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_direct_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_direct_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                 rcu_assign_pointer(root->rnode, node);
 255                 root->height = newheight;
 256         } while (height > root->height);
 257 out:
 258         return 0;
 259 }
 260
 261 /**
 262  *      radix_tree_insert    -    insert into a radix tree
 263  *      @root:          radix tree root
 264  *      @index:         index key
 265  *      @item:          item to insert
 266  *
 267  *      Insert an item into the radix tree at position @index.
 268  */
 269 int radix_tree_insert(struct radix_tree_root *root,
 270                         unsigned long index, void *item)
 271 {
 272         struct radix_tree_node *node = NULL, *slot;
 273         unsigned int height, shift;
 274         int offset;
 275         int error;
 276
 277         BUG_ON(radix_tree_is_direct_ptr(item));
 278
 279         /* Make sure the tree is high enough.  */
 280         if (index > radix_tree_maxindex(root->height)) {
 281                 error = radix_tree_extend(root, index);
 282                 if (error)
 283                         return error;
 284         }
 285
 286         slot = root->rnode;
 287         height = root->height;
 288         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
 289
 290         offset = 0;                     /* uninitialised var warning */
 291         while (height > 0) {
 292                 if (slot == NULL) {
 293                         /* Have to add a child node.  */
 294                         if (!(slot = radix_tree_node_alloc(root)))
 295                                 return -ENOMEM;
 296                         slot->height = height;
 297                         if (node) {
 298                                 rcu_assign_pointer(node->slots[offset], slot);
 299                                 node->count++;
 300                         } else
 301                                 rcu_assign_pointer(root->rnode, slot);
 302                 }
 303
 304                 /* Go a level down */
 305                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
 306                 node = slot;
 307                 slot = node->slots[offset];
 308                 shift -= RADIX_TREE_MAP_SHIFT;
 309                 height--;
 310         }
 311
 312         if (slot != NULL)
 313                 return -EEXIST;
 314
 315         if (node) {
 316                 node->count++;
 317                 rcu_assign_pointer(node->slots[offset], item);
 318                 BUG_ON(tag_get(node, 0, offset));
 319                 BUG_ON(tag_get(node, 1, offset));
 320         } else {
 321                 rcu_assign_pointer(root->rnode, radix_tree_ptr_to_direct(item));
 322                 BUG_ON(root_tag_get(root, 0));
 323                 BUG_ON(root_tag_get(root, 1));
 324         }
 325
 326         return 0;
 327 }
 328 EXPORT_SYMBOL(radix_tree_insert);
 329
 330 /**
 331  *      radix_tree_lookup_slot    -    lookup a slot in a radix tree
 332  *      @root:          radix tree root
 333  *      @index:         index key
 334  *
 335  *      Returns:  the slot corresponding to the position @index in the
 336  *      radix tree @root. This is useful for update-if-exists operations.
 337  *
 338  *      This function cannot be called under rcu_read_lock, it must be
 339  *      excluded from writers, as must the returned slot for subsequent
 340  *      use by radix_tree_deref_slot() and radix_tree_replace slot.
 341  *      Caller must hold tree write locked across slot lookup and
 342  *      replace.
 343  */
 344 void **radix_tree_lookup_slot(struct radix_tree_root *root, unsigned long index)
 345 {
 346         unsigned int height, shift;
 347         struct radix_tree_node *node, **slot;
 348
 349         node = root->rnode;
 350         if (node == NULL)
 351                 return NULL;
 352
 353         if (radix_tree_is_direct_ptr(node)) {
 354                 if (index > 0)
 355                         return NULL;
 356                 return (void **)&root->rnode;
 357         }
 358
 359         height = node->height;
 360         if (index > radix_tree_maxindex(height))
 361                 return NULL;
 362
 363         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
 364
 365         do {
 366                 slot = (struct radix_tree_node **)
 367                         (node->slots + ((index>>shift) & RADIX_TREE_MAP_MASK));
 368                 node = *slot;
 369                 if (node == NULL)
 370                         return NULL;
 371
 372                 shift -= RADIX_TREE_MAP_SHIFT;
 373                 height--;
 374         } while (height > 0);
 375
 376         return (void **)slot;
 377 }
 378 EXPORT_SYMBOL(radix_tree_lookup_slot);
 379
 380 /**
 381  *      radix_tree_lookup    -    perform lookup operation on a radix tree
 382  *      @root:          radix tree root
 383  *      @index:         index key
 384  *
 385  *      Lookup the item at the position @index in the radix tree @root.
 386  *
 387  *      This function can be called under rcu_read_lock, however the caller
 388  *      must manage lifetimes of leaf nodes (eg. RCU may also be used to free
 389  *      them safely). No RCU barriers are required to access or modify the
 390  *      returned item, however.
 391  */
 392 void *radix_tree_lookup(struct radix_tree_root *root, unsigned long index)
 393 {
 394         unsigned int height, shift;
 395         struct radix_tree_node *node, **slot;
 396
 397         node = rcu_dereference(root->rnode);
 398         if (node == NULL)
 399                 return NULL;
 400
 401         if (radix_tree_is_direct_ptr(node)) {
 402                 if (index > 0)
 403                         return NULL;
 404                 return radix_tree_direct_to_ptr(node);
 405         }
 406
 407         height = node->height;
 408         if (index > radix_tree_maxindex(height))
 409                 return NULL;
 410
 411         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
 412
 413         do {
 414                 slot = (struct radix_tree_node **)
 415                         (node->slots + ((index>>shift) & RADIX_TREE_MAP_MASK));
 416                 node = rcu_dereference(*slot);
 417                 if (node == NULL)
 418                         return NULL;
 419
 420                 shift -= RADIX_TREE_MAP_SHIFT;
 421                 height--;
 422         } while (height > 0);
 423
 424         return node;
 425 }
 426 EXPORT_SYMBOL(radix_tree_lookup);
 427
 428 /**
 429  *      radix_tree_tag_set - set a tag on a radix tree node
 430  *      @root:          radix tree root
 431  *      @index:         index key
 432  *      @tag:           tag index
 433  *
 434  *      Set the search tag (which must be < RADIX_TREE_MAX_TAGS)
 435  *      corresponding to @index in the radix tree.  From
 436  *      the root all the way down to the leaf node.
 437  *
 438  *      Returns the address of the tagged item.   Setting a tag on a not-present
 439  *      item is a bug.
 440  */
 441 void *radix_tree_tag_set(struct radix_tree_root *root,
 442                         unsigned long index, unsigned int tag)
 443 {
 444         unsigned int height, shift;
 445         struct radix_tree_node *slot;
 446
 447         height = root->height;
 448         BUG_ON(index > radix_tree_maxindex(height));
 449
 450         slot = root->rnode;
 451         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
 452
 453         while (height > 0) {
 454                 int offset;
 455
 456                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
 457                 if (!tag_get(slot, tag, offset))
 458                         tag_set(slot, tag, offset);
 459                 slot = slot->slots[offset];
 460                 BUG_ON(slot == NULL);
 461                 shift -= RADIX_TREE_MAP_SHIFT;
 462                 height--;
 463         }
 464
 465         /* set the root's tag bit */
 466         if (slot && !root_tag_get(root, tag))
 467                 root_tag_set(root, tag);
 468
 469         return slot;
 470 }
 471 EXPORT_SYMBOL(radix_tree_tag_set);
 472
 473 /**
 474  *      radix_tree_tag_clear - clear a tag on a radix tree node
 475  *      @root:          radix tree root
 476  *      @index:         index key
 477  *      @tag:           tag index
 478  *
 479  *      Clear the search tag (which must be < RADIX_TREE_MAX_TAGS)
 480  *      corresponding to @index in the radix tree.  If
 481  *      this causes the leaf node to have no tags set then clear the tag in the
 482  *      next-to-leaf node, etc.
 483  *
 484  *      Returns the address of the tagged item on success, else NULL.  ie:
 485  *      has the same return value and semantics as radix_tree_lookup().
 486  */
 487 void *radix_tree_tag_clear(struct radix_tree_root *root,
 488                         unsigned long index, unsigned int tag)
 489 {
 490         struct radix_tree_path path[RADIX_TREE_MAX_PATH], *pathp = path;
 491         struct radix_tree_node *slot = NULL;
 492         unsigned int height, shift;
 493
 494         height = root->height;
 495         if (index > radix_tree_maxindex(height))
 496                 goto out;
 497
 498         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
 499         pathp->node = NULL;
 500         slot = root->rnode;
 501
 502         while (height > 0) {
 503                 int offset;
 504
 505                 if (slot == NULL)
 506                         goto out;
 507
 508                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
 509                 pathp[1].offset = offset;
 510                 pathp[1].node = slot;
 511                 slot = slot->slots[offset];
 512                 pathp++;
 513                 shift -= RADIX_TREE_MAP_SHIFT;
 514                 height--;
 515         }
 516
 517         if (slot == NULL)
 518                 goto out;
 519
 520         while (pathp->node) {
 521                 if (!tag_get(pathp->node, tag, pathp->offset))
 522                         goto out;
 523                 tag_clear(pathp->node, tag, pathp->offset);
 524                 if (any_tag_set(pathp->node, tag))
 525                         goto out;
 526                 pathp--;
 527         }
 528
 529         /* clear the root's tag bit */
 530         if (root_tag_get(root, tag))
 531                 root_tag_clear(root, tag);
 532
 533 out:
 534         return slot;
 535 }
 536 EXPORT_SYMBOL(radix_tree_tag_clear);
 537
 538 #ifndef __KERNEL__      /* Only the test harness uses this at present */
 539 /**
 540  * radix_tree_tag_get - get a tag on a radix tree node
 541  * @root:               radix tree root
 542  * @index:              index key
 543  * @tag:                tag index (< RADIX_TREE_MAX_TAGS)
 544  *
 545  * Return values:
 546  *
 547  *  0: tag not present or not set
 548  *  1: tag set
 549  */
 550 int radix_tree_tag_get(struct radix_tree_root *root,
 551                         unsigned long index, unsigned int tag)
 552 {
 553         unsigned int height, shift;
 554         struct radix_tree_node *node;
 555         int saw_unset_tag = 0;
 556
 557         /* check the root's tag bit */
 558         if (!root_tag_get(root, tag))
 559                 return 0;
 560
 561         node = rcu_dereference(root->rnode);
 562         if (node == NULL)
 563                 return 0;
 564
 565         if (radix_tree_is_direct_ptr(node))
 566                 return (index == 0);
 567
 568         height = node->height;
 569         if (index > radix_tree_maxindex(height))
 570                 return 0;
 571
 572         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
 573
 574         for ( ; ; ) {
 575                 int offset;
 576
 577                 if (node == NULL)
 578                         return 0;
 579
 580                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
 581
 582                 /*
 583                  * This is just a debug check.  Later, we can bale as soon as
 584                  * we see an unset tag.
 585                  */
 586                 if (!tag_get(node, tag, offset))
 587                         saw_unset_tag = 1;
 588                 if (height == 1) {
 589                         int ret = tag_get(node, tag, offset);
 590
 591                         BUG_ON(ret && saw_unset_tag);
 592                         return !!ret;
 593                 }
 594                 node = rcu_dereference(node->slots[offset]);
 595                 shift -= RADIX_TREE_MAP_SHIFT;
 596                 height--;
 597         }
 598 }
 599 EXPORT_SYMBOL(radix_tree_tag_get);
 600 #endif
 601
 602 /**
 603  *      radix_tree_next_hole    -    find the next hole (not-present entry)
 604  *      @root:          tree root
 605  *      @index:         index key
 606  *      @max_scan:      maximum range to search
 607  *
 608  *      Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the lowest
 609  *      indexed hole.
 610  *
 611  *      Returns: the index of the hole if found, otherwise returns an index
 612  *      outside of the set specified (in which case 'return - index >= max_scan'
 613  *      will be true).
 614  *
 615  *      radix_tree_next_hole may be called under rcu_read_lock. However, like
 616  *      radix_tree_gang_lookup, this will not atomically search a snapshot of the
 617  *      tree at a single point in time. For example, if a hole is created at index
 618  *      5, then subsequently a hole is created at index 10, radix_tree_next_hole
 619  *      covering both indexes may return 10 if called under rcu_read_lock.
 620  */
 621 unsigned long radix_tree_next_hole(struct radix_tree_root *root,
 622                                 unsigned long index, unsigned long max_scan)
 623 {
 624         unsigned long i;
 625
 626         for (i = 0; i < max_scan; i++) {
 627                 if (!radix_tree_lookup(root, index))
 628                         break;
 629                 index++;
 630                 if (index == 0)
 631                         break;
 632         }
 633
 634         return index;
 635 }
 636 EXPORT_SYMBOL(radix_tree_next_hole);
 637
 638 static unsigned int
 639 __lookup(struct radix_tree_node *slot, void **results, unsigned long index,
 640         unsigned int max_items, unsigned long *next_index)
 641 {
 642         unsigned int nr_found = 0;
 643         unsigned int shift, height;
 644         unsigned long i;
 645
 646         height = slot->height;
 647         if (height == 0)
 648                 goto out;
 649         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
 650
 651         for ( ; height > 1; height--) {
 652                 i = (index >> shift) & RADIX_TREE_MAP_MASK;
 653                 for (;;) {
 654                         if (slot->slots[i] != NULL)
 655                                 break;
 656                         index &= ~((1UL << shift) - 1);
 657                         index += 1UL << shift;
 658                         if (index == 0)
 659                                 goto out;       /* 32-bit wraparound */
 660                         i++;
 661                         if (i == RADIX_TREE_MAP_SIZE)
 662                                 goto out;
 663                 }
 664
 665                 shift -= RADIX_TREE_MAP_SHIFT;
 666                 slot = rcu_dereference(slot->slots[i]);
 667                 if (slot == NULL)
 668                         goto out;
 669         }
 670
 671         /* Bottom level: grab some items */
 672         for (i = index & RADIX_TREE_MAP_MASK; i < RADIX_TREE_MAP_SIZE; i++) {
 673                 struct radix_tree_node *node;
 674                 index++;
 675                 node = slot->slots[i];
 676                 if (node) {
 677                         results[nr_found++] = rcu_dereference(node);
 678                         if (nr_found == max_items)
 679                                 goto out;
 680                 }
 681         }
 682 out:
 683         *next_index = index;
 684         return nr_found;
 685 }
 686
 687 /**
 688  *      radix_tree_gang_lookup - perform multiple lookup on a radix tree
 689  *      @root:          radix tree root
 690  *      @results:       where the results of the lookup are placed
 691  *      @first_index:   start the lookup from this key
 692  *      @max_items:     place up to this many items at *results
 693  *
 694  *      Performs an index-ascending scan of the tree for present items.  Places
 695  *      them at *@results and returns the number of items which were placed at
 696  *      *@results.
 697  *
 698  *      The implementation is naive.
 699  *
 700  *      Like radix_tree_lookup, radix_tree_gang_lookup may be called under
 701  *      rcu_read_lock. In this case, rather than the returned results being
 702  *      an atomic snapshot of the tree at a single point in time, the semantics
 703  *      of an RCU protected gang lookup are as though multiple radix_tree_lookups
 704  *      have been issued in individual locks, and results stored in 'results'.
 705  */
 706 unsigned int
 707 radix_tree_gang_lookup(struct radix_tree_root *root, void **results,
 708                         unsigned long first_index, unsigned int max_items)
 709 {
 710         unsigned long max_index;
 711         struct radix_tree_node *node;
 712         unsigned long cur_index = first_index;
 713         unsigned int ret;
 714
 715         node = rcu_dereference(root->rnode);
 716         if (!node)
 717                 return 0;
 718
 719         if (radix_tree_is_direct_ptr(node)) {
 720                 if (first_index > 0)
 721                         return 0;
 722                 node = radix_tree_direct_to_ptr(node);
 723                 results[0] = rcu_dereference(node);
 724                 return 1;
 725         }
 726
 727         max_index = radix_tree_maxindex(node->height);
 728
 729         ret = 0;
 730         while (ret < max_items) {
 731                 unsigned int nr_found;
 732                 unsigned long next_index;       /* Index of next search */
 733
 734                 if (cur_index > max_index)
 735                         break;
 736                 nr_found = __lookup(node, results + ret, cur_index,
 737                                         max_items - ret, &next_index);
 738                 ret += nr_found;
 739                 if (next_index == 0)
 740                         break;
 741                 cur_index = next_index;
 742         }
 743
 744         return ret;
 745 }
 746 EXPORT_SYMBOL(radix_tree_gang_lookup);
 747
 748 /*
 749  * FIXME: the two tag_get()s here should use find_next_bit() instead of
 750  * open-coding the search.
 751  */
 752 static unsigned int
 753 __lookup_tag(struct radix_tree_node *slot, void **results, unsigned long index,
 754         unsigned int max_items, unsigned long *next_index, unsigned int tag)
 755 {
 756         unsigned int nr_found = 0;
 757         unsigned int shift, height;
 758
 759         height = slot->height;
 760         if (height == 0)
 761                 goto out;
 762         shift = (height-1) * RADIX_TREE_MAP_SHIFT;
 763
 764         while (height > 0) {
 765                 unsigned long i = (index >> shift) & RADIX_TREE_MAP_MASK ;
 766
 767                 for (;;) {
 768                         if (tag_get(slot, tag, i))
 769                                 break;
 770                         index &= ~((1UL << shift) - 1);
 771                         index += 1UL << shift;
 772                         if (index == 0)
 773                                 goto out;       /* 32-bit wraparound */
 774                         i++;
 775                         if (i == RADIX_TREE_MAP_SIZE)
 776                                 goto out;
 777                 }
 778                 height--;
 779                 if (height == 0) {      /* Bottom level: grab some items */
 780                         unsigned long j = index & RADIX_TREE_MAP_MASK;
 781
 782                         for ( ; j < RADIX_TREE_MAP_SIZE; j++) {
 783                                 struct radix_tree_node *node;
 784                                 index++;
 785                                 if (!tag_get(slot, tag, j))
 786                                         continue;
 787                                 node = slot->slots[j];
 788                                 /*
 789                                  * Even though the tag was found set, we need to
 790                                  * recheck that we have a non-NULL node, because
 791                                  * if this lookup is lockless, it may have been
 792                                  * subsequently deleted.
 793                                  *
 794                                  * Similar care must be taken in any place that
 795                                  * lookup ->slots[x] without a lock (ie. can't
 796                                  * rely on its value remaining the same).
 797                                  */
 798                                 if (node) {
 799                                         node = rcu_dereference(node);
 800                                         results[nr_found++] = node;
 801                                         if (nr_found == max_items)
 802                                                 goto out;
 803                                 }
 804                         }
 805                 }
 806                 shift -= RADIX_TREE_MAP_SHIFT;
 807                 slot = rcu_dereference(slot->slots[i]);
 808                 if (slot == NULL)
 809                         break;
 810         }
 811 out:
 812         *next_index = index;
 813         return nr_found;
 814 }
 815
 816 /**
 817  *      radix_tree_gang_lookup_tag - perform multiple lookup on a radix tree
 818  *                                   based on a tag
 819  *      @root:          radix tree root
 820  *      @results:       where the results of the lookup are placed
 821  *      @first_index:   start the lookup from this key
 822  *      @max_items:     place up to this many items at *results
 823  *      @tag:           the tag index (< RADIX_TREE_MAX_TAGS)
 824  *
 825  *      Performs an index-ascending scan of the tree for present items which
 826  *      have the tag indexed by @tag set.  Places the items at *@results and
 827  *      returns the number of items which were placed at *@results.
 828  */
 829 unsigned int
 830 radix_tree_gang_lookup_tag(struct radix_tree_root *root, void **results,
 831                 unsigned long first_index, unsigned int max_items,
 832                 unsigned int tag)
 833 {
 834         struct radix_tree_node *node;
 835         unsigned long max_index;
 836         unsigned long cur_index = first_index;
 837         unsigned int ret;
 838
 839         /* check the root's tag bit */
 840         if (!root_tag_get(root, tag))
 841                 return 0;
 842
 843         node = rcu_dereference(root->rnode);
 844         if (!node)
 845                 return 0;
 846
 847         if (radix_tree_is_direct_ptr(node)) {
 848                 if (first_index > 0)
 849                         return 0;
 850                 node = radix_tree_direct_to_ptr(node);
 851                 results[0] = rcu_dereference(node);
 852                 return 1;
 853         }
 854
 855         max_index = radix_tree_maxindex(node->height);
 856
 857         ret = 0;
 858         while (ret < max_items) {
 859                 unsigned int nr_found;
 860                 unsigned long next_index;       /* Index of next search */
 861
 862                 if (cur_index > max_index)
 863                         break;
 864                 nr_found = __lookup_tag(node, results + ret, cur_index,
 865                                         max_items - ret, &next_index, tag);
 866                 ret += nr_found;
 867                 if (next_index == 0)
 868                         break;
 869                 cur_index = next_index;
 870         }
 871
 872         return ret;
 873 }
 874 EXPORT_SYMBOL(radix_tree_gang_lookup_tag);
 875
 876 /**
 877  *      radix_tree_shrink    -    shrink height of a radix tree to minimal
 878  *      @root           radix tree root
 879  */
 880 static inline void radix_tree_shrink(struct radix_tree_root *root)
 881 {
 882         /* try to shrink tree height */
 883         while (root->height > 0 &&
 884                         root->rnode->count == 1 &&
 885                         root->rnode->slots[0]) {
 886                 struct radix_tree_node *to_free = root->rnode;
 887                 void *newptr;
 888
 889                 /*
 890                  * We don't need rcu_assign_pointer(), since we are simply
 891                  * moving the node from one part of the tree to another. If
 892                  * it was safe to dereference the old pointer to it
 893                  * (to_free->slots[0]), it will be safe to dereference the new
 894                  * one (root->rnode).
 895                  */
 896                 newptr = to_free->slots[0];
 897                 if (root->height == 1)
 898                         newptr = radix_tree_ptr_to_direct(newptr);
 899                 root->rnode = newptr;
 900                 root->height--;
 901                 /* must only free zeroed nodes into the slab */
 902                 tag_clear(to_free, 0, 0);
 903                 tag_clear(to_free, 1, 0);
 904                 to_free->slots[0] = NULL;
 905                 to_free->count = 0;
 906                 radix_tree_node_free(to_free);
 907         }
 908 }
 909
 910 /**
 911  *      radix_tree_delete    -    delete an item from a radix tree
 912  *      @root:          radix tree root
 913  *      @index:         index key
 914  *
 915  *      Remove the item at @index from the radix tree rooted at @root.
 916  *
 917  *      Returns the address of the deleted item, or NULL if it was not present.
 918  */
 919 void *radix_tree_delete(struct radix_tree_root *root, unsigned long index)
 920 {
 921         struct radix_tree_path path[RADIX_TREE_MAX_PATH], *pathp = path;
 922         struct radix_tree_node *slot = NULL;
 923         struct radix_tree_node *to_free;
 924         unsigned int height, shift;
 925         int tag;
 926         int offset;
 927
 928         height = root->height;
 929         if (index > radix_tree_maxindex(height))
 930                 goto out;
 931
 932         slot = root->rnode;
 933         if (height == 0 && root->rnode) {
 934                 slot = radix_tree_direct_to_ptr(slot);
 935                 root_tag_clear_all(root);
 936                 root->rnode = NULL;
 937                 goto out;
 938         }
 939
 940         shift = (height - 1) * RADIX_TREE_MAP_SHIFT;
 941         pathp->node = NULL;
 942
 943         do {
 944                 if (slot == NULL)
 945                         goto out;
 946
 947                 pathp++;
 948                 offset = (index >> shift) & RADIX_TREE_MAP_MASK;
 949                 pathp->offset = offset;
 950                 pathp->node = slot;
 951                 slot = slot->slots[offset];
 952                 shift -= RADIX_TREE_MAP_SHIFT;
 953                 height--;
 954         } while (height > 0);
 955
 956         if (slot == NULL)
 957                 goto out;
 958
 959         /*
 960          * Clear all tags associated with the just-deleted item
 961          */
 962         for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
 963                 if (tag_get(pathp->node, tag, pathp->offset))
 964                         radix_tree_tag_clear(root, index, tag);
 965         }
 966
 967         to_free = NULL;
 968         /* Now free the nodes we do not need anymore */
 969         while (pathp->node) {
 970                 pathp->node->slots[pathp->offset] = NULL;
 971                 pathp->node->count--;
 972                 /*
 973                  * Queue the node for deferred freeing after the
 974                  * last reference to it disappears (set NULL, above).
 975                  */
 976                 if (to_free)
 977                         radix_tree_node_free(to_free);
 978
 979                 if (pathp->node->count) {
 980                         if (pathp->node == root->rnode)
 981                                 radix_tree_shrink(root);
 982                         goto out;
 983                 }
 984
 985                 /* Node with zero slots in use so free it */
 986                 to_free = pathp->node;
 987                 pathp--;
 988
 989         }
 990         root_tag_clear_all(root);
 991         root->height = 0;
 992         root->rnode = NULL;
 993         if (to_free)
 994                 radix_tree_node_free(to_free);
 995
 996 out:
 997         return slot;
 998 }
 999 EXPORT_SYMBOL(radix_tree_delete);
1000
1001 /**
1002  *      radix_tree_tagged - test whether any items in the tree are tagged
1003  *      @root:          radix tree root
1004  *      @tag:           tag to test
1005  */
1006 int radix_tree_tagged(struct radix_tree_root *root, unsigned int tag)
1007 {
1008         return root_tag_get(root, tag);
1009 }
1010 EXPORT_SYMBOL(radix_tree_tagged);
1011
1012 static void
1013 radix_tree_node_ctor(void *node, struct kmem_cache *cachep, unsigned long flags)
1014 {
1015         memset(node, 0, sizeof(struct radix_tree_node));
1016 }
1017
1018 static __init unsigned long __maxindex(unsigned int height)
1019 {
1020         unsigned int tmp = height * RADIX_TREE_MAP_SHIFT;
1021         unsigned long index = (~0UL >> (RADIX_TREE_INDEX_BITS - tmp - 1)) >> 1;
1022
1023         if (tmp >= RADIX_TREE_INDEX_BITS)
1024                 index = ~0UL;
1025         return index;
1026 }
1027
1028 static __init void radix_tree_init_maxindex(void)
1029 {
1030         unsigned int i;
1031
1032         for (i = 0; i < ARRAY_SIZE(height_to_maxindex); i++)
1033                 height_to_maxindex[i] = __maxindex(i);
1034 }
1035
1036 static int radix_tree_callback(struct notifier_block *nfb,
1037                             unsigned long action,
1038                             void *hcpu)
1039 {
1040        int cpu = (long)hcpu;
1041        struct radix_tree_preload *rtp;
1042
1043        /* Free per-cpu pool of perloaded nodes */
1044        if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
1045                rtp = &per_cpu(radix_tree_preloads, cpu);
1046                while (rtp->nr) {
1047                        kmem_cache_free(radix_tree_node_cachep,
1048                                        rtp->nodes[rtp->nr-1]);
1049                        rtp->nodes[rtp->nr-1] = NULL;
1050                        rtp->nr--;
1051                }
1052        }
1053        return NOTIFY_OK;
1054 }
1055
1056 void __init radix_tree_init(void)
1057 {
1058         radix_tree_node_cachep = kmem_cache_create("radix_tree_node",
1059                         sizeof(struct radix_tree_node), 0,
1060                         SLAB_PANIC, radix_tree_node_ctor);
1061         radix_tree_init_maxindex();
1062         hotcpu_notifier(radix_tree_callback, 0);
1063 }