]> git.proxmox.com Git - mirror_ubuntu-eoan-kernel.git/blame - lib/radix-tree.c
radix-tree: fix radix_tree_create for sibling entries
[mirror_ubuntu-eoan-kernel.git] / lib / radix-tree.c
CommitLineData
1da177e4
LT
1/*
2 * Copyright (C) 2001 Momchil Velikov
3 * Portions Copyright (C) 2001 Christoph Hellwig
cde53535 4 * Copyright (C) 2005 SGI, Christoph Lameter
7cf9c2c7 5 * Copyright (C) 2006 Nick Piggin
78c1d784 6 * Copyright (C) 2012 Konstantin Khlebnikov
1da177e4
LT
7 *
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public License as
10 * published by the Free Software Foundation; either version 2, or (at
11 * your option) any later version.
12 *
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 */
22
23#include <linux/errno.h>
24#include <linux/init.h>
25#include <linux/kernel.h>
8bc3bcc9 26#include <linux/export.h>
1da177e4
LT
27#include <linux/radix-tree.h>
28#include <linux/percpu.h>
29#include <linux/slab.h>
ce80b067 30#include <linux/kmemleak.h>
1da177e4
LT
31#include <linux/notifier.h>
32#include <linux/cpu.h>
1da177e4
LT
33#include <linux/string.h>
34#include <linux/bitops.h>
7cf9c2c7 35#include <linux/rcupdate.h>
92cf2118 36#include <linux/preempt.h> /* in_interrupt() */
1da177e4
LT
37
38
26fb1589
JM
39/*
40 * The height_to_maxindex array needs to be one deeper than the maximum
41 * path as height 0 holds only 1 entry.
42 */
43static unsigned long height_to_maxindex[RADIX_TREE_MAX_PATH + 1] __read_mostly;
1da177e4
LT
44
45/*
46 * Radix tree node cache.
47 */
e18b890b 48static struct kmem_cache *radix_tree_node_cachep;
1da177e4 49
55368052
NP
50/*
51 * The radix tree is variable-height, so an insert operation not only has
52 * to build the branch to its corresponding item, it also has to build the
53 * branch to existing items if the size has to be increased (by
54 * radix_tree_extend).
55 *
56 * The worst case is a zero height tree with just a single item at index 0,
57 * and then inserting an item at index ULONG_MAX. This requires 2 new branches
58 * of RADIX_TREE_MAX_PATH size to be created, with only the root node shared.
59 * Hence:
60 */
61#define RADIX_TREE_PRELOAD_SIZE (RADIX_TREE_MAX_PATH * 2 - 1)
62
1da177e4
LT
63/*
64 * Per-cpu pool of preloaded nodes
65 */
66struct radix_tree_preload {
67 int nr;
9d2a8da0
KS
68 /* nodes->private_data points to next preallocated node */
69 struct radix_tree_node *nodes;
1da177e4 70};
8cef7d57 71static DEFINE_PER_CPU(struct radix_tree_preload, radix_tree_preloads) = { 0, };
1da177e4 72
27d20fdd
NP
73static inline void *ptr_to_indirect(void *ptr)
74{
75 return (void *)((unsigned long)ptr | RADIX_TREE_INDIRECT_PTR);
76}
77
afe0e395
MW
78#define RADIX_TREE_RETRY ptr_to_indirect(NULL)
79
db050f29
MW
80#ifdef CONFIG_RADIX_TREE_MULTIORDER
81/* Sibling slots point directly to another slot in the same node */
82static inline bool is_sibling_entry(struct radix_tree_node *parent, void *node)
83{
84 void **ptr = node;
85 return (parent->slots <= ptr) &&
86 (ptr < parent->slots + RADIX_TREE_MAP_SIZE);
87}
88#else
89static inline bool is_sibling_entry(struct radix_tree_node *parent, void *node)
90{
91 return false;
92}
93#endif
94
95static inline unsigned long get_slot_offset(struct radix_tree_node *parent,
96 void **slot)
97{
98 return slot - parent->slots;
99}
100
101static unsigned radix_tree_descend(struct radix_tree_node *parent,
102 struct radix_tree_node **nodep, unsigned offset)
103{
104 void **entry = rcu_dereference_raw(parent->slots[offset]);
105
106#ifdef CONFIG_RADIX_TREE_MULTIORDER
107 if (radix_tree_is_indirect_ptr(entry)) {
108 unsigned long siboff = get_slot_offset(parent, entry);
109 if (siboff < RADIX_TREE_MAP_SIZE) {
110 offset = siboff;
111 entry = rcu_dereference_raw(parent->slots[offset]);
112 }
113 }
114#endif
115
116 *nodep = (void *)entry;
117 return offset;
118}
119
612d6c19
NP
120static inline gfp_t root_gfp_mask(struct radix_tree_root *root)
121{
122 return root->gfp_mask & __GFP_BITS_MASK;
123}
124
643b52b9
NP
125static inline void tag_set(struct radix_tree_node *node, unsigned int tag,
126 int offset)
127{
128 __set_bit(offset, node->tags[tag]);
129}
130
131static inline void tag_clear(struct radix_tree_node *node, unsigned int tag,
132 int offset)
133{
134 __clear_bit(offset, node->tags[tag]);
135}
136
137static inline int tag_get(struct radix_tree_node *node, unsigned int tag,
138 int offset)
139{
140 return test_bit(offset, node->tags[tag]);
141}
142
143static inline void root_tag_set(struct radix_tree_root *root, unsigned int tag)
144{
145 root->gfp_mask |= (__force gfp_t)(1 << (tag + __GFP_BITS_SHIFT));
146}
147
148static inline void root_tag_clear(struct radix_tree_root *root, unsigned int tag)
149{
150 root->gfp_mask &= (__force gfp_t)~(1 << (tag + __GFP_BITS_SHIFT));
151}
152
153static inline void root_tag_clear_all(struct radix_tree_root *root)
154{
155 root->gfp_mask &= __GFP_BITS_MASK;
156}
157
158static inline int root_tag_get(struct radix_tree_root *root, unsigned int tag)
159{
160 return (__force unsigned)root->gfp_mask & (1 << (tag + __GFP_BITS_SHIFT));
161}
162
7b60e9ad
MW
163static inline unsigned root_tags_get(struct radix_tree_root *root)
164{
165 return (__force unsigned)root->gfp_mask >> __GFP_BITS_SHIFT;
166}
167
643b52b9
NP
168/*
169 * Returns 1 if any slot in the node has this tag set.
170 * Otherwise returns 0.
171 */
172static inline int any_tag_set(struct radix_tree_node *node, unsigned int tag)
173{
174 int idx;
175 for (idx = 0; idx < RADIX_TREE_TAG_LONGS; idx++) {
176 if (node->tags[tag][idx])
177 return 1;
178 }
179 return 0;
180}
78c1d784
KK
181
182/**
183 * radix_tree_find_next_bit - find the next set bit in a memory region
184 *
185 * @addr: The address to base the search on
186 * @size: The bitmap size in bits
187 * @offset: The bitnumber to start searching at
188 *
189 * Unrollable variant of find_next_bit() for constant size arrays.
190 * Tail bits starting from size to roundup(size, BITS_PER_LONG) must be zero.
191 * Returns next bit offset, or size if nothing found.
192 */
193static __always_inline unsigned long
194radix_tree_find_next_bit(const unsigned long *addr,
195 unsigned long size, unsigned long offset)
196{
197 if (!__builtin_constant_p(size))
198 return find_next_bit(addr, size, offset);
199
200 if (offset < size) {
201 unsigned long tmp;
202
203 addr += offset / BITS_PER_LONG;
204 tmp = *addr >> (offset % BITS_PER_LONG);
205 if (tmp)
206 return __ffs(tmp) + offset;
207 offset = (offset + BITS_PER_LONG) & ~(BITS_PER_LONG - 1);
208 while (offset < size) {
209 tmp = *++addr;
210 if (tmp)
211 return __ffs(tmp) + offset;
212 offset += BITS_PER_LONG;
213 }
214 }
215 return size;
216}
217
7cf19af4
MW
218#if 0
219static void dump_node(void *slot, int height, int offset)
220{
221 struct radix_tree_node *node;
222 int i;
223
224 if (!slot)
225 return;
226
227 if (height == 0) {
228 pr_debug("radix entry %p offset %d\n", slot, offset);
229 return;
230 }
231
232 node = indirect_to_ptr(slot);
233 pr_debug("radix node: %p offset %d tags %lx %lx %lx path %x count %d parent %p\n",
234 slot, offset, node->tags[0][0], node->tags[1][0],
235 node->tags[2][0], node->path, node->count, node->parent);
236
237 for (i = 0; i < RADIX_TREE_MAP_SIZE; i++)
238 dump_node(node->slots[i], height - 1, i);
239}
240
241/* For debug */
242static void radix_tree_dump(struct radix_tree_root *root)
243{
244 pr_debug("radix root: %p height %d rnode %p tags %x\n",
245 root, root->height, root->rnode,
246 root->gfp_mask >> __GFP_BITS_SHIFT);
247 if (!radix_tree_is_indirect_ptr(root->rnode))
248 return;
249 dump_node(root->rnode, root->height, 0);
250}
251#endif
252
1da177e4
LT
253/*
254 * This assumes that the caller has performed appropriate preallocation, and
255 * that the caller has pinned this thread of control to the current CPU.
256 */
257static struct radix_tree_node *
258radix_tree_node_alloc(struct radix_tree_root *root)
259{
e2848a0e 260 struct radix_tree_node *ret = NULL;
612d6c19 261 gfp_t gfp_mask = root_gfp_mask(root);
1da177e4 262
5e4c0d97
JK
263 /*
264 * Preload code isn't irq safe and it doesn't make sence to use
265 * preloading in the interrupt anyway as all the allocations have to
266 * be atomic. So just do normal allocation when in interrupt.
267 */
d0164adc 268 if (!gfpflags_allow_blocking(gfp_mask) && !in_interrupt()) {
1da177e4
LT
269 struct radix_tree_preload *rtp;
270
58e698af
VD
271 /*
272 * Even if the caller has preloaded, try to allocate from the
273 * cache first for the new node to get accounted.
274 */
275 ret = kmem_cache_alloc(radix_tree_node_cachep,
276 gfp_mask | __GFP_ACCOUNT | __GFP_NOWARN);
277 if (ret)
278 goto out;
279
e2848a0e
NP
280 /*
281 * Provided the caller has preloaded here, we will always
282 * succeed in getting a node here (and never reach
283 * kmem_cache_alloc)
284 */
7c8e0181 285 rtp = this_cpu_ptr(&radix_tree_preloads);
1da177e4 286 if (rtp->nr) {
9d2a8da0
KS
287 ret = rtp->nodes;
288 rtp->nodes = ret->private_data;
289 ret->private_data = NULL;
1da177e4
LT
290 rtp->nr--;
291 }
ce80b067
CM
292 /*
293 * Update the allocation stack trace as this is more useful
294 * for debugging.
295 */
296 kmemleak_update_trace(ret);
58e698af 297 goto out;
1da177e4 298 }
58e698af
VD
299 ret = kmem_cache_alloc(radix_tree_node_cachep,
300 gfp_mask | __GFP_ACCOUNT);
301out:
c0bc9875 302 BUG_ON(radix_tree_is_indirect_ptr(ret));
1da177e4
LT
303 return ret;
304}
305
7cf9c2c7
NP
306static void radix_tree_node_rcu_free(struct rcu_head *head)
307{
308 struct radix_tree_node *node =
309 container_of(head, struct radix_tree_node, rcu_head);
b6dd0865 310 int i;
643b52b9
NP
311
312 /*
313 * must only free zeroed nodes into the slab. radix_tree_shrink
314 * can leave us with a non-NULL entry in the first slot, so clear
315 * that here to make sure.
316 */
b6dd0865
DC
317 for (i = 0; i < RADIX_TREE_MAX_TAGS; i++)
318 tag_clear(node, i, 0);
319
643b52b9
NP
320 node->slots[0] = NULL;
321 node->count = 0;
322
7cf9c2c7
NP
323 kmem_cache_free(radix_tree_node_cachep, node);
324}
325
1da177e4
LT
326static inline void
327radix_tree_node_free(struct radix_tree_node *node)
328{
7cf9c2c7 329 call_rcu(&node->rcu_head, radix_tree_node_rcu_free);
1da177e4
LT
330}
331
332/*
333 * Load up this CPU's radix_tree_node buffer with sufficient objects to
334 * ensure that the addition of a single element in the tree cannot fail. On
335 * success, return zero, with preemption disabled. On error, return -ENOMEM
336 * with preemption not disabled.
b34df792
DH
337 *
338 * To make use of this facility, the radix tree must be initialised without
d0164adc 339 * __GFP_DIRECT_RECLAIM being passed to INIT_RADIX_TREE().
1da177e4 340 */
5e4c0d97 341static int __radix_tree_preload(gfp_t gfp_mask)
1da177e4
LT
342{
343 struct radix_tree_preload *rtp;
344 struct radix_tree_node *node;
345 int ret = -ENOMEM;
346
347 preempt_disable();
7c8e0181 348 rtp = this_cpu_ptr(&radix_tree_preloads);
9d2a8da0 349 while (rtp->nr < RADIX_TREE_PRELOAD_SIZE) {
1da177e4 350 preempt_enable();
488514d1 351 node = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
1da177e4
LT
352 if (node == NULL)
353 goto out;
354 preempt_disable();
7c8e0181 355 rtp = this_cpu_ptr(&radix_tree_preloads);
9d2a8da0
KS
356 if (rtp->nr < RADIX_TREE_PRELOAD_SIZE) {
357 node->private_data = rtp->nodes;
358 rtp->nodes = node;
359 rtp->nr++;
360 } else {
1da177e4 361 kmem_cache_free(radix_tree_node_cachep, node);
9d2a8da0 362 }
1da177e4
LT
363 }
364 ret = 0;
365out:
366 return ret;
367}
5e4c0d97
JK
368
369/*
370 * Load up this CPU's radix_tree_node buffer with sufficient objects to
371 * ensure that the addition of a single element in the tree cannot fail. On
372 * success, return zero, with preemption disabled. On error, return -ENOMEM
373 * with preemption not disabled.
374 *
375 * To make use of this facility, the radix tree must be initialised without
d0164adc 376 * __GFP_DIRECT_RECLAIM being passed to INIT_RADIX_TREE().
5e4c0d97
JK
377 */
378int radix_tree_preload(gfp_t gfp_mask)
379{
380 /* Warn on non-sensical use... */
d0164adc 381 WARN_ON_ONCE(!gfpflags_allow_blocking(gfp_mask));
5e4c0d97
JK
382 return __radix_tree_preload(gfp_mask);
383}
d7f0923d 384EXPORT_SYMBOL(radix_tree_preload);
1da177e4 385
5e4c0d97
JK
386/*
387 * The same as above function, except we don't guarantee preloading happens.
388 * We do it, if we decide it helps. On success, return zero with preemption
389 * disabled. On error, return -ENOMEM with preemption not disabled.
390 */
391int radix_tree_maybe_preload(gfp_t gfp_mask)
392{
d0164adc 393 if (gfpflags_allow_blocking(gfp_mask))
5e4c0d97
JK
394 return __radix_tree_preload(gfp_mask);
395 /* Preloading doesn't help anything with this gfp mask, skip it */
396 preempt_disable();
397 return 0;
398}
399EXPORT_SYMBOL(radix_tree_maybe_preload);
400
1da177e4
LT
401/*
402 * Return the maximum key which can be store into a
403 * radix tree with height HEIGHT.
404 */
405static inline unsigned long radix_tree_maxindex(unsigned int height)
406{
407 return height_to_maxindex[height];
408}
409
1456a439
MW
410static inline unsigned long node_maxindex(struct radix_tree_node *node)
411{
412 return radix_tree_maxindex(node->path & RADIX_TREE_HEIGHT_MASK);
413}
414
415static unsigned radix_tree_load_root(struct radix_tree_root *root,
416 struct radix_tree_node **nodep, unsigned long *maxindex)
417{
418 struct radix_tree_node *node = rcu_dereference_raw(root->rnode);
419
420 *nodep = node;
421
422 if (likely(radix_tree_is_indirect_ptr(node))) {
423 node = indirect_to_ptr(node);
424 *maxindex = node_maxindex(node);
425 return (node->path & RADIX_TREE_HEIGHT_MASK) *
426 RADIX_TREE_MAP_SHIFT;
427 }
428
429 *maxindex = 0;
430 return 0;
431}
432
1da177e4
LT
433/*
434 * Extend a radix tree so it can store key @index.
435 */
e6145236 436static int radix_tree_extend(struct radix_tree_root *root,
49ea6ebc 437 unsigned long index)
1da177e4
LT
438{
439 struct radix_tree_node *node;
e2bdb933 440 struct radix_tree_node *slot;
1da177e4 441 unsigned int height;
1da177e4
LT
442 int tag;
443
444 /* Figure out what the height should be. */
445 height = root->height + 1;
446 while (index > radix_tree_maxindex(height))
447 height++;
448
49ea6ebc 449 if (root->rnode == NULL) {
1da177e4
LT
450 root->height = height;
451 goto out;
452 }
453
1da177e4 454 do {
7cf9c2c7 455 unsigned int newheight;
1da177e4
LT
456 if (!(node = radix_tree_node_alloc(root)))
457 return -ENOMEM;
458
1da177e4 459 /* Propagate the aggregated tag info into the new root */
daff89f3 460 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
612d6c19 461 if (root_tag_get(root, tag))
1da177e4
LT
462 tag_set(node, tag, 0);
463 }
464
e2bdb933 465 /* Increase the height. */
7cf9c2c7 466 newheight = root->height+1;
449dd698
JW
467 BUG_ON(newheight & ~RADIX_TREE_HEIGHT_MASK);
468 node->path = newheight;
1da177e4 469 node->count = 1;
e2bdb933
HD
470 node->parent = NULL;
471 slot = root->rnode;
49ea6ebc 472 if (radix_tree_is_indirect_ptr(slot)) {
e2bdb933
HD
473 slot = indirect_to_ptr(slot);
474 slot->parent = node;
339e6353 475 slot = ptr_to_indirect(slot);
e2bdb933
HD
476 }
477 node->slots[0] = slot;
27d20fdd 478 node = ptr_to_indirect(node);
7cf9c2c7
NP
479 rcu_assign_pointer(root->rnode, node);
480 root->height = newheight;
1da177e4
LT
481 } while (height > root->height);
482out:
49ea6ebc 483 return height * RADIX_TREE_MAP_SHIFT;
1da177e4
LT
484}
485
486/**
139e5616 487 * __radix_tree_create - create a slot in a radix tree
1da177e4
LT
488 * @root: radix tree root
489 * @index: index key
e6145236 490 * @order: index occupies 2^order aligned slots
139e5616
JW
491 * @nodep: returns node
492 * @slotp: returns slot
1da177e4 493 *
139e5616
JW
494 * Create, if necessary, and return the node and slot for an item
495 * at position @index in the radix tree @root.
496 *
497 * Until there is more than one item in the tree, no nodes are
498 * allocated and @root->rnode is used as a direct slot instead of
499 * pointing to a node, in which case *@nodep will be NULL.
500 *
501 * Returns -ENOMEM, or 0 for success.
1da177e4 502 */
139e5616 503int __radix_tree_create(struct radix_tree_root *root, unsigned long index,
e6145236
MW
504 unsigned order, struct radix_tree_node **nodep,
505 void ***slotp)
1da177e4 506{
201b6264 507 struct radix_tree_node *node = NULL, *slot;
49ea6ebc 508 unsigned long maxindex;
139e5616 509 unsigned int height, shift, offset;
49ea6ebc
MW
510 unsigned long max = index | ((1UL << order) - 1);
511
512 shift = radix_tree_load_root(root, &slot, &maxindex);
1da177e4
LT
513
514 /* Make sure the tree is high enough. */
49ea6ebc
MW
515 if (max > maxindex) {
516 int error = radix_tree_extend(root, max);
517 if (error < 0)
1da177e4 518 return error;
49ea6ebc
MW
519 shift = error;
520 slot = root->rnode;
521 if (order == shift) {
522 shift += RADIX_TREE_MAP_SHIFT;
523 root->height++;
524 }
1da177e4
LT
525 }
526
1da177e4 527 height = root->height;
1da177e4
LT
528
529 offset = 0; /* uninitialised var warning */
e6145236 530 while (shift > order) {
201b6264 531 if (slot == NULL) {
1da177e4 532 /* Have to add a child node. */
201b6264 533 if (!(slot = radix_tree_node_alloc(root)))
1da177e4 534 return -ENOMEM;
449dd698 535 slot->path = height;
e2bdb933 536 slot->parent = node;
201b6264 537 if (node) {
339e6353
MW
538 rcu_assign_pointer(node->slots[offset],
539 ptr_to_indirect(slot));
1da177e4 540 node->count++;
449dd698 541 slot->path |= offset << RADIX_TREE_HEIGHT_SHIFT;
201b6264 542 } else
339e6353
MW
543 rcu_assign_pointer(root->rnode,
544 ptr_to_indirect(slot));
e6145236
MW
545 } else if (!radix_tree_is_indirect_ptr(slot))
546 break;
1da177e4
LT
547
548 /* Go a level down */
e6145236 549 height--;
0070e28d 550 shift -= RADIX_TREE_MAP_SHIFT;
e6145236 551 node = indirect_to_ptr(slot);
8a14f4d8
MW
552 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
553 offset = radix_tree_descend(node, &slot, offset);
e6145236
MW
554 }
555
57578c2e 556#ifdef CONFIG_RADIX_TREE_MULTIORDER
e6145236 557 /* Insert pointers to the canonical entry */
3b8c00f6
MW
558 if (order > shift) {
559 int i, n = 1 << (order - shift);
e6145236
MW
560 offset = offset & ~(n - 1);
561 slot = ptr_to_indirect(&node->slots[offset]);
562 for (i = 0; i < n; i++) {
563 if (node->slots[offset + i])
564 return -EEXIST;
565 }
566
567 for (i = 1; i < n; i++) {
568 rcu_assign_pointer(node->slots[offset + i], slot);
569 node->count++;
570 }
612d6c19 571 }
57578c2e 572#endif
1da177e4 573
139e5616
JW
574 if (nodep)
575 *nodep = node;
576 if (slotp)
577 *slotp = node ? node->slots + offset : (void **)&root->rnode;
578 return 0;
579}
580
581/**
e6145236 582 * __radix_tree_insert - insert into a radix tree
139e5616
JW
583 * @root: radix tree root
584 * @index: index key
e6145236 585 * @order: key covers the 2^order indices around index
139e5616
JW
586 * @item: item to insert
587 *
588 * Insert an item into the radix tree at position @index.
589 */
e6145236
MW
590int __radix_tree_insert(struct radix_tree_root *root, unsigned long index,
591 unsigned order, void *item)
139e5616
JW
592{
593 struct radix_tree_node *node;
594 void **slot;
595 int error;
596
597 BUG_ON(radix_tree_is_indirect_ptr(item));
598
e6145236 599 error = __radix_tree_create(root, index, order, &node, &slot);
139e5616
JW
600 if (error)
601 return error;
602 if (*slot != NULL)
1da177e4 603 return -EEXIST;
139e5616 604 rcu_assign_pointer(*slot, item);
201b6264 605
612d6c19 606 if (node) {
7b60e9ad 607 unsigned offset = get_slot_offset(node, slot);
612d6c19 608 node->count++;
7b60e9ad
MW
609 BUG_ON(tag_get(node, 0, offset));
610 BUG_ON(tag_get(node, 1, offset));
611 BUG_ON(tag_get(node, 2, offset));
612d6c19 612 } else {
7b60e9ad 613 BUG_ON(root_tags_get(root));
612d6c19 614 }
1da177e4 615
1da177e4
LT
616 return 0;
617}
e6145236 618EXPORT_SYMBOL(__radix_tree_insert);
1da177e4 619
139e5616
JW
620/**
621 * __radix_tree_lookup - lookup an item in a radix tree
622 * @root: radix tree root
623 * @index: index key
624 * @nodep: returns node
625 * @slotp: returns slot
626 *
627 * Lookup and return the item at position @index in the radix
628 * tree @root.
629 *
630 * Until there is more than one item in the tree, no nodes are
631 * allocated and @root->rnode is used as a direct slot instead of
632 * pointing to a node, in which case *@nodep will be NULL.
7cf9c2c7 633 */
139e5616
JW
634void *__radix_tree_lookup(struct radix_tree_root *root, unsigned long index,
635 struct radix_tree_node **nodep, void ***slotp)
1da177e4 636{
139e5616 637 struct radix_tree_node *node, *parent;
85829954
MW
638 unsigned long maxindex;
639 unsigned int shift;
139e5616 640 void **slot;
612d6c19 641
85829954
MW
642 restart:
643 parent = NULL;
644 slot = (void **)&root->rnode;
645 shift = radix_tree_load_root(root, &node, &maxindex);
646 if (index > maxindex)
1da177e4
LT
647 return NULL;
648
85829954
MW
649 while (radix_tree_is_indirect_ptr(node)) {
650 unsigned offset;
1da177e4 651
85829954
MW
652 if (node == RADIX_TREE_RETRY)
653 goto restart;
654 parent = indirect_to_ptr(node);
1da177e4 655 shift -= RADIX_TREE_MAP_SHIFT;
85829954
MW
656 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
657 offset = radix_tree_descend(parent, &node, offset);
658 slot = parent->slots + offset;
659 }
1da177e4 660
139e5616
JW
661 if (nodep)
662 *nodep = parent;
663 if (slotp)
664 *slotp = slot;
665 return node;
b72b71c6
HS
666}
667
668/**
669 * radix_tree_lookup_slot - lookup a slot in a radix tree
670 * @root: radix tree root
671 * @index: index key
672 *
673 * Returns: the slot corresponding to the position @index in the
674 * radix tree @root. This is useful for update-if-exists operations.
675 *
676 * This function can be called under rcu_read_lock iff the slot is not
677 * modified by radix_tree_replace_slot, otherwise it must be called
678 * exclusive from other writers. Any dereference of the slot must be done
679 * using radix_tree_deref_slot.
680 */
681void **radix_tree_lookup_slot(struct radix_tree_root *root, unsigned long index)
682{
139e5616
JW
683 void **slot;
684
685 if (!__radix_tree_lookup(root, index, NULL, &slot))
686 return NULL;
687 return slot;
a4331366 688}
a4331366
HR
689EXPORT_SYMBOL(radix_tree_lookup_slot);
690
691/**
692 * radix_tree_lookup - perform lookup operation on a radix tree
693 * @root: radix tree root
694 * @index: index key
695 *
696 * Lookup the item at the position @index in the radix tree @root.
7cf9c2c7
NP
697 *
698 * This function can be called under rcu_read_lock, however the caller
699 * must manage lifetimes of leaf nodes (eg. RCU may also be used to free
700 * them safely). No RCU barriers are required to access or modify the
701 * returned item, however.
a4331366
HR
702 */
703void *radix_tree_lookup(struct radix_tree_root *root, unsigned long index)
704{
139e5616 705 return __radix_tree_lookup(root, index, NULL, NULL);
1da177e4
LT
706}
707EXPORT_SYMBOL(radix_tree_lookup);
708
709/**
710 * radix_tree_tag_set - set a tag on a radix tree node
711 * @root: radix tree root
712 * @index: index key
713 * @tag: tag index
714 *
daff89f3
JC
715 * Set the search tag (which must be < RADIX_TREE_MAX_TAGS)
716 * corresponding to @index in the radix tree. From
1da177e4
LT
717 * the root all the way down to the leaf node.
718 *
719 * Returns the address of the tagged item. Setting a tag on a not-present
720 * item is a bug.
721 */
722void *radix_tree_tag_set(struct radix_tree_root *root,
daff89f3 723 unsigned long index, unsigned int tag)
1da177e4 724{
fb969909
RZ
725 struct radix_tree_node *node, *parent;
726 unsigned long maxindex;
727 unsigned int shift;
1da177e4 728
fb969909
RZ
729 shift = radix_tree_load_root(root, &node, &maxindex);
730 BUG_ON(index > maxindex);
1da177e4 731
fb969909
RZ
732 while (radix_tree_is_indirect_ptr(node)) {
733 unsigned offset;
1da177e4 734
1da177e4 735 shift -= RADIX_TREE_MAP_SHIFT;
fb969909
RZ
736 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
737
738 parent = indirect_to_ptr(node);
739 offset = radix_tree_descend(parent, &node, offset);
740 BUG_ON(!node);
741
742 if (!tag_get(parent, tag, offset))
743 tag_set(parent, tag, offset);
1da177e4
LT
744 }
745
612d6c19 746 /* set the root's tag bit */
fb969909 747 if (!root_tag_get(root, tag))
612d6c19
NP
748 root_tag_set(root, tag);
749
fb969909 750 return node;
1da177e4
LT
751}
752EXPORT_SYMBOL(radix_tree_tag_set);
753
754/**
755 * radix_tree_tag_clear - clear a tag on a radix tree node
756 * @root: radix tree root
757 * @index: index key
758 * @tag: tag index
759 *
daff89f3
JC
760 * Clear the search tag (which must be < RADIX_TREE_MAX_TAGS)
761 * corresponding to @index in the radix tree. If
1da177e4
LT
762 * this causes the leaf node to have no tags set then clear the tag in the
763 * next-to-leaf node, etc.
764 *
765 * Returns the address of the tagged item on success, else NULL. ie:
766 * has the same return value and semantics as radix_tree_lookup().
767 */
768void *radix_tree_tag_clear(struct radix_tree_root *root,
daff89f3 769 unsigned long index, unsigned int tag)
1da177e4 770{
00f47b58
RZ
771 struct radix_tree_node *node, *parent;
772 unsigned long maxindex;
773 unsigned int shift;
e2bdb933 774 int uninitialized_var(offset);
1da177e4 775
00f47b58
RZ
776 shift = radix_tree_load_root(root, &node, &maxindex);
777 if (index > maxindex)
778 return NULL;
1da177e4 779
00f47b58 780 parent = NULL;
1da177e4 781
00f47b58 782 while (radix_tree_is_indirect_ptr(node)) {
e2bdb933 783 shift -= RADIX_TREE_MAP_SHIFT;
1da177e4 784 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
00f47b58
RZ
785
786 parent = indirect_to_ptr(node);
787 offset = radix_tree_descend(parent, &node, offset);
1da177e4
LT
788 }
789
00f47b58 790 if (node == NULL)
1da177e4
LT
791 goto out;
792
00f47b58
RZ
793 index >>= shift;
794
795 while (parent) {
796 if (!tag_get(parent, tag, offset))
d5274261 797 goto out;
00f47b58
RZ
798 tag_clear(parent, tag, offset);
799 if (any_tag_set(parent, tag))
6e954b9e 800 goto out;
e2bdb933
HD
801
802 index >>= RADIX_TREE_MAP_SHIFT;
803 offset = index & RADIX_TREE_MAP_MASK;
00f47b58 804 parent = parent->parent;
612d6c19
NP
805 }
806
807 /* clear the root's tag bit */
808 if (root_tag_get(root, tag))
809 root_tag_clear(root, tag);
810
1da177e4 811out:
00f47b58 812 return node;
1da177e4
LT
813}
814EXPORT_SYMBOL(radix_tree_tag_clear);
815
1da177e4 816/**
32605a18
MT
817 * radix_tree_tag_get - get a tag on a radix tree node
818 * @root: radix tree root
819 * @index: index key
daff89f3 820 * @tag: tag index (< RADIX_TREE_MAX_TAGS)
1da177e4 821 *
32605a18 822 * Return values:
1da177e4 823 *
612d6c19
NP
824 * 0: tag not present or not set
825 * 1: tag set
ce82653d
DH
826 *
827 * Note that the return value of this function may not be relied on, even if
828 * the RCU lock is held, unless tag modification and node deletion are excluded
829 * from concurrency.
1da177e4
LT
830 */
831int radix_tree_tag_get(struct radix_tree_root *root,
daff89f3 832 unsigned long index, unsigned int tag)
1da177e4 833{
4589ba6d
RZ
834 struct radix_tree_node *node, *parent;
835 unsigned long maxindex;
836 unsigned int shift;
1da177e4 837
612d6c19
NP
838 if (!root_tag_get(root, tag))
839 return 0;
840
4589ba6d
RZ
841 shift = radix_tree_load_root(root, &node, &maxindex);
842 if (index > maxindex)
843 return 0;
7cf9c2c7
NP
844 if (node == NULL)
845 return 0;
846
4589ba6d
RZ
847 while (radix_tree_is_indirect_ptr(node)) {
848 int offset;
612d6c19 849
4589ba6d
RZ
850 shift -= RADIX_TREE_MAP_SHIFT;
851 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
1da177e4 852
4589ba6d
RZ
853 parent = indirect_to_ptr(node);
854 offset = radix_tree_descend(parent, &node, offset);
1da177e4 855
4589ba6d 856 if (!node)
1da177e4 857 return 0;
4589ba6d 858 if (!tag_get(parent, tag, offset))
3fa36acb 859 return 0;
4589ba6d
RZ
860 if (node == RADIX_TREE_RETRY)
861 break;
1da177e4 862 }
4589ba6d
RZ
863
864 return 1;
1da177e4
LT
865}
866EXPORT_SYMBOL(radix_tree_tag_get);
1da177e4 867
21ef5339
RZ
868static inline void __set_iter_shift(struct radix_tree_iter *iter,
869 unsigned int shift)
870{
871#ifdef CONFIG_RADIX_TREE_MULTIORDER
872 iter->shift = shift;
873#endif
874}
875
78c1d784
KK
876/**
877 * radix_tree_next_chunk - find next chunk of slots for iteration
878 *
879 * @root: radix tree root
880 * @iter: iterator state
881 * @flags: RADIX_TREE_ITER_* flags and tag index
882 * Returns: pointer to chunk first slot, or NULL if iteration is over
883 */
884void **radix_tree_next_chunk(struct radix_tree_root *root,
885 struct radix_tree_iter *iter, unsigned flags)
886{
887 unsigned shift, tag = flags & RADIX_TREE_ITER_TAG_MASK;
888 struct radix_tree_node *rnode, *node;
21ef5339 889 unsigned long index, offset, maxindex;
78c1d784
KK
890
891 if ((flags & RADIX_TREE_ITER_TAGGED) && !root_tag_get(root, tag))
892 return NULL;
893
894 /*
895 * Catch next_index overflow after ~0UL. iter->index never overflows
896 * during iterating; it can be zero only at the beginning.
897 * And we cannot overflow iter->next_index in a single step,
898 * because RADIX_TREE_MAP_SHIFT < BITS_PER_LONG.
fffaee36
KK
899 *
900 * This condition also used by radix_tree_next_slot() to stop
901 * contiguous iterating, and forbid swithing to the next chunk.
78c1d784
KK
902 */
903 index = iter->next_index;
904 if (!index && iter->index)
905 return NULL;
906
21ef5339
RZ
907 restart:
908 shift = radix_tree_load_root(root, &rnode, &maxindex);
909 if (index > maxindex)
910 return NULL;
911
78c1d784
KK
912 if (radix_tree_is_indirect_ptr(rnode)) {
913 rnode = indirect_to_ptr(rnode);
21ef5339 914 } else if (rnode) {
78c1d784 915 /* Single-slot tree */
21ef5339
RZ
916 iter->index = index;
917 iter->next_index = maxindex + 1;
78c1d784 918 iter->tags = 1;
21ef5339 919 __set_iter_shift(iter, shift);
78c1d784
KK
920 return (void **)&root->rnode;
921 } else
922 return NULL;
923
21ef5339 924 shift -= RADIX_TREE_MAP_SHIFT;
78c1d784
KK
925 offset = index >> shift;
926
78c1d784
KK
927 node = rnode;
928 while (1) {
e6145236 929 struct radix_tree_node *slot;
21ef5339
RZ
930 unsigned new_off = radix_tree_descend(node, &slot, offset);
931
932 if (new_off < offset) {
933 offset = new_off;
934 index &= ~((RADIX_TREE_MAP_SIZE << shift) - 1);
935 index |= offset << shift;
936 }
937
78c1d784 938 if ((flags & RADIX_TREE_ITER_TAGGED) ?
21ef5339 939 !tag_get(node, tag, offset) : !slot) {
78c1d784
KK
940 /* Hole detected */
941 if (flags & RADIX_TREE_ITER_CONTIG)
942 return NULL;
943
944 if (flags & RADIX_TREE_ITER_TAGGED)
945 offset = radix_tree_find_next_bit(
946 node->tags[tag],
947 RADIX_TREE_MAP_SIZE,
948 offset + 1);
949 else
950 while (++offset < RADIX_TREE_MAP_SIZE) {
21ef5339
RZ
951 void *slot = node->slots[offset];
952 if (is_sibling_entry(node, slot))
953 continue;
954 if (slot)
78c1d784
KK
955 break;
956 }
957 index &= ~((RADIX_TREE_MAP_SIZE << shift) - 1);
958 index += offset << shift;
959 /* Overflow after ~0UL */
960 if (!index)
961 return NULL;
962 if (offset == RADIX_TREE_MAP_SIZE)
963 goto restart;
21ef5339 964 slot = rcu_dereference_raw(node->slots[offset]);
78c1d784
KK
965 }
966
21ef5339 967 if ((slot == NULL) || (slot == RADIX_TREE_RETRY))
78c1d784 968 goto restart;
e6145236
MW
969 if (!radix_tree_is_indirect_ptr(slot))
970 break;
21ef5339 971
e6145236 972 node = indirect_to_ptr(slot);
78c1d784
KK
973 shift -= RADIX_TREE_MAP_SHIFT;
974 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
975 }
976
977 /* Update the iterator state */
21ef5339
RZ
978 iter->index = index & ~((1 << shift) - 1);
979 iter->next_index = (index | ((RADIX_TREE_MAP_SIZE << shift) - 1)) + 1;
980 __set_iter_shift(iter, shift);
78c1d784
KK
981
982 /* Construct iter->tags bit-mask from node->tags[tag] array */
983 if (flags & RADIX_TREE_ITER_TAGGED) {
984 unsigned tag_long, tag_bit;
985
986 tag_long = offset / BITS_PER_LONG;
987 tag_bit = offset % BITS_PER_LONG;
988 iter->tags = node->tags[tag][tag_long] >> tag_bit;
989 /* This never happens if RADIX_TREE_TAG_LONGS == 1 */
990 if (tag_long < RADIX_TREE_TAG_LONGS - 1) {
991 /* Pick tags from next element */
992 if (tag_bit)
993 iter->tags |= node->tags[tag][tag_long + 1] <<
994 (BITS_PER_LONG - tag_bit);
995 /* Clip chunk size, here only BITS_PER_LONG tags */
996 iter->next_index = index + BITS_PER_LONG;
997 }
998 }
999
1000 return node->slots + offset;
1001}
1002EXPORT_SYMBOL(radix_tree_next_chunk);
1003
ebf8aa44
JK
1004/**
1005 * radix_tree_range_tag_if_tagged - for each item in given range set given
1006 * tag if item has another tag set
1007 * @root: radix tree root
1008 * @first_indexp: pointer to a starting index of a range to scan
1009 * @last_index: last index of a range to scan
1010 * @nr_to_tag: maximum number items to tag
1011 * @iftag: tag index to test
1012 * @settag: tag index to set if tested tag is set
1013 *
1014 * This function scans range of radix tree from first_index to last_index
1015 * (inclusive). For each item in the range if iftag is set, the function sets
1016 * also settag. The function stops either after tagging nr_to_tag items or
1017 * after reaching last_index.
1018 *
144dcfc0
DC
1019 * The tags must be set from the leaf level only and propagated back up the
1020 * path to the root. We must do this so that we resolve the full path before
1021 * setting any tags on intermediate nodes. If we set tags as we descend, then
1022 * we can get to the leaf node and find that the index that has the iftag
1023 * set is outside the range we are scanning. This reults in dangling tags and
1024 * can lead to problems with later tag operations (e.g. livelocks on lookups).
1025 *
ebf8aa44
JK
1026 * The function returns number of leaves where the tag was set and sets
1027 * *first_indexp to the first unscanned index.
d5ed3a4a
JK
1028 * WARNING! *first_indexp can wrap if last_index is ULONG_MAX. Caller must
1029 * be prepared to handle that.
ebf8aa44
JK
1030 */
1031unsigned long radix_tree_range_tag_if_tagged(struct radix_tree_root *root,
1032 unsigned long *first_indexp, unsigned long last_index,
1033 unsigned long nr_to_tag,
1034 unsigned int iftag, unsigned int settag)
1035{
144dcfc0 1036 unsigned int height = root->height;
e2bdb933 1037 struct radix_tree_node *node = NULL;
144dcfc0
DC
1038 struct radix_tree_node *slot;
1039 unsigned int shift;
1040 unsigned long tagged = 0;
1041 unsigned long index = *first_indexp;
ebf8aa44
JK
1042
1043 last_index = min(last_index, radix_tree_maxindex(height));
1044 if (index > last_index)
1045 return 0;
1046 if (!nr_to_tag)
1047 return 0;
1048 if (!root_tag_get(root, iftag)) {
1049 *first_indexp = last_index + 1;
1050 return 0;
1051 }
1052 if (height == 0) {
1053 *first_indexp = last_index + 1;
1054 root_tag_set(root, settag);
1055 return 1;
1056 }
1057
1058 shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
27d20fdd 1059 slot = indirect_to_ptr(root->rnode);
ebf8aa44
JK
1060
1061 for (;;) {
e2bdb933 1062 unsigned long upindex;
ebf8aa44
JK
1063 int offset;
1064
1065 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
1066 if (!slot->slots[offset])
1067 goto next;
1068 if (!tag_get(slot, iftag, offset))
1069 goto next;
e2bdb933 1070 if (shift) {
e2bdb933 1071 node = slot;
144dcfc0 1072 slot = slot->slots[offset];
e6145236
MW
1073 if (radix_tree_is_indirect_ptr(slot)) {
1074 slot = indirect_to_ptr(slot);
1075 shift -= RADIX_TREE_MAP_SHIFT;
1076 continue;
1077 } else {
1078 slot = node;
1079 node = node->parent;
1080 }
144dcfc0
DC
1081 }
1082
1083 /* tag the leaf */
e6145236 1084 tagged += 1 << shift;
ebf8aa44 1085 tag_set(slot, settag, offset);
144dcfc0
DC
1086
1087 /* walk back up the path tagging interior nodes */
e2bdb933
HD
1088 upindex = index;
1089 while (node) {
1090 upindex >>= RADIX_TREE_MAP_SHIFT;
1091 offset = upindex & RADIX_TREE_MAP_MASK;
1092
144dcfc0 1093 /* stop if we find a node with the tag already set */
e2bdb933 1094 if (tag_get(node, settag, offset))
144dcfc0 1095 break;
e2bdb933
HD
1096 tag_set(node, settag, offset);
1097 node = node->parent;
ebf8aa44 1098 }
144dcfc0 1099
e2bdb933
HD
1100 /*
1101 * Small optimization: now clear that node pointer.
1102 * Since all of this slot's ancestors now have the tag set
1103 * from setting it above, we have no further need to walk
1104 * back up the tree setting tags, until we update slot to
1105 * point to another radix_tree_node.
1106 */
1107 node = NULL;
1108
ebf8aa44
JK
1109next:
1110 /* Go to next item at level determined by 'shift' */
1111 index = ((index >> shift) + 1) << shift;
d5ed3a4a
JK
1112 /* Overflow can happen when last_index is ~0UL... */
1113 if (index > last_index || !index)
ebf8aa44
JK
1114 break;
1115 if (tagged >= nr_to_tag)
1116 break;
1117 while (((index >> shift) & RADIX_TREE_MAP_MASK) == 0) {
1118 /*
1119 * We've fully scanned this node. Go up. Because
1120 * last_index is guaranteed to be in the tree, what
1121 * we do below cannot wander astray.
1122 */
e2bdb933 1123 slot = slot->parent;
ebf8aa44
JK
1124 shift += RADIX_TREE_MAP_SHIFT;
1125 }
1126 }
1127 /*
ac15ee69
TO
1128 * We need not to tag the root tag if there is no tag which is set with
1129 * settag within the range from *first_indexp to last_index.
ebf8aa44 1130 */
ac15ee69
TO
1131 if (tagged > 0)
1132 root_tag_set(root, settag);
ebf8aa44
JK
1133 *first_indexp = index;
1134
1135 return tagged;
1136}
1137EXPORT_SYMBOL(radix_tree_range_tag_if_tagged);
1138
1da177e4
LT
1139/**
1140 * radix_tree_gang_lookup - perform multiple lookup on a radix tree
1141 * @root: radix tree root
1142 * @results: where the results of the lookup are placed
1143 * @first_index: start the lookup from this key
1144 * @max_items: place up to this many items at *results
1145 *
1146 * Performs an index-ascending scan of the tree for present items. Places
1147 * them at *@results and returns the number of items which were placed at
1148 * *@results.
1149 *
1150 * The implementation is naive.
7cf9c2c7
NP
1151 *
1152 * Like radix_tree_lookup, radix_tree_gang_lookup may be called under
1153 * rcu_read_lock. In this case, rather than the returned results being
1154 * an atomic snapshot of the tree at a single point in time, the semantics
1155 * of an RCU protected gang lookup are as though multiple radix_tree_lookups
1156 * have been issued in individual locks, and results stored in 'results'.
1da177e4
LT
1157 */
1158unsigned int
1159radix_tree_gang_lookup(struct radix_tree_root *root, void **results,
1160 unsigned long first_index, unsigned int max_items)
1161{
cebbd29e
KK
1162 struct radix_tree_iter iter;
1163 void **slot;
1164 unsigned int ret = 0;
7cf9c2c7 1165
cebbd29e 1166 if (unlikely(!max_items))
7cf9c2c7 1167 return 0;
1da177e4 1168
cebbd29e 1169 radix_tree_for_each_slot(slot, root, &iter, first_index) {
46437f9a 1170 results[ret] = rcu_dereference_raw(*slot);
cebbd29e
KK
1171 if (!results[ret])
1172 continue;
46437f9a
MW
1173 if (radix_tree_is_indirect_ptr(results[ret])) {
1174 slot = radix_tree_iter_retry(&iter);
1175 continue;
1176 }
cebbd29e 1177 if (++ret == max_items)
1da177e4 1178 break;
1da177e4 1179 }
7cf9c2c7 1180
1da177e4
LT
1181 return ret;
1182}
1183EXPORT_SYMBOL(radix_tree_gang_lookup);
1184
47feff2c
NP
1185/**
1186 * radix_tree_gang_lookup_slot - perform multiple slot lookup on radix tree
1187 * @root: radix tree root
1188 * @results: where the results of the lookup are placed
6328650b 1189 * @indices: where their indices should be placed (but usually NULL)
47feff2c
NP
1190 * @first_index: start the lookup from this key
1191 * @max_items: place up to this many items at *results
1192 *
1193 * Performs an index-ascending scan of the tree for present items. Places
1194 * their slots at *@results and returns the number of items which were
1195 * placed at *@results.
1196 *
1197 * The implementation is naive.
1198 *
1199 * Like radix_tree_gang_lookup as far as RCU and locking goes. Slots must
1200 * be dereferenced with radix_tree_deref_slot, and if using only RCU
1201 * protection, radix_tree_deref_slot may fail requiring a retry.
1202 */
1203unsigned int
6328650b
HD
1204radix_tree_gang_lookup_slot(struct radix_tree_root *root,
1205 void ***results, unsigned long *indices,
47feff2c
NP
1206 unsigned long first_index, unsigned int max_items)
1207{
cebbd29e
KK
1208 struct radix_tree_iter iter;
1209 void **slot;
1210 unsigned int ret = 0;
47feff2c 1211
cebbd29e 1212 if (unlikely(!max_items))
47feff2c
NP
1213 return 0;
1214
cebbd29e
KK
1215 radix_tree_for_each_slot(slot, root, &iter, first_index) {
1216 results[ret] = slot;
6328650b 1217 if (indices)
cebbd29e
KK
1218 indices[ret] = iter.index;
1219 if (++ret == max_items)
47feff2c 1220 break;
47feff2c
NP
1221 }
1222
1223 return ret;
1224}
1225EXPORT_SYMBOL(radix_tree_gang_lookup_slot);
1226
1da177e4
LT
1227/**
1228 * radix_tree_gang_lookup_tag - perform multiple lookup on a radix tree
1229 * based on a tag
1230 * @root: radix tree root
1231 * @results: where the results of the lookup are placed
1232 * @first_index: start the lookup from this key
1233 * @max_items: place up to this many items at *results
daff89f3 1234 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1da177e4
LT
1235 *
1236 * Performs an index-ascending scan of the tree for present items which
1237 * have the tag indexed by @tag set. Places the items at *@results and
1238 * returns the number of items which were placed at *@results.
1239 */
1240unsigned int
1241radix_tree_gang_lookup_tag(struct radix_tree_root *root, void **results,
daff89f3
JC
1242 unsigned long first_index, unsigned int max_items,
1243 unsigned int tag)
1da177e4 1244{
cebbd29e
KK
1245 struct radix_tree_iter iter;
1246 void **slot;
1247 unsigned int ret = 0;
612d6c19 1248
cebbd29e 1249 if (unlikely(!max_items))
7cf9c2c7
NP
1250 return 0;
1251
cebbd29e 1252 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
46437f9a 1253 results[ret] = rcu_dereference_raw(*slot);
cebbd29e
KK
1254 if (!results[ret])
1255 continue;
46437f9a
MW
1256 if (radix_tree_is_indirect_ptr(results[ret])) {
1257 slot = radix_tree_iter_retry(&iter);
1258 continue;
1259 }
cebbd29e 1260 if (++ret == max_items)
1da177e4 1261 break;
1da177e4 1262 }
7cf9c2c7 1263
1da177e4
LT
1264 return ret;
1265}
1266EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
1267
47feff2c
NP
1268/**
1269 * radix_tree_gang_lookup_tag_slot - perform multiple slot lookup on a
1270 * radix tree based on a tag
1271 * @root: radix tree root
1272 * @results: where the results of the lookup are placed
1273 * @first_index: start the lookup from this key
1274 * @max_items: place up to this many items at *results
1275 * @tag: the tag index (< RADIX_TREE_MAX_TAGS)
1276 *
1277 * Performs an index-ascending scan of the tree for present items which
1278 * have the tag indexed by @tag set. Places the slots at *@results and
1279 * returns the number of slots which were placed at *@results.
1280 */
1281unsigned int
1282radix_tree_gang_lookup_tag_slot(struct radix_tree_root *root, void ***results,
1283 unsigned long first_index, unsigned int max_items,
1284 unsigned int tag)
1285{
cebbd29e
KK
1286 struct radix_tree_iter iter;
1287 void **slot;
1288 unsigned int ret = 0;
47feff2c 1289
cebbd29e 1290 if (unlikely(!max_items))
47feff2c
NP
1291 return 0;
1292
cebbd29e
KK
1293 radix_tree_for_each_tagged(slot, root, &iter, first_index, tag) {
1294 results[ret] = slot;
1295 if (++ret == max_items)
47feff2c 1296 break;
47feff2c
NP
1297 }
1298
1299 return ret;
1300}
1301EXPORT_SYMBOL(radix_tree_gang_lookup_tag_slot);
1302
e504f3fd
HD
1303#if defined(CONFIG_SHMEM) && defined(CONFIG_SWAP)
1304#include <linux/sched.h> /* for cond_resched() */
1305
1306/*
1307 * This linear search is at present only useful to shmem_unuse_inode().
1308 */
1309static unsigned long __locate(struct radix_tree_node *slot, void *item,
1310 unsigned long index, unsigned long *found_index)
1311{
1312 unsigned int shift, height;
1313 unsigned long i;
1314
449dd698 1315 height = slot->path & RADIX_TREE_HEIGHT_MASK;
e504f3fd
HD
1316 shift = (height-1) * RADIX_TREE_MAP_SHIFT;
1317
1318 for ( ; height > 1; height--) {
1319 i = (index >> shift) & RADIX_TREE_MAP_MASK;
1320 for (;;) {
1321 if (slot->slots[i] != NULL)
1322 break;
1323 index &= ~((1UL << shift) - 1);
1324 index += 1UL << shift;
1325 if (index == 0)
1326 goto out; /* 32-bit wraparound */
1327 i++;
1328 if (i == RADIX_TREE_MAP_SIZE)
1329 goto out;
1330 }
1331
e504f3fd
HD
1332 slot = rcu_dereference_raw(slot->slots[i]);
1333 if (slot == NULL)
1334 goto out;
e6145236
MW
1335 if (!radix_tree_is_indirect_ptr(slot)) {
1336 if (slot == item) {
1337 *found_index = index + i;
1338 index = 0;
1339 } else {
1340 index += shift;
1341 }
1342 goto out;
1343 }
339e6353 1344 slot = indirect_to_ptr(slot);
e6145236 1345 shift -= RADIX_TREE_MAP_SHIFT;
e504f3fd
HD
1346 }
1347
1348 /* Bottom level: check items */
1349 for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
1350 if (slot->slots[i] == item) {
1351 *found_index = index + i;
1352 index = 0;
1353 goto out;
1354 }
1355 }
1356 index += RADIX_TREE_MAP_SIZE;
1357out:
1358 return index;
1359}
1360
1361/**
1362 * radix_tree_locate_item - search through radix tree for item
1363 * @root: radix tree root
1364 * @item: item to be found
1365 *
1366 * Returns index where item was found, or -1 if not found.
1367 * Caller must hold no lock (since this time-consuming function needs
1368 * to be preemptible), and must check afterwards if item is still there.
1369 */
1370unsigned long radix_tree_locate_item(struct radix_tree_root *root, void *item)
1371{
1372 struct radix_tree_node *node;
1373 unsigned long max_index;
1374 unsigned long cur_index = 0;
1375 unsigned long found_index = -1;
1376
1377 do {
1378 rcu_read_lock();
1379 node = rcu_dereference_raw(root->rnode);
1380 if (!radix_tree_is_indirect_ptr(node)) {
1381 rcu_read_unlock();
1382 if (node == item)
1383 found_index = 0;
1384 break;
1385 }
1386
1387 node = indirect_to_ptr(node);
449dd698
JW
1388 max_index = radix_tree_maxindex(node->path &
1389 RADIX_TREE_HEIGHT_MASK);
5f30fc94
HD
1390 if (cur_index > max_index) {
1391 rcu_read_unlock();
e504f3fd 1392 break;
5f30fc94 1393 }
e504f3fd
HD
1394
1395 cur_index = __locate(node, item, cur_index, &found_index);
1396 rcu_read_unlock();
1397 cond_resched();
1398 } while (cur_index != 0 && cur_index <= max_index);
1399
1400 return found_index;
1401}
1402#else
1403unsigned long radix_tree_locate_item(struct radix_tree_root *root, void *item)
1404{
1405 return -1;
1406}
1407#endif /* CONFIG_SHMEM && CONFIG_SWAP */
47feff2c 1408
a5f51c96
NP
1409/**
1410 * radix_tree_shrink - shrink height of a radix tree to minimal
1411 * @root radix tree root
1412 */
1413static inline void radix_tree_shrink(struct radix_tree_root *root)
1414{
1415 /* try to shrink tree height */
c0bc9875 1416 while (root->height > 0) {
a5f51c96 1417 struct radix_tree_node *to_free = root->rnode;
e2bdb933 1418 struct radix_tree_node *slot;
a5f51c96 1419
c0bc9875 1420 BUG_ON(!radix_tree_is_indirect_ptr(to_free));
27d20fdd 1421 to_free = indirect_to_ptr(to_free);
c0bc9875
NP
1422
1423 /*
1424 * The candidate node has more than one child, or its child
e6145236
MW
1425 * is not at the leftmost slot, or it is a multiorder entry,
1426 * we cannot shrink.
c0bc9875
NP
1427 */
1428 if (to_free->count != 1)
1429 break;
339e6353
MW
1430 slot = to_free->slots[0];
1431 if (!slot)
c0bc9875 1432 break;
afe0e395
MW
1433 if (!radix_tree_is_indirect_ptr(slot) && (root->height > 1))
1434 break;
1435
1436 if (radix_tree_is_indirect_ptr(slot)) {
1437 slot = indirect_to_ptr(slot);
1438 slot->parent = NULL;
1439 slot = ptr_to_indirect(slot);
1440 }
c0bc9875 1441
7cf9c2c7
NP
1442 /*
1443 * We don't need rcu_assign_pointer(), since we are simply
27d20fdd
NP
1444 * moving the node from one part of the tree to another: if it
1445 * was safe to dereference the old pointer to it
7cf9c2c7 1446 * (to_free->slots[0]), it will be safe to dereference the new
27d20fdd 1447 * one (root->rnode) as far as dependent read barriers go.
7cf9c2c7 1448 */
e2bdb933 1449 root->rnode = slot;
a5f51c96 1450 root->height--;
27d20fdd
NP
1451
1452 /*
1453 * We have a dilemma here. The node's slot[0] must not be
1454 * NULLed in case there are concurrent lookups expecting to
1455 * find the item. However if this was a bottom-level node,
1456 * then it may be subject to the slot pointer being visible
1457 * to callers dereferencing it. If item corresponding to
1458 * slot[0] is subsequently deleted, these callers would expect
1459 * their slot to become empty sooner or later.
1460 *
1461 * For example, lockless pagecache will look up a slot, deref
1462 * the page pointer, and if the page is 0 refcount it means it
1463 * was concurrently deleted from pagecache so try the deref
1464 * again. Fortunately there is already a requirement for logic
1465 * to retry the entire slot lookup -- the indirect pointer
1466 * problem (replacing direct root node with an indirect pointer
1467 * also results in a stale slot). So tag the slot as indirect
1468 * to force callers to retry.
1469 */
afe0e395
MW
1470 if (!radix_tree_is_indirect_ptr(slot))
1471 to_free->slots[0] = RADIX_TREE_RETRY;
27d20fdd 1472
a5f51c96
NP
1473 radix_tree_node_free(to_free);
1474 }
1475}
1476
139e5616
JW
1477/**
1478 * __radix_tree_delete_node - try to free node after clearing a slot
1479 * @root: radix tree root
139e5616
JW
1480 * @node: node containing @index
1481 *
1482 * After clearing the slot at @index in @node from radix tree
1483 * rooted at @root, call this function to attempt freeing the
1484 * node and shrinking the tree.
1485 *
1486 * Returns %true if @node was freed, %false otherwise.
1487 */
449dd698 1488bool __radix_tree_delete_node(struct radix_tree_root *root,
139e5616
JW
1489 struct radix_tree_node *node)
1490{
1491 bool deleted = false;
1492
1493 do {
1494 struct radix_tree_node *parent;
1495
1496 if (node->count) {
1497 if (node == indirect_to_ptr(root->rnode)) {
1498 radix_tree_shrink(root);
1499 if (root->height == 0)
1500 deleted = true;
1501 }
1502 return deleted;
1503 }
1504
1505 parent = node->parent;
1506 if (parent) {
449dd698 1507 unsigned int offset;
139e5616 1508
449dd698
JW
1509 offset = node->path >> RADIX_TREE_HEIGHT_SHIFT;
1510 parent->slots[offset] = NULL;
139e5616
JW
1511 parent->count--;
1512 } else {
1513 root_tag_clear_all(root);
1514 root->height = 0;
1515 root->rnode = NULL;
1516 }
1517
1518 radix_tree_node_free(node);
1519 deleted = true;
1520
1521 node = parent;
1522 } while (node);
1523
1524 return deleted;
1525}
1526
57578c2e
MW
1527static inline void delete_sibling_entries(struct radix_tree_node *node,
1528 void *ptr, unsigned offset)
1529{
1530#ifdef CONFIG_RADIX_TREE_MULTIORDER
1531 int i;
1532 for (i = 1; offset + i < RADIX_TREE_MAP_SIZE; i++) {
1533 if (node->slots[offset + i] != ptr)
1534 break;
1535 node->slots[offset + i] = NULL;
1536 node->count--;
1537 }
1538#endif
1539}
1540
1da177e4 1541/**
53c59f26 1542 * radix_tree_delete_item - delete an item from a radix tree
1da177e4
LT
1543 * @root: radix tree root
1544 * @index: index key
53c59f26 1545 * @item: expected item
1da177e4 1546 *
53c59f26 1547 * Remove @item at @index from the radix tree rooted at @root.
1da177e4 1548 *
53c59f26
JW
1549 * Returns the address of the deleted item, or NULL if it was not present
1550 * or the entry at the given @index was not @item.
1da177e4 1551 */
53c59f26
JW
1552void *radix_tree_delete_item(struct radix_tree_root *root,
1553 unsigned long index, void *item)
1da177e4 1554{
139e5616 1555 struct radix_tree_node *node;
57578c2e 1556 unsigned int offset;
139e5616
JW
1557 void **slot;
1558 void *entry;
d5274261 1559 int tag;
1da177e4 1560
139e5616
JW
1561 entry = __radix_tree_lookup(root, index, &node, &slot);
1562 if (!entry)
1563 return NULL;
1da177e4 1564
139e5616
JW
1565 if (item && entry != item)
1566 return NULL;
1567
1568 if (!node) {
612d6c19
NP
1569 root_tag_clear_all(root);
1570 root->rnode = NULL;
139e5616 1571 return entry;
612d6c19 1572 }
1da177e4 1573
29e0967c 1574 offset = get_slot_offset(node, slot);
53c59f26 1575
1da177e4 1576 /*
e2bdb933
HD
1577 * Clear all tags associated with the item to be deleted.
1578 * This way of doing it would be inefficient, but seldom is any set.
1da177e4 1579 */
daff89f3 1580 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
e2bdb933 1581 if (tag_get(node, tag, offset))
612d6c19 1582 radix_tree_tag_clear(root, index, tag);
d5274261 1583 }
1da177e4 1584
57578c2e 1585 delete_sibling_entries(node, ptr_to_indirect(slot), offset);
139e5616
JW
1586 node->slots[offset] = NULL;
1587 node->count--;
e2bdb933 1588
449dd698 1589 __radix_tree_delete_node(root, node);
612d6c19 1590
139e5616 1591 return entry;
1da177e4 1592}
53c59f26
JW
1593EXPORT_SYMBOL(radix_tree_delete_item);
1594
1595/**
1596 * radix_tree_delete - delete an item from a radix tree
1597 * @root: radix tree root
1598 * @index: index key
1599 *
1600 * Remove the item at @index from the radix tree rooted at @root.
1601 *
1602 * Returns the address of the deleted item, or NULL if it was not present.
1603 */
1604void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
1605{
1606 return radix_tree_delete_item(root, index, NULL);
1607}
1da177e4
LT
1608EXPORT_SYMBOL(radix_tree_delete);
1609
1610/**
1611 * radix_tree_tagged - test whether any items in the tree are tagged
1612 * @root: radix tree root
1613 * @tag: tag to test
1614 */
daff89f3 1615int radix_tree_tagged(struct radix_tree_root *root, unsigned int tag)
1da177e4 1616{
612d6c19 1617 return root_tag_get(root, tag);
1da177e4
LT
1618}
1619EXPORT_SYMBOL(radix_tree_tagged);
1620
1621static void
449dd698 1622radix_tree_node_ctor(void *arg)
1da177e4 1623{
449dd698
JW
1624 struct radix_tree_node *node = arg;
1625
1626 memset(node, 0, sizeof(*node));
1627 INIT_LIST_HEAD(&node->private_list);
1da177e4
LT
1628}
1629
1630static __init unsigned long __maxindex(unsigned int height)
1631{
430d275a
PL
1632 unsigned int width = height * RADIX_TREE_MAP_SHIFT;
1633 int shift = RADIX_TREE_INDEX_BITS - width;
1634
1635 if (shift < 0)
1636 return ~0UL;
1637 if (shift >= BITS_PER_LONG)
1638 return 0UL;
1639 return ~0UL >> shift;
1da177e4
LT
1640}
1641
1642static __init void radix_tree_init_maxindex(void)
1643{
1644 unsigned int i;
1645
1646 for (i = 0; i < ARRAY_SIZE(height_to_maxindex); i++)
1647 height_to_maxindex[i] = __maxindex(i);
1648}
1649
1da177e4
LT
1650static int radix_tree_callback(struct notifier_block *nfb,
1651 unsigned long action,
1652 void *hcpu)
1653{
1654 int cpu = (long)hcpu;
1655 struct radix_tree_preload *rtp;
9d2a8da0 1656 struct radix_tree_node *node;
1da177e4
LT
1657
1658 /* Free per-cpu pool of perloaded nodes */
8bb78442 1659 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
1da177e4
LT
1660 rtp = &per_cpu(radix_tree_preloads, cpu);
1661 while (rtp->nr) {
9d2a8da0
KS
1662 node = rtp->nodes;
1663 rtp->nodes = node->private_data;
1664 kmem_cache_free(radix_tree_node_cachep, node);
1665 rtp->nr--;
1da177e4
LT
1666 }
1667 }
1668 return NOTIFY_OK;
1669}
1da177e4
LT
1670
1671void __init radix_tree_init(void)
1672{
1673 radix_tree_node_cachep = kmem_cache_create("radix_tree_node",
1674 sizeof(struct radix_tree_node), 0,
488514d1
CL
1675 SLAB_PANIC | SLAB_RECLAIM_ACCOUNT,
1676 radix_tree_node_ctor);
1da177e4
LT
1677 radix_tree_init_maxindex();
1678 hotcpu_notifier(radix_tree_callback, 0);
1679}