linux/mm/slab.c
<<
>>
Prefs
   1/*
   2 * linux/mm/slab.c
   3 * Written by Mark Hemment, 1996/97.
   4 * (markhe@nextd.demon.co.uk)
   5 *
   6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
   7 *
   8 * Major cleanup, different bufctl logic, per-cpu arrays
   9 *      (c) 2000 Manfred Spraul
  10 *
  11 * Cleanup, make the head arrays unconditional, preparation for NUMA
  12 *      (c) 2002 Manfred Spraul
  13 *
  14 * An implementation of the Slab Allocator as described in outline in;
  15 *      UNIX Internals: The New Frontiers by Uresh Vahalia
  16 *      Pub: Prentice Hall      ISBN 0-13-101908-2
  17 * or with a little more detail in;
  18 *      The Slab Allocator: An Object-Caching Kernel Memory Allocator
  19 *      Jeff Bonwick (Sun Microsystems).
  20 *      Presented at: USENIX Summer 1994 Technical Conference
  21 *
  22 * The memory is organized in caches, one cache for each object type.
  23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
  24 * Each cache consists out of many slabs (they are small (usually one
  25 * page long) and always contiguous), and each slab contains multiple
  26 * initialized objects.
  27 *
  28 * This means, that your constructor is used only for newly allocated
  29 * slabs and you must pass objects with the same initializations to
  30 * kmem_cache_free.
  31 *
  32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
  33 * normal). If you need a special memory type, then must create a new
  34 * cache for that memory type.
  35 *
  36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
  37 *   full slabs with 0 free objects
  38 *   partial slabs
  39 *   empty slabs with no allocated objects
  40 *
  41 * If partial slabs exist, then new allocations come from these slabs,
  42 * otherwise from empty slabs or new slabs are allocated.
  43 *
  44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
  45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
  46 *
  47 * Each cache has a short per-cpu head array, most allocs
  48 * and frees go into that array, and if that array overflows, then 1/2
  49 * of the entries in the array are given back into the global cache.
  50 * The head array is strictly LIFO and should improve the cache hit rates.
  51 * On SMP, it additionally reduces the spinlock operations.
  52 *
  53 * The c_cpuarray may not be read with enabled local interrupts -
  54 * it's changed with a smp_call_function().
  55 *
  56 * SMP synchronization:
  57 *  constructors and destructors are called without any locking.
  58 *  Several members in struct kmem_cache and struct slab never change, they
  59 *      are accessed without any locking.
  60 *  The per-cpu arrays are never accessed from the wrong cpu, no locking,
  61 *      and local interrupts are disabled so slab code is preempt-safe.
  62 *  The non-constant members are protected with a per-cache irq spinlock.
  63 *
  64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
  65 * in 2000 - many ideas in the current implementation are derived from
  66 * his patch.
  67 *
  68 * Further notes from the original documentation:
  69 *
  70 * 11 April '97.  Started multi-threading - markhe
  71 *      The global cache-chain is protected by the mutex 'slab_mutex'.
  72 *      The sem is only needed when accessing/extending the cache-chain, which
  73 *      can never happen inside an interrupt (kmem_cache_create(),
  74 *      kmem_cache_shrink() and kmem_cache_reap()).
  75 *
  76 *      At present, each engine can be growing a cache.  This should be blocked.
  77 *
  78 * 15 March 2005. NUMA slab allocator.
  79 *      Shai Fultheim <shai@scalex86.org>.
  80 *      Shobhit Dayal <shobhit@calsoftinc.com>
  81 *      Alok N Kataria <alokk@calsoftinc.com>
  82 *      Christoph Lameter <christoph@lameter.com>
  83 *
  84 *      Modified the slab allocator to be node aware on NUMA systems.
  85 *      Each node has its own list of partial, free and full slabs.
  86 *      All object allocations for a node occur from node specific slab lists.
  87 */
  88
  89#include        <linux/slab.h>
  90#include        "slab.h"
  91#include        <linux/mm.h>
  92#include        <linux/poison.h>
  93#include        <linux/swap.h>
  94#include        <linux/cache.h>
  95#include        <linux/interrupt.h>
  96#include        <linux/init.h>
  97#include        <linux/compiler.h>
  98#include        <linux/cpuset.h>
  99#include        <linux/proc_fs.h>
 100#include        <linux/seq_file.h>
 101#include        <linux/notifier.h>
 102#include        <linux/kallsyms.h>
 103#include        <linux/cpu.h>
 104#include        <linux/sysctl.h>
 105#include        <linux/module.h>
 106#include        <linux/rcupdate.h>
 107#include        <linux/string.h>
 108#include        <linux/uaccess.h>
 109#include        <linux/nodemask.h>
 110#include        <linux/kmemleak.h>
 111#include        <linux/mempolicy.h>
 112#include        <linux/mutex.h>
 113#include        <linux/fault-inject.h>
 114#include        <linux/rtmutex.h>
 115#include        <linux/reciprocal_div.h>
 116#include        <linux/debugobjects.h>
 117#include        <linux/kmemcheck.h>
 118#include        <linux/memory.h>
 119#include        <linux/prefetch.h>
 120
 121#include        <net/sock.h>
 122
 123#include        <asm/cacheflush.h>
 124#include        <asm/tlbflush.h>
 125#include        <asm/page.h>
 126
 127#include <trace/events/kmem.h>
 128
 129#include        "internal.h"
 130
 131/*
 132 * DEBUG        - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
 133 *                0 for faster, smaller code (especially in the critical paths).
 134 *
 135 * STATS        - 1 to collect stats for /proc/slabinfo.
 136 *                0 for faster, smaller code (especially in the critical paths).
 137 *
 138 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
 139 */
 140
 141#ifdef CONFIG_DEBUG_SLAB
 142#define DEBUG           1
 143#define STATS           1
 144#define FORCED_DEBUG    1
 145#else
 146#define DEBUG           0
 147#define STATS           0
 148#define FORCED_DEBUG    0
 149#endif
 150
 151/* Shouldn't this be in a header file somewhere? */
 152#define BYTES_PER_WORD          sizeof(void *)
 153#define REDZONE_ALIGN           max(BYTES_PER_WORD, __alignof__(unsigned long long))
 154
 155#ifndef ARCH_KMALLOC_FLAGS
 156#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
 157#endif
 158
 159/*
 160 * true if a page was allocated from pfmemalloc reserves for network-based
 161 * swap
 162 */
 163static bool pfmemalloc_active __read_mostly;
 164
 165/* Legal flag mask for kmem_cache_create(). */
 166#if DEBUG
 167# define CREATE_MASK    (SLAB_RED_ZONE | \
 168                         SLAB_POISON | SLAB_HWCACHE_ALIGN | \
 169                         SLAB_CACHE_DMA | \
 170                         SLAB_STORE_USER | \
 171                         SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
 172                         SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
 173                         SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
 174#else
 175# define CREATE_MASK    (SLAB_HWCACHE_ALIGN | \
 176                         SLAB_CACHE_DMA | \
 177                         SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
 178                         SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
 179                         SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
 180#endif
 181
 182/*
 183 * kmem_bufctl_t:
 184 *
 185 * Bufctl's are used for linking objs within a slab
 186 * linked offsets.
 187 *
 188 * This implementation relies on "struct page" for locating the cache &
 189 * slab an object belongs to.
 190 * This allows the bufctl structure to be small (one int), but limits
 191 * the number of objects a slab (not a cache) can contain when off-slab
 192 * bufctls are used. The limit is the size of the largest general cache
 193 * that does not use off-slab slabs.
 194 * For 32bit archs with 4 kB pages, is this 56.
 195 * This is not serious, as it is only for large objects, when it is unwise
 196 * to have too many per slab.
 197 * Note: This limit can be raised by introducing a general cache whose size
 198 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
 199 */
 200
 201typedef unsigned int kmem_bufctl_t;
 202#define BUFCTL_END      (((kmem_bufctl_t)(~0U))-0)
 203#define BUFCTL_FREE     (((kmem_bufctl_t)(~0U))-1)
 204#define BUFCTL_ACTIVE   (((kmem_bufctl_t)(~0U))-2)
 205#define SLAB_LIMIT      (((kmem_bufctl_t)(~0U))-3)
 206
 207/*
 208 * struct slab_rcu
 209 *
 210 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
 211 * arrange for kmem_freepages to be called via RCU.  This is useful if
 212 * we need to approach a kernel structure obliquely, from its address
 213 * obtained without the usual locking.  We can lock the structure to
 214 * stabilize it and check it's still at the given address, only if we
 215 * can be sure that the memory has not been meanwhile reused for some
 216 * other kind of object (which our subsystem's lock might corrupt).
 217 *
 218 * rcu_read_lock before reading the address, then rcu_read_unlock after
 219 * taking the spinlock within the structure expected at that address.
 220 */
 221struct slab_rcu {
 222        struct rcu_head head;
 223        struct kmem_cache *cachep;
 224        void *addr;
 225};
 226
 227/*
 228 * struct slab
 229 *
 230 * Manages the objs in a slab. Placed either at the beginning of mem allocated
 231 * for a slab, or allocated from an general cache.
 232 * Slabs are chained into three list: fully used, partial, fully free slabs.
 233 */
 234struct slab {
 235        union {
 236                struct {
 237                        struct list_head list;
 238                        unsigned long colouroff;
 239                        void *s_mem;            /* including colour offset */
 240                        unsigned int inuse;     /* num of objs active in slab */
 241                        kmem_bufctl_t free;
 242                        unsigned short nodeid;
 243                };
 244                struct slab_rcu __slab_cover_slab_rcu;
 245        };
 246};
 247
 248/*
 249 * struct array_cache
 250 *
 251 * Purpose:
 252 * - LIFO ordering, to hand out cache-warm objects from _alloc
 253 * - reduce the number of linked list operations
 254 * - reduce spinlock operations
 255 *
 256 * The limit is stored in the per-cpu structure to reduce the data cache
 257 * footprint.
 258 *
 259 */
 260struct array_cache {
 261        unsigned int avail;
 262        unsigned int limit;
 263        unsigned int batchcount;
 264        unsigned int touched;
 265        spinlock_t lock;
 266        void *entry[];  /*
 267                         * Must have this definition in here for the proper
 268                         * alignment of array_cache. Also simplifies accessing
 269                         * the entries.
 270                         *
 271                         * Entries should not be directly dereferenced as
 272                         * entries belonging to slabs marked pfmemalloc will
 273                         * have the lower bits set SLAB_OBJ_PFMEMALLOC
 274                         */
 275};
 276
 277#define SLAB_OBJ_PFMEMALLOC     1
 278static inline bool is_obj_pfmemalloc(void *objp)
 279{
 280        return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
 281}
 282
 283static inline void set_obj_pfmemalloc(void **objp)
 284{
 285        *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
 286        return;
 287}
 288
 289static inline void clear_obj_pfmemalloc(void **objp)
 290{
 291        *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
 292}
 293
 294/*
 295 * bootstrap: The caches do not work without cpuarrays anymore, but the
 296 * cpuarrays are allocated from the generic caches...
 297 */
 298#define BOOT_CPUCACHE_ENTRIES   1
 299struct arraycache_init {
 300        struct array_cache cache;
 301        void *entries[BOOT_CPUCACHE_ENTRIES];
 302};
 303
 304/*
 305 * The slab lists for all objects.
 306 */
 307struct kmem_list3 {
 308        struct list_head slabs_partial; /* partial list first, better asm code */
 309        struct list_head slabs_full;
 310        struct list_head slabs_free;
 311        unsigned long free_objects;
 312        unsigned int free_limit;
 313        unsigned int colour_next;       /* Per-node cache coloring */
 314        spinlock_t list_lock;
 315        struct array_cache *shared;     /* shared per node */
 316        struct array_cache **alien;     /* on other nodes */
 317        unsigned long next_reap;        /* updated without locking */
 318        int free_touched;               /* updated without locking */
 319};
 320
 321/*
 322 * Need this for bootstrapping a per node allocator.
 323 */
 324#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
 325static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
 326#define CACHE_CACHE 0
 327#define SIZE_AC MAX_NUMNODES
 328#define SIZE_L3 (2 * MAX_NUMNODES)
 329
 330static int drain_freelist(struct kmem_cache *cache,
 331                        struct kmem_list3 *l3, int tofree);
 332static void free_block(struct kmem_cache *cachep, void **objpp, int len,
 333                        int node);
 334static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
 335static void cache_reap(struct work_struct *unused);
 336
 337/*
 338 * This function must be completely optimized away if a constant is passed to
 339 * it.  Mostly the same as what is in linux/slab.h except it returns an index.
 340 */
 341static __always_inline int index_of(const size_t size)
 342{
 343        extern void __bad_size(void);
 344
 345        if (__builtin_constant_p(size)) {
 346                int i = 0;
 347
 348#define CACHE(x) \
 349        if (size <=x) \
 350                return i; \
 351        else \
 352                i++;
 353#include <linux/kmalloc_sizes.h>
 354#undef CACHE
 355                __bad_size();
 356        } else
 357                __bad_size();
 358        return 0;
 359}
 360
 361static int slab_early_init = 1;
 362
 363#define INDEX_AC index_of(sizeof(struct arraycache_init))
 364#define INDEX_L3 index_of(sizeof(struct kmem_list3))
 365
 366static void kmem_list3_init(struct kmem_list3 *parent)
 367{
 368        INIT_LIST_HEAD(&parent->slabs_full);
 369        INIT_LIST_HEAD(&parent->slabs_partial);
 370        INIT_LIST_HEAD(&parent->slabs_free);
 371        parent->shared = NULL;
 372        parent->alien = NULL;
 373        parent->colour_next = 0;
 374        spin_lock_init(&parent->list_lock);
 375        parent->free_objects = 0;
 376        parent->free_touched = 0;
 377}
 378
 379#define MAKE_LIST(cachep, listp, slab, nodeid)                          \
 380        do {                                                            \
 381                INIT_LIST_HEAD(listp);                                  \
 382                list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
 383        } while (0)
 384
 385#define MAKE_ALL_LISTS(cachep, ptr, nodeid)                             \
 386        do {                                                            \
 387        MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid);  \
 388        MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
 389        MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid);  \
 390        } while (0)
 391
 392#define CFLGS_OFF_SLAB          (0x80000000UL)
 393#define OFF_SLAB(x)     ((x)->flags & CFLGS_OFF_SLAB)
 394
 395#define BATCHREFILL_LIMIT       16
 396/*
 397 * Optimization question: fewer reaps means less probability for unnessary
 398 * cpucache drain/refill cycles.
 399 *
 400 * OTOH the cpuarrays can contain lots of objects,
 401 * which could lock up otherwise freeable slabs.
 402 */
 403#define REAPTIMEOUT_CPUC        (2*HZ)
 404#define REAPTIMEOUT_LIST3       (4*HZ)
 405
 406#if STATS
 407#define STATS_INC_ACTIVE(x)     ((x)->num_active++)
 408#define STATS_DEC_ACTIVE(x)     ((x)->num_active--)
 409#define STATS_INC_ALLOCED(x)    ((x)->num_allocations++)
 410#define STATS_INC_GROWN(x)      ((x)->grown++)
 411#define STATS_ADD_REAPED(x,y)   ((x)->reaped += (y))
 412#define STATS_SET_HIGH(x)                                               \
 413        do {                                                            \
 414                if ((x)->num_active > (x)->high_mark)                   \
 415                        (x)->high_mark = (x)->num_active;               \
 416        } while (0)
 417#define STATS_INC_ERR(x)        ((x)->errors++)
 418#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
 419#define STATS_INC_NODEFREES(x)  ((x)->node_frees++)
 420#define STATS_INC_ACOVERFLOW(x)   ((x)->node_overflow++)
 421#define STATS_SET_FREEABLE(x, i)                                        \
 422        do {                                                            \
 423                if ((x)->max_freeable < i)                              \
 424                        (x)->max_freeable = i;                          \
 425        } while (0)
 426#define STATS_INC_ALLOCHIT(x)   atomic_inc(&(x)->allochit)
 427#define STATS_INC_ALLOCMISS(x)  atomic_inc(&(x)->allocmiss)
 428#define STATS_INC_FREEHIT(x)    atomic_inc(&(x)->freehit)
 429#define STATS_INC_FREEMISS(x)   atomic_inc(&(x)->freemiss)
 430#else
 431#define STATS_INC_ACTIVE(x)     do { } while (0)
 432#define STATS_DEC_ACTIVE(x)     do { } while (0)
 433#define STATS_INC_ALLOCED(x)    do { } while (0)
 434#define STATS_INC_GROWN(x)      do { } while (0)
 435#define STATS_ADD_REAPED(x,y)   do { (void)(y); } while (0)
 436#define STATS_SET_HIGH(x)       do { } while (0)
 437#define STATS_INC_ERR(x)        do { } while (0)
 438#define STATS_INC_NODEALLOCS(x) do { } while (0)
 439#define STATS_INC_NODEFREES(x)  do { } while (0)
 440#define STATS_INC_ACOVERFLOW(x)   do { } while (0)
 441#define STATS_SET_FREEABLE(x, i) do { } while (0)
 442#define STATS_INC_ALLOCHIT(x)   do { } while (0)
 443#define STATS_INC_ALLOCMISS(x)  do { } while (0)
 444#define STATS_INC_FREEHIT(x)    do { } while (0)
 445#define STATS_INC_FREEMISS(x)   do { } while (0)
 446#endif
 447
 448#if DEBUG
 449
 450/*
 451 * memory layout of objects:
 452 * 0            : objp
 453 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
 454 *              the end of an object is aligned with the end of the real
 455 *              allocation. Catches writes behind the end of the allocation.
 456 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
 457 *              redzone word.
 458 * cachep->obj_offset: The real object.
 459 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
 460 * cachep->size - 1* BYTES_PER_WORD: last caller address
 461 *                                      [BYTES_PER_WORD long]
 462 */
 463static int obj_offset(struct kmem_cache *cachep)
 464{
 465        return cachep->obj_offset;
 466}
 467
 468static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
 469{
 470        BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
 471        return (unsigned long long*) (objp + obj_offset(cachep) -
 472                                      sizeof(unsigned long long));
 473}
 474
 475static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
 476{
 477        BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
 478        if (cachep->flags & SLAB_STORE_USER)
 479                return (unsigned long long *)(objp + cachep->size -
 480                                              sizeof(unsigned long long) -
 481                                              REDZONE_ALIGN);
 482        return (unsigned long long *) (objp + cachep->size -
 483                                       sizeof(unsigned long long));
 484}
 485
 486static void **dbg_userword(struct kmem_cache *cachep, void *objp)
 487{
 488        BUG_ON(!(cachep->flags & SLAB_STORE_USER));
 489        return (void **)(objp + cachep->size - BYTES_PER_WORD);
 490}
 491
 492#else
 493
 494#define obj_offset(x)                   0
 495#define dbg_redzone1(cachep, objp)      ({BUG(); (unsigned long long *)NULL;})
 496#define dbg_redzone2(cachep, objp)      ({BUG(); (unsigned long long *)NULL;})
 497#define dbg_userword(cachep, objp)      ({BUG(); (void **)NULL;})
 498
 499#endif
 500
 501#ifdef CONFIG_TRACING
 502size_t slab_buffer_size(struct kmem_cache *cachep)
 503{
 504        return cachep->size;
 505}
 506EXPORT_SYMBOL(slab_buffer_size);
 507#endif
 508
 509/*
 510 * Do not go above this order unless 0 objects fit into the slab or
 511 * overridden on the command line.
 512 */
 513#define SLAB_MAX_ORDER_HI       1
 514#define SLAB_MAX_ORDER_LO       0
 515static int slab_max_order = SLAB_MAX_ORDER_LO;
 516static bool slab_max_order_set __initdata;
 517
 518static inline struct kmem_cache *page_get_cache(struct page *page)
 519{
 520        page = compound_head(page);
 521        BUG_ON(!PageSlab(page));
 522        return page->slab_cache;
 523}
 524
 525static inline struct kmem_cache *virt_to_cache(const void *obj)
 526{
 527        struct page *page = virt_to_head_page(obj);
 528        return page->slab_cache;
 529}
 530
 531static inline struct slab *virt_to_slab(const void *obj)
 532{
 533        struct page *page = virt_to_head_page(obj);
 534
 535        VM_BUG_ON(!PageSlab(page));
 536        return page->slab_page;
 537}
 538
 539static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
 540                                 unsigned int idx)
 541{
 542        return slab->s_mem + cache->size * idx;
 543}
 544
 545/*
 546 * We want to avoid an expensive divide : (offset / cache->size)
 547 *   Using the fact that size is a constant for a particular cache,
 548 *   we can replace (offset / cache->size) by
 549 *   reciprocal_divide(offset, cache->reciprocal_buffer_size)
 550 */
 551static inline unsigned int obj_to_index(const struct kmem_cache *cache,
 552                                        const struct slab *slab, void *obj)
 553{
 554        u32 offset = (obj - slab->s_mem);
 555        return reciprocal_divide(offset, cache->reciprocal_buffer_size);
 556}
 557
 558/*
 559 * These are the default caches for kmalloc. Custom caches can have other sizes.
 560 */
 561struct cache_sizes malloc_sizes[] = {
 562#define CACHE(x) { .cs_size = (x) },
 563#include <linux/kmalloc_sizes.h>
 564        CACHE(ULONG_MAX)
 565#undef CACHE
 566};
 567EXPORT_SYMBOL(malloc_sizes);
 568
 569/* Must match cache_sizes above. Out of line to keep cache footprint low. */
 570struct cache_names {
 571        char *name;
 572        char *name_dma;
 573};
 574
 575static struct cache_names __initdata cache_names[] = {
 576#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
 577#include <linux/kmalloc_sizes.h>
 578        {NULL,}
 579#undef CACHE
 580};
 581
 582static struct arraycache_init initarray_cache __initdata =
 583    { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
 584static struct arraycache_init initarray_generic =
 585    { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
 586
 587/* internal cache of cache description objs */
 588static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
 589static struct kmem_cache cache_cache = {
 590        .nodelists = cache_cache_nodelists,
 591        .batchcount = 1,
 592        .limit = BOOT_CPUCACHE_ENTRIES,
 593        .shared = 1,
 594        .size = sizeof(struct kmem_cache),
 595        .name = "kmem_cache",
 596};
 597
 598#define BAD_ALIEN_MAGIC 0x01020304ul
 599
 600#ifdef CONFIG_LOCKDEP
 601
 602/*
 603 * Slab sometimes uses the kmalloc slabs to store the slab headers
 604 * for other slabs "off slab".
 605 * The locking for this is tricky in that it nests within the locks
 606 * of all other slabs in a few places; to deal with this special
 607 * locking we put on-slab caches into a separate lock-class.
 608 *
 609 * We set lock class for alien array caches which are up during init.
 610 * The lock annotation will be lost if all cpus of a node goes down and
 611 * then comes back up during hotplug
 612 */
 613static struct lock_class_key on_slab_l3_key;
 614static struct lock_class_key on_slab_alc_key;
 615
 616static struct lock_class_key debugobj_l3_key;
 617static struct lock_class_key debugobj_alc_key;
 618
 619static void slab_set_lock_classes(struct kmem_cache *cachep,
 620                struct lock_class_key *l3_key, struct lock_class_key *alc_key,
 621                int q)
 622{
 623        struct array_cache **alc;
 624        struct kmem_list3 *l3;
 625        int r;
 626
 627        l3 = cachep->nodelists[q];
 628        if (!l3)
 629                return;
 630
 631        lockdep_set_class(&l3->list_lock, l3_key);
 632        alc = l3->alien;
 633        /*
 634         * FIXME: This check for BAD_ALIEN_MAGIC
 635         * should go away when common slab code is taught to
 636         * work even without alien caches.
 637         * Currently, non NUMA code returns BAD_ALIEN_MAGIC
 638         * for alloc_alien_cache,
 639         */
 640        if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
 641                return;
 642        for_each_node(r) {
 643                if (alc[r])
 644                        lockdep_set_class(&alc[r]->lock, alc_key);
 645        }
 646}
 647
 648static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
 649{
 650        slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
 651}
 652
 653static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
 654{
 655        int node;
 656
 657        for_each_online_node(node)
 658                slab_set_debugobj_lock_classes_node(cachep, node);
 659}
 660
 661static void init_node_lock_keys(int q)
 662{
 663        struct cache_sizes *s = malloc_sizes;
 664
 665        if (slab_state < UP)
 666                return;
 667
 668        for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
 669                struct kmem_list3 *l3;
 670
 671                l3 = s->cs_cachep->nodelists[q];
 672                if (!l3 || OFF_SLAB(s->cs_cachep))
 673                        continue;
 674
 675                slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
 676                                &on_slab_alc_key, q);
 677        }
 678}
 679
 680static inline void init_lock_keys(void)
 681{
 682        int node;
 683
 684        for_each_node(node)
 685                init_node_lock_keys(node);
 686}
 687#else
 688static void init_node_lock_keys(int q)
 689{
 690}
 691
 692static inline void init_lock_keys(void)
 693{
 694}
 695
 696static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
 697{
 698}
 699
 700static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
 701{
 702}
 703#endif
 704
 705static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
 706
 707static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
 708{
 709        return cachep->array[smp_processor_id()];
 710}
 711
 712static inline struct kmem_cache *__find_general_cachep(size_t size,
 713                                                        gfp_t gfpflags)
 714{
 715        struct cache_sizes *csizep = malloc_sizes;
 716
 717#if DEBUG
 718        /* This happens if someone tries to call
 719         * kmem_cache_create(), or __kmalloc(), before
 720         * the generic caches are initialized.
 721         */
 722        BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
 723#endif
 724        if (!size)
 725                return ZERO_SIZE_PTR;
 726
 727        while (size > csizep->cs_size)
 728                csizep++;
 729
 730        /*
 731         * Really subtle: The last entry with cs->cs_size==ULONG_MAX
 732         * has cs_{dma,}cachep==NULL. Thus no special case
 733         * for large kmalloc calls required.
 734         */
 735#ifdef CONFIG_ZONE_DMA
 736        if (unlikely(gfpflags & GFP_DMA))
 737                return csizep->cs_dmacachep;
 738#endif
 739        return csizep->cs_cachep;
 740}
 741
 742static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
 743{
 744        return __find_general_cachep(size, gfpflags);
 745}
 746
 747static size_t slab_mgmt_size(size_t nr_objs, size_t align)
 748{
 749        return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
 750}
 751
 752/*
 753 * Calculate the number of objects and left-over bytes for a given buffer size.
 754 */
 755static void cache_estimate(unsigned long gfporder, size_t buffer_size,
 756                           size_t align, int flags, size_t *left_over,
 757                           unsigned int *num)
 758{
 759        int nr_objs;
 760        size_t mgmt_size;
 761        size_t slab_size = PAGE_SIZE << gfporder;
 762
 763        /*
 764         * The slab management structure can be either off the slab or
 765         * on it. For the latter case, the memory allocated for a
 766         * slab is used for:
 767         *
 768         * - The struct slab
 769         * - One kmem_bufctl_t for each object
 770         * - Padding to respect alignment of @align
 771         * - @buffer_size bytes for each object
 772         *
 773         * If the slab management structure is off the slab, then the
 774         * alignment will already be calculated into the size. Because
 775         * the slabs are all pages aligned, the objects will be at the
 776         * correct alignment when allocated.
 777         */
 778        if (flags & CFLGS_OFF_SLAB) {
 779                mgmt_size = 0;
 780                nr_objs = slab_size / buffer_size;
 781
 782                if (nr_objs > SLAB_LIMIT)
 783                        nr_objs = SLAB_LIMIT;
 784        } else {
 785                /*
 786                 * Ignore padding for the initial guess. The padding
 787                 * is at most @align-1 bytes, and @buffer_size is at
 788                 * least @align. In the worst case, this result will
 789                 * be one greater than the number of objects that fit
 790                 * into the memory allocation when taking the padding
 791                 * into account.
 792                 */
 793                nr_objs = (slab_size - sizeof(struct slab)) /
 794                          (buffer_size + sizeof(kmem_bufctl_t));
 795
 796                /*
 797                 * This calculated number will be either the right
 798                 * amount, or one greater than what we want.
 799                 */
 800                if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
 801                       > slab_size)
 802                        nr_objs--;
 803
 804                if (nr_objs > SLAB_LIMIT)
 805                        nr_objs = SLAB_LIMIT;
 806
 807                mgmt_size = slab_mgmt_size(nr_objs, align);
 808        }
 809        *num = nr_objs;
 810        *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
 811}
 812
 813#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
 814
 815static void __slab_error(const char *function, struct kmem_cache *cachep,
 816                        char *msg)
 817{
 818        printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
 819               function, cachep->name, msg);
 820        dump_stack();
 821}
 822
 823/*
 824 * By default on NUMA we use alien caches to stage the freeing of
 825 * objects allocated from other nodes. This causes massive memory
 826 * inefficiencies when using fake NUMA setup to split memory into a
 827 * large number of small nodes, so it can be disabled on the command
 828 * line
 829  */
 830
 831static int use_alien_caches __read_mostly = 1;
 832static int __init noaliencache_setup(char *s)
 833{
 834        use_alien_caches = 0;
 835        return 1;
 836}
 837__setup("noaliencache", noaliencache_setup);
 838
 839static int __init slab_max_order_setup(char *str)
 840{
 841        get_option(&str, &slab_max_order);
 842        slab_max_order = slab_max_order < 0 ? 0 :
 843                                min(slab_max_order, MAX_ORDER - 1);
 844        slab_max_order_set = true;
 845
 846        return 1;
 847}
 848__setup("slab_max_order=", slab_max_order_setup);
 849
 850#ifdef CONFIG_NUMA
 851/*
 852 * Special reaping functions for NUMA systems called from cache_reap().
 853 * These take care of doing round robin flushing of alien caches (containing
 854 * objects freed on different nodes from which they were allocated) and the
 855 * flushing of remote pcps by calling drain_node_pages.
 856 */
 857static DEFINE_PER_CPU(unsigned long, slab_reap_node);
 858
 859static void init_reap_node(int cpu)
 860{
 861        int node;
 862
 863        node = next_node(cpu_to_mem(cpu), node_online_map);
 864        if (node == MAX_NUMNODES)
 865                node = first_node(node_online_map);
 866
 867        per_cpu(slab_reap_node, cpu) = node;
 868}
 869
 870static void next_reap_node(void)
 871{
 872        int node = __this_cpu_read(slab_reap_node);
 873
 874        node = next_node(node, node_online_map);
 875        if (unlikely(node >= MAX_NUMNODES))
 876                node = first_node(node_online_map);
 877        __this_cpu_write(slab_reap_node, node);
 878}
 879
 880#else
 881#define init_reap_node(cpu) do { } while (0)
 882#define next_reap_node(void) do { } while (0)
 883#endif
 884
 885/*
 886 * Initiate the reap timer running on the target CPU.  We run at around 1 to 2Hz
 887 * via the workqueue/eventd.
 888 * Add the CPU number into the expiration time to minimize the possibility of
 889 * the CPUs getting into lockstep and contending for the global cache chain
 890 * lock.
 891 */
 892static void __cpuinit start_cpu_timer(int cpu)
 893{
 894        struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
 895
 896        /*
 897         * When this gets called from do_initcalls via cpucache_init(),
 898         * init_workqueues() has already run, so keventd will be setup
 899         * at that time.
 900         */
 901        if (keventd_up() && reap_work->work.func == NULL) {
 902                init_reap_node(cpu);
 903                INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
 904                schedule_delayed_work_on(cpu, reap_work,
 905                                        __round_jiffies_relative(HZ, cpu));
 906        }
 907}
 908
 909static struct array_cache *alloc_arraycache(int node, int entries,
 910                                            int batchcount, gfp_t gfp)
 911{
 912        int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
 913        struct array_cache *nc = NULL;
 914
 915        nc = kmalloc_node(memsize, gfp, node);
 916        /*
 917         * The array_cache structures contain pointers to free object.
 918         * However, when such objects are allocated or transferred to another
 919         * cache the pointers are not cleared and they could be counted as
 920         * valid references during a kmemleak scan. Therefore, kmemleak must
 921         * not scan such objects.
 922         */
 923        kmemleak_no_scan(nc);
 924        if (nc) {
 925                nc->avail = 0;
 926                nc->limit = entries;
 927                nc->batchcount = batchcount;
 928                nc->touched = 0;
 929                spin_lock_init(&nc->lock);
 930        }
 931        return nc;
 932}
 933
 934static inline bool is_slab_pfmemalloc(struct slab *slabp)
 935{
 936        struct page *page = virt_to_page(slabp->s_mem);
 937
 938        return PageSlabPfmemalloc(page);
 939}
 940
 941/* Clears pfmemalloc_active if no slabs have pfmalloc set */
 942static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
 943                                                struct array_cache *ac)
 944{
 945        struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()];
 946        struct slab *slabp;
 947        unsigned long flags;
 948
 949        if (!pfmemalloc_active)
 950                return;
 951
 952        spin_lock_irqsave(&l3->list_lock, flags);
 953        list_for_each_entry(slabp, &l3->slabs_full, list)
 954                if (is_slab_pfmemalloc(slabp))
 955                        goto out;
 956
 957        list_for_each_entry(slabp, &l3->slabs_partial, list)
 958                if (is_slab_pfmemalloc(slabp))
 959                        goto out;
 960
 961        list_for_each_entry(slabp, &l3->slabs_free, list)
 962                if (is_slab_pfmemalloc(slabp))
 963                        goto out;
 964
 965        pfmemalloc_active = false;
 966out:
 967        spin_unlock_irqrestore(&l3->list_lock, flags);
 968}
 969
 970static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
 971                                                gfp_t flags, bool force_refill)
 972{
 973        int i;
 974        void *objp = ac->entry[--ac->avail];
 975
 976        /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
 977        if (unlikely(is_obj_pfmemalloc(objp))) {
 978                struct kmem_list3 *l3;
 979
 980                if (gfp_pfmemalloc_allowed(flags)) {
 981                        clear_obj_pfmemalloc(&objp);
 982                        return objp;
 983                }
 984
 985                /* The caller cannot use PFMEMALLOC objects, find another one */
 986                for (i = 0; i < ac->avail; i++) {
 987                        /* If a !PFMEMALLOC object is found, swap them */
 988                        if (!is_obj_pfmemalloc(ac->entry[i])) {
 989                                objp = ac->entry[i];
 990                                ac->entry[i] = ac->entry[ac->avail];
 991                                ac->entry[ac->avail] = objp;
 992                                return objp;
 993                        }
 994                }
 995
 996                /*
 997                 * If there are empty slabs on the slabs_free list and we are
 998                 * being forced to refill the cache, mark this one !pfmemalloc.
 999                 */
1000                l3 = cachep->nodelists[numa_mem_id()];
1001                if (!list_empty(&l3->slabs_free) && force_refill) {
1002                        struct slab *slabp = virt_to_slab(objp);
1003                        ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
1004                        clear_obj_pfmemalloc(&objp);
1005                        recheck_pfmemalloc_active(cachep, ac);
1006                        return objp;
1007                }
1008
1009                /* No !PFMEMALLOC objects available */
1010                ac->avail++;
1011                objp = NULL;
1012        }
1013
1014        return objp;
1015}
1016
1017static inline void *ac_get_obj(struct kmem_cache *cachep,
1018                        struct array_cache *ac, gfp_t flags, bool force_refill)
1019{
1020        void *objp;
1021
1022        if (unlikely(sk_memalloc_socks()))
1023                objp = __ac_get_obj(cachep, ac, flags, force_refill);
1024        else
1025                objp = ac->entry[--ac->avail];
1026
1027        return objp;
1028}
1029
1030static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1031                                                                void *objp)
1032{
1033        if (unlikely(pfmemalloc_active)) {
1034                /* Some pfmemalloc slabs exist, check if this is one */
1035                struct page *page = virt_to_head_page(objp);
1036                if (PageSlabPfmemalloc(page))
1037                        set_obj_pfmemalloc(&objp);
1038        }
1039
1040        return objp;
1041}
1042
1043static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1044                                                                void *objp)
1045{
1046        if (unlikely(sk_memalloc_socks()))
1047                objp = __ac_put_obj(cachep, ac, objp);
1048
1049        ac->entry[ac->avail++] = objp;
1050}
1051
1052/*
1053 * Transfer objects in one arraycache to another.
1054 * Locking must be handled by the caller.
1055 *
1056 * Return the number of entries transferred.
1057 */
1058static int transfer_objects(struct array_cache *to,
1059                struct array_cache *from, unsigned int max)
1060{
1061        /* Figure out how many entries to transfer */
1062        int nr = min3(from->avail, max, to->limit - to->avail);
1063
1064        if (!nr)
1065                return 0;
1066
1067        memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1068                        sizeof(void *) *nr);
1069
1070        from->avail -= nr;
1071        to->avail += nr;
1072        return nr;
1073}
1074
1075#ifndef CONFIG_NUMA
1076
1077#define drain_alien_cache(cachep, alien) do { } while (0)
1078#define reap_alien(cachep, l3) do { } while (0)
1079
1080static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1081{
1082        return (struct array_cache **)BAD_ALIEN_MAGIC;
1083}
1084
1085static inline void free_alien_cache(struct array_cache **ac_ptr)
1086{
1087}
1088
1089static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1090{
1091        return 0;
1092}
1093
1094static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1095                gfp_t flags)
1096{
1097        return NULL;
1098}
1099
1100static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1101                 gfp_t flags, int nodeid)
1102{
1103        return NULL;
1104}
1105
1106#else   /* CONFIG_NUMA */
1107
1108static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1109static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1110
1111static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1112{
1113        struct array_cache **ac_ptr;
1114        int memsize = sizeof(void *) * nr_node_ids;
1115        int i;
1116
1117        if (limit > 1)
1118                limit = 12;
1119        ac_ptr = kzalloc_node(memsize, gfp, node);
1120        if (ac_ptr) {
1121                for_each_node(i) {
1122                        if (i == node || !node_online(i))
1123                                continue;
1124                        ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1125                        if (!ac_ptr[i]) {
1126                                for (i--; i >= 0; i--)
1127                                        kfree(ac_ptr[i]);
1128                                kfree(ac_ptr);
1129                                return NULL;
1130                        }
1131                }
1132        }
1133        return ac_ptr;
1134}
1135
1136static void free_alien_cache(struct array_cache **ac_ptr)
1137{
1138        int i;
1139
1140        if (!ac_ptr)
1141                return;
1142        for_each_node(i)
1143            kfree(ac_ptr[i]);
1144        kfree(ac_ptr);
1145}
1146
1147static void __drain_alien_cache(struct kmem_cache *cachep,
1148                                struct array_cache *ac, int node)
1149{
1150        struct kmem_list3 *rl3 = cachep->nodelists[node];
1151
1152        if (ac->avail) {
1153                spin_lock(&rl3->list_lock);
1154                /*
1155                 * Stuff objects into the remote nodes shared array first.
1156                 * That way we could avoid the overhead of putting the objects
1157                 * into the free lists and getting them back later.
1158                 */
1159                if (rl3->shared)
1160                        transfer_objects(rl3->shared, ac, ac->limit);
1161
1162                free_block(cachep, ac->entry, ac->avail, node);
1163                ac->avail = 0;
1164                spin_unlock(&rl3->list_lock);
1165        }
1166}
1167
1168/*
1169 * Called from cache_reap() to regularly drain alien caches round robin.
1170 */
1171static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1172{
1173        int node = __this_cpu_read(slab_reap_node);
1174
1175        if (l3->alien) {
1176                struct array_cache *ac = l3->alien[node];
1177
1178                if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1179                        __drain_alien_cache(cachep, ac, node);
1180                        spin_unlock_irq(&ac->lock);
1181                }
1182        }
1183}
1184
1185static void drain_alien_cache(struct kmem_cache *cachep,
1186                                struct array_cache **alien)
1187{
1188        int i = 0;
1189        struct array_cache *ac;
1190        unsigned long flags;
1191
1192        for_each_online_node(i) {
1193                ac = alien[i];
1194                if (ac) {
1195                        spin_lock_irqsave(&ac->lock, flags);
1196                        __drain_alien_cache(cachep, ac, i);
1197                        spin_unlock_irqrestore(&ac->lock, flags);
1198                }
1199        }
1200}
1201
1202static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1203{
1204        struct slab *slabp = virt_to_slab(objp);
1205        int nodeid = slabp->nodeid;
1206        struct kmem_list3 *l3;
1207        struct array_cache *alien = NULL;
1208        int node;
1209
1210        node = numa_mem_id();
1211
1212        /*
1213         * Make sure we are not freeing a object from another node to the array
1214         * cache on this cpu.
1215         */
1216        if (likely(slabp->nodeid == node))
1217                return 0;
1218
1219        l3 = cachep->nodelists[node];
1220        STATS_INC_NODEFREES(cachep);
1221        if (l3->alien && l3->alien[nodeid]) {
1222                alien = l3->alien[nodeid];
1223                spin_lock(&alien->lock);
1224                if (unlikely(alien->avail == alien->limit)) {
1225                        STATS_INC_ACOVERFLOW(cachep);
1226                        __drain_alien_cache(cachep, alien, nodeid);
1227                }
1228                ac_put_obj(cachep, alien, objp);
1229                spin_unlock(&alien->lock);
1230        } else {
1231                spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1232                free_block(cachep, &objp, 1, nodeid);
1233                spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1234        }
1235        return 1;
1236}
1237#endif
1238
1239/*
1240 * Allocates and initializes nodelists for a node on each slab cache, used for
1241 * either memory or cpu hotplug.  If memory is being hot-added, the kmem_list3
1242 * will be allocated off-node since memory is not yet online for the new node.
1243 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1244 * already in use.
1245 *
1246 * Must hold slab_mutex.
1247 */
1248static int init_cache_nodelists_node(int node)
1249{
1250        struct kmem_cache *cachep;
1251        struct kmem_list3 *l3;
1252        const int memsize = sizeof(struct kmem_list3);
1253
1254        list_for_each_entry(cachep, &slab_caches, list) {
1255                /*
1256                 * Set up the size64 kmemlist for cpu before we can
1257                 * begin anything. Make sure some other cpu on this
1258                 * node has not already allocated this
1259                 */
1260                if (!cachep->nodelists[node]) {
1261                        l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1262                        if (!l3)
1263                                return -ENOMEM;
1264                        kmem_list3_init(l3);
1265                        l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1266                            ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1267
1268                        /*
1269                         * The l3s don't come and go as CPUs come and
1270                         * go.  slab_mutex is sufficient
1271                         * protection here.
1272                         */
1273                        cachep->nodelists[node] = l3;
1274                }
1275
1276                spin_lock_irq(&cachep->nodelists[node]->list_lock);
1277                cachep->nodelists[node]->free_limit =
1278                        (1 + nr_cpus_node(node)) *
1279                        cachep->batchcount + cachep->num;
1280                spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1281        }
1282        return 0;
1283}
1284
1285static void __cpuinit cpuup_canceled(long cpu)
1286{
1287        struct kmem_cache *cachep;
1288        struct kmem_list3 *l3 = NULL;
1289        int node = cpu_to_mem(cpu);
1290        const struct cpumask *mask = cpumask_of_node(node);
1291
1292        list_for_each_entry(cachep, &slab_caches, list) {
1293                struct array_cache *nc;
1294                struct array_cache *shared;
1295                struct array_cache **alien;
1296
1297                /* cpu is dead; no one can alloc from it. */
1298                nc = cachep->array[cpu];
1299                cachep->array[cpu] = NULL;
1300                l3 = cachep->nodelists[node];
1301
1302                if (!l3)
1303                        goto free_array_cache;
1304
1305                spin_lock_irq(&l3->list_lock);
1306
1307                /* Free limit for this kmem_list3 */
1308                l3->free_limit -= cachep->batchcount;
1309                if (nc)
1310                        free_block(cachep, nc->entry, nc->avail, node);
1311
1312                if (!cpumask_empty(mask)) {
1313                        spin_unlock_irq(&l3->list_lock);
1314                        goto free_array_cache;
1315                }
1316
1317                shared = l3->shared;
1318                if (shared) {
1319                        free_block(cachep, shared->entry,
1320                                   shared->avail, node);
1321                        l3->shared = NULL;
1322                }
1323
1324                alien = l3->alien;
1325                l3->alien = NULL;
1326
1327                spin_unlock_irq(&l3->list_lock);
1328
1329                kfree(shared);
1330                if (alien) {
1331                        drain_alien_cache(cachep, alien);
1332                        free_alien_cache(alien);
1333                }
1334free_array_cache:
1335                kfree(nc);
1336        }
1337        /*
1338         * In the previous loop, all the objects were freed to
1339         * the respective cache's slabs,  now we can go ahead and
1340         * shrink each nodelist to its limit.
1341         */
1342        list_for_each_entry(cachep, &slab_caches, list) {
1343                l3 = cachep->nodelists[node];
1344                if (!l3)
1345                        continue;
1346                drain_freelist(cachep, l3, l3->free_objects);
1347        }
1348}
1349
1350static int __cpuinit cpuup_prepare(long cpu)
1351{
1352        struct kmem_cache *cachep;
1353        struct kmem_list3 *l3 = NULL;
1354        int node = cpu_to_mem(cpu);
1355        int err;
1356
1357        /*
1358         * We need to do this right in the beginning since
1359         * alloc_arraycache's are going to use this list.
1360         * kmalloc_node allows us to add the slab to the right
1361         * kmem_list3 and not this cpu's kmem_list3
1362         */
1363        err = init_cache_nodelists_node(node);
1364        if (err < 0)
1365                goto bad;
1366
1367        /*
1368         * Now we can go ahead with allocating the shared arrays and
1369         * array caches
1370         */
1371        list_for_each_entry(cachep, &slab_caches, list) {
1372                struct array_cache *nc;
1373                struct array_cache *shared = NULL;
1374                struct array_cache **alien = NULL;
1375
1376                nc = alloc_arraycache(node, cachep->limit,
1377                                        cachep->batchcount, GFP_KERNEL);
1378                if (!nc)
1379                        goto bad;
1380                if (cachep->shared) {
1381                        shared = alloc_arraycache(node,
1382                                cachep->shared * cachep->batchcount,
1383                                0xbaadf00d, GFP_KERNEL);
1384                        if (!shared) {
1385                                kfree(nc);
1386                                goto bad;
1387                        }
1388                }
1389                if (use_alien_caches) {
1390                        alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1391                        if (!alien) {
1392                                kfree(shared);
1393                                kfree(nc);
1394                                goto bad;
1395                        }
1396                }
1397                cachep->array[cpu] = nc;
1398                l3 = cachep->nodelists[node];
1399                BUG_ON(!l3);
1400
1401                spin_lock_irq(&l3->list_lock);
1402                if (!l3->shared) {
1403                        /*
1404                         * We are serialised from CPU_DEAD or
1405                         * CPU_UP_CANCELLED by the cpucontrol lock
1406                         */
1407                        l3->shared = shared;
1408                        shared = NULL;
1409                }
1410#ifdef CONFIG_NUMA
1411                if (!l3->alien) {
1412                        l3->alien = alien;
1413                        alien = NULL;
1414                }
1415#endif
1416                spin_unlock_irq(&l3->list_lock);
1417                kfree(shared);
1418                free_alien_cache(alien);
1419                if (cachep->flags & SLAB_DEBUG_OBJECTS)
1420                        slab_set_debugobj_lock_classes_node(cachep, node);
1421        }
1422        init_node_lock_keys(node);
1423
1424        return 0;
1425bad:
1426        cpuup_canceled(cpu);
1427        return -ENOMEM;
1428}
1429
1430static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1431                                    unsigned long action, void *hcpu)
1432{
1433        long cpu = (long)hcpu;
1434        int err = 0;
1435
1436        switch (action) {
1437        case CPU_UP_PREPARE:
1438        case CPU_UP_PREPARE_FROZEN:
1439                mutex_lock(&slab_mutex);
1440                err = cpuup_prepare(cpu);
1441                mutex_unlock(&slab_mutex);
1442                break;
1443        case CPU_ONLINE:
1444        case CPU_ONLINE_FROZEN:
1445                start_cpu_timer(cpu);
1446                break;
1447#ifdef CONFIG_HOTPLUG_CPU
1448        case CPU_DOWN_PREPARE:
1449        case CPU_DOWN_PREPARE_FROZEN:
1450                /*
1451                 * Shutdown cache reaper. Note that the slab_mutex is
1452                 * held so that if cache_reap() is invoked it cannot do
1453                 * anything expensive but will only modify reap_work
1454                 * and reschedule the timer.
1455                */
1456                cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1457                /* Now the cache_reaper is guaranteed to be not running. */
1458                per_cpu(slab_reap_work, cpu).work.func = NULL;
1459                break;
1460        case CPU_DOWN_FAILED:
1461        case CPU_DOWN_FAILED_FROZEN:
1462                start_cpu_timer(cpu);
1463                break;
1464        case CPU_DEAD:
1465        case CPU_DEAD_FROZEN:
1466                /*
1467                 * Even if all the cpus of a node are down, we don't free the
1468                 * kmem_list3 of any cache. This to avoid a race between
1469                 * cpu_down, and a kmalloc allocation from another cpu for
1470                 * memory from the node of the cpu going down.  The list3
1471                 * structure is usually allocated from kmem_cache_create() and
1472                 * gets destroyed at kmem_cache_destroy().
1473                 */
1474                /* fall through */
1475#endif
1476        case CPU_UP_CANCELED:
1477        case CPU_UP_CANCELED_FROZEN:
1478                mutex_lock(&slab_mutex);
1479                cpuup_canceled(cpu);
1480                mutex_unlock(&slab_mutex);
1481                break;
1482        }
1483        return notifier_from_errno(err);
1484}
1485
1486static struct notifier_block __cpuinitdata cpucache_notifier = {
1487        &cpuup_callback, NULL, 0
1488};
1489
1490#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1491/*
1492 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1493 * Returns -EBUSY if all objects cannot be drained so that the node is not
1494 * removed.
1495 *
1496 * Must hold slab_mutex.
1497 */
1498static int __meminit drain_cache_nodelists_node(int node)
1499{
1500        struct kmem_cache *cachep;
1501        int ret = 0;
1502
1503        list_for_each_entry(cachep, &slab_caches, list) {
1504                struct kmem_list3 *l3;
1505
1506                l3 = cachep->nodelists[node];
1507                if (!l3)
1508                        continue;
1509
1510                drain_freelist(cachep, l3, l3->free_objects);
1511
1512                if (!list_empty(&l3->slabs_full) ||
1513                    !list_empty(&l3->slabs_partial)) {
1514                        ret = -EBUSY;
1515                        break;
1516                }
1517        }
1518        return ret;
1519}
1520
1521static int __meminit slab_memory_callback(struct notifier_block *self,
1522                                        unsigned long action, void *arg)
1523{
1524        struct memory_notify *mnb = arg;
1525        int ret = 0;
1526        int nid;
1527
1528        nid = mnb->status_change_nid;
1529        if (nid < 0)
1530                goto out;
1531
1532        switch (action) {
1533        case MEM_GOING_ONLINE:
1534                mutex_lock(&slab_mutex);
1535                ret = init_cache_nodelists_node(nid);
1536                mutex_unlock(&slab_mutex);
1537                break;
1538        case MEM_GOING_OFFLINE:
1539                mutex_lock(&slab_mutex);
1540                ret = drain_cache_nodelists_node(nid);
1541                mutex_unlock(&slab_mutex);
1542                break;
1543        case MEM_ONLINE:
1544        case MEM_OFFLINE:
1545        case MEM_CANCEL_ONLINE:
1546        case MEM_CANCEL_OFFLINE:
1547                break;
1548        }
1549out:
1550        return notifier_from_errno(ret);
1551}
1552#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1553
1554/*
1555 * swap the static kmem_list3 with kmalloced memory
1556 */
1557static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1558                                int nodeid)
1559{
1560        struct kmem_list3 *ptr;
1561
1562        ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1563        BUG_ON(!ptr);
1564
1565        memcpy(ptr, list, sizeof(struct kmem_list3));
1566        /*
1567         * Do not assume that spinlocks can be initialized via memcpy:
1568         */
1569        spin_lock_init(&ptr->list_lock);
1570
1571        MAKE_ALL_LISTS(cachep, ptr, nodeid);
1572        cachep->nodelists[nodeid] = ptr;
1573}
1574
1575/*
1576 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1577 * size of kmem_list3.
1578 */
1579static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1580{
1581        int node;
1582
1583        for_each_online_node(node) {
1584                cachep->nodelists[node] = &initkmem_list3[index + node];
1585                cachep->nodelists[node]->next_reap = jiffies +
1586                    REAPTIMEOUT_LIST3 +
1587                    ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1588        }
1589}
1590
1591/*
1592 * Initialisation.  Called after the page allocator have been initialised and
1593 * before smp_init().
1594 */
1595void __init kmem_cache_init(void)
1596{
1597        size_t left_over;
1598        struct cache_sizes *sizes;
1599        struct cache_names *names;
1600        int i;
1601        int order;
1602        int node;
1603
1604        if (num_possible_nodes() == 1)
1605                use_alien_caches = 0;
1606
1607        for (i = 0; i < NUM_INIT_LISTS; i++) {
1608                kmem_list3_init(&initkmem_list3[i]);
1609                if (i < MAX_NUMNODES)
1610                        cache_cache.nodelists[i] = NULL;
1611        }
1612        set_up_list3s(&cache_cache, CACHE_CACHE);
1613
1614        /*
1615         * Fragmentation resistance on low memory - only use bigger
1616         * page orders on machines with more than 32MB of memory if
1617         * not overridden on the command line.
1618         */
1619        if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1620                slab_max_order = SLAB_MAX_ORDER_HI;
1621
1622        /* Bootstrap is tricky, because several objects are allocated
1623         * from caches that do not exist yet:
1624         * 1) initialize the cache_cache cache: it contains the struct
1625         *    kmem_cache structures of all caches, except cache_cache itself:
1626         *    cache_cache is statically allocated.
1627         *    Initially an __init data area is used for the head array and the
1628         *    kmem_list3 structures, it's replaced with a kmalloc allocated
1629         *    array at the end of the bootstrap.
1630         * 2) Create the first kmalloc cache.
1631         *    The struct kmem_cache for the new cache is allocated normally.
1632         *    An __init data area is used for the head array.
1633         * 3) Create the remaining kmalloc caches, with minimally sized
1634         *    head arrays.
1635         * 4) Replace the __init data head arrays for cache_cache and the first
1636         *    kmalloc cache with kmalloc allocated arrays.
1637         * 5) Replace the __init data for kmem_list3 for cache_cache and
1638         *    the other cache's with kmalloc allocated memory.
1639         * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1640         */
1641
1642        node = numa_mem_id();
1643
1644        /* 1) create the cache_cache */
1645        INIT_LIST_HEAD(&slab_caches);
1646        list_add(&cache_cache.list, &slab_caches);
1647        cache_cache.colour_off = cache_line_size();
1648        cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1649        cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1650
1651        /*
1652         * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1653         */
1654        cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1655                                  nr_node_ids * sizeof(struct kmem_list3 *);
1656        cache_cache.object_size = cache_cache.size;
1657        cache_cache.size = ALIGN(cache_cache.size,
1658                                        cache_line_size());
1659        cache_cache.reciprocal_buffer_size =
1660                reciprocal_value(cache_cache.size);
1661
1662        for (order = 0; order < MAX_ORDER; order++) {
1663                cache_estimate(order, cache_cache.size,
1664                        cache_line_size(), 0, &left_over, &cache_cache.num);
1665                if (cache_cache.num)
1666                        break;
1667        }
1668        BUG_ON(!cache_cache.num);
1669        cache_cache.gfporder = order;
1670        cache_cache.colour = left_over / cache_cache.colour_off;
1671        cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1672                                      sizeof(struct slab), cache_line_size());
1673
1674        /* 2+3) create the kmalloc caches */
1675        sizes = malloc_sizes;
1676        names = cache_names;
1677
1678        /*
1679         * Initialize the caches that provide memory for the array cache and the
1680         * kmem_list3 structures first.  Without this, further allocations will
1681         * bug.
1682         */
1683
1684        sizes[INDEX_AC].cs_cachep = __kmem_cache_create(names[INDEX_AC].name,
1685                                        sizes[INDEX_AC].cs_size,
1686                                        ARCH_KMALLOC_MINALIGN,
1687                                        ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1688                                        NULL);
1689
1690        if (INDEX_AC != INDEX_L3) {
1691                sizes[INDEX_L3].cs_cachep =
1692                        __kmem_cache_create(names[INDEX_L3].name,
1693                                sizes[INDEX_L3].cs_size,
1694                                ARCH_KMALLOC_MINALIGN,
1695                                ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1696                                NULL);
1697        }
1698
1699        slab_early_init = 0;
1700
1701        while (sizes->cs_size != ULONG_MAX) {
1702                /*
1703                 * For performance, all the general caches are L1 aligned.
1704                 * This should be particularly beneficial on SMP boxes, as it
1705                 * eliminates "false sharing".
1706                 * Note for systems short on memory removing the alignment will
1707                 * allow tighter packing of the smaller caches.
1708                 */
1709                if (!sizes->cs_cachep) {
1710                        sizes->cs_cachep = __kmem_cache_create(names->name,
1711                                        sizes->cs_size,
1712                                        ARCH_KMALLOC_MINALIGN,
1713                                        ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1714                                        NULL);
1715                }
1716#ifdef CONFIG_ZONE_DMA
1717                sizes->cs_dmacachep = __kmem_cache_create(
1718                                        names->name_dma,
1719                                        sizes->cs_size,
1720                                        ARCH_KMALLOC_MINALIGN,
1721                                        ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1722                                                SLAB_PANIC,
1723                                        NULL);
1724#endif
1725                sizes++;
1726                names++;
1727        }
1728        /* 4) Replace the bootstrap head arrays */
1729        {
1730                struct array_cache *ptr;
1731
1732                ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1733
1734                BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1735                memcpy(ptr, cpu_cache_get(&cache_cache),
1736                       sizeof(struct arraycache_init));
1737                /*
1738                 * Do not assume that spinlocks can be initialized via memcpy:
1739                 */
1740                spin_lock_init(&ptr->lock);
1741
1742                cache_cache.array[smp_processor_id()] = ptr;
1743
1744                ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1745
1746                BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1747                       != &initarray_generic.cache);
1748                memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1749                       sizeof(struct arraycache_init));
1750                /*
1751                 * Do not assume that spinlocks can be initialized via memcpy:
1752                 */
1753                spin_lock_init(&ptr->lock);
1754
1755                malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1756                    ptr;
1757        }
1758        /* 5) Replace the bootstrap kmem_list3's */
1759        {
1760                int nid;
1761
1762                for_each_online_node(nid) {
1763                        init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1764
1765                        init_list(malloc_sizes[INDEX_AC].cs_cachep,
1766                                  &initkmem_list3[SIZE_AC + nid], nid);
1767
1768                        if (INDEX_AC != INDEX_L3) {
1769                                init_list(malloc_sizes[INDEX_L3].cs_cachep,
1770                                          &initkmem_list3[SIZE_L3 + nid], nid);
1771                        }
1772                }
1773        }
1774
1775        slab_state = UP;
1776}
1777
1778void __init kmem_cache_init_late(void)
1779{
1780        struct kmem_cache *cachep;
1781
1782        slab_state = UP;
1783
1784        /* 6) resize the head arrays to their final sizes */
1785        mutex_lock(&slab_mutex);
1786        list_for_each_entry(cachep, &slab_caches, list)
1787                if (enable_cpucache(cachep, GFP_NOWAIT))
1788                        BUG();
1789        mutex_unlock(&slab_mutex);
1790
1791        /* Annotate slab for lockdep -- annotate the malloc caches */
1792        init_lock_keys();
1793
1794        /* Done! */
1795        slab_state = FULL;
1796
1797        /*
1798         * Register a cpu startup notifier callback that initializes
1799         * cpu_cache_get for all new cpus
1800         */
1801        register_cpu_notifier(&cpucache_notifier);
1802
1803#ifdef CONFIG_NUMA
1804        /*
1805         * Register a memory hotplug callback that initializes and frees
1806         * nodelists.
1807         */
1808        hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1809#endif
1810
1811        /*
1812         * The reap timers are started later, with a module init call: That part
1813         * of the kernel is not yet operational.
1814         */
1815}
1816
1817static int __init cpucache_init(void)
1818{
1819        int cpu;
1820
1821        /*
1822         * Register the timers that return unneeded pages to the page allocator
1823         */
1824        for_each_online_cpu(cpu)
1825                start_cpu_timer(cpu);
1826
1827        /* Done! */
1828        slab_state = FULL;
1829        return 0;
1830}
1831__initcall(cpucache_init);
1832
1833static noinline void
1834slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1835{
1836        struct kmem_list3 *l3;
1837        struct slab *slabp;
1838        unsigned long flags;
1839        int node;
1840
1841        printk(KERN_WARNING
1842                "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1843                nodeid, gfpflags);
1844        printk(KERN_WARNING "  cache: %s, object size: %d, order: %d\n",
1845                cachep->name, cachep->size, cachep->gfporder);
1846
1847        for_each_online_node(node) {
1848                unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1849                unsigned long active_slabs = 0, num_slabs = 0;
1850
1851                l3 = cachep->nodelists[node];
1852                if (!l3)
1853                        continue;
1854
1855                spin_lock_irqsave(&l3->list_lock, flags);
1856                list_for_each_entry(slabp, &l3->slabs_full, list) {
1857                        active_objs += cachep->num;
1858                        active_slabs++;
1859                }
1860                list_for_each_entry(slabp, &l3->slabs_partial, list) {
1861                        active_objs += slabp->inuse;
1862                        active_slabs++;
1863                }
1864                list_for_each_entry(slabp, &l3->slabs_free, list)
1865                        num_slabs++;
1866
1867                free_objects += l3->free_objects;
1868                spin_unlock_irqrestore(&l3->list_lock, flags);
1869
1870                num_slabs += active_slabs;
1871                num_objs = num_slabs * cachep->num;
1872                printk(KERN_WARNING
1873                        "  node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1874                        node, active_slabs, num_slabs, active_objs, num_objs,
1875                        free_objects);
1876        }
1877}
1878
1879/*
1880 * Interface to system's page allocator. No need to hold the cache-lock.
1881 *
1882 * If we requested dmaable memory, we will get it. Even if we
1883 * did not request dmaable memory, we might get it, but that
1884 * would be relatively rare and ignorable.
1885 */
1886static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1887{
1888        struct page *page;
1889        int nr_pages;
1890        int i;
1891
1892#ifndef CONFIG_MMU
1893        /*
1894         * Nommu uses slab's for process anonymous memory allocations, and thus
1895         * requires __GFP_COMP to properly refcount higher order allocations
1896         */
1897        flags |= __GFP_COMP;
1898#endif
1899
1900        flags |= cachep->allocflags;
1901        if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1902                flags |= __GFP_RECLAIMABLE;
1903
1904        page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1905        if (!page) {
1906                if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1907                        slab_out_of_memory(cachep, flags, nodeid);
1908                return NULL;
1909        }
1910
1911        /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1912        if (unlikely(page->pfmemalloc))
1913                pfmemalloc_active = true;
1914
1915        nr_pages = (1 << cachep->gfporder);
1916        if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1917                add_zone_page_state(page_zone(page),
1918                        NR_SLAB_RECLAIMABLE, nr_pages);
1919        else
1920                add_zone_page_state(page_zone(page),
1921                        NR_SLAB_UNRECLAIMABLE, nr_pages);
1922        for (i = 0; i < nr_pages; i++) {
1923                __SetPageSlab(page + i);
1924
1925                if (page->pfmemalloc)
1926                        SetPageSlabPfmemalloc(page + i);
1927        }
1928
1929        if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1930                kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1931
1932                if (cachep->ctor)
1933                        kmemcheck_mark_uninitialized_pages(page, nr_pages);
1934                else
1935                        kmemcheck_mark_unallocated_pages(page, nr_pages);
1936        }
1937
1938        return page_address(page);
1939}
1940
1941/*
1942 * Interface to system's page release.
1943 */
1944static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1945{
1946        unsigned long i = (1 << cachep->gfporder);
1947        struct page *page = virt_to_page(addr);
1948        const unsigned long nr_freed = i;
1949
1950        kmemcheck_free_shadow(page, cachep->gfporder);
1951
1952        if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1953                sub_zone_page_state(page_zone(page),
1954                                NR_SLAB_RECLAIMABLE, nr_freed);
1955        else
1956                sub_zone_page_state(page_zone(page),
1957                                NR_SLAB_UNRECLAIMABLE, nr_freed);
1958        while (i--) {
1959                BUG_ON(!PageSlab(page));
1960                __ClearPageSlabPfmemalloc(page);
1961                __ClearPageSlab(page);
1962                page++;
1963        }
1964        if (current->reclaim_state)
1965                current->reclaim_state->reclaimed_slab += nr_freed;
1966        free_pages((unsigned long)addr, cachep->gfporder);
1967}
1968
1969static void kmem_rcu_free(struct rcu_head *head)
1970{
1971        struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1972        struct kmem_cache *cachep = slab_rcu->cachep;
1973
1974        kmem_freepages(cachep, slab_rcu->addr);
1975        if (OFF_SLAB(cachep))
1976                kmem_cache_free(cachep->slabp_cache, slab_rcu);
1977}
1978
1979#if DEBUG
1980
1981#ifdef CONFIG_DEBUG_PAGEALLOC
1982static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1983                            unsigned long caller)
1984{
1985        int size = cachep->object_size;
1986
1987        addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1988
1989        if (size < 5 * sizeof(unsigned long))
1990                return;
1991
1992        *addr++ = 0x12345678;
1993        *addr++ = caller;
1994        *addr++ = smp_processor_id();
1995        size -= 3 * sizeof(unsigned long);
1996        {
1997                unsigned long *sptr = &caller;
1998                unsigned long svalue;
1999
2000                while (!kstack_end(sptr)) {
2001                        svalue = *sptr++;
2002                        if (kernel_text_address(svalue)) {
2003                                *addr++ = svalue;
2004                                size -= sizeof(unsigned long);
2005                                if (size <= sizeof(unsigned long))
2006                                        break;
2007                        }
2008                }
2009
2010        }
2011        *addr++ = 0x87654321;
2012}
2013#endif
2014
2015static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
2016{
2017        int size = cachep->object_size;
2018        addr = &((char *)addr)[obj_offset(cachep)];
2019
2020        memset(addr, val, size);
2021        *(unsigned char *)(addr + size - 1) = POISON_END;
2022}
2023
2024static void dump_line(char *data, int offset, int limit)
2025{
2026        int i;
2027        unsigned char error = 0;
2028        int bad_count = 0;
2029
2030        printk(KERN_ERR "%03x: ", offset);
2031        for (i = 0; i < limit; i++) {
2032                if (data[offset + i] != POISON_FREE) {
2033                        error = data[offset + i];
2034                        bad_count++;
2035                }
2036        }
2037        print_hex_dump(KERN_CONT, "", 0, 16, 1,
2038                        &data[offset], limit, 1);
2039
2040        if (bad_count == 1) {
2041                error ^= POISON_FREE;
2042                if (!(error & (error - 1))) {
2043                        printk(KERN_ERR "Single bit error detected. Probably "
2044                                        "bad RAM.\n");
2045#ifdef CONFIG_X86
2046                        printk(KERN_ERR "Run memtest86+ or a similar memory "
2047                                        "test tool.\n");
2048#else
2049                        printk(KERN_ERR "Run a memory test tool.\n");
2050#endif
2051                }
2052        }
2053}
2054#endif
2055
2056#if DEBUG
2057
2058static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
2059{
2060        int i, size;
2061        char *realobj;
2062
2063        if (cachep->flags & SLAB_RED_ZONE) {
2064                printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
2065                        *dbg_redzone1(cachep, objp),
2066                        *dbg_redzone2(cachep, objp));
2067        }
2068
2069        if (cachep->flags & SLAB_STORE_USER) {
2070                printk(KERN_ERR "Last user: [<%p>]",
2071                        *dbg_userword(cachep, objp));
2072                print_symbol("(%s)",
2073                                (unsigned long)*dbg_userword(cachep, objp));
2074                printk("\n");
2075        }
2076        realobj = (char *)objp + obj_offset(cachep);
2077        size = cachep->object_size;
2078        for (i = 0; i < size && lines; i += 16, lines--) {
2079                int limit;
2080                limit = 16;
2081                if (i + limit > size)
2082                        limit = size - i;
2083                dump_line(realobj, i, limit);
2084        }
2085}
2086
2087static void check_poison_obj(struct kmem_cache *cachep, void *objp)
2088{
2089        char *realobj;
2090        int size, i;
2091        int lines = 0;
2092
2093        realobj = (char *)objp + obj_offset(cachep);
2094        size = cachep->object_size;
2095
2096        for (i = 0; i < size; i++) {
2097                char exp = POISON_FREE;
2098                if (i == size - 1)
2099                        exp = POISON_END;
2100                if (realobj[i] != exp) {
2101                        int limit;
2102                        /* Mismatch ! */
2103                        /* Print header */
2104                        if (lines == 0) {
2105                                printk(KERN_ERR
2106                                        "Slab corruption (%s): %s start=%p, len=%d\n",
2107                                        print_tainted(), cachep->name, realobj, size);
2108                                print_objinfo(cachep, objp, 0);
2109                        }
2110                        /* Hexdump the affected line */
2111                        i = (i / 16) * 16;
2112                        limit = 16;
2113                        if (i + limit > size)
2114                                limit = size - i;
2115                        dump_line(realobj, i, limit);
2116                        i += 16;
2117                        lines++;
2118                        /* Limit to 5 lines */
2119                        if (lines > 5)
2120                                break;
2121                }
2122        }
2123        if (lines != 0) {
2124                /* Print some data about the neighboring objects, if they
2125                 * exist:
2126                 */
2127                struct slab *slabp = virt_to_slab(objp);
2128                unsigned int objnr;
2129
2130                objnr = obj_to_index(cachep, slabp, objp);
2131                if (objnr) {
2132                        objp = index_to_obj(cachep, slabp, objnr - 1);
2133                        realobj = (char *)objp + obj_offset(cachep);
2134                        printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2135                               realobj, size);
2136                        print_objinfo(cachep, objp, 2);
2137                }
2138                if (objnr + 1 < cachep->num) {
2139                        objp = index_to_obj(cachep, slabp, objnr + 1);
2140                        realobj = (char *)objp + obj_offset(cachep);
2141                        printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2142                               realobj, size);
2143                        print_objinfo(cachep, objp, 2);
2144                }
2145        }
2146}
2147#endif
2148
2149#if DEBUG
2150static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2151{
2152        int i;
2153        for (i = 0; i < cachep->num; i++) {
2154                void *objp = index_to_obj(cachep, slabp, i);
2155
2156                if (cachep->flags & SLAB_POISON) {
2157#ifdef CONFIG_DEBUG_PAGEALLOC
2158                        if (cachep->size % PAGE_SIZE == 0 &&
2159                                        OFF_SLAB(cachep))
2160                                kernel_map_pages(virt_to_page(objp),
2161                                        cachep->size / PAGE_SIZE, 1);
2162                        else
2163                                check_poison_obj(cachep, objp);
2164#else
2165                        check_poison_obj(cachep, objp);
2166#endif
2167                }
2168                if (cachep->flags & SLAB_RED_ZONE) {
2169                        if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2170                                slab_error(cachep, "start of a freed object "
2171                                           "was overwritten");
2172                        if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2173                                slab_error(cachep, "end of a freed object "
2174                                           "was overwritten");
2175                }
2176        }
2177}
2178#else
2179static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2180{
2181}
2182#endif
2183
2184/**
2185 * slab_destroy - destroy and release all objects in a slab
2186 * @cachep: cache pointer being destroyed
2187 * @slabp: slab pointer being destroyed
2188 *
2189 * Destroy all the objs in a slab, and release the mem back to the system.
2190 * Before calling the slab must have been unlinked from the cache.  The
2191 * cache-lock is not held/needed.
2192 */
2193static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2194{
2195        void *addr = slabp->s_mem - slabp->colouroff;
2196
2197        slab_destroy_debugcheck(cachep, slabp);
2198        if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2199                struct slab_rcu *slab_rcu;
2200
2201                slab_rcu = (struct slab_rcu *)slabp;
2202                slab_rcu->cachep = cachep;
2203                slab_rcu->addr = addr;
2204                call_rcu(&slab_rcu->head, kmem_rcu_free);
2205        } else {
2206                kmem_freepages(cachep, addr);
2207                if (OFF_SLAB(cachep))
2208                        kmem_cache_free(cachep->slabp_cache, slabp);
2209        }
2210}
2211
2212static void __kmem_cache_destroy(struct kmem_cache *cachep)
2213{
2214        int i;
2215        struct kmem_list3 *l3;
2216
2217        for_each_online_cpu(i)
2218            kfree(cachep->array[i]);
2219
2220        /* NUMA: free the list3 structures */
2221        for_each_online_node(i) {
2222                l3 = cachep->nodelists[i];
2223                if (l3) {
2224                        kfree(l3->shared);
2225                        free_alien_cache(l3->alien);
2226                        kfree(l3);
2227                }
2228        }
2229        kmem_cache_free(&cache_cache, cachep);
2230}
2231
2232
2233/**
2234 * calculate_slab_order - calculate size (page order) of slabs
2235 * @cachep: pointer to the cache that is being created
2236 * @size: size of objects to be created in this cache.
2237 * @align: required alignment for the objects.
2238 * @flags: slab allocation flags
2239 *
2240 * Also calculates the number of objects per slab.
2241 *
2242 * This could be made much more intelligent.  For now, try to avoid using
2243 * high order pages for slabs.  When the gfp() functions are more friendly
2244 * towards high-order requests, this should be changed.
2245 */
2246static size_t calculate_slab_order(struct kmem_cache *cachep,
2247                        size_t size, size_t align, unsigned long flags)
2248{
2249        unsigned long offslab_limit;
2250        size_t left_over = 0;
2251        int gfporder;
2252
2253        for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2254                unsigned int num;
2255                size_t remainder;
2256
2257                cache_estimate(gfporder, size, align, flags, &remainder, &num);
2258                if (!num)
2259                        continue;
2260
2261                if (flags & CFLGS_OFF_SLAB) {
2262                        /*
2263                         * Max number of objs-per-slab for caches which
2264                         * use off-slab slabs. Needed to avoid a possible
2265                         * looping condition in cache_grow().
2266                         */
2267                        offslab_limit = size - sizeof(struct slab);
2268                        offslab_limit /= sizeof(kmem_bufctl_t);
2269
2270                        if (num > offslab_limit)
2271                                break;
2272                }
2273
2274                /* Found something acceptable - save it away */
2275                cachep->num = num;
2276                cachep->gfporder = gfporder;
2277                left_over = remainder;
2278
2279                /*
2280                 * A VFS-reclaimable slab tends to have most allocations
2281                 * as GFP_NOFS and we really don't want to have to be allocating
2282                 * higher-order pages when we are unable to shrink dcache.
2283                 */
2284                if (flags & SLAB_RECLAIM_ACCOUNT)
2285                        break;
2286
2287                /*
2288                 * Large number of objects is good, but very large slabs are
2289                 * currently bad for the gfp()s.
2290                 */
2291                if (gfporder >= slab_max_order)
2292                        break;
2293
2294                /*
2295                 * Acceptable internal fragmentation?
2296                 */
2297                if (left_over * 8 <= (PAGE_SIZE << gfporder))
2298                        break;
2299        }
2300        return left_over;
2301}
2302
2303static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2304{
2305        if (slab_state >= FULL)
2306                return enable_cpucache(cachep, gfp);
2307
2308        if (slab_state == DOWN) {
2309                /*
2310                 * Note: the first kmem_cache_create must create the cache
2311                 * that's used by kmalloc(24), otherwise the creation of
2312                 * further caches will BUG().
2313                 */
2314                cachep->array[smp_processor_id()] = &initarray_generic.cache;
2315
2316                /*
2317                 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2318                 * the first cache, then we need to set up all its list3s,
2319                 * otherwise the creation of further caches will BUG().
2320                 */
2321                set_up_list3s(cachep, SIZE_AC);
2322                if (INDEX_AC == INDEX_L3)
2323                        slab_state = PARTIAL_L3;
2324                else
2325                        slab_state = PARTIAL_ARRAYCACHE;
2326        } else {
2327                cachep->array[smp_processor_id()] =
2328                        kmalloc(sizeof(struct arraycache_init), gfp);
2329
2330                if (slab_state == PARTIAL_ARRAYCACHE) {
2331                        set_up_list3s(cachep, SIZE_L3);
2332                        slab_state = PARTIAL_L3;
2333                } else {
2334                        int node;
2335                        for_each_online_node(node) {
2336                                cachep->nodelists[node] =
2337                                    kmalloc_node(sizeof(struct kmem_list3),
2338                                                gfp, node);
2339                                BUG_ON(!cachep->nodelists[node]);
2340                                kmem_list3_init(cachep->nodelists[node]);
2341                        }
2342                }
2343        }
2344        cachep->nodelists[numa_mem_id()]->next_reap =
2345                        jiffies + REAPTIMEOUT_LIST3 +
2346                        ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2347
2348        cpu_cache_get(cachep)->avail = 0;
2349        cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2350        cpu_cache_get(cachep)->batchcount = 1;
2351        cpu_cache_get(cachep)->touched = 0;
2352        cachep->batchcount = 1;
2353        cachep->limit = BOOT_CPUCACHE_ENTRIES;
2354        return 0;
2355}
2356
2357/**
2358 * __kmem_cache_create - Create a cache.
2359 * @name: A string which is used in /proc/slabinfo to identify this cache.
2360 * @size: The size of objects to be created in this cache.
2361 * @align: The required alignment for the objects.
2362 * @flags: SLAB flags
2363 * @ctor: A constructor for the objects.
2364 *
2365 * Returns a ptr to the cache on success, NULL on failure.
2366 * Cannot be called within a int, but can be interrupted.
2367 * The @ctor is run when new pages are allocated by the cache.
2368 *
2369 * @name must be valid until the cache is destroyed. This implies that
2370 * the module calling this has to destroy the cache before getting unloaded.
2371 *
2372 * The flags are
2373 *
2374 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2375 * to catch references to uninitialised memory.
2376 *
2377 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2378 * for buffer overruns.
2379 *
2380 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2381 * cacheline.  This can be beneficial if you're counting cycles as closely
2382 * as davem.
2383 */
2384struct kmem_cache *
2385__kmem_cache_create (const char *name, size_t size, size_t align,
2386        unsigned long flags, void (*ctor)(void *))
2387{
2388        size_t left_over, slab_size, ralign;
2389        struct kmem_cache *cachep = NULL;
2390        gfp_t gfp;
2391
2392#if DEBUG
2393#if FORCED_DEBUG
2394        /*
2395         * Enable redzoning and last user accounting, except for caches with
2396         * large objects, if the increased size would increase the object size
2397         * above the next power of two: caches with object sizes just above a
2398         * power of two have a significant amount of internal fragmentation.
2399         */
2400        if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2401                                                2 * sizeof(unsigned long long)))
2402                flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2403        if (!(flags & SLAB_DESTROY_BY_RCU))
2404                flags |= SLAB_POISON;
2405#endif
2406        if (flags & SLAB_DESTROY_BY_RCU)
2407                BUG_ON(flags & SLAB_POISON);
2408#endif
2409        /*
2410         * Always checks flags, a caller might be expecting debug support which
2411         * isn't available.
2412         */
2413        BUG_ON(flags & ~CREATE_MASK);
2414
2415        /*
2416         * Check that size is in terms of words.  This is needed to avoid
2417         * unaligned accesses for some archs when redzoning is used, and makes
2418         * sure any on-slab bufctl's are also correctly aligned.
2419         */
2420        if (size & (BYTES_PER_WORD - 1)) {
2421                size += (BYTES_PER_WORD - 1);
2422                size &= ~(BYTES_PER_WORD - 1);
2423        }
2424
2425        /* calculate the final buffer alignment: */
2426
2427        /* 1) arch recommendation: can be overridden for debug */
2428        if (flags & SLAB_HWCACHE_ALIGN) {
2429                /*
2430                 * Default alignment: as specified by the arch code.  Except if
2431                 * an object is really small, then squeeze multiple objects into
2432                 * one cacheline.
2433                 */
2434                ralign = cache_line_size();
2435                while (size <= ralign / 2)
2436                        ralign /= 2;
2437        } else {
2438                ralign = BYTES_PER_WORD;
2439        }
2440
2441        /*
2442         * Redzoning and user store require word alignment or possibly larger.
2443         * Note this will be overridden by architecture or caller mandated
2444         * alignment if either is greater than BYTES_PER_WORD.
2445         */
2446        if (flags & SLAB_STORE_USER)
2447                ralign = BYTES_PER_WORD;
2448
2449        if (flags & SLAB_RED_ZONE) {
2450                ralign = REDZONE_ALIGN;
2451                /* If redzoning, ensure that the second redzone is suitably
2452                 * aligned, by adjusting the object size accordingly. */
2453                size += REDZONE_ALIGN - 1;
2454                size &= ~(REDZONE_ALIGN - 1);
2455        }
2456
2457        /* 2) arch mandated alignment */
2458        if (ralign < ARCH_SLAB_MINALIGN) {
2459                ralign = ARCH_SLAB_MINALIGN;
2460        }
2461        /* 3) caller mandated alignment */
2462        if (ralign < align) {
2463                ralign = align;
2464        }
2465        /* disable debug if necessary */
2466        if (ralign > __alignof__(unsigned long long))
2467                flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2468        /*
2469         * 4) Store it.
2470         */
2471        align = ralign;
2472
2473        if (slab_is_available())
2474                gfp = GFP_KERNEL;
2475        else
2476                gfp = GFP_NOWAIT;
2477
2478        /* Get cache's description obj. */
2479        cachep = kmem_cache_zalloc(&cache_cache, gfp);
2480        if (!cachep)
2481                return NULL;
2482
2483        cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2484        cachep->object_size = size;
2485        cachep->align = align;
2486#if DEBUG
2487
2488        /*
2489         * Both debugging options require word-alignment which is calculated
2490         * into align above.
2491         */
2492        if (flags & SLAB_RED_ZONE) {
2493                /* add space for red zone words */
2494                cachep->obj_offset += sizeof(unsigned long long);
2495                size += 2 * sizeof(unsigned long long);
2496        }
2497        if (flags & SLAB_STORE_USER) {
2498                /* user store requires one word storage behind the end of
2499                 * the real object. But if the second red zone needs to be
2500                 * aligned to 64 bits, we must allow that much space.
2501                 */
2502                if (flags & SLAB_RED_ZONE)
2503                        size += REDZONE_ALIGN;
2504                else
2505                        size += BYTES_PER_WORD;
2506        }
2507#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2508        if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2509            && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2510                cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2511                size = PAGE_SIZE;
2512        }
2513#endif
2514#endif
2515
2516        /*
2517         * Determine if the slab management is 'on' or 'off' slab.
2518         * (bootstrapping cannot cope with offslab caches so don't do
2519         * it too early on. Always use on-slab management when
2520         * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2521         */
2522        if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2523            !(flags & SLAB_NOLEAKTRACE))
2524                /*
2525                 * Size is large, assume best to place the slab management obj
2526                 * off-slab (should allow better packing of objs).
2527                 */
2528                flags |= CFLGS_OFF_SLAB;
2529
2530        size = ALIGN(size, align);
2531
2532        left_over = calculate_slab_order(cachep, size, align, flags);
2533
2534        if (!cachep->num) {
2535                printk(KERN_ERR
2536                       "kmem_cache_create: couldn't create cache %s.\n", name);
2537                kmem_cache_free(&cache_cache, cachep);
2538                return NULL;
2539        }
2540        slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2541                          + sizeof(struct slab), align);
2542
2543        /*
2544         * If the slab has been placed off-slab, and we have enough space then
2545         * move it on-slab. This is at the expense of any extra colouring.
2546         */
2547        if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2548                flags &= ~CFLGS_OFF_SLAB;
2549                left_over -= slab_size;
2550        }
2551
2552        if (flags & CFLGS_OFF_SLAB) {
2553                /* really off slab. No need for manual alignment */
2554                slab_size =
2555                    cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2556
2557#ifdef CONFIG_PAGE_POISONING
2558                /* If we're going to use the generic kernel_map_pages()
2559                 * poisoning, then it's going to smash the contents of
2560                 * the redzone and userword anyhow, so switch them off.
2561                 */
2562                if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2563                        flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2564#endif
2565        }
2566
2567        cachep->colour_off = cache_line_size();
2568        /* Offset must be a multiple of the alignment. */
2569        if (cachep->colour_off < align)
2570                cachep->colour_off = align;
2571        cachep->colour = left_over / cachep->colour_off;
2572        cachep->slab_size = slab_size;
2573        cachep->flags = flags;
2574        cachep->allocflags = 0;
2575        if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2576                cachep->allocflags |= GFP_DMA;
2577        cachep->size = size;
2578        cachep->reciprocal_buffer_size = reciprocal_value(size);
2579
2580        if (flags & CFLGS_OFF_SLAB) {
2581                cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2582                /*
2583                 * This is a possibility for one of the malloc_sizes caches.
2584                 * But since we go off slab only for object size greater than
2585                 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2586                 * this should not happen at all.
2587                 * But leave a BUG_ON for some lucky dude.
2588                 */
2589                BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2590        }
2591        cachep->ctor = ctor;
2592        cachep->name = name;
2593
2594        if (setup_cpu_cache(cachep, gfp)) {
2595                __kmem_cache_destroy(cachep);
2596                return NULL;
2597        }
2598
2599        if (flags & SLAB_DEBUG_OBJECTS) {
2600                /*
2601                 * Would deadlock through slab_destroy()->call_rcu()->
2602                 * debug_object_activate()->kmem_cache_alloc().
2603                 */
2604                WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2605
2606                slab_set_debugobj_lock_classes(cachep);
2607        }
2608
2609        /* cache setup completed, link it into the list */
2610        list_add(&cachep->list, &slab_caches);
2611        return cachep;
2612}
2613
2614#if DEBUG
2615static void check_irq_off(void)
2616{
2617        BUG_ON(!irqs_disabled());
2618}
2619
2620static void check_irq_on(void)
2621{
2622        BUG_ON(irqs_disabled());
2623}
2624
2625static void check_spinlock_acquired(struct kmem_cache *cachep)
2626{
2627#ifdef CONFIG_SMP
2628        check_irq_off();
2629        assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2630#endif
2631}
2632
2633static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2634{
2635#ifdef CONFIG_SMP
2636        check_irq_off();
2637        assert_spin_locked(&cachep->nodelists[node]->list_lock);
2638#endif
2639}
2640
2641#else
2642#define check_irq_off() do { } while(0)
2643#define check_irq_on()  do { } while(0)
2644#define check_spinlock_acquired(x) do { } while(0)
2645#define check_spinlock_acquired_node(x, y) do { } while(0)
2646#endif
2647
2648static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2649                        struct array_cache *ac,
2650                        int force, int node);
2651
2652static void do_drain(void *arg)
2653{
2654        struct kmem_cache *cachep = arg;
2655        struct array_cache *ac;
2656        int node = numa_mem_id();
2657
2658        check_irq_off();
2659        ac = cpu_cache_get(cachep);
2660        spin_lock(&cachep->nodelists[node]->list_lock);
2661        free_block(cachep, ac->entry, ac->avail, node);
2662        spin_unlock(&cachep->nodelists[node]->list_lock);
2663        ac->avail = 0;
2664}
2665
2666static void drain_cpu_caches(struct kmem_cache *cachep)
2667{
2668        struct kmem_list3 *l3;
2669        int node;
2670
2671        on_each_cpu(do_drain, cachep, 1);
2672        check_irq_on();
2673        for_each_online_node(node) {
2674                l3 = cachep->nodelists[node];
2675                if (l3 && l3->alien)
2676                        drain_alien_cache(cachep, l3->alien);
2677        }
2678
2679        for_each_online_node(node) {
2680                l3 = cachep->nodelists[node];
2681                if (l3)
2682                        drain_array(cachep, l3, l3->shared, 1, node);
2683        }
2684}
2685
2686/*
2687 * Remove slabs from the list of free slabs.
2688 * Specify the number of slabs to drain in tofree.
2689 *
2690 * Returns the actual number of slabs released.
2691 */
2692static int drain_freelist(struct kmem_cache *cache,
2693                        struct kmem_list3 *l3, int tofree)
2694{
2695        struct list_head *p;
2696        int nr_freed;
2697        struct slab *slabp;
2698
2699        nr_freed = 0;
2700        while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2701
2702                spin_lock_irq(&l3->list_lock);
2703                p = l3->slabs_free.prev;
2704                if (p == &l3->slabs_free) {
2705                        spin_unlock_irq(&l3->list_lock);
2706                        goto out;
2707                }
2708
2709                slabp = list_entry(p, struct slab, list);
2710#if DEBUG
2711                BUG_ON(slabp->inuse);
2712#endif
2713                list_del(&slabp->list);
2714                /*
2715                 * Safe to drop the lock. The slab is no longer linked
2716                 * to the cache.
2717                 */
2718                l3->free_objects -= cache->num;
2719                spin_unlock_irq(&l3->list_lock);
2720                slab_destroy(cache, slabp);
2721                nr_freed++;
2722        }
2723out:
2724        return nr_freed;
2725}
2726
2727/* Called with slab_mutex held to protect against cpu hotplug */
2728static int __cache_shrink(struct kmem_cache *cachep)
2729{
2730        int ret = 0, i = 0;
2731        struct kmem_list3 *l3;
2732
2733        drain_cpu_caches(cachep);
2734
2735        check_irq_on();
2736        for_each_online_node(i) {
2737                l3 = cachep->nodelists[i];
2738                if (!l3)
2739                        continue;
2740
2741                drain_freelist(cachep, l3, l3->free_objects);
2742
2743                ret += !list_empty(&l3->slabs_full) ||
2744                        !list_empty(&l3->slabs_partial);
2745        }
2746        return (ret ? 1 : 0);
2747}
2748
2749/**
2750 * kmem_cache_shrink - Shrink a cache.
2751 * @cachep: The cache to shrink.
2752 *
2753 * Releases as many slabs as possible for a cache.
2754 * To help debugging, a zero exit status indicates all slabs were released.
2755 */
2756int kmem_cache_shrink(struct kmem_cache *cachep)
2757{
2758        int ret;
2759        BUG_ON(!cachep || in_interrupt());
2760
2761        get_online_cpus();
2762        mutex_lock(&slab_mutex);
2763        ret = __cache_shrink(cachep);
2764        mutex_unlock(&slab_mutex);
2765        put_online_cpus();
2766        return ret;
2767}
2768EXPORT_SYMBOL(kmem_cache_shrink);
2769
2770/**
2771 * kmem_cache_destroy - delete a cache
2772 * @cachep: the cache to destroy
2773 *
2774 * Remove a &struct kmem_cache object from the slab cache.
2775 *
2776 * It is expected this function will be called by a module when it is
2777 * unloaded.  This will remove the cache completely, and avoid a duplicate
2778 * cache being allocated each time a module is loaded and unloaded, if the
2779 * module doesn't have persistent in-kernel storage across loads and unloads.
2780 *
2781 * The cache must be empty before calling this function.
2782 *
2783 * The caller must guarantee that no one will allocate memory from the cache
2784 * during the kmem_cache_destroy().
2785 */
2786void kmem_cache_destroy(struct kmem_cache *cachep)
2787{
2788        BUG_ON(!cachep || in_interrupt());
2789
2790        /* Find the cache in the chain of caches. */
2791        get_online_cpus();
2792        mutex_lock(&slab_mutex);
2793        /*
2794         * the chain is never empty, cache_cache is never destroyed
2795         */
2796        list_del(&cachep->list);
2797        if (__cache_shrink(cachep)) {
2798                slab_error(cachep, "Can't free all objects");
2799                list_add(&cachep->list, &slab_caches);
2800                mutex_unlock(&slab_mutex);
2801                put_online_cpus();
2802                return;
2803        }
2804
2805        if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2806                rcu_barrier();
2807
2808        __kmem_cache_destroy(cachep);
2809        mutex_unlock(&slab_mutex);
2810        put_online_cpus();
2811}
2812EXPORT_SYMBOL(kmem_cache_destroy);
2813
2814/*
2815 * Get the memory for a slab management obj.
2816 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2817 * always come from malloc_sizes caches.  The slab descriptor cannot
2818 * come from the same cache which is getting created because,
2819 * when we are searching for an appropriate cache for these
2820 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2821 * If we are creating a malloc_sizes cache here it would not be visible to
2822 * kmem_find_general_cachep till the initialization is complete.
2823 * Hence we cannot have slabp_cache same as the original cache.
2824 */
2825static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2826                                   int colour_off, gfp_t local_flags,
2827                                   int nodeid)
2828{
2829        struct slab *slabp;
2830
2831        if (OFF_SLAB(cachep)) {
2832                /* Slab management obj is off-slab. */
2833                slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2834                                              local_flags, nodeid);
2835                /*
2836                 * If the first object in the slab is leaked (it's allocated
2837                 * but no one has a reference to it), we want to make sure
2838                 * kmemleak does not treat the ->s_mem pointer as a reference
2839                 * to the object. Otherwise we will not report the leak.
2840                 */
2841                kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2842                                   local_flags);
2843                if (!slabp)
2844                        return NULL;
2845        } else {
2846                slabp = objp + colour_off;
2847                colour_off += cachep->slab_size;
2848        }
2849        slabp->inuse = 0;
2850        slabp->colouroff = colour_off;
2851        slabp->s_mem = objp + colour_off;
2852        slabp->nodeid = nodeid;
2853        slabp->free = 0;
2854        return slabp;
2855}
2856
2857static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2858{
2859        return (kmem_bufctl_t *) (slabp + 1);
2860}
2861
2862static void cache_init_objs(struct kmem_cache *cachep,
2863                            struct slab *slabp)
2864{
2865        int i;
2866
2867        for (i = 0; i < cachep->num; i++) {
2868                void *objp = index_to_obj(cachep, slabp, i);
2869#if DEBUG
2870                /* need to poison the objs? */
2871                if (cachep->flags & SLAB_POISON)
2872                        poison_obj(cachep, objp, POISON_FREE);
2873                if (cachep->flags & SLAB_STORE_USER)
2874                        *dbg_userword(cachep, objp) = NULL;
2875
2876                if (cachep->flags & SLAB_RED_ZONE) {
2877                        *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2878                        *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2879                }
2880                /*
2881                 * Constructors are not allowed to allocate memory from the same
2882                 * cache which they are a constructor for.  Otherwise, deadlock.
2883                 * They must also be threaded.
2884                 */
2885                if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2886                        cachep->ctor(objp + obj_offset(cachep));
2887
2888                if (cachep->flags & SLAB_RED_ZONE) {
2889                        if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2890                                slab_error(cachep, "constructor overwrote the"
2891                                           " end of an object");
2892                        if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2893                                slab_error(cachep, "constructor overwrote the"
2894                                           " start of an object");
2895                }
2896                if ((cachep->size % PAGE_SIZE) == 0 &&
2897                            OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2898                        kernel_map_pages(virt_to_page(objp),
2899                                         cachep->size / PAGE_SIZE, 0);
2900#else
2901                if (cachep->ctor)
2902                        cachep->ctor(objp);
2903#endif
2904                slab_bufctl(slabp)[i] = i + 1;
2905        }
2906        slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2907}
2908
2909static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2910{
2911        if (CONFIG_ZONE_DMA_FLAG) {
2912                if (flags & GFP_DMA)
2913                        BUG_ON(!(cachep->allocflags & GFP_DMA));
2914                else
2915                        BUG_ON(cachep->allocflags & GFP_DMA);
2916        }
2917}
2918
2919static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2920                                int nodeid)
2921{
2922        void *objp = index_to_obj(cachep, slabp, slabp->free);
2923        kmem_bufctl_t next;
2924
2925        slabp->inuse++;
2926        next = slab_bufctl(slabp)[slabp->free];
2927#if DEBUG
2928        slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2929        WARN_ON(slabp->nodeid != nodeid);
2930#endif
2931        slabp->free = next;
2932
2933        return objp;
2934}
2935
2936static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2937                                void *objp, int nodeid)
2938{
2939        unsigned int objnr = obj_to_index(cachep, slabp, objp);
2940
2941#if DEBUG
2942        /* Verify that the slab belongs to the intended node */
2943        WARN_ON(slabp->nodeid != nodeid);
2944
2945        if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2946                printk(KERN_ERR "slab: double free detected in cache "
2947                                "'%s', objp %p\n", cachep->name, objp);
2948                BUG();
2949        }
2950#endif
2951        slab_bufctl(slabp)[objnr] = slabp->free;
2952        slabp->free = objnr;
2953        slabp->inuse--;
2954}
2955
2956/*
2957 * Map pages beginning at addr to the given cache and slab. This is required
2958 * for the slab allocator to be able to lookup the cache and slab of a
2959 * virtual address for kfree, ksize, and slab debugging.
2960 */
2961static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2962                           void *addr)
2963{
2964        int nr_pages;
2965        struct page *page;
2966
2967        page = virt_to_page(addr);
2968
2969        nr_pages = 1;
2970        if (likely(!PageCompound(page)))
2971                nr_pages <<= cache->gfporder;
2972
2973        do {
2974                page->slab_cache = cache;
2975                page->slab_page = slab;
2976                page++;
2977        } while (--nr_pages);
2978}
2979
2980/*
2981 * Grow (by 1) the number of slabs within a cache.  This is called by
2982 * kmem_cache_alloc() when there are no active objs left in a cache.
2983 */
2984static int cache_grow(struct kmem_cache *cachep,
2985                gfp_t flags, int nodeid, void *objp)
2986{
2987        struct slab *slabp;
2988        size_t offset;
2989        gfp_t local_flags;
2990        struct kmem_list3 *l3;
2991
2992        /*
2993         * Be lazy and only check for valid flags here,  keeping it out of the
2994         * critical path in kmem_cache_alloc().
2995         */
2996        BUG_ON(flags & GFP_SLAB_BUG_MASK);
2997        local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2998
2999        /* Take the l3 list lock to change the colour_next on this node */
3000        check_irq_off();
3001        l3 = cachep->nodelists[nodeid];
3002        spin_lock(&l3->list_lock);
3003
3004        /* Get colour for the slab, and cal the next value. */
3005        offset = l3->colour_next;
3006        l3->colour_next++;
3007        if (l3->colour_next >= cachep->colour)
3008                l3->colour_next = 0;
3009        spin_unlock(&l3->list_lock);
3010
3011        offset *= cachep->colour_off;
3012
3013        if (local_flags & __GFP_WAIT)
3014                local_irq_enable();
3015
3016        /*
3017         * The test for missing atomic flag is performed here, rather than
3018         * the more obvious place, simply to reduce the critical path length
3019         * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
3020         * will eventually be caught here (where it matters).
3021         */
3022        kmem_flagcheck(cachep, flags);
3023
3024        /*
3025         * Get mem for the objs.  Attempt to allocate a physical page from
3026         * 'nodeid'.
3027         */
3028        if (!objp)
3029                objp = kmem_getpages(cachep, local_flags, nodeid);
3030        if (!objp)
3031                goto failed;
3032
3033        /* Get slab management. */
3034        slabp = alloc_slabmgmt(cachep, objp, offset,
3035                        local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
3036        if (!slabp)
3037                goto opps1;
3038
3039        slab_map_pages(cachep, slabp, objp);
3040
3041        cache_init_objs(cachep, slabp);
3042
3043        if (local_flags & __GFP_WAIT)
3044                local_irq_disable();
3045        check_irq_off();
3046        spin_lock(&l3->list_lock);
3047
3048        /* Make slab active. */
3049        list_add_tail(&slabp->list, &(l3->slabs_free));
3050        STATS_INC_GROWN(cachep);
3051        l3->free_objects += cachep->num;
3052        spin_unlock(&l3->list_lock);
3053        return 1;
3054opps1:
3055        kmem_freepages(cachep, objp);
3056failed:
3057        if (local_flags & __GFP_WAIT)
3058                local_irq_disable();
3059        return 0;
3060}
3061
3062#if DEBUG
3063
3064/*
3065 * Perform extra freeing checks:
3066 * - detect bad pointers.
3067 * - POISON/RED_ZONE checking
3068 */
3069static void kfree_debugcheck(const void *objp)
3070{
3071        if (!virt_addr_valid(objp)) {
3072                printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
3073                       (unsigned long)objp);
3074                BUG();
3075        }
3076}
3077
3078static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
3079{
3080        unsigned long long redzone1, redzone2;
3081
3082        redzone1 = *dbg_redzone1(cache, obj);
3083        redzone2 = *dbg_redzone2(cache, obj);
3084
3085        /*
3086         * Redzone is ok.
3087         */
3088        if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3089                return;
3090
3091        if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3092                slab_error(cache, "double free detected");
3093        else
3094                slab_error(cache, "memory outside object was overwritten");
3095
3096        printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3097                        obj, redzone1, redzone2);
3098}
3099
3100static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
3101                                   void *caller)
3102{
3103        struct page *page;
3104        unsigned int objnr;
3105        struct slab *slabp;
3106
3107        BUG_ON(virt_to_cache(objp) != cachep);
3108
3109        objp -= obj_offset(cachep);
3110        kfree_debugcheck(objp);
3111        page = virt_to_head_page(objp);
3112
3113        slabp = page->slab_page;
3114
3115        if (cachep->flags & SLAB_RED_ZONE) {
3116                verify_redzone_free(cachep, objp);
3117                *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3118                *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3119        }
3120        if (cachep->flags & SLAB_STORE_USER)
3121                *dbg_userword(cachep, objp) = caller;
3122
3123        objnr = obj_to_index(cachep, slabp, objp);
3124
3125        BUG_ON(objnr >= cachep->num);
3126        BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3127
3128#ifdef CONFIG_DEBUG_SLAB_LEAK
3129        slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3130#endif
3131        if (cachep->flags & SLAB_POISON) {
3132#ifdef CONFIG_DEBUG_PAGEALLOC
3133                if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3134                        store_stackinfo(cachep, objp, (unsigned long)caller);
3135                        kernel_map_pages(virt_to_page(objp),
3136                                         cachep->size / PAGE_SIZE, 0);
3137                } else {
3138                        poison_obj(cachep, objp, POISON_FREE);
3139                }
3140#else
3141                poison_obj(cachep, objp, POISON_FREE);
3142#endif
3143        }
3144        return objp;
3145}
3146
3147static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3148{
3149        kmem_bufctl_t i;
3150        int entries = 0;
3151
3152        /* Check slab's freelist to see if this obj is there. */
3153        for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3154                entries++;
3155                if (entries > cachep->num || i >= cachep->num)
3156                        goto bad;
3157        }
3158        if (entries != cachep->num - slabp->inuse) {
3159bad:
3160                printk(KERN_ERR "slab: Internal list corruption detected in "
3161                        "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3162                        cachep->name, cachep->num, slabp, slabp->inuse,
3163                        print_tainted());
3164                print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3165                        sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3166                        1);
3167                BUG();
3168        }
3169}
3170#else
3171#define kfree_debugcheck(x) do { } while(0)
3172#define cache_free_debugcheck(x,objp,z) (objp)
3173#define check_slabp(x,y) do { } while(0)
3174#endif
3175
3176static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3177                                                        bool force_refill)
3178{
3179        int batchcount;
3180        struct kmem_list3 *l3;
3181        struct array_cache *ac;
3182        int node;
3183
3184        check_irq_off();
3185        node = numa_mem_id();
3186        if (unlikely(force_refill))
3187                goto force_grow;
3188retry:
3189        ac = cpu_cache_get(cachep);
3190        batchcount = ac->batchcount;
3191        if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3192                /*
3193                 * If there was little recent activity on this cache, then
3194                 * perform only a partial refill.  Otherwise we could generate
3195                 * refill bouncing.
3196                 */
3197                batchcount = BATCHREFILL_LIMIT;
3198        }
3199        l3 = cachep->nodelists[node];
3200
3201        BUG_ON(ac->avail > 0 || !l3);
3202        spin_lock(&l3->list_lock);
3203
3204        /* See if we can refill from the shared array */
3205        if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3206                l3->shared->touched = 1;
3207                goto alloc_done;
3208        }
3209
3210        while (batchcount > 0) {
3211                struct list_head *entry;
3212                struct slab *slabp;
3213                /* Get slab alloc is to come from. */
3214                entry = l3->slabs_partial.next;
3215                if (entry == &l3->slabs_partial) {
3216                        l3->free_touched = 1;
3217                        entry = l3->slabs_free.next;
3218                        if (entry == &l3->slabs_free)
3219                                goto must_grow;
3220                }
3221
3222                slabp = list_entry(entry, struct slab, list);
3223                check_slabp(cachep, slabp);
3224                check_spinlock_acquired(cachep);
3225
3226                /*
3227                 * The slab was either on partial or free list so
3228                 * there must be at least one object available for
3229                 * allocation.
3230                 */
3231                BUG_ON(slabp->inuse >= cachep->num);
3232
3233                while (slabp->inuse < cachep->num && batchcount--) {
3234                        STATS_INC_ALLOCED(cachep);
3235                        STATS_INC_ACTIVE(cachep);
3236                        STATS_SET_HIGH(cachep);
3237
3238                        ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3239                                                                        node));
3240                }
3241                check_slabp(cachep, slabp);
3242
3243                /* move slabp to correct slabp list: */
3244                list_del(&slabp->list);
3245                if (slabp->free == BUFCTL_END)
3246                        list_add(&slabp->list, &l3->slabs_full);
3247                else
3248                        list_add(&slabp->list, &l3->slabs_partial);
3249        }
3250
3251must_grow:
3252        l3->free_objects -= ac->avail;
3253alloc_done:
3254        spin_unlock(&l3->list_lock);
3255
3256        if (unlikely(!ac->avail)) {
3257                int x;
3258force_grow:
3259                x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3260
3261                /* cache_grow can reenable interrupts, then ac could change. */
3262                ac = cpu_cache_get(cachep);
3263                node = numa_mem_id();
3264
3265                /* no objects in sight? abort */
3266                if (!x && (ac->avail == 0 || force_refill))
3267                        return NULL;
3268
3269                if (!ac->avail)         /* objects refilled by interrupt? */
3270                        goto retry;
3271        }
3272        ac->touched = 1;
3273
3274        return ac_get_obj(cachep, ac, flags, force_refill);
3275}
3276
3277static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3278                                                gfp_t flags)
3279{
3280        might_sleep_if(flags & __GFP_WAIT);
3281#if DEBUG
3282        kmem_flagcheck(cachep, flags);
3283#endif
3284}
3285
3286#if DEBUG
3287static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3288                                gfp_t flags, void *objp, void *caller)
3289{
3290        if (!objp)
3291                return objp;
3292        if (cachep->flags & SLAB_POISON) {
3293#ifdef CONFIG_DEBUG_PAGEALLOC
3294                if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3295                        kernel_map_pages(virt_to_page(objp),
3296                                         cachep->size / PAGE_SIZE, 1);
3297                else
3298                        check_poison_obj(cachep, objp);
3299#else
3300                check_poison_obj(cachep, objp);
3301#endif
3302                poison_obj(cachep, objp, POISON_INUSE);
3303        }
3304        if (cachep->flags & SLAB_STORE_USER)
3305                *dbg_userword(cachep, objp) = caller;
3306
3307        if (cachep->flags & SLAB_RED_ZONE) {
3308                if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3309                                *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3310                        slab_error(cachep, "double free, or memory outside"
3311                                                " object was overwritten");
3312                        printk(KERN_ERR
3313                                "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3314                                objp, *dbg_redzone1(cachep, objp),
3315                                *dbg_redzone2(cachep, objp));
3316                }
3317                *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3318                *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3319        }
3320#ifdef CONFIG_DEBUG_SLAB_LEAK
3321        {
3322                struct slab *slabp;
3323                unsigned objnr;
3324
3325                slabp = virt_to_head_page(objp)->slab_page;
3326                objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3327                slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3328        }
3329#endif
3330        objp += obj_offset(cachep);
3331        if (cachep->ctor && cachep->flags & SLAB_POISON)
3332                cachep->ctor(objp);
3333        if (ARCH_SLAB_MINALIGN &&
3334            ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3335                printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3336                       objp, (int)ARCH_SLAB_MINALIGN);
3337        }
3338        return objp;
3339}
3340#else
3341#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3342#endif
3343
3344static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3345{
3346        if (cachep == &cache_cache)
3347                return false;
3348
3349        return should_failslab(cachep->object_size, flags, cachep->flags);
3350}
3351
3352static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3353{
3354        void *objp;
3355        struct array_cache *ac;
3356        bool force_refill = false;
3357
3358        check_irq_off();
3359
3360        ac = cpu_cache_get(cachep);
3361        if (likely(ac->avail)) {
3362                ac->touched = 1;
3363                objp = ac_get_obj(cachep, ac, flags, false);
3364
3365                /*
3366                 * Allow for the possibility all avail objects are not allowed
3367                 * by the current flags
3368                 */
3369                if (objp) {
3370                        STATS_INC_ALLOCHIT(cachep);
3371                        goto out;
3372                }
3373                force_refill = true;
3374        }
3375
3376        STATS_INC_ALLOCMISS(cachep);
3377        objp = cache_alloc_refill(cachep, flags, force_refill);
3378        /*
3379         * the 'ac' may be updated by cache_alloc_refill(),
3380         * and kmemleak_erase() requires its correct value.
3381         */
3382        ac = cpu_cache_get(cachep);
3383
3384out:
3385        /*
3386         * To avoid a false negative, if an object that is in one of the
3387         * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3388         * treat the array pointers as a reference to the object.
3389         */
3390        if (objp)
3391                kmemleak_erase(&ac->entry[ac->avail]);
3392        return objp;
3393}
3394
3395#ifdef CONFIG_NUMA
3396/*
3397 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3398 *
3399 * If we are in_interrupt, then process context, including cpusets and
3400 * mempolicy, may not apply and should not be used for allocation policy.
3401 */
3402static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3403{
3404        int nid_alloc, nid_here;
3405
3406        if (in_interrupt() || (flags & __GFP_THISNODE))
3407                return NULL;
3408        nid_alloc = nid_here = numa_mem_id();
3409        if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3410                nid_alloc = cpuset_slab_spread_node();
3411        else if (current->mempolicy)
3412                nid_alloc = slab_node();
3413        if (nid_alloc != nid_here)
3414                return ____cache_alloc_node(cachep, flags, nid_alloc);
3415        return NULL;
3416}
3417
3418/*
3419 * Fallback function if there was no memory available and no objects on a
3420 * certain node and fall back is permitted. First we scan all the
3421 * available nodelists for available objects. If that fails then we
3422 * perform an allocation without specifying a node. This allows the page
3423 * allocator to do its reclaim / fallback magic. We then insert the
3424 * slab into the proper nodelist and then allocate from it.
3425 */
3426static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3427{
3428        struct zonelist *zonelist;
3429        gfp_t local_flags;
3430        struct zoneref *z;
3431        struct zone *zone;
3432        enum zone_type high_zoneidx = gfp_zone(flags);
3433        void *obj = NULL;
3434        int nid;
3435        unsigned int cpuset_mems_cookie;
3436
3437        if (flags & __GFP_THISNODE)
3438                return NULL;
3439
3440        local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3441
3442retry_cpuset:
3443        cpuset_mems_cookie = get_mems_allowed();
3444        zonelist = node_zonelist(slab_node(), flags);
3445
3446retry:
3447        /*
3448         * Look through allowed nodes for objects available
3449         * from existing per node queues.
3450         */
3451        for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3452                nid = zone_to_nid(zone);
3453
3454                if (cpuset_zone_allowed_hardwall(zone, flags) &&
3455                        cache->nodelists[nid] &&
3456                        cache->nodelists[nid]->free_objects) {
3457                                obj = ____cache_alloc_node(cache,
3458                                        flags | GFP_THISNODE, nid);
3459                                if (obj)
3460                                        break;
3461                }
3462        }
3463
3464        if (!obj) {
3465                /*
3466                 * This allocation will be performed within the constraints
3467                 * of the current cpuset / memory policy requirements.
3468                 * We may trigger various forms of reclaim on the allowed
3469                 * set and go into memory reserves if necessary.
3470                 */
3471                if (local_flags & __GFP_WAIT)
3472                        local_irq_enable();
3473                kmem_flagcheck(cache, flags);
3474                obj = kmem_getpages(cache, local_flags, numa_mem_id());
3475                if (local_flags & __GFP_WAIT)
3476                        local_irq_disable();
3477                if (obj) {
3478                        /*
3479                         * Insert into the appropriate per node queues
3480                         */
3481                        nid = page_to_nid(virt_to_page(obj));
3482                        if (cache_grow(cache, flags, nid, obj)) {
3483                                obj = ____cache_alloc_node(cache,
3484                                        flags | GFP_THISNODE, nid);
3485                                if (!obj)
3486                                        /*
3487                                         * Another processor may allocate the
3488                                         * objects in the slab since we are
3489                                         * not holding any locks.
3490                                         */
3491                                        goto retry;
3492                        } else {
3493                                /* cache_grow already freed obj */
3494                                obj = NULL;
3495                        }
3496                }
3497        }
3498
3499        if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3500                goto retry_cpuset;
3501        return obj;
3502}
3503
3504/*
3505 * A interface to enable slab creation on nodeid
3506 */
3507static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3508                                int nodeid)
3509{
3510        struct list_head *entry;
3511        struct slab *slabp;
3512        struct kmem_list3 *l3;
3513        void *obj;
3514        int x;
3515
3516        l3 = cachep->nodelists[nodeid];
3517        BUG_ON(!l3);
3518
3519retry:
3520        check_irq_off();
3521        spin_lock(&l3->list_lock);
3522        entry = l3->slabs_partial.next;
3523        if (entry == &l3->slabs_partial) {
3524                l3->free_touched = 1;
3525                entry = l3->slabs_free.next;
3526                if (entry == &l3->slabs_free)
3527                        goto must_grow;
3528        }
3529
3530        slabp = list_entry(entry, struct slab, list);
3531        check_spinlock_acquired_node(cachep, nodeid);
3532        check_slabp(cachep, slabp);
3533
3534        STATS_INC_NODEALLOCS(cachep);
3535        STATS_INC_ACTIVE(cachep);
3536        STATS_SET_HIGH(cachep);
3537
3538        BUG_ON(slabp->inuse == cachep->num);
3539
3540        obj = slab_get_obj(cachep, slabp, nodeid);
3541        check_slabp(cachep, slabp);
3542        l3->free_objects--;
3543        /* move slabp to correct slabp list: */
3544        list_del(&slabp->list);
3545
3546        if (slabp->free == BUFCTL_END)
3547                list_add(&slabp->list, &l3->slabs_full);
3548        else
3549                list_add(&slabp->list, &l3->slabs_partial);
3550
3551        spin_unlock(&l3->list_lock);
3552        goto done;
3553
3554must_grow:
3555        spin_unlock(&l3->list_lock);
3556        x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3557        if (x)
3558                goto retry;
3559
3560        return fallback_alloc(cachep, flags);
3561
3562done:
3563        return obj;
3564}
3565
3566/**
3567 * kmem_cache_alloc_node - Allocate an object on the specified node
3568 * @cachep: The cache to allocate from.
3569 * @flags: See kmalloc().
3570 * @nodeid: node number of the target node.
3571 * @caller: return address of caller, used for debug information
3572 *
3573 * Identical to kmem_cache_alloc but it will allocate memory on the given
3574 * node, which can improve the performance for cpu bound structures.
3575 *
3576 * Fallback to other node is possible if __GFP_THISNODE is not set.
3577 */
3578static __always_inline void *
3579__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3580                   void *caller)
3581{
3582        unsigned long save_flags;
3583        void *ptr;
3584        int slab_node = numa_mem_id();
3585
3586        flags &= gfp_allowed_mask;
3587
3588        lockdep_trace_alloc(flags);
3589
3590        if (slab_should_failslab(cachep, flags))
3591                return NULL;
3592
3593        cache_alloc_debugcheck_before(cachep, flags);
3594        local_irq_save(save_flags);
3595
3596        if (nodeid == NUMA_NO_NODE)
3597                nodeid = slab_node;
3598
3599        if (unlikely(!cachep->nodelists[nodeid])) {
3600                /* Node not bootstrapped yet */
3601                ptr = fallback_alloc(cachep, flags);
3602                goto out;
3603        }
3604
3605        if (nodeid == slab_node) {
3606                /*
3607                 * Use the locally cached objects if possible.
3608                 * However ____cache_alloc does not allow fallback
3609                 * to other nodes. It may fail while we still have
3610                 * objects on other nodes available.
3611                 */
3612                ptr = ____cache_alloc(cachep, flags);
3613                if (ptr)
3614                        goto out;
3615        }
3616        /* ___cache_alloc_node can fall back to other nodes */
3617        ptr = ____cache_alloc_node(cachep, flags, nodeid);
3618  out:
3619        local_irq_restore(save_flags);
3620        ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3621        kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3622                                 flags);
3623
3624        if (likely(ptr))
3625                kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3626
3627        if (unlikely((flags & __GFP_ZERO) && ptr))
3628                memset(ptr, 0, cachep->object_size);
3629
3630        return ptr;
3631}
3632
3633static __always_inline void *
3634__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3635{
3636        void *objp;
3637
3638        if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3639                objp = alternate_node_alloc(cache, flags);
3640                if (objp)
3641                        goto out;
3642        }
3643        objp = ____cache_alloc(cache, flags);
3644
3645        /*
3646         * We may just have run out of memory on the local node.
3647         * ____cache_alloc_node() knows how to locate memory on other nodes
3648         */
3649        if (!objp)
3650                objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3651
3652  out:
3653        return objp;
3654}
3655#else
3656
3657static __always_inline void *
3658__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3659{
3660        return ____cache_alloc(cachep, flags);
3661}
3662
3663#endif /* CONFIG_NUMA */
3664
3665static __always_inline void *
3666__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3667{
3668        unsigned long save_flags;
3669        void *objp;
3670
3671        flags &= gfp_allowed_mask;
3672
3673        lockdep_trace_alloc(flags);
3674
3675        if (slab_should_failslab(cachep, flags))
3676                return NULL;
3677
3678        cache_alloc_debugcheck_before(cachep, flags);
3679        local_irq_save(save_flags);
3680        objp = __do_cache_alloc(cachep, flags);
3681        local_irq_restore(save_flags);
3682        objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3683        kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3684                                 flags);
3685        prefetchw(objp);
3686
3687        if (likely(objp))
3688                kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3689
3690        if (unlikely((flags & __GFP_ZERO) && objp))
3691                memset(objp, 0, cachep->object_size);
3692
3693        return objp;
3694}
3695
3696/*
3697 * Caller needs to acquire correct kmem_list's list_lock
3698 */
3699static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3700                       int node)
3701{
3702        int i;
3703        struct kmem_list3 *l3;
3704
3705        for (i = 0; i < nr_objects; i++) {
3706                void *objp;
3707                struct slab *slabp;
3708
3709                clear_obj_pfmemalloc(&objpp[i]);
3710                objp = objpp[i];
3711
3712                slabp = virt_to_slab(objp);
3713                l3 = cachep->nodelists[node];
3714                list_del(&slabp->list);
3715                check_spinlock_acquired_node(cachep, node);
3716                check_slabp(cachep, slabp);
3717                slab_put_obj(cachep, slabp, objp, node);
3718                STATS_DEC_ACTIVE(cachep);
3719                l3->free_objects++;
3720                check_slabp(cachep, slabp);
3721
3722                /* fixup slab chains */
3723                if (slabp->inuse == 0) {
3724                        if (l3->free_objects > l3->free_limit) {
3725                                l3->free_objects -= cachep->num;
3726                                /* No need to drop any previously held
3727                                 * lock here, even if we have a off-slab slab
3728                                 * descriptor it is guaranteed to come from
3729                                 * a different cache, refer to comments before
3730                                 * alloc_slabmgmt.
3731                                 */
3732                                slab_destroy(cachep, slabp);
3733                        } else {
3734                                list_add(&slabp->list, &l3->slabs_free);
3735                        }
3736                } else {
3737                        /* Unconditionally move a slab to the end of the
3738                         * partial list on free - maximum time for the
3739                         * other objects to be freed, too.
3740                         */
3741                        list_add_tail(&slabp->list, &l3->slabs_partial);
3742                }
3743        }
3744}
3745
3746static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3747{
3748        int batchcount;
3749        struct kmem_list3 *l3;
3750        int node = numa_mem_id();
3751
3752        batchcount = ac->batchcount;
3753#if DEBUG
3754        BUG_ON(!batchcount || batchcount > ac->avail);
3755#endif
3756        check_irq_off();
3757        l3 = cachep->nodelists[node];
3758        spin_lock(&l3->list_lock);
3759        if (l3->shared) {
3760                struct array_cache *shared_array = l3->shared;
3761                int max = shared_array->limit - shared_array->avail;
3762                if (max) {
3763                        if (batchcount > max)
3764                                batchcount = max;
3765                        memcpy(&(shared_array->entry[shared_array->avail]),
3766                               ac->entry, sizeof(void *) * batchcount);
3767                        shared_array->avail += batchcount;
3768                        goto free_done;
3769                }
3770        }
3771
3772        free_block(cachep, ac->entry, batchcount, node);
3773free_done:
3774#if STATS
3775        {
3776                int i = 0;
3777                struct list_head *p;
3778
3779                p = l3->slabs_free.next;
3780                while (p != &(l3->slabs_free)) {
3781                        struct slab *slabp;
3782
3783                        slabp = list_entry(p, struct slab, list);
3784                        BUG_ON(slabp->inuse);
3785
3786                        i++;
3787                        p = p->next;
3788                }
3789                STATS_SET_FREEABLE(cachep, i);
3790        }
3791#endif
3792        spin_unlock(&l3->list_lock);
3793        ac->avail -= batchcount;
3794        memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3795}
3796
3797/*
3798 * Release an obj back to its cache. If the obj has a constructed state, it must
3799 * be in this state _before_ it is released.  Called with disabled ints.
3800 */
3801static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3802    void *caller)
3803{
3804        struct array_cache *ac = cpu_cache_get(cachep);
3805
3806        check_irq_off();
3807        kmemleak_free_recursive(objp, cachep->flags);
3808        objp = cache_free_debugcheck(cachep, objp, caller);
3809
3810        kmemcheck_slab_free(cachep, objp, cachep->object_size);
3811
3812        /*
3813         * Skip calling cache_free_alien() when the platform is not numa.
3814         * This will avoid cache misses that happen while accessing slabp (which
3815         * is per page memory  reference) to get nodeid. Instead use a global
3816         * variable to skip the call, which is mostly likely to be present in
3817         * the cache.
3818         */
3819        if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3820                return;
3821
3822        if (likely(ac->avail < ac->limit)) {
3823                STATS_INC_FREEHIT(cachep);
3824        } else {
3825                STATS_INC_FREEMISS(cachep);
3826                cache_flusharray(cachep, ac);
3827        }
3828
3829        ac_put_obj(cachep, ac, objp);
3830}
3831
3832/**
3833 * kmem_cache_alloc - Allocate an object
3834 * @cachep: The cache to allocate from.
3835 * @flags: See kmalloc().
3836 *
3837 * Allocate an object from this cache.  The flags are only relevant
3838 * if the cache has no available objects.
3839 */
3840void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3841{
3842        void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3843
3844        trace_kmem_cache_alloc(_RET_IP_, ret,
3845                               cachep->object_size, cachep->size, flags);
3846
3847        return ret;
3848}
3849EXPORT_SYMBOL(kmem_cache_alloc);
3850
3851#ifdef CONFIG_TRACING
3852void *
3853kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3854{
3855        void *ret;
3856
3857        ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3858
3859        trace_kmalloc(_RET_IP_, ret,
3860                      size, slab_buffer_size(cachep), flags);
3861        return ret;
3862}
3863EXPORT_SYMBOL(kmem_cache_alloc_trace);
3864#endif
3865
3866#ifdef CONFIG_NUMA
3867void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3868{
3869        void *ret = __cache_alloc_node(cachep, flags, nodeid,
3870                                       __builtin_return_address(0));
3871
3872        trace_kmem_cache_alloc_node(_RET_IP_, ret,
3873                                    cachep->object_size, cachep->size,
3874                                    flags, nodeid);
3875
3876        return ret;
3877}
3878EXPORT_SYMBOL(kmem_cache_alloc_node);
3879
3880#ifdef CONFIG_TRACING
3881void *kmem_cache_alloc_node_trace(size_t size,
3882                                  struct kmem_cache *cachep,
3883                                  gfp_t flags,
3884                                  int nodeid)
3885{
3886        void *ret;
3887
3888        ret = __cache_alloc_node(cachep, flags, nodeid,
3889                                  __builtin_return_address(0));
3890        trace_kmalloc_node(_RET_IP_, ret,
3891                           size, slab_buffer_size(cachep),
3892                           flags, nodeid);
3893        return ret;
3894}
3895EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3896#endif
3897
3898static __always_inline void *
3899__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3900{
3901        struct kmem_cache *cachep;
3902
3903        cachep = kmem_find_general_cachep(size, flags);
3904        if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3905                return cachep;
3906        return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3907}
3908
3909#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3910void *__kmalloc_node(size_t size, gfp_t flags, int node)
3911{
3912        return __do_kmalloc_node(size, flags, node,
3913                        __builtin_return_address(0));
3914}
3915EXPORT_SYMBOL(__kmalloc_node);
3916
3917void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3918                int node, unsigned long caller)
3919{
3920        return __do_kmalloc_node(size, flags, node, (void *)caller);
3921}
3922EXPORT_SYMBOL(__kmalloc_node_track_caller);
3923#else
3924void *__kmalloc_node(size_t size, gfp_t flags, int node)
3925{
3926        return __do_kmalloc_node(size, flags, node, NULL);
3927}
3928EXPORT_SYMBOL(__kmalloc_node);
3929#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3930#endif /* CONFIG_NUMA */
3931
3932/**
3933 * __do_kmalloc - allocate memory
3934 * @size: how many bytes of memory are required.
3935 * @flags: the type of memory to allocate (see kmalloc).
3936 * @caller: function caller for debug tracking of the caller
3937 */
3938static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3939                                          void *caller)
3940{
3941        struct kmem_cache *cachep;
3942        void *ret;
3943
3944        /* If you want to save a few bytes .text space: replace
3945         * __ with kmem_.
3946         * Then kmalloc uses the uninlined functions instead of the inline
3947         * functions.
3948         */
3949        cachep = __find_general_cachep(size, flags);
3950        if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3951                return cachep;
3952        ret = __cache_alloc(cachep, flags, caller);
3953
3954        trace_kmalloc((unsigned long) caller, ret,
3955                      size, cachep->size, flags);
3956
3957        return ret;
3958}
3959
3960
3961#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3962void *__kmalloc(size_t size, gfp_t flags)
3963{
3964        return __do_kmalloc(size, flags, __builtin_return_address(0));
3965}
3966EXPORT_SYMBOL(__kmalloc);
3967
3968void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3969{
3970        return __do_kmalloc(size, flags, (void *)caller);
3971}
3972EXPORT_SYMBOL(__kmalloc_track_caller);
3973
3974#else
3975void *__kmalloc(size_t size, gfp_t flags)
3976{
3977        return __do_kmalloc(size, flags, NULL);
3978}
3979EXPORT_SYMBOL(__kmalloc);
3980#endif
3981
3982/**
3983 * kmem_cache_free - Deallocate an object
3984 * @cachep: The cache the allocation was from.
3985 * @objp: The previously allocated object.
3986 *
3987 * Free an object which was previously allocated from this
3988 * cache.
3989 */
3990void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3991{
3992        unsigned long flags;
3993
3994        local_irq_save(flags);
3995        debug_check_no_locks_freed(objp, cachep->object_size);
3996        if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3997                debug_check_no_obj_freed(objp, cachep->object_size);
3998        __cache_free(cachep, objp, __builtin_return_address(0));
3999        local_irq_restore(flags);
4000
4001        trace_kmem_cache_free(_RET_IP_, objp);
4002}
4003EXPORT_SYMBOL(kmem_cache_free);
4004
4005/**
4006 * kfree - free previously allocated memory
4007 * @objp: pointer returned by kmalloc.
4008 *
4009 * If @objp is NULL, no operation is performed.
4010 *
4011 * Don't free memory not originally allocated by kmalloc()
4012 * or you will run into trouble.
4013 */
4014void kfree(const void *objp)
4015{
4016        struct kmem_cache *c;
4017        unsigned long flags;
4018
4019        trace_kfree(_RET_IP_, objp);
4020
4021        if (unlikely(ZERO_OR_NULL_PTR(objp)))
4022                return;
4023        local_irq_save(flags);
4024        kfree_debugcheck(objp);
4025        c = virt_to_cache(objp);
4026        debug_check_no_locks_freed(objp, c->object_size);
4027
4028        debug_check_no_obj_freed(objp, c->object_size);
4029        __cache_free(c, (void *)objp, __builtin_return_address(0));
4030        local_irq_restore(flags);
4031}
4032EXPORT_SYMBOL(kfree);
4033
4034unsigned int kmem_cache_size(struct kmem_cache *cachep)
4035{
4036        return cachep->object_size;
4037}
4038EXPORT_SYMBOL(kmem_cache_size);
4039
4040/*
4041 * This initializes kmem_list3 or resizes various caches for all nodes.
4042 */
4043static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
4044{
4045        int node;
4046        struct kmem_list3 *l3;
4047        struct array_cache *new_shared;
4048        struct array_cache **new_alien = NULL;
4049
4050        for_each_online_node(node) {
4051
4052                if (use_alien_caches) {
4053                        new_alien = alloc_alien_cache(node, cachep->limit, gfp);
4054                        if (!new_alien)
4055                                goto fail;
4056                }
4057
4058                new_shared = NULL;
4059                if (cachep->shared) {
4060                        new_shared = alloc_arraycache(node,
4061                                cachep->shared*cachep->batchcount,
4062                                        0xbaadf00d, gfp);
4063                        if (!new_shared) {
4064                                free_alien_cache(new_alien);
4065                                goto fail;
4066                        }
4067                }
4068
4069                l3 = cachep->nodelists[node];
4070                if (l3) {
4071                        struct array_cache *shared = l3->shared;
4072
4073                        spin_lock_irq(&l3->list_lock);
4074
4075                        if (shared)
4076                                free_block(cachep, shared->entry,
4077                                                shared->avail, node);
4078
4079                        l3->shared = new_shared;
4080                        if (!l3->alien) {
4081                                l3->alien = new_alien;
4082                                new_alien = NULL;
4083                        }
4084                        l3->free_limit = (1 + nr_cpus_node(node)) *
4085                                        cachep->batchcount + cachep->num;
4086                        spin_unlock_irq(&l3->list_lock);
4087                        kfree(shared);
4088                        free_alien_cache(new_alien);
4089                        continue;
4090                }
4091                l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
4092                if (!l3) {
4093                        free_alien_cache(new_alien);
4094                        kfree(new_shared);
4095                        goto fail;
4096                }
4097
4098                kmem_list3_init(l3);
4099                l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
4100                                ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
4101                l3->shared = new_shared;
4102                l3->alien = new_alien;
4103                l3->free_limit = (1 + nr_cpus_node(node)) *
4104                                        cachep->batchcount + cachep->num;
4105                cachep->nodelists[node] = l3;
4106        }
4107        return 0;
4108
4109fail:
4110        if (!cachep->list.next) {
4111                /* Cache is not active yet. Roll back what we did */
4112                node--;
4113                while (node >= 0) {
4114                        if (cachep->nodelists[node]) {
4115                                l3 = cachep->nodelists[node];
4116
4117                                kfree(l3->shared);
4118                                free_alien_cache(l3->alien);
4119                                kfree(l3);
4120                                cachep->nodelists[node] = NULL;
4121                        }
4122                        node--;
4123                }
4124        }
4125        return -ENOMEM;
4126}
4127
4128struct ccupdate_struct {
4129        struct kmem_cache *cachep;
4130        struct array_cache *new[0];
4131};
4132
4133static void do_ccupdate_local(void *info)
4134{
4135        struct ccupdate_struct *new = info;
4136        struct array_cache *old;
4137
4138        check_irq_off();
4139        old = cpu_cache_get(new->cachep);
4140
4141        new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4142        new->new[smp_processor_id()] = old;
4143}
4144
4145/* Always called with the slab_mutex held */
4146static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4147                                int batchcount, int shared, gfp_t gfp)
4148{
4149        struct ccupdate_struct *new;
4150        int i;
4151
4152        new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4153                      gfp);
4154        if (!new)
4155                return -ENOMEM;
4156
4157        for_each_online_cpu(i) {
4158                new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4159                                                batchcount, gfp);
4160                if (!new->new[i]) {
4161                        for (i--; i >= 0; i--)
4162                                kfree(new->new[i]);
4163                        kfree(new);
4164                        return -ENOMEM;
4165                }
4166        }
4167        new->cachep = cachep;
4168
4169        on_each_cpu(do_ccupdate_local, (void *)new, 1);
4170
4171        check_irq_on();
4172        cachep->batchcount = batchcount;
4173        cachep->limit = limit;
4174        cachep->shared = shared;
4175
4176        for_each_online_cpu(i) {
4177                struct array_cache *ccold = new->new[i];
4178                if (!ccold)
4179                        continue;
4180                spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4181                free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4182                spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4183                kfree(ccold);
4184        }
4185        kfree(new);
4186        return alloc_kmemlist(cachep, gfp);
4187}
4188
4189/* Called with slab_mutex held always */
4190static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4191{
4192        int err;
4193        int limit, shared;
4194
4195        /*
4196         * The head array serves three purposes:
4197         * - create a LIFO ordering, i.e. return objects that are cache-warm
4198         * - reduce the number of spinlock operations.
4199         * - reduce the number of linked list operations on the slab and
4200         *   bufctl chains: array operations are cheaper.
4201         * The numbers are guessed, we should auto-tune as described by
4202         * Bonwick.
4203         */
4204        if (cachep->size > 131072)
4205                limit = 1;
4206        else if (cachep->size > PAGE_SIZE)
4207                limit = 8;
4208        else if (cachep->size > 1024)
4209                limit = 24;
4210        else if (cachep->size > 256)
4211                limit = 54;
4212        else
4213                limit = 120;
4214
4215        /*
4216         * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4217         * allocation behaviour: Most allocs on one cpu, most free operations
4218         * on another cpu. For these cases, an efficient object passing between
4219         * cpus is necessary. This is provided by a shared array. The array
4220         * replaces Bonwick's magazine layer.
4221         * On uniprocessor, it's functionally equivalent (but less efficient)
4222         * to a larger limit. Thus disabled by default.
4223         */
4224        shared = 0;
4225        if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4226                shared = 8;
4227
4228#if DEBUG
4229        /*
4230         * With debugging enabled, large batchcount lead to excessively long
4231         * periods with disabled local interrupts. Limit the batchcount
4232         */
4233        if (limit > 32)
4234                limit = 32;
4235#endif
4236        err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4237        if (err)
4238                printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4239                       cachep->name, -err);
4240        return err;
4241}
4242
4243/*
4244 * Drain an array if it contains any elements taking the l3 lock only if
4245 * necessary. Note that the l3 listlock also protects the array_cache
4246 * if drain_array() is used on the shared array.
4247 */
4248static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4249                         struct array_cache *ac, int force, int node)
4250{
4251        int tofree;
4252
4253        if (!ac || !ac->avail)
4254                return;
4255        if (ac->touched && !force) {
4256                ac->touched = 0;
4257        } else {
4258                spin_lock_irq(&l3->list_lock);
4259                if (ac->avail) {
4260                        tofree = force ? ac->avail : (ac->limit + 4) / 5;
4261                        if (tofree > ac->avail)
4262                                tofree = (ac->avail + 1) / 2;
4263                        free_block(cachep, ac->entry, tofree, node);
4264                        ac->avail -= tofree;
4265                        memmove(ac->entry, &(ac->entry[tofree]),
4266                                sizeof(void *) * ac->avail);
4267                }
4268                spin_unlock_irq(&l3->list_lock);
4269        }
4270}
4271
4272/**
4273 * cache_reap - Reclaim memory from caches.
4274 * @w: work descriptor
4275 *
4276 * Called from workqueue/eventd every few seconds.
4277 * Purpose:
4278 * - clear the per-cpu caches for this CPU.
4279 * - return freeable pages to the main free memory pool.
4280 *
4281 * If we cannot acquire the cache chain mutex then just give up - we'll try
4282 * again on the next iteration.
4283 */
4284static void cache_reap(struct work_struct *w)
4285{
4286        struct kmem_cache *searchp;
4287        struct kmem_list3 *l3;
4288        int node = numa_mem_id();
4289        struct delayed_work *work = to_delayed_work(w);
4290
4291        if (!mutex_trylock(&slab_mutex))
4292                /* Give up. Setup the next iteration. */
4293                goto out;
4294
4295        list_for_each_entry(searchp, &slab_caches, list) {
4296                check_irq_on();
4297
4298                /*
4299                 * We only take the l3 lock if absolutely necessary and we
4300                 * have established with reasonable certainty that
4301                 * we can do some work if the lock was obtained.
4302                 */
4303                l3 = searchp->nodelists[node];
4304
4305                reap_alien(searchp, l3);
4306
4307                drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4308
4309                /*
4310                 * These are racy checks but it does not matter
4311                 * if we skip one check or scan twice.
4312                 */
4313                if (time_after(l3->next_reap, jiffies))
4314                        goto next;
4315
4316                l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4317
4318                drain_array(searchp, l3, l3->shared, 0, node);
4319
4320                if (l3->free_touched)
4321                        l3->free_touched = 0;
4322                else {
4323                        int freed;
4324
4325                        freed = drain_freelist(searchp, l3, (l3->free_limit +
4326                                5 * searchp->num - 1) / (5 * searchp->num));
4327                        STATS_ADD_REAPED(searchp, freed);
4328                }
4329next:
4330                cond_resched();
4331        }
4332        check_irq_on();
4333        mutex_unlock(&slab_mutex);
4334        next_reap_node();
4335out:
4336        /* Set up the next iteration */
4337        schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4338}
4339
4340#ifdef CONFIG_SLABINFO
4341
4342static void print_slabinfo_header(struct seq_file *m)
4343{
4344        /*
4345         * Output format version, so at least we can change it
4346         * without _too_ many complaints.
4347         */
4348#if STATS
4349        seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4350#else
4351        seq_puts(m, "slabinfo - version: 2.1\n");
4352#endif
4353        seq_puts(m, "# name            <active_objs> <num_objs> <objsize> "
4354                 "<objperslab> <pagesperslab>");
4355        seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4356        seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4357#if STATS
4358        seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4359                 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4360        seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4361#endif
4362        seq_putc(m, '\n');
4363}
4364
4365static void *s_start(struct seq_file *m, loff_t *pos)
4366{
4367        loff_t n = *pos;
4368
4369        mutex_lock(&slab_mutex);
4370        if (!n)
4371                print_slabinfo_header(m);
4372
4373        return seq_list_start(&slab_caches, *pos);
4374}
4375
4376static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4377{
4378        return seq_list_next(p, &slab_caches, pos);
4379}
4380
4381static void s_stop(struct seq_file *m, void *p)
4382{
4383        mutex_unlock(&slab_mutex);
4384}
4385
4386static int s_show(struct seq_file *m, void *p)
4387{
4388        struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4389        struct slab *slabp;
4390        unsigned long active_objs;
4391        unsigned long num_objs;
4392        unsigned long active_slabs = 0;
4393        unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4394        const char *name;
4395        char *error = NULL;
4396        int node;
4397        struct kmem_list3 *l3;
4398
4399        active_objs = 0;
4400        num_slabs = 0;
4401        for_each_online_node(node) {
4402                l3 = cachep->nodelists[node];
4403                if (!l3)
4404                        continue;
4405
4406                check_irq_on();
4407                spin_lock_irq(&l3->list_lock);
4408
4409                list_for_each_entry(slabp, &l3->slabs_full, list) {
4410                        if (slabp->inuse != cachep->num && !error)
4411                                error = "slabs_full accounting error";
4412                        active_objs += cachep->num;
4413                        active_slabs++;
4414                }
4415                list_for_each_entry(slabp, &l3->slabs_partial, list) {
4416                        if (slabp->inuse == cachep->num && !error)
4417                                error = "slabs_partial inuse accounting error";
4418                        if (!slabp->inuse && !error)
4419                                error = "slabs_partial/inuse accounting error";
4420                        active_objs += slabp->inuse;
4421                        active_slabs++;
4422                }
4423                list_for_each_entry(slabp, &l3->slabs_free, list) {
4424                        if (slabp->inuse && !error)
4425                                error = "slabs_free/inuse accounting error";
4426                        num_slabs++;
4427                }
4428                free_objects += l3->free_objects;
4429                if (l3->shared)
4430                        shared_avail += l3->shared->avail;
4431
4432                spin_unlock_irq(&l3->list_lock);
4433        }
4434        num_slabs += active_slabs;
4435        num_objs = num_slabs * cachep->num;
4436        if (num_objs - active_objs != free_objects && !error)
4437                error = "free_objects accounting error";
4438
4439        name = cachep->name;
4440        if (error)
4441                printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4442
4443        seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4444                   name, active_objs, num_objs, cachep->size,
4445                   cachep->num, (1 << cachep->gfporder));
4446        seq_printf(m, " : tunables %4u %4u %4u",
4447                   cachep->limit, cachep->batchcount, cachep->shared);
4448        seq_printf(m, " : slabdata %6lu %6lu %6lu",
4449                   active_slabs, num_slabs, shared_avail);
4450#if STATS
4451        {                       /* list3 stats */
4452                unsigned long high = cachep->high_mark;
4453                unsigned long allocs = cachep->num_allocations;
4454                unsigned long grown = cachep->grown;
4455                unsigned long reaped = cachep->reaped;
4456                unsigned long errors = cachep->errors;
4457                unsigned long max_freeable = cachep->max_freeable;
4458                unsigned long node_allocs = cachep->node_allocs;
4459                unsigned long node_frees = cachep->node_frees;
4460                unsigned long overflows = cachep->node_overflow;
4461
4462                seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4463                           "%4lu %4lu %4lu %4lu %4lu",
4464                           allocs, high, grown,
4465                           reaped, errors, max_freeable, node_allocs,
4466                           node_frees, overflows);
4467        }
4468        /* cpu stats */
4469        {
4470                unsigned long allochit = atomic_read(&cachep->allochit);
4471                unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4472                unsigned long freehit = atomic_read(&cachep->freehit);
4473                unsigned long freemiss = atomic_read(&cachep->freemiss);
4474
4475                seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4476                           allochit, allocmiss, freehit, freemiss);
4477        }
4478#endif
4479        seq_putc(m, '\n');
4480        return 0;
4481}
4482
4483/*
4484 * slabinfo_op - iterator that generates /proc/slabinfo
4485 *
4486 * Output layout:
4487 * cache-name
4488 * num-active-objs
4489 * total-objs
4490 * object size
4491 * num-active-slabs
4492 * total-slabs
4493 * num-pages-per-slab
4494 * + further values on SMP and with statistics enabled
4495 */
4496
4497static const struct seq_operations slabinfo_op = {
4498        .start = s_start,
4499        .next = s_next,
4500        .stop = s_stop,
4501        .show = s_show,
4502};
4503
4504#define MAX_SLABINFO_WRITE 128
4505/**
4506 * slabinfo_write - Tuning for the slab allocator
4507 * @file: unused
4508 * @buffer: user buffer
4509 * @count: data length
4510 * @ppos: unused
4511 */
4512static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4513                       size_t count, loff_t *ppos)
4514{
4515        char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4516        int limit, batchcount, shared, res;
4517        struct kmem_cache *cachep;
4518
4519        if (count > MAX_SLABINFO_WRITE)
4520                return -EINVAL;
4521        if (copy_from_user(&kbuf, buffer, count))
4522                return -EFAULT;
4523        kbuf[MAX_SLABINFO_WRITE] = '\0';
4524
4525        tmp = strchr(kbuf, ' ');
4526        if (!tmp)
4527                return -EINVAL;
4528        *tmp = '\0';
4529        tmp++;
4530        if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4531                return -EINVAL;
4532
4533        /* Find the cache in the chain of caches. */
4534        mutex_lock(&slab_mutex);
4535        res = -EINVAL;
4536        list_for_each_entry(cachep, &slab_caches, list) {
4537                if (!strcmp(cachep->name, kbuf)) {
4538                        if (limit < 1 || batchcount < 1 ||
4539                                        batchcount > limit || shared < 0) {
4540                                res = 0;
4541                        } else {
4542                                res = do_tune_cpucache(cachep, limit,
4543                                                       batchcount, shared,
4544                                                       GFP_KERNEL);
4545                        }
4546                        break;
4547                }
4548        }
4549        mutex_unlock(&slab_mutex);
4550        if (res >= 0)
4551                res = count;
4552        return res;
4553}
4554
4555static int slabinfo_open(struct inode *inode, struct file *file)
4556{
4557        return seq_open(file, &slabinfo_op);
4558}
4559
4560static const struct file_operations proc_slabinfo_operations = {
4561        .open           = slabinfo_open,
4562        .read           = seq_read,
4563        .write          = slabinfo_write,
4564        .llseek         = seq_lseek,
4565        .release        = seq_release,
4566};
4567
4568#ifdef CONFIG_DEBUG_SLAB_LEAK
4569
4570static void *leaks_start(struct seq_file *m, loff_t *pos)
4571{
4572        mutex_lock(&slab_mutex);
4573        return seq_list_start(&slab_caches, *pos);
4574}
4575
4576static inline int add_caller(unsigned long *n, unsigned long v)
4577{
4578        unsigned long *p;
4579        int l;
4580        if (!v)
4581                return 1;
4582        l = n[1];
4583        p = n + 2;
4584        while (l) {
4585                int i = l/2;
4586                unsigned long *q = p + 2 * i;
4587                if (*q == v) {
4588                        q[1]++;
4589                        return 1;
4590                }
4591                if (*q > v) {
4592                        l = i;
4593                } else {
4594                        p = q + 2;
4595                        l -= i + 1;
4596                }
4597        }
4598        if (++n[1] == n[0])
4599                return 0;
4600        memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4601        p[0] = v;
4602        p[1] = 1;
4603        return 1;
4604}
4605
4606static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4607{
4608        void *p;
4609        int i;
4610        if (n[0] == n[1])
4611                return;
4612        for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4613                if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4614                        continue;
4615                if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4616                        return;
4617        }
4618}
4619
4620static void show_symbol(struct seq_file *m, unsigned long address)
4621{
4622#ifdef CONFIG_KALLSYMS
4623        unsigned long offset, size;
4624        char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4625
4626        if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4627                seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4628                if (modname[0])
4629                        seq_printf(m, " [%s]", modname);
4630                return;
4631        }
4632#endif
4633        seq_printf(m, "%p", (void *)address);
4634}
4635
4636static int leaks_show(struct seq_file *m, void *p)
4637{
4638        struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4639        struct slab *slabp;
4640        struct kmem_list3 *l3;
4641        const char *name;
4642        unsigned long *n = m->private;
4643        int node;
4644        int i;
4645
4646        if (!(cachep->flags & SLAB_STORE_USER))
4647                return 0;
4648        if (!(cachep->flags & SLAB_RED_ZONE))
4649                return 0;
4650
4651        /* OK, we can do it */
4652
4653        n[1] = 0;
4654
4655        for_each_online_node(node) {
4656                l3 = cachep->nodelists[node];
4657                if (!l3)
4658                        continue;
4659
4660                check_irq_on();
4661                spin_lock_irq(&l3->list_lock);
4662
4663                list_for_each_entry(slabp, &l3->slabs_full, list)
4664                        handle_slab(n, cachep, slabp);
4665                list_for_each_entry(slabp, &l3->slabs_partial, list)
4666                        handle_slab(n, cachep, slabp);
4667                spin_unlock_irq(&l3->list_lock);
4668        }
4669        name = cachep->name;
4670        if (n[0] == n[1]) {
4671                /* Increase the buffer size */
4672                mutex_unlock(&slab_mutex);
4673                m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4674                if (!m->private) {
4675                        /* Too bad, we are really out */
4676                        m->private = n;
4677                        mutex_lock(&slab_mutex);
4678                        return -ENOMEM;
4679                }
4680                *(unsigned long *)m->private = n[0] * 2;
4681                kfree(n);
4682                mutex_lock(&slab_mutex);
4683                /* Now make sure this entry will be retried */
4684                m->count = m->size;
4685                return 0;
4686        }
4687        for (i = 0; i < n[1]; i++) {
4688                seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4689                show_symbol(m, n[2*i+2]);
4690                seq_putc(m, '\n');
4691        }
4692
4693        return 0;
4694}
4695
4696static const struct seq_operations slabstats_op = {
4697        .start = leaks_start,
4698        .next = s_next,
4699        .stop = s_stop,
4700        .show = leaks_show,
4701};
4702
4703static int slabstats_open(struct inode *inode, struct file *file)
4704{
4705        unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4706        int ret = -ENOMEM;
4707        if (n) {
4708                ret = seq_open(file, &slabstats_op);
4709                if (!ret) {
4710                        struct seq_file *m = file->private_data;
4711                        *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4712                        m->private = n;
4713                        n = NULL;
4714                }
4715                kfree(n);
4716        }
4717        return ret;
4718}
4719
4720static const struct file_operations proc_slabstats_operations = {
4721        .open           = slabstats_open,
4722        .read           = seq_read,
4723        .llseek         = seq_lseek,
4724        .release        = seq_release_private,
4725};
4726#endif
4727
4728static int __init slab_proc_init(void)
4729{
4730        proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4731#ifdef CONFIG_DEBUG_SLAB_LEAK
4732        proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4733#endif
4734        return 0;
4735}
4736module_init(slab_proc_init);
4737#endif
4738
4739/**
4740 * ksize - get the actual amount of memory allocated for a given object
4741 * @objp: Pointer to the object
4742 *
4743 * kmalloc may internally round up allocations and return more memory
4744 * than requested. ksize() can be used to determine the actual amount of
4745 * memory allocated. The caller may use this additional memory, even though
4746 * a smaller amount of memory was initially specified with the kmalloc call.
4747 * The caller must guarantee that objp points to a valid object previously
4748 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4749 * must not be freed during the duration of the call.
4750 */
4751size_t ksize(const void *objp)
4752{
4753        BUG_ON(!objp);
4754        if (unlikely(objp == ZERO_SIZE_PTR))
4755                return 0;
4756
4757        return virt_to_cache(objp)->object_size;
4758}
4759EXPORT_SYMBOL(ksize);
4760
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.