Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (C) 2001 Momchil Velikov
4 * Portions Copyright (C) 2001 Christoph Hellwig
5 * Copyright (C) 2005 SGI, Christoph Lameter
6 * Copyright (C) 2006 Nick Piggin
7 * Copyright (C) 2012 Konstantin Khlebnikov
8 * Copyright (C) 2016 Intel, Matthew Wilcox
9 * Copyright (C) 2016 Intel, Ross Zwisler
10 */
11
12#include <linux/bitmap.h>
13#include <linux/bitops.h>
14#include <linux/bug.h>
15#include <linux/cpu.h>
16#include <linux/errno.h>
17#include <linux/export.h>
18#include <linux/idr.h>
19#include <linux/init.h>
20#include <linux/kernel.h>
21#include <linux/kmemleak.h>
22#include <linux/percpu.h>
23#include <linux/preempt.h> /* in_interrupt() */
24#include <linux/radix-tree.h>
25#include <linux/rcupdate.h>
26#include <linux/slab.h>
27#include <linux/string.h>
28#include <linux/xarray.h>
29
30/*
31 * Radix tree node cache.
32 */
33struct kmem_cache *radix_tree_node_cachep;
34
35/*
36 * The radix tree is variable-height, so an insert operation not only has
37 * to build the branch to its corresponding item, it also has to build the
38 * branch to existing items if the size has to be increased (by
39 * radix_tree_extend).
40 *
41 * The worst case is a zero height tree with just a single item at index 0,
42 * and then inserting an item at index ULONG_MAX. This requires 2 new branches
43 * of RADIX_TREE_MAX_PATH size to be created, with only the root node shared.
44 * Hence:
45 */
46#define RADIX_TREE_PRELOAD_SIZE (RADIX_TREE_MAX_PATH * 2 - 1)
47
48/*
49 * The IDR does not have to be as high as the radix tree since it uses
50 * signed integers, not unsigned longs.
51 */
52#define IDR_INDEX_BITS (8 /* CHAR_BIT */ * sizeof(int) - 1)
53#define IDR_MAX_PATH (DIV_ROUND_UP(IDR_INDEX_BITS, \
54 RADIX_TREE_MAP_SHIFT))
55#define IDR_PRELOAD_SIZE (IDR_MAX_PATH * 2 - 1)
56
57/*
58 * Per-cpu pool of preloaded nodes
59 */
60DEFINE_PER_CPU(struct radix_tree_preload, radix_tree_preloads) = {
61 .lock = INIT_LOCAL_LOCK(lock),
62};
63EXPORT_PER_CPU_SYMBOL_GPL(radix_tree_preloads);
64
65static inline struct radix_tree_node *entry_to_node(void *ptr)
66{
67 return (void *)((unsigned long)ptr & ~RADIX_TREE_INTERNAL_NODE);
68}
69
70static inline void *node_to_entry(void *ptr)
71{
72 return (void *)((unsigned long)ptr | RADIX_TREE_INTERNAL_NODE);
73}
74
75#define RADIX_TREE_RETRY XA_RETRY_ENTRY
76
77static inline unsigned long
78get_slot_offset(const struct radix_tree_node *parent, void __rcu **slot)
79{
80 return parent ? slot - parent->slots : 0;
81}
82
83static unsigned int radix_tree_descend(const struct radix_tree_node *parent,
84 struct radix_tree_node **nodep, unsigned long index)
85{
86 unsigned int offset = (index >> parent->shift) & RADIX_TREE_MAP_MASK;
87 void __rcu **entry = rcu_dereference_raw(parent->slots[offset]);
88
89 *nodep = (void *)entry;
90 return offset;
91}
92
93static inline gfp_t root_gfp_mask(const struct radix_tree_root *root)
94{
95 return root->xa_flags & (__GFP_BITS_MASK & ~GFP_ZONEMASK);
96}
97
98static inline void tag_set(struct radix_tree_node *node, unsigned int tag,
99 int offset)
100{
101 __set_bit(offset, node->tags[tag]);
102}
103
104static inline void tag_clear(struct radix_tree_node *node, unsigned int tag,
105 int offset)
106{
107 __clear_bit(offset, node->tags[tag]);
108}
109
110static inline int tag_get(const struct radix_tree_node *node, unsigned int tag,
111 int offset)
112{
113 return test_bit(offset, node->tags[tag]);
114}
115
116static inline void root_tag_set(struct radix_tree_root *root, unsigned tag)
117{
118 root->xa_flags |= (__force gfp_t)(1 << (tag + ROOT_TAG_SHIFT));
119}
120
121static inline void root_tag_clear(struct radix_tree_root *root, unsigned tag)
122{
123 root->xa_flags &= (__force gfp_t)~(1 << (tag + ROOT_TAG_SHIFT));
124}
125
126static inline void root_tag_clear_all(struct radix_tree_root *root)
127{
128 root->xa_flags &= (__force gfp_t)((1 << ROOT_TAG_SHIFT) - 1);
129}
130
131static inline int root_tag_get(const struct radix_tree_root *root, unsigned tag)
132{
133 return (__force int)root->xa_flags & (1 << (tag + ROOT_TAG_SHIFT));
134}
135
136static inline unsigned root_tags_get(const struct radix_tree_root *root)
137{
138 return (__force unsigned)root->xa_flags >> ROOT_TAG_SHIFT;
139}
140
141static inline bool is_idr(const struct radix_tree_root *root)
142{
143 return !!(root->xa_flags & ROOT_IS_IDR);
144}
145
146/*
147 * Returns 1 if any slot in the node has this tag set.
148 * Otherwise returns 0.
149 */
150static inline int any_tag_set(const struct radix_tree_node *node,
151 unsigned int tag)
152{
153 unsigned idx;
154 for (idx = 0; idx < RADIX_TREE_TAG_LONGS; idx++) {
155 if (node->tags[tag][idx])
156 return 1;
157 }
158 return 0;
159}
160
161static inline void all_tag_set(struct radix_tree_node *node, unsigned int tag)
162{
163 bitmap_fill(node->tags[tag], RADIX_TREE_MAP_SIZE);
164}
165
166/**
167 * radix_tree_find_next_bit - find the next set bit in a memory region
168 *
169 * @node: where to begin the search
170 * @tag: the tag index
171 * @offset: the bitnumber to start searching at
172 *
173 * Unrollable variant of find_next_bit() for constant size arrays.
174 * Tail bits starting from size to roundup(size, BITS_PER_LONG) must be zero.
175 * Returns next bit offset, or size if nothing found.
176 */
177static __always_inline unsigned long
178radix_tree_find_next_bit(struct radix_tree_node *node, unsigned int tag,
179 unsigned long offset)
180{
181 const unsigned long *addr = node->tags[tag];
182
183 if (offset < RADIX_TREE_MAP_SIZE) {
184 unsigned long tmp;
185
186 addr += offset / BITS_PER_LONG;
187 tmp = *addr >> (offset % BITS_PER_LONG);
188 if (tmp)
189 return __ffs(tmp) + offset;
190 offset = (offset + BITS_PER_LONG) & ~(BITS_PER_LONG - 1);
191 while (offset < RADIX_TREE_MAP_SIZE) {
192 tmp = *++addr;
193 if (tmp)
194 return __ffs(tmp) + offset;
195 offset += BITS_PER_LONG;
196 }
197 }
198 return RADIX_TREE_MAP_SIZE;
199}
200
201static unsigned int iter_offset(const struct radix_tree_iter *iter)
202{
203 return iter->index & RADIX_TREE_MAP_MASK;
204}
205
206/*
207 * The maximum index which can be stored in a radix tree
208 */
209static inline unsigned long shift_maxindex(unsigned int shift)
210{
211 return (RADIX_TREE_MAP_SIZE << shift) - 1;
212}
213
214static inline unsigned long node_maxindex(const struct radix_tree_node *node)
215{
216 return shift_maxindex(node->shift);
217}
218
219static unsigned long next_index(unsigned long index,
220 const struct radix_tree_node *node,
221 unsigned long offset)
222{
223 return (index & ~node_maxindex(node)) + (offset << node->shift);
224}
225
226/*
227 * This assumes that the caller has performed appropriate preallocation, and
228 * that the caller has pinned this thread of control to the current CPU.
229 */
230static struct radix_tree_node *
231radix_tree_node_alloc(gfp_t gfp_mask, struct radix_tree_node *parent,
232 struct radix_tree_root *root,
233 unsigned int shift, unsigned int offset,
234 unsigned int count, unsigned int nr_values)
235{
236 struct radix_tree_node *ret = NULL;
237
238 /*
239 * Preload code isn't irq safe and it doesn't make sense to use
240 * preloading during an interrupt anyway as all the allocations have
241 * to be atomic. So just do normal allocation when in interrupt.
242 */
243 if (!gfpflags_allow_blocking(gfp_mask) && !in_interrupt()) {
244 struct radix_tree_preload *rtp;
245
246 /*
247 * Even if the caller has preloaded, try to allocate from the
248 * cache first for the new node to get accounted to the memory
249 * cgroup.
250 */
251 ret = kmem_cache_alloc(radix_tree_node_cachep,
252 gfp_mask | __GFP_NOWARN);
253 if (ret)
254 goto out;
255
256 /*
257 * Provided the caller has preloaded here, we will always
258 * succeed in getting a node here (and never reach
259 * kmem_cache_alloc)
260 */
261 rtp = this_cpu_ptr(&radix_tree_preloads);
262 if (rtp->nr) {
263 ret = rtp->nodes;
264 rtp->nodes = ret->parent;
265 rtp->nr--;
266 }
267 /*
268 * Update the allocation stack trace as this is more useful
269 * for debugging.
270 */
271 kmemleak_update_trace(ret);
272 goto out;
273 }
274 ret = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
275out:
276 BUG_ON(radix_tree_is_internal_node(ret));
277 if (ret) {
278 ret->shift = shift;
279 ret->offset = offset;
280 ret->count = count;
281 ret->nr_values = nr_values;
282 ret->parent = parent;
283 ret->array = root;
284 }
285 return ret;
286}
287
288void radix_tree_node_rcu_free(struct rcu_head *head)
289{
290 struct radix_tree_node *node =
291 container_of(head, struct radix_tree_node, rcu_head);
292
293 /*
294 * Must only free zeroed nodes into the slab. We can be left with
295 * non-NULL entries by radix_tree_free_nodes, so clear the entries
296 * and tags here.
297 */
298 memset(node->slots, 0, sizeof(node->slots));
299 memset(node->tags, 0, sizeof(node->tags));
300 INIT_LIST_HEAD(&node->private_list);
301
302 kmem_cache_free(radix_tree_node_cachep, node);
303}
304
305static inline void
306radix_tree_node_free(struct radix_tree_node *node)
307{
308 call_rcu(&node->rcu_head, radix_tree_node_rcu_free);
309}
310
311/*
312 * Load up this CPU's radix_tree_node buffer with sufficient objects to
313 * ensure that the addition of a single element in the tree cannot fail. On
314 * success, return zero, with preemption disabled. On error, return -ENOMEM
315 * with preemption not disabled.
316 *
317 * To make use of this facility, the radix tree must be initialised without
318 * __GFP_DIRECT_RECLAIM being passed to INIT_RADIX_TREE().
319 */
320static __must_check int __radix_tree_preload(gfp_t gfp_mask, unsigned nr)
321{
322 struct radix_tree_preload *rtp;
323 struct radix_tree_node *node;
324 int ret = -ENOMEM;
325
326 /*
327 * Nodes preloaded by one cgroup can be used by another cgroup, so
328 * they should never be accounted to any particular memory cgroup.
329 */
330 gfp_mask &= ~__GFP_ACCOUNT;
331
332 local_lock(&radix_tree_preloads.lock);
333 rtp = this_cpu_ptr(&radix_tree_preloads);
334 while (rtp->nr < nr) {
335 local_unlock(&radix_tree_preloads.lock);
336 node = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
337 if (node == NULL)
338 goto out;
339 local_lock(&radix_tree_preloads.lock);
340 rtp = this_cpu_ptr(&radix_tree_preloads);
341 if (rtp->nr < nr) {
342 node->parent = rtp->nodes;
343 rtp->nodes = node;
344 rtp->nr++;
345 } else {
346 kmem_cache_free(radix_tree_node_cachep, node);
347 }
348 }
349 ret = 0;
350out:
351 return ret;
352}
353
354/*
355 * Load up this CPU's radix_tree_node buffer with sufficient objects to
356 * ensure that the addition of a single element in the tree cannot fail. On
357 * success, return zero, with preemption disabled. On error, return -ENOMEM
358 * with preemption not disabled.
359 *
360 * To make use of this facility, the radix tree must be initialised without
361 * __GFP_DIRECT_RECLAIM being passed to INIT_RADIX_TREE().
362 */
363int radix_tree_preload(gfp_t gfp_mask)
364{
365 /* Warn on non-sensical use... */
366 WARN_ON_ONCE(!gfpflags_allow_blocking(gfp_mask));
367 return __radix_tree_preload(gfp_mask, RADIX_TREE_PRELOAD_SIZE);
368}
369EXPORT_SYMBOL(radix_tree_preload);
370
371/*
372 * The same as above function, except we don't guarantee preloading happens.
373 * We do it, if we decide it helps. On success, return zero with preemption
374 * disabled. On error, return -ENOMEM with preemption not disabled.
375 */
376int radix_tree_maybe_preload(gfp_t gfp_mask)
377{
378 if (gfpflags_allow_blocking(gfp_mask))
379 return __radix_tree_preload(gfp_mask, RADIX_TREE_PRELOAD_SIZE);
380 /* Preloading doesn't help anything with this gfp mask, skip it */
381 local_lock(&radix_tree_preloads.lock);
382 return 0;
383}
384EXPORT_SYMBOL(radix_tree_maybe_preload);
385
386static unsigned radix_tree_load_root(const struct radix_tree_root *root,
387 struct radix_tree_node **nodep, unsigned long *maxindex)
388{
389 struct radix_tree_node *node = rcu_dereference_raw(root->xa_head);
390
391 *nodep = node;
392
393 if (likely(radix_tree_is_internal_node(node))) {
394 node = entry_to_node(node);
395 *maxindex = node_maxindex(node);
396 return node->shift + RADIX_TREE_MAP_SHIFT;
397 }
398
399 *maxindex = 0;
400 return 0;
401}
402
403/*
404 * Extend a radix tree so it can store key @index.
405 */
406static int radix_tree_extend(struct radix_tree_root *root, gfp_t gfp,
407 unsigned long index, unsigned int shift)
408{
409 void *entry;
410 unsigned int maxshift;
411 int tag;
412
413 /* Figure out what the shift should be. */
414 maxshift = shift;
415 while (index > shift_maxindex(maxshift))
416 maxshift += RADIX_TREE_MAP_SHIFT;
417
418 entry = rcu_dereference_raw(root->xa_head);
419 if (!entry && (!is_idr(root) || root_tag_get(root, IDR_FREE)))
420 goto out;
421
422 do {
423 struct radix_tree_node *node = radix_tree_node_alloc(gfp, NULL,
424 root, shift, 0, 1, 0);
425 if (!node)
426 return -ENOMEM;
427
428 if (is_idr(root)) {
429 all_tag_set(node, IDR_FREE);
430 if (!root_tag_get(root, IDR_FREE)) {
431 tag_clear(node, IDR_FREE, 0);
432 root_tag_set(root, IDR_FREE);
433 }
434 } else {
435 /* Propagate the aggregated tag info to the new child */
436 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
437 if (root_tag_get(root, tag))
438 tag_set(node, tag, 0);
439 }
440 }
441
442 BUG_ON(shift > BITS_PER_LONG);
443 if (radix_tree_is_internal_node(entry)) {
444 entry_to_node(entry)->parent = node;
445 } else if (xa_is_value(entry)) {
446 /* Moving a value entry root->xa_head to a node */
447 node->nr_values = 1;
448 }
449 /*
450 * entry was already in the radix tree, so we do not need
451 * rcu_assign_pointer here
452 */
453 node->slots[0] = (void __rcu *)entry;
454 entry = node_to_entry(node);
455 rcu_assign_pointer(root->xa_head, entry);
456 shift += RADIX_TREE_MAP_SHIFT;
457 } while (shift <= maxshift);
458out:
459 return maxshift + RADIX_TREE_MAP_SHIFT;
460}
461
462/**
463 * radix_tree_shrink - shrink radix tree to minimum height
464 * @root: radix tree root
465 */
466static inline bool radix_tree_shrink(struct radix_tree_root *root)
467{
468 bool shrunk = false;
469
470 for (;;) {
471 struct radix_tree_node *node = rcu_dereference_raw(root->xa_head);
472 struct radix_tree_node *child;
473
474 if (!radix_tree_is_internal_node(node))
475 break;
476 node = entry_to_node(node);
477
478 /*
479 * The candidate node has more than one child, or its child
480 * is not at the leftmost slot, we cannot shrink.
481 */
482 if (node->count != 1)
483 break;
484 child = rcu_dereference_raw(node->slots[0]);
485 if (!child)
486 break;
487
488 /*
489 * For an IDR, we must not shrink entry 0 into the root in
490 * case somebody calls idr_replace() with a pointer that
491 * appears to be an internal entry
492 */
493 if (!node->shift && is_idr(root))
494 break;
495
496 if (radix_tree_is_internal_node(child))
497 entry_to_node(child)->parent = NULL;
498
499 /*
500 * We don't need rcu_assign_pointer(), since we are simply
501 * moving the node from one part of the tree to another: if it
502 * was safe to dereference the old pointer to it
503 * (node->slots[0]), it will be safe to dereference the new
504 * one (root->xa_head) as far as dependent read barriers go.
505 */
506 root->xa_head = (void __rcu *)child;
507 if (is_idr(root) && !tag_get(node, IDR_FREE, 0))
508 root_tag_clear(root, IDR_FREE);
509
510 /*
511 * We have a dilemma here. The node's slot[0] must not be
512 * NULLed in case there are concurrent lookups expecting to
513 * find the item. However if this was a bottom-level node,
514 * then it may be subject to the slot pointer being visible
515 * to callers dereferencing it. If item corresponding to
516 * slot[0] is subsequently deleted, these callers would expect
517 * their slot to become empty sooner or later.
518 *
519 * For example, lockless pagecache will look up a slot, deref
520 * the page pointer, and if the page has 0 refcount it means it
521 * was concurrently deleted from pagecache so try the deref
522 * again. Fortunately there is already a requirement for logic
523 * to retry the entire slot lookup -- the indirect pointer
524 * problem (replacing direct root node with an indirect pointer
525 * also results in a stale slot). So tag the slot as indirect
526 * to force callers to retry.
527 */
528 node->count = 0;
529 if (!radix_tree_is_internal_node(child)) {
530 node->slots[0] = (void __rcu *)RADIX_TREE_RETRY;
531 }
532
533 WARN_ON_ONCE(!list_empty(&node->private_list));
534 radix_tree_node_free(node);
535 shrunk = true;
536 }
537
538 return shrunk;
539}
540
541static bool delete_node(struct radix_tree_root *root,
542 struct radix_tree_node *node)
543{
544 bool deleted = false;
545
546 do {
547 struct radix_tree_node *parent;
548
549 if (node->count) {
550 if (node_to_entry(node) ==
551 rcu_dereference_raw(root->xa_head))
552 deleted |= radix_tree_shrink(root);
553 return deleted;
554 }
555
556 parent = node->parent;
557 if (parent) {
558 parent->slots[node->offset] = NULL;
559 parent->count--;
560 } else {
561 /*
562 * Shouldn't the tags already have all been cleared
563 * by the caller?
564 */
565 if (!is_idr(root))
566 root_tag_clear_all(root);
567 root->xa_head = NULL;
568 }
569
570 WARN_ON_ONCE(!list_empty(&node->private_list));
571 radix_tree_node_free(node);
572 deleted = true;
573
574 node = parent;
575 } while (node);
576
577 return deleted;
578}
579
580/**
581 * __radix_tree_create - create a slot in a radix tree
582 * @root: radix tree root
583 * @index: index key
584 * @nodep: returns node
585 * @slotp: returns slot
586 *
587 * Create, if necessary, and return the node and slot for an item
588 * at position @index in the radix tree @root.
589 *
590 * Until there is more than one item in the tree, no nodes are
591 * allocated and @root->xa_head is used as a direct slot instead of
592 * pointing to a node, in which case *@nodep will be NULL.
593 *
594 * Returns -ENOMEM, or 0 for success.
595 */
596static int __radix_tree_create(struct radix_tree_root *root,
597 unsigned long index, struct radix_tree_node **nodep,
598 void __rcu ***slotp)
599{
600 struct radix_tree_node *node = NULL, *child;
601 void __rcu **slot = (void __rcu **)&root->xa_head;
602 unsigned long maxindex;
603 unsigned int shift, offset = 0;
604 unsigned long max = index;
605 gfp_t gfp = root_gfp_mask(root);
606
607 shift = radix_tree_load_root(root, &child, &maxindex);
608
609 /* Make sure the tree is high enough. */
610 if (max > maxindex) {
611 int error = radix_tree_extend(root, gfp, max, shift);
612 if (error < 0)
613 return error;
614 shift = error;
615 child = rcu_dereference_raw(root->xa_head);
616 }
617
618 while (shift > 0) {
619 shift -= RADIX_TREE_MAP_SHIFT;
620 if (child == NULL) {
621 /* Have to add a child node. */
622 child = radix_tree_node_alloc(gfp, node, root, shift,
623 offset, 0, 0);
624 if (!child)
625 return -ENOMEM;
626 rcu_assign_pointer(*slot, node_to_entry(child));
627 if (node)
628 node->count++;
629 } else if (!radix_tree_is_internal_node(child))
630 break;
631
632 /* Go a level down */
633 node = entry_to_node(child);
634 offset = radix_tree_descend(node, &child, index);
635 slot = &node->slots[offset];
636 }
637
638 if (nodep)
639 *nodep = node;
640 if (slotp)
641 *slotp = slot;
642 return 0;
643}
644
645/*
646 * Free any nodes below this node. The tree is presumed to not need
647 * shrinking, and any user data in the tree is presumed to not need a
648 * destructor called on it. If we need to add a destructor, we can
649 * add that functionality later. Note that we may not clear tags or
650 * slots from the tree as an RCU walker may still have a pointer into
651 * this subtree. We could replace the entries with RADIX_TREE_RETRY,
652 * but we'll still have to clear those in rcu_free.
653 */
654static void radix_tree_free_nodes(struct radix_tree_node *node)
655{
656 unsigned offset = 0;
657 struct radix_tree_node *child = entry_to_node(node);
658
659 for (;;) {
660 void *entry = rcu_dereference_raw(child->slots[offset]);
661 if (xa_is_node(entry) && child->shift) {
662 child = entry_to_node(entry);
663 offset = 0;
664 continue;
665 }
666 offset++;
667 while (offset == RADIX_TREE_MAP_SIZE) {
668 struct radix_tree_node *old = child;
669 offset = child->offset + 1;
670 child = child->parent;
671 WARN_ON_ONCE(!list_empty(&old->private_list));
672 radix_tree_node_free(old);
673 if (old == entry_to_node(node))
674 return;
675 }
676 }
677}
678
679static inline int insert_entries(struct radix_tree_node *node,
680 void __rcu **slot, void *item, bool replace)
681{
682 if (*slot)
683 return -EEXIST;
684 rcu_assign_pointer(*slot, item);
685 if (node) {
686 node->count++;
687 if (xa_is_value(item))
688 node->nr_values++;
689 }
690 return 1;
691}
692
693/**
694 * radix_tree_insert - insert into a radix tree
695 * @root: radix tree root
696 * @index: index key
697 * @item: item to insert
698 *
699 * Insert an item into the radix tree at position @index.
700 */
701int radix_tree_insert(struct radix_tree_root *root, unsigned long index,
702 void *item)
703{
704 struct radix_tree_node *node;
705 void __rcu **slot;
706 int error;
707
708 BUG_ON(radix_tree_is_internal_node(item));
709
710 error = __radix_tree_create(root, index, &node, &slot);
711 if (error)
712 return error;
713
714 error = insert_entries(node, slot, item, false);
715 if (error < 0)
716 return error;
717
718 if (node) {
719 unsigned offset = get_slot_offset(node, slot);
720 BUG_ON(tag_get(node, 0, offset));
721 BUG_ON(tag_get(node, 1, offset));
722 BUG_ON(tag_get(node, 2, offset));
723 } else {
724 BUG_ON(root_tags_get(root));
725 }
726
727 return 0;
728}
729EXPORT_SYMBOL(radix_tree_insert);
730
731/**
732 * __radix_tree_lookup - lookup an item in a radix tree
733 * @root: radix tree root
734 * @index: index key
735 * @nodep: returns node
736 * @slotp: returns slot
737 *
738 * Lookup and return the item at position @index in the radix
739 * tree @root.
740 *
741 * Until there is more than one item in the tree, no nodes are
742 * allocated and @root->xa_head is used as a direct slot instead of
743 * pointing to a node, in which case *@nodep will be NULL.
744 */
745void *__radix_tree_lookup(const struct radix_tree_root *root,
746 unsigned long index, struct radix_tree_node **nodep,
747 void __rcu ***slotp)
748{
749 struct radix_tree_node *node, *parent;
750 unsigned long maxindex;
751 void __rcu **slot;
752
753 restart:
754 parent = NULL;
755 slot = (void __rcu **)&root->xa_head;
756 radix_tree_load_root(root, &node, &maxindex);
757 if (index > maxindex)
758 return NULL;
759
760 while (radix_tree_is_internal_node(node)) {
761 unsigned offset;
762
763 parent = entry_to_node(node);
764 offset = radix_tree_descend(parent, &node, index);
765 slot = parent->slots + offset;
766 if (node == RADIX_TREE_RETRY)
767 goto restart;
768 if (parent->shift == 0)
769 break;
770 }
771
772 if (nodep)
773 *nodep = parent;
774 if (slotp)
775 *slotp = slot;
776 return node;
777}
778
779/**
780 * radix_tree_lookup_slot - lookup a slot in a radix tree
781 * @root: radix tree root
782 * @index: index key
783 *
784 * Returns: the slot corresponding to the position @index in the
785 * radix tree @root. This is useful for update-if-exists operations.
786 *
787 * This function can be called under rcu_read_lock iff the slot is not
788 * modified by radix_tree_replace_slot, otherwise it must be called
789 * exclusive from other writers. Any dereference of the slot must be done
790 * using radix_tree_deref_slot.
791 */
792void __rcu **radix_tree_lookup_slot(const struct radix_tree_root *root,
793 unsigned long index)
794{
795 void __rcu **slot;
796
797 if (!__radix_tree_lookup(root, index, NULL, &slot))
798 return NULL;
799 return slot;
800}
801EXPORT_SYMBOL(radix_tree_lookup_slot);
802
803/**
804 * radix_tree_lookup - perform lookup operation on a radix tree
805 * @root: radix tree root
806 * @index: index key
807 *
808 * Lookup the item at the position @index in the radix tree @root.
809 *
810 * This function can be called under rcu_read_lock, however the caller
811 * must manage lifetimes of leaf nodes (eg. RCU may also be used to free
812 * them safely). No RCU barriers are required to access or modify the
813 * returned item, however.
814 */
815void *radix_tree_lookup(const struct radix_tree_root *root, unsigned long index)
816{
817 return __radix_tree_lookup(root, index, NULL, NULL);
818}
819EXPORT_SYMBOL(radix_tree_lookup);
820
821static void replace_slot(void __rcu **slot, void *item,
822 struct radix_tree_node *node, int count, int values)
823{
824 if (node && (count || values)) {
825 node->count += count;
826 node->nr_values += values;
827 }
828
829 rcu_assign_pointer(*slot, item);
830}
831
832static bool node_tag_get(const struct radix_tree_root *root,
833 const struct radix_tree_node *node,
834 unsigned int tag, unsigned int offset)
835{
836 if (node)
837 return tag_get(node, tag, offset);
838 return root_tag_get(root, tag);
839}
840
841/*
842 * IDR users want to be able to store NULL in the tree, so if the slot isn't
843 * free, don't adjust the count, even if it's transitioning between NULL and
844 * non-NULL. For the IDA, we mark slots as being IDR_FREE while they still
845 * have empty bits, but it only stores NULL in slots when they're being
846 * deleted.
847 */
848static int calculate_count(struct radix_tree_root *root,
849 struct radix_tree_node *node, void __rcu **slot,
850 void *item, void *old)
851{
852 if (is_idr(root)) {
853 unsigned offset = get_slot_offset(node, slot);
854 bool free = node_tag_get(root, node, IDR_FREE, offset);
855 if (!free)
856 return 0;
857 if (!old)
858 return 1;
859 }
860 return !!item - !!old;
861}
862
863/**
864 * __radix_tree_replace - replace item in a slot
865 * @root: radix tree root
866 * @node: pointer to tree node
867 * @slot: pointer to slot in @node
868 * @item: new item to store in the slot.
869 *
870 * For use with __radix_tree_lookup(). Caller must hold tree write locked
871 * across slot lookup and replacement.
872 */
873void __radix_tree_replace(struct radix_tree_root *root,
874 struct radix_tree_node *node,
875 void __rcu **slot, void *item)
876{
877 void *old = rcu_dereference_raw(*slot);
878 int values = !!xa_is_value(item) - !!xa_is_value(old);
879 int count = calculate_count(root, node, slot, item, old);
880
881 /*
882 * This function supports replacing value entries and
883 * deleting entries, but that needs accounting against the
884 * node unless the slot is root->xa_head.
885 */
886 WARN_ON_ONCE(!node && (slot != (void __rcu **)&root->xa_head) &&
887 (count || values));
888 replace_slot(slot, item, node, count, values);
889
890 if (!node)
891 return;
892
893 delete_node(root, node);
894}
895
896/**
897 * radix_tree_replace_slot - replace item in a slot
898 * @root: radix tree root
899 * @slot: pointer to slot
900 * @item: new item to store in the slot.
901 *
902 * For use with radix_tree_lookup_slot() and
903 * radix_tree_gang_lookup_tag_slot(). Caller must hold tree write locked
904 * across slot lookup and replacement.
905 *
906 * NOTE: This cannot be used to switch between non-entries (empty slots),
907 * regular entries, and value entries, as that requires accounting
908 * inside the radix tree node. When switching from one type of entry or
909 * deleting, use __radix_tree_lookup() and __radix_tree_replace() or
910 * radix_tree_iter_replace().
911 */
912void radix_tree_replace_slot(struct radix_tree_root *root,
913 void __rcu **slot, void *item)
914{
915 __radix_tree_replace(root, NULL, slot, item);
916}
917EXPORT_SYMBOL(radix_tree_replace_slot);
918
919/**
920 * radix_tree_iter_replace - replace item in a slot
921 * @root: radix tree root
922 * @iter: iterator state
923 * @slot: pointer to slot
924 * @item: new item to store in the slot.
925 *
926 * For use with radix_tree_for_each_slot().
927 * Caller must hold tree write locked.
928 */
929void radix_tree_iter_replace(struct radix_tree_root *root,
930 const struct radix_tree_iter *iter,
931 void __rcu **slot, void *item)
932{
933 __radix_tree_replace(root, iter->node, slot, item);
934}
935
936static void node_tag_set(struct radix_tree_root *root,
937 struct radix_tree_node *node,
938 unsigned int tag, unsigned int offset)
939{
940 while (node) {
941 if (tag_get(node, tag, offset))
942 return;
943 tag_set(node, tag, offset);
944 offset = node->offset;
945 node = node->parent;
946 }
947
948 if (!root_tag_get(root, tag))
949 root_tag_set(root, tag);
950}
951
952/**
953 * radix_tree_tag_set - set a tag on a radix tree node
954 * @root: radix tree root
955 * @index: index key
956 * @tag: tag index
957 *
958 * Set the search tag (which must be < RADIX_TREE_MAX_TAGS)
959 * corresponding to @index in the radix tree. From
960 * the root all the way down to the leaf node.
961 *
962 * Returns the address of the tagged item. Setting a tag on a not-present
963 * item is a bug.
964 */
965void *radix_tree_tag_set(struct radix_tree_root *root,
966 unsigned long index, unsigned int tag)
967{
968 struct radix_tree_node *node, *parent;
969 unsigned long maxindex;
970
971 radix_tree_load_root(root, &node, &maxindex);
972 BUG_ON(index > maxindex);
973
974 while (radix_tree_is_internal_node(node)) {
975 unsigned offset;
976
977 parent = entry_to_node(node);
978 offset = radix_tree_descend(parent, &node, index);
979 BUG_ON(!node);
980
981 if (!tag_get(parent, tag, offset))
982 tag_set(parent, tag, offset);
983 }
984
985 /* set the root's tag bit */
986 if (!root_tag_get(root, tag))
987 root_tag_set(root, tag);
988
989 return node;
990}
991EXPORT_SYMBOL(radix_tree_tag_set);
992
993static void node_tag_clear(struct radix_tree_root *root,
994 struct radix_tree_node *node,
995 unsigned int tag, unsigned int offset)
996{
997 while (node) {
998 if (!tag_get(node, tag, offset))
999 return;
1000 tag_clear(node, tag, offset);
1001 if (any_tag_set(node, tag))
1002 return;
1003
1004 offset = node->offset;
1005 node = node->parent;
1006 }
1007
1008 /* clear the root's tag bit */
1009 if (root_tag_get(root, tag))
1010 root_tag_clear(root, tag);
1011}
1012
1013/**
1014 * radix_tree_tag_clear - clear a tag on a radix tree node
1015 * @root: radix tree root
1016 * @index: index key
1017 * @tag: tag index
1018 *
1019 * Clear the search tag (which must be < RADIX_TREE_MAX_TAGS)
1020 * corresponding to @index in the radix tree. If this causes
1021 * the leaf node to have no tags set then clear the tag in the
1022 * next-to-leaf node, etc.
1023 *
1024 * Returns the address of the tagged item on success, else NULL. ie:
1025 * has the same return value and semantics as radix_tree_lookup().
1026 */
1027void *radix_tree_tag_clear(struct radix_tree_root *root,
1028 unsigned long index, unsigned int tag)
1029{
1030 struct radix_tree_node *node, *parent;
1031 unsigned long maxindex;
1032 int offset;
1033
1034 radix_tree_load_root(root, &node, &maxindex);
1035 if (index > maxindex)
1036 return NULL;
1037
1038 parent = NULL;
1039
1040 while (radix_tree_is_internal_node(node)) {
1041 parent = entry_to_node(node);
1042 offset = radix_tree_descend(parent, &node, index);
1043 }
1044
1045 if (node)
1046 node_tag_clear(root, parent, tag, offset);
1047
1048 return node;
1049}
1050EXPORT_SYMBOL(radix_tree_tag_clear);
1051
1052/**
1053 * radix_tree_iter_tag_clear - clear a tag on the current iterator entry
1054 * @root: radix tree root
1055 * @iter: iterator state
1056 * @tag: tag to clear
1057 */
1058void radix_tree_iter_tag_clear(struct radix_tree_root *root,
1059 const struct radix_tree_iter *iter, unsigned int tag)
1060{
1061 node_tag_clear(root, iter->node, tag, iter_offset(iter));
1062}
1063
1064/**
1065 * radix_tree_tag_get - get a tag on a radix tree node
1066 * @root: radix tree root
1067 * @index: index key
1068 * @tag: tag index (< RADIX_TREE_MAX_TAGS)
1069 *
1070 * Return values:
1071 *
1072 * 0: tag not present or not set
1073 * 1: tag set
1074 *
1075 * Note that the return value of this function may not be relied on, even if
1076 * the RCU lock is held, unless tag modification and node deletion are excluded
1077 * from concurrency.
1078 */
1079int radix_tree_tag_get(const struct radix_tree_root *root,
1080 unsigned long index, unsigned int tag)
1081{
1082 struct radix_tree_node *node, *parent;
1083 unsigned long maxindex;
1084
1085 if (!root_tag_get(root, tag))
1086 return 0;
1087
1088 radix_tree_load_root(root, &node, &maxindex);
1089 if (index > maxindex)
1090 return 0;
1091
1092 while (radix_tree_is_internal_node(node)) {
1093 unsigned offset;
1094
1095 parent = entry_to_node(node);
1096 offset = radix_tree_descend(parent, &node, index);
1097
1098 if (!tag_get(parent, tag, offset))
1099 return 0;
1100 if (node == RADIX_TREE_RETRY)
1101 break;
1102 }
1103
1104 return 1;
1105}
1106EXPORT_SYMBOL(radix_tree_tag_get);
1107
1108/* Construct iter->tags bit-mask from node->tags[tag] array */
1109static void set_iter_tags(struct radix_tree_iter *iter,
1110 struct radix_tree_node *node, unsigned offset,
1111 unsigned tag)
1112{
1113 unsigned tag_long = offset / BITS_PER_LONG;
1114 unsigned tag_bit = offset % BITS_PER_LONG;
1115
1116 if (!node) {
1117 iter->tags = 1;
1118 return;
1119 }
1120
1121 iter->tags = node->tags[tag][tag_long] >> tag_bit;
1122
1123 /* This never happens if RADIX_TREE_TAG_LONGS == 1 */
1124 if (tag_long < RADIX_TREE_TAG_LONGS - 1) {
1125 /* Pick tags from next element */
1126 if (tag_bit)
1127 iter->tags |= node->tags[tag][tag_long + 1] <<
1128 (BITS_PER_LONG - tag_bit);
1129 /* Clip chunk size, here only BITS_PER_LONG tags */
1130 iter->next_index = __radix_tree_iter_add(iter, BITS_PER_LONG);
1131 }
1132}
1133
1134void __rcu **radix_tree_iter_resume(void __rcu **slot,
1135 struct radix_tree_iter *iter)
1136{
1137 slot++;
1138 iter->index = __radix_tree_iter_add(iter, 1);
1139 iter->next_index = iter->index;
1140 iter->tags = 0;
1141 return NULL;
1142}
1143EXPORT_SYMBOL(radix_tree_iter_resume);
1144
1145/**
1146 * radix_tree_next_chunk - find next chunk of slots for iteration
1147 *
1148 * @root: radix tree root
1149 * @iter: iterator state
1150 * @flags: RADIX_TREE_ITER_* flags and tag index
1151 * Returns: pointer to chunk first slot, or NULL if iteration is over
1152 */
1153void __rcu **radix_tree_next_chunk(const struct radix_tree_root *root,
1154 struct radix_tree_iter *iter, unsigned flags)
1155{
1156 unsigned tag = flags & RADIX_TREE_ITER_TAG_MASK;
1157 struct radix_tree_node *node, *child;
1158 unsigned long index, offset, maxindex;
1159
1160 if ((flags & RADIX_TREE_ITER_TAGGED) && !root_tag_get(root, tag))
1161 return NULL;
1162
1163 /*
1164 * Catch next_index overflow after ~0UL. iter->index never overflows
1165 * during iterating; it can be zero only at the beginning.
1166 * And we cannot overflow iter->next_index in a single step,
1167 * because RADIX_TREE_MAP_SHIFT < BITS_PER_LONG.
1168 *
1169 * This condition also used by radix_tree_next_slot() to stop
1170 * contiguous iterating, and forbid switching to the next chunk.
1171 */
1172 index = iter->next_index;
1173 if (!index && iter->index)
1174 return NULL;
1175
1176 restart:
1177 radix_tree_load_root(root, &child, &maxindex);
1178 if (index > maxindex)
1179 return NULL;
1180 if (!child)
1181 return NULL;
1182
1183 if (!radix_tree_is_internal_node(child)) {
1184 /* Single-slot tree */
1185 iter->index = index;
1186 iter->next_index = maxindex + 1;
1187 iter->tags = 1;
1188 iter->node = NULL;
1189 return (void __rcu **)&root->xa_head;
1190 }
1191
1192 do {
1193 node = entry_to_node(child);
1194 offset = radix_tree_descend(node, &child, index);
1195
1196 if ((flags & RADIX_TREE_ITER_TAGGED) ?
1197 !tag_get(node, tag, offset) : !child) {
1198 /* Hole detected */
1199 if (flags & RADIX_TREE_ITER_CONTIG)
1200 return NULL;
1201
1202 if (flags & RADIX_TREE_ITER_TAGGED)
1203 offset = radix_tree_find_next_bit(node, tag,
1204 offset + 1);
1205 else
1206 while (++offset < RADIX_TREE_MAP_SIZE) {
1207 void *slot = rcu_dereference_raw(
1208 node->slots[offset]);
1209 if (slot)
1210 break;
1211 }
1212 index &= ~node_maxindex(node);
1213 index += offset << node->shift;
1214 /* Overflow after ~0UL */
1215 if (!index)
1216 return NULL;
1217 if (offset == RADIX_TREE_MAP_SIZE)
1218 goto restart;
1219 child = rcu_dereference_raw(node->slots[offset]);
1220 }
1221
1222 if (!child)
1223 goto restart;
1224 if (child == RADIX_TREE_RETRY)
1225 break;
1226 } while (node->shift && radix_tree_is_internal_node(child));
1227
1228 /* Update the iterator state */
1229 iter->index = (index &~ node_maxindex(node)) | offset;
1230 iter->next_index = (index | node_maxindex(node)) + 1;
1231 iter->node = node;
1232
1233 if (flags & RADIX_TREE_ITER_TAGGED)
1234 set_iter_tags(iter, node, offset, tag);
1235
1236 return node->slots + offset;
1237}
1238EXPORT_SYMBOL(radix_tree_next_chunk);
1239
1240/**
1241 * radix_tree_gang_lookup - perform multiple lookup on a radix tree
1242 * @root: radix tree root
1243 * @results: where the results of the lookup are placed
1244 * @first_index: start the lookup from this key
1245 * @max_items: place up to this many items at *results
1246 *
1247 * Performs an index-ascending scan of the tree for present items. Places
1248 * them at *@results and returns the number of items which were placed at
1249 * *@results.
1250 *
1251 * The implementation is naive.
1252 *
1253 * Like radix_tree_lookup, radix_tree_gang_lookup may be called under
1254 * rcu_read_lock. In this case, rather than the returned results being
1255 * an atomic snapshot of the tree at a single point in time, the
1256 * semantics of an RCU protected gang lookup are as though multiple
1257 * radix_tree_lookups have been issued in individual locks, and results
1258 * stored in 'results'.
1259 */
1260unsigned int
1261radix_tree_gang_lookup(const struct radix_tree_root *root, void **results,
1262 unsigned long first_index, unsigned int max_items)
1263{
1264 struct radix_tree_iter iter;
1265 void __rcu **slot;
1266 unsigned int ret = 0;
1267
1268 if (unlikely(!max_items))
1269 return 0;
1270
1271 radix_tree_for_each_slot(slot, root, &iter, first_index) {
1272 results[ret] = rcu_dereference_raw(*slot);
1273 if (!results[ret])
1274 continue;
1275 if (radix_tree_is_internal_node(results[ret])) {
1276 slot = radix_tree_iter_retry(&iter);
1277 continue;
1278 }
1279 if (++ret == max_items)
1280 break;
1281 }
1282
1283 return ret;
1284}
1285EXPORT_SYMBOL(radix_tree_gang_lookup);
1286
1287/**
1288 * radix_tree_gang_lookup_tag - perform multiple lookup on a radix tree
1289 * based on a tag
1290 * @root: radix tree root
1291 * @results: where the results of the lookup are placed
1292 * @first_index: start the lookup from this key
1293 * @max_items: place up to this many items at *results
1294 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1295 *
1296 * Performs an index-ascending scan of the tree for present items which
1297 * have the tag indexed by @tag set. Places the items at *@results and
1298 * returns the number of items which were placed at *@results.
1299 */
1300unsigned int
1301radix_tree_gang_lookup_tag(const struct radix_tree_root *root, void **results,
1302 unsigned long first_index, unsigned int max_items,
1303 unsigned int tag)
1304{
1305 struct radix_tree_iter iter;
1306 void __rcu **slot;
1307 unsigned int ret = 0;
1308
1309 if (unlikely(!max_items))
1310 return 0;
1311
1312 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1313 results[ret] = rcu_dereference_raw(*slot);
1314 if (!results[ret])
1315 continue;
1316 if (radix_tree_is_internal_node(results[ret])) {
1317 slot = radix_tree_iter_retry(&iter);
1318 continue;
1319 }
1320 if (++ret == max_items)
1321 break;
1322 }
1323
1324 return ret;
1325}
1326EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
1327
1328/**
1329 * radix_tree_gang_lookup_tag_slot - perform multiple slot lookup on a
1330 * radix tree based on a tag
1331 * @root: radix tree root
1332 * @results: where the results of the lookup are placed
1333 * @first_index: start the lookup from this key
1334 * @max_items: place up to this many items at *results
1335 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1336 *
1337 * Performs an index-ascending scan of the tree for present items which
1338 * have the tag indexed by @tag set. Places the slots at *@results and
1339 * returns the number of slots which were placed at *@results.
1340 */
1341unsigned int
1342radix_tree_gang_lookup_tag_slot(const struct radix_tree_root *root,
1343 void __rcu ***results, unsigned long first_index,
1344 unsigned int max_items, unsigned int tag)
1345{
1346 struct radix_tree_iter iter;
1347 void __rcu **slot;
1348 unsigned int ret = 0;
1349
1350 if (unlikely(!max_items))
1351 return 0;
1352
1353 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1354 results[ret] = slot;
1355 if (++ret == max_items)
1356 break;
1357 }
1358
1359 return ret;
1360}
1361EXPORT_SYMBOL(radix_tree_gang_lookup_tag_slot);
1362
1363static bool __radix_tree_delete(struct radix_tree_root *root,
1364 struct radix_tree_node *node, void __rcu **slot)
1365{
1366 void *old = rcu_dereference_raw(*slot);
1367 int values = xa_is_value(old) ? -1 : 0;
1368 unsigned offset = get_slot_offset(node, slot);
1369 int tag;
1370
1371 if (is_idr(root))
1372 node_tag_set(root, node, IDR_FREE, offset);
1373 else
1374 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++)
1375 node_tag_clear(root, node, tag, offset);
1376
1377 replace_slot(slot, NULL, node, -1, values);
1378 return node && delete_node(root, node);
1379}
1380
1381/**
1382 * radix_tree_iter_delete - delete the entry at this iterator position
1383 * @root: radix tree root
1384 * @iter: iterator state
1385 * @slot: pointer to slot
1386 *
1387 * Delete the entry at the position currently pointed to by the iterator.
1388 * This may result in the current node being freed; if it is, the iterator
1389 * is advanced so that it will not reference the freed memory. This
1390 * function may be called without any locking if there are no other threads
1391 * which can access this tree.
1392 */
1393void radix_tree_iter_delete(struct radix_tree_root *root,
1394 struct radix_tree_iter *iter, void __rcu **slot)
1395{
1396 if (__radix_tree_delete(root, iter->node, slot))
1397 iter->index = iter->next_index;
1398}
1399EXPORT_SYMBOL(radix_tree_iter_delete);
1400
1401/**
1402 * radix_tree_delete_item - delete an item from a radix tree
1403 * @root: radix tree root
1404 * @index: index key
1405 * @item: expected item
1406 *
1407 * Remove @item at @index from the radix tree rooted at @root.
1408 *
1409 * Return: the deleted entry, or %NULL if it was not present
1410 * or the entry at the given @index was not @item.
1411 */
1412void *radix_tree_delete_item(struct radix_tree_root *root,
1413 unsigned long index, void *item)
1414{
1415 struct radix_tree_node *node = NULL;
1416 void __rcu **slot = NULL;
1417 void *entry;
1418
1419 entry = __radix_tree_lookup(root, index, &node, &slot);
1420 if (!slot)
1421 return NULL;
1422 if (!entry && (!is_idr(root) || node_tag_get(root, node, IDR_FREE,
1423 get_slot_offset(node, slot))))
1424 return NULL;
1425
1426 if (item && entry != item)
1427 return NULL;
1428
1429 __radix_tree_delete(root, node, slot);
1430
1431 return entry;
1432}
1433EXPORT_SYMBOL(radix_tree_delete_item);
1434
1435/**
1436 * radix_tree_delete - delete an entry from a radix tree
1437 * @root: radix tree root
1438 * @index: index key
1439 *
1440 * Remove the entry at @index from the radix tree rooted at @root.
1441 *
1442 * Return: The deleted entry, or %NULL if it was not present.
1443 */
1444void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
1445{
1446 return radix_tree_delete_item(root, index, NULL);
1447}
1448EXPORT_SYMBOL(radix_tree_delete);
1449
1450/**
1451 * radix_tree_tagged - test whether any items in the tree are tagged
1452 * @root: radix tree root
1453 * @tag: tag to test
1454 */
1455int radix_tree_tagged(const struct radix_tree_root *root, unsigned int tag)
1456{
1457 return root_tag_get(root, tag);
1458}
1459EXPORT_SYMBOL(radix_tree_tagged);
1460
1461/**
1462 * idr_preload - preload for idr_alloc()
1463 * @gfp_mask: allocation mask to use for preloading
1464 *
1465 * Preallocate memory to use for the next call to idr_alloc(). This function
1466 * returns with preemption disabled. It will be enabled by idr_preload_end().
1467 */
1468void idr_preload(gfp_t gfp_mask)
1469{
1470 if (__radix_tree_preload(gfp_mask, IDR_PRELOAD_SIZE))
1471 local_lock(&radix_tree_preloads.lock);
1472}
1473EXPORT_SYMBOL(idr_preload);
1474
1475void __rcu **idr_get_free(struct radix_tree_root *root,
1476 struct radix_tree_iter *iter, gfp_t gfp,
1477 unsigned long max)
1478{
1479 struct radix_tree_node *node = NULL, *child;
1480 void __rcu **slot = (void __rcu **)&root->xa_head;
1481 unsigned long maxindex, start = iter->next_index;
1482 unsigned int shift, offset = 0;
1483
1484 grow:
1485 shift = radix_tree_load_root(root, &child, &maxindex);
1486 if (!radix_tree_tagged(root, IDR_FREE))
1487 start = max(start, maxindex + 1);
1488 if (start > max)
1489 return ERR_PTR(-ENOSPC);
1490
1491 if (start > maxindex) {
1492 int error = radix_tree_extend(root, gfp, start, shift);
1493 if (error < 0)
1494 return ERR_PTR(error);
1495 shift = error;
1496 child = rcu_dereference_raw(root->xa_head);
1497 }
1498 if (start == 0 && shift == 0)
1499 shift = RADIX_TREE_MAP_SHIFT;
1500
1501 while (shift) {
1502 shift -= RADIX_TREE_MAP_SHIFT;
1503 if (child == NULL) {
1504 /* Have to add a child node. */
1505 child = radix_tree_node_alloc(gfp, node, root, shift,
1506 offset, 0, 0);
1507 if (!child)
1508 return ERR_PTR(-ENOMEM);
1509 all_tag_set(child, IDR_FREE);
1510 rcu_assign_pointer(*slot, node_to_entry(child));
1511 if (node)
1512 node->count++;
1513 } else if (!radix_tree_is_internal_node(child))
1514 break;
1515
1516 node = entry_to_node(child);
1517 offset = radix_tree_descend(node, &child, start);
1518 if (!tag_get(node, IDR_FREE, offset)) {
1519 offset = radix_tree_find_next_bit(node, IDR_FREE,
1520 offset + 1);
1521 start = next_index(start, node, offset);
1522 if (start > max || start == 0)
1523 return ERR_PTR(-ENOSPC);
1524 while (offset == RADIX_TREE_MAP_SIZE) {
1525 offset = node->offset + 1;
1526 node = node->parent;
1527 if (!node)
1528 goto grow;
1529 shift = node->shift;
1530 }
1531 child = rcu_dereference_raw(node->slots[offset]);
1532 }
1533 slot = &node->slots[offset];
1534 }
1535
1536 iter->index = start;
1537 if (node)
1538 iter->next_index = 1 + min(max, (start | node_maxindex(node)));
1539 else
1540 iter->next_index = 1;
1541 iter->node = node;
1542 set_iter_tags(iter, node, offset, IDR_FREE);
1543
1544 return slot;
1545}
1546
1547/**
1548 * idr_destroy - release all internal memory from an IDR
1549 * @idr: idr handle
1550 *
1551 * After this function is called, the IDR is empty, and may be reused or
1552 * the data structure containing it may be freed.
1553 *
1554 * A typical clean-up sequence for objects stored in an idr tree will use
1555 * idr_for_each() to free all objects, if necessary, then idr_destroy() to
1556 * free the memory used to keep track of those objects.
1557 */
1558void idr_destroy(struct idr *idr)
1559{
1560 struct radix_tree_node *node = rcu_dereference_raw(idr->idr_rt.xa_head);
1561 if (radix_tree_is_internal_node(node))
1562 radix_tree_free_nodes(node);
1563 idr->idr_rt.xa_head = NULL;
1564 root_tag_set(&idr->idr_rt, IDR_FREE);
1565}
1566EXPORT_SYMBOL(idr_destroy);
1567
1568static void
1569radix_tree_node_ctor(void *arg)
1570{
1571 struct radix_tree_node *node = arg;
1572
1573 memset(node, 0, sizeof(*node));
1574 INIT_LIST_HEAD(&node->private_list);
1575}
1576
1577static int radix_tree_cpu_dead(unsigned int cpu)
1578{
1579 struct radix_tree_preload *rtp;
1580 struct radix_tree_node *node;
1581
1582 /* Free per-cpu pool of preloaded nodes */
1583 rtp = &per_cpu(radix_tree_preloads, cpu);
1584 while (rtp->nr) {
1585 node = rtp->nodes;
1586 rtp->nodes = node->parent;
1587 kmem_cache_free(radix_tree_node_cachep, node);
1588 rtp->nr--;
1589 }
1590 return 0;
1591}
1592
1593void __init radix_tree_init(void)
1594{
1595 int ret;
1596
1597 BUILD_BUG_ON(RADIX_TREE_MAX_TAGS + __GFP_BITS_SHIFT > 32);
1598 BUILD_BUG_ON(ROOT_IS_IDR & ~GFP_ZONEMASK);
1599 BUILD_BUG_ON(XA_CHUNK_SIZE > 255);
1600 radix_tree_node_cachep = kmem_cache_create("radix_tree_node",
1601 sizeof(struct radix_tree_node), 0,
1602 SLAB_PANIC | SLAB_RECLAIM_ACCOUNT,
1603 radix_tree_node_ctor);
1604 ret = cpuhp_setup_state_nocalls(CPUHP_RADIX_DEAD, "lib/radix:dead",
1605 NULL, radix_tree_cpu_dead);
1606 WARN_ON(ret < 0);
1607}
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (C) 2001 Momchil Velikov
4 * Portions Copyright (C) 2001 Christoph Hellwig
5 * Copyright (C) 2005 SGI, Christoph Lameter
6 * Copyright (C) 2006 Nick Piggin
7 * Copyright (C) 2012 Konstantin Khlebnikov
8 * Copyright (C) 2016 Intel, Matthew Wilcox
9 * Copyright (C) 2016 Intel, Ross Zwisler
10 */
11
12#include <linux/bitmap.h>
13#include <linux/bitops.h>
14#include <linux/bug.h>
15#include <linux/cpu.h>
16#include <linux/errno.h>
17#include <linux/export.h>
18#include <linux/idr.h>
19#include <linux/init.h>
20#include <linux/kernel.h>
21#include <linux/kmemleak.h>
22#include <linux/percpu.h>
23#include <linux/preempt.h> /* in_interrupt() */
24#include <linux/radix-tree.h>
25#include <linux/rcupdate.h>
26#include <linux/slab.h>
27#include <linux/string.h>
28#include <linux/xarray.h>
29
30#include "radix-tree.h"
31
32/*
33 * Radix tree node cache.
34 */
35struct kmem_cache *radix_tree_node_cachep;
36
37/*
38 * The radix tree is variable-height, so an insert operation not only has
39 * to build the branch to its corresponding item, it also has to build the
40 * branch to existing items if the size has to be increased (by
41 * radix_tree_extend).
42 *
43 * The worst case is a zero height tree with just a single item at index 0,
44 * and then inserting an item at index ULONG_MAX. This requires 2 new branches
45 * of RADIX_TREE_MAX_PATH size to be created, with only the root node shared.
46 * Hence:
47 */
48#define RADIX_TREE_PRELOAD_SIZE (RADIX_TREE_MAX_PATH * 2 - 1)
49
50/*
51 * The IDR does not have to be as high as the radix tree since it uses
52 * signed integers, not unsigned longs.
53 */
54#define IDR_INDEX_BITS (8 /* CHAR_BIT */ * sizeof(int) - 1)
55#define IDR_MAX_PATH (DIV_ROUND_UP(IDR_INDEX_BITS, \
56 RADIX_TREE_MAP_SHIFT))
57#define IDR_PRELOAD_SIZE (IDR_MAX_PATH * 2 - 1)
58
59/*
60 * Per-cpu pool of preloaded nodes
61 */
62DEFINE_PER_CPU(struct radix_tree_preload, radix_tree_preloads) = {
63 .lock = INIT_LOCAL_LOCK(lock),
64};
65EXPORT_PER_CPU_SYMBOL_GPL(radix_tree_preloads);
66
67static inline struct radix_tree_node *entry_to_node(void *ptr)
68{
69 return (void *)((unsigned long)ptr & ~RADIX_TREE_INTERNAL_NODE);
70}
71
72static inline void *node_to_entry(void *ptr)
73{
74 return (void *)((unsigned long)ptr | RADIX_TREE_INTERNAL_NODE);
75}
76
77#define RADIX_TREE_RETRY XA_RETRY_ENTRY
78
79static inline unsigned long
80get_slot_offset(const struct radix_tree_node *parent, void __rcu **slot)
81{
82 return parent ? slot - parent->slots : 0;
83}
84
85static unsigned int radix_tree_descend(const struct radix_tree_node *parent,
86 struct radix_tree_node **nodep, unsigned long index)
87{
88 unsigned int offset = (index >> parent->shift) & RADIX_TREE_MAP_MASK;
89 void __rcu **entry = rcu_dereference_raw(parent->slots[offset]);
90
91 *nodep = (void *)entry;
92 return offset;
93}
94
95static inline gfp_t root_gfp_mask(const struct radix_tree_root *root)
96{
97 return root->xa_flags & (__GFP_BITS_MASK & ~GFP_ZONEMASK);
98}
99
100static inline void tag_set(struct radix_tree_node *node, unsigned int tag,
101 int offset)
102{
103 __set_bit(offset, node->tags[tag]);
104}
105
106static inline void tag_clear(struct radix_tree_node *node, unsigned int tag,
107 int offset)
108{
109 __clear_bit(offset, node->tags[tag]);
110}
111
112static inline int tag_get(const struct radix_tree_node *node, unsigned int tag,
113 int offset)
114{
115 return test_bit(offset, node->tags[tag]);
116}
117
118static inline void root_tag_set(struct radix_tree_root *root, unsigned tag)
119{
120 root->xa_flags |= (__force gfp_t)(1 << (tag + ROOT_TAG_SHIFT));
121}
122
123static inline void root_tag_clear(struct radix_tree_root *root, unsigned tag)
124{
125 root->xa_flags &= (__force gfp_t)~(1 << (tag + ROOT_TAG_SHIFT));
126}
127
128static inline void root_tag_clear_all(struct radix_tree_root *root)
129{
130 root->xa_flags &= (__force gfp_t)((1 << ROOT_TAG_SHIFT) - 1);
131}
132
133static inline int root_tag_get(const struct radix_tree_root *root, unsigned tag)
134{
135 return (__force int)root->xa_flags & (1 << (tag + ROOT_TAG_SHIFT));
136}
137
138static inline unsigned root_tags_get(const struct radix_tree_root *root)
139{
140 return (__force unsigned)root->xa_flags >> ROOT_TAG_SHIFT;
141}
142
143static inline bool is_idr(const struct radix_tree_root *root)
144{
145 return !!(root->xa_flags & ROOT_IS_IDR);
146}
147
148/*
149 * Returns 1 if any slot in the node has this tag set.
150 * Otherwise returns 0.
151 */
152static inline int any_tag_set(const struct radix_tree_node *node,
153 unsigned int tag)
154{
155 unsigned idx;
156 for (idx = 0; idx < RADIX_TREE_TAG_LONGS; idx++) {
157 if (node->tags[tag][idx])
158 return 1;
159 }
160 return 0;
161}
162
163static inline void all_tag_set(struct radix_tree_node *node, unsigned int tag)
164{
165 bitmap_fill(node->tags[tag], RADIX_TREE_MAP_SIZE);
166}
167
168/**
169 * radix_tree_find_next_bit - find the next set bit in a memory region
170 *
171 * @node: where to begin the search
172 * @tag: the tag index
173 * @offset: the bitnumber to start searching at
174 *
175 * Unrollable variant of find_next_bit() for constant size arrays.
176 * Tail bits starting from size to roundup(size, BITS_PER_LONG) must be zero.
177 * Returns next bit offset, or size if nothing found.
178 */
179static __always_inline unsigned long
180radix_tree_find_next_bit(struct radix_tree_node *node, unsigned int tag,
181 unsigned long offset)
182{
183 const unsigned long *addr = node->tags[tag];
184
185 if (offset < RADIX_TREE_MAP_SIZE) {
186 unsigned long tmp;
187
188 addr += offset / BITS_PER_LONG;
189 tmp = *addr >> (offset % BITS_PER_LONG);
190 if (tmp)
191 return __ffs(tmp) + offset;
192 offset = (offset + BITS_PER_LONG) & ~(BITS_PER_LONG - 1);
193 while (offset < RADIX_TREE_MAP_SIZE) {
194 tmp = *++addr;
195 if (tmp)
196 return __ffs(tmp) + offset;
197 offset += BITS_PER_LONG;
198 }
199 }
200 return RADIX_TREE_MAP_SIZE;
201}
202
203static unsigned int iter_offset(const struct radix_tree_iter *iter)
204{
205 return iter->index & RADIX_TREE_MAP_MASK;
206}
207
208/*
209 * The maximum index which can be stored in a radix tree
210 */
211static inline unsigned long shift_maxindex(unsigned int shift)
212{
213 return (RADIX_TREE_MAP_SIZE << shift) - 1;
214}
215
216static inline unsigned long node_maxindex(const struct radix_tree_node *node)
217{
218 return shift_maxindex(node->shift);
219}
220
221static unsigned long next_index(unsigned long index,
222 const struct radix_tree_node *node,
223 unsigned long offset)
224{
225 return (index & ~node_maxindex(node)) + (offset << node->shift);
226}
227
228/*
229 * This assumes that the caller has performed appropriate preallocation, and
230 * that the caller has pinned this thread of control to the current CPU.
231 */
232static struct radix_tree_node *
233radix_tree_node_alloc(gfp_t gfp_mask, struct radix_tree_node *parent,
234 struct radix_tree_root *root,
235 unsigned int shift, unsigned int offset,
236 unsigned int count, unsigned int nr_values)
237{
238 struct radix_tree_node *ret = NULL;
239
240 /*
241 * Preload code isn't irq safe and it doesn't make sense to use
242 * preloading during an interrupt anyway as all the allocations have
243 * to be atomic. So just do normal allocation when in interrupt.
244 */
245 if (!gfpflags_allow_blocking(gfp_mask) && !in_interrupt()) {
246 struct radix_tree_preload *rtp;
247
248 /*
249 * Even if the caller has preloaded, try to allocate from the
250 * cache first for the new node to get accounted to the memory
251 * cgroup.
252 */
253 ret = kmem_cache_alloc(radix_tree_node_cachep,
254 gfp_mask | __GFP_NOWARN);
255 if (ret)
256 goto out;
257
258 /*
259 * Provided the caller has preloaded here, we will always
260 * succeed in getting a node here (and never reach
261 * kmem_cache_alloc)
262 */
263 rtp = this_cpu_ptr(&radix_tree_preloads);
264 if (rtp->nr) {
265 ret = rtp->nodes;
266 rtp->nodes = ret->parent;
267 rtp->nr--;
268 }
269 /*
270 * Update the allocation stack trace as this is more useful
271 * for debugging.
272 */
273 kmemleak_update_trace(ret);
274 goto out;
275 }
276 ret = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
277out:
278 BUG_ON(radix_tree_is_internal_node(ret));
279 if (ret) {
280 ret->shift = shift;
281 ret->offset = offset;
282 ret->count = count;
283 ret->nr_values = nr_values;
284 ret->parent = parent;
285 ret->array = root;
286 }
287 return ret;
288}
289
290void radix_tree_node_rcu_free(struct rcu_head *head)
291{
292 struct radix_tree_node *node =
293 container_of(head, struct radix_tree_node, rcu_head);
294
295 /*
296 * Must only free zeroed nodes into the slab. We can be left with
297 * non-NULL entries by radix_tree_free_nodes, so clear the entries
298 * and tags here.
299 */
300 memset(node->slots, 0, sizeof(node->slots));
301 memset(node->tags, 0, sizeof(node->tags));
302 INIT_LIST_HEAD(&node->private_list);
303
304 kmem_cache_free(radix_tree_node_cachep, node);
305}
306
307static inline void
308radix_tree_node_free(struct radix_tree_node *node)
309{
310 call_rcu(&node->rcu_head, radix_tree_node_rcu_free);
311}
312
313/*
314 * Load up this CPU's radix_tree_node buffer with sufficient objects to
315 * ensure that the addition of a single element in the tree cannot fail. On
316 * success, return zero, with preemption disabled. On error, return -ENOMEM
317 * with preemption not disabled.
318 *
319 * To make use of this facility, the radix tree must be initialised without
320 * __GFP_DIRECT_RECLAIM being passed to INIT_RADIX_TREE().
321 */
322static __must_check int __radix_tree_preload(gfp_t gfp_mask, unsigned nr)
323{
324 struct radix_tree_preload *rtp;
325 struct radix_tree_node *node;
326 int ret = -ENOMEM;
327
328 /*
329 * Nodes preloaded by one cgroup can be used by another cgroup, so
330 * they should never be accounted to any particular memory cgroup.
331 */
332 gfp_mask &= ~__GFP_ACCOUNT;
333
334 local_lock(&radix_tree_preloads.lock);
335 rtp = this_cpu_ptr(&radix_tree_preloads);
336 while (rtp->nr < nr) {
337 local_unlock(&radix_tree_preloads.lock);
338 node = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
339 if (node == NULL)
340 goto out;
341 local_lock(&radix_tree_preloads.lock);
342 rtp = this_cpu_ptr(&radix_tree_preloads);
343 if (rtp->nr < nr) {
344 node->parent = rtp->nodes;
345 rtp->nodes = node;
346 rtp->nr++;
347 } else {
348 kmem_cache_free(radix_tree_node_cachep, node);
349 }
350 }
351 ret = 0;
352out:
353 return ret;
354}
355
356/*
357 * Load up this CPU's radix_tree_node buffer with sufficient objects to
358 * ensure that the addition of a single element in the tree cannot fail. On
359 * success, return zero, with preemption disabled. On error, return -ENOMEM
360 * with preemption not disabled.
361 *
362 * To make use of this facility, the radix tree must be initialised without
363 * __GFP_DIRECT_RECLAIM being passed to INIT_RADIX_TREE().
364 */
365int radix_tree_preload(gfp_t gfp_mask)
366{
367 /* Warn on non-sensical use... */
368 WARN_ON_ONCE(!gfpflags_allow_blocking(gfp_mask));
369 return __radix_tree_preload(gfp_mask, RADIX_TREE_PRELOAD_SIZE);
370}
371EXPORT_SYMBOL(radix_tree_preload);
372
373/*
374 * The same as above function, except we don't guarantee preloading happens.
375 * We do it, if we decide it helps. On success, return zero with preemption
376 * disabled. On error, return -ENOMEM with preemption not disabled.
377 */
378int radix_tree_maybe_preload(gfp_t gfp_mask)
379{
380 if (gfpflags_allow_blocking(gfp_mask))
381 return __radix_tree_preload(gfp_mask, RADIX_TREE_PRELOAD_SIZE);
382 /* Preloading doesn't help anything with this gfp mask, skip it */
383 local_lock(&radix_tree_preloads.lock);
384 return 0;
385}
386EXPORT_SYMBOL(radix_tree_maybe_preload);
387
388static unsigned radix_tree_load_root(const struct radix_tree_root *root,
389 struct radix_tree_node **nodep, unsigned long *maxindex)
390{
391 struct radix_tree_node *node = rcu_dereference_raw(root->xa_head);
392
393 *nodep = node;
394
395 if (likely(radix_tree_is_internal_node(node))) {
396 node = entry_to_node(node);
397 *maxindex = node_maxindex(node);
398 return node->shift + RADIX_TREE_MAP_SHIFT;
399 }
400
401 *maxindex = 0;
402 return 0;
403}
404
405/*
406 * Extend a radix tree so it can store key @index.
407 */
408static int radix_tree_extend(struct radix_tree_root *root, gfp_t gfp,
409 unsigned long index, unsigned int shift)
410{
411 void *entry;
412 unsigned int maxshift;
413 int tag;
414
415 /* Figure out what the shift should be. */
416 maxshift = shift;
417 while (index > shift_maxindex(maxshift))
418 maxshift += RADIX_TREE_MAP_SHIFT;
419
420 entry = rcu_dereference_raw(root->xa_head);
421 if (!entry && (!is_idr(root) || root_tag_get(root, IDR_FREE)))
422 goto out;
423
424 do {
425 struct radix_tree_node *node = radix_tree_node_alloc(gfp, NULL,
426 root, shift, 0, 1, 0);
427 if (!node)
428 return -ENOMEM;
429
430 if (is_idr(root)) {
431 all_tag_set(node, IDR_FREE);
432 if (!root_tag_get(root, IDR_FREE)) {
433 tag_clear(node, IDR_FREE, 0);
434 root_tag_set(root, IDR_FREE);
435 }
436 } else {
437 /* Propagate the aggregated tag info to the new child */
438 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
439 if (root_tag_get(root, tag))
440 tag_set(node, tag, 0);
441 }
442 }
443
444 BUG_ON(shift > BITS_PER_LONG);
445 if (radix_tree_is_internal_node(entry)) {
446 entry_to_node(entry)->parent = node;
447 } else if (xa_is_value(entry)) {
448 /* Moving a value entry root->xa_head to a node */
449 node->nr_values = 1;
450 }
451 /*
452 * entry was already in the radix tree, so we do not need
453 * rcu_assign_pointer here
454 */
455 node->slots[0] = (void __rcu *)entry;
456 entry = node_to_entry(node);
457 rcu_assign_pointer(root->xa_head, entry);
458 shift += RADIX_TREE_MAP_SHIFT;
459 } while (shift <= maxshift);
460out:
461 return maxshift + RADIX_TREE_MAP_SHIFT;
462}
463
464/**
465 * radix_tree_shrink - shrink radix tree to minimum height
466 * @root: radix tree root
467 */
468static inline bool radix_tree_shrink(struct radix_tree_root *root)
469{
470 bool shrunk = false;
471
472 for (;;) {
473 struct radix_tree_node *node = rcu_dereference_raw(root->xa_head);
474 struct radix_tree_node *child;
475
476 if (!radix_tree_is_internal_node(node))
477 break;
478 node = entry_to_node(node);
479
480 /*
481 * The candidate node has more than one child, or its child
482 * is not at the leftmost slot, we cannot shrink.
483 */
484 if (node->count != 1)
485 break;
486 child = rcu_dereference_raw(node->slots[0]);
487 if (!child)
488 break;
489
490 /*
491 * For an IDR, we must not shrink entry 0 into the root in
492 * case somebody calls idr_replace() with a pointer that
493 * appears to be an internal entry
494 */
495 if (!node->shift && is_idr(root))
496 break;
497
498 if (radix_tree_is_internal_node(child))
499 entry_to_node(child)->parent = NULL;
500
501 /*
502 * We don't need rcu_assign_pointer(), since we are simply
503 * moving the node from one part of the tree to another: if it
504 * was safe to dereference the old pointer to it
505 * (node->slots[0]), it will be safe to dereference the new
506 * one (root->xa_head) as far as dependent read barriers go.
507 */
508 root->xa_head = (void __rcu *)child;
509 if (is_idr(root) && !tag_get(node, IDR_FREE, 0))
510 root_tag_clear(root, IDR_FREE);
511
512 /*
513 * We have a dilemma here. The node's slot[0] must not be
514 * NULLed in case there are concurrent lookups expecting to
515 * find the item. However if this was a bottom-level node,
516 * then it may be subject to the slot pointer being visible
517 * to callers dereferencing it. If item corresponding to
518 * slot[0] is subsequently deleted, these callers would expect
519 * their slot to become empty sooner or later.
520 *
521 * For example, lockless pagecache will look up a slot, deref
522 * the page pointer, and if the page has 0 refcount it means it
523 * was concurrently deleted from pagecache so try the deref
524 * again. Fortunately there is already a requirement for logic
525 * to retry the entire slot lookup -- the indirect pointer
526 * problem (replacing direct root node with an indirect pointer
527 * also results in a stale slot). So tag the slot as indirect
528 * to force callers to retry.
529 */
530 node->count = 0;
531 if (!radix_tree_is_internal_node(child)) {
532 node->slots[0] = (void __rcu *)RADIX_TREE_RETRY;
533 }
534
535 WARN_ON_ONCE(!list_empty(&node->private_list));
536 radix_tree_node_free(node);
537 shrunk = true;
538 }
539
540 return shrunk;
541}
542
543static bool delete_node(struct radix_tree_root *root,
544 struct radix_tree_node *node)
545{
546 bool deleted = false;
547
548 do {
549 struct radix_tree_node *parent;
550
551 if (node->count) {
552 if (node_to_entry(node) ==
553 rcu_dereference_raw(root->xa_head))
554 deleted |= radix_tree_shrink(root);
555 return deleted;
556 }
557
558 parent = node->parent;
559 if (parent) {
560 parent->slots[node->offset] = NULL;
561 parent->count--;
562 } else {
563 /*
564 * Shouldn't the tags already have all been cleared
565 * by the caller?
566 */
567 if (!is_idr(root))
568 root_tag_clear_all(root);
569 root->xa_head = NULL;
570 }
571
572 WARN_ON_ONCE(!list_empty(&node->private_list));
573 radix_tree_node_free(node);
574 deleted = true;
575
576 node = parent;
577 } while (node);
578
579 return deleted;
580}
581
582/**
583 * __radix_tree_create - create a slot in a radix tree
584 * @root: radix tree root
585 * @index: index key
586 * @nodep: returns node
587 * @slotp: returns slot
588 *
589 * Create, if necessary, and return the node and slot for an item
590 * at position @index in the radix tree @root.
591 *
592 * Until there is more than one item in the tree, no nodes are
593 * allocated and @root->xa_head is used as a direct slot instead of
594 * pointing to a node, in which case *@nodep will be NULL.
595 *
596 * Returns -ENOMEM, or 0 for success.
597 */
598static int __radix_tree_create(struct radix_tree_root *root,
599 unsigned long index, struct radix_tree_node **nodep,
600 void __rcu ***slotp)
601{
602 struct radix_tree_node *node = NULL, *child;
603 void __rcu **slot = (void __rcu **)&root->xa_head;
604 unsigned long maxindex;
605 unsigned int shift, offset = 0;
606 unsigned long max = index;
607 gfp_t gfp = root_gfp_mask(root);
608
609 shift = radix_tree_load_root(root, &child, &maxindex);
610
611 /* Make sure the tree is high enough. */
612 if (max > maxindex) {
613 int error = radix_tree_extend(root, gfp, max, shift);
614 if (error < 0)
615 return error;
616 shift = error;
617 child = rcu_dereference_raw(root->xa_head);
618 }
619
620 while (shift > 0) {
621 shift -= RADIX_TREE_MAP_SHIFT;
622 if (child == NULL) {
623 /* Have to add a child node. */
624 child = radix_tree_node_alloc(gfp, node, root, shift,
625 offset, 0, 0);
626 if (!child)
627 return -ENOMEM;
628 rcu_assign_pointer(*slot, node_to_entry(child));
629 if (node)
630 node->count++;
631 } else if (!radix_tree_is_internal_node(child))
632 break;
633
634 /* Go a level down */
635 node = entry_to_node(child);
636 offset = radix_tree_descend(node, &child, index);
637 slot = &node->slots[offset];
638 }
639
640 if (nodep)
641 *nodep = node;
642 if (slotp)
643 *slotp = slot;
644 return 0;
645}
646
647/*
648 * Free any nodes below this node. The tree is presumed to not need
649 * shrinking, and any user data in the tree is presumed to not need a
650 * destructor called on it. If we need to add a destructor, we can
651 * add that functionality later. Note that we may not clear tags or
652 * slots from the tree as an RCU walker may still have a pointer into
653 * this subtree. We could replace the entries with RADIX_TREE_RETRY,
654 * but we'll still have to clear those in rcu_free.
655 */
656static void radix_tree_free_nodes(struct radix_tree_node *node)
657{
658 unsigned offset = 0;
659 struct radix_tree_node *child = entry_to_node(node);
660
661 for (;;) {
662 void *entry = rcu_dereference_raw(child->slots[offset]);
663 if (xa_is_node(entry) && child->shift) {
664 child = entry_to_node(entry);
665 offset = 0;
666 continue;
667 }
668 offset++;
669 while (offset == RADIX_TREE_MAP_SIZE) {
670 struct radix_tree_node *old = child;
671 offset = child->offset + 1;
672 child = child->parent;
673 WARN_ON_ONCE(!list_empty(&old->private_list));
674 radix_tree_node_free(old);
675 if (old == entry_to_node(node))
676 return;
677 }
678 }
679}
680
681static inline int insert_entries(struct radix_tree_node *node,
682 void __rcu **slot, void *item)
683{
684 if (*slot)
685 return -EEXIST;
686 rcu_assign_pointer(*slot, item);
687 if (node) {
688 node->count++;
689 if (xa_is_value(item))
690 node->nr_values++;
691 }
692 return 1;
693}
694
695/**
696 * radix_tree_insert - insert into a radix tree
697 * @root: radix tree root
698 * @index: index key
699 * @item: item to insert
700 *
701 * Insert an item into the radix tree at position @index.
702 */
703int radix_tree_insert(struct radix_tree_root *root, unsigned long index,
704 void *item)
705{
706 struct radix_tree_node *node;
707 void __rcu **slot;
708 int error;
709
710 BUG_ON(radix_tree_is_internal_node(item));
711
712 error = __radix_tree_create(root, index, &node, &slot);
713 if (error)
714 return error;
715
716 error = insert_entries(node, slot, item);
717 if (error < 0)
718 return error;
719
720 if (node) {
721 unsigned offset = get_slot_offset(node, slot);
722 BUG_ON(tag_get(node, 0, offset));
723 BUG_ON(tag_get(node, 1, offset));
724 BUG_ON(tag_get(node, 2, offset));
725 } else {
726 BUG_ON(root_tags_get(root));
727 }
728
729 return 0;
730}
731EXPORT_SYMBOL(radix_tree_insert);
732
733/**
734 * __radix_tree_lookup - lookup an item in a radix tree
735 * @root: radix tree root
736 * @index: index key
737 * @nodep: returns node
738 * @slotp: returns slot
739 *
740 * Lookup and return the item at position @index in the radix
741 * tree @root.
742 *
743 * Until there is more than one item in the tree, no nodes are
744 * allocated and @root->xa_head is used as a direct slot instead of
745 * pointing to a node, in which case *@nodep will be NULL.
746 */
747void *__radix_tree_lookup(const struct radix_tree_root *root,
748 unsigned long index, struct radix_tree_node **nodep,
749 void __rcu ***slotp)
750{
751 struct radix_tree_node *node, *parent;
752 unsigned long maxindex;
753 void __rcu **slot;
754
755 restart:
756 parent = NULL;
757 slot = (void __rcu **)&root->xa_head;
758 radix_tree_load_root(root, &node, &maxindex);
759 if (index > maxindex)
760 return NULL;
761
762 while (radix_tree_is_internal_node(node)) {
763 unsigned offset;
764
765 parent = entry_to_node(node);
766 offset = radix_tree_descend(parent, &node, index);
767 slot = parent->slots + offset;
768 if (node == RADIX_TREE_RETRY)
769 goto restart;
770 if (parent->shift == 0)
771 break;
772 }
773
774 if (nodep)
775 *nodep = parent;
776 if (slotp)
777 *slotp = slot;
778 return node;
779}
780
781/**
782 * radix_tree_lookup_slot - lookup a slot in a radix tree
783 * @root: radix tree root
784 * @index: index key
785 *
786 * Returns: the slot corresponding to the position @index in the
787 * radix tree @root. This is useful for update-if-exists operations.
788 *
789 * This function can be called under rcu_read_lock iff the slot is not
790 * modified by radix_tree_replace_slot, otherwise it must be called
791 * exclusive from other writers. Any dereference of the slot must be done
792 * using radix_tree_deref_slot.
793 */
794void __rcu **radix_tree_lookup_slot(const struct radix_tree_root *root,
795 unsigned long index)
796{
797 void __rcu **slot;
798
799 if (!__radix_tree_lookup(root, index, NULL, &slot))
800 return NULL;
801 return slot;
802}
803EXPORT_SYMBOL(radix_tree_lookup_slot);
804
805/**
806 * radix_tree_lookup - perform lookup operation on a radix tree
807 * @root: radix tree root
808 * @index: index key
809 *
810 * Lookup the item at the position @index in the radix tree @root.
811 *
812 * This function can be called under rcu_read_lock, however the caller
813 * must manage lifetimes of leaf nodes (eg. RCU may also be used to free
814 * them safely). No RCU barriers are required to access or modify the
815 * returned item, however.
816 */
817void *radix_tree_lookup(const struct radix_tree_root *root, unsigned long index)
818{
819 return __radix_tree_lookup(root, index, NULL, NULL);
820}
821EXPORT_SYMBOL(radix_tree_lookup);
822
823static void replace_slot(void __rcu **slot, void *item,
824 struct radix_tree_node *node, int count, int values)
825{
826 if (node && (count || values)) {
827 node->count += count;
828 node->nr_values += values;
829 }
830
831 rcu_assign_pointer(*slot, item);
832}
833
834static bool node_tag_get(const struct radix_tree_root *root,
835 const struct radix_tree_node *node,
836 unsigned int tag, unsigned int offset)
837{
838 if (node)
839 return tag_get(node, tag, offset);
840 return root_tag_get(root, tag);
841}
842
843/*
844 * IDR users want to be able to store NULL in the tree, so if the slot isn't
845 * free, don't adjust the count, even if it's transitioning between NULL and
846 * non-NULL. For the IDA, we mark slots as being IDR_FREE while they still
847 * have empty bits, but it only stores NULL in slots when they're being
848 * deleted.
849 */
850static int calculate_count(struct radix_tree_root *root,
851 struct radix_tree_node *node, void __rcu **slot,
852 void *item, void *old)
853{
854 if (is_idr(root)) {
855 unsigned offset = get_slot_offset(node, slot);
856 bool free = node_tag_get(root, node, IDR_FREE, offset);
857 if (!free)
858 return 0;
859 if (!old)
860 return 1;
861 }
862 return !!item - !!old;
863}
864
865/**
866 * __radix_tree_replace - replace item in a slot
867 * @root: radix tree root
868 * @node: pointer to tree node
869 * @slot: pointer to slot in @node
870 * @item: new item to store in the slot.
871 *
872 * For use with __radix_tree_lookup(). Caller must hold tree write locked
873 * across slot lookup and replacement.
874 */
875void __radix_tree_replace(struct radix_tree_root *root,
876 struct radix_tree_node *node,
877 void __rcu **slot, void *item)
878{
879 void *old = rcu_dereference_raw(*slot);
880 int values = !!xa_is_value(item) - !!xa_is_value(old);
881 int count = calculate_count(root, node, slot, item, old);
882
883 /*
884 * This function supports replacing value entries and
885 * deleting entries, but that needs accounting against the
886 * node unless the slot is root->xa_head.
887 */
888 WARN_ON_ONCE(!node && (slot != (void __rcu **)&root->xa_head) &&
889 (count || values));
890 replace_slot(slot, item, node, count, values);
891
892 if (!node)
893 return;
894
895 delete_node(root, node);
896}
897
898/**
899 * radix_tree_replace_slot - replace item in a slot
900 * @root: radix tree root
901 * @slot: pointer to slot
902 * @item: new item to store in the slot.
903 *
904 * For use with radix_tree_lookup_slot() and
905 * radix_tree_gang_lookup_tag_slot(). Caller must hold tree write locked
906 * across slot lookup and replacement.
907 *
908 * NOTE: This cannot be used to switch between non-entries (empty slots),
909 * regular entries, and value entries, as that requires accounting
910 * inside the radix tree node. When switching from one type of entry or
911 * deleting, use __radix_tree_lookup() and __radix_tree_replace() or
912 * radix_tree_iter_replace().
913 */
914void radix_tree_replace_slot(struct radix_tree_root *root,
915 void __rcu **slot, void *item)
916{
917 __radix_tree_replace(root, NULL, slot, item);
918}
919EXPORT_SYMBOL(radix_tree_replace_slot);
920
921/**
922 * radix_tree_iter_replace - replace item in a slot
923 * @root: radix tree root
924 * @iter: iterator state
925 * @slot: pointer to slot
926 * @item: new item to store in the slot.
927 *
928 * For use with radix_tree_for_each_slot().
929 * Caller must hold tree write locked.
930 */
931void radix_tree_iter_replace(struct radix_tree_root *root,
932 const struct radix_tree_iter *iter,
933 void __rcu **slot, void *item)
934{
935 __radix_tree_replace(root, iter->node, slot, item);
936}
937
938static void node_tag_set(struct radix_tree_root *root,
939 struct radix_tree_node *node,
940 unsigned int tag, unsigned int offset)
941{
942 while (node) {
943 if (tag_get(node, tag, offset))
944 return;
945 tag_set(node, tag, offset);
946 offset = node->offset;
947 node = node->parent;
948 }
949
950 if (!root_tag_get(root, tag))
951 root_tag_set(root, tag);
952}
953
954/**
955 * radix_tree_tag_set - set a tag on a radix tree node
956 * @root: radix tree root
957 * @index: index key
958 * @tag: tag index
959 *
960 * Set the search tag (which must be < RADIX_TREE_MAX_TAGS)
961 * corresponding to @index in the radix tree. From
962 * the root all the way down to the leaf node.
963 *
964 * Returns the address of the tagged item. Setting a tag on a not-present
965 * item is a bug.
966 */
967void *radix_tree_tag_set(struct radix_tree_root *root,
968 unsigned long index, unsigned int tag)
969{
970 struct radix_tree_node *node, *parent;
971 unsigned long maxindex;
972
973 radix_tree_load_root(root, &node, &maxindex);
974 BUG_ON(index > maxindex);
975
976 while (radix_tree_is_internal_node(node)) {
977 unsigned offset;
978
979 parent = entry_to_node(node);
980 offset = radix_tree_descend(parent, &node, index);
981 BUG_ON(!node);
982
983 if (!tag_get(parent, tag, offset))
984 tag_set(parent, tag, offset);
985 }
986
987 /* set the root's tag bit */
988 if (!root_tag_get(root, tag))
989 root_tag_set(root, tag);
990
991 return node;
992}
993EXPORT_SYMBOL(radix_tree_tag_set);
994
995static void node_tag_clear(struct radix_tree_root *root,
996 struct radix_tree_node *node,
997 unsigned int tag, unsigned int offset)
998{
999 while (node) {
1000 if (!tag_get(node, tag, offset))
1001 return;
1002 tag_clear(node, tag, offset);
1003 if (any_tag_set(node, tag))
1004 return;
1005
1006 offset = node->offset;
1007 node = node->parent;
1008 }
1009
1010 /* clear the root's tag bit */
1011 if (root_tag_get(root, tag))
1012 root_tag_clear(root, tag);
1013}
1014
1015/**
1016 * radix_tree_tag_clear - clear a tag on a radix tree node
1017 * @root: radix tree root
1018 * @index: index key
1019 * @tag: tag index
1020 *
1021 * Clear the search tag (which must be < RADIX_TREE_MAX_TAGS)
1022 * corresponding to @index in the radix tree. If this causes
1023 * the leaf node to have no tags set then clear the tag in the
1024 * next-to-leaf node, etc.
1025 *
1026 * Returns the address of the tagged item on success, else NULL. ie:
1027 * has the same return value and semantics as radix_tree_lookup().
1028 */
1029void *radix_tree_tag_clear(struct radix_tree_root *root,
1030 unsigned long index, unsigned int tag)
1031{
1032 struct radix_tree_node *node, *parent;
1033 unsigned long maxindex;
1034 int offset = 0;
1035
1036 radix_tree_load_root(root, &node, &maxindex);
1037 if (index > maxindex)
1038 return NULL;
1039
1040 parent = NULL;
1041
1042 while (radix_tree_is_internal_node(node)) {
1043 parent = entry_to_node(node);
1044 offset = radix_tree_descend(parent, &node, index);
1045 }
1046
1047 if (node)
1048 node_tag_clear(root, parent, tag, offset);
1049
1050 return node;
1051}
1052EXPORT_SYMBOL(radix_tree_tag_clear);
1053
1054/**
1055 * radix_tree_iter_tag_clear - clear a tag on the current iterator entry
1056 * @root: radix tree root
1057 * @iter: iterator state
1058 * @tag: tag to clear
1059 */
1060void radix_tree_iter_tag_clear(struct radix_tree_root *root,
1061 const struct radix_tree_iter *iter, unsigned int tag)
1062{
1063 node_tag_clear(root, iter->node, tag, iter_offset(iter));
1064}
1065
1066/**
1067 * radix_tree_tag_get - get a tag on a radix tree node
1068 * @root: radix tree root
1069 * @index: index key
1070 * @tag: tag index (< RADIX_TREE_MAX_TAGS)
1071 *
1072 * Return values:
1073 *
1074 * 0: tag not present or not set
1075 * 1: tag set
1076 *
1077 * Note that the return value of this function may not be relied on, even if
1078 * the RCU lock is held, unless tag modification and node deletion are excluded
1079 * from concurrency.
1080 */
1081int radix_tree_tag_get(const struct radix_tree_root *root,
1082 unsigned long index, unsigned int tag)
1083{
1084 struct radix_tree_node *node, *parent;
1085 unsigned long maxindex;
1086
1087 if (!root_tag_get(root, tag))
1088 return 0;
1089
1090 radix_tree_load_root(root, &node, &maxindex);
1091 if (index > maxindex)
1092 return 0;
1093
1094 while (radix_tree_is_internal_node(node)) {
1095 unsigned offset;
1096
1097 parent = entry_to_node(node);
1098 offset = radix_tree_descend(parent, &node, index);
1099
1100 if (!tag_get(parent, tag, offset))
1101 return 0;
1102 if (node == RADIX_TREE_RETRY)
1103 break;
1104 }
1105
1106 return 1;
1107}
1108EXPORT_SYMBOL(radix_tree_tag_get);
1109
1110/* Construct iter->tags bit-mask from node->tags[tag] array */
1111static void set_iter_tags(struct radix_tree_iter *iter,
1112 struct radix_tree_node *node, unsigned offset,
1113 unsigned tag)
1114{
1115 unsigned tag_long = offset / BITS_PER_LONG;
1116 unsigned tag_bit = offset % BITS_PER_LONG;
1117
1118 if (!node) {
1119 iter->tags = 1;
1120 return;
1121 }
1122
1123 iter->tags = node->tags[tag][tag_long] >> tag_bit;
1124
1125 /* This never happens if RADIX_TREE_TAG_LONGS == 1 */
1126 if (tag_long < RADIX_TREE_TAG_LONGS - 1) {
1127 /* Pick tags from next element */
1128 if (tag_bit)
1129 iter->tags |= node->tags[tag][tag_long + 1] <<
1130 (BITS_PER_LONG - tag_bit);
1131 /* Clip chunk size, here only BITS_PER_LONG tags */
1132 iter->next_index = __radix_tree_iter_add(iter, BITS_PER_LONG);
1133 }
1134}
1135
1136void __rcu **radix_tree_iter_resume(void __rcu **slot,
1137 struct radix_tree_iter *iter)
1138{
1139 iter->index = __radix_tree_iter_add(iter, 1);
1140 iter->next_index = iter->index;
1141 iter->tags = 0;
1142 return NULL;
1143}
1144EXPORT_SYMBOL(radix_tree_iter_resume);
1145
1146/**
1147 * radix_tree_next_chunk - find next chunk of slots for iteration
1148 *
1149 * @root: radix tree root
1150 * @iter: iterator state
1151 * @flags: RADIX_TREE_ITER_* flags and tag index
1152 * Returns: pointer to chunk first slot, or NULL if iteration is over
1153 */
1154void __rcu **radix_tree_next_chunk(const struct radix_tree_root *root,
1155 struct radix_tree_iter *iter, unsigned flags)
1156{
1157 unsigned tag = flags & RADIX_TREE_ITER_TAG_MASK;
1158 struct radix_tree_node *node, *child;
1159 unsigned long index, offset, maxindex;
1160
1161 if ((flags & RADIX_TREE_ITER_TAGGED) && !root_tag_get(root, tag))
1162 return NULL;
1163
1164 /*
1165 * Catch next_index overflow after ~0UL. iter->index never overflows
1166 * during iterating; it can be zero only at the beginning.
1167 * And we cannot overflow iter->next_index in a single step,
1168 * because RADIX_TREE_MAP_SHIFT < BITS_PER_LONG.
1169 *
1170 * This condition also used by radix_tree_next_slot() to stop
1171 * contiguous iterating, and forbid switching to the next chunk.
1172 */
1173 index = iter->next_index;
1174 if (!index && iter->index)
1175 return NULL;
1176
1177 restart:
1178 radix_tree_load_root(root, &child, &maxindex);
1179 if (index > maxindex)
1180 return NULL;
1181 if (!child)
1182 return NULL;
1183
1184 if (!radix_tree_is_internal_node(child)) {
1185 /* Single-slot tree */
1186 iter->index = index;
1187 iter->next_index = maxindex + 1;
1188 iter->tags = 1;
1189 iter->node = NULL;
1190 return (void __rcu **)&root->xa_head;
1191 }
1192
1193 do {
1194 node = entry_to_node(child);
1195 offset = radix_tree_descend(node, &child, index);
1196
1197 if ((flags & RADIX_TREE_ITER_TAGGED) ?
1198 !tag_get(node, tag, offset) : !child) {
1199 /* Hole detected */
1200 if (flags & RADIX_TREE_ITER_CONTIG)
1201 return NULL;
1202
1203 if (flags & RADIX_TREE_ITER_TAGGED)
1204 offset = radix_tree_find_next_bit(node, tag,
1205 offset + 1);
1206 else
1207 while (++offset < RADIX_TREE_MAP_SIZE) {
1208 void *slot = rcu_dereference_raw(
1209 node->slots[offset]);
1210 if (slot)
1211 break;
1212 }
1213 index &= ~node_maxindex(node);
1214 index += offset << node->shift;
1215 /* Overflow after ~0UL */
1216 if (!index)
1217 return NULL;
1218 if (offset == RADIX_TREE_MAP_SIZE)
1219 goto restart;
1220 child = rcu_dereference_raw(node->slots[offset]);
1221 }
1222
1223 if (!child)
1224 goto restart;
1225 if (child == RADIX_TREE_RETRY)
1226 break;
1227 } while (node->shift && radix_tree_is_internal_node(child));
1228
1229 /* Update the iterator state */
1230 iter->index = (index &~ node_maxindex(node)) | offset;
1231 iter->next_index = (index | node_maxindex(node)) + 1;
1232 iter->node = node;
1233
1234 if (flags & RADIX_TREE_ITER_TAGGED)
1235 set_iter_tags(iter, node, offset, tag);
1236
1237 return node->slots + offset;
1238}
1239EXPORT_SYMBOL(radix_tree_next_chunk);
1240
1241/**
1242 * radix_tree_gang_lookup - perform multiple lookup on a radix tree
1243 * @root: radix tree root
1244 * @results: where the results of the lookup are placed
1245 * @first_index: start the lookup from this key
1246 * @max_items: place up to this many items at *results
1247 *
1248 * Performs an index-ascending scan of the tree for present items. Places
1249 * them at *@results and returns the number of items which were placed at
1250 * *@results.
1251 *
1252 * The implementation is naive.
1253 *
1254 * Like radix_tree_lookup, radix_tree_gang_lookup may be called under
1255 * rcu_read_lock. In this case, rather than the returned results being
1256 * an atomic snapshot of the tree at a single point in time, the
1257 * semantics of an RCU protected gang lookup are as though multiple
1258 * radix_tree_lookups have been issued in individual locks, and results
1259 * stored in 'results'.
1260 */
1261unsigned int
1262radix_tree_gang_lookup(const struct radix_tree_root *root, void **results,
1263 unsigned long first_index, unsigned int max_items)
1264{
1265 struct radix_tree_iter iter;
1266 void __rcu **slot;
1267 unsigned int ret = 0;
1268
1269 if (unlikely(!max_items))
1270 return 0;
1271
1272 radix_tree_for_each_slot(slot, root, &iter, first_index) {
1273 results[ret] = rcu_dereference_raw(*slot);
1274 if (!results[ret])
1275 continue;
1276 if (radix_tree_is_internal_node(results[ret])) {
1277 slot = radix_tree_iter_retry(&iter);
1278 continue;
1279 }
1280 if (++ret == max_items)
1281 break;
1282 }
1283
1284 return ret;
1285}
1286EXPORT_SYMBOL(radix_tree_gang_lookup);
1287
1288/**
1289 * radix_tree_gang_lookup_tag - perform multiple lookup on a radix tree
1290 * based on a tag
1291 * @root: radix tree root
1292 * @results: where the results of the lookup are placed
1293 * @first_index: start the lookup from this key
1294 * @max_items: place up to this many items at *results
1295 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1296 *
1297 * Performs an index-ascending scan of the tree for present items which
1298 * have the tag indexed by @tag set. Places the items at *@results and
1299 * returns the number of items which were placed at *@results.
1300 */
1301unsigned int
1302radix_tree_gang_lookup_tag(const struct radix_tree_root *root, void **results,
1303 unsigned long first_index, unsigned int max_items,
1304 unsigned int tag)
1305{
1306 struct radix_tree_iter iter;
1307 void __rcu **slot;
1308 unsigned int ret = 0;
1309
1310 if (unlikely(!max_items))
1311 return 0;
1312
1313 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1314 results[ret] = rcu_dereference_raw(*slot);
1315 if (!results[ret])
1316 continue;
1317 if (radix_tree_is_internal_node(results[ret])) {
1318 slot = radix_tree_iter_retry(&iter);
1319 continue;
1320 }
1321 if (++ret == max_items)
1322 break;
1323 }
1324
1325 return ret;
1326}
1327EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
1328
1329/**
1330 * radix_tree_gang_lookup_tag_slot - perform multiple slot lookup on a
1331 * radix tree based on a tag
1332 * @root: radix tree root
1333 * @results: where the results of the lookup are placed
1334 * @first_index: start the lookup from this key
1335 * @max_items: place up to this many items at *results
1336 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1337 *
1338 * Performs an index-ascending scan of the tree for present items which
1339 * have the tag indexed by @tag set. Places the slots at *@results and
1340 * returns the number of slots which were placed at *@results.
1341 */
1342unsigned int
1343radix_tree_gang_lookup_tag_slot(const struct radix_tree_root *root,
1344 void __rcu ***results, unsigned long first_index,
1345 unsigned int max_items, unsigned int tag)
1346{
1347 struct radix_tree_iter iter;
1348 void __rcu **slot;
1349 unsigned int ret = 0;
1350
1351 if (unlikely(!max_items))
1352 return 0;
1353
1354 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1355 results[ret] = slot;
1356 if (++ret == max_items)
1357 break;
1358 }
1359
1360 return ret;
1361}
1362EXPORT_SYMBOL(radix_tree_gang_lookup_tag_slot);
1363
1364static bool __radix_tree_delete(struct radix_tree_root *root,
1365 struct radix_tree_node *node, void __rcu **slot)
1366{
1367 void *old = rcu_dereference_raw(*slot);
1368 int values = xa_is_value(old) ? -1 : 0;
1369 unsigned offset = get_slot_offset(node, slot);
1370 int tag;
1371
1372 if (is_idr(root))
1373 node_tag_set(root, node, IDR_FREE, offset);
1374 else
1375 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++)
1376 node_tag_clear(root, node, tag, offset);
1377
1378 replace_slot(slot, NULL, node, -1, values);
1379 return node && delete_node(root, node);
1380}
1381
1382/**
1383 * radix_tree_iter_delete - delete the entry at this iterator position
1384 * @root: radix tree root
1385 * @iter: iterator state
1386 * @slot: pointer to slot
1387 *
1388 * Delete the entry at the position currently pointed to by the iterator.
1389 * This may result in the current node being freed; if it is, the iterator
1390 * is advanced so that it will not reference the freed memory. This
1391 * function may be called without any locking if there are no other threads
1392 * which can access this tree.
1393 */
1394void radix_tree_iter_delete(struct radix_tree_root *root,
1395 struct radix_tree_iter *iter, void __rcu **slot)
1396{
1397 if (__radix_tree_delete(root, iter->node, slot))
1398 iter->index = iter->next_index;
1399}
1400EXPORT_SYMBOL(radix_tree_iter_delete);
1401
1402/**
1403 * radix_tree_delete_item - delete an item from a radix tree
1404 * @root: radix tree root
1405 * @index: index key
1406 * @item: expected item
1407 *
1408 * Remove @item at @index from the radix tree rooted at @root.
1409 *
1410 * Return: the deleted entry, or %NULL if it was not present
1411 * or the entry at the given @index was not @item.
1412 */
1413void *radix_tree_delete_item(struct radix_tree_root *root,
1414 unsigned long index, void *item)
1415{
1416 struct radix_tree_node *node = NULL;
1417 void __rcu **slot = NULL;
1418 void *entry;
1419
1420 entry = __radix_tree_lookup(root, index, &node, &slot);
1421 if (!slot)
1422 return NULL;
1423 if (!entry && (!is_idr(root) || node_tag_get(root, node, IDR_FREE,
1424 get_slot_offset(node, slot))))
1425 return NULL;
1426
1427 if (item && entry != item)
1428 return NULL;
1429
1430 __radix_tree_delete(root, node, slot);
1431
1432 return entry;
1433}
1434EXPORT_SYMBOL(radix_tree_delete_item);
1435
1436/**
1437 * radix_tree_delete - delete an entry from a radix tree
1438 * @root: radix tree root
1439 * @index: index key
1440 *
1441 * Remove the entry at @index from the radix tree rooted at @root.
1442 *
1443 * Return: The deleted entry, or %NULL if it was not present.
1444 */
1445void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
1446{
1447 return radix_tree_delete_item(root, index, NULL);
1448}
1449EXPORT_SYMBOL(radix_tree_delete);
1450
1451/**
1452 * radix_tree_tagged - test whether any items in the tree are tagged
1453 * @root: radix tree root
1454 * @tag: tag to test
1455 */
1456int radix_tree_tagged(const struct radix_tree_root *root, unsigned int tag)
1457{
1458 return root_tag_get(root, tag);
1459}
1460EXPORT_SYMBOL(radix_tree_tagged);
1461
1462/**
1463 * idr_preload - preload for idr_alloc()
1464 * @gfp_mask: allocation mask to use for preloading
1465 *
1466 * Preallocate memory to use for the next call to idr_alloc(). This function
1467 * returns with preemption disabled. It will be enabled by idr_preload_end().
1468 */
1469void idr_preload(gfp_t gfp_mask)
1470{
1471 if (__radix_tree_preload(gfp_mask, IDR_PRELOAD_SIZE))
1472 local_lock(&radix_tree_preloads.lock);
1473}
1474EXPORT_SYMBOL(idr_preload);
1475
1476void __rcu **idr_get_free(struct radix_tree_root *root,
1477 struct radix_tree_iter *iter, gfp_t gfp,
1478 unsigned long max)
1479{
1480 struct radix_tree_node *node = NULL, *child;
1481 void __rcu **slot = (void __rcu **)&root->xa_head;
1482 unsigned long maxindex, start = iter->next_index;
1483 unsigned int shift, offset = 0;
1484
1485 grow:
1486 shift = radix_tree_load_root(root, &child, &maxindex);
1487 if (!radix_tree_tagged(root, IDR_FREE))
1488 start = max(start, maxindex + 1);
1489 if (start > max)
1490 return ERR_PTR(-ENOSPC);
1491
1492 if (start > maxindex) {
1493 int error = radix_tree_extend(root, gfp, start, shift);
1494 if (error < 0)
1495 return ERR_PTR(error);
1496 shift = error;
1497 child = rcu_dereference_raw(root->xa_head);
1498 }
1499 if (start == 0 && shift == 0)
1500 shift = RADIX_TREE_MAP_SHIFT;
1501
1502 while (shift) {
1503 shift -= RADIX_TREE_MAP_SHIFT;
1504 if (child == NULL) {
1505 /* Have to add a child node. */
1506 child = radix_tree_node_alloc(gfp, node, root, shift,
1507 offset, 0, 0);
1508 if (!child)
1509 return ERR_PTR(-ENOMEM);
1510 all_tag_set(child, IDR_FREE);
1511 rcu_assign_pointer(*slot, node_to_entry(child));
1512 if (node)
1513 node->count++;
1514 } else if (!radix_tree_is_internal_node(child))
1515 break;
1516
1517 node = entry_to_node(child);
1518 offset = radix_tree_descend(node, &child, start);
1519 if (!tag_get(node, IDR_FREE, offset)) {
1520 offset = radix_tree_find_next_bit(node, IDR_FREE,
1521 offset + 1);
1522 start = next_index(start, node, offset);
1523 if (start > max || start == 0)
1524 return ERR_PTR(-ENOSPC);
1525 while (offset == RADIX_TREE_MAP_SIZE) {
1526 offset = node->offset + 1;
1527 node = node->parent;
1528 if (!node)
1529 goto grow;
1530 shift = node->shift;
1531 }
1532 child = rcu_dereference_raw(node->slots[offset]);
1533 }
1534 slot = &node->slots[offset];
1535 }
1536
1537 iter->index = start;
1538 if (node)
1539 iter->next_index = 1 + min(max, (start | node_maxindex(node)));
1540 else
1541 iter->next_index = 1;
1542 iter->node = node;
1543 set_iter_tags(iter, node, offset, IDR_FREE);
1544
1545 return slot;
1546}
1547
1548/**
1549 * idr_destroy - release all internal memory from an IDR
1550 * @idr: idr handle
1551 *
1552 * After this function is called, the IDR is empty, and may be reused or
1553 * the data structure containing it may be freed.
1554 *
1555 * A typical clean-up sequence for objects stored in an idr tree will use
1556 * idr_for_each() to free all objects, if necessary, then idr_destroy() to
1557 * free the memory used to keep track of those objects.
1558 */
1559void idr_destroy(struct idr *idr)
1560{
1561 struct radix_tree_node *node = rcu_dereference_raw(idr->idr_rt.xa_head);
1562 if (radix_tree_is_internal_node(node))
1563 radix_tree_free_nodes(node);
1564 idr->idr_rt.xa_head = NULL;
1565 root_tag_set(&idr->idr_rt, IDR_FREE);
1566}
1567EXPORT_SYMBOL(idr_destroy);
1568
1569static void
1570radix_tree_node_ctor(void *arg)
1571{
1572 struct radix_tree_node *node = arg;
1573
1574 memset(node, 0, sizeof(*node));
1575 INIT_LIST_HEAD(&node->private_list);
1576}
1577
1578static int radix_tree_cpu_dead(unsigned int cpu)
1579{
1580 struct radix_tree_preload *rtp;
1581 struct radix_tree_node *node;
1582
1583 /* Free per-cpu pool of preloaded nodes */
1584 rtp = &per_cpu(radix_tree_preloads, cpu);
1585 while (rtp->nr) {
1586 node = rtp->nodes;
1587 rtp->nodes = node->parent;
1588 kmem_cache_free(radix_tree_node_cachep, node);
1589 rtp->nr--;
1590 }
1591 return 0;
1592}
1593
1594void __init radix_tree_init(void)
1595{
1596 int ret;
1597
1598 BUILD_BUG_ON(RADIX_TREE_MAX_TAGS + __GFP_BITS_SHIFT > 32);
1599 BUILD_BUG_ON(ROOT_IS_IDR & ~GFP_ZONEMASK);
1600 BUILD_BUG_ON(XA_CHUNK_SIZE > 255);
1601 radix_tree_node_cachep = kmem_cache_create("radix_tree_node",
1602 sizeof(struct radix_tree_node), 0,
1603 SLAB_PANIC | SLAB_RECLAIM_ACCOUNT,
1604 radix_tree_node_ctor);
1605 ret = cpuhp_setup_state_nocalls(CPUHP_RADIX_DEAD, "lib/radix:dead",
1606 NULL, radix_tree_cpu_dead);
1607 WARN_ON(ret < 0);
1608}