radix-tree: use indirect bit
[linux-2.6.git] / lib / radix-tree.c
1 /*
2  * Copyright (C) 2001 Momchil Velikov
3  * Portions Copyright (C) 2001 Christoph Hellwig
4  * Copyright (C) 2005 SGI, Christoph Lameter <clameter@sgi.com>
5  * Copyright (C) 2006 Nick Piggin
6  *
7  * This program is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU General Public License as
9  * published by the Free Software Foundation; either version 2, or (at
10  * your option) any later version.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public License
18  * along with this program; if not, write to the Free Software
19  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20  */
21
22 #include <linux/errno.h>
23 #include <linux/init.h>
24 #include <linux/kernel.h>
25 #include <linux/module.h>
26 #include <linux/radix-tree.h>
27 #include <linux/percpu.h>
28 #include <linux/slab.h>
29 #include <linux/notifier.h>
30 #include <linux/cpu.h>
31 #include <linux/gfp.h>
32 #include <linux/string.h>
33 #include <linux/bitops.h>
34 #include <linux/rcupdate.h>
35
36
37 #ifdef __KERNEL__
38 #define RADIX_TREE_MAP_SHIFT    (CONFIG_BASE_SMALL ? 4 : 6)
39 #else
40 #define RADIX_TREE_MAP_SHIFT    3       /* For more stressful testing */
41 #endif
42
43 #define RADIX_TREE_MAP_SIZE     (1UL << RADIX_TREE_MAP_SHIFT)
44 #define RADIX_TREE_MAP_MASK     (RADIX_TREE_MAP_SIZE-1)
45
46 #define RADIX_TREE_TAG_LONGS    \
47         ((RADIX_TREE_MAP_SIZE + BITS_PER_LONG - 1) / BITS_PER_LONG)
48
49 struct radix_tree_node {
50         unsigned int    height;         /* Height from the bottom */
51         unsigned int    count;
52         struct rcu_head rcu_head;
53         void            *slots[RADIX_TREE_MAP_SIZE];
54         unsigned long   tags[RADIX_TREE_MAX_TAGS][RADIX_TREE_TAG_LONGS];
55 };
56
57 struct radix_tree_path {
58         struct radix_tree_node *node;
59         int offset;
60 };
61
62 #define RADIX_TREE_INDEX_BITS  (8 /* CHAR_BIT */ * sizeof(unsigned long))
63 #define RADIX_TREE_MAX_PATH (RADIX_TREE_INDEX_BITS/RADIX_TREE_MAP_SHIFT + 2)
64
65 static unsigned long height_to_maxindex[RADIX_TREE_MAX_PATH] __read_mostly;
66
67 /*
68  * Radix tree node cache.
69  */
70 static struct kmem_cache *radix_tree_node_cachep;
71
72 /*
73  * Per-cpu pool of preloaded nodes
74  */
75 struct radix_tree_preload {
76         int nr;
77         struct radix_tree_node *nodes[RADIX_TREE_MAX_PATH];
78 };
79 DEFINE_PER_CPU(struct radix_tree_preload, radix_tree_preloads) = { 0, };
80
81 static inline gfp_t root_gfp_mask(struct radix_tree_root *root)
82 {
83         return root->gfp_mask & __GFP_BITS_MASK;
84 }
85
86 /*
87  * This assumes that the caller has performed appropriate preallocation, and
88  * that the caller has pinned this thread of control to the current CPU.
89  */
90 static struct radix_tree_node *
91 radix_tree_node_alloc(struct radix_tree_root *root)
92 {
93         struct radix_tree_node *ret;
94         gfp_t gfp_mask = root_gfp_mask(root);
95
96         ret = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
97         if (ret == NULL && !(gfp_mask & __GFP_WAIT)) {
98                 struct radix_tree_preload *rtp;
99
100                 rtp = &__get_cpu_var(radix_tree_preloads);
101                 if (rtp->nr) {
102                         ret = rtp->nodes[rtp->nr - 1];
103                         rtp->nodes[rtp->nr - 1] = NULL;
104                         rtp->nr--;
105                 }
106         }
107         BUG_ON(radix_tree_is_indirect_ptr(ret));
108         return ret;
109 }
110
111 static void radix_tree_node_rcu_free(struct rcu_head *head)
112 {
113         struct radix_tree_node *node =
114                         container_of(head, struct radix_tree_node, rcu_head);
115         kmem_cache_free(radix_tree_node_cachep, node);
116 }
117
118 static inline void
119 radix_tree_node_free(struct radix_tree_node *node)
120 {
121         call_rcu(&node->rcu_head, radix_tree_node_rcu_free);
122 }
123
124 /*
125  * Load up this CPU's radix_tree_node buffer with sufficient objects to
126  * ensure that the addition of a single element in the tree cannot fail.  On
127  * success, return zero, with preemption disabled.  On error, return -ENOMEM
128  * with preemption not disabled.
129  */
130 int radix_tree_preload(gfp_t gfp_mask)
131 {
132         struct radix_tree_preload *rtp;
133         struct radix_tree_node *node;
134         int ret = -ENOMEM;
135
136         preempt_disable();
137         rtp = &__get_cpu_var(radix_tree_preloads);
138         while (rtp->nr < ARRAY_SIZE(rtp->nodes)) {
139                 preempt_enable();
140                 node = kmem_cache_alloc(radix_tree_node_cachep, gfp_mask);
141                 if (node == NULL)
142                         goto out;
143                 preempt_disable();
144                 rtp = &__get_cpu_var(radix_tree_preloads);
145                 if (rtp->nr < ARRAY_SIZE(rtp->nodes))
146                         rtp->nodes[rtp->nr++] = node;
147                 else
148                         kmem_cache_free(radix_tree_node_cachep, node);
149         }
150         ret = 0;
151 out:
152         return ret;
153 }
154 EXPORT_SYMBOL(radix_tree_preload);
155
156 static inline void tag_set(struct radix_tree_node *node, unsigned int tag,
157                 int offset)
158 {
159         __set_bit(offset, node->tags[tag]);
160 }
161
162 static inline void tag_clear(struct radix_tree_node *node, unsigned int tag,
163                 int offset)
164 {
165         __clear_bit(offset, node->tags[tag]);
166 }
167
168 static inline int tag_get(struct radix_tree_node *node, unsigned int tag,
169                 int offset)
170 {
171         return test_bit(offset, node->tags[tag]);
172 }
173
174 static inline void root_tag_set(struct radix_tree_root *root, unsigned int tag)
175 {
176         root->gfp_mask |= (__force gfp_t)(1 << (tag + __GFP_BITS_SHIFT));
177 }
178
179
180 static inline void root_tag_clear(struct radix_tree_root *root, unsigned int tag)
181 {
182         root->gfp_mask &= (__force gfp_t)~(1 << (tag + __GFP_BITS_SHIFT));
183 }
184
185 static inline void root_tag_clear_all(struct radix_tree_root *root)
186 {
187         root->gfp_mask &= __GFP_BITS_MASK;
188 }
189
190 static inline int root_tag_get(struct radix_tree_root *root, unsigned int tag)
191 {
192         return (__force unsigned)root->gfp_mask & (1 << (tag + __GFP_BITS_SHIFT));
193 }
194
195 /*
196  * Returns 1 if any slot in the node has this tag set.
197  * Otherwise returns 0.
198  */
199 static inline int any_tag_set(struct radix_tree_node *node, unsigned int tag)
200 {
201         int idx;
202         for (idx = 0; idx < RADIX_TREE_TAG_LONGS; idx++) {
203                 if (node->tags[tag][idx])
204                         return 1;
205         }
206         return 0;
207 }
208
209 /*
210  *      Return the maximum key which can be store into a
211  *      radix tree with height HEIGHT.
212  */
213 static inline unsigned long radix_tree_maxindex(unsigned int height)
214 {
215         return height_to_maxindex[height];
216 }
217
218 /*
219  *      Extend a radix tree so it can store key @index.
220  */
221 static int radix_tree_extend(struct radix_tree_root *root, unsigned long index)
222 {
223         struct radix_tree_node *node;
224         unsigned int height;
225         int tag;
226
227         /* Figure out what the height should be.  */
228         height = root->height + 1;
229         while (index > radix_tree_maxindex(height))
230                 height++;
231
232         if (root->rnode == NULL) {
233                 root->height = height;
234                 goto out;
235         }
236
237         do {
238                 unsigned int newheight;
239                 if (!(node = radix_tree_node_alloc(root)))
240                         return -ENOMEM;
241
242                 /* Increase the height.  */
243                 node->slots[0] = radix_tree_indirect_to_ptr(root->rnode);
244
245                 /* Propagate the aggregated tag info into the new root */
246                 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
247                         if (root_tag_get(root, tag))
248                                 tag_set(node, tag, 0);
249                 }
250
251                 newheight = root->height+1;
252                 node->height = newheight;
253                 node->count = 1;
254                 node = radix_tree_ptr_to_indirect(node);
255                 rcu_assign_pointer(root->rnode, node);
256                 root->height = newheight;
257         } while (height > root->height);
258 out:
259         return 0;
260 }
261
262 /**
263  *      radix_tree_insert    -    insert into a radix tree
264  *      @root:          radix tree root
265  *      @index:         index key
266  *      @item:          item to insert
267  *
268  *      Insert an item into the radix tree at position @index.
269  */
270 int radix_tree_insert(struct radix_tree_root *root,
271                         unsigned long index, void *item)
272 {
273         struct radix_tree_node *node = NULL, *slot;
274         unsigned int height, shift;
275         int offset;
276         int error;
277
278         BUG_ON(radix_tree_is_indirect_ptr(item));
279
280         /* Make sure the tree is high enough.  */
281         if (index > radix_tree_maxindex(root->height)) {
282                 error = radix_tree_extend(root, index);
283                 if (error)
284                         return error;
285         }
286
287         slot = radix_tree_indirect_to_ptr(root->rnode);
288
289         height = root->height;
290         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
291
292         offset = 0;                     /* uninitialised var warning */
293         while (height > 0) {
294                 if (slot == NULL) {
295                         /* Have to add a child node.  */
296                         if (!(slot = radix_tree_node_alloc(root)))
297                                 return -ENOMEM;
298                         slot->height = height;
299                         if (node) {
300                                 rcu_assign_pointer(node->slots[offset], slot);
301                                 node->count++;
302                         } else
303                                 rcu_assign_pointer(root->rnode,
304                                         radix_tree_ptr_to_indirect(slot));
305                 }
306
307                 /* Go a level down */
308                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
309                 node = slot;
310                 slot = node->slots[offset];
311                 shift -= RADIX_TREE_MAP_SHIFT;
312                 height--;
313         }
314
315         if (slot != NULL)
316                 return -EEXIST;
317
318         if (node) {
319                 node->count++;
320                 rcu_assign_pointer(node->slots[offset], item);
321                 BUG_ON(tag_get(node, 0, offset));
322                 BUG_ON(tag_get(node, 1, offset));
323         } else {
324                 rcu_assign_pointer(root->rnode, item);
325                 BUG_ON(root_tag_get(root, 0));
326                 BUG_ON(root_tag_get(root, 1));
327         }
328
329         return 0;
330 }
331 EXPORT_SYMBOL(radix_tree_insert);
332
333 /**
334  *      radix_tree_lookup_slot    -    lookup a slot in a radix tree
335  *      @root:          radix tree root
336  *      @index:         index key
337  *
338  *      Returns:  the slot corresponding to the position @index in the
339  *      radix tree @root. This is useful for update-if-exists operations.
340  *
341  *      This function cannot be called under rcu_read_lock, it must be
342  *      excluded from writers, as must the returned slot for subsequent
343  *      use by radix_tree_deref_slot() and radix_tree_replace slot.
344  *      Caller must hold tree write locked across slot lookup and
345  *      replace.
346  */
347 void **radix_tree_lookup_slot(struct radix_tree_root *root, unsigned long index)
348 {
349         unsigned int height, shift;
350         struct radix_tree_node *node, **slot;
351
352         node = root->rnode;
353         if (node == NULL)
354                 return NULL;
355
356         if (!radix_tree_is_indirect_ptr(node)) {
357                 if (index > 0)
358                         return NULL;
359                 return (void **)&root->rnode;
360         }
361         node = radix_tree_indirect_to_ptr(node);
362
363         height = node->height;
364         if (index > radix_tree_maxindex(height))
365                 return NULL;
366
367         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
368
369         do {
370                 slot = (struct radix_tree_node **)
371                         (node->slots + ((index>>shift) & RADIX_TREE_MAP_MASK));
372                 node = *slot;
373                 if (node == NULL)
374                         return NULL;
375
376                 shift -= RADIX_TREE_MAP_SHIFT;
377                 height--;
378         } while (height > 0);
379
380         return (void **)slot;
381 }
382 EXPORT_SYMBOL(radix_tree_lookup_slot);
383
384 /**
385  *      radix_tree_lookup    -    perform lookup operation on a radix tree
386  *      @root:          radix tree root
387  *      @index:         index key
388  *
389  *      Lookup the item at the position @index in the radix tree @root.
390  *
391  *      This function can be called under rcu_read_lock, however the caller
392  *      must manage lifetimes of leaf nodes (eg. RCU may also be used to free
393  *      them safely). No RCU barriers are required to access or modify the
394  *      returned item, however.
395  */
396 void *radix_tree_lookup(struct radix_tree_root *root, unsigned long index)
397 {
398         unsigned int height, shift;
399         struct radix_tree_node *node, **slot;
400
401         node = rcu_dereference(root->rnode);
402         if (node == NULL)
403                 return NULL;
404
405         if (!radix_tree_is_indirect_ptr(node)) {
406                 if (index > 0)
407                         return NULL;
408                 return node;
409         }
410         node = radix_tree_indirect_to_ptr(node);
411
412         height = node->height;
413         if (index > radix_tree_maxindex(height))
414                 return NULL;
415
416         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
417
418         do {
419                 slot = (struct radix_tree_node **)
420                         (node->slots + ((index>>shift) & RADIX_TREE_MAP_MASK));
421                 node = rcu_dereference(*slot);
422                 if (node == NULL)
423                         return NULL;
424
425                 shift -= RADIX_TREE_MAP_SHIFT;
426                 height--;
427         } while (height > 0);
428
429         return node;
430 }
431 EXPORT_SYMBOL(radix_tree_lookup);
432
433 /**
434  *      radix_tree_tag_set - set a tag on a radix tree node
435  *      @root:          radix tree root
436  *      @index:         index key
437  *      @tag:           tag index
438  *
439  *      Set the search tag (which must be < RADIX_TREE_MAX_TAGS)
440  *      corresponding to @index in the radix tree.  From
441  *      the root all the way down to the leaf node.
442  *
443  *      Returns the address of the tagged item.   Setting a tag on a not-present
444  *      item is a bug.
445  */
446 void *radix_tree_tag_set(struct radix_tree_root *root,
447                         unsigned long index, unsigned int tag)
448 {
449         unsigned int height, shift;
450         struct radix_tree_node *slot;
451
452         height = root->height;
453         BUG_ON(index > radix_tree_maxindex(height));
454
455         slot = radix_tree_indirect_to_ptr(root->rnode);
456         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
457
458         while (height > 0) {
459                 int offset;
460
461                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
462                 if (!tag_get(slot, tag, offset))
463                         tag_set(slot, tag, offset);
464                 slot = slot->slots[offset];
465                 BUG_ON(slot == NULL);
466                 shift -= RADIX_TREE_MAP_SHIFT;
467                 height--;
468         }
469
470         /* set the root's tag bit */
471         if (slot && !root_tag_get(root, tag))
472                 root_tag_set(root, tag);
473
474         return slot;
475 }
476 EXPORT_SYMBOL(radix_tree_tag_set);
477
478 /**
479  *      radix_tree_tag_clear - clear a tag on a radix tree node
480  *      @root:          radix tree root
481  *      @index:         index key
482  *      @tag:           tag index
483  *
484  *      Clear the search tag (which must be < RADIX_TREE_MAX_TAGS)
485  *      corresponding to @index in the radix tree.  If
486  *      this causes the leaf node to have no tags set then clear the tag in the
487  *      next-to-leaf node, etc.
488  *
489  *      Returns the address of the tagged item on success, else NULL.  ie:
490  *      has the same return value and semantics as radix_tree_lookup().
491  */
492 void *radix_tree_tag_clear(struct radix_tree_root *root,
493                         unsigned long index, unsigned int tag)
494 {
495         struct radix_tree_path path[RADIX_TREE_MAX_PATH], *pathp = path;
496         struct radix_tree_node *slot = NULL;
497         unsigned int height, shift;
498
499         height = root->height;
500         if (index > radix_tree_maxindex(height))
501                 goto out;
502
503         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
504         pathp->node = NULL;
505         slot = radix_tree_indirect_to_ptr(root->rnode);
506
507         while (height > 0) {
508                 int offset;
509
510                 if (slot == NULL)
511                         goto out;
512
513                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
514                 pathp[1].offset = offset;
515                 pathp[1].node = slot;
516                 slot = slot->slots[offset];
517                 pathp++;
518                 shift -= RADIX_TREE_MAP_SHIFT;
519                 height--;
520         }
521
522         if (slot == NULL)
523                 goto out;
524
525         while (pathp->node) {
526                 if (!tag_get(pathp->node, tag, pathp->offset))
527                         goto out;
528                 tag_clear(pathp->node, tag, pathp->offset);
529                 if (any_tag_set(pathp->node, tag))
530                         goto out;
531                 pathp--;
532         }
533
534         /* clear the root's tag bit */
535         if (root_tag_get(root, tag))
536                 root_tag_clear(root, tag);
537
538 out:
539         return slot;
540 }
541 EXPORT_SYMBOL(radix_tree_tag_clear);
542
543 #ifndef __KERNEL__      /* Only the test harness uses this at present */
544 /**
545  * radix_tree_tag_get - get a tag on a radix tree node
546  * @root:               radix tree root
547  * @index:              index key
548  * @tag:                tag index (< RADIX_TREE_MAX_TAGS)
549  *
550  * Return values:
551  *
552  *  0: tag not present or not set
553  *  1: tag set
554  */
555 int radix_tree_tag_get(struct radix_tree_root *root,
556                         unsigned long index, unsigned int tag)
557 {
558         unsigned int height, shift;
559         struct radix_tree_node *node;
560         int saw_unset_tag = 0;
561
562         /* check the root's tag bit */
563         if (!root_tag_get(root, tag))
564                 return 0;
565
566         node = rcu_dereference(root->rnode);
567         if (node == NULL)
568                 return 0;
569
570         if (!radix_tree_is_indirect_ptr(node))
571                 return (index == 0);
572         node = radix_tree_indirect_to_ptr(node);
573
574         height = node->height;
575         if (index > radix_tree_maxindex(height))
576                 return 0;
577
578         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
579
580         for ( ; ; ) {
581                 int offset;
582
583                 if (node == NULL)
584                         return 0;
585
586                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
587
588                 /*
589                  * This is just a debug check.  Later, we can bale as soon as
590                  * we see an unset tag.
591                  */
592                 if (!tag_get(node, tag, offset))
593                         saw_unset_tag = 1;
594                 if (height == 1) {
595                         int ret = tag_get(node, tag, offset);
596
597                         BUG_ON(ret && saw_unset_tag);
598                         return !!ret;
599                 }
600                 node = rcu_dereference(node->slots[offset]);
601                 shift -= RADIX_TREE_MAP_SHIFT;
602                 height--;
603         }
604 }
605 EXPORT_SYMBOL(radix_tree_tag_get);
606 #endif
607
608 /**
609  *      radix_tree_next_hole    -    find the next hole (not-present entry)
610  *      @root:          tree root
611  *      @index:         index key
612  *      @max_scan:      maximum range to search
613  *
614  *      Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the lowest
615  *      indexed hole.
616  *
617  *      Returns: the index of the hole if found, otherwise returns an index
618  *      outside of the set specified (in which case 'return - index >= max_scan'
619  *      will be true).
620  *
621  *      radix_tree_next_hole may be called under rcu_read_lock. However, like
622  *      radix_tree_gang_lookup, this will not atomically search a snapshot of the
623  *      tree at a single point in time. For example, if a hole is created at index
624  *      5, then subsequently a hole is created at index 10, radix_tree_next_hole
625  *      covering both indexes may return 10 if called under rcu_read_lock.
626  */
627 unsigned long radix_tree_next_hole(struct radix_tree_root *root,
628                                 unsigned long index, unsigned long max_scan)
629 {
630         unsigned long i;
631
632         for (i = 0; i < max_scan; i++) {
633                 if (!radix_tree_lookup(root, index))
634                         break;
635                 index++;
636                 if (index == 0)
637                         break;
638         }
639
640         return index;
641 }
642 EXPORT_SYMBOL(radix_tree_next_hole);
643
644 static unsigned int
645 __lookup(struct radix_tree_node *slot, void **results, unsigned long index,
646         unsigned int max_items, unsigned long *next_index)
647 {
648         unsigned int nr_found = 0;
649         unsigned int shift, height;
650         unsigned long i;
651
652         height = slot->height;
653         if (height == 0)
654                 goto out;
655         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
656
657         for ( ; height > 1; height--) {
658                 i = (index >> shift) & RADIX_TREE_MAP_MASK;
659                 for (;;) {
660                         if (slot->slots[i] != NULL)
661                                 break;
662                         index &= ~((1UL << shift) - 1);
663                         index += 1UL << shift;
664                         if (index == 0)
665                                 goto out;       /* 32-bit wraparound */
666                         i++;
667                         if (i == RADIX_TREE_MAP_SIZE)
668                                 goto out;
669                 }
670
671                 shift -= RADIX_TREE_MAP_SHIFT;
672                 slot = rcu_dereference(slot->slots[i]);
673                 if (slot == NULL)
674                         goto out;
675         }
676
677         /* Bottom level: grab some items */
678         for (i = index & RADIX_TREE_MAP_MASK; i < RADIX_TREE_MAP_SIZE; i++) {
679                 struct radix_tree_node *node;
680                 index++;
681                 node = slot->slots[i];
682                 if (node) {
683                         results[nr_found++] = rcu_dereference(node);
684                         if (nr_found == max_items)
685                                 goto out;
686                 }
687         }
688 out:
689         *next_index = index;
690         return nr_found;
691 }
692
693 /**
694  *      radix_tree_gang_lookup - perform multiple lookup on a radix tree
695  *      @root:          radix tree root
696  *      @results:       where the results of the lookup are placed
697  *      @first_index:   start the lookup from this key
698  *      @max_items:     place up to this many items at *results
699  *
700  *      Performs an index-ascending scan of the tree for present items.  Places
701  *      them at *@results and returns the number of items which were placed at
702  *      *@results.
703  *
704  *      The implementation is naive.
705  *
706  *      Like radix_tree_lookup, radix_tree_gang_lookup may be called under
707  *      rcu_read_lock. In this case, rather than the returned results being
708  *      an atomic snapshot of the tree at a single point in time, the semantics
709  *      of an RCU protected gang lookup are as though multiple radix_tree_lookups
710  *      have been issued in individual locks, and results stored in 'results'.
711  */
712 unsigned int
713 radix_tree_gang_lookup(struct radix_tree_root *root, void **results,
714                         unsigned long first_index, unsigned int max_items)
715 {
716         unsigned long max_index;
717         struct radix_tree_node *node;
718         unsigned long cur_index = first_index;
719         unsigned int ret;
720
721         node = rcu_dereference(root->rnode);
722         if (!node)
723                 return 0;
724
725         if (!radix_tree_is_indirect_ptr(node)) {
726                 if (first_index > 0)
727                         return 0;
728                 results[0] = node;
729                 return 1;
730         }
731         node = radix_tree_indirect_to_ptr(node);
732
733         max_index = radix_tree_maxindex(node->height);
734
735         ret = 0;
736         while (ret < max_items) {
737                 unsigned int nr_found;
738                 unsigned long next_index;       /* Index of next search */
739
740                 if (cur_index > max_index)
741                         break;
742                 nr_found = __lookup(node, results + ret, cur_index,
743                                         max_items - ret, &next_index);
744                 ret += nr_found;
745                 if (next_index == 0)
746                         break;
747                 cur_index = next_index;
748         }
749
750         return ret;
751 }
752 EXPORT_SYMBOL(radix_tree_gang_lookup);
753
754 /*
755  * FIXME: the two tag_get()s here should use find_next_bit() instead of
756  * open-coding the search.
757  */
758 static unsigned int
759 __lookup_tag(struct radix_tree_node *slot, void **results, unsigned long index,
760         unsigned int max_items, unsigned long *next_index, unsigned int tag)
761 {
762         unsigned int nr_found = 0;
763         unsigned int shift, height;
764
765         height = slot->height;
766         if (height == 0)
767                 goto out;
768         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
769
770         while (height > 0) {
771                 unsigned long i = (index >> shift) & RADIX_TREE_MAP_MASK ;
772
773                 for (;;) {
774                         if (tag_get(slot, tag, i))
775                                 break;
776                         index &= ~((1UL << shift) - 1);
777                         index += 1UL << shift;
778                         if (index == 0)
779                                 goto out;       /* 32-bit wraparound */
780                         i++;
781                         if (i == RADIX_TREE_MAP_SIZE)
782                                 goto out;
783                 }
784                 height--;
785                 if (height == 0) {      /* Bottom level: grab some items */
786                         unsigned long j = index & RADIX_TREE_MAP_MASK;
787
788                         for ( ; j < RADIX_TREE_MAP_SIZE; j++) {
789                                 struct radix_tree_node *node;
790                                 index++;
791                                 if (!tag_get(slot, tag, j))
792                                         continue;
793                                 node = slot->slots[j];
794                                 /*
795                                  * Even though the tag was found set, we need to
796                                  * recheck that we have a non-NULL node, because
797                                  * if this lookup is lockless, it may have been
798                                  * subsequently deleted.
799                                  *
800                                  * Similar care must be taken in any place that
801                                  * lookup ->slots[x] without a lock (ie. can't
802                                  * rely on its value remaining the same).
803                                  */
804                                 if (node) {
805                                         node = rcu_dereference(node);
806                                         results[nr_found++] = node;
807                                         if (nr_found == max_items)
808                                                 goto out;
809                                 }
810                         }
811                 }
812                 shift -= RADIX_TREE_MAP_SHIFT;
813                 slot = rcu_dereference(slot->slots[i]);
814                 if (slot == NULL)
815                         break;
816         }
817 out:
818         *next_index = index;
819         return nr_found;
820 }
821
822 /**
823  *      radix_tree_gang_lookup_tag - perform multiple lookup on a radix tree
824  *                                   based on a tag
825  *      @root:          radix tree root
826  *      @results:       where the results of the lookup are placed
827  *      @first_index:   start the lookup from this key
828  *      @max_items:     place up to this many items at *results
829  *      @tag:           the tag index (< RADIX_TREE_MAX_TAGS)
830  *
831  *      Performs an index-ascending scan of the tree for present items which
832  *      have the tag indexed by @tag set.  Places the items at *@results and
833  *      returns the number of items which were placed at *@results.
834  */
835 unsigned int
836 radix_tree_gang_lookup_tag(struct radix_tree_root *root, void **results,
837                 unsigned long first_index, unsigned int max_items,
838                 unsigned int tag)
839 {
840         struct radix_tree_node *node;
841         unsigned long max_index;
842         unsigned long cur_index = first_index;
843         unsigned int ret;
844
845         /* check the root's tag bit */
846         if (!root_tag_get(root, tag))
847                 return 0;
848
849         node = rcu_dereference(root->rnode);
850         if (!node)
851                 return 0;
852
853         if (!radix_tree_is_indirect_ptr(node)) {
854                 if (first_index > 0)
855                         return 0;
856                 results[0] = node;
857                 return 1;
858         }
859         node = radix_tree_indirect_to_ptr(node);
860
861         max_index = radix_tree_maxindex(node->height);
862
863         ret = 0;
864         while (ret < max_items) {
865                 unsigned int nr_found;
866                 unsigned long next_index;       /* Index of next search */
867
868                 if (cur_index > max_index)
869                         break;
870                 nr_found = __lookup_tag(node, results + ret, cur_index,
871                                         max_items - ret, &next_index, tag);
872                 ret += nr_found;
873                 if (next_index == 0)
874                         break;
875                 cur_index = next_index;
876         }
877
878         return ret;
879 }
880 EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
881
882 /**
883  *      radix_tree_shrink    -    shrink height of a radix tree to minimal
884  *      @root           radix tree root
885  */
886 static inline void radix_tree_shrink(struct radix_tree_root *root)
887 {
888         /* try to shrink tree height */
889         while (root->height > 0) {
890                 struct radix_tree_node *to_free = root->rnode;
891                 void *newptr;
892
893                 BUG_ON(!radix_tree_is_indirect_ptr(to_free));
894                 to_free = radix_tree_indirect_to_ptr(to_free);
895
896                 /*
897                  * The candidate node has more than one child, or its child
898                  * is not at the leftmost slot, we cannot shrink.
899                  */
900                 if (to_free->count != 1)
901                         break;
902                 if (!to_free->slots[0])
903                         break;
904
905                 /*
906                  * We don't need rcu_assign_pointer(), since we are simply
907                  * moving the node from one part of the tree to another. If
908                  * it was safe to dereference the old pointer to it
909                  * (to_free->slots[0]), it will be safe to dereference the new
910                  * one (root->rnode).
911                  */
912                 newptr = to_free->slots[0];
913                 if (root->height > 1)
914                         newptr = radix_tree_ptr_to_indirect(newptr);
915                 root->rnode = newptr;
916                 root->height--;
917                 /* must only free zeroed nodes into the slab */
918                 tag_clear(to_free, 0, 0);
919                 tag_clear(to_free, 1, 0);
920                 to_free->slots[0] = NULL;
921                 to_free->count = 0;
922                 radix_tree_node_free(to_free);
923         }
924 }
925
926 /**
927  *      radix_tree_delete    -    delete an item from a radix tree
928  *      @root:          radix tree root
929  *      @index:         index key
930  *
931  *      Remove the item at @index from the radix tree rooted at @root.
932  *
933  *      Returns the address of the deleted item, or NULL if it was not present.
934  */
935 void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
936 {
937         struct radix_tree_path path[RADIX_TREE_MAX_PATH], *pathp = path;
938         struct radix_tree_node *slot = NULL;
939         struct radix_tree_node *to_free;
940         unsigned int height, shift;
941         int tag;
942         int offset;
943
944         height = root->height;
945         if (index > radix_tree_maxindex(height))
946                 goto out;
947
948         slot = root->rnode;
949         if (height == 0) {
950                 root_tag_clear_all(root);
951                 root->rnode = NULL;
952                 goto out;
953         }
954         slot = radix_tree_indirect_to_ptr(slot);
955
956         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
957         pathp->node = NULL;
958
959         do {
960                 if (slot == NULL)
961                         goto out;
962
963                 pathp++;
964                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
965                 pathp->offset = offset;
966                 pathp->node = slot;
967                 slot = slot->slots[offset];
968                 shift -= RADIX_TREE_MAP_SHIFT;
969                 height--;
970         } while (height > 0);
971
972         if (slot == NULL)
973                 goto out;
974
975         /*
976          * Clear all tags associated with the just-deleted item
977          */
978         for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
979                 if (tag_get(pathp->node, tag, pathp->offset))
980                         radix_tree_tag_clear(root, index, tag);
981         }
982
983         to_free = NULL;
984         /* Now free the nodes we do not need anymore */
985         while (pathp->node) {
986                 pathp->node->slots[pathp->offset] = NULL;
987                 pathp->node->count--;
988                 /*
989                  * Queue the node for deferred freeing after the
990                  * last reference to it disappears (set NULL, above).
991                  */
992                 if (to_free)
993                         radix_tree_node_free(to_free);
994
995                 if (pathp->node->count) {
996                         if (pathp->node ==
997                                         radix_tree_indirect_to_ptr(root->rnode))
998                                 radix_tree_shrink(root);
999                         goto out;
1000                 }
1001
1002                 /* Node with zero slots in use so free it */
1003                 to_free = pathp->node;
1004                 pathp--;
1005
1006         }
1007         root_tag_clear_all(root);
1008         root->height = 0;
1009         root->rnode = NULL;
1010         if (to_free)
1011                 radix_tree_node_free(to_free);
1012
1013 out:
1014         return slot;
1015 }
1016 EXPORT_SYMBOL(radix_tree_delete);
1017
1018 /**
1019  *      radix_tree_tagged - test whether any items in the tree are tagged
1020  *      @root:          radix tree root
1021  *      @tag:           tag to test
1022  */
1023 int radix_tree_tagged(struct radix_tree_root *root, unsigned int tag)
1024 {
1025         return root_tag_get(root, tag);
1026 }
1027 EXPORT_SYMBOL(radix_tree_tagged);
1028
1029 static void
1030 radix_tree_node_ctor(void *node, struct kmem_cache *cachep, unsigned long flags)
1031 {
1032         memset(node, 0, sizeof(struct radix_tree_node));
1033 }
1034
1035 static __init unsigned long __maxindex(unsigned int height)
1036 {
1037         unsigned int tmp = height * RADIX_TREE_MAP_SHIFT;
1038         unsigned long index = (~0UL >> (RADIX_TREE_INDEX_BITS - tmp - 1)) >> 1;
1039
1040         if (tmp >= RADIX_TREE_INDEX_BITS)
1041                 index = ~0UL;
1042         return index;
1043 }
1044
1045 static __init void radix_tree_init_maxindex(void)
1046 {
1047         unsigned int i;
1048
1049         for (i = 0; i < ARRAY_SIZE(height_to_maxindex); i++)
1050                 height_to_maxindex[i] = __maxindex(i);
1051 }
1052
1053 static int radix_tree_callback(struct notifier_block *nfb,
1054                             unsigned long action,
1055                             void *hcpu)
1056 {
1057        int cpu = (long)hcpu;
1058        struct radix_tree_preload *rtp;
1059
1060        /* Free per-cpu pool of perloaded nodes */
1061        if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
1062                rtp = &per_cpu(radix_tree_preloads, cpu);
1063                while (rtp->nr) {
1064                        kmem_cache_free(radix_tree_node_cachep,
1065                                        rtp->nodes[rtp->nr-1]);
1066                        rtp->nodes[rtp->nr-1] = NULL;
1067                        rtp->nr--;
1068                }
1069        }
1070        return NOTIFY_OK;
1071 }
1072
1073 void __init radix_tree_init(void)
1074 {
1075         radix_tree_node_cachep = kmem_cache_create("radix_tree_node",
1076                         sizeof(struct radix_tree_node), 0,
1077                         SLAB_PANIC, radix_tree_node_ctor);
1078         radix_tree_init_maxindex();
1079         hotcpu_notifier(radix_tree_callback, 0);
1080 }