Loading...
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 */
22
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() */
36
37
38/*
39 * The height_to_maxindex array needs to be one deeper than the maximum
40 * path as height 0 holds only 1 entry.
41 */
42static unsigned long height_to_maxindex[RADIX_TREE_MAX_PATH + 1] __read_mostly;
43
44/*
45 * Radix tree node cache.
46 */
47static struct kmem_cache *radix_tree_node_cachep;
48
49/*
50 * The radix tree is variable-height, so an insert operation not only has
51 * to build the branch to its corresponding item, it also has to build the
52 * branch to existing items if the size has to be increased (by
53 * radix_tree_extend).
54 *
55 * The worst case is a zero height tree with just a single item at index 0,
56 * and then inserting an item at index ULONG_MAX. This requires 2 new branches
57 * of RADIX_TREE_MAX_PATH size to be created, with only the root node shared.
58 * Hence:
59 */
60#define RADIX_TREE_PRELOAD_SIZE (RADIX_TREE_MAX_PATH * 2 - 1)
61
62/*
63 * Per-cpu pool of preloaded nodes
64 */
65struct radix_tree_preload {
66 int nr;
67 struct radix_tree_node *nodes[RADIX_TREE_PRELOAD_SIZE];
68};
69static DEFINE_PER_CPU(struct radix_tree_preload, radix_tree_preloads) = { 0, };
70
71static inline void *ptr_to_indirect(void *ptr)
72{
73 return (void *)((unsigned long)ptr | RADIX_TREE_INDIRECT_PTR);
74}
75
76static inline void *indirect_to_ptr(void *ptr)
77{
78 return (void *)((unsigned long)ptr & ~RADIX_TREE_INDIRECT_PTR);
79}
80
81static inline gfp_t root_gfp_mask(struct radix_tree_root *root)
82{
83 return root->gfp_mask & __GFP_BITS_MASK;
84}
85
86static inline void tag_set(struct radix_tree_node *node, unsigned int tag,
87 int offset)
88{
89 __set_bit(offset, node->tags[tag]);
90}
91
92static inline void tag_clear(struct radix_tree_node *node, unsigned int tag,
93 int offset)
94{
95 __clear_bit(offset, node->tags[tag]);
96}
97
98static inline int tag_get(struct radix_tree_node *node, unsigned int tag,
99 int offset)
100{
101 return test_bit(offset, node->tags[tag]);
102}
103
104static inline void root_tag_set(struct radix_tree_root *root, unsigned int tag)
105{
106 root->gfp_mask |= (__force gfp_t)(1 << (tag + __GFP_BITS_SHIFT));
107}
108
109static inline void root_tag_clear(struct radix_tree_root *root, unsigned int tag)
110{
111 root->gfp_mask &= (__force gfp_t)~(1 << (tag + __GFP_BITS_SHIFT));
112}
113
114static inline void root_tag_clear_all(struct radix_tree_root *root)
115{
116 root->gfp_mask &= __GFP_BITS_MASK;
117}
118
119static inline int root_tag_get(struct radix_tree_root *root, unsigned int tag)
120{
121 return (__force unsigned)root->gfp_mask & (1 << (tag + __GFP_BITS_SHIFT));
122}
123
124/*
125 * Returns 1 if any slot in the node has this tag set.
126 * Otherwise returns 0.
127 */
128static inline int any_tag_set(struct radix_tree_node *node, unsigned int tag)
129{
130 int idx;
131 for (idx = 0; idx < RADIX_TREE_TAG_LONGS; idx++) {
132 if (node->tags[tag][idx])
133 return 1;
134 }
135 return 0;
136}
137
138/**
139 * radix_tree_find_next_bit - find the next set bit in a memory region
140 *
141 * @addr: The address to base the search on
142 * @size: The bitmap size in bits
143 * @offset: The bitnumber to start searching at
144 *
145 * Unrollable variant of find_next_bit() for constant size arrays.
146 * Tail bits starting from size to roundup(size, BITS_PER_LONG) must be zero.
147 * Returns next bit offset, or size if nothing found.
148 */
149static __always_inline unsigned long
150radix_tree_find_next_bit(const unsigned long *addr,
151 unsigned long size, unsigned long offset)
152{
153 if (!__builtin_constant_p(size))
154 return find_next_bit(addr, size, offset);
155
156 if (offset < size) {
157 unsigned long tmp;
158
159 addr += offset / BITS_PER_LONG;
160 tmp = *addr >> (offset % BITS_PER_LONG);
161 if (tmp)
162 return __ffs(tmp) + offset;
163 offset = (offset + BITS_PER_LONG) & ~(BITS_PER_LONG - 1);
164 while (offset < size) {
165 tmp = *++addr;
166 if (tmp)
167 return __ffs(tmp) + offset;
168 offset += BITS_PER_LONG;
169 }
170 }
171 return size;
172}
173
174/*
175 * This assumes that the caller has performed appropriate preallocation, and
176 * that the caller has pinned this thread of control to the current CPU.
177 */
178static struct radix_tree_node *
179radix_tree_node_alloc(struct radix_tree_root *root)
180{
181 struct radix_tree_node *ret = NULL;
182 gfp_t gfp_mask = root_gfp_mask(root);
183
184 /*
185 * Preload code isn't irq safe and it doesn't make sence to use
186 * preloading in the interrupt anyway as all the allocations have to
187 * be atomic. So just do normal allocation when in interrupt.
188 */
189 if (!(gfp_mask & __GFP_WAIT) && !in_interrupt()) {
190 struct radix_tree_preload *rtp;
191
192 /*
193 * Provided the caller has preloaded here, we will always
194 * succeed in getting a node here (and never reach
195 * kmem_cache_alloc)
196 */
197 rtp = &__get_cpu_var(radix_tree_preloads);
198 if (rtp->nr) {
199 ret = rtp->nodes[rtp->nr - 1];
200 rtp->nodes[rtp->nr - 1] = NULL;
201 rtp->nr--;
202 }
203 }
204 if (ret == NULL)
205 ret = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
206
207 BUG_ON(radix_tree_is_indirect_ptr(ret));
208 return ret;
209}
210
211static void radix_tree_node_rcu_free(struct rcu_head *head)
212{
213 struct radix_tree_node *node =
214 container_of(head, struct radix_tree_node, rcu_head);
215 int i;
216
217 /*
218 * must only free zeroed nodes into the slab. radix_tree_shrink
219 * can leave us with a non-NULL entry in the first slot, so clear
220 * that here to make sure.
221 */
222 for (i = 0; i < RADIX_TREE_MAX_TAGS; i++)
223 tag_clear(node, i, 0);
224
225 node->slots[0] = NULL;
226 node->count = 0;
227
228 kmem_cache_free(radix_tree_node_cachep, node);
229}
230
231static inline void
232radix_tree_node_free(struct radix_tree_node *node)
233{
234 call_rcu(&node->rcu_head, radix_tree_node_rcu_free);
235}
236
237/*
238 * Load up this CPU's radix_tree_node buffer with sufficient objects to
239 * ensure that the addition of a single element in the tree cannot fail. On
240 * success, return zero, with preemption disabled. On error, return -ENOMEM
241 * with preemption not disabled.
242 *
243 * To make use of this facility, the radix tree must be initialised without
244 * __GFP_WAIT being passed to INIT_RADIX_TREE().
245 */
246static int __radix_tree_preload(gfp_t gfp_mask)
247{
248 struct radix_tree_preload *rtp;
249 struct radix_tree_node *node;
250 int ret = -ENOMEM;
251
252 preempt_disable();
253 rtp = &__get_cpu_var(radix_tree_preloads);
254 while (rtp->nr < ARRAY_SIZE(rtp->nodes)) {
255 preempt_enable();
256 node = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
257 if (node == NULL)
258 goto out;
259 preempt_disable();
260 rtp = &__get_cpu_var(radix_tree_preloads);
261 if (rtp->nr < ARRAY_SIZE(rtp->nodes))
262 rtp->nodes[rtp->nr++] = node;
263 else
264 kmem_cache_free(radix_tree_node_cachep, node);
265 }
266 ret = 0;
267out:
268 return ret;
269}
270
271/*
272 * Load up this CPU's radix_tree_node buffer with sufficient objects to
273 * ensure that the addition of a single element in the tree cannot fail. On
274 * success, return zero, with preemption disabled. On error, return -ENOMEM
275 * with preemption not disabled.
276 *
277 * To make use of this facility, the radix tree must be initialised without
278 * __GFP_WAIT being passed to INIT_RADIX_TREE().
279 */
280int radix_tree_preload(gfp_t gfp_mask)
281{
282 /* Warn on non-sensical use... */
283 WARN_ON_ONCE(!(gfp_mask & __GFP_WAIT));
284 return __radix_tree_preload(gfp_mask);
285}
286EXPORT_SYMBOL(radix_tree_preload);
287
288/*
289 * The same as above function, except we don't guarantee preloading happens.
290 * We do it, if we decide it helps. On success, return zero with preemption
291 * disabled. On error, return -ENOMEM with preemption not disabled.
292 */
293int radix_tree_maybe_preload(gfp_t gfp_mask)
294{
295 if (gfp_mask & __GFP_WAIT)
296 return __radix_tree_preload(gfp_mask);
297 /* Preloading doesn't help anything with this gfp mask, skip it */
298 preempt_disable();
299 return 0;
300}
301EXPORT_SYMBOL(radix_tree_maybe_preload);
302
303/*
304 * Return the maximum key which can be store into a
305 * radix tree with height HEIGHT.
306 */
307static inline unsigned long radix_tree_maxindex(unsigned int height)
308{
309 return height_to_maxindex[height];
310}
311
312/*
313 * Extend a radix tree so it can store key @index.
314 */
315static int radix_tree_extend(struct radix_tree_root *root, unsigned long index)
316{
317 struct radix_tree_node *node;
318 struct radix_tree_node *slot;
319 unsigned int height;
320 int tag;
321
322 /* Figure out what the height should be. */
323 height = root->height + 1;
324 while (index > radix_tree_maxindex(height))
325 height++;
326
327 if (root->rnode == NULL) {
328 root->height = height;
329 goto out;
330 }
331
332 do {
333 unsigned int newheight;
334 if (!(node = radix_tree_node_alloc(root)))
335 return -ENOMEM;
336
337 /* Propagate the aggregated tag info into the new root */
338 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
339 if (root_tag_get(root, tag))
340 tag_set(node, tag, 0);
341 }
342
343 /* Increase the height. */
344 newheight = root->height+1;
345 BUG_ON(newheight & ~RADIX_TREE_HEIGHT_MASK);
346 node->path = newheight;
347 node->count = 1;
348 node->parent = NULL;
349 slot = root->rnode;
350 if (newheight > 1) {
351 slot = indirect_to_ptr(slot);
352 slot->parent = node;
353 }
354 node->slots[0] = slot;
355 node = ptr_to_indirect(node);
356 rcu_assign_pointer(root->rnode, node);
357 root->height = newheight;
358 } while (height > root->height);
359out:
360 return 0;
361}
362
363/**
364 * __radix_tree_create - create a slot in a radix tree
365 * @root: radix tree root
366 * @index: index key
367 * @nodep: returns node
368 * @slotp: returns slot
369 *
370 * Create, if necessary, and return the node and slot for an item
371 * at position @index in the radix tree @root.
372 *
373 * Until there is more than one item in the tree, no nodes are
374 * allocated and @root->rnode is used as a direct slot instead of
375 * pointing to a node, in which case *@nodep will be NULL.
376 *
377 * Returns -ENOMEM, or 0 for success.
378 */
379int __radix_tree_create(struct radix_tree_root *root, unsigned long index,
380 struct radix_tree_node **nodep, void ***slotp)
381{
382 struct radix_tree_node *node = NULL, *slot;
383 unsigned int height, shift, offset;
384 int error;
385
386 /* Make sure the tree is high enough. */
387 if (index > radix_tree_maxindex(root->height)) {
388 error = radix_tree_extend(root, index);
389 if (error)
390 return error;
391 }
392
393 slot = indirect_to_ptr(root->rnode);
394
395 height = root->height;
396 shift = (height-1) * RADIX_TREE_MAP_SHIFT;
397
398 offset = 0; /* uninitialised var warning */
399 while (height > 0) {
400 if (slot == NULL) {
401 /* Have to add a child node. */
402 if (!(slot = radix_tree_node_alloc(root)))
403 return -ENOMEM;
404 slot->path = height;
405 slot->parent = node;
406 if (node) {
407 rcu_assign_pointer(node->slots[offset], slot);
408 node->count++;
409 slot->path |= offset << RADIX_TREE_HEIGHT_SHIFT;
410 } else
411 rcu_assign_pointer(root->rnode, ptr_to_indirect(slot));
412 }
413
414 /* Go a level down */
415 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
416 node = slot;
417 slot = node->slots[offset];
418 shift -= RADIX_TREE_MAP_SHIFT;
419 height--;
420 }
421
422 if (nodep)
423 *nodep = node;
424 if (slotp)
425 *slotp = node ? node->slots + offset : (void **)&root->rnode;
426 return 0;
427}
428
429/**
430 * radix_tree_insert - insert into a radix tree
431 * @root: radix tree root
432 * @index: index key
433 * @item: item to insert
434 *
435 * Insert an item into the radix tree at position @index.
436 */
437int radix_tree_insert(struct radix_tree_root *root,
438 unsigned long index, void *item)
439{
440 struct radix_tree_node *node;
441 void **slot;
442 int error;
443
444 BUG_ON(radix_tree_is_indirect_ptr(item));
445
446 error = __radix_tree_create(root, index, &node, &slot);
447 if (error)
448 return error;
449 if (*slot != NULL)
450 return -EEXIST;
451 rcu_assign_pointer(*slot, item);
452
453 if (node) {
454 node->count++;
455 BUG_ON(tag_get(node, 0, index & RADIX_TREE_MAP_MASK));
456 BUG_ON(tag_get(node, 1, index & RADIX_TREE_MAP_MASK));
457 } else {
458 BUG_ON(root_tag_get(root, 0));
459 BUG_ON(root_tag_get(root, 1));
460 }
461
462 return 0;
463}
464EXPORT_SYMBOL(radix_tree_insert);
465
466/**
467 * __radix_tree_lookup - lookup an item in a radix tree
468 * @root: radix tree root
469 * @index: index key
470 * @nodep: returns node
471 * @slotp: returns slot
472 *
473 * Lookup and return the item at position @index in the radix
474 * tree @root.
475 *
476 * Until there is more than one item in the tree, no nodes are
477 * allocated and @root->rnode is used as a direct slot instead of
478 * pointing to a node, in which case *@nodep will be NULL.
479 */
480void *__radix_tree_lookup(struct radix_tree_root *root, unsigned long index,
481 struct radix_tree_node **nodep, void ***slotp)
482{
483 struct radix_tree_node *node, *parent;
484 unsigned int height, shift;
485 void **slot;
486
487 node = rcu_dereference_raw(root->rnode);
488 if (node == NULL)
489 return NULL;
490
491 if (!radix_tree_is_indirect_ptr(node)) {
492 if (index > 0)
493 return NULL;
494
495 if (nodep)
496 *nodep = NULL;
497 if (slotp)
498 *slotp = (void **)&root->rnode;
499 return node;
500 }
501 node = indirect_to_ptr(node);
502
503 height = node->path & RADIX_TREE_HEIGHT_MASK;
504 if (index > radix_tree_maxindex(height))
505 return NULL;
506
507 shift = (height-1) * RADIX_TREE_MAP_SHIFT;
508
509 do {
510 parent = node;
511 slot = node->slots + ((index >> shift) & RADIX_TREE_MAP_MASK);
512 node = rcu_dereference_raw(*slot);
513 if (node == NULL)
514 return NULL;
515
516 shift -= RADIX_TREE_MAP_SHIFT;
517 height--;
518 } while (height > 0);
519
520 if (nodep)
521 *nodep = parent;
522 if (slotp)
523 *slotp = slot;
524 return node;
525}
526
527/**
528 * radix_tree_lookup_slot - lookup a slot in a radix tree
529 * @root: radix tree root
530 * @index: index key
531 *
532 * Returns: the slot corresponding to the position @index in the
533 * radix tree @root. This is useful for update-if-exists operations.
534 *
535 * This function can be called under rcu_read_lock iff the slot is not
536 * modified by radix_tree_replace_slot, otherwise it must be called
537 * exclusive from other writers. Any dereference of the slot must be done
538 * using radix_tree_deref_slot.
539 */
540void **radix_tree_lookup_slot(struct radix_tree_root *root, unsigned long index)
541{
542 void **slot;
543
544 if (!__radix_tree_lookup(root, index, NULL, &slot))
545 return NULL;
546 return slot;
547}
548EXPORT_SYMBOL(radix_tree_lookup_slot);
549
550/**
551 * radix_tree_lookup - perform lookup operation on a radix tree
552 * @root: radix tree root
553 * @index: index key
554 *
555 * Lookup the item at the position @index in the radix tree @root.
556 *
557 * This function can be called under rcu_read_lock, however the caller
558 * must manage lifetimes of leaf nodes (eg. RCU may also be used to free
559 * them safely). No RCU barriers are required to access or modify the
560 * returned item, however.
561 */
562void *radix_tree_lookup(struct radix_tree_root *root, unsigned long index)
563{
564 return __radix_tree_lookup(root, index, NULL, NULL);
565}
566EXPORT_SYMBOL(radix_tree_lookup);
567
568/**
569 * radix_tree_tag_set - set a tag on a radix tree node
570 * @root: radix tree root
571 * @index: index key
572 * @tag: tag index
573 *
574 * Set the search tag (which must be < RADIX_TREE_MAX_TAGS)
575 * corresponding to @index in the radix tree. From
576 * the root all the way down to the leaf node.
577 *
578 * Returns the address of the tagged item. Setting a tag on a not-present
579 * item is a bug.
580 */
581void *radix_tree_tag_set(struct radix_tree_root *root,
582 unsigned long index, unsigned int tag)
583{
584 unsigned int height, shift;
585 struct radix_tree_node *slot;
586
587 height = root->height;
588 BUG_ON(index > radix_tree_maxindex(height));
589
590 slot = indirect_to_ptr(root->rnode);
591 shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
592
593 while (height > 0) {
594 int offset;
595
596 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
597 if (!tag_get(slot, tag, offset))
598 tag_set(slot, tag, offset);
599 slot = slot->slots[offset];
600 BUG_ON(slot == NULL);
601 shift -= RADIX_TREE_MAP_SHIFT;
602 height--;
603 }
604
605 /* set the root's tag bit */
606 if (slot && !root_tag_get(root, tag))
607 root_tag_set(root, tag);
608
609 return slot;
610}
611EXPORT_SYMBOL(radix_tree_tag_set);
612
613/**
614 * radix_tree_tag_clear - clear a tag on a radix tree node
615 * @root: radix tree root
616 * @index: index key
617 * @tag: tag index
618 *
619 * Clear the search tag (which must be < RADIX_TREE_MAX_TAGS)
620 * corresponding to @index in the radix tree. If
621 * this causes the leaf node to have no tags set then clear the tag in the
622 * next-to-leaf node, etc.
623 *
624 * Returns the address of the tagged item on success, else NULL. ie:
625 * has the same return value and semantics as radix_tree_lookup().
626 */
627void *radix_tree_tag_clear(struct radix_tree_root *root,
628 unsigned long index, unsigned int tag)
629{
630 struct radix_tree_node *node = NULL;
631 struct radix_tree_node *slot = NULL;
632 unsigned int height, shift;
633 int uninitialized_var(offset);
634
635 height = root->height;
636 if (index > radix_tree_maxindex(height))
637 goto out;
638
639 shift = height * RADIX_TREE_MAP_SHIFT;
640 slot = indirect_to_ptr(root->rnode);
641
642 while (shift) {
643 if (slot == NULL)
644 goto out;
645
646 shift -= RADIX_TREE_MAP_SHIFT;
647 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
648 node = slot;
649 slot = slot->slots[offset];
650 }
651
652 if (slot == NULL)
653 goto out;
654
655 while (node) {
656 if (!tag_get(node, tag, offset))
657 goto out;
658 tag_clear(node, tag, offset);
659 if (any_tag_set(node, tag))
660 goto out;
661
662 index >>= RADIX_TREE_MAP_SHIFT;
663 offset = index & RADIX_TREE_MAP_MASK;
664 node = node->parent;
665 }
666
667 /* clear the root's tag bit */
668 if (root_tag_get(root, tag))
669 root_tag_clear(root, tag);
670
671out:
672 return slot;
673}
674EXPORT_SYMBOL(radix_tree_tag_clear);
675
676/**
677 * radix_tree_tag_get - get a tag on a radix tree node
678 * @root: radix tree root
679 * @index: index key
680 * @tag: tag index (< RADIX_TREE_MAX_TAGS)
681 *
682 * Return values:
683 *
684 * 0: tag not present or not set
685 * 1: tag set
686 *
687 * Note that the return value of this function may not be relied on, even if
688 * the RCU lock is held, unless tag modification and node deletion are excluded
689 * from concurrency.
690 */
691int radix_tree_tag_get(struct radix_tree_root *root,
692 unsigned long index, unsigned int tag)
693{
694 unsigned int height, shift;
695 struct radix_tree_node *node;
696
697 /* check the root's tag bit */
698 if (!root_tag_get(root, tag))
699 return 0;
700
701 node = rcu_dereference_raw(root->rnode);
702 if (node == NULL)
703 return 0;
704
705 if (!radix_tree_is_indirect_ptr(node))
706 return (index == 0);
707 node = indirect_to_ptr(node);
708
709 height = node->path & RADIX_TREE_HEIGHT_MASK;
710 if (index > radix_tree_maxindex(height))
711 return 0;
712
713 shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
714
715 for ( ; ; ) {
716 int offset;
717
718 if (node == NULL)
719 return 0;
720
721 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
722 if (!tag_get(node, tag, offset))
723 return 0;
724 if (height == 1)
725 return 1;
726 node = rcu_dereference_raw(node->slots[offset]);
727 shift -= RADIX_TREE_MAP_SHIFT;
728 height--;
729 }
730}
731EXPORT_SYMBOL(radix_tree_tag_get);
732
733/**
734 * radix_tree_next_chunk - find next chunk of slots for iteration
735 *
736 * @root: radix tree root
737 * @iter: iterator state
738 * @flags: RADIX_TREE_ITER_* flags and tag index
739 * Returns: pointer to chunk first slot, or NULL if iteration is over
740 */
741void **radix_tree_next_chunk(struct radix_tree_root *root,
742 struct radix_tree_iter *iter, unsigned flags)
743{
744 unsigned shift, tag = flags & RADIX_TREE_ITER_TAG_MASK;
745 struct radix_tree_node *rnode, *node;
746 unsigned long index, offset, height;
747
748 if ((flags & RADIX_TREE_ITER_TAGGED) && !root_tag_get(root, tag))
749 return NULL;
750
751 /*
752 * Catch next_index overflow after ~0UL. iter->index never overflows
753 * during iterating; it can be zero only at the beginning.
754 * And we cannot overflow iter->next_index in a single step,
755 * because RADIX_TREE_MAP_SHIFT < BITS_PER_LONG.
756 *
757 * This condition also used by radix_tree_next_slot() to stop
758 * contiguous iterating, and forbid swithing to the next chunk.
759 */
760 index = iter->next_index;
761 if (!index && iter->index)
762 return NULL;
763
764 rnode = rcu_dereference_raw(root->rnode);
765 if (radix_tree_is_indirect_ptr(rnode)) {
766 rnode = indirect_to_ptr(rnode);
767 } else if (rnode && !index) {
768 /* Single-slot tree */
769 iter->index = 0;
770 iter->next_index = 1;
771 iter->tags = 1;
772 return (void **)&root->rnode;
773 } else
774 return NULL;
775
776restart:
777 height = rnode->path & RADIX_TREE_HEIGHT_MASK;
778 shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
779 offset = index >> shift;
780
781 /* Index outside of the tree */
782 if (offset >= RADIX_TREE_MAP_SIZE)
783 return NULL;
784
785 node = rnode;
786 while (1) {
787 if ((flags & RADIX_TREE_ITER_TAGGED) ?
788 !test_bit(offset, node->tags[tag]) :
789 !node->slots[offset]) {
790 /* Hole detected */
791 if (flags & RADIX_TREE_ITER_CONTIG)
792 return NULL;
793
794 if (flags & RADIX_TREE_ITER_TAGGED)
795 offset = radix_tree_find_next_bit(
796 node->tags[tag],
797 RADIX_TREE_MAP_SIZE,
798 offset + 1);
799 else
800 while (++offset < RADIX_TREE_MAP_SIZE) {
801 if (node->slots[offset])
802 break;
803 }
804 index &= ~((RADIX_TREE_MAP_SIZE << shift) - 1);
805 index += offset << shift;
806 /* Overflow after ~0UL */
807 if (!index)
808 return NULL;
809 if (offset == RADIX_TREE_MAP_SIZE)
810 goto restart;
811 }
812
813 /* This is leaf-node */
814 if (!shift)
815 break;
816
817 node = rcu_dereference_raw(node->slots[offset]);
818 if (node == NULL)
819 goto restart;
820 shift -= RADIX_TREE_MAP_SHIFT;
821 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
822 }
823
824 /* Update the iterator state */
825 iter->index = index;
826 iter->next_index = (index | RADIX_TREE_MAP_MASK) + 1;
827
828 /* Construct iter->tags bit-mask from node->tags[tag] array */
829 if (flags & RADIX_TREE_ITER_TAGGED) {
830 unsigned tag_long, tag_bit;
831
832 tag_long = offset / BITS_PER_LONG;
833 tag_bit = offset % BITS_PER_LONG;
834 iter->tags = node->tags[tag][tag_long] >> tag_bit;
835 /* This never happens if RADIX_TREE_TAG_LONGS == 1 */
836 if (tag_long < RADIX_TREE_TAG_LONGS - 1) {
837 /* Pick tags from next element */
838 if (tag_bit)
839 iter->tags |= node->tags[tag][tag_long + 1] <<
840 (BITS_PER_LONG - tag_bit);
841 /* Clip chunk size, here only BITS_PER_LONG tags */
842 iter->next_index = index + BITS_PER_LONG;
843 }
844 }
845
846 return node->slots + offset;
847}
848EXPORT_SYMBOL(radix_tree_next_chunk);
849
850/**
851 * radix_tree_range_tag_if_tagged - for each item in given range set given
852 * tag if item has another tag set
853 * @root: radix tree root
854 * @first_indexp: pointer to a starting index of a range to scan
855 * @last_index: last index of a range to scan
856 * @nr_to_tag: maximum number items to tag
857 * @iftag: tag index to test
858 * @settag: tag index to set if tested tag is set
859 *
860 * This function scans range of radix tree from first_index to last_index
861 * (inclusive). For each item in the range if iftag is set, the function sets
862 * also settag. The function stops either after tagging nr_to_tag items or
863 * after reaching last_index.
864 *
865 * The tags must be set from the leaf level only and propagated back up the
866 * path to the root. We must do this so that we resolve the full path before
867 * setting any tags on intermediate nodes. If we set tags as we descend, then
868 * we can get to the leaf node and find that the index that has the iftag
869 * set is outside the range we are scanning. This reults in dangling tags and
870 * can lead to problems with later tag operations (e.g. livelocks on lookups).
871 *
872 * The function returns number of leaves where the tag was set and sets
873 * *first_indexp to the first unscanned index.
874 * WARNING! *first_indexp can wrap if last_index is ULONG_MAX. Caller must
875 * be prepared to handle that.
876 */
877unsigned long radix_tree_range_tag_if_tagged(struct radix_tree_root *root,
878 unsigned long *first_indexp, unsigned long last_index,
879 unsigned long nr_to_tag,
880 unsigned int iftag, unsigned int settag)
881{
882 unsigned int height = root->height;
883 struct radix_tree_node *node = NULL;
884 struct radix_tree_node *slot;
885 unsigned int shift;
886 unsigned long tagged = 0;
887 unsigned long index = *first_indexp;
888
889 last_index = min(last_index, radix_tree_maxindex(height));
890 if (index > last_index)
891 return 0;
892 if (!nr_to_tag)
893 return 0;
894 if (!root_tag_get(root, iftag)) {
895 *first_indexp = last_index + 1;
896 return 0;
897 }
898 if (height == 0) {
899 *first_indexp = last_index + 1;
900 root_tag_set(root, settag);
901 return 1;
902 }
903
904 shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
905 slot = indirect_to_ptr(root->rnode);
906
907 for (;;) {
908 unsigned long upindex;
909 int offset;
910
911 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
912 if (!slot->slots[offset])
913 goto next;
914 if (!tag_get(slot, iftag, offset))
915 goto next;
916 if (shift) {
917 /* Go down one level */
918 shift -= RADIX_TREE_MAP_SHIFT;
919 node = slot;
920 slot = slot->slots[offset];
921 continue;
922 }
923
924 /* tag the leaf */
925 tagged++;
926 tag_set(slot, settag, offset);
927
928 /* walk back up the path tagging interior nodes */
929 upindex = index;
930 while (node) {
931 upindex >>= RADIX_TREE_MAP_SHIFT;
932 offset = upindex & RADIX_TREE_MAP_MASK;
933
934 /* stop if we find a node with the tag already set */
935 if (tag_get(node, settag, offset))
936 break;
937 tag_set(node, settag, offset);
938 node = node->parent;
939 }
940
941 /*
942 * Small optimization: now clear that node pointer.
943 * Since all of this slot's ancestors now have the tag set
944 * from setting it above, we have no further need to walk
945 * back up the tree setting tags, until we update slot to
946 * point to another radix_tree_node.
947 */
948 node = NULL;
949
950next:
951 /* Go to next item at level determined by 'shift' */
952 index = ((index >> shift) + 1) << shift;
953 /* Overflow can happen when last_index is ~0UL... */
954 if (index > last_index || !index)
955 break;
956 if (tagged >= nr_to_tag)
957 break;
958 while (((index >> shift) & RADIX_TREE_MAP_MASK) == 0) {
959 /*
960 * We've fully scanned this node. Go up. Because
961 * last_index is guaranteed to be in the tree, what
962 * we do below cannot wander astray.
963 */
964 slot = slot->parent;
965 shift += RADIX_TREE_MAP_SHIFT;
966 }
967 }
968 /*
969 * We need not to tag the root tag if there is no tag which is set with
970 * settag within the range from *first_indexp to last_index.
971 */
972 if (tagged > 0)
973 root_tag_set(root, settag);
974 *first_indexp = index;
975
976 return tagged;
977}
978EXPORT_SYMBOL(radix_tree_range_tag_if_tagged);
979
980/**
981 * radix_tree_gang_lookup - perform multiple lookup on a radix tree
982 * @root: radix tree root
983 * @results: where the results of the lookup are placed
984 * @first_index: start the lookup from this key
985 * @max_items: place up to this many items at *results
986 *
987 * Performs an index-ascending scan of the tree for present items. Places
988 * them at *@results and returns the number of items which were placed at
989 * *@results.
990 *
991 * The implementation is naive.
992 *
993 * Like radix_tree_lookup, radix_tree_gang_lookup may be called under
994 * rcu_read_lock. In this case, rather than the returned results being
995 * an atomic snapshot of the tree at a single point in time, the semantics
996 * of an RCU protected gang lookup are as though multiple radix_tree_lookups
997 * have been issued in individual locks, and results stored in 'results'.
998 */
999unsigned int
1000radix_tree_gang_lookup(struct radix_tree_root *root, void **results,
1001 unsigned long first_index, unsigned int max_items)
1002{
1003 struct radix_tree_iter iter;
1004 void **slot;
1005 unsigned int ret = 0;
1006
1007 if (unlikely(!max_items))
1008 return 0;
1009
1010 radix_tree_for_each_slot(slot, root, &iter, first_index) {
1011 results[ret] = indirect_to_ptr(rcu_dereference_raw(*slot));
1012 if (!results[ret])
1013 continue;
1014 if (++ret == max_items)
1015 break;
1016 }
1017
1018 return ret;
1019}
1020EXPORT_SYMBOL(radix_tree_gang_lookup);
1021
1022/**
1023 * radix_tree_gang_lookup_slot - perform multiple slot lookup on radix tree
1024 * @root: radix tree root
1025 * @results: where the results of the lookup are placed
1026 * @indices: where their indices should be placed (but usually NULL)
1027 * @first_index: start the lookup from this key
1028 * @max_items: place up to this many items at *results
1029 *
1030 * Performs an index-ascending scan of the tree for present items. Places
1031 * their slots at *@results and returns the number of items which were
1032 * placed at *@results.
1033 *
1034 * The implementation is naive.
1035 *
1036 * Like radix_tree_gang_lookup as far as RCU and locking goes. Slots must
1037 * be dereferenced with radix_tree_deref_slot, and if using only RCU
1038 * protection, radix_tree_deref_slot may fail requiring a retry.
1039 */
1040unsigned int
1041radix_tree_gang_lookup_slot(struct radix_tree_root *root,
1042 void ***results, unsigned long *indices,
1043 unsigned long first_index, unsigned int max_items)
1044{
1045 struct radix_tree_iter iter;
1046 void **slot;
1047 unsigned int ret = 0;
1048
1049 if (unlikely(!max_items))
1050 return 0;
1051
1052 radix_tree_for_each_slot(slot, root, &iter, first_index) {
1053 results[ret] = slot;
1054 if (indices)
1055 indices[ret] = iter.index;
1056 if (++ret == max_items)
1057 break;
1058 }
1059
1060 return ret;
1061}
1062EXPORT_SYMBOL(radix_tree_gang_lookup_slot);
1063
1064/**
1065 * radix_tree_gang_lookup_tag - perform multiple lookup on a radix tree
1066 * based on a tag
1067 * @root: radix tree root
1068 * @results: where the results of the lookup are placed
1069 * @first_index: start the lookup from this key
1070 * @max_items: place up to this many items at *results
1071 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1072 *
1073 * Performs an index-ascending scan of the tree for present items which
1074 * have the tag indexed by @tag set. Places the items at *@results and
1075 * returns the number of items which were placed at *@results.
1076 */
1077unsigned int
1078radix_tree_gang_lookup_tag(struct radix_tree_root *root, void **results,
1079 unsigned long first_index, unsigned int max_items,
1080 unsigned int tag)
1081{
1082 struct radix_tree_iter iter;
1083 void **slot;
1084 unsigned int ret = 0;
1085
1086 if (unlikely(!max_items))
1087 return 0;
1088
1089 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1090 results[ret] = indirect_to_ptr(rcu_dereference_raw(*slot));
1091 if (!results[ret])
1092 continue;
1093 if (++ret == max_items)
1094 break;
1095 }
1096
1097 return ret;
1098}
1099EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
1100
1101/**
1102 * radix_tree_gang_lookup_tag_slot - perform multiple slot lookup on a
1103 * radix tree based on a tag
1104 * @root: radix tree root
1105 * @results: where the results of the lookup are placed
1106 * @first_index: start the lookup from this key
1107 * @max_items: place up to this many items at *results
1108 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1109 *
1110 * Performs an index-ascending scan of the tree for present items which
1111 * have the tag indexed by @tag set. Places the slots at *@results and
1112 * returns the number of slots which were placed at *@results.
1113 */
1114unsigned int
1115radix_tree_gang_lookup_tag_slot(struct radix_tree_root *root, void ***results,
1116 unsigned long first_index, unsigned int max_items,
1117 unsigned int tag)
1118{
1119 struct radix_tree_iter iter;
1120 void **slot;
1121 unsigned int ret = 0;
1122
1123 if (unlikely(!max_items))
1124 return 0;
1125
1126 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1127 results[ret] = slot;
1128 if (++ret == max_items)
1129 break;
1130 }
1131
1132 return ret;
1133}
1134EXPORT_SYMBOL(radix_tree_gang_lookup_tag_slot);
1135
1136#if defined(CONFIG_SHMEM) && defined(CONFIG_SWAP)
1137#include <linux/sched.h> /* for cond_resched() */
1138
1139/*
1140 * This linear search is at present only useful to shmem_unuse_inode().
1141 */
1142static unsigned long __locate(struct radix_tree_node *slot, void *item,
1143 unsigned long index, unsigned long *found_index)
1144{
1145 unsigned int shift, height;
1146 unsigned long i;
1147
1148 height = slot->path & RADIX_TREE_HEIGHT_MASK;
1149 shift = (height-1) * RADIX_TREE_MAP_SHIFT;
1150
1151 for ( ; height > 1; height--) {
1152 i = (index >> shift) & RADIX_TREE_MAP_MASK;
1153 for (;;) {
1154 if (slot->slots[i] != NULL)
1155 break;
1156 index &= ~((1UL << shift) - 1);
1157 index += 1UL << shift;
1158 if (index == 0)
1159 goto out; /* 32-bit wraparound */
1160 i++;
1161 if (i == RADIX_TREE_MAP_SIZE)
1162 goto out;
1163 }
1164
1165 shift -= RADIX_TREE_MAP_SHIFT;
1166 slot = rcu_dereference_raw(slot->slots[i]);
1167 if (slot == NULL)
1168 goto out;
1169 }
1170
1171 /* Bottom level: check items */
1172 for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
1173 if (slot->slots[i] == item) {
1174 *found_index = index + i;
1175 index = 0;
1176 goto out;
1177 }
1178 }
1179 index += RADIX_TREE_MAP_SIZE;
1180out:
1181 return index;
1182}
1183
1184/**
1185 * radix_tree_locate_item - search through radix tree for item
1186 * @root: radix tree root
1187 * @item: item to be found
1188 *
1189 * Returns index where item was found, or -1 if not found.
1190 * Caller must hold no lock (since this time-consuming function needs
1191 * to be preemptible), and must check afterwards if item is still there.
1192 */
1193unsigned long radix_tree_locate_item(struct radix_tree_root *root, void *item)
1194{
1195 struct radix_tree_node *node;
1196 unsigned long max_index;
1197 unsigned long cur_index = 0;
1198 unsigned long found_index = -1;
1199
1200 do {
1201 rcu_read_lock();
1202 node = rcu_dereference_raw(root->rnode);
1203 if (!radix_tree_is_indirect_ptr(node)) {
1204 rcu_read_unlock();
1205 if (node == item)
1206 found_index = 0;
1207 break;
1208 }
1209
1210 node = indirect_to_ptr(node);
1211 max_index = radix_tree_maxindex(node->path &
1212 RADIX_TREE_HEIGHT_MASK);
1213 if (cur_index > max_index) {
1214 rcu_read_unlock();
1215 break;
1216 }
1217
1218 cur_index = __locate(node, item, cur_index, &found_index);
1219 rcu_read_unlock();
1220 cond_resched();
1221 } while (cur_index != 0 && cur_index <= max_index);
1222
1223 return found_index;
1224}
1225#else
1226unsigned long radix_tree_locate_item(struct radix_tree_root *root, void *item)
1227{
1228 return -1;
1229}
1230#endif /* CONFIG_SHMEM && CONFIG_SWAP */
1231
1232/**
1233 * radix_tree_shrink - shrink height of a radix tree to minimal
1234 * @root radix tree root
1235 */
1236static inline void radix_tree_shrink(struct radix_tree_root *root)
1237{
1238 /* try to shrink tree height */
1239 while (root->height > 0) {
1240 struct radix_tree_node *to_free = root->rnode;
1241 struct radix_tree_node *slot;
1242
1243 BUG_ON(!radix_tree_is_indirect_ptr(to_free));
1244 to_free = indirect_to_ptr(to_free);
1245
1246 /*
1247 * The candidate node has more than one child, or its child
1248 * is not at the leftmost slot, we cannot shrink.
1249 */
1250 if (to_free->count != 1)
1251 break;
1252 if (!to_free->slots[0])
1253 break;
1254
1255 /*
1256 * We don't need rcu_assign_pointer(), since we are simply
1257 * moving the node from one part of the tree to another: if it
1258 * was safe to dereference the old pointer to it
1259 * (to_free->slots[0]), it will be safe to dereference the new
1260 * one (root->rnode) as far as dependent read barriers go.
1261 */
1262 slot = to_free->slots[0];
1263 if (root->height > 1) {
1264 slot->parent = NULL;
1265 slot = ptr_to_indirect(slot);
1266 }
1267 root->rnode = slot;
1268 root->height--;
1269
1270 /*
1271 * We have a dilemma here. The node's slot[0] must not be
1272 * NULLed in case there are concurrent lookups expecting to
1273 * find the item. However if this was a bottom-level node,
1274 * then it may be subject to the slot pointer being visible
1275 * to callers dereferencing it. If item corresponding to
1276 * slot[0] is subsequently deleted, these callers would expect
1277 * their slot to become empty sooner or later.
1278 *
1279 * For example, lockless pagecache will look up a slot, deref
1280 * the page pointer, and if the page is 0 refcount it means it
1281 * was concurrently deleted from pagecache so try the deref
1282 * again. Fortunately there is already a requirement for logic
1283 * to retry the entire slot lookup -- the indirect pointer
1284 * problem (replacing direct root node with an indirect pointer
1285 * also results in a stale slot). So tag the slot as indirect
1286 * to force callers to retry.
1287 */
1288 if (root->height == 0)
1289 *((unsigned long *)&to_free->slots[0]) |=
1290 RADIX_TREE_INDIRECT_PTR;
1291
1292 radix_tree_node_free(to_free);
1293 }
1294}
1295
1296/**
1297 * __radix_tree_delete_node - try to free node after clearing a slot
1298 * @root: radix tree root
1299 * @index: index key
1300 * @node: node containing @index
1301 *
1302 * After clearing the slot at @index in @node from radix tree
1303 * rooted at @root, call this function to attempt freeing the
1304 * node and shrinking the tree.
1305 *
1306 * Returns %true if @node was freed, %false otherwise.
1307 */
1308bool __radix_tree_delete_node(struct radix_tree_root *root,
1309 struct radix_tree_node *node)
1310{
1311 bool deleted = false;
1312
1313 do {
1314 struct radix_tree_node *parent;
1315
1316 if (node->count) {
1317 if (node == indirect_to_ptr(root->rnode)) {
1318 radix_tree_shrink(root);
1319 if (root->height == 0)
1320 deleted = true;
1321 }
1322 return deleted;
1323 }
1324
1325 parent = node->parent;
1326 if (parent) {
1327 unsigned int offset;
1328
1329 offset = node->path >> RADIX_TREE_HEIGHT_SHIFT;
1330 parent->slots[offset] = NULL;
1331 parent->count--;
1332 } else {
1333 root_tag_clear_all(root);
1334 root->height = 0;
1335 root->rnode = NULL;
1336 }
1337
1338 radix_tree_node_free(node);
1339 deleted = true;
1340
1341 node = parent;
1342 } while (node);
1343
1344 return deleted;
1345}
1346
1347/**
1348 * radix_tree_delete_item - delete an item from a radix tree
1349 * @root: radix tree root
1350 * @index: index key
1351 * @item: expected item
1352 *
1353 * Remove @item at @index from the radix tree rooted at @root.
1354 *
1355 * Returns the address of the deleted item, or NULL if it was not present
1356 * or the entry at the given @index was not @item.
1357 */
1358void *radix_tree_delete_item(struct radix_tree_root *root,
1359 unsigned long index, void *item)
1360{
1361 struct radix_tree_node *node;
1362 unsigned int offset;
1363 void **slot;
1364 void *entry;
1365 int tag;
1366
1367 entry = __radix_tree_lookup(root, index, &node, &slot);
1368 if (!entry)
1369 return NULL;
1370
1371 if (item && entry != item)
1372 return NULL;
1373
1374 if (!node) {
1375 root_tag_clear_all(root);
1376 root->rnode = NULL;
1377 return entry;
1378 }
1379
1380 offset = index & RADIX_TREE_MAP_MASK;
1381
1382 /*
1383 * Clear all tags associated with the item to be deleted.
1384 * This way of doing it would be inefficient, but seldom is any set.
1385 */
1386 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
1387 if (tag_get(node, tag, offset))
1388 radix_tree_tag_clear(root, index, tag);
1389 }
1390
1391 node->slots[offset] = NULL;
1392 node->count--;
1393
1394 __radix_tree_delete_node(root, node);
1395
1396 return entry;
1397}
1398EXPORT_SYMBOL(radix_tree_delete_item);
1399
1400/**
1401 * radix_tree_delete - delete an item from a radix tree
1402 * @root: radix tree root
1403 * @index: index key
1404 *
1405 * Remove the item at @index from the radix tree rooted at @root.
1406 *
1407 * Returns the address of the deleted item, or NULL if it was not present.
1408 */
1409void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
1410{
1411 return radix_tree_delete_item(root, index, NULL);
1412}
1413EXPORT_SYMBOL(radix_tree_delete);
1414
1415/**
1416 * radix_tree_tagged - test whether any items in the tree are tagged
1417 * @root: radix tree root
1418 * @tag: tag to test
1419 */
1420int radix_tree_tagged(struct radix_tree_root *root, unsigned int tag)
1421{
1422 return root_tag_get(root, tag);
1423}
1424EXPORT_SYMBOL(radix_tree_tagged);
1425
1426static void
1427radix_tree_node_ctor(void *arg)
1428{
1429 struct radix_tree_node *node = arg;
1430
1431 memset(node, 0, sizeof(*node));
1432 INIT_LIST_HEAD(&node->private_list);
1433}
1434
1435static __init unsigned long __maxindex(unsigned int height)
1436{
1437 unsigned int width = height * RADIX_TREE_MAP_SHIFT;
1438 int shift = RADIX_TREE_INDEX_BITS - width;
1439
1440 if (shift < 0)
1441 return ~0UL;
1442 if (shift >= BITS_PER_LONG)
1443 return 0UL;
1444 return ~0UL >> shift;
1445}
1446
1447static __init void radix_tree_init_maxindex(void)
1448{
1449 unsigned int i;
1450
1451 for (i = 0; i < ARRAY_SIZE(height_to_maxindex); i++)
1452 height_to_maxindex[i] = __maxindex(i);
1453}
1454
1455static int radix_tree_callback(struct notifier_block *nfb,
1456 unsigned long action,
1457 void *hcpu)
1458{
1459 int cpu = (long)hcpu;
1460 struct radix_tree_preload *rtp;
1461
1462 /* Free per-cpu pool of perloaded nodes */
1463 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
1464 rtp = &per_cpu(radix_tree_preloads, cpu);
1465 while (rtp->nr) {
1466 kmem_cache_free(radix_tree_node_cachep,
1467 rtp->nodes[rtp->nr-1]);
1468 rtp->nodes[rtp->nr-1] = NULL;
1469 rtp->nr--;
1470 }
1471 }
1472 return NOTIFY_OK;
1473}
1474
1475void __init radix_tree_init(void)
1476{
1477 radix_tree_node_cachep = kmem_cache_create("radix_tree_node",
1478 sizeof(struct radix_tree_node), 0,
1479 SLAB_PANIC | SLAB_RECLAIM_ACCOUNT,
1480 radix_tree_node_ctor);
1481 radix_tree_init_maxindex();
1482 hotcpu_notifier(radix_tree_callback, 0);
1483}
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 * Copyright (C) 2016 Intel, Matthew Wilcox
8 * Copyright (C) 2016 Intel, Ross Zwisler
9 *
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.
14 *
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.
19 *
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.
23 */
24
25#include <linux/cpu.h>
26#include <linux/errno.h>
27#include <linux/init.h>
28#include <linux/kernel.h>
29#include <linux/export.h>
30#include <linux/radix-tree.h>
31#include <linux/percpu.h>
32#include <linux/slab.h>
33#include <linux/kmemleak.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() */
39
40
41/* Number of nodes in fully populated tree of given height */
42static unsigned long height_to_maxnodes[RADIX_TREE_MAX_PATH + 1] __read_mostly;
43
44/*
45 * Radix tree node cache.
46 */
47static struct kmem_cache *radix_tree_node_cachep;
48
49/*
50 * The radix tree is variable-height, so an insert operation not only has
51 * to build the branch to its corresponding item, it also has to build the
52 * branch to existing items if the size has to be increased (by
53 * radix_tree_extend).
54 *
55 * The worst case is a zero height tree with just a single item at index 0,
56 * and then inserting an item at index ULONG_MAX. This requires 2 new branches
57 * of RADIX_TREE_MAX_PATH size to be created, with only the root node shared.
58 * Hence:
59 */
60#define RADIX_TREE_PRELOAD_SIZE (RADIX_TREE_MAX_PATH * 2 - 1)
61
62/*
63 * Per-cpu pool of preloaded nodes
64 */
65struct radix_tree_preload {
66 unsigned nr;
67 /* nodes->private_data points to next preallocated node */
68 struct radix_tree_node *nodes;
69};
70static DEFINE_PER_CPU(struct radix_tree_preload, radix_tree_preloads) = { 0, };
71
72static inline struct radix_tree_node *entry_to_node(void *ptr)
73{
74 return (void *)((unsigned long)ptr & ~RADIX_TREE_INTERNAL_NODE);
75}
76
77static inline void *node_to_entry(void *ptr)
78{
79 return (void *)((unsigned long)ptr | RADIX_TREE_INTERNAL_NODE);
80}
81
82#define RADIX_TREE_RETRY node_to_entry(NULL)
83
84#ifdef CONFIG_RADIX_TREE_MULTIORDER
85/* Sibling slots point directly to another slot in the same node */
86static inline bool is_sibling_entry(struct radix_tree_node *parent, void *node)
87{
88 void **ptr = node;
89 return (parent->slots <= ptr) &&
90 (ptr < parent->slots + RADIX_TREE_MAP_SIZE);
91}
92#else
93static inline bool is_sibling_entry(struct radix_tree_node *parent, void *node)
94{
95 return false;
96}
97#endif
98
99static inline unsigned long get_slot_offset(struct radix_tree_node *parent,
100 void **slot)
101{
102 return slot - parent->slots;
103}
104
105static unsigned int radix_tree_descend(struct radix_tree_node *parent,
106 struct radix_tree_node **nodep, unsigned long index)
107{
108 unsigned int offset = (index >> parent->shift) & RADIX_TREE_MAP_MASK;
109 void **entry = rcu_dereference_raw(parent->slots[offset]);
110
111#ifdef CONFIG_RADIX_TREE_MULTIORDER
112 if (radix_tree_is_internal_node(entry)) {
113 if (is_sibling_entry(parent, entry)) {
114 void **sibentry = (void **) entry_to_node(entry);
115 offset = get_slot_offset(parent, sibentry);
116 entry = rcu_dereference_raw(*sibentry);
117 }
118 }
119#endif
120
121 *nodep = (void *)entry;
122 return offset;
123}
124
125static inline gfp_t root_gfp_mask(struct radix_tree_root *root)
126{
127 return root->gfp_mask & __GFP_BITS_MASK;
128}
129
130static inline void tag_set(struct radix_tree_node *node, unsigned int tag,
131 int offset)
132{
133 __set_bit(offset, node->tags[tag]);
134}
135
136static inline void tag_clear(struct radix_tree_node *node, unsigned int tag,
137 int offset)
138{
139 __clear_bit(offset, node->tags[tag]);
140}
141
142static inline int tag_get(struct radix_tree_node *node, unsigned int tag,
143 int offset)
144{
145 return test_bit(offset, node->tags[tag]);
146}
147
148static inline void root_tag_set(struct radix_tree_root *root, unsigned int tag)
149{
150 root->gfp_mask |= (__force gfp_t)(1 << (tag + __GFP_BITS_SHIFT));
151}
152
153static inline void root_tag_clear(struct radix_tree_root *root, unsigned tag)
154{
155 root->gfp_mask &= (__force gfp_t)~(1 << (tag + __GFP_BITS_SHIFT));
156}
157
158static inline void root_tag_clear_all(struct radix_tree_root *root)
159{
160 root->gfp_mask &= __GFP_BITS_MASK;
161}
162
163static inline int root_tag_get(struct radix_tree_root *root, unsigned int tag)
164{
165 return (__force int)root->gfp_mask & (1 << (tag + __GFP_BITS_SHIFT));
166}
167
168static inline unsigned root_tags_get(struct radix_tree_root *root)
169{
170 return (__force unsigned)root->gfp_mask >> __GFP_BITS_SHIFT;
171}
172
173/*
174 * Returns 1 if any slot in the node has this tag set.
175 * Otherwise returns 0.
176 */
177static inline int any_tag_set(struct radix_tree_node *node, unsigned int tag)
178{
179 unsigned idx;
180 for (idx = 0; idx < RADIX_TREE_TAG_LONGS; idx++) {
181 if (node->tags[tag][idx])
182 return 1;
183 }
184 return 0;
185}
186
187/**
188 * radix_tree_find_next_bit - find the next set bit in a memory region
189 *
190 * @addr: The address to base the search on
191 * @size: The bitmap size in bits
192 * @offset: The bitnumber to start searching at
193 *
194 * Unrollable variant of find_next_bit() for constant size arrays.
195 * Tail bits starting from size to roundup(size, BITS_PER_LONG) must be zero.
196 * Returns next bit offset, or size if nothing found.
197 */
198static __always_inline unsigned long
199radix_tree_find_next_bit(struct radix_tree_node *node, unsigned int tag,
200 unsigned long offset)
201{
202 const unsigned long *addr = node->tags[tag];
203
204 if (offset < RADIX_TREE_MAP_SIZE) {
205 unsigned long tmp;
206
207 addr += offset / BITS_PER_LONG;
208 tmp = *addr >> (offset % BITS_PER_LONG);
209 if (tmp)
210 return __ffs(tmp) + offset;
211 offset = (offset + BITS_PER_LONG) & ~(BITS_PER_LONG - 1);
212 while (offset < RADIX_TREE_MAP_SIZE) {
213 tmp = *++addr;
214 if (tmp)
215 return __ffs(tmp) + offset;
216 offset += BITS_PER_LONG;
217 }
218 }
219 return RADIX_TREE_MAP_SIZE;
220}
221
222static unsigned int iter_offset(const struct radix_tree_iter *iter)
223{
224 return (iter->index >> iter_shift(iter)) & RADIX_TREE_MAP_MASK;
225}
226
227/*
228 * The maximum index which can be stored in a radix tree
229 */
230static inline unsigned long shift_maxindex(unsigned int shift)
231{
232 return (RADIX_TREE_MAP_SIZE << shift) - 1;
233}
234
235static inline unsigned long node_maxindex(struct radix_tree_node *node)
236{
237 return shift_maxindex(node->shift);
238}
239
240#ifndef __KERNEL__
241static void dump_node(struct radix_tree_node *node, unsigned long index)
242{
243 unsigned long i;
244
245 pr_debug("radix node: %p offset %d indices %lu-%lu parent %p tags %lx %lx %lx shift %d count %d exceptional %d\n",
246 node, node->offset, index, index | node_maxindex(node),
247 node->parent,
248 node->tags[0][0], node->tags[1][0], node->tags[2][0],
249 node->shift, node->count, node->exceptional);
250
251 for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
252 unsigned long first = index | (i << node->shift);
253 unsigned long last = first | ((1UL << node->shift) - 1);
254 void *entry = node->slots[i];
255 if (!entry)
256 continue;
257 if (entry == RADIX_TREE_RETRY) {
258 pr_debug("radix retry offset %ld indices %lu-%lu parent %p\n",
259 i, first, last, node);
260 } else if (!radix_tree_is_internal_node(entry)) {
261 pr_debug("radix entry %p offset %ld indices %lu-%lu parent %p\n",
262 entry, i, first, last, node);
263 } else if (is_sibling_entry(node, entry)) {
264 pr_debug("radix sblng %p offset %ld indices %lu-%lu parent %p val %p\n",
265 entry, i, first, last, node,
266 *(void **)entry_to_node(entry));
267 } else {
268 dump_node(entry_to_node(entry), first);
269 }
270 }
271}
272
273/* For debug */
274static void radix_tree_dump(struct radix_tree_root *root)
275{
276 pr_debug("radix root: %p rnode %p tags %x\n",
277 root, root->rnode,
278 root->gfp_mask >> __GFP_BITS_SHIFT);
279 if (!radix_tree_is_internal_node(root->rnode))
280 return;
281 dump_node(entry_to_node(root->rnode), 0);
282}
283#endif
284
285/*
286 * This assumes that the caller has performed appropriate preallocation, and
287 * that the caller has pinned this thread of control to the current CPU.
288 */
289static struct radix_tree_node *
290radix_tree_node_alloc(struct radix_tree_root *root,
291 struct radix_tree_node *parent,
292 unsigned int shift, unsigned int offset,
293 unsigned int count, unsigned int exceptional)
294{
295 struct radix_tree_node *ret = NULL;
296 gfp_t gfp_mask = root_gfp_mask(root);
297
298 /*
299 * Preload code isn't irq safe and it doesn't make sense to use
300 * preloading during an interrupt anyway as all the allocations have
301 * to be atomic. So just do normal allocation when in interrupt.
302 */
303 if (!gfpflags_allow_blocking(gfp_mask) && !in_interrupt()) {
304 struct radix_tree_preload *rtp;
305
306 /*
307 * Even if the caller has preloaded, try to allocate from the
308 * cache first for the new node to get accounted to the memory
309 * cgroup.
310 */
311 ret = kmem_cache_alloc(radix_tree_node_cachep,
312 gfp_mask | __GFP_NOWARN);
313 if (ret)
314 goto out;
315
316 /*
317 * Provided the caller has preloaded here, we will always
318 * succeed in getting a node here (and never reach
319 * kmem_cache_alloc)
320 */
321 rtp = this_cpu_ptr(&radix_tree_preloads);
322 if (rtp->nr) {
323 ret = rtp->nodes;
324 rtp->nodes = ret->private_data;
325 ret->private_data = NULL;
326 rtp->nr--;
327 }
328 /*
329 * Update the allocation stack trace as this is more useful
330 * for debugging.
331 */
332 kmemleak_update_trace(ret);
333 goto out;
334 }
335 ret = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
336out:
337 BUG_ON(radix_tree_is_internal_node(ret));
338 if (ret) {
339 ret->parent = parent;
340 ret->shift = shift;
341 ret->offset = offset;
342 ret->count = count;
343 ret->exceptional = exceptional;
344 }
345 return ret;
346}
347
348static void radix_tree_node_rcu_free(struct rcu_head *head)
349{
350 struct radix_tree_node *node =
351 container_of(head, struct radix_tree_node, rcu_head);
352
353 /*
354 * Must only free zeroed nodes into the slab. We can be left with
355 * non-NULL entries by radix_tree_free_nodes, so clear the entries
356 * and tags here.
357 */
358 memset(node->slots, 0, sizeof(node->slots));
359 memset(node->tags, 0, sizeof(node->tags));
360 INIT_LIST_HEAD(&node->private_list);
361
362 kmem_cache_free(radix_tree_node_cachep, node);
363}
364
365static inline void
366radix_tree_node_free(struct radix_tree_node *node)
367{
368 call_rcu(&node->rcu_head, radix_tree_node_rcu_free);
369}
370
371/*
372 * Load up this CPU's radix_tree_node buffer with sufficient objects to
373 * ensure that the addition of a single element in the tree cannot fail. On
374 * success, return zero, with preemption disabled. On error, return -ENOMEM
375 * with preemption not disabled.
376 *
377 * To make use of this facility, the radix tree must be initialised without
378 * __GFP_DIRECT_RECLAIM being passed to INIT_RADIX_TREE().
379 */
380static int __radix_tree_preload(gfp_t gfp_mask, unsigned nr)
381{
382 struct radix_tree_preload *rtp;
383 struct radix_tree_node *node;
384 int ret = -ENOMEM;
385
386 /*
387 * Nodes preloaded by one cgroup can be be used by another cgroup, so
388 * they should never be accounted to any particular memory cgroup.
389 */
390 gfp_mask &= ~__GFP_ACCOUNT;
391
392 preempt_disable();
393 rtp = this_cpu_ptr(&radix_tree_preloads);
394 while (rtp->nr < nr) {
395 preempt_enable();
396 node = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
397 if (node == NULL)
398 goto out;
399 preempt_disable();
400 rtp = this_cpu_ptr(&radix_tree_preloads);
401 if (rtp->nr < nr) {
402 node->private_data = rtp->nodes;
403 rtp->nodes = node;
404 rtp->nr++;
405 } else {
406 kmem_cache_free(radix_tree_node_cachep, node);
407 }
408 }
409 ret = 0;
410out:
411 return ret;
412}
413
414/*
415 * Load up this CPU's radix_tree_node buffer with sufficient objects to
416 * ensure that the addition of a single element in the tree cannot fail. On
417 * success, return zero, with preemption disabled. On error, return -ENOMEM
418 * with preemption not disabled.
419 *
420 * To make use of this facility, the radix tree must be initialised without
421 * __GFP_DIRECT_RECLAIM being passed to INIT_RADIX_TREE().
422 */
423int radix_tree_preload(gfp_t gfp_mask)
424{
425 /* Warn on non-sensical use... */
426 WARN_ON_ONCE(!gfpflags_allow_blocking(gfp_mask));
427 return __radix_tree_preload(gfp_mask, RADIX_TREE_PRELOAD_SIZE);
428}
429EXPORT_SYMBOL(radix_tree_preload);
430
431/*
432 * The same as above function, except we don't guarantee preloading happens.
433 * We do it, if we decide it helps. On success, return zero with preemption
434 * disabled. On error, return -ENOMEM with preemption not disabled.
435 */
436int radix_tree_maybe_preload(gfp_t gfp_mask)
437{
438 if (gfpflags_allow_blocking(gfp_mask))
439 return __radix_tree_preload(gfp_mask, RADIX_TREE_PRELOAD_SIZE);
440 /* Preloading doesn't help anything with this gfp mask, skip it */
441 preempt_disable();
442 return 0;
443}
444EXPORT_SYMBOL(radix_tree_maybe_preload);
445
446#ifdef CONFIG_RADIX_TREE_MULTIORDER
447/*
448 * Preload with enough objects to ensure that we can split a single entry
449 * of order @old_order into many entries of size @new_order
450 */
451int radix_tree_split_preload(unsigned int old_order, unsigned int new_order,
452 gfp_t gfp_mask)
453{
454 unsigned top = 1 << (old_order % RADIX_TREE_MAP_SHIFT);
455 unsigned layers = (old_order / RADIX_TREE_MAP_SHIFT) -
456 (new_order / RADIX_TREE_MAP_SHIFT);
457 unsigned nr = 0;
458
459 WARN_ON_ONCE(!gfpflags_allow_blocking(gfp_mask));
460 BUG_ON(new_order >= old_order);
461
462 while (layers--)
463 nr = nr * RADIX_TREE_MAP_SIZE + 1;
464 return __radix_tree_preload(gfp_mask, top * nr);
465}
466#endif
467
468/*
469 * The same as function above, but preload number of nodes required to insert
470 * (1 << order) continuous naturally-aligned elements.
471 */
472int radix_tree_maybe_preload_order(gfp_t gfp_mask, int order)
473{
474 unsigned long nr_subtrees;
475 int nr_nodes, subtree_height;
476
477 /* Preloading doesn't help anything with this gfp mask, skip it */
478 if (!gfpflags_allow_blocking(gfp_mask)) {
479 preempt_disable();
480 return 0;
481 }
482
483 /*
484 * Calculate number and height of fully populated subtrees it takes to
485 * store (1 << order) elements.
486 */
487 nr_subtrees = 1 << order;
488 for (subtree_height = 0; nr_subtrees > RADIX_TREE_MAP_SIZE;
489 subtree_height++)
490 nr_subtrees >>= RADIX_TREE_MAP_SHIFT;
491
492 /*
493 * The worst case is zero height tree with a single item at index 0 and
494 * then inserting items starting at ULONG_MAX - (1 << order).
495 *
496 * This requires RADIX_TREE_MAX_PATH nodes to build branch from root to
497 * 0-index item.
498 */
499 nr_nodes = RADIX_TREE_MAX_PATH;
500
501 /* Plus branch to fully populated subtrees. */
502 nr_nodes += RADIX_TREE_MAX_PATH - subtree_height;
503
504 /* Root node is shared. */
505 nr_nodes--;
506
507 /* Plus nodes required to build subtrees. */
508 nr_nodes += nr_subtrees * height_to_maxnodes[subtree_height];
509
510 return __radix_tree_preload(gfp_mask, nr_nodes);
511}
512
513static unsigned radix_tree_load_root(struct radix_tree_root *root,
514 struct radix_tree_node **nodep, unsigned long *maxindex)
515{
516 struct radix_tree_node *node = rcu_dereference_raw(root->rnode);
517
518 *nodep = node;
519
520 if (likely(radix_tree_is_internal_node(node))) {
521 node = entry_to_node(node);
522 *maxindex = node_maxindex(node);
523 return node->shift + RADIX_TREE_MAP_SHIFT;
524 }
525
526 *maxindex = 0;
527 return 0;
528}
529
530/*
531 * Extend a radix tree so it can store key @index.
532 */
533static int radix_tree_extend(struct radix_tree_root *root,
534 unsigned long index, unsigned int shift)
535{
536 struct radix_tree_node *slot;
537 unsigned int maxshift;
538 int tag;
539
540 /* Figure out what the shift should be. */
541 maxshift = shift;
542 while (index > shift_maxindex(maxshift))
543 maxshift += RADIX_TREE_MAP_SHIFT;
544
545 slot = root->rnode;
546 if (!slot)
547 goto out;
548
549 do {
550 struct radix_tree_node *node = radix_tree_node_alloc(root,
551 NULL, shift, 0, 1, 0);
552 if (!node)
553 return -ENOMEM;
554
555 /* Propagate the aggregated tag info into the new root */
556 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
557 if (root_tag_get(root, tag))
558 tag_set(node, tag, 0);
559 }
560
561 BUG_ON(shift > BITS_PER_LONG);
562 if (radix_tree_is_internal_node(slot)) {
563 entry_to_node(slot)->parent = node;
564 } else if (radix_tree_exceptional_entry(slot)) {
565 /* Moving an exceptional root->rnode to a node */
566 node->exceptional = 1;
567 }
568 node->slots[0] = slot;
569 slot = node_to_entry(node);
570 rcu_assign_pointer(root->rnode, slot);
571 shift += RADIX_TREE_MAP_SHIFT;
572 } while (shift <= maxshift);
573out:
574 return maxshift + RADIX_TREE_MAP_SHIFT;
575}
576
577/**
578 * radix_tree_shrink - shrink radix tree to minimum height
579 * @root radix tree root
580 */
581static inline void radix_tree_shrink(struct radix_tree_root *root,
582 radix_tree_update_node_t update_node,
583 void *private)
584{
585 for (;;) {
586 struct radix_tree_node *node = root->rnode;
587 struct radix_tree_node *child;
588
589 if (!radix_tree_is_internal_node(node))
590 break;
591 node = entry_to_node(node);
592
593 /*
594 * The candidate node has more than one child, or its child
595 * is not at the leftmost slot, or the child is a multiorder
596 * entry, we cannot shrink.
597 */
598 if (node->count != 1)
599 break;
600 child = node->slots[0];
601 if (!child)
602 break;
603 if (!radix_tree_is_internal_node(child) && node->shift)
604 break;
605
606 if (radix_tree_is_internal_node(child))
607 entry_to_node(child)->parent = NULL;
608
609 /*
610 * We don't need rcu_assign_pointer(), since we are simply
611 * moving the node from one part of the tree to another: if it
612 * was safe to dereference the old pointer to it
613 * (node->slots[0]), it will be safe to dereference the new
614 * one (root->rnode) as far as dependent read barriers go.
615 */
616 root->rnode = child;
617
618 /*
619 * We have a dilemma here. The node's slot[0] must not be
620 * NULLed in case there are concurrent lookups expecting to
621 * find the item. However if this was a bottom-level node,
622 * then it may be subject to the slot pointer being visible
623 * to callers dereferencing it. If item corresponding to
624 * slot[0] is subsequently deleted, these callers would expect
625 * their slot to become empty sooner or later.
626 *
627 * For example, lockless pagecache will look up a slot, deref
628 * the page pointer, and if the page has 0 refcount it means it
629 * was concurrently deleted from pagecache so try the deref
630 * again. Fortunately there is already a requirement for logic
631 * to retry the entire slot lookup -- the indirect pointer
632 * problem (replacing direct root node with an indirect pointer
633 * also results in a stale slot). So tag the slot as indirect
634 * to force callers to retry.
635 */
636 node->count = 0;
637 if (!radix_tree_is_internal_node(child)) {
638 node->slots[0] = RADIX_TREE_RETRY;
639 if (update_node)
640 update_node(node, private);
641 }
642
643 WARN_ON_ONCE(!list_empty(&node->private_list));
644 radix_tree_node_free(node);
645 }
646}
647
648static void delete_node(struct radix_tree_root *root,
649 struct radix_tree_node *node,
650 radix_tree_update_node_t update_node, void *private)
651{
652 do {
653 struct radix_tree_node *parent;
654
655 if (node->count) {
656 if (node == entry_to_node(root->rnode))
657 radix_tree_shrink(root, update_node, private);
658 return;
659 }
660
661 parent = node->parent;
662 if (parent) {
663 parent->slots[node->offset] = NULL;
664 parent->count--;
665 } else {
666 root_tag_clear_all(root);
667 root->rnode = NULL;
668 }
669
670 WARN_ON_ONCE(!list_empty(&node->private_list));
671 radix_tree_node_free(node);
672
673 node = parent;
674 } while (node);
675}
676
677/**
678 * __radix_tree_create - create a slot in a radix tree
679 * @root: radix tree root
680 * @index: index key
681 * @order: index occupies 2^order aligned slots
682 * @nodep: returns node
683 * @slotp: returns slot
684 *
685 * Create, if necessary, and return the node and slot for an item
686 * at position @index in the radix tree @root.
687 *
688 * Until there is more than one item in the tree, no nodes are
689 * allocated and @root->rnode is used as a direct slot instead of
690 * pointing to a node, in which case *@nodep will be NULL.
691 *
692 * Returns -ENOMEM, or 0 for success.
693 */
694int __radix_tree_create(struct radix_tree_root *root, unsigned long index,
695 unsigned order, struct radix_tree_node **nodep,
696 void ***slotp)
697{
698 struct radix_tree_node *node = NULL, *child;
699 void **slot = (void **)&root->rnode;
700 unsigned long maxindex;
701 unsigned int shift, offset = 0;
702 unsigned long max = index | ((1UL << order) - 1);
703
704 shift = radix_tree_load_root(root, &child, &maxindex);
705
706 /* Make sure the tree is high enough. */
707 if (order > 0 && max == ((1UL << order) - 1))
708 max++;
709 if (max > maxindex) {
710 int error = radix_tree_extend(root, max, shift);
711 if (error < 0)
712 return error;
713 shift = error;
714 child = root->rnode;
715 }
716
717 while (shift > order) {
718 shift -= RADIX_TREE_MAP_SHIFT;
719 if (child == NULL) {
720 /* Have to add a child node. */
721 child = radix_tree_node_alloc(root, node, shift,
722 offset, 0, 0);
723 if (!child)
724 return -ENOMEM;
725 rcu_assign_pointer(*slot, node_to_entry(child));
726 if (node)
727 node->count++;
728 } else if (!radix_tree_is_internal_node(child))
729 break;
730
731 /* Go a level down */
732 node = entry_to_node(child);
733 offset = radix_tree_descend(node, &child, index);
734 slot = &node->slots[offset];
735 }
736
737 if (nodep)
738 *nodep = node;
739 if (slotp)
740 *slotp = slot;
741 return 0;
742}
743
744#ifdef CONFIG_RADIX_TREE_MULTIORDER
745/*
746 * Free any nodes below this node. The tree is presumed to not need
747 * shrinking, and any user data in the tree is presumed to not need a
748 * destructor called on it. If we need to add a destructor, we can
749 * add that functionality later. Note that we may not clear tags or
750 * slots from the tree as an RCU walker may still have a pointer into
751 * this subtree. We could replace the entries with RADIX_TREE_RETRY,
752 * but we'll still have to clear those in rcu_free.
753 */
754static void radix_tree_free_nodes(struct radix_tree_node *node)
755{
756 unsigned offset = 0;
757 struct radix_tree_node *child = entry_to_node(node);
758
759 for (;;) {
760 void *entry = child->slots[offset];
761 if (radix_tree_is_internal_node(entry) &&
762 !is_sibling_entry(child, entry)) {
763 child = entry_to_node(entry);
764 offset = 0;
765 continue;
766 }
767 offset++;
768 while (offset == RADIX_TREE_MAP_SIZE) {
769 struct radix_tree_node *old = child;
770 offset = child->offset + 1;
771 child = child->parent;
772 WARN_ON_ONCE(!list_empty(&old->private_list));
773 radix_tree_node_free(old);
774 if (old == entry_to_node(node))
775 return;
776 }
777 }
778}
779
780static inline int insert_entries(struct radix_tree_node *node, void **slot,
781 void *item, unsigned order, bool replace)
782{
783 struct radix_tree_node *child;
784 unsigned i, n, tag, offset, tags = 0;
785
786 if (node) {
787 if (order > node->shift)
788 n = 1 << (order - node->shift);
789 else
790 n = 1;
791 offset = get_slot_offset(node, slot);
792 } else {
793 n = 1;
794 offset = 0;
795 }
796
797 if (n > 1) {
798 offset = offset & ~(n - 1);
799 slot = &node->slots[offset];
800 }
801 child = node_to_entry(slot);
802
803 for (i = 0; i < n; i++) {
804 if (slot[i]) {
805 if (replace) {
806 node->count--;
807 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++)
808 if (tag_get(node, tag, offset + i))
809 tags |= 1 << tag;
810 } else
811 return -EEXIST;
812 }
813 }
814
815 for (i = 0; i < n; i++) {
816 struct radix_tree_node *old = slot[i];
817 if (i) {
818 rcu_assign_pointer(slot[i], child);
819 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++)
820 if (tags & (1 << tag))
821 tag_clear(node, tag, offset + i);
822 } else {
823 rcu_assign_pointer(slot[i], item);
824 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++)
825 if (tags & (1 << tag))
826 tag_set(node, tag, offset);
827 }
828 if (radix_tree_is_internal_node(old) &&
829 !is_sibling_entry(node, old) &&
830 (old != RADIX_TREE_RETRY))
831 radix_tree_free_nodes(old);
832 if (radix_tree_exceptional_entry(old))
833 node->exceptional--;
834 }
835 if (node) {
836 node->count += n;
837 if (radix_tree_exceptional_entry(item))
838 node->exceptional += n;
839 }
840 return n;
841}
842#else
843static inline int insert_entries(struct radix_tree_node *node, void **slot,
844 void *item, unsigned order, bool replace)
845{
846 if (*slot)
847 return -EEXIST;
848 rcu_assign_pointer(*slot, item);
849 if (node) {
850 node->count++;
851 if (radix_tree_exceptional_entry(item))
852 node->exceptional++;
853 }
854 return 1;
855}
856#endif
857
858/**
859 * __radix_tree_insert - insert into a radix tree
860 * @root: radix tree root
861 * @index: index key
862 * @order: key covers the 2^order indices around index
863 * @item: item to insert
864 *
865 * Insert an item into the radix tree at position @index.
866 */
867int __radix_tree_insert(struct radix_tree_root *root, unsigned long index,
868 unsigned order, void *item)
869{
870 struct radix_tree_node *node;
871 void **slot;
872 int error;
873
874 BUG_ON(radix_tree_is_internal_node(item));
875
876 error = __radix_tree_create(root, index, order, &node, &slot);
877 if (error)
878 return error;
879
880 error = insert_entries(node, slot, item, order, false);
881 if (error < 0)
882 return error;
883
884 if (node) {
885 unsigned offset = get_slot_offset(node, slot);
886 BUG_ON(tag_get(node, 0, offset));
887 BUG_ON(tag_get(node, 1, offset));
888 BUG_ON(tag_get(node, 2, offset));
889 } else {
890 BUG_ON(root_tags_get(root));
891 }
892
893 return 0;
894}
895EXPORT_SYMBOL(__radix_tree_insert);
896
897/**
898 * __radix_tree_lookup - lookup an item in a radix tree
899 * @root: radix tree root
900 * @index: index key
901 * @nodep: returns node
902 * @slotp: returns slot
903 *
904 * Lookup and return the item at position @index in the radix
905 * tree @root.
906 *
907 * Until there is more than one item in the tree, no nodes are
908 * allocated and @root->rnode is used as a direct slot instead of
909 * pointing to a node, in which case *@nodep will be NULL.
910 */
911void *__radix_tree_lookup(struct radix_tree_root *root, unsigned long index,
912 struct radix_tree_node **nodep, void ***slotp)
913{
914 struct radix_tree_node *node, *parent;
915 unsigned long maxindex;
916 void **slot;
917
918 restart:
919 parent = NULL;
920 slot = (void **)&root->rnode;
921 radix_tree_load_root(root, &node, &maxindex);
922 if (index > maxindex)
923 return NULL;
924
925 while (radix_tree_is_internal_node(node)) {
926 unsigned offset;
927
928 if (node == RADIX_TREE_RETRY)
929 goto restart;
930 parent = entry_to_node(node);
931 offset = radix_tree_descend(parent, &node, index);
932 slot = parent->slots + offset;
933 }
934
935 if (nodep)
936 *nodep = parent;
937 if (slotp)
938 *slotp = slot;
939 return node;
940}
941
942/**
943 * radix_tree_lookup_slot - lookup a slot in a radix tree
944 * @root: radix tree root
945 * @index: index key
946 *
947 * Returns: the slot corresponding to the position @index in the
948 * radix tree @root. This is useful for update-if-exists operations.
949 *
950 * This function can be called under rcu_read_lock iff the slot is not
951 * modified by radix_tree_replace_slot, otherwise it must be called
952 * exclusive from other writers. Any dereference of the slot must be done
953 * using radix_tree_deref_slot.
954 */
955void **radix_tree_lookup_slot(struct radix_tree_root *root, unsigned long index)
956{
957 void **slot;
958
959 if (!__radix_tree_lookup(root, index, NULL, &slot))
960 return NULL;
961 return slot;
962}
963EXPORT_SYMBOL(radix_tree_lookup_slot);
964
965/**
966 * radix_tree_lookup - perform lookup operation on a radix tree
967 * @root: radix tree root
968 * @index: index key
969 *
970 * Lookup the item at the position @index in the radix tree @root.
971 *
972 * This function can be called under rcu_read_lock, however the caller
973 * must manage lifetimes of leaf nodes (eg. RCU may also be used to free
974 * them safely). No RCU barriers are required to access or modify the
975 * returned item, however.
976 */
977void *radix_tree_lookup(struct radix_tree_root *root, unsigned long index)
978{
979 return __radix_tree_lookup(root, index, NULL, NULL);
980}
981EXPORT_SYMBOL(radix_tree_lookup);
982
983static inline int slot_count(struct radix_tree_node *node,
984 void **slot)
985{
986 int n = 1;
987#ifdef CONFIG_RADIX_TREE_MULTIORDER
988 void *ptr = node_to_entry(slot);
989 unsigned offset = get_slot_offset(node, slot);
990 int i;
991
992 for (i = 1; offset + i < RADIX_TREE_MAP_SIZE; i++) {
993 if (node->slots[offset + i] != ptr)
994 break;
995 n++;
996 }
997#endif
998 return n;
999}
1000
1001static void replace_slot(struct radix_tree_root *root,
1002 struct radix_tree_node *node,
1003 void **slot, void *item,
1004 bool warn_typeswitch)
1005{
1006 void *old = rcu_dereference_raw(*slot);
1007 int count, exceptional;
1008
1009 WARN_ON_ONCE(radix_tree_is_internal_node(item));
1010
1011 count = !!item - !!old;
1012 exceptional = !!radix_tree_exceptional_entry(item) -
1013 !!radix_tree_exceptional_entry(old);
1014
1015 WARN_ON_ONCE(warn_typeswitch && (count || exceptional));
1016
1017 if (node) {
1018 node->count += count;
1019 if (exceptional) {
1020 exceptional *= slot_count(node, slot);
1021 node->exceptional += exceptional;
1022 }
1023 }
1024
1025 rcu_assign_pointer(*slot, item);
1026}
1027
1028static inline void delete_sibling_entries(struct radix_tree_node *node,
1029 void **slot)
1030{
1031#ifdef CONFIG_RADIX_TREE_MULTIORDER
1032 bool exceptional = radix_tree_exceptional_entry(*slot);
1033 void *ptr = node_to_entry(slot);
1034 unsigned offset = get_slot_offset(node, slot);
1035 int i;
1036
1037 for (i = 1; offset + i < RADIX_TREE_MAP_SIZE; i++) {
1038 if (node->slots[offset + i] != ptr)
1039 break;
1040 node->slots[offset + i] = NULL;
1041 node->count--;
1042 if (exceptional)
1043 node->exceptional--;
1044 }
1045#endif
1046}
1047
1048/**
1049 * __radix_tree_replace - replace item in a slot
1050 * @root: radix tree root
1051 * @node: pointer to tree node
1052 * @slot: pointer to slot in @node
1053 * @item: new item to store in the slot.
1054 * @update_node: callback for changing leaf nodes
1055 * @private: private data to pass to @update_node
1056 *
1057 * For use with __radix_tree_lookup(). Caller must hold tree write locked
1058 * across slot lookup and replacement.
1059 */
1060void __radix_tree_replace(struct radix_tree_root *root,
1061 struct radix_tree_node *node,
1062 void **slot, void *item,
1063 radix_tree_update_node_t update_node, void *private)
1064{
1065 if (!item)
1066 delete_sibling_entries(node, slot);
1067 /*
1068 * This function supports replacing exceptional entries and
1069 * deleting entries, but that needs accounting against the
1070 * node unless the slot is root->rnode.
1071 */
1072 replace_slot(root, node, slot, item,
1073 !node && slot != (void **)&root->rnode);
1074
1075 if (!node)
1076 return;
1077
1078 if (update_node)
1079 update_node(node, private);
1080
1081 delete_node(root, node, update_node, private);
1082}
1083
1084/**
1085 * radix_tree_replace_slot - replace item in a slot
1086 * @root: radix tree root
1087 * @slot: pointer to slot
1088 * @item: new item to store in the slot.
1089 *
1090 * For use with radix_tree_lookup_slot(), radix_tree_gang_lookup_slot(),
1091 * radix_tree_gang_lookup_tag_slot(). Caller must hold tree write locked
1092 * across slot lookup and replacement.
1093 *
1094 * NOTE: This cannot be used to switch between non-entries (empty slots),
1095 * regular entries, and exceptional entries, as that requires accounting
1096 * inside the radix tree node. When switching from one type of entry or
1097 * deleting, use __radix_tree_lookup() and __radix_tree_replace() or
1098 * radix_tree_iter_replace().
1099 */
1100void radix_tree_replace_slot(struct radix_tree_root *root,
1101 void **slot, void *item)
1102{
1103 replace_slot(root, NULL, slot, item, true);
1104}
1105
1106/**
1107 * radix_tree_iter_replace - replace item in a slot
1108 * @root: radix tree root
1109 * @slot: pointer to slot
1110 * @item: new item to store in the slot.
1111 *
1112 * For use with radix_tree_split() and radix_tree_for_each_slot().
1113 * Caller must hold tree write locked across split and replacement.
1114 */
1115void radix_tree_iter_replace(struct radix_tree_root *root,
1116 const struct radix_tree_iter *iter, void **slot, void *item)
1117{
1118 __radix_tree_replace(root, iter->node, slot, item, NULL, NULL);
1119}
1120
1121#ifdef CONFIG_RADIX_TREE_MULTIORDER
1122/**
1123 * radix_tree_join - replace multiple entries with one multiorder entry
1124 * @root: radix tree root
1125 * @index: an index inside the new entry
1126 * @order: order of the new entry
1127 * @item: new entry
1128 *
1129 * Call this function to replace several entries with one larger entry.
1130 * The existing entries are presumed to not need freeing as a result of
1131 * this call.
1132 *
1133 * The replacement entry will have all the tags set on it that were set
1134 * on any of the entries it is replacing.
1135 */
1136int radix_tree_join(struct radix_tree_root *root, unsigned long index,
1137 unsigned order, void *item)
1138{
1139 struct radix_tree_node *node;
1140 void **slot;
1141 int error;
1142
1143 BUG_ON(radix_tree_is_internal_node(item));
1144
1145 error = __radix_tree_create(root, index, order, &node, &slot);
1146 if (!error)
1147 error = insert_entries(node, slot, item, order, true);
1148 if (error > 0)
1149 error = 0;
1150
1151 return error;
1152}
1153
1154/**
1155 * radix_tree_split - Split an entry into smaller entries
1156 * @root: radix tree root
1157 * @index: An index within the large entry
1158 * @order: Order of new entries
1159 *
1160 * Call this function as the first step in replacing a multiorder entry
1161 * with several entries of lower order. After this function returns,
1162 * loop over the relevant portion of the tree using radix_tree_for_each_slot()
1163 * and call radix_tree_iter_replace() to set up each new entry.
1164 *
1165 * The tags from this entry are replicated to all the new entries.
1166 *
1167 * The radix tree should be locked against modification during the entire
1168 * replacement operation. Lock-free lookups will see RADIX_TREE_RETRY which
1169 * should prompt RCU walkers to restart the lookup from the root.
1170 */
1171int radix_tree_split(struct radix_tree_root *root, unsigned long index,
1172 unsigned order)
1173{
1174 struct radix_tree_node *parent, *node, *child;
1175 void **slot;
1176 unsigned int offset, end;
1177 unsigned n, tag, tags = 0;
1178
1179 if (!__radix_tree_lookup(root, index, &parent, &slot))
1180 return -ENOENT;
1181 if (!parent)
1182 return -ENOENT;
1183
1184 offset = get_slot_offset(parent, slot);
1185
1186 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++)
1187 if (tag_get(parent, tag, offset))
1188 tags |= 1 << tag;
1189
1190 for (end = offset + 1; end < RADIX_TREE_MAP_SIZE; end++) {
1191 if (!is_sibling_entry(parent, parent->slots[end]))
1192 break;
1193 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++)
1194 if (tags & (1 << tag))
1195 tag_set(parent, tag, end);
1196 /* rcu_assign_pointer ensures tags are set before RETRY */
1197 rcu_assign_pointer(parent->slots[end], RADIX_TREE_RETRY);
1198 }
1199 rcu_assign_pointer(parent->slots[offset], RADIX_TREE_RETRY);
1200 parent->exceptional -= (end - offset);
1201
1202 if (order == parent->shift)
1203 return 0;
1204 if (order > parent->shift) {
1205 while (offset < end)
1206 offset += insert_entries(parent, &parent->slots[offset],
1207 RADIX_TREE_RETRY, order, true);
1208 return 0;
1209 }
1210
1211 node = parent;
1212
1213 for (;;) {
1214 if (node->shift > order) {
1215 child = radix_tree_node_alloc(root, node,
1216 node->shift - RADIX_TREE_MAP_SHIFT,
1217 offset, 0, 0);
1218 if (!child)
1219 goto nomem;
1220 if (node != parent) {
1221 node->count++;
1222 node->slots[offset] = node_to_entry(child);
1223 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++)
1224 if (tags & (1 << tag))
1225 tag_set(node, tag, offset);
1226 }
1227
1228 node = child;
1229 offset = 0;
1230 continue;
1231 }
1232
1233 n = insert_entries(node, &node->slots[offset],
1234 RADIX_TREE_RETRY, order, false);
1235 BUG_ON(n > RADIX_TREE_MAP_SIZE);
1236
1237 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++)
1238 if (tags & (1 << tag))
1239 tag_set(node, tag, offset);
1240 offset += n;
1241
1242 while (offset == RADIX_TREE_MAP_SIZE) {
1243 if (node == parent)
1244 break;
1245 offset = node->offset;
1246 child = node;
1247 node = node->parent;
1248 rcu_assign_pointer(node->slots[offset],
1249 node_to_entry(child));
1250 offset++;
1251 }
1252 if ((node == parent) && (offset == end))
1253 return 0;
1254 }
1255
1256 nomem:
1257 /* Shouldn't happen; did user forget to preload? */
1258 /* TODO: free all the allocated nodes */
1259 WARN_ON(1);
1260 return -ENOMEM;
1261}
1262#endif
1263
1264/**
1265 * radix_tree_tag_set - set a tag on a radix tree node
1266 * @root: radix tree root
1267 * @index: index key
1268 * @tag: tag index
1269 *
1270 * Set the search tag (which must be < RADIX_TREE_MAX_TAGS)
1271 * corresponding to @index in the radix tree. From
1272 * the root all the way down to the leaf node.
1273 *
1274 * Returns the address of the tagged item. Setting a tag on a not-present
1275 * item is a bug.
1276 */
1277void *radix_tree_tag_set(struct radix_tree_root *root,
1278 unsigned long index, unsigned int tag)
1279{
1280 struct radix_tree_node *node, *parent;
1281 unsigned long maxindex;
1282
1283 radix_tree_load_root(root, &node, &maxindex);
1284 BUG_ON(index > maxindex);
1285
1286 while (radix_tree_is_internal_node(node)) {
1287 unsigned offset;
1288
1289 parent = entry_to_node(node);
1290 offset = radix_tree_descend(parent, &node, index);
1291 BUG_ON(!node);
1292
1293 if (!tag_get(parent, tag, offset))
1294 tag_set(parent, tag, offset);
1295 }
1296
1297 /* set the root's tag bit */
1298 if (!root_tag_get(root, tag))
1299 root_tag_set(root, tag);
1300
1301 return node;
1302}
1303EXPORT_SYMBOL(radix_tree_tag_set);
1304
1305static void node_tag_clear(struct radix_tree_root *root,
1306 struct radix_tree_node *node,
1307 unsigned int tag, unsigned int offset)
1308{
1309 while (node) {
1310 if (!tag_get(node, tag, offset))
1311 return;
1312 tag_clear(node, tag, offset);
1313 if (any_tag_set(node, tag))
1314 return;
1315
1316 offset = node->offset;
1317 node = node->parent;
1318 }
1319
1320 /* clear the root's tag bit */
1321 if (root_tag_get(root, tag))
1322 root_tag_clear(root, tag);
1323}
1324
1325static void node_tag_set(struct radix_tree_root *root,
1326 struct radix_tree_node *node,
1327 unsigned int tag, unsigned int offset)
1328{
1329 while (node) {
1330 if (tag_get(node, tag, offset))
1331 return;
1332 tag_set(node, tag, offset);
1333 offset = node->offset;
1334 node = node->parent;
1335 }
1336
1337 if (!root_tag_get(root, tag))
1338 root_tag_set(root, tag);
1339}
1340
1341/**
1342 * radix_tree_iter_tag_set - set a tag on the current iterator entry
1343 * @root: radix tree root
1344 * @iter: iterator state
1345 * @tag: tag to set
1346 */
1347void radix_tree_iter_tag_set(struct radix_tree_root *root,
1348 const struct radix_tree_iter *iter, unsigned int tag)
1349{
1350 node_tag_set(root, iter->node, tag, iter_offset(iter));
1351}
1352
1353/**
1354 * radix_tree_tag_clear - clear a tag on a radix tree node
1355 * @root: radix tree root
1356 * @index: index key
1357 * @tag: tag index
1358 *
1359 * Clear the search tag (which must be < RADIX_TREE_MAX_TAGS)
1360 * corresponding to @index in the radix tree. If this causes
1361 * the leaf node to have no tags set then clear the tag in the
1362 * next-to-leaf node, etc.
1363 *
1364 * Returns the address of the tagged item on success, else NULL. ie:
1365 * has the same return value and semantics as radix_tree_lookup().
1366 */
1367void *radix_tree_tag_clear(struct radix_tree_root *root,
1368 unsigned long index, unsigned int tag)
1369{
1370 struct radix_tree_node *node, *parent;
1371 unsigned long maxindex;
1372 int uninitialized_var(offset);
1373
1374 radix_tree_load_root(root, &node, &maxindex);
1375 if (index > maxindex)
1376 return NULL;
1377
1378 parent = NULL;
1379
1380 while (radix_tree_is_internal_node(node)) {
1381 parent = entry_to_node(node);
1382 offset = radix_tree_descend(parent, &node, index);
1383 }
1384
1385 if (node)
1386 node_tag_clear(root, parent, tag, offset);
1387
1388 return node;
1389}
1390EXPORT_SYMBOL(radix_tree_tag_clear);
1391
1392/**
1393 * radix_tree_tag_get - get a tag on a radix tree node
1394 * @root: radix tree root
1395 * @index: index key
1396 * @tag: tag index (< RADIX_TREE_MAX_TAGS)
1397 *
1398 * Return values:
1399 *
1400 * 0: tag not present or not set
1401 * 1: tag set
1402 *
1403 * Note that the return value of this function may not be relied on, even if
1404 * the RCU lock is held, unless tag modification and node deletion are excluded
1405 * from concurrency.
1406 */
1407int radix_tree_tag_get(struct radix_tree_root *root,
1408 unsigned long index, unsigned int tag)
1409{
1410 struct radix_tree_node *node, *parent;
1411 unsigned long maxindex;
1412
1413 if (!root_tag_get(root, tag))
1414 return 0;
1415
1416 radix_tree_load_root(root, &node, &maxindex);
1417 if (index > maxindex)
1418 return 0;
1419 if (node == NULL)
1420 return 0;
1421
1422 while (radix_tree_is_internal_node(node)) {
1423 unsigned offset;
1424
1425 parent = entry_to_node(node);
1426 offset = radix_tree_descend(parent, &node, index);
1427
1428 if (!node)
1429 return 0;
1430 if (!tag_get(parent, tag, offset))
1431 return 0;
1432 if (node == RADIX_TREE_RETRY)
1433 break;
1434 }
1435
1436 return 1;
1437}
1438EXPORT_SYMBOL(radix_tree_tag_get);
1439
1440static inline void __set_iter_shift(struct radix_tree_iter *iter,
1441 unsigned int shift)
1442{
1443#ifdef CONFIG_RADIX_TREE_MULTIORDER
1444 iter->shift = shift;
1445#endif
1446}
1447
1448/* Construct iter->tags bit-mask from node->tags[tag] array */
1449static void set_iter_tags(struct radix_tree_iter *iter,
1450 struct radix_tree_node *node, unsigned offset,
1451 unsigned tag)
1452{
1453 unsigned tag_long = offset / BITS_PER_LONG;
1454 unsigned tag_bit = offset % BITS_PER_LONG;
1455
1456 iter->tags = node->tags[tag][tag_long] >> tag_bit;
1457
1458 /* This never happens if RADIX_TREE_TAG_LONGS == 1 */
1459 if (tag_long < RADIX_TREE_TAG_LONGS - 1) {
1460 /* Pick tags from next element */
1461 if (tag_bit)
1462 iter->tags |= node->tags[tag][tag_long + 1] <<
1463 (BITS_PER_LONG - tag_bit);
1464 /* Clip chunk size, here only BITS_PER_LONG tags */
1465 iter->next_index = __radix_tree_iter_add(iter, BITS_PER_LONG);
1466 }
1467}
1468
1469#ifdef CONFIG_RADIX_TREE_MULTIORDER
1470static void **skip_siblings(struct radix_tree_node **nodep,
1471 void **slot, struct radix_tree_iter *iter)
1472{
1473 void *sib = node_to_entry(slot - 1);
1474
1475 while (iter->index < iter->next_index) {
1476 *nodep = rcu_dereference_raw(*slot);
1477 if (*nodep && *nodep != sib)
1478 return slot;
1479 slot++;
1480 iter->index = __radix_tree_iter_add(iter, 1);
1481 iter->tags >>= 1;
1482 }
1483
1484 *nodep = NULL;
1485 return NULL;
1486}
1487
1488void ** __radix_tree_next_slot(void **slot, struct radix_tree_iter *iter,
1489 unsigned flags)
1490{
1491 unsigned tag = flags & RADIX_TREE_ITER_TAG_MASK;
1492 struct radix_tree_node *node = rcu_dereference_raw(*slot);
1493
1494 slot = skip_siblings(&node, slot, iter);
1495
1496 while (radix_tree_is_internal_node(node)) {
1497 unsigned offset;
1498 unsigned long next_index;
1499
1500 if (node == RADIX_TREE_RETRY)
1501 return slot;
1502 node = entry_to_node(node);
1503 iter->node = node;
1504 iter->shift = node->shift;
1505
1506 if (flags & RADIX_TREE_ITER_TAGGED) {
1507 offset = radix_tree_find_next_bit(node, tag, 0);
1508 if (offset == RADIX_TREE_MAP_SIZE)
1509 return NULL;
1510 slot = &node->slots[offset];
1511 iter->index = __radix_tree_iter_add(iter, offset);
1512 set_iter_tags(iter, node, offset, tag);
1513 node = rcu_dereference_raw(*slot);
1514 } else {
1515 offset = 0;
1516 slot = &node->slots[0];
1517 for (;;) {
1518 node = rcu_dereference_raw(*slot);
1519 if (node)
1520 break;
1521 slot++;
1522 offset++;
1523 if (offset == RADIX_TREE_MAP_SIZE)
1524 return NULL;
1525 }
1526 iter->index = __radix_tree_iter_add(iter, offset);
1527 }
1528 if ((flags & RADIX_TREE_ITER_CONTIG) && (offset > 0))
1529 goto none;
1530 next_index = (iter->index | shift_maxindex(iter->shift)) + 1;
1531 if (next_index < iter->next_index)
1532 iter->next_index = next_index;
1533 }
1534
1535 return slot;
1536 none:
1537 iter->next_index = 0;
1538 return NULL;
1539}
1540EXPORT_SYMBOL(__radix_tree_next_slot);
1541#else
1542static void **skip_siblings(struct radix_tree_node **nodep,
1543 void **slot, struct radix_tree_iter *iter)
1544{
1545 return slot;
1546}
1547#endif
1548
1549void **radix_tree_iter_resume(void **slot, struct radix_tree_iter *iter)
1550{
1551 struct radix_tree_node *node;
1552
1553 slot++;
1554 iter->index = __radix_tree_iter_add(iter, 1);
1555 node = rcu_dereference_raw(*slot);
1556 skip_siblings(&node, slot, iter);
1557 iter->next_index = iter->index;
1558 iter->tags = 0;
1559 return NULL;
1560}
1561EXPORT_SYMBOL(radix_tree_iter_resume);
1562
1563/**
1564 * radix_tree_next_chunk - find next chunk of slots for iteration
1565 *
1566 * @root: radix tree root
1567 * @iter: iterator state
1568 * @flags: RADIX_TREE_ITER_* flags and tag index
1569 * Returns: pointer to chunk first slot, or NULL if iteration is over
1570 */
1571void **radix_tree_next_chunk(struct radix_tree_root *root,
1572 struct radix_tree_iter *iter, unsigned flags)
1573{
1574 unsigned tag = flags & RADIX_TREE_ITER_TAG_MASK;
1575 struct radix_tree_node *node, *child;
1576 unsigned long index, offset, maxindex;
1577
1578 if ((flags & RADIX_TREE_ITER_TAGGED) && !root_tag_get(root, tag))
1579 return NULL;
1580
1581 /*
1582 * Catch next_index overflow after ~0UL. iter->index never overflows
1583 * during iterating; it can be zero only at the beginning.
1584 * And we cannot overflow iter->next_index in a single step,
1585 * because RADIX_TREE_MAP_SHIFT < BITS_PER_LONG.
1586 *
1587 * This condition also used by radix_tree_next_slot() to stop
1588 * contiguous iterating, and forbid switching to the next chunk.
1589 */
1590 index = iter->next_index;
1591 if (!index && iter->index)
1592 return NULL;
1593
1594 restart:
1595 radix_tree_load_root(root, &child, &maxindex);
1596 if (index > maxindex)
1597 return NULL;
1598 if (!child)
1599 return NULL;
1600
1601 if (!radix_tree_is_internal_node(child)) {
1602 /* Single-slot tree */
1603 iter->index = index;
1604 iter->next_index = maxindex + 1;
1605 iter->tags = 1;
1606 iter->node = NULL;
1607 __set_iter_shift(iter, 0);
1608 return (void **)&root->rnode;
1609 }
1610
1611 do {
1612 node = entry_to_node(child);
1613 offset = radix_tree_descend(node, &child, index);
1614
1615 if ((flags & RADIX_TREE_ITER_TAGGED) ?
1616 !tag_get(node, tag, offset) : !child) {
1617 /* Hole detected */
1618 if (flags & RADIX_TREE_ITER_CONTIG)
1619 return NULL;
1620
1621 if (flags & RADIX_TREE_ITER_TAGGED)
1622 offset = radix_tree_find_next_bit(node, tag,
1623 offset + 1);
1624 else
1625 while (++offset < RADIX_TREE_MAP_SIZE) {
1626 void *slot = node->slots[offset];
1627 if (is_sibling_entry(node, slot))
1628 continue;
1629 if (slot)
1630 break;
1631 }
1632 index &= ~node_maxindex(node);
1633 index += offset << node->shift;
1634 /* Overflow after ~0UL */
1635 if (!index)
1636 return NULL;
1637 if (offset == RADIX_TREE_MAP_SIZE)
1638 goto restart;
1639 child = rcu_dereference_raw(node->slots[offset]);
1640 }
1641
1642 if (!child)
1643 goto restart;
1644 if (child == RADIX_TREE_RETRY)
1645 break;
1646 } while (radix_tree_is_internal_node(child));
1647
1648 /* Update the iterator state */
1649 iter->index = (index &~ node_maxindex(node)) | (offset << node->shift);
1650 iter->next_index = (index | node_maxindex(node)) + 1;
1651 iter->node = node;
1652 __set_iter_shift(iter, node->shift);
1653
1654 if (flags & RADIX_TREE_ITER_TAGGED)
1655 set_iter_tags(iter, node, offset, tag);
1656
1657 return node->slots + offset;
1658}
1659EXPORT_SYMBOL(radix_tree_next_chunk);
1660
1661/**
1662 * radix_tree_gang_lookup - perform multiple lookup on a radix tree
1663 * @root: radix tree root
1664 * @results: where the results of the lookup are placed
1665 * @first_index: start the lookup from this key
1666 * @max_items: place up to this many items at *results
1667 *
1668 * Performs an index-ascending scan of the tree for present items. Places
1669 * them at *@results and returns the number of items which were placed at
1670 * *@results.
1671 *
1672 * The implementation is naive.
1673 *
1674 * Like radix_tree_lookup, radix_tree_gang_lookup may be called under
1675 * rcu_read_lock. In this case, rather than the returned results being
1676 * an atomic snapshot of the tree at a single point in time, the
1677 * semantics of an RCU protected gang lookup are as though multiple
1678 * radix_tree_lookups have been issued in individual locks, and results
1679 * stored in 'results'.
1680 */
1681unsigned int
1682radix_tree_gang_lookup(struct radix_tree_root *root, void **results,
1683 unsigned long first_index, unsigned int max_items)
1684{
1685 struct radix_tree_iter iter;
1686 void **slot;
1687 unsigned int ret = 0;
1688
1689 if (unlikely(!max_items))
1690 return 0;
1691
1692 radix_tree_for_each_slot(slot, root, &iter, first_index) {
1693 results[ret] = rcu_dereference_raw(*slot);
1694 if (!results[ret])
1695 continue;
1696 if (radix_tree_is_internal_node(results[ret])) {
1697 slot = radix_tree_iter_retry(&iter);
1698 continue;
1699 }
1700 if (++ret == max_items)
1701 break;
1702 }
1703
1704 return ret;
1705}
1706EXPORT_SYMBOL(radix_tree_gang_lookup);
1707
1708/**
1709 * radix_tree_gang_lookup_slot - perform multiple slot lookup on radix tree
1710 * @root: radix tree root
1711 * @results: where the results of the lookup are placed
1712 * @indices: where their indices should be placed (but usually NULL)
1713 * @first_index: start the lookup from this key
1714 * @max_items: place up to this many items at *results
1715 *
1716 * Performs an index-ascending scan of the tree for present items. Places
1717 * their slots at *@results and returns the number of items which were
1718 * placed at *@results.
1719 *
1720 * The implementation is naive.
1721 *
1722 * Like radix_tree_gang_lookup as far as RCU and locking goes. Slots must
1723 * be dereferenced with radix_tree_deref_slot, and if using only RCU
1724 * protection, radix_tree_deref_slot may fail requiring a retry.
1725 */
1726unsigned int
1727radix_tree_gang_lookup_slot(struct radix_tree_root *root,
1728 void ***results, unsigned long *indices,
1729 unsigned long first_index, unsigned int max_items)
1730{
1731 struct radix_tree_iter iter;
1732 void **slot;
1733 unsigned int ret = 0;
1734
1735 if (unlikely(!max_items))
1736 return 0;
1737
1738 radix_tree_for_each_slot(slot, root, &iter, first_index) {
1739 results[ret] = slot;
1740 if (indices)
1741 indices[ret] = iter.index;
1742 if (++ret == max_items)
1743 break;
1744 }
1745
1746 return ret;
1747}
1748EXPORT_SYMBOL(radix_tree_gang_lookup_slot);
1749
1750/**
1751 * radix_tree_gang_lookup_tag - perform multiple lookup on a radix tree
1752 * based on a tag
1753 * @root: radix tree root
1754 * @results: where the results of the lookup are placed
1755 * @first_index: start the lookup from this key
1756 * @max_items: place up to this many items at *results
1757 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1758 *
1759 * Performs an index-ascending scan of the tree for present items which
1760 * have the tag indexed by @tag set. Places the items at *@results and
1761 * returns the number of items which were placed at *@results.
1762 */
1763unsigned int
1764radix_tree_gang_lookup_tag(struct radix_tree_root *root, void **results,
1765 unsigned long first_index, unsigned int max_items,
1766 unsigned int tag)
1767{
1768 struct radix_tree_iter iter;
1769 void **slot;
1770 unsigned int ret = 0;
1771
1772 if (unlikely(!max_items))
1773 return 0;
1774
1775 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1776 results[ret] = rcu_dereference_raw(*slot);
1777 if (!results[ret])
1778 continue;
1779 if (radix_tree_is_internal_node(results[ret])) {
1780 slot = radix_tree_iter_retry(&iter);
1781 continue;
1782 }
1783 if (++ret == max_items)
1784 break;
1785 }
1786
1787 return ret;
1788}
1789EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
1790
1791/**
1792 * radix_tree_gang_lookup_tag_slot - perform multiple slot lookup on a
1793 * radix tree based on a tag
1794 * @root: radix tree root
1795 * @results: where the results of the lookup are placed
1796 * @first_index: start the lookup from this key
1797 * @max_items: place up to this many items at *results
1798 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1799 *
1800 * Performs an index-ascending scan of the tree for present items which
1801 * have the tag indexed by @tag set. Places the slots at *@results and
1802 * returns the number of slots which were placed at *@results.
1803 */
1804unsigned int
1805radix_tree_gang_lookup_tag_slot(struct radix_tree_root *root, void ***results,
1806 unsigned long first_index, unsigned int max_items,
1807 unsigned int tag)
1808{
1809 struct radix_tree_iter iter;
1810 void **slot;
1811 unsigned int ret = 0;
1812
1813 if (unlikely(!max_items))
1814 return 0;
1815
1816 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1817 results[ret] = slot;
1818 if (++ret == max_items)
1819 break;
1820 }
1821
1822 return ret;
1823}
1824EXPORT_SYMBOL(radix_tree_gang_lookup_tag_slot);
1825
1826/**
1827 * __radix_tree_delete_node - try to free node after clearing a slot
1828 * @root: radix tree root
1829 * @node: node containing @index
1830 * @update_node: callback for changing leaf nodes
1831 * @private: private data to pass to @update_node
1832 *
1833 * After clearing the slot at @index in @node from radix tree
1834 * rooted at @root, call this function to attempt freeing the
1835 * node and shrinking the tree.
1836 */
1837void __radix_tree_delete_node(struct radix_tree_root *root,
1838 struct radix_tree_node *node,
1839 radix_tree_update_node_t update_node,
1840 void *private)
1841{
1842 delete_node(root, node, update_node, private);
1843}
1844
1845/**
1846 * radix_tree_delete_item - delete an item from a radix tree
1847 * @root: radix tree root
1848 * @index: index key
1849 * @item: expected item
1850 *
1851 * Remove @item at @index from the radix tree rooted at @root.
1852 *
1853 * Returns the address of the deleted item, or NULL if it was not present
1854 * or the entry at the given @index was not @item.
1855 */
1856void *radix_tree_delete_item(struct radix_tree_root *root,
1857 unsigned long index, void *item)
1858{
1859 struct radix_tree_node *node;
1860 unsigned int offset;
1861 void **slot;
1862 void *entry;
1863 int tag;
1864
1865 entry = __radix_tree_lookup(root, index, &node, &slot);
1866 if (!entry)
1867 return NULL;
1868
1869 if (item && entry != item)
1870 return NULL;
1871
1872 if (!node) {
1873 root_tag_clear_all(root);
1874 root->rnode = NULL;
1875 return entry;
1876 }
1877
1878 offset = get_slot_offset(node, slot);
1879
1880 /* Clear all tags associated with the item to be deleted. */
1881 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++)
1882 node_tag_clear(root, node, tag, offset);
1883
1884 __radix_tree_replace(root, node, slot, NULL, NULL, NULL);
1885
1886 return entry;
1887}
1888EXPORT_SYMBOL(radix_tree_delete_item);
1889
1890/**
1891 * radix_tree_delete - delete an item from a radix tree
1892 * @root: radix tree root
1893 * @index: index key
1894 *
1895 * Remove the item at @index from the radix tree rooted at @root.
1896 *
1897 * Returns the address of the deleted item, or NULL if it was not present.
1898 */
1899void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
1900{
1901 return radix_tree_delete_item(root, index, NULL);
1902}
1903EXPORT_SYMBOL(radix_tree_delete);
1904
1905void radix_tree_clear_tags(struct radix_tree_root *root,
1906 struct radix_tree_node *node,
1907 void **slot)
1908{
1909 if (node) {
1910 unsigned int tag, offset = get_slot_offset(node, slot);
1911 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++)
1912 node_tag_clear(root, node, tag, offset);
1913 } else {
1914 /* Clear root node tags */
1915 root->gfp_mask &= __GFP_BITS_MASK;
1916 }
1917}
1918
1919/**
1920 * radix_tree_tagged - test whether any items in the tree are tagged
1921 * @root: radix tree root
1922 * @tag: tag to test
1923 */
1924int radix_tree_tagged(struct radix_tree_root *root, unsigned int tag)
1925{
1926 return root_tag_get(root, tag);
1927}
1928EXPORT_SYMBOL(radix_tree_tagged);
1929
1930static void
1931radix_tree_node_ctor(void *arg)
1932{
1933 struct radix_tree_node *node = arg;
1934
1935 memset(node, 0, sizeof(*node));
1936 INIT_LIST_HEAD(&node->private_list);
1937}
1938
1939static __init unsigned long __maxindex(unsigned int height)
1940{
1941 unsigned int width = height * RADIX_TREE_MAP_SHIFT;
1942 int shift = RADIX_TREE_INDEX_BITS - width;
1943
1944 if (shift < 0)
1945 return ~0UL;
1946 if (shift >= BITS_PER_LONG)
1947 return 0UL;
1948 return ~0UL >> shift;
1949}
1950
1951static __init void radix_tree_init_maxnodes(void)
1952{
1953 unsigned long height_to_maxindex[RADIX_TREE_MAX_PATH + 1];
1954 unsigned int i, j;
1955
1956 for (i = 0; i < ARRAY_SIZE(height_to_maxindex); i++)
1957 height_to_maxindex[i] = __maxindex(i);
1958 for (i = 0; i < ARRAY_SIZE(height_to_maxnodes); i++) {
1959 for (j = i; j > 0; j--)
1960 height_to_maxnodes[i] += height_to_maxindex[j - 1] + 1;
1961 }
1962}
1963
1964static int radix_tree_cpu_dead(unsigned int cpu)
1965{
1966 struct radix_tree_preload *rtp;
1967 struct radix_tree_node *node;
1968
1969 /* Free per-cpu pool of preloaded nodes */
1970 rtp = &per_cpu(radix_tree_preloads, cpu);
1971 while (rtp->nr) {
1972 node = rtp->nodes;
1973 rtp->nodes = node->private_data;
1974 kmem_cache_free(radix_tree_node_cachep, node);
1975 rtp->nr--;
1976 }
1977 return 0;
1978}
1979
1980void __init radix_tree_init(void)
1981{
1982 int ret;
1983 radix_tree_node_cachep = kmem_cache_create("radix_tree_node",
1984 sizeof(struct radix_tree_node), 0,
1985 SLAB_PANIC | SLAB_RECLAIM_ACCOUNT,
1986 radix_tree_node_ctor);
1987 radix_tree_init_maxnodes();
1988 ret = cpuhp_setup_state_nocalls(CPUHP_RADIX_DEAD, "lib/radix:dead",
1989 NULL, radix_tree_cpu_dead);
1990 WARN_ON(ret < 0);
1991}