48c250fe2233d82c6d8fd563e83c234cdf1109ee
[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 (DIV_ROUND_UP(RADIX_TREE_INDEX_BITS, \
64                                           RADIX_TREE_MAP_SHIFT))
65
66 /*
67  * The height_to_maxindex array needs to be one deeper than the maximum
68  * path as height 0 holds only 1 entry.
69  */
70 static unsigned long height_to_maxindex[RADIX_TREE_MAX_PATH + 1] __read_mostly;
71
72 /*
73  * Radix tree node cache.
74  */
75 static struct kmem_cache *radix_tree_node_cachep;
76
77 /*
78  * Per-cpu pool of preloaded nodes
79  */
80 struct radix_tree_preload {
81         int nr;
82         struct radix_tree_node *nodes[RADIX_TREE_MAX_PATH];
83 };
84 DEFINE_PER_CPU(struct radix_tree_preload, radix_tree_preloads) = { 0, };
85
86 static inline gfp_t root_gfp_mask(struct radix_tree_root *root)
87 {
88         return root->gfp_mask & __GFP_BITS_MASK;
89 }
90
91 /*
92  * This assumes that the caller has performed appropriate preallocation, and
93  * that the caller has pinned this thread of control to the current CPU.
94  */
95 static struct radix_tree_node *
96 radix_tree_node_alloc(struct radix_tree_root *root)
97 {
98         struct radix_tree_node *ret;
99         gfp_t gfp_mask = root_gfp_mask(root);
100
101         ret = kmem_cache_alloc(radix_tree_node_cachep,
102                                 set_migrateflags(gfp_mask, __GFP_RECLAIMABLE));
103         if (ret == NULL && !(gfp_mask & __GFP_WAIT)) {
104                 struct radix_tree_preload *rtp;
105
106                 rtp = &__get_cpu_var(radix_tree_preloads);
107                 if (rtp->nr) {
108                         ret = rtp->nodes[rtp->nr - 1];
109                         rtp->nodes[rtp->nr - 1] = NULL;
110                         rtp->nr--;
111                 }
112         }
113         BUG_ON(radix_tree_is_indirect_ptr(ret));
114         return ret;
115 }
116
117 static void radix_tree_node_rcu_free(struct rcu_head *head)
118 {
119         struct radix_tree_node *node =
120                         container_of(head, struct radix_tree_node, rcu_head);
121         kmem_cache_free(radix_tree_node_cachep, node);
122 }
123
124 static inline void
125 radix_tree_node_free(struct radix_tree_node *node)
126 {
127         call_rcu(&node->rcu_head, radix_tree_node_rcu_free);
128 }
129
130 /*
131  * Load up this CPU's radix_tree_node buffer with sufficient objects to
132  * ensure that the addition of a single element in the tree cannot fail.  On
133  * success, return zero, with preemption disabled.  On error, return -ENOMEM
134  * with preemption not disabled.
135  */
136 int radix_tree_preload(gfp_t gfp_mask)
137 {
138         struct radix_tree_preload *rtp;
139         struct radix_tree_node *node;
140         int ret = -ENOMEM;
141
142         preempt_disable();
143         rtp = &__get_cpu_var(radix_tree_preloads);
144         while (rtp->nr < ARRAY_SIZE(rtp->nodes)) {
145                 preempt_enable();
146                 node = kmem_cache_alloc(radix_tree_node_cachep,
147                                 set_migrateflags(gfp_mask, __GFP_RECLAIMABLE));
148                 if (node == NULL)
149                         goto out;
150                 preempt_disable();
151                 rtp = &__get_cpu_var(radix_tree_preloads);
152                 if (rtp->nr < ARRAY_SIZE(rtp->nodes))
153                         rtp->nodes[rtp->nr++] = node;
154                 else
155                         kmem_cache_free(radix_tree_node_cachep, node);
156         }
157         ret = 0;
158 out:
159         return ret;
160 }
161 EXPORT_SYMBOL(radix_tree_preload);
162
163 static inline void tag_set(struct radix_tree_node *node, unsigned int tag,
164                 int offset)
165 {
166         __set_bit(offset, node->tags[tag]);
167 }
168
169 static inline void tag_clear(struct radix_tree_node *node, unsigned int tag,
170                 int offset)
171 {
172         __clear_bit(offset, node->tags[tag]);
173 }
174
175 static inline int tag_get(struct radix_tree_node *node, unsigned int tag,
176                 int offset)
177 {
178         return test_bit(offset, node->tags[tag]);
179 }
180
181 static inline void root_tag_set(struct radix_tree_root *root, unsigned int tag)
182 {
183         root->gfp_mask |= (__force gfp_t)(1 << (tag + __GFP_BITS_SHIFT));
184 }
185
186
187 static inline void root_tag_clear(struct radix_tree_root *root, unsigned int tag)
188 {
189         root->gfp_mask &= (__force gfp_t)~(1 << (tag + __GFP_BITS_SHIFT));
190 }
191
192 static inline void root_tag_clear_all(struct radix_tree_root *root)
193 {
194         root->gfp_mask &= __GFP_BITS_MASK;
195 }
196
197 static inline int root_tag_get(struct radix_tree_root *root, unsigned int tag)
198 {
199         return (__force unsigned)root->gfp_mask & (1 << (tag + __GFP_BITS_SHIFT));
200 }
201
202 /*
203  * Returns 1 if any slot in the node has this tag set.
204  * Otherwise returns 0.
205  */
206 static inline int any_tag_set(struct radix_tree_node *node, unsigned int tag)
207 {
208         int idx;
209         for (idx = 0; idx < RADIX_TREE_TAG_LONGS; idx++) {
210                 if (node->tags[tag][idx])
211                         return 1;
212         }
213         return 0;
214 }
215
216 /*
217  *      Return the maximum key which can be store into a
218  *      radix tree with height HEIGHT.
219  */
220 static inline unsigned long radix_tree_maxindex(unsigned int height)
221 {
222         return height_to_maxindex[height];
223 }
224
225 /*
226  *      Extend a radix tree so it can store key @index.
227  */
228 static int radix_tree_extend(struct radix_tree_root *root, unsigned long index)
229 {
230         struct radix_tree_node *node;
231         unsigned int height;
232         int tag;
233
234         /* Figure out what the height should be.  */
235         height = root->height + 1;
236         while (index > radix_tree_maxindex(height))
237                 height++;
238
239         if (root->rnode == NULL) {
240                 root->height = height;
241                 goto out;
242         }
243
244         do {
245                 unsigned int newheight;
246                 if (!(node = radix_tree_node_alloc(root)))
247                         return -ENOMEM;
248
249                 /* Increase the height.  */
250                 node->slots[0] = radix_tree_indirect_to_ptr(root->rnode);
251
252                 /* Propagate the aggregated tag info into the new root */
253                 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
254                         if (root_tag_get(root, tag))
255                                 tag_set(node, tag, 0);
256                 }
257
258                 newheight = root->height+1;
259                 node->height = newheight;
260                 node->count = 1;
261                 node = radix_tree_ptr_to_indirect(node);
262                 rcu_assign_pointer(root->rnode, node);
263                 root->height = newheight;
264         } while (height > root->height);
265 out:
266         return 0;
267 }
268
269 /**
270  *      radix_tree_insert    -    insert into a radix tree
271  *      @root:          radix tree root
272  *      @index:         index key
273  *      @item:          item to insert
274  *
275  *      Insert an item into the radix tree at position @index.
276  */
277 int radix_tree_insert(struct radix_tree_root *root,
278                         unsigned long index, void *item)
279 {
280         struct radix_tree_node *node = NULL, *slot;
281         unsigned int height, shift;
282         int offset;
283         int error;
284
285         BUG_ON(radix_tree_is_indirect_ptr(item));
286
287         /* Make sure the tree is high enough.  */
288         if (index > radix_tree_maxindex(root->height)) {
289                 error = radix_tree_extend(root, index);
290                 if (error)
291                         return error;
292         }
293
294         slot = radix_tree_indirect_to_ptr(root->rnode);
295
296         height = root->height;
297         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
298
299         offset = 0;                     /* uninitialised var warning */
300         while (height > 0) {
301                 if (slot == NULL) {
302                         /* Have to add a child node.  */
303                         if (!(slot = radix_tree_node_alloc(root)))
304                                 return -ENOMEM;
305                         slot->height = height;
306                         if (node) {
307                                 rcu_assign_pointer(node->slots[offset], slot);
308                                 node->count++;
309                         } else
310                                 rcu_assign_pointer(root->rnode,
311                                         radix_tree_ptr_to_indirect(slot));
312                 }
313
314                 /* Go a level down */
315                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
316                 node = slot;
317                 slot = node->slots[offset];
318                 shift -= RADIX_TREE_MAP_SHIFT;
319                 height--;
320         }
321
322         if (slot != NULL)
323                 return -EEXIST;
324
325         if (node) {
326                 node->count++;
327                 rcu_assign_pointer(node->slots[offset], item);
328                 BUG_ON(tag_get(node, 0, offset));
329                 BUG_ON(tag_get(node, 1, offset));
330         } else {
331                 rcu_assign_pointer(root->rnode, item);
332                 BUG_ON(root_tag_get(root, 0));
333                 BUG_ON(root_tag_get(root, 1));
334         }
335
336         return 0;
337 }
338 EXPORT_SYMBOL(radix_tree_insert);
339
340 /**
341  *      radix_tree_lookup_slot    -    lookup a slot in a radix tree
342  *      @root:          radix tree root
343  *      @index:         index key
344  *
345  *      Returns:  the slot corresponding to the position @index in the
346  *      radix tree @root. This is useful for update-if-exists operations.
347  *
348  *      This function cannot be called under rcu_read_lock, it must be
349  *      excluded from writers, as must the returned slot for subsequent
350  *      use by radix_tree_deref_slot() and radix_tree_replace slot.
351  *      Caller must hold tree write locked across slot lookup and
352  *      replace.
353  */
354 void **radix_tree_lookup_slot(struct radix_tree_root *root, unsigned long index)
355 {
356         unsigned int height, shift;
357         struct radix_tree_node *node, **slot;
358
359         node = root->rnode;
360         if (node == NULL)
361                 return NULL;
362
363         if (!radix_tree_is_indirect_ptr(node)) {
364                 if (index > 0)
365                         return NULL;
366                 return (void **)&root->rnode;
367         }
368         node = radix_tree_indirect_to_ptr(node);
369
370         height = node->height;
371         if (index > radix_tree_maxindex(height))
372                 return NULL;
373
374         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
375
376         do {
377                 slot = (struct radix_tree_node **)
378                         (node->slots + ((index>>shift) & RADIX_TREE_MAP_MASK));
379                 node = *slot;
380                 if (node == NULL)
381                         return NULL;
382
383                 shift -= RADIX_TREE_MAP_SHIFT;
384                 height--;
385         } while (height > 0);
386
387         return (void **)slot;
388 }
389 EXPORT_SYMBOL(radix_tree_lookup_slot);
390
391 /**
392  *      radix_tree_lookup    -    perform lookup operation on a radix tree
393  *      @root:          radix tree root
394  *      @index:         index key
395  *
396  *      Lookup the item at the position @index in the radix tree @root.
397  *
398  *      This function can be called under rcu_read_lock, however the caller
399  *      must manage lifetimes of leaf nodes (eg. RCU may also be used to free
400  *      them safely). No RCU barriers are required to access or modify the
401  *      returned item, however.
402  */
403 void *radix_tree_lookup(struct radix_tree_root *root, unsigned long index)
404 {
405         unsigned int height, shift;
406         struct radix_tree_node *node, **slot;
407
408         node = rcu_dereference(root->rnode);
409         if (node == NULL)
410                 return NULL;
411
412         if (!radix_tree_is_indirect_ptr(node)) {
413                 if (index > 0)
414                         return NULL;
415                 return node;
416         }
417         node = radix_tree_indirect_to_ptr(node);
418
419         height = node->height;
420         if (index > radix_tree_maxindex(height))
421                 return NULL;
422
423         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
424
425         do {
426                 slot = (struct radix_tree_node **)
427                         (node->slots + ((index>>shift) & RADIX_TREE_MAP_MASK));
428                 node = rcu_dereference(*slot);
429                 if (node == NULL)
430                         return NULL;
431
432                 shift -= RADIX_TREE_MAP_SHIFT;
433                 height--;
434         } while (height > 0);
435
436         return node;
437 }
438 EXPORT_SYMBOL(radix_tree_lookup);
439
440 /**
441  *      radix_tree_tag_set - set a tag on a radix tree node
442  *      @root:          radix tree root
443  *      @index:         index key
444  *      @tag:           tag index
445  *
446  *      Set the search tag (which must be < RADIX_TREE_MAX_TAGS)
447  *      corresponding to @index in the radix tree.  From
448  *      the root all the way down to the leaf node.
449  *
450  *      Returns the address of the tagged item.   Setting a tag on a not-present
451  *      item is a bug.
452  */
453 void *radix_tree_tag_set(struct radix_tree_root *root,
454                         unsigned long index, unsigned int tag)
455 {
456         unsigned int height, shift;
457         struct radix_tree_node *slot;
458
459         height = root->height;
460         BUG_ON(index > radix_tree_maxindex(height));
461
462         slot = radix_tree_indirect_to_ptr(root->rnode);
463         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
464
465         while (height > 0) {
466                 int offset;
467
468                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
469                 if (!tag_get(slot, tag, offset))
470                         tag_set(slot, tag, offset);
471                 slot = slot->slots[offset];
472                 BUG_ON(slot == NULL);
473                 shift -= RADIX_TREE_MAP_SHIFT;
474                 height--;
475         }
476
477         /* set the root's tag bit */
478         if (slot && !root_tag_get(root, tag))
479                 root_tag_set(root, tag);
480
481         return slot;
482 }
483 EXPORT_SYMBOL(radix_tree_tag_set);
484
485 /**
486  *      radix_tree_tag_clear - clear a tag on a radix tree node
487  *      @root:          radix tree root
488  *      @index:         index key
489  *      @tag:           tag index
490  *
491  *      Clear the search tag (which must be < RADIX_TREE_MAX_TAGS)
492  *      corresponding to @index in the radix tree.  If
493  *      this causes the leaf node to have no tags set then clear the tag in the
494  *      next-to-leaf node, etc.
495  *
496  *      Returns the address of the tagged item on success, else NULL.  ie:
497  *      has the same return value and semantics as radix_tree_lookup().
498  */
499 void *radix_tree_tag_clear(struct radix_tree_root *root,
500                         unsigned long index, unsigned int tag)
501 {
502         /*
503          * The radix tree path needs to be one longer than the maximum path
504          * since the "list" is null terminated.
505          */
506         struct radix_tree_path path[RADIX_TREE_MAX_PATH + 1], *pathp = path;
507         struct radix_tree_node *slot = NULL;
508         unsigned int height, shift;
509
510         height = root->height;
511         if (index > radix_tree_maxindex(height))
512                 goto out;
513
514         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
515         pathp->node = NULL;
516         slot = radix_tree_indirect_to_ptr(root->rnode);
517
518         while (height > 0) {
519                 int offset;
520
521                 if (slot == NULL)
522                         goto out;
523
524                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
525                 pathp[1].offset = offset;
526                 pathp[1].node = slot;
527                 slot = slot->slots[offset];
528                 pathp++;
529                 shift -= RADIX_TREE_MAP_SHIFT;
530                 height--;
531         }
532
533         if (slot == NULL)
534                 goto out;
535
536         while (pathp->node) {
537                 if (!tag_get(pathp->node, tag, pathp->offset))
538                         goto out;
539                 tag_clear(pathp->node, tag, pathp->offset);
540                 if (any_tag_set(pathp->node, tag))
541                         goto out;
542                 pathp--;
543         }
544
545         /* clear the root's tag bit */
546         if (root_tag_get(root, tag))
547                 root_tag_clear(root, tag);
548
549 out:
550         return slot;
551 }
552 EXPORT_SYMBOL(radix_tree_tag_clear);
553
554 #ifndef __KERNEL__      /* Only the test harness uses this at present */
555 /**
556  * radix_tree_tag_get - get a tag on a radix tree node
557  * @root:               radix tree root
558  * @index:              index key
559  * @tag:                tag index (< RADIX_TREE_MAX_TAGS)
560  *
561  * Return values:
562  *
563  *  0: tag not present or not set
564  *  1: tag set
565  */
566 int radix_tree_tag_get(struct radix_tree_root *root,
567                         unsigned long index, unsigned int tag)
568 {
569         unsigned int height, shift;
570         struct radix_tree_node *node;
571         int saw_unset_tag = 0;
572
573         /* check the root's tag bit */
574         if (!root_tag_get(root, tag))
575                 return 0;
576
577         node = rcu_dereference(root->rnode);
578         if (node == NULL)
579                 return 0;
580
581         if (!radix_tree_is_indirect_ptr(node))
582                 return (index == 0);
583         node = radix_tree_indirect_to_ptr(node);
584
585         height = node->height;
586         if (index > radix_tree_maxindex(height))
587                 return 0;
588
589         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
590
591         for ( ; ; ) {
592                 int offset;
593
594                 if (node == NULL)
595                         return 0;
596
597                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
598
599                 /*
600                  * This is just a debug check.  Later, we can bale as soon as
601                  * we see an unset tag.
602                  */
603                 if (!tag_get(node, tag, offset))
604                         saw_unset_tag = 1;
605                 if (height == 1) {
606                         int ret = tag_get(node, tag, offset);
607
608                         BUG_ON(ret && saw_unset_tag);
609                         return !!ret;
610                 }
611                 node = rcu_dereference(node->slots[offset]);
612                 shift -= RADIX_TREE_MAP_SHIFT;
613                 height--;
614         }
615 }
616 EXPORT_SYMBOL(radix_tree_tag_get);
617 #endif
618
619 /**
620  *      radix_tree_next_hole    -    find the next hole (not-present entry)
621  *      @root:          tree root
622  *      @index:         index key
623  *      @max_scan:      maximum range to search
624  *
625  *      Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the lowest
626  *      indexed hole.
627  *
628  *      Returns: the index of the hole if found, otherwise returns an index
629  *      outside of the set specified (in which case 'return - index >= max_scan'
630  *      will be true).
631  *
632  *      radix_tree_next_hole may be called under rcu_read_lock. However, like
633  *      radix_tree_gang_lookup, this will not atomically search a snapshot of the
634  *      tree at a single point in time. For example, if a hole is created at index
635  *      5, then subsequently a hole is created at index 10, radix_tree_next_hole
636  *      covering both indexes may return 10 if called under rcu_read_lock.
637  */
638 unsigned long radix_tree_next_hole(struct radix_tree_root *root,
639                                 unsigned long index, unsigned long max_scan)
640 {
641         unsigned long i;
642
643         for (i = 0; i < max_scan; i++) {
644                 if (!radix_tree_lookup(root, index))
645                         break;
646                 index++;
647                 if (index == 0)
648                         break;
649         }
650
651         return index;
652 }
653 EXPORT_SYMBOL(radix_tree_next_hole);
654
655 static unsigned int
656 __lookup(struct radix_tree_node *slot, void **results, unsigned long index,
657         unsigned int max_items, unsigned long *next_index)
658 {
659         unsigned int nr_found = 0;
660         unsigned int shift, height;
661         unsigned long i;
662
663         height = slot->height;
664         if (height == 0)
665                 goto out;
666         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
667
668         for ( ; height > 1; height--) {
669                 i = (index >> shift) & RADIX_TREE_MAP_MASK;
670                 for (;;) {
671                         if (slot->slots[i] != NULL)
672                                 break;
673                         index &= ~((1UL << shift) - 1);
674                         index += 1UL << shift;
675                         if (index == 0)
676                                 goto out;       /* 32-bit wraparound */
677                         i++;
678                         if (i == RADIX_TREE_MAP_SIZE)
679                                 goto out;
680                 }
681
682                 shift -= RADIX_TREE_MAP_SHIFT;
683                 slot = rcu_dereference(slot->slots[i]);
684                 if (slot == NULL)
685                         goto out;
686         }
687
688         /* Bottom level: grab some items */
689         for (i = index & RADIX_TREE_MAP_MASK; i < RADIX_TREE_MAP_SIZE; i++) {
690                 struct radix_tree_node *node;
691                 index++;
692                 node = slot->slots[i];
693                 if (node) {
694                         results[nr_found++] = rcu_dereference(node);
695                         if (nr_found == max_items)
696                                 goto out;
697                 }
698         }
699 out:
700         *next_index = index;
701         return nr_found;
702 }
703
704 /**
705  *      radix_tree_gang_lookup - perform multiple lookup on a radix tree
706  *      @root:          radix tree root
707  *      @results:       where the results of the lookup are placed
708  *      @first_index:   start the lookup from this key
709  *      @max_items:     place up to this many items at *results
710  *
711  *      Performs an index-ascending scan of the tree for present items.  Places
712  *      them at *@results and returns the number of items which were placed at
713  *      *@results.
714  *
715  *      The implementation is naive.
716  *
717  *      Like radix_tree_lookup, radix_tree_gang_lookup may be called under
718  *      rcu_read_lock. In this case, rather than the returned results being
719  *      an atomic snapshot of the tree at a single point in time, the semantics
720  *      of an RCU protected gang lookup are as though multiple radix_tree_lookups
721  *      have been issued in individual locks, and results stored in 'results'.
722  */
723 unsigned int
724 radix_tree_gang_lookup(struct radix_tree_root *root, void **results,
725                         unsigned long first_index, unsigned int max_items)
726 {
727         unsigned long max_index;
728         struct radix_tree_node *node;
729         unsigned long cur_index = first_index;
730         unsigned int ret;
731
732         node = rcu_dereference(root->rnode);
733         if (!node)
734                 return 0;
735
736         if (!radix_tree_is_indirect_ptr(node)) {
737                 if (first_index > 0)
738                         return 0;
739                 results[0] = node;
740                 return 1;
741         }
742         node = radix_tree_indirect_to_ptr(node);
743
744         max_index = radix_tree_maxindex(node->height);
745
746         ret = 0;
747         while (ret < max_items) {
748                 unsigned int nr_found;
749                 unsigned long next_index;       /* Index of next search */
750
751                 if (cur_index > max_index)
752                         break;
753                 nr_found = __lookup(node, results + ret, cur_index,
754                                         max_items - ret, &next_index);
755                 ret += nr_found;
756                 if (next_index == 0)
757                         break;
758                 cur_index = next_index;
759         }
760
761         return ret;
762 }
763 EXPORT_SYMBOL(radix_tree_gang_lookup);
764
765 /*
766  * FIXME: the two tag_get()s here should use find_next_bit() instead of
767  * open-coding the search.
768  */
769 static unsigned int
770 __lookup_tag(struct radix_tree_node *slot, void **results, unsigned long index,
771         unsigned int max_items, unsigned long *next_index, unsigned int tag)
772 {
773         unsigned int nr_found = 0;
774         unsigned int shift, height;
775
776         height = slot->height;
777         if (height == 0)
778                 goto out;
779         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
780
781         while (height > 0) {
782                 unsigned long i = (index >> shift) & RADIX_TREE_MAP_MASK ;
783
784                 for (;;) {
785                         if (tag_get(slot, tag, i))
786                                 break;
787                         index &= ~((1UL << shift) - 1);
788                         index += 1UL << shift;
789                         if (index == 0)
790                                 goto out;       /* 32-bit wraparound */
791                         i++;
792                         if (i == RADIX_TREE_MAP_SIZE)
793                                 goto out;
794                 }
795                 height--;
796                 if (height == 0) {      /* Bottom level: grab some items */
797                         unsigned long j = index & RADIX_TREE_MAP_MASK;
798
799                         for ( ; j < RADIX_TREE_MAP_SIZE; j++) {
800                                 struct radix_tree_node *node;
801                                 index++;
802                                 if (!tag_get(slot, tag, j))
803                                         continue;
804                                 node = slot->slots[j];
805                                 /*
806                                  * Even though the tag was found set, we need to
807                                  * recheck that we have a non-NULL node, because
808                                  * if this lookup is lockless, it may have been
809                                  * subsequently deleted.
810                                  *
811                                  * Similar care must be taken in any place that
812                                  * lookup ->slots[x] without a lock (ie. can't
813                                  * rely on its value remaining the same).
814                                  */
815                                 if (node) {
816                                         node = rcu_dereference(node);
817                                         results[nr_found++] = node;
818                                         if (nr_found == max_items)
819                                                 goto out;
820                                 }
821                         }
822                 }
823                 shift -= RADIX_TREE_MAP_SHIFT;
824                 slot = rcu_dereference(slot->slots[i]);
825                 if (slot == NULL)
826                         break;
827         }
828 out:
829         *next_index = index;
830         return nr_found;
831 }
832
833 /**
834  *      radix_tree_gang_lookup_tag - perform multiple lookup on a radix tree
835  *                                   based on a tag
836  *      @root:          radix tree root
837  *      @results:       where the results of the lookup are placed
838  *      @first_index:   start the lookup from this key
839  *      @max_items:     place up to this many items at *results
840  *      @tag:           the tag index (< RADIX_TREE_MAX_TAGS)
841  *
842  *      Performs an index-ascending scan of the tree for present items which
843  *      have the tag indexed by @tag set.  Places the items at *@results and
844  *      returns the number of items which were placed at *@results.
845  */
846 unsigned int
847 radix_tree_gang_lookup_tag(struct radix_tree_root *root, void **results,
848                 unsigned long first_index, unsigned int max_items,
849                 unsigned int tag)
850 {
851         struct radix_tree_node *node;
852         unsigned long max_index;
853         unsigned long cur_index = first_index;
854         unsigned int ret;
855
856         /* check the root's tag bit */
857         if (!root_tag_get(root, tag))
858                 return 0;
859
860         node = rcu_dereference(root->rnode);
861         if (!node)
862                 return 0;
863
864         if (!radix_tree_is_indirect_ptr(node)) {
865                 if (first_index > 0)
866                         return 0;
867                 results[0] = node;
868                 return 1;
869         }
870         node = radix_tree_indirect_to_ptr(node);
871
872         max_index = radix_tree_maxindex(node->height);
873
874         ret = 0;
875         while (ret < max_items) {
876                 unsigned int nr_found;
877                 unsigned long next_index;       /* Index of next search */
878
879                 if (cur_index > max_index)
880                         break;
881                 nr_found = __lookup_tag(node, results + ret, cur_index,
882                                         max_items - ret, &next_index, tag);
883                 ret += nr_found;
884                 if (next_index == 0)
885                         break;
886                 cur_index = next_index;
887         }
888
889         return ret;
890 }
891 EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
892
893 /**
894  *      radix_tree_shrink    -    shrink height of a radix tree to minimal
895  *      @root           radix tree root
896  */
897 static inline void radix_tree_shrink(struct radix_tree_root *root)
898 {
899         /* try to shrink tree height */
900         while (root->height > 0) {
901                 struct radix_tree_node *to_free = root->rnode;
902                 void *newptr;
903
904                 BUG_ON(!radix_tree_is_indirect_ptr(to_free));
905                 to_free = radix_tree_indirect_to_ptr(to_free);
906
907                 /*
908                  * The candidate node has more than one child, or its child
909                  * is not at the leftmost slot, we cannot shrink.
910                  */
911                 if (to_free->count != 1)
912                         break;
913                 if (!to_free->slots[0])
914                         break;
915
916                 /*
917                  * We don't need rcu_assign_pointer(), since we are simply
918                  * moving the node from one part of the tree to another. If
919                  * it was safe to dereference the old pointer to it
920                  * (to_free->slots[0]), it will be safe to dereference the new
921                  * one (root->rnode).
922                  */
923                 newptr = to_free->slots[0];
924                 if (root->height > 1)
925                         newptr = radix_tree_ptr_to_indirect(newptr);
926                 root->rnode = newptr;
927                 root->height--;
928                 /* must only free zeroed nodes into the slab */
929                 tag_clear(to_free, 0, 0);
930                 tag_clear(to_free, 1, 0);
931                 to_free->slots[0] = NULL;
932                 to_free->count = 0;
933                 radix_tree_node_free(to_free);
934         }
935 }
936
937 /**
938  *      radix_tree_delete    -    delete an item from a radix tree
939  *      @root:          radix tree root
940  *      @index:         index key
941  *
942  *      Remove the item at @index from the radix tree rooted at @root.
943  *
944  *      Returns the address of the deleted item, or NULL if it was not present.
945  */
946 void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
947 {
948         /*
949          * The radix tree path needs to be one longer than the maximum path
950          * since the "list" is null terminated.
951          */
952         struct radix_tree_path path[RADIX_TREE_MAX_PATH + 1], *pathp = path;
953         struct radix_tree_node *slot = NULL;
954         struct radix_tree_node *to_free;
955         unsigned int height, shift;
956         int tag;
957         int offset;
958
959         height = root->height;
960         if (index > radix_tree_maxindex(height))
961                 goto out;
962
963         slot = root->rnode;
964         if (height == 0) {
965                 root_tag_clear_all(root);
966                 root->rnode = NULL;
967                 goto out;
968         }
969         slot = radix_tree_indirect_to_ptr(slot);
970
971         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
972         pathp->node = NULL;
973
974         do {
975                 if (slot == NULL)
976                         goto out;
977
978                 pathp++;
979                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
980                 pathp->offset = offset;
981                 pathp->node = slot;
982                 slot = slot->slots[offset];
983                 shift -= RADIX_TREE_MAP_SHIFT;
984                 height--;
985         } while (height > 0);
986
987         if (slot == NULL)
988                 goto out;
989
990         /*
991          * Clear all tags associated with the just-deleted item
992          */
993         for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
994                 if (tag_get(pathp->node, tag, pathp->offset))
995                         radix_tree_tag_clear(root, index, tag);
996         }
997
998         to_free = NULL;
999         /* Now free the nodes we do not need anymore */
1000         while (pathp->node) {
1001                 pathp->node->slots[pathp->offset] = NULL;
1002                 pathp->node->count--;
1003                 /*
1004                  * Queue the node for deferred freeing after the
1005                  * last reference to it disappears (set NULL, above).
1006                  */
1007                 if (to_free)
1008                         radix_tree_node_free(to_free);
1009
1010                 if (pathp->node->count) {
1011                         if (pathp->node ==
1012                                         radix_tree_indirect_to_ptr(root->rnode))
1013                                 radix_tree_shrink(root);
1014                         goto out;
1015                 }
1016
1017                 /* Node with zero slots in use so free it */
1018                 to_free = pathp->node;
1019                 pathp--;
1020
1021         }
1022         root_tag_clear_all(root);
1023         root->height = 0;
1024         root->rnode = NULL;
1025         if (to_free)
1026                 radix_tree_node_free(to_free);
1027
1028 out:
1029         return slot;
1030 }
1031 EXPORT_SYMBOL(radix_tree_delete);
1032
1033 /**
1034  *      radix_tree_tagged - test whether any items in the tree are tagged
1035  *      @root:          radix tree root
1036  *      @tag:           tag to test
1037  */
1038 int radix_tree_tagged(struct radix_tree_root *root, unsigned int tag)
1039 {
1040         return root_tag_get(root, tag);
1041 }
1042 EXPORT_SYMBOL(radix_tree_tagged);
1043
1044 static void
1045 radix_tree_node_ctor(struct kmem_cache *cachep, void *node)
1046 {
1047         memset(node, 0, sizeof(struct radix_tree_node));
1048 }
1049
1050 static __init unsigned long __maxindex(unsigned int height)
1051 {
1052         unsigned int width = height * RADIX_TREE_MAP_SHIFT;
1053         int shift = RADIX_TREE_INDEX_BITS - width;
1054
1055         if (shift < 0)
1056                 return ~0UL;
1057         if (shift >= BITS_PER_LONG)
1058                 return 0UL;
1059         return ~0UL >> shift;
1060 }
1061
1062 static __init void radix_tree_init_maxindex(void)
1063 {
1064         unsigned int i;
1065
1066         for (i = 0; i < ARRAY_SIZE(height_to_maxindex); i++)
1067                 height_to_maxindex[i] = __maxindex(i);
1068 }
1069
1070 static int radix_tree_callback(struct notifier_block *nfb,
1071                             unsigned long action,
1072                             void *hcpu)
1073 {
1074        int cpu = (long)hcpu;
1075        struct radix_tree_preload *rtp;
1076
1077        /* Free per-cpu pool of perloaded nodes */
1078        if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
1079                rtp = &per_cpu(radix_tree_preloads, cpu);
1080                while (rtp->nr) {
1081                        kmem_cache_free(radix_tree_node_cachep,
1082                                        rtp->nodes[rtp->nr-1]);
1083                        rtp->nodes[rtp->nr-1] = NULL;
1084                        rtp->nr--;
1085                }
1086        }
1087        return NOTIFY_OK;
1088 }
1089
1090 void __init radix_tree_init(void)
1091 {
1092         radix_tree_node_cachep = kmem_cache_create("radix_tree_node",
1093                         sizeof(struct radix_tree_node), 0,
1094                         SLAB_PANIC, radix_tree_node_ctor);
1095         radix_tree_init_maxindex();
1096         hotcpu_notifier(radix_tree_callback, 0);
1097 }