else
431 rcu_assign_pointer(root->rnode, ptr_to_indirect(slot));
432 }
434 /* Go a level down */
435 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
436 node = slot;
437 slot = node->slots[offset];
438 shift -= RADIX_TREE_MAP_SHIFT;
439 height--;
440 }
442 if (slot != NULL)
443 return -EEXIST;
445 if (node) {
446 node->count++;
447 rcu_assign_pointer(node->slots[offset], item);
448 BUG_ON(tag_get(node, 0, offset));
449 BUG_ON(tag_get(node, 1, offset));
450 } else {
451 rcu_assign_pointer(root->rnode, item);
452 BUG_ON(root_tag_get(root, 0));
453 BUG_ON(root_tag_get(root, 1));
454 }
456 return 0;
457 }
458 EXPORT_SYMBOL(radix_tree_insert);
460 /*
461 * is_slot == 1 : search for the slot.
462 * is_slot == 0 : search for the node.
463 */
464 static void *radix_tree_lookup_element(struct radix_tree_root *root,
465 unsigned long index, int is_slot)
466 {
467 unsigned int height, shift;
468 struct radix_tree_node *node, **slot;
470 node = rcu_dereference_raw(root->rnode);
471 if (node == NULL)
472 return NULL;
474 if (!radix_tree_is_indirect_ptr(node)) {
475 if (index > 0)
476 return NULL;
477 return is_slot ? (void *)&root->rnode : node;
478 }
479 node = indirect_to_ptr(node);
481 height = node->height;
482 if (index > radix_tree_maxindex(height))
483 return NULL;
485 shift = (height-1) * RADIX_TREE_MAP_SHIFT;
487 do {
488 slot = (struct radix_tree_node **)
489 (node->slots + ((index>>shift) & RADIX_TREE_MAP_MASK));
490 node = rcu_dereference_raw(*slot);
491 if (node == NULL)
492 return NULL;
494 shift -= RADIX_TREE_MAP_SHIFT;
495 height--;
496 } while (height > 0);
498 return is_slot ? (void *)slot : indirect_to_ptr(node);
499 }
501 /**
502 * radix_tree_lookup_slot - lookup a slot in a radix tree
503 * @root: radix tree root
504 * @index: index key
505 *
506 * Returns: the slot corresponding to the position @index in the
507 * radix tree @root. This is useful for update-if-exists operations.
508 *
509 * This function can be called under rcu_read_lock iff the slot is not
510 * modified by radix_tree_replace_slot, otherwise it must be called
511 * exclusive from other writers. Any dereference of the slot must be done
512 * using radix_tree_deref_slot.
513 */
514 void **radix_tree_lookup_slot(struct radix_tree_root *root, unsigned long index)
515 {
516 return (void **)radix_tree_lookup_element(root, index, 1);
517 }
518 EXPORT_SYMBOL(radix_tree_lookup_slot);
520 /**
521 * radix_tree_lookup - perform lookup operation on a radix tree
522 * @root: radix tree root
523 * @index: index key
524 *
525 * Lookup the item at the position @index in the radix tree @root.
526 *
527 * This function can be called under rcu_read_lock, however the caller
528 * must manage lifetimes of leaf nodes (eg. RCU may also be used to free
529 * them safely). No RCU barriers are required to access or modify the
530 * returned item, however.
531 */
532 void *radix_tree_lookup(struct radix_tree_root *root, unsigned long index)
533 {
534 return radix_tree_lookup_element(root, index, 0);
535 }
536 EXPORT_SYMBOL(radix_tree_lookup);
538 /**
539 * radix_tree_tag_set - set a tag on a radix tree node
540 * @root: radix tree root
541 * @index: index key
542 * @tag: tag index
543 *
544 * Set the search tag (which must be < RADIX_TREE_MAX_TAGS)
545 * corresponding to @index in the radix tree. From
546 * the root all the way down to the leaf node.
547 *
548 * Returns the address of the tagged item. Setting a tag on a not-present
549 * item is a bug.
550 */
551 void *radix_tree_tag_set(struct radix_tree_root *root,
552 unsigned long index, unsigned int tag)
553 {
554 unsigned int height, shift;
555 struct radix_tree_node *slot;
557 height = root->height;
558 BUG_ON(index > radix_tree_maxindex(height));
560 slot = indirect_to_ptr(root->rnode);
561 shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
563 while (height > 0) {
564 int offset;
566 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
567 if (!tag_get(slot, tag, offset))
568 tag_set(slot, tag, offset);
569 slot = slot->slots[offset];
570 BUG_ON(slot == NULL);
571 shift -= RADIX_TREE_MAP_SHIFT;
572 height--;
573 }
575 /* set the root's tag bit */
576 if (slot && !root_tag_get(root, tag))
577 root_tag_set(root, tag);
579 return slot;
580 }
581 EXPORT_SYMBOL(radix_tree_tag_set);
583 /**
584 * radix_tree_tag_clear - clear a tag on a radix tree node
585 * @root: radix tree root
586 * @index: index key
587 * @tag: tag index
588 *
589 * Clear the search tag (which must be < RADIX_TREE_MAX_TAGS)
590 * corresponding to @index in the radix tree. If
591 * this causes the leaf node to have no tags set then clear the tag in the
592 * next-to-leaf node, etc.
593 *
594 * Returns the address of the tagged item on success, else NULL. ie:
595 * has the same return value and semantics as radix_tree_lookup().
596 */
597 void *radix_tree_tag_clear(struct radix_tree_root *root,
598 unsigned long index, unsigned int tag)
599 {
600 struct radix_tree_node *node = NULL;
601 struct radix_tree_node *slot = NULL;
602 unsigned int height, shift;
603 int uninitialized_var(offset);
605 height = root->height;
606 if (index > radix_tree_maxindex(height))
607 goto out;
609 shift = height * RADIX_TREE_MAP_SHIFT;
610 slot = indirect_to_ptr(root->rnode);
612 while (shift) {
613 if (slot == NULL)
614 goto out;
616 shift -= RADIX_TREE_MAP_SHIFT;
617 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
618 node = slot;
619 slot = slot->slots[offset];
620 }
622 if (slot == NULL)
623 goto out;
625 while (node) {
626 if (!tag_get(node, tag, offset))
627 goto out;
628 tag_clear(node, tag, offset);
629 if (any_tag_set(node, tag))
630 goto out;
632 index >>= RADIX_TREE_MAP_SHIFT;
633 offset = index & RADIX_TREE_MAP_MASK;
634 node = node->parent;
635 }
637 /* clear the root's tag bit */
638 if (root_tag_get(root, tag))
639 root_tag_clear(root, tag);
641 out:
642 return slot;
643 }
644 EXPORT_SYMBOL(radix_tree_tag_clear);
646 /**
647 * radix_tree_tag_get - get a tag on a radix tree node
648 * @root: radix tree root
649 * @index: index key
650 * @tag: tag index (< RADIX_TREE_MAX_TAGS)
651 *
652 * Return values:
653 *
654 * 0: tag not present or not set
655 * 1: tag set
656 *
657 * Note that the return value of this function may not be relied on, even if
658 * the RCU lock is held, unless tag modification and node deletion are excluded
659 * from concurrency.
660 */
661 int radix_tree_tag_get(struct radix_tree_root *root,
662 unsigned long index, unsigned int tag)
663 {
664 unsigned int height, shift;
665 struct radix_tree_node *node;
667 /* check the root's tag bit */
668 if (!root_tag_get(root, tag))
669 return 0;
671 node = rcu_dereference_raw(root->rnode);
672 if (node == NULL)
673 return 0;
675 if (!radix_tree_is_indirect_ptr(node))
676 return (index == 0);
677 node = indirect_to_ptr(node);
679 height = node->height;
680 if (index > radix_tree_maxindex(height))
681 return 0;
683 shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
685 for ( ; ; ) {
686 int offset;
688 if (node == NULL)
689 return 0;
691 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
692 if (!tag_get(node, tag, offset))
693 return 0;
694 if (height == 1)
695 return 1;
696 node = rcu_dereference_raw(node->slots[offset]);
697 shift -= RADIX_TREE_MAP_SHIFT;
698 height--;
699 }
700 }
701 EXPORT_SYMBOL(radix_tree_tag_get);
703 /**
704 * radix_tree_next_chunk - find next chunk of slots for iteration
705 *
706 * @root: radix tree root
707 * @iter: iterator state
708 * @flags: RADIX_TREE_ITER_* flags and tag index
709 * Returns: pointer to chunk first slot, or NULL if iteration is over
710 */
711 void **radix_tree_next_chunk(struct radix_tree_root *root,
712 struct radix_tree_iter *iter, unsigned flags)
713 {
714 unsigned shift, tag = flags & RADIX_TREE_ITER_TAG_MASK;
715 struct radix_tree_node *rnode, *node;
716 unsigned long index, offset;
718 if ((flags & RADIX_TREE_ITER_TAGGED) && !root_tag_get(root, tag))
719 return NULL;
721 /*
722 * Catch next_index overflow after ~0UL. iter->index never overflows
723 * during iterating; it can be zero only at the beginning.
724 * And we cannot overflow iter->next_index in a single step,
725 * because RADIX_TREE_MAP_SHIFT < BITS_PER_LONG.
726 *
727 * This condition also used by radix_tree_next_slot() to stop
728 * contiguous iterating, and forbid swithing to the next chunk.
729 */
730 index = iter->next_index;
731 if (!index && iter->index)
732 return NULL;
734 rnode = rcu_dereference_raw(root->rnode);
735 if (radix_tree_is_indirect_ptr(rnode)) {
736 rnode = indirect_to_ptr(rnode);
737 } else if (rnode && !index) {
738 /* Single-slot tree */
739 iter->index = 0;
740 iter->next_index = 1;
741 iter->tags = 1;
742 return (void **)&root->rnode;
743 } else
744 return NULL;
746 restart:
747 shift = (rnode->height - 1) * RADIX_TREE_MAP_SHIFT;
748 offset = index >> shift;
750 /* Index outside of the tree */
751 if (offset >= RADIX_TREE_MAP_SIZE)
752 return NULL;
754 node = rnode;
755 while (1) {
756 if ((flags & RADIX_TREE_ITER_TAGGED) ?
757 !test_bit(offset, node->tags[tag]) :
758 !node->slots[offset]) {
759 /* Hole detected */
760 if (flags & RADIX_TREE_ITER_CONTIG)
761 return NULL;
763 if (flags & RADIX_TREE_ITER_TAGGED)
764 offset = radix_tree_find_next_bit(
765 node->tags[tag],
766 RADIX_TREE_MAP_SIZE,
767 offset + 1);
768 else
769 while (++offset < RADIX_TREE_MAP_SIZE) {
770 if (node->slots[offset])
771 break;
772 }
773 index &= ~((RADIX_TREE_MAP_SIZE << shift) - 1);
774 index += offset << shift;
775 /* Overflow after ~0UL */
776 if (!index)
777 return NULL;
778 if (offset == RADIX_TREE_MAP_SIZE)
779 goto restart;
780 }
782 /* This is leaf-node */
783 if (!shift)
784 break;
786 node = rcu_dereference_raw(node->slots[offset]);
787 if (node == NULL)
788 goto restart;
789 shift -= RADIX_TREE_MAP_SHIFT;
790 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
791 }
793 /* Update the iterator state */
794 iter->index = index;
795 iter->next_index = (index | RADIX_TREE_MAP_MASK) + 1;
797 /* Construct iter->tags bit-mask from node->tags[tag] array */
798 if (flags & RADIX_TREE_ITER_TAGGED) {
799 unsigned tag_long, tag_bit;
801 tag_long = offset / BITS_PER_LONG;
802 tag_bit = offset % BITS_PER_LONG;
803 iter->tags = node->tags[tag][tag_long] >> tag_bit;
804 /* This never happens if RADIX_TREE_TAG_LONGS == 1 */
805 if (tag_long < RADIX_TREE_TAG_LONGS - 1) {
806 /* Pick tags from next element */
807 if (tag_bit)
808 iter->tags |= node->tags[tag][tag_long + 1] <<
809 (BITS_PER_LONG - tag_bit);
810 /* Clip chunk size, here only BITS_PER_LONG tags */
811 iter->next_index = index + BITS_PER_LONG;
812 }
813 }
815 return node->slots + offset;
816 }
817 EXPORT_SYMBOL(radix_tree_next_chunk);
819 /**
820 * radix_tree_range_tag_if_tagged - for each item in given range set given
821 * tag if item has another tag set
822 * @root: radix tree root
823 * @first_indexp: pointer to a starting index of a range to scan
824 * @last_index: last index of a range to scan
825 * @nr_to_tag: maximum number items to tag
826 * @iftag: tag index to test
827 * @settag: tag index to set if tested tag is set
828 *
829 * This function scans range of radix tree from first_index to last_index
830 * (inclusive). For each item in the range if iftag is set, the function sets
831 * also settag. The function stops either after tagging nr_to_tag items or
832 * after reaching last_index.
833 *
834 * The tags must be set from the leaf level only and propagated back up the
835 * path to the root. We must do this so that we resolve the full path before
836 * setting any tags on intermediate nodes. If we set tags as we descend, then
837 * we can get to the leaf node and find that the index that has the iftag
838 * set is outside the range we are scanning. This reults in dangling tags and
839 * can lead to problems with later tag operations (e.g. livelocks on lookups).
840 *
841 * The function returns number of leaves where the tag was set and sets
842 * *first_indexp to the first unscanned index.
843 * WARNING! *first_indexp can wrap if last_index is ULONG_MAX. Caller must
844 * be prepared to handle that.
845 */
846 unsigned long radix_tree_range_tag_if_tagged(struct radix_tree_root *root,
847 unsigned long *first_indexp, unsigned long last_index,
848 unsigned long nr_to_tag,
849 unsigned int iftag, unsigned int settag)
850 {
851 unsigned int height = root->height;
852 struct radix_tree_node *node = NULL;
853 struct radix_tree_node *slot;
854 unsigned int shift;
855 unsigned long tagged = 0;
856 unsigned long index = *first_indexp;
858 last_index = min(last_index, radix_tree_maxindex(height));
859 if (index > last_index)
860 return 0;
861 if (!nr_to_tag)
862 return 0;
863 if (!root_tag_get(root, iftag)) {
864 *first_indexp = last_index + 1;
865 return 0;
866 }
867 if (height == 0) {
868 *first_indexp = last_index + 1;
869 root_tag_set(root, settag);
870 return 1;
871 }
873 shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
874 slot = indirect_to_ptr(root->rnode);
876 for (;;) {
877 unsigned long upindex;
878 int offset;
880 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
881 if (!slot->slots[offset])
882 goto next;
883 if (!tag_get(slot, iftag, offset))
884 goto next;
885 if (shift) {
886 /* Go down one level */
887 shift -= RADIX_TREE_MAP_SHIFT;
888 node = slot;
889 slot = slot->slots[offset];
890 continue;
891 }
893 /* tag the leaf */
894 tagged++;
895 tag_set(slot, settag, offset);
897 /* walk back up the path tagging interior nodes */
898 upindex = index;
899 while (node) {
900 upindex >>= RADIX_TREE_MAP_SHIFT;
901 offset = upindex & RADIX_TREE_MAP_MASK;
903 /* stop if we find a node with the tag already set */
904 if (tag_get(node, settag, offset))
905 break;
906 tag_set(node, settag, offset);
907 node = node->parent;
908 }
910 /*
911 * Small optimization: now clear that node pointer.
912 * Since all of this slot's ancestors now have the tag set
913 * from setting it above, we have no further need to walk
914 * back up the tree setting tags, until we update slot to
915 * point to another radix_tree_node.
916 */
917 node = NULL;
919 next:
920 /* Go to next item at level determined by 'shift' */
921 index = ((index >> shift) + 1) << shift;
922 /* Overflow can happen when last_index is ~0UL... */
923 if (index > last_index || !index)
924 break;
925 if (tagged >= nr_to_tag)
926 break;
927 while (((index >> shift) & RADIX_TREE_MAP_MASK) == 0) {
928 /*
929 * We've fully scanned this node. Go up. Because
930 * last_index is guaranteed to be in the tree, what
931 * we do below cannot wander astray.
932 */
933 slot = slot->parent;
934 shift += RADIX_TREE_MAP_SHIFT;
935 }
936 }
937 /*
938 * We need not to tag the root tag if there is no tag which is set with
939 * settag within the range from *first_indexp to last_index.
940 */
941 if (tagged > 0)
942 root_tag_set(root, settag);
943 *first_indexp = index;
945 return tagged;
946 }
947 EXPORT_SYMBOL(radix_tree_range_tag_if_tagged);
949 /**
950 * radix_tree_gang_lookup - perform multiple lookup on a radix tree
951 * @root: radix tree root
952 * @results: where the results of the lookup are placed
953 * @first_index: start the lookup from this key
954 * @max_items: place up to this many items at *results
955 *
956 * Performs an index-ascending scan of the tree for present items. Places
957 * them at *@results and returns the number of items which were placed at
958 * *@results.
959 *
960 * The implementation is naive.
961 *
962 * Like radix_tree_lookup, radix_tree_gang_lookup may be called under
963 * rcu_read_lock. In this case, rather than the returned results being
964 * an atomic snapshot of the tree at a single point in time, the semantics
965 * of an RCU protected gang lookup are as though multiple radix_tree_lookups
966 * have been issued in individual locks, and results stored in 'results'.
967 */
968 unsigned int
969 radix_tree_gang_lookup(struct radix_tree_root *root, void **results,
970 unsigned long first_index, unsigned int max_items)
971 {
972 struct radix_tree_iter iter;
973 void **slot;
974 unsigned int ret = 0;
976 if (unlikely(!max_items))
977 return 0;
979 radix_tree_for_each_slot(slot, root, &iter, first_index) {
980 results[ret] = indirect_to_ptr(rcu_dereference_raw(*slot));
981 if (!results[ret])
982 continue;
983 if (++ret == max_items)
984 break;
985 }
987 return ret;
988 }
989 EXPORT_SYMBOL(radix_tree_gang_lookup);
991 /**
992 * radix_tree_gang_lookup_slot - perform multiple slot lookup on radix tree
993 * @root: radix tree root
994 * @results: where the results of the lookup are placed
995 * @indices: where their indices should be placed (but usually NULL)
996 * @first_index: start the lookup from this key
997 * @max_items: place up to this many items at *results
998 *
999 * Performs an index-ascending scan of the tree for present items. Places
1000 * their slots at *@results and returns the number of items which were
1001 * placed at *@results.
1002 *
1003 * The implementation is naive.
1004 *
1005 * Like radix_tree_gang_lookup as far as RCU and locking goes. Slots must
1006 * be dereferenced with radix_tree_deref_slot, and if using only RCU
1007 * protection, radix_tree_deref_slot may fail requiring a retry.
1008 */
1009 unsigned int
1010 radix_tree_gang_lookup_slot(struct radix_tree_root *root,
1011 void ***results, unsigned long *indices,
1012 unsigned long first_index, unsigned int max_items)
1013 {
1014 struct radix_tree_iter iter;
1015 void **slot;
1016 unsigned int ret = 0;
1018 if (unlikely(!max_items))
1019 return 0;
1021 radix_tree_for_each_slot(slot, root, &iter, first_index) {
1022 results[ret] = slot;
1023 if (indices)
1024 indices[ret] = iter.index;
1025 if (++ret == max_items)
1026 break;
1027 }
1029 return ret;
1030 }
1031 EXPORT_SYMBOL(radix_tree_gang_lookup_slot);
1033 /**
1034 * radix_tree_gang_lookup_tag - perform multiple lookup on a radix tree
1035 * based on a tag
1036 * @root: radix tree root
1037 * @results: where the results of the lookup are placed
1038 * @first_index: start the lookup from this key
1039 * @max_items: place up to this many items at *results
1040 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1041 *
1042 * Performs an index-ascending scan of the tree for present items which
1043 * have the tag indexed by @tag set. Places the items at *@results and
1044 * returns the number of items which were placed at *@results.
1045 */
1046 unsigned int
1047 radix_tree_gang_lookup_tag(struct radix_tree_root *root, void **results,
1048 unsigned long first_index, unsigned int max_items,
1049 unsigned int tag)
1050 {
1051 struct radix_tree_iter iter;
1052 void **slot;
1053 unsigned int ret = 0;
1055 if (unlikely(!max_items))
1056 return 0;
1058 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1059 results[ret] = indirect_to_ptr(rcu_dereference_raw(*slot));
1060 if (!results[ret])
1061 continue;
1062 if (++ret == max_items)
1063 break;
1064 }
1066 return ret;
1067 }
1068 EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
1070 /**
1071 * radix_tree_gang_lookup_tag_slot - perform multiple slot lookup on a
1072 * radix tree based on a tag
1073 * @root: radix tree root
1074 * @results: where the results of the lookup are placed
1075 * @first_index: start the lookup from this key
1076 * @max_items: place up to this many items at *results
1077 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1078 *
1079 * Performs an index-ascending scan of the tree for present items which
1080 * have the tag indexed by @tag set. Places the slots at *@results and
1081 * returns the number of slots which were placed at *@results.
1082 */
1083 unsigned int
1084 radix_tree_gang_lookup_tag_slot(struct radix_tree_root *root, void ***results,
1085 unsigned long first_index, unsigned int max_items,
1086 unsigned int tag)
1087 {
1088 struct radix_tree_iter iter;
1089 void **slot;
1090 unsigned int ret = 0;
1092 if (unlikely(!max_items))
1093 return 0;
1095 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1096 results[ret] = slot;
1097 if (++ret == max_items)
1098 break;
1099 }
1101 return ret;
1102 }
1103 EXPORT_SYMBOL(radix_tree_gang_lookup_tag_slot);
1105 #if defined(CONFIG_SHMEM) && defined(CONFIG_SWAP)
1106 #include <linux/sched.h> /* for cond_resched() */
1108 /*
1109 * This linear search is at present only useful to shmem_unuse_inode().
1110 */
1111 static unsigned long __locate(struct radix_tree_node *slot, void *item,
1112 unsigned long index, unsigned long *found_index)
1113 {
1114 unsigned int shift, height;
1115 unsigned long i;
1117 height = slot->height;
1118 shift = (height-1) * RADIX_TREE_MAP_SHIFT;
1120 for ( ; height > 1; height--) {
1121 i = (index >> shift) & RADIX_TREE_MAP_MASK;
1122 for (;;) {
1123 if (slot->slots[i] != NULL)
1124 break;
1125 index &= ~((1UL << shift) - 1);
1126 index += 1UL << shift;
1127 if (index == 0)
1128 goto out; /* 32-bit wraparound */
1129 i++;
1130 if (i == RADIX_TREE_MAP_SIZE)
1131 goto out;
1132 }
1134 shift -= RADIX_TREE_MAP_SHIFT;
1135 slot = rcu_dereference_raw(slot->slots[i]);
1136 if (slot == NULL)
1137 goto out;
1138 }
1140 /* Bottom level: check items */
1141 for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
1142 if (slot->slots[i] == item) {
1143 *found_index = index + i;
1144 index = 0;
1145 goto out;
1146 }
1147 }
1148 index += RADIX_TREE_MAP_SIZE;
1149 out:
1150 return index;
1151 }
1153 /**
1154 * radix_tree_locate_item - search through radix tree for item
1155 * @root: radix tree root
1156 * @item: item to be found
1157 *
1158 * Returns index where item was found, or -1 if not found.
1159 * Caller must hold no lock (since this time-consuming function needs
1160 * to be preemptible), and must check afterwards if item is still there.
1161 */
1162 unsigned long radix_tree_locate_item(struct radix_tree_root *root, void *item)
1163 {
1164 struct radix_tree_node *node;
1165 unsigned long max_index;
1166 unsigned long cur_index = 0;
1167 unsigned long found_index = -1;
1169 do {
1170 rcu_read_lock();
1171 node = rcu_dereference_raw(root->rnode);
1172 if (!radix_tree_is_indirect_ptr(node)) {
1173 rcu_read_unlock();
1174 if (node == item)
1175 found_index = 0;
1176 break;
1177 }
1179 node = indirect_to_ptr(node);
1180 max_index = radix_tree_maxindex(node->height);
1181 if (cur_index > max_index) {
1182 rcu_read_unlock();
1183 break;
1184 }
1186 cur_index = __locate(node, item, cur_index, &found_index);
1187 rcu_read_unlock();
1188 cond_resched();
1189 } while (cur_index != 0 && cur_index <= max_index);
1191 return found_index;
1192 }
1193 #else
1194 unsigned long radix_tree_locate_item(struct radix_tree_root *root, void *item)
1195 {
1196 return -1;
1197 }
1198 #endif /* CONFIG_SHMEM && CONFIG_SWAP */
1200 /**
1201 * radix_tree_shrink - shrink height of a radix tree to minimal
1202 * @root radix tree root
1203 */
1204 static inline void radix_tree_shrink(struct radix_tree_root *root)
1205 {
1206 /* try to shrink tree height */
1207 while (root->height > 0) {
1208 struct radix_tree_node *to_free = root->rnode;
1209 struct radix_tree_node *slot;
1211 BUG_ON(!radix_tree_is_indirect_ptr(to_free));
1212 to_free = indirect_to_ptr(to_free);
1214 /*
1215 * The candidate node has more than one child, or its child
1216 * is not at the leftmost slot, we cannot shrink.
1217 */
1218 if (to_free->count != 1)
1219 break;
1220 if (!to_free->slots[0])
1221 break;
1223 /*
1224 * We don't need rcu_assign_pointer(), since we are simply
1225 * moving the node from one part of the tree to another: if it
1226 * was safe to dereference the old pointer to it
1227 * (to_free->slots[0]), it will be safe to dereference the new
1228 * one (root->rnode) as far as dependent read barriers go.
1229 */
1230 slot = to_free->slots[0];
1231 if (root->height > 1) {
1232 slot->parent = NULL;
1233 slot = ptr_to_indirect(slot);
1234 }
1235 root->rnode = slot;
1236 root->height--;
1238 /*
1239 * We have a dilemma here. The node's slot[0] must not be
1240 * NULLed in case there are concurrent lookups expecting to
1241 * find the item. However if this was a bottom-level node,
1242 * then it may be subject to the slot pointer being visible
1243 * to callers dereferencing it. If item corresponding to
1244 * slot[0] is subsequently deleted, these callers would expect
1245 * their slot to become empty sooner or later.
1246 *
1247 * For example, lockless pagecache will look up a slot, deref
1248 * the page pointer, and if the page is 0 refcount it means it
1249 * was concurrently deleted from pagecache so try the deref
1250 * again. Fortunately there is already a requirement for logic
1251 * to retry the entire slot lookup -- the indirect pointer
1252 * problem (replacing direct root node with an indirect pointer
1253 * also results in a stale slot). So tag the slot as indirect
1254 * to force callers to retry.
1255 */
1256 if (root->height == 0)
1257 *((unsigned long *)&to_free->slots[0]) |=
1258 RADIX_TREE_INDIRECT_PTR;
1260 radix_tree_node_free(to_free);
1261 }
1262 }
1264 /**
1265 * radix_tree_delete_item - delete an item from a radix tree
1266 * @root: radix tree root
1267 * @index: index key
1268 * @item: expected item
1269 *
1270 * Remove @item at @index from the radix tree rooted at @root.
1271 *
1272 * Returns the address of the deleted item, or NULL if it was not present
1273 * or the entry at the given @index was not @item.
1274 */
1275 void *radix_tree_delete_item(struct radix_tree_root *root,
1276 unsigned long index, void *item)
1277 {
1278 struct radix_tree_node *node = NULL;
1279 struct radix_tree_node *slot = NULL;
1280 struct radix_tree_node *to_free;
1281 unsigned int height, shift;
1282 int tag;
1283 int uninitialized_var(offset);
1285 height = root->height;
1286 if (index > radix_tree_maxindex(height))
1287 goto out;
1289 slot = root->rnode;
1290 if (height == 0) {
1291 root_tag_clear_all(root);
1292 root->rnode = NULL;
1293 goto out;
1294 }
1295 slot = indirect_to_ptr(slot);
1296 shift = height * RADIX_TREE_MAP_SHIFT;
1298 do {
1299 if (slot == NULL)
1300 goto out;
1302 shift -= RADIX_TREE_MAP_SHIFT;
1303 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
1304 node = slot;
1305 slot = slot->slots[offset];
1306 } while (shift);
1308 if (slot == NULL)
1309 goto out;
1311 if (item && slot != item) {
1312 slot = NULL;
1313 goto out;
1314 }
1316 /*
1317 * Clear all tags associated with the item to be deleted.
1318 * This way of doing it would be inefficient, but seldom is any set.
1319 */
1320 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
1321 if (tag_get(node, tag, offset))
1322 radix_tree_tag_clear(root, index, tag);
1323 }
1325 to_free = NULL;
1326 /* Now free the nodes we do not need anymore */
1327 while (node) {
1328 node->slots[offset] = NULL;
1329 node->count--;
1330 /*
1331 * Queue the node for deferred freeing after the
1332 * last reference to it disappears (set NULL, above).
1333 */
1334 if (to_free)
1335 radix_tree_node_free(to_free);
1337 if (node->count) {
1338 if (node == indirect_to_ptr(root->rnode))
1339 radix_tree_shrink(root);
1340 goto out;
1341 }
1343 /* Node with zero slots in use so free it */
1344 to_free = node;
1346 index >>= RADIX_TREE_MAP_SHIFT;
1347 offset = index & RADIX_TREE_MAP_MASK;
1348 node = node->parent;
1349 }
1351 root_tag_clear_all(root);
1352 root->height = 0;
1353 root->rnode = NULL;
1354 if (to_free)
1355 radix_tree_node_free(to_free);
1357 out:
1358 return slot;
1359 }
1360 EXPORT_SYMBOL(radix_tree_delete_item);
1362 /**
1363 * radix_tree_delete - delete an item from a radix tree
1364 * @root: radix tree root
1365 * @index: index key
1366 *
1367 * Remove the item at @index from the radix tree rooted at @root.
1368 *
1369 * Returns the address of the deleted item, or NULL if it was not present.
1370 */
1371 void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
1372 {
1373 return radix_tree_delete_item(root, index, NULL);
1374 }
1375 EXPORT_SYMBOL(radix_tree_delete);
1377 /**
1378 * radix_tree_tagged - test whether any items in the tree are tagged
1379 * @root: radix tree root
1380 * @tag: tag to test
1381 */
1382 int radix_tree_tagged(struct radix_tree_root *root, unsigned int tag)
1383 {
1384 return root_tag_get(root, tag);
1385 }
1386 EXPORT_SYMBOL(radix_tree_tagged);
1388 static void
1389 radix_tree_node_ctor(void *node)
1390 {
1391 memset(node, 0, sizeof(struct radix_tree_node));
1392 }
1394 static __init unsigned long __maxindex(unsigned int height)
1395 {
1396 unsigned int width = height * RADIX_TREE_MAP_SHIFT;
1397 int shift = RADIX_TREE_INDEX_BITS - width;
1399 if (shift < 0)
1400 return ~0UL;
1401 if (shift >= BITS_PER_LONG)
1402 return 0UL;
1403 return ~0UL >> shift;
1404 }
1406 static __init void radix_tree_init_maxindex(void)
1407 {
1408 unsigned int i;
1410 for (i = 0; i < ARRAY_SIZE(height_to_maxindex); i++)
1411 height_to_maxindex[i] = __maxindex(i);
1412 }
1414 static int radix_tree_callback(struct notifier_block *nfb,
1415 unsigned long action,
1416 void *hcpu)
1417 {
1418 int cpu = (long)hcpu;
1419 struct radix_tree_preload *rtp;
1421 /* Free per-cpu pool of perloaded nodes */
1422 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
1423 rtp = &per_cpu(radix_tree_preloads, cpu);
1424 while (rtp->nr) {
1425 kmem_cache_free(radix_tree_node_cachep,
1426 rtp->nodes[rtp->nr-1]);
1427 rtp->nodes[rtp->nr-1] = NULL;
1428 rtp->nr--;
1429 }
1430 }
1431 return NOTIFY_OK;
1432 }
1434 void __init radix_tree_init(void)
1435 {
1436 radix_tree_node_cachep = kmem_cache_create("radix_tree_node",
1437 sizeof(struct radix_tree_node), 0,
1438 SLAB_PANIC | SLAB_RECLAIM_ACCOUNT,
1439 radix_tree_node_ctor);
1440 radix_tree_init_maxindex();
1441 hotcpu_notifier(radix_tree_callback, 0);
1442 }