linux/lib/idr.c
<<
>>
Prefs
   1/*
   2 * 2002-10-18  written by Jim Houston jim.houston@ccur.com
   3 *      Copyright (C) 2002 by Concurrent Computer Corporation
   4 *      Distributed under the GNU GPL license version 2.
   5 *
   6 * Modified by George Anzinger to reuse immediately and to use
   7 * find bit instructions.  Also removed _irq on spinlocks.
   8 *
   9 * Modified by Nadia Derbey to make it RCU safe.
  10 *
  11 * Small id to pointer translation service.
  12 *
  13 * It uses a radix tree like structure as a sparse array indexed
  14 * by the id to obtain the pointer.  The bitmap makes allocating
  15 * a new id quick.
  16 *
  17 * You call it to allocate an id (an int) an associate with that id a
  18 * pointer or what ever, we treat it as a (void *).  You can pass this
  19 * id to a user for him to pass back at a later time.  You then pass
  20 * that id to this code and it returns your pointer.
  21
  22 * You can release ids at any time. When all ids are released, most of
  23 * the memory is returned (we keep MAX_IDR_FREE) in a local pool so we
  24 * don't need to go to the memory "store" during an id allocate, just
  25 * so you don't need to be too concerned about locking and conflicts
  26 * with the slab allocator.
  27 */
  28
  29#ifndef TEST                        // to test in user space...
  30#include <linux/slab.h>
  31#include <linux/init.h>
  32#include <linux/export.h>
  33#endif
  34#include <linux/err.h>
  35#include <linux/string.h>
  36#include <linux/idr.h>
  37#include <linux/spinlock.h>
  38#include <linux/percpu.h>
  39#include <linux/hardirq.h>
  40
  41#define MAX_IDR_SHIFT           (sizeof(int) * 8 - 1)
  42#define MAX_IDR_BIT             (1U << MAX_IDR_SHIFT)
  43
  44/* Leave the possibility of an incomplete final layer */
  45#define MAX_IDR_LEVEL ((MAX_IDR_SHIFT + IDR_BITS - 1) / IDR_BITS)
  46
  47/* Number of id_layer structs to leave in free list */
  48#define MAX_IDR_FREE (MAX_IDR_LEVEL * 2)
  49
  50static struct kmem_cache *idr_layer_cache;
  51static DEFINE_PER_CPU(struct idr_layer *, idr_preload_head);
  52static DEFINE_PER_CPU(int, idr_preload_cnt);
  53static DEFINE_SPINLOCK(simple_ida_lock);
  54
  55/* the maximum ID which can be allocated given idr->layers */
  56static int idr_max(int layers)
  57{
  58        int bits = min_t(int, layers * IDR_BITS, MAX_IDR_SHIFT);
  59
  60        return (1 << bits) - 1;
  61}
  62
  63/*
  64 * Prefix mask for an idr_layer at @layer.  For layer 0, the prefix mask is
  65 * all bits except for the lower IDR_BITS.  For layer 1, 2 * IDR_BITS, and
  66 * so on.
  67 */
  68static int idr_layer_prefix_mask(int layer)
  69{
  70        return ~idr_max(layer + 1);
  71}
  72
  73static struct idr_layer *get_from_free_list(struct idr *idp)
  74{
  75        struct idr_layer *p;
  76        unsigned long flags;
  77
  78        spin_lock_irqsave(&idp->lock, flags);
  79        if ((p = idp->id_free)) {
  80                idp->id_free = p->ary[0];
  81                idp->id_free_cnt--;
  82                p->ary[0] = NULL;
  83        }
  84        spin_unlock_irqrestore(&idp->lock, flags);
  85        return(p);
  86}
  87
  88/**
  89 * idr_layer_alloc - allocate a new idr_layer
  90 * @gfp_mask: allocation mask
  91 * @layer_idr: optional idr to allocate from
  92 *
  93 * If @layer_idr is %NULL, directly allocate one using @gfp_mask or fetch
  94 * one from the per-cpu preload buffer.  If @layer_idr is not %NULL, fetch
  95 * an idr_layer from @idr->id_free.
  96 *
  97 * @layer_idr is to maintain backward compatibility with the old alloc
  98 * interface - idr_pre_get() and idr_get_new*() - and will be removed
  99 * together with per-pool preload buffer.
 100 */
 101static struct idr_layer *idr_layer_alloc(gfp_t gfp_mask, struct idr *layer_idr)
 102{
 103        struct idr_layer *new;
 104
 105        /* this is the old path, bypass to get_from_free_list() */
 106        if (layer_idr)
 107                return get_from_free_list(layer_idr);
 108
 109        /*
 110         * Try to allocate directly from kmem_cache.  We want to try this
 111         * before preload buffer; otherwise, non-preloading idr_alloc()
 112         * users will end up taking advantage of preloading ones.  As the
 113         * following is allowed to fail for preloaded cases, suppress
 114         * warning this time.
 115         */
 116        new = kmem_cache_zalloc(idr_layer_cache, gfp_mask | __GFP_NOWARN);
 117        if (new)
 118                return new;
 119
 120        /*
 121         * Try to fetch one from the per-cpu preload buffer if in process
 122         * context.  See idr_preload() for details.
 123         */
 124        if (!in_interrupt()) {
 125                preempt_disable();
 126                new = __this_cpu_read(idr_preload_head);
 127                if (new) {
 128                        __this_cpu_write(idr_preload_head, new->ary[0]);
 129                        __this_cpu_dec(idr_preload_cnt);
 130                        new->ary[0] = NULL;
 131                }
 132                preempt_enable();
 133                if (new)
 134                        return new;
 135        }
 136
 137        /*
 138         * Both failed.  Try kmem_cache again w/o adding __GFP_NOWARN so
 139         * that memory allocation failure warning is printed as intended.
 140         */
 141        return kmem_cache_zalloc(idr_layer_cache, gfp_mask);
 142}
 143
 144static void idr_layer_rcu_free(struct rcu_head *head)
 145{
 146        struct idr_layer *layer;
 147
 148        layer = container_of(head, struct idr_layer, rcu_head);
 149        kmem_cache_free(idr_layer_cache, layer);
 150}
 151
 152static inline void free_layer(struct idr *idr, struct idr_layer *p)
 153{
 154        if (idr->hint && idr->hint == p)
 155                RCU_INIT_POINTER(idr->hint, NULL);
 156        call_rcu(&p->rcu_head, idr_layer_rcu_free);
 157}
 158
 159/* only called when idp->lock is held */
 160static void __move_to_free_list(struct idr *idp, struct idr_layer *p)
 161{
 162        p->ary[0] = idp->id_free;
 163        idp->id_free = p;
 164        idp->id_free_cnt++;
 165}
 166
 167static void move_to_free_list(struct idr *idp, struct idr_layer *p)
 168{
 169        unsigned long flags;
 170
 171        /*
 172         * Depends on the return element being zeroed.
 173         */
 174        spin_lock_irqsave(&idp->lock, flags);
 175        __move_to_free_list(idp, p);
 176        spin_unlock_irqrestore(&idp->lock, flags);
 177}
 178
 179static void idr_mark_full(struct idr_layer **pa, int id)
 180{
 181        struct idr_layer *p = pa[0];
 182        int l = 0;
 183
 184        __set_bit(id & IDR_MASK, p->bitmap);
 185        /*
 186         * If this layer is full mark the bit in the layer above to
 187         * show that this part of the radix tree is full.  This may
 188         * complete the layer above and require walking up the radix
 189         * tree.
 190         */
 191        while (bitmap_full(p->bitmap, IDR_SIZE)) {
 192                if (!(p = pa[++l]))
 193                        break;
 194                id = id >> IDR_BITS;
 195                __set_bit((id & IDR_MASK), p->bitmap);
 196        }
 197}
 198
 199int __idr_pre_get(struct idr *idp, gfp_t gfp_mask)
 200{
 201        while (idp->id_free_cnt < MAX_IDR_FREE) {
 202                struct idr_layer *new;
 203                new = kmem_cache_zalloc(idr_layer_cache, gfp_mask);
 204                if (new == NULL)
 205                        return (0);
 206                move_to_free_list(idp, new);
 207        }
 208        return 1;
 209}
 210EXPORT_SYMBOL(__idr_pre_get);
 211
 212/**
 213 * sub_alloc - try to allocate an id without growing the tree depth
 214 * @idp: idr handle
 215 * @starting_id: id to start search at
 216 * @pa: idr_layer[MAX_IDR_LEVEL] used as backtrack buffer
 217 * @gfp_mask: allocation mask for idr_layer_alloc()
 218 * @layer_idr: optional idr passed to idr_layer_alloc()
 219 *
 220 * Allocate an id in range [@starting_id, INT_MAX] from @idp without
 221 * growing its depth.  Returns
 222 *
 223 *  the allocated id >= 0 if successful,
 224 *  -EAGAIN if the tree needs to grow for allocation to succeed,
 225 *  -ENOSPC if the id space is exhausted,
 226 *  -ENOMEM if more idr_layers need to be allocated.
 227 */
 228static int sub_alloc(struct idr *idp, int *starting_id, struct idr_layer **pa,
 229                     gfp_t gfp_mask, struct idr *layer_idr)
 230{
 231        int n, m, sh;
 232        struct idr_layer *p, *new;
 233        int l, id, oid;
 234
 235        id = *starting_id;
 236 restart:
 237        p = idp->top;
 238        l = idp->layers;
 239        pa[l--] = NULL;
 240        while (1) {
 241                /*
 242                 * We run around this while until we reach the leaf node...
 243                 */
 244                n = (id >> (IDR_BITS*l)) & IDR_MASK;
 245                m = find_next_zero_bit(p->bitmap, IDR_SIZE, n);
 246                if (m == IDR_SIZE) {
 247                        /* no space available go back to previous layer. */
 248                        l++;
 249                        oid = id;
 250                        id = (id | ((1 << (IDR_BITS * l)) - 1)) + 1;
 251
 252                        /* if already at the top layer, we need to grow */
 253                        if (id >= 1 << (idp->layers * IDR_BITS)) {
 254                                *starting_id = id;
 255                                return -EAGAIN;
 256                        }
 257                        p = pa[l];
 258                        BUG_ON(!p);
 259
 260                        /* If we need to go up one layer, continue the
 261                         * loop; otherwise, restart from the top.
 262                         */
 263                        sh = IDR_BITS * (l + 1);
 264                        if (oid >> sh == id >> sh)
 265                                continue;
 266                        else
 267                                goto restart;
 268                }
 269                if (m != n) {
 270                        sh = IDR_BITS*l;
 271                        id = ((id >> sh) ^ n ^ m) << sh;
 272                }
 273                if ((id >= MAX_IDR_BIT) || (id < 0))
 274                        return -ENOSPC;
 275                if (l == 0)
 276                        break;
 277                /*
 278                 * Create the layer below if it is missing.
 279                 */
 280                if (!p->ary[m]) {
 281                        new = idr_layer_alloc(gfp_mask, layer_idr);
 282                        if (!new)
 283                                return -ENOMEM;
 284                        new->layer = l-1;
 285                        new->prefix = id & idr_layer_prefix_mask(new->layer);
 286                        rcu_assign_pointer(p->ary[m], new);
 287                        p->count++;
 288                }
 289                pa[l--] = p;
 290                p = p->ary[m];
 291        }
 292
 293        pa[l] = p;
 294        return id;
 295}
 296
 297static int idr_get_empty_slot(struct idr *idp, int starting_id,
 298                              struct idr_layer **pa, gfp_t gfp_mask,
 299                              struct idr *layer_idr)
 300{
 301        struct idr_layer *p, *new;
 302        int layers, v, id;
 303        unsigned long flags;
 304
 305        id = starting_id;
 306build_up:
 307        p = idp->top;
 308        layers = idp->layers;
 309        if (unlikely(!p)) {
 310                if (!(p = idr_layer_alloc(gfp_mask, layer_idr)))
 311                        return -ENOMEM;
 312                p->layer = 0;
 313                layers = 1;
 314        }
 315        /*
 316         * Add a new layer to the top of the tree if the requested
 317         * id is larger than the currently allocated space.
 318         */
 319        while (id > idr_max(layers)) {
 320                layers++;
 321                if (!p->count) {
 322                        /* special case: if the tree is currently empty,
 323                         * then we grow the tree by moving the top node
 324                         * upwards.
 325                         */
 326                        p->layer++;
 327                        WARN_ON_ONCE(p->prefix);
 328                        continue;
 329                }
 330                if (!(new = idr_layer_alloc(gfp_mask, layer_idr))) {
 331                        /*
 332                         * The allocation failed.  If we built part of
 333                         * the structure tear it down.
 334                         */
 335                        spin_lock_irqsave(&idp->lock, flags);
 336                        for (new = p; p && p != idp->top; new = p) {
 337                                p = p->ary[0];
 338                                new->ary[0] = NULL;
 339                                new->count = 0;
 340                                bitmap_clear(new->bitmap, 0, IDR_SIZE);
 341                                __move_to_free_list(idp, new);
 342                        }
 343                        spin_unlock_irqrestore(&idp->lock, flags);
 344                        return -ENOMEM;
 345                }
 346                new->ary[0] = p;
 347                new->count = 1;
 348                new->layer = layers-1;
 349                new->prefix = id & idr_layer_prefix_mask(new->layer);
 350                if (bitmap_full(p->bitmap, IDR_SIZE))
 351                        __set_bit(0, new->bitmap);
 352                p = new;
 353        }
 354        rcu_assign_pointer(idp->top, p);
 355        idp->layers = layers;
 356        v = sub_alloc(idp, &id, pa, gfp_mask, layer_idr);
 357        if (v == -EAGAIN)
 358                goto build_up;
 359        return(v);
 360}
 361
 362/*
 363 * @id and @pa are from a successful allocation from idr_get_empty_slot().
 364 * Install the user pointer @ptr and mark the slot full.
 365 */
 366static void idr_fill_slot(struct idr *idr, void *ptr, int id,
 367                          struct idr_layer **pa)
 368{
 369        /* update hint used for lookup, cleared from free_layer() */
 370        rcu_assign_pointer(idr->hint, pa[0]);
 371
 372        rcu_assign_pointer(pa[0]->ary[id & IDR_MASK], (struct idr_layer *)ptr);
 373        pa[0]->count++;
 374        idr_mark_full(pa, id);
 375}
 376
 377int __idr_get_new_above(struct idr *idp, void *ptr, int starting_id, int *id)
 378{
 379        struct idr_layer *pa[MAX_IDR_LEVEL + 1];
 380        int rv;
 381
 382        rv = idr_get_empty_slot(idp, starting_id, pa, 0, idp);
 383        if (rv < 0)
 384                return rv == -ENOMEM ? -EAGAIN : rv;
 385
 386        idr_fill_slot(idp, ptr, rv, pa);
 387        *id = rv;
 388        return 0;
 389}
 390EXPORT_SYMBOL(__idr_get_new_above);
 391
 392/**
 393 * idr_preload - preload for idr_alloc()
 394 * @gfp_mask: allocation mask to use for preloading
 395 *
 396 * Preload per-cpu layer buffer for idr_alloc().  Can only be used from
 397 * process context and each idr_preload() invocation should be matched with
 398 * idr_preload_end().  Note that preemption is disabled while preloaded.
 399 *
 400 * The first idr_alloc() in the preloaded section can be treated as if it
 401 * were invoked with @gfp_mask used for preloading.  This allows using more
 402 * permissive allocation masks for idrs protected by spinlocks.
 403 *
 404 * For example, if idr_alloc() below fails, the failure can be treated as
 405 * if idr_alloc() were called with GFP_KERNEL rather than GFP_NOWAIT.
 406 *
 407 *      idr_preload(GFP_KERNEL);
 408 *      spin_lock(lock);
 409 *
 410 *      id = idr_alloc(idr, ptr, start, end, GFP_NOWAIT);
 411 *
 412 *      spin_unlock(lock);
 413 *      idr_preload_end();
 414 *      if (id < 0)
 415 *              error;
 416 */
 417void idr_preload(gfp_t gfp_mask)
 418{
 419        /*
 420         * Consuming preload buffer from non-process context breaks preload
 421         * allocation guarantee.  Disallow usage from those contexts.
 422         */
 423        WARN_ON_ONCE(in_interrupt());
 424        might_sleep_if(gfp_mask & __GFP_WAIT);
 425
 426        preempt_disable();
 427
 428        /*
 429         * idr_alloc() is likely to succeed w/o full idr_layer buffer and
 430         * return value from idr_alloc() needs to be checked for failure
 431         * anyway.  Silently give up if allocation fails.  The caller can
 432         * treat failures from idr_alloc() as if idr_alloc() were called
 433         * with @gfp_mask which should be enough.
 434         */
 435        while (__this_cpu_read(idr_preload_cnt) < MAX_IDR_FREE) {
 436                struct idr_layer *new;
 437
 438                preempt_enable();
 439                new = kmem_cache_zalloc(idr_layer_cache, gfp_mask);
 440                preempt_disable();
 441                if (!new)
 442                        break;
 443
 444                /* link the new one to per-cpu preload list */
 445                new->ary[0] = __this_cpu_read(idr_preload_head);
 446                __this_cpu_write(idr_preload_head, new);
 447                __this_cpu_inc(idr_preload_cnt);
 448        }
 449}
 450EXPORT_SYMBOL(idr_preload);
 451
 452/**
 453 * idr_alloc - allocate new idr entry
 454 * @idr: the (initialized) idr
 455 * @ptr: pointer to be associated with the new id
 456 * @start: the minimum id (inclusive)
 457 * @end: the maximum id (exclusive, <= 0 for max)
 458 * @gfp_mask: memory allocation flags
 459 *
 460 * Allocate an id in [start, end) and associate it with @ptr.  If no ID is
 461 * available in the specified range, returns -ENOSPC.  On memory allocation
 462 * failure, returns -ENOMEM.
 463 *
 464 * Note that @end is treated as max when <= 0.  This is to always allow
 465 * using @start + N as @end as long as N is inside integer range.
 466 *
 467 * The user is responsible for exclusively synchronizing all operations
 468 * which may modify @idr.  However, read-only accesses such as idr_find()
 469 * or iteration can be performed under RCU read lock provided the user
 470 * destroys @ptr in RCU-safe way after removal from idr.
 471 */
 472int idr_alloc(struct idr *idr, void *ptr, int start, int end, gfp_t gfp_mask)
 473{
 474        int max = end > 0 ? end - 1 : INT_MAX;  /* inclusive upper limit */
 475        struct idr_layer *pa[MAX_IDR_LEVEL + 1];
 476        int id;
 477
 478        might_sleep_if(gfp_mask & __GFP_WAIT);
 479
 480        /* sanity checks */
 481        if (WARN_ON_ONCE(start < 0))
 482                return -EINVAL;
 483        if (unlikely(max < start))
 484                return -ENOSPC;
 485
 486        /* allocate id */
 487        id = idr_get_empty_slot(idr, start, pa, gfp_mask, NULL);
 488        if (unlikely(id < 0))
 489                return id;
 490        if (unlikely(id > max))
 491                return -ENOSPC;
 492
 493        idr_fill_slot(idr, ptr, id, pa);
 494        return id;
 495}
 496EXPORT_SYMBOL_GPL(idr_alloc);
 497
 498/**
 499 * idr_alloc_cyclic - allocate new idr entry in a cyclical fashion
 500 * @idr: the (initialized) idr
 501 * @ptr: pointer to be associated with the new id
 502 * @start: the minimum id (inclusive)
 503 * @end: the maximum id (exclusive, <= 0 for max)
 504 * @gfp_mask: memory allocation flags
 505 *
 506 * Essentially the same as idr_alloc, but prefers to allocate progressively
 507 * higher ids if it can. If the "cur" counter wraps, then it will start again
 508 * at the "start" end of the range and allocate one that has already been used.
 509 */
 510int idr_alloc_cyclic(struct idr *idr, void *ptr, int start, int end,
 511                        gfp_t gfp_mask)
 512{
 513        int id;
 514
 515        id = idr_alloc(idr, ptr, max(start, idr->cur), end, gfp_mask);
 516        if (id == -ENOSPC)
 517                id = idr_alloc(idr, ptr, start, end, gfp_mask);
 518
 519        if (likely(id >= 0))
 520                idr->cur = id + 1;
 521        return id;
 522}
 523EXPORT_SYMBOL(idr_alloc_cyclic);
 524
 525static void idr_remove_warning(int id)
 526{
 527        WARN(1, "idr_remove called for id=%d which is not allocated.\n", id);
 528}
 529
 530static void sub_remove(struct idr *idp, int shift, int id)
 531{
 532        struct idr_layer *p = idp->top;
 533        struct idr_layer **pa[MAX_IDR_LEVEL + 1];
 534        struct idr_layer ***paa = &pa[0];
 535        struct idr_layer *to_free;
 536        int n;
 537
 538        *paa = NULL;
 539        *++paa = &idp->top;
 540
 541        while ((shift > 0) && p) {
 542                n = (id >> shift) & IDR_MASK;
 543                __clear_bit(n, p->bitmap);
 544                *++paa = &p->ary[n];
 545                p = p->ary[n];
 546                shift -= IDR_BITS;
 547        }
 548        n = id & IDR_MASK;
 549        if (likely(p != NULL && test_bit(n, p->bitmap))) {
 550                __clear_bit(n, p->bitmap);
 551                rcu_assign_pointer(p->ary[n], NULL);
 552                to_free = NULL;
 553                while(*paa && ! --((**paa)->count)){
 554                        if (to_free)
 555                                free_layer(idp, to_free);
 556                        to_free = **paa;
 557                        **paa-- = NULL;
 558                }
 559                if (!*paa)
 560                        idp->layers = 0;
 561                if (to_free)
 562                        free_layer(idp, to_free);
 563        } else
 564                idr_remove_warning(id);
 565}
 566
 567/**
 568 * idr_remove - remove the given id and free its slot
 569 * @idp: idr handle
 570 * @id: unique key
 571 */
 572void idr_remove(struct idr *idp, int id)
 573{
 574        struct idr_layer *p;
 575        struct idr_layer *to_free;
 576
 577        if (id < 0)
 578                return;
 579
 580        sub_remove(idp, (idp->layers - 1) * IDR_BITS, id);
 581        if (idp->top && idp->top->count == 1 && (idp->layers > 1) &&
 582            idp->top->ary[0]) {
 583                /*
 584                 * Single child at leftmost slot: we can shrink the tree.
 585                 * This level is not needed anymore since when layers are
 586                 * inserted, they are inserted at the top of the existing
 587                 * tree.
 588                 */
 589                to_free = idp->top;
 590                p = idp->top->ary[0];
 591                rcu_assign_pointer(idp->top, p);
 592                --idp->layers;
 593                to_free->count = 0;
 594                bitmap_clear(to_free->bitmap, 0, IDR_SIZE);
 595                free_layer(idp, to_free);
 596        }
 597        while (idp->id_free_cnt >= MAX_IDR_FREE) {
 598                p = get_from_free_list(idp);
 599                /*
 600                 * Note: we don't call the rcu callback here, since the only
 601                 * layers that fall into the freelist are those that have been
 602                 * preallocated.
 603                 */
 604                kmem_cache_free(idr_layer_cache, p);
 605        }
 606        return;
 607}
 608EXPORT_SYMBOL(idr_remove);
 609
 610void __idr_remove_all(struct idr *idp)
 611{
 612        int n, id, max;
 613        int bt_mask;
 614        struct idr_layer *p;
 615        struct idr_layer *pa[MAX_IDR_LEVEL + 1];
 616        struct idr_layer **paa = &pa[0];
 617
 618        n = idp->layers * IDR_BITS;
 619        p = idp->top;
 620        rcu_assign_pointer(idp->top, NULL);
 621        max = idr_max(idp->layers);
 622
 623        id = 0;
 624        while (id >= 0 && id <= max) {
 625                while (n > IDR_BITS && p) {
 626                        n -= IDR_BITS;
 627                        *paa++ = p;
 628                        p = p->ary[(id >> n) & IDR_MASK];
 629                }
 630
 631                bt_mask = id;
 632                id += 1 << n;
 633                /* Get the highest bit that the above add changed from 0->1. */
 634                while (n < fls(id ^ bt_mask)) {
 635                        if (p)
 636                                free_layer(idp, p);
 637                        n += IDR_BITS;
 638                        p = *--paa;
 639                }
 640        }
 641        idp->layers = 0;
 642}
 643EXPORT_SYMBOL(__idr_remove_all);
 644
 645/**
 646 * idr_destroy - release all cached layers within an idr tree
 647 * @idp: idr handle
 648 *
 649 * Free all id mappings and all idp_layers.  After this function, @idp is
 650 * completely unused and can be freed / recycled.  The caller is
 651 * responsible for ensuring that no one else accesses @idp during or after
 652 * idr_destroy().
 653 *
 654 * A typical clean-up sequence for objects stored in an idr tree will use
 655 * idr_for_each() to free all objects, if necessay, then idr_destroy() to
 656 * free up the id mappings and cached idr_layers.
 657 */
 658void idr_destroy(struct idr *idp)
 659{
 660        __idr_remove_all(idp);
 661
 662        while (idp->id_free_cnt) {
 663                struct idr_layer *p = get_from_free_list(idp);
 664                kmem_cache_free(idr_layer_cache, p);
 665        }
 666}
 667EXPORT_SYMBOL(idr_destroy);
 668
 669void *idr_find_slowpath(struct idr *idp, int id)
 670{
 671        int n;
 672        struct idr_layer *p;
 673
 674        if (id < 0)
 675                return NULL;
 676
 677        p = rcu_dereference_raw(idp->top);
 678        if (!p)
 679                return NULL;
 680        n = (p->layer+1) * IDR_BITS;
 681
 682        if (id > idr_max(p->layer + 1))
 683                return NULL;
 684        BUG_ON(n == 0);
 685
 686        while (n > 0 && p) {
 687                n -= IDR_BITS;
 688                BUG_ON(n != p->layer*IDR_BITS);
 689                p = rcu_dereference_raw(p->ary[(id >> n) & IDR_MASK]);
 690        }
 691        return((void *)p);
 692}
 693EXPORT_SYMBOL(idr_find_slowpath);
 694
 695/**
 696 * idr_for_each - iterate through all stored pointers
 697 * @idp: idr handle
 698 * @fn: function to be called for each pointer
 699 * @data: data passed back to callback function
 700 *
 701 * Iterate over the pointers registered with the given idr.  The
 702 * callback function will be called for each pointer currently
 703 * registered, passing the id, the pointer and the data pointer passed
 704 * to this function.  It is not safe to modify the idr tree while in
 705 * the callback, so functions such as idr_get_new and idr_remove are
 706 * not allowed.
 707 *
 708 * We check the return of @fn each time. If it returns anything other
 709 * than %0, we break out and return that value.
 710 *
 711 * The caller must serialize idr_for_each() vs idr_get_new() and idr_remove().
 712 */
 713int idr_for_each(struct idr *idp,
 714                 int (*fn)(int id, void *p, void *data), void *data)
 715{
 716        int n, id, max, error = 0;
 717        struct idr_layer *p;
 718        struct idr_layer *pa[MAX_IDR_LEVEL + 1];
 719        struct idr_layer **paa = &pa[0];
 720
 721        n = idp->layers * IDR_BITS;
 722        p = rcu_dereference_raw(idp->top);
 723        max = idr_max(idp->layers);
 724
 725        id = 0;
 726        while (id >= 0 && id <= max) {
 727                while (n > 0 && p) {
 728                        n -= IDR_BITS;
 729                        *paa++ = p;
 730                        p = rcu_dereference_raw(p->ary[(id >> n) & IDR_MASK]);
 731                }
 732
 733                if (p) {
 734                        error = fn(id, (void *)p, data);
 735                        if (error)
 736                                break;
 737                }
 738
 739                id += 1 << n;
 740                while (n < fls(id)) {
 741                        n += IDR_BITS;
 742                        p = *--paa;
 743                }
 744        }
 745
 746        return error;
 747}
 748EXPORT_SYMBOL(idr_for_each);
 749
 750/**
 751 * idr_get_next - lookup next object of id to given id.
 752 * @idp: idr handle
 753 * @nextidp:  pointer to lookup key
 754 *
 755 * Returns pointer to registered object with id, which is next number to
 756 * given id. After being looked up, *@nextidp will be updated for the next
 757 * iteration.
 758 *
 759 * This function can be called under rcu_read_lock(), given that the leaf
 760 * pointers lifetimes are correctly managed.
 761 */
 762void *idr_get_next(struct idr *idp, int *nextidp)
 763{
 764        struct idr_layer *p, *pa[MAX_IDR_LEVEL + 1];
 765        struct idr_layer **paa = &pa[0];
 766        int id = *nextidp;
 767        int n, max;
 768
 769        /* find first ent */
 770        p = rcu_dereference_raw(idp->top);
 771        if (!p)
 772                return NULL;
 773        n = (p->layer + 1) * IDR_BITS;
 774        max = idr_max(p->layer + 1);
 775
 776        while (id >= 0 && id <= max) {
 777                while (n > 0 && p) {
 778                        n -= IDR_BITS;
 779                        *paa++ = p;
 780                        p = rcu_dereference_raw(p->ary[(id >> n) & IDR_MASK]);
 781                }
 782
 783                if (p) {
 784                        *nextidp = id;
 785                        return p;
 786                }
 787
 788                /*
 789                 * Proceed to the next layer at the current level.  Unlike
 790                 * idr_for_each(), @id isn't guaranteed to be aligned to
 791                 * layer boundary at this point and adding 1 << n may
 792                 * incorrectly skip IDs.  Make sure we jump to the
 793                 * beginning of the next layer using round_up().
 794                 */
 795                id = round_up(id + 1, 1 << n);
 796                while (n < fls(id)) {
 797                        n += IDR_BITS;
 798                        p = *--paa;
 799                }
 800        }
 801        return NULL;
 802}
 803EXPORT_SYMBOL(idr_get_next);
 804
 805
 806/**
 807 * idr_replace - replace pointer for given id
 808 * @idp: idr handle
 809 * @ptr: pointer you want associated with the id
 810 * @id: lookup key
 811 *
 812 * Replace the pointer registered with an id and return the old value.
 813 * A %-ENOENT return indicates that @id was not found.
 814 * A %-EINVAL return indicates that @id was not within valid constraints.
 815 *
 816 * The caller must serialize with writers.
 817 */
 818void *idr_replace(struct idr *idp, void *ptr, int id)
 819{
 820        int n;
 821        struct idr_layer *p, *old_p;
 822
 823        if (id < 0)
 824                return ERR_PTR(-EINVAL);
 825
 826        p = idp->top;
 827        if (!p)
 828                return ERR_PTR(-EINVAL);
 829
 830        n = (p->layer+1) * IDR_BITS;
 831
 832        if (id >= (1 << n))
 833                return ERR_PTR(-EINVAL);
 834
 835        n -= IDR_BITS;
 836        while ((n > 0) && p) {
 837                p = p->ary[(id >> n) & IDR_MASK];
 838                n -= IDR_BITS;
 839        }
 840
 841        n = id & IDR_MASK;
 842        if (unlikely(p == NULL || !test_bit(n, p->bitmap)))
 843                return ERR_PTR(-ENOENT);
 844
 845        old_p = p->ary[n];
 846        rcu_assign_pointer(p->ary[n], ptr);
 847
 848        return old_p;
 849}
 850EXPORT_SYMBOL(idr_replace);
 851
 852void __init idr_init_cache(void)
 853{
 854        idr_layer_cache = kmem_cache_create("idr_layer_cache",
 855                                sizeof(struct idr_layer), 0, SLAB_PANIC, NULL);
 856}
 857
 858/**
 859 * idr_init - initialize idr handle
 860 * @idp:        idr handle
 861 *
 862 * This function is use to set up the handle (@idp) that you will pass
 863 * to the rest of the functions.
 864 */
 865void idr_init(struct idr *idp)
 866{
 867        memset(idp, 0, sizeof(struct idr));
 868        spin_lock_init(&idp->lock);
 869}
 870EXPORT_SYMBOL(idr_init);
 871
 872
 873/**
 874 * DOC: IDA description
 875 * IDA - IDR based ID allocator
 876 *
 877 * This is id allocator without id -> pointer translation.  Memory
 878 * usage is much lower than full blown idr because each id only
 879 * occupies a bit.  ida uses a custom leaf node which contains
 880 * IDA_BITMAP_BITS slots.
 881 *
 882 * 2007-04-25  written by Tejun Heo <htejun@gmail.com>
 883 */
 884
 885static void free_bitmap(struct ida *ida, struct ida_bitmap *bitmap)
 886{
 887        unsigned long flags;
 888
 889        if (!ida->free_bitmap) {
 890                spin_lock_irqsave(&ida->idr.lock, flags);
 891                if (!ida->free_bitmap) {
 892                        ida->free_bitmap = bitmap;
 893                        bitmap = NULL;
 894                }
 895                spin_unlock_irqrestore(&ida->idr.lock, flags);
 896        }
 897
 898        kfree(bitmap);
 899}
 900
 901/**
 902 * ida_pre_get - reserve resources for ida allocation
 903 * @ida:        ida handle
 904 * @gfp_mask:   memory allocation flag
 905 *
 906 * This function should be called prior to locking and calling the
 907 * following function.  It preallocates enough memory to satisfy the
 908 * worst possible allocation.
 909 *
 910 * If the system is REALLY out of memory this function returns %0,
 911 * otherwise %1.
 912 */
 913int ida_pre_get(struct ida *ida, gfp_t gfp_mask)
 914{
 915        /* allocate idr_layers */
 916        if (!__idr_pre_get(&ida->idr, gfp_mask))
 917                return 0;
 918
 919        /* allocate free_bitmap */
 920        if (!ida->free_bitmap) {
 921                struct ida_bitmap *bitmap;
 922
 923                bitmap = kmalloc(sizeof(struct ida_bitmap), gfp_mask);
 924                if (!bitmap)
 925                        return 0;
 926
 927                free_bitmap(ida, bitmap);
 928        }
 929
 930        return 1;
 931}
 932EXPORT_SYMBOL(ida_pre_get);
 933
 934/**
 935 * ida_get_new_above - allocate new ID above or equal to a start id
 936 * @ida:        ida handle
 937 * @starting_id: id to start search at
 938 * @p_id:       pointer to the allocated handle
 939 *
 940 * Allocate new ID above or equal to @starting_id.  It should be called
 941 * with any required locks.
 942 *
 943 * If memory is required, it will return %-EAGAIN, you should unlock
 944 * and go back to the ida_pre_get() call.  If the ida is full, it will
 945 * return %-ENOSPC.
 946 *
 947 * @p_id returns a value in the range @starting_id ... %0x7fffffff.
 948 */
 949int ida_get_new_above(struct ida *ida, int starting_id, int *p_id)
 950{
 951        struct idr_layer *pa[MAX_IDR_LEVEL + 1];
 952        struct ida_bitmap *bitmap;
 953        unsigned long flags;
 954        int idr_id = starting_id / IDA_BITMAP_BITS;
 955        int offset = starting_id % IDA_BITMAP_BITS;
 956        int t, id;
 957
 958 restart:
 959        /* get vacant slot */
 960        t = idr_get_empty_slot(&ida->idr, idr_id, pa, 0, &ida->idr);
 961        if (t < 0)
 962                return t == -ENOMEM ? -EAGAIN : t;
 963
 964        if (t * IDA_BITMAP_BITS >= MAX_IDR_BIT)
 965                return -ENOSPC;
 966
 967        if (t != idr_id)
 968                offset = 0;
 969        idr_id = t;
 970
 971        /* if bitmap isn't there, create a new one */
 972        bitmap = (void *)pa[0]->ary[idr_id & IDR_MASK];
 973        if (!bitmap) {
 974                spin_lock_irqsave(&ida->idr.lock, flags);
 975                bitmap = ida->free_bitmap;
 976                ida->free_bitmap = NULL;
 977                spin_unlock_irqrestore(&ida->idr.lock, flags);
 978
 979                if (!bitmap)
 980                        return -EAGAIN;
 981
 982                memset(bitmap, 0, sizeof(struct ida_bitmap));
 983                rcu_assign_pointer(pa[0]->ary[idr_id & IDR_MASK],
 984                                (void *)bitmap);
 985                pa[0]->count++;
 986        }
 987
 988        /* lookup for empty slot */
 989        t = find_next_zero_bit(bitmap->bitmap, IDA_BITMAP_BITS, offset);
 990        if (t == IDA_BITMAP_BITS) {
 991                /* no empty slot after offset, continue to the next chunk */
 992                idr_id++;
 993                offset = 0;
 994                goto restart;
 995        }
 996
 997        id = idr_id * IDA_BITMAP_BITS + t;
 998        if (id >= MAX_IDR_BIT)
 999                return -ENOSPC;
1000
1001        __set_bit(t, bitmap->bitmap);
1002        if (++bitmap->nr_busy == IDA_BITMAP_BITS)
1003                idr_mark_full(pa, idr_id);
1004
1005        *p_id = id;
1006
1007        /* Each leaf node can handle nearly a thousand slots and the
1008         * whole idea of ida is to have small memory foot print.
1009         * Throw away extra resources one by one after each successful
1010         * allocation.
1011         */
1012        if (ida->idr.id_free_cnt || ida->free_bitmap) {
1013                struct idr_layer *p = get_from_free_list(&ida->idr);
1014                if (p)
1015                        kmem_cache_free(idr_layer_cache, p);
1016        }
1017
1018        return 0;
1019}
1020EXPORT_SYMBOL(ida_get_new_above);
1021
1022/**
1023 * ida_remove - remove the given ID
1024 * @ida:        ida handle
1025 * @id:         ID to free
1026 */
1027void ida_remove(struct ida *ida, int id)
1028{
1029        struct idr_layer *p = ida->idr.top;
1030        int shift = (ida->idr.layers - 1) * IDR_BITS;
1031        int idr_id = id / IDA_BITMAP_BITS;
1032        int offset = id % IDA_BITMAP_BITS;
1033        int n;
1034        struct ida_bitmap *bitmap;
1035
1036        /* clear full bits while looking up the leaf idr_layer */
1037        while ((shift > 0) && p) {
1038                n = (idr_id >> shift) & IDR_MASK;
1039                __clear_bit(n, p->bitmap);
1040                p = p->ary[n];
1041                shift -= IDR_BITS;
1042        }
1043
1044        if (p == NULL)
1045                goto err;
1046
1047        n = idr_id & IDR_MASK;
1048        __clear_bit(n, p->bitmap);
1049
1050        bitmap = (void *)p->ary[n];
1051        if (!test_bit(offset, bitmap->bitmap))
1052                goto err;
1053
1054        /* update bitmap and remove it if empty */
1055        __clear_bit(offset, bitmap->bitmap);
1056        if (--bitmap->nr_busy == 0) {
1057                __set_bit(n, p->bitmap);        /* to please idr_remove() */
1058                idr_remove(&ida->idr, idr_id);
1059                free_bitmap(ida, bitmap);
1060        }
1061
1062        return;
1063
1064 err:
1065        WARN(1, "ida_remove called for id=%d which is not allocated.\n", id);
1066}
1067EXPORT_SYMBOL(ida_remove);
1068
1069/**
1070 * ida_destroy - release all cached layers within an ida tree
1071 * @ida:                ida handle
1072 */
1073void ida_destroy(struct ida *ida)
1074{
1075        idr_destroy(&ida->idr);
1076        kfree(ida->free_bitmap);
1077}
1078EXPORT_SYMBOL(ida_destroy);
1079
1080/**
1081 * ida_simple_get - get a new id.
1082 * @ida: the (initialized) ida.
1083 * @start: the minimum id (inclusive, < 0x8000000)
1084 * @end: the maximum id (exclusive, < 0x8000000 or 0)
1085 * @gfp_mask: memory allocation flags
1086 *
1087 * Allocates an id in the range start <= id < end, or returns -ENOSPC.
1088 * On memory allocation failure, returns -ENOMEM.
1089 *
1090 * Use ida_simple_remove() to get rid of an id.
1091 */
1092int ida_simple_get(struct ida *ida, unsigned int start, unsigned int end,
1093                   gfp_t gfp_mask)
1094{
1095        int ret, id;
1096        unsigned int max;
1097        unsigned long flags;
1098
1099        BUG_ON((int)start < 0);
1100        BUG_ON((int)end < 0);
1101
1102        if (end == 0)
1103                max = 0x80000000;
1104        else {
1105                BUG_ON(end < start);
1106                max = end - 1;
1107        }
1108
1109again:
1110        if (!ida_pre_get(ida, gfp_mask))
1111                return -ENOMEM;
1112
1113        spin_lock_irqsave(&simple_ida_lock, flags);
1114        ret = ida_get_new_above(ida, start, &id);
1115        if (!ret) {
1116                if (id > max) {
1117                        ida_remove(ida, id);
1118                        ret = -ENOSPC;
1119                } else {
1120                        ret = id;
1121                }
1122        }
1123        spin_unlock_irqrestore(&simple_ida_lock, flags);
1124
1125        if (unlikely(ret == -EAGAIN))
1126                goto again;
1127
1128        return ret;
1129}
1130EXPORT_SYMBOL(ida_simple_get);
1131
1132/**
1133 * ida_simple_remove - remove an allocated id.
1134 * @ida: the (initialized) ida.
1135 * @id: the id returned by ida_simple_get.
1136 */
1137void ida_simple_remove(struct ida *ida, unsigned int id)
1138{
1139        unsigned long flags;
1140
1141        BUG_ON((int)id < 0);
1142        spin_lock_irqsave(&simple_ida_lock, flags);
1143        ida_remove(ida, id);
1144        spin_unlock_irqrestore(&simple_ida_lock, flags);
1145}
1146EXPORT_SYMBOL(ida_simple_remove);
1147
1148/**
1149 * ida_init - initialize ida handle
1150 * @ida:        ida handle
1151 *
1152 * This function is use to set up the handle (@ida) that you will pass
1153 * to the rest of the functions.
1154 */
1155void ida_init(struct ida *ida)
1156{
1157        memset(ida, 0, sizeof(struct ida));
1158        idr_init(&ida->idr);
1159
1160}
1161EXPORT_SYMBOL(ida_init);
1162