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        <linux/mm.h>
  91#include        <linux/poison.h>
  92#include        <linux/swap.h>
  93#include        <linux/cache.h>
  94#include        <linux/interrupt.h>
  95#include        <linux/init.h>
  96#include        <linux/compiler.h>
  97#include        <linux/cpuset.h>
  98#include        <linux/proc_fs.h>
  99#include        <linux/seq_file.h>
 100#include        <linux/notifier.h>
 101#include        <linux/kallsyms.h>
 102#include        <linux/cpu.h>
 103#include        <linux/sysctl.h>
 104#include        <linux/module.h>
 105#include        <linux/rcupdate.h>
 106#include        <linux/string.h>
 107#include        <linux/uaccess.h>
 108#include        <linux/nodemask.h>
 109#include        <linux/kmemleak.h>
 110#include        <linux/mempolicy.h>
 111#include        <linux/mutex.h>
 112#include        <linux/fault-inject.h>
 113#include        <linux/rtmutex.h>
 114#include        <linux/reciprocal_div.h>
 115#include        <linux/debugobjects.h>
 116#include        <linux/kmemcheck.h>
 117#include        <linux/memory.h>
 118#include        <linux/prefetch.h>
 119
 120#include        <net/sock.h>
 121
 122#include        <asm/cacheflush.h>
 123#include        <asm/tlbflush.h>
 124#include        <asm/page.h>
 125
 126#include <trace/events/kmem.h>
 127
 128#include        "internal.h"
 129
 130#include        "slab.h"
 131
 132/*
 133 * DEBUG        - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
 134 *                0 for faster, smaller code (especially in the critical paths).
 135 *
 136 * STATS        - 1 to collect stats for /proc/slabinfo.
 137 *                0 for faster, smaller code (especially in the critical paths).
 138 *
 139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
 140 */
 141
 142#ifdef CONFIG_DEBUG_SLAB
 143#define DEBUG           1
 144#define STATS           1
 145#define FORCED_DEBUG    1
 146#else
 147#define DEBUG           0
 148#define STATS           0
 149#define FORCED_DEBUG    0
 150#endif
 151
 152/* Shouldn't this be in a header file somewhere? */
 153#define BYTES_PER_WORD          sizeof(void *)
 154#define REDZONE_ALIGN           max(BYTES_PER_WORD, __alignof__(unsigned long long))
 155
 156#ifndef ARCH_KMALLOC_FLAGS
 157#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
 158#endif
 159
 160/*
 161 * true if a page was allocated from pfmemalloc reserves for network-based
 162 * swap
 163 */
 164static bool pfmemalloc_active __read_mostly;
 165
 166/*
 167 * kmem_bufctl_t:
 168 *
 169 * Bufctl's are used for linking objs within a slab
 170 * linked offsets.
 171 *
 172 * This implementation relies on "struct page" for locating the cache &
 173 * slab an object belongs to.
 174 * This allows the bufctl structure to be small (one int), but limits
 175 * the number of objects a slab (not a cache) can contain when off-slab
 176 * bufctls are used. The limit is the size of the largest general cache
 177 * that does not use off-slab slabs.
 178 * For 32bit archs with 4 kB pages, is this 56.
 179 * This is not serious, as it is only for large objects, when it is unwise
 180 * to have too many per slab.
 181 * Note: This limit can be raised by introducing a general cache whose size
 182 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
 183 */
 184
 185typedef unsigned int kmem_bufctl_t;
 186#define BUFCTL_END      (((kmem_bufctl_t)(~0U))-0)
 187#define BUFCTL_FREE     (((kmem_bufctl_t)(~0U))-1)
 188#define BUFCTL_ACTIVE   (((kmem_bufctl_t)(~0U))-2)
 189#define SLAB_LIMIT      (((kmem_bufctl_t)(~0U))-3)
 190
 191/*
 192 * struct slab_rcu
 193 *
 194 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
 195 * arrange for kmem_freepages to be called via RCU.  This is useful if
 196 * we need to approach a kernel structure obliquely, from its address
 197 * obtained without the usual locking.  We can lock the structure to
 198 * stabilize it and check it's still at the given address, only if we
 199 * can be sure that the memory has not been meanwhile reused for some
 200 * other kind of object (which our subsystem's lock might corrupt).
 201 *
 202 * rcu_read_lock before reading the address, then rcu_read_unlock after
 203 * taking the spinlock within the structure expected at that address.
 204 */
 205struct slab_rcu {
 206        struct rcu_head head;
 207        struct kmem_cache *cachep;
 208        void *addr;
 209};
 210
 211/*
 212 * struct slab
 213 *
 214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
 215 * for a slab, or allocated from an general cache.
 216 * Slabs are chained into three list: fully used, partial, fully free slabs.
 217 */
 218struct slab {
 219        union {
 220                struct {
 221                        struct list_head list;
 222                        unsigned long colouroff;
 223                        void *s_mem;            /* including colour offset */
 224                        unsigned int inuse;     /* num of objs active in slab */
 225                        kmem_bufctl_t free;
 226                        unsigned short nodeid;
 227                };
 228                struct slab_rcu __slab_cover_slab_rcu;
 229        };
 230};
 231
 232/*
 233 * struct array_cache
 234 *
 235 * Purpose:
 236 * - LIFO ordering, to hand out cache-warm objects from _alloc
 237 * - reduce the number of linked list operations
 238 * - reduce spinlock operations
 239 *
 240 * The limit is stored in the per-cpu structure to reduce the data cache
 241 * footprint.
 242 *
 243 */
 244struct array_cache {
 245        unsigned int avail;
 246        unsigned int limit;
 247        unsigned int batchcount;
 248        unsigned int touched;
 249        spinlock_t lock;
 250        void *entry[];  /*
 251                         * Must have this definition in here for the proper
 252                         * alignment of array_cache. Also simplifies accessing
 253                         * the entries.
 254                         *
 255                         * Entries should not be directly dereferenced as
 256                         * entries belonging to slabs marked pfmemalloc will
 257                         * have the lower bits set SLAB_OBJ_PFMEMALLOC
 258                         */
 259};
 260
 261#define SLAB_OBJ_PFMEMALLOC     1
 262static inline bool is_obj_pfmemalloc(void *objp)
 263{
 264        return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
 265}
 266
 267static inline void set_obj_pfmemalloc(void **objp)
 268{
 269        *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
 270        return;
 271}
 272
 273static inline void clear_obj_pfmemalloc(void **objp)
 274{
 275        *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
 276}
 277
 278/*
 279 * bootstrap: The caches do not work without cpuarrays anymore, but the
 280 * cpuarrays are allocated from the generic caches...
 281 */
 282#define BOOT_CPUCACHE_ENTRIES   1
 283struct arraycache_init {
 284        struct array_cache cache;
 285        void *entries[BOOT_CPUCACHE_ENTRIES];
 286};
 287
 288/*
 289 * Need this for bootstrapping a per node allocator.
 290 */
 291#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
 292static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
 293#define CACHE_CACHE 0
 294#define SIZE_AC MAX_NUMNODES
 295#define SIZE_NODE (2 * MAX_NUMNODES)
 296
 297static int drain_freelist(struct kmem_cache *cache,
 298                        struct kmem_cache_node *n, int tofree);
 299static void free_block(struct kmem_cache *cachep, void **objpp, int len,
 300                        int node);
 301static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
 302static void cache_reap(struct work_struct *unused);
 303
 304static int slab_early_init = 1;
 305
 306#define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
 307#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
 308
 309static void kmem_cache_node_init(struct kmem_cache_node *parent)
 310{
 311        INIT_LIST_HEAD(&parent->slabs_full);
 312        INIT_LIST_HEAD(&parent->slabs_partial);
 313        INIT_LIST_HEAD(&parent->slabs_free);
 314        parent->shared = NULL;
 315        parent->alien = NULL;
 316        parent->colour_next = 0;
 317        spin_lock_init(&parent->list_lock);
 318        parent->free_objects = 0;
 319        parent->free_touched = 0;
 320}
 321
 322#define MAKE_LIST(cachep, listp, slab, nodeid)                          \
 323        do {                                                            \
 324                INIT_LIST_HEAD(listp);                                  \
 325                list_splice(&(cachep->node[nodeid]->slab), listp);      \
 326        } while (0)
 327
 328#define MAKE_ALL_LISTS(cachep, ptr, nodeid)                             \
 329        do {                                                            \
 330        MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid);  \
 331        MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
 332        MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid);  \
 333        } while (0)
 334
 335#define CFLGS_OFF_SLAB          (0x80000000UL)
 336#define OFF_SLAB(x)     ((x)->flags & CFLGS_OFF_SLAB)
 337
 338#define BATCHREFILL_LIMIT       16
 339/*
 340 * Optimization question: fewer reaps means less probability for unnessary
 341 * cpucache drain/refill cycles.
 342 *
 343 * OTOH the cpuarrays can contain lots of objects,
 344 * which could lock up otherwise freeable slabs.
 345 */
 346#define REAPTIMEOUT_CPUC        (2*HZ)
 347#define REAPTIMEOUT_LIST3       (4*HZ)
 348
 349#if STATS
 350#define STATS_INC_ACTIVE(x)     ((x)->num_active++)
 351#define STATS_DEC_ACTIVE(x)     ((x)->num_active--)
 352#define STATS_INC_ALLOCED(x)    ((x)->num_allocations++)
 353#define STATS_INC_GROWN(x)      ((x)->grown++)
 354#define STATS_ADD_REAPED(x,y)   ((x)->reaped += (y))
 355#define STATS_SET_HIGH(x)                                               \
 356        do {                                                            \
 357                if ((x)->num_active > (x)->high_mark)                   \
 358                        (x)->high_mark = (x)->num_active;               \
 359        } while (0)
 360#define STATS_INC_ERR(x)        ((x)->errors++)
 361#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
 362#define STATS_INC_NODEFREES(x)  ((x)->node_frees++)
 363#define STATS_INC_ACOVERFLOW(x)   ((x)->node_overflow++)
 364#define STATS_SET_FREEABLE(x, i)                                        \
 365        do {                                                            \
 366                if ((x)->max_freeable < i)                              \
 367                        (x)->max_freeable = i;                          \
 368        } while (0)
 369#define STATS_INC_ALLOCHIT(x)   atomic_inc(&(x)->allochit)
 370#define STATS_INC_ALLOCMISS(x)  atomic_inc(&(x)->allocmiss)
 371#define STATS_INC_FREEHIT(x)    atomic_inc(&(x)->freehit)
 372#define STATS_INC_FREEMISS(x)   atomic_inc(&(x)->freemiss)
 373#else
 374#define STATS_INC_ACTIVE(x)     do { } while (0)
 375#define STATS_DEC_ACTIVE(x)     do { } while (0)
 376#define STATS_INC_ALLOCED(x)    do { } while (0)
 377#define STATS_INC_GROWN(x)      do { } while (0)
 378#define STATS_ADD_REAPED(x,y)   do { (void)(y); } while (0)
 379#define STATS_SET_HIGH(x)       do { } while (0)
 380#define STATS_INC_ERR(x)        do { } while (0)
 381#define STATS_INC_NODEALLOCS(x) do { } while (0)
 382#define STATS_INC_NODEFREES(x)  do { } while (0)
 383#define STATS_INC_ACOVERFLOW(x)   do { } while (0)
 384#define STATS_SET_FREEABLE(x, i) do { } while (0)
 385#define STATS_INC_ALLOCHIT(x)   do { } while (0)
 386#define STATS_INC_ALLOCMISS(x)  do { } while (0)
 387#define STATS_INC_FREEHIT(x)    do { } while (0)
 388#define STATS_INC_FREEMISS(x)   do { } while (0)
 389#endif
 390
 391#if DEBUG
 392
 393/*
 394 * memory layout of objects:
 395 * 0            : objp
 396 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
 397 *              the end of an object is aligned with the end of the real
 398 *              allocation. Catches writes behind the end of the allocation.
 399 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
 400 *              redzone word.
 401 * cachep->obj_offset: The real object.
 402 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
 403 * cachep->size - 1* BYTES_PER_WORD: last caller address
 404 *                                      [BYTES_PER_WORD long]
 405 */
 406static int obj_offset(struct kmem_cache *cachep)
 407{
 408        return cachep->obj_offset;
 409}
 410
 411static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
 412{
 413        BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
 414        return (unsigned long long*) (objp + obj_offset(cachep) -
 415                                      sizeof(unsigned long long));
 416}
 417
 418static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
 419{
 420        BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
 421        if (cachep->flags & SLAB_STORE_USER)
 422                return (unsigned long long *)(objp + cachep->size -
 423                                              sizeof(unsigned long long) -
 424                                              REDZONE_ALIGN);
 425        return (unsigned long long *) (objp + cachep->size -
 426                                       sizeof(unsigned long long));
 427}
 428
 429static void **dbg_userword(struct kmem_cache *cachep, void *objp)
 430{
 431        BUG_ON(!(cachep->flags & SLAB_STORE_USER));
 432        return (void **)(objp + cachep->size - BYTES_PER_WORD);
 433}
 434
 435#else
 436
 437#define obj_offset(x)                   0
 438#define dbg_redzone1(cachep, objp)      ({BUG(); (unsigned long long *)NULL;})
 439#define dbg_redzone2(cachep, objp)      ({BUG(); (unsigned long long *)NULL;})
 440#define dbg_userword(cachep, objp)      ({BUG(); (void **)NULL;})
 441
 442#endif
 443
 444/*
 445 * Do not go above this order unless 0 objects fit into the slab or
 446 * overridden on the command line.
 447 */
 448#define SLAB_MAX_ORDER_HI       1
 449#define SLAB_MAX_ORDER_LO       0
 450static int slab_max_order = SLAB_MAX_ORDER_LO;
 451static bool slab_max_order_set __initdata;
 452
 453static inline struct kmem_cache *virt_to_cache(const void *obj)
 454{
 455        struct page *page = virt_to_head_page(obj);
 456        return page->slab_cache;
 457}
 458
 459static inline struct slab *virt_to_slab(const void *obj)
 460{
 461        struct page *page = virt_to_head_page(obj);
 462
 463        VM_BUG_ON(!PageSlab(page));
 464        return page->slab_page;
 465}
 466
 467static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
 468                                 unsigned int idx)
 469{
 470        return slab->s_mem + cache->size * idx;
 471}
 472
 473/*
 474 * We want to avoid an expensive divide : (offset / cache->size)
 475 *   Using the fact that size is a constant for a particular cache,
 476 *   we can replace (offset / cache->size) by
 477 *   reciprocal_divide(offset, cache->reciprocal_buffer_size)
 478 */
 479static inline unsigned int obj_to_index(const struct kmem_cache *cache,
 480                                        const struct slab *slab, void *obj)
 481{
 482        u32 offset = (obj - slab->s_mem);
 483        return reciprocal_divide(offset, cache->reciprocal_buffer_size);
 484}
 485
 486static struct arraycache_init initarray_generic =
 487    { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
 488
 489/* internal cache of cache description objs */
 490static struct kmem_cache kmem_cache_boot = {
 491        .batchcount = 1,
 492        .limit = BOOT_CPUCACHE_ENTRIES,
 493        .shared = 1,
 494        .size = sizeof(struct kmem_cache),
 495        .name = "kmem_cache",
 496};
 497
 498#define BAD_ALIEN_MAGIC 0x01020304ul
 499
 500#ifdef CONFIG_LOCKDEP
 501
 502/*
 503 * Slab sometimes uses the kmalloc slabs to store the slab headers
 504 * for other slabs "off slab".
 505 * The locking for this is tricky in that it nests within the locks
 506 * of all other slabs in a few places; to deal with this special
 507 * locking we put on-slab caches into a separate lock-class.
 508 *
 509 * We set lock class for alien array caches which are up during init.
 510 * The lock annotation will be lost if all cpus of a node goes down and
 511 * then comes back up during hotplug
 512 */
 513static struct lock_class_key on_slab_l3_key;
 514static struct lock_class_key on_slab_alc_key;
 515
 516static struct lock_class_key debugobj_l3_key;
 517static struct lock_class_key debugobj_alc_key;
 518
 519static void slab_set_lock_classes(struct kmem_cache *cachep,
 520                struct lock_class_key *l3_key, struct lock_class_key *alc_key,
 521                int q)
 522{
 523        struct array_cache **alc;
 524        struct kmem_cache_node *n;
 525        int r;
 526
 527        n = cachep->node[q];
 528        if (!n)
 529                return;
 530
 531        lockdep_set_class(&n->list_lock, l3_key);
 532        alc = n->alien;
 533        /*
 534         * FIXME: This check for BAD_ALIEN_MAGIC
 535         * should go away when common slab code is taught to
 536         * work even without alien caches.
 537         * Currently, non NUMA code returns BAD_ALIEN_MAGIC
 538         * for alloc_alien_cache,
 539         */
 540        if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
 541                return;
 542        for_each_node(r) {
 543                if (alc[r])
 544                        lockdep_set_class(&alc[r]->lock, alc_key);
 545        }
 546}
 547
 548static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
 549{
 550        slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
 551}
 552
 553static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
 554{
 555        int node;
 556
 557        for_each_online_node(node)
 558                slab_set_debugobj_lock_classes_node(cachep, node);
 559}
 560
 561static void init_node_lock_keys(int q)
 562{
 563        int i;
 564
 565        if (slab_state < UP)
 566                return;
 567
 568        for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
 569                struct kmem_cache_node *n;
 570                struct kmem_cache *cache = kmalloc_caches[i];
 571
 572                if (!cache)
 573                        continue;
 574
 575                n = cache->node[q];
 576                if (!n || OFF_SLAB(cache))
 577                        continue;
 578
 579                slab_set_lock_classes(cache, &on_slab_l3_key,
 580                                &on_slab_alc_key, q);
 581        }
 582}
 583
 584static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
 585{
 586        if (!cachep->node[q])
 587                return;
 588
 589        slab_set_lock_classes(cachep, &on_slab_l3_key,
 590                        &on_slab_alc_key, q);
 591}
 592
 593static inline void on_slab_lock_classes(struct kmem_cache *cachep)
 594{
 595        int node;
 596
 597        VM_BUG_ON(OFF_SLAB(cachep));
 598        for_each_node(node)
 599                on_slab_lock_classes_node(cachep, node);
 600}
 601
 602static inline void init_lock_keys(void)
 603{
 604        int node;
 605
 606        for_each_node(node)
 607                init_node_lock_keys(node);
 608}
 609#else
 610static void init_node_lock_keys(int q)
 611{
 612}
 613
 614static inline void init_lock_keys(void)
 615{
 616}
 617
 618static inline void on_slab_lock_classes(struct kmem_cache *cachep)
 619{
 620}
 621
 622static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
 623{
 624}
 625
 626static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
 627{
 628}
 629
 630static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
 631{
 632}
 633#endif
 634
 635static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
 636
 637static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
 638{
 639        return cachep->array[smp_processor_id()];
 640}
 641
 642static size_t slab_mgmt_size(size_t nr_objs, size_t align)
 643{
 644        return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
 645}
 646
 647/*
 648 * Calculate the number of objects and left-over bytes for a given buffer size.
 649 */
 650static void cache_estimate(unsigned long gfporder, size_t buffer_size,
 651                           size_t align, int flags, size_t *left_over,
 652                           unsigned int *num)
 653{
 654        int nr_objs;
 655        size_t mgmt_size;
 656        size_t slab_size = PAGE_SIZE << gfporder;
 657
 658        /*
 659         * The slab management structure can be either off the slab or
 660         * on it. For the latter case, the memory allocated for a
 661         * slab is used for:
 662         *
 663         * - The struct slab
 664         * - One kmem_bufctl_t for each object
 665         * - Padding to respect alignment of @align
 666         * - @buffer_size bytes for each object
 667         *
 668         * If the slab management structure is off the slab, then the
 669         * alignment will already be calculated into the size. Because
 670         * the slabs are all pages aligned, the objects will be at the
 671         * correct alignment when allocated.
 672         */
 673        if (flags & CFLGS_OFF_SLAB) {
 674                mgmt_size = 0;
 675                nr_objs = slab_size / buffer_size;
 676
 677                if (nr_objs > SLAB_LIMIT)
 678                        nr_objs = SLAB_LIMIT;
 679        } else {
 680                /*
 681                 * Ignore padding for the initial guess. The padding
 682                 * is at most @align-1 bytes, and @buffer_size is at
 683                 * least @align. In the worst case, this result will
 684                 * be one greater than the number of objects that fit
 685                 * into the memory allocation when taking the padding
 686                 * into account.
 687                 */
 688                nr_objs = (slab_size - sizeof(struct slab)) /
 689                          (buffer_size + sizeof(kmem_bufctl_t));
 690
 691                /*
 692                 * This calculated number will be either the right
 693                 * amount, or one greater than what we want.
 694                 */
 695                if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
 696                       > slab_size)
 697                        nr_objs--;
 698
 699                if (nr_objs > SLAB_LIMIT)
 700                        nr_objs = SLAB_LIMIT;
 701
 702                mgmt_size = slab_mgmt_size(nr_objs, align);
 703        }
 704        *num = nr_objs;
 705        *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
 706}
 707
 708#if DEBUG
 709#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
 710
 711static void __slab_error(const char *function, struct kmem_cache *cachep,
 712                        char *msg)
 713{
 714        printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
 715               function, cachep->name, msg);
 716        dump_stack();
 717        add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 718}
 719#endif
 720
 721/*
 722 * By default on NUMA we use alien caches to stage the freeing of
 723 * objects allocated from other nodes. This causes massive memory
 724 * inefficiencies when using fake NUMA setup to split memory into a
 725 * large number of small nodes, so it can be disabled on the command
 726 * line
 727  */
 728
 729static int use_alien_caches __read_mostly = 1;
 730static int __init noaliencache_setup(char *s)
 731{
 732        use_alien_caches = 0;
 733        return 1;
 734}
 735__setup("noaliencache", noaliencache_setup);
 736
 737static int __init slab_max_order_setup(char *str)
 738{
 739        get_option(&str, &slab_max_order);
 740        slab_max_order = slab_max_order < 0 ? 0 :
 741                                min(slab_max_order, MAX_ORDER - 1);
 742        slab_max_order_set = true;
 743
 744        return 1;
 745}
 746__setup("slab_max_order=", slab_max_order_setup);
 747
 748#ifdef CONFIG_NUMA
 749/*
 750 * Special reaping functions for NUMA systems called from cache_reap().
 751 * These take care of doing round robin flushing of alien caches (containing
 752 * objects freed on different nodes from which they were allocated) and the
 753 * flushing of remote pcps by calling drain_node_pages.
 754 */
 755static DEFINE_PER_CPU(unsigned long, slab_reap_node);
 756
 757static void init_reap_node(int cpu)
 758{
 759        int node;
 760
 761        node = next_node(cpu_to_mem(cpu), node_online_map);
 762        if (node == MAX_NUMNODES)
 763                node = first_node(node_online_map);
 764
 765        per_cpu(slab_reap_node, cpu) = node;
 766}
 767
 768static void next_reap_node(void)
 769{
 770        int node = __this_cpu_read(slab_reap_node);
 771
 772        node = next_node(node, node_online_map);
 773        if (unlikely(node >= MAX_NUMNODES))
 774                node = first_node(node_online_map);
 775        __this_cpu_write(slab_reap_node, node);
 776}
 777
 778#else
 779#define init_reap_node(cpu) do { } while (0)
 780#define next_reap_node(void) do { } while (0)
 781#endif
 782
 783/*
 784 * Initiate the reap timer running on the target CPU.  We run at around 1 to 2Hz
 785 * via the workqueue/eventd.
 786 * Add the CPU number into the expiration time to minimize the possibility of
 787 * the CPUs getting into lockstep and contending for the global cache chain
 788 * lock.
 789 */
 790static void __cpuinit start_cpu_timer(int cpu)
 791{
 792        struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
 793
 794        /*
 795         * When this gets called from do_initcalls via cpucache_init(),
 796         * init_workqueues() has already run, so keventd will be setup
 797         * at that time.
 798         */
 799        if (keventd_up() && reap_work->work.func == NULL) {
 800                init_reap_node(cpu);
 801                INIT_DEFERRABLE_WORK(reap_work, cache_reap);
 802                schedule_delayed_work_on(cpu, reap_work,
 803                                        __round_jiffies_relative(HZ, cpu));
 804        }
 805}
 806
 807static struct array_cache *alloc_arraycache(int node, int entries,
 808                                            int batchcount, gfp_t gfp)
 809{
 810        int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
 811        struct array_cache *nc = NULL;
 812
 813        nc = kmalloc_node(memsize, gfp, node);
 814        /*
 815         * The array_cache structures contain pointers to free object.
 816         * However, when such objects are allocated or transferred to another
 817         * cache the pointers are not cleared and they could be counted as
 818         * valid references during a kmemleak scan. Therefore, kmemleak must
 819         * not scan such objects.
 820         */
 821        kmemleak_no_scan(nc);
 822        if (nc) {
 823                nc->avail = 0;
 824                nc->limit = entries;
 825                nc->batchcount = batchcount;
 826                nc->touched = 0;
 827                spin_lock_init(&nc->lock);
 828        }
 829        return nc;
 830}
 831
 832static inline bool is_slab_pfmemalloc(struct slab *slabp)
 833{
 834        struct page *page = virt_to_page(slabp->s_mem);
 835
 836        return PageSlabPfmemalloc(page);
 837}
 838
 839/* Clears pfmemalloc_active if no slabs have pfmalloc set */
 840static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
 841                                                struct array_cache *ac)
 842{
 843        struct kmem_cache_node *n = cachep->node[numa_mem_id()];
 844        struct slab *slabp;
 845        unsigned long flags;
 846
 847        if (!pfmemalloc_active)
 848                return;
 849
 850        spin_lock_irqsave(&n->list_lock, flags);
 851        list_for_each_entry(slabp, &n->slabs_full, list)
 852                if (is_slab_pfmemalloc(slabp))
 853                        goto out;
 854
 855        list_for_each_entry(slabp, &n->slabs_partial, list)
 856                if (is_slab_pfmemalloc(slabp))
 857                        goto out;
 858
 859        list_for_each_entry(slabp, &n->slabs_free, list)
 860                if (is_slab_pfmemalloc(slabp))
 861                        goto out;
 862
 863        pfmemalloc_active = false;
 864out:
 865        spin_unlock_irqrestore(&n->list_lock, flags);
 866}
 867
 868static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
 869                                                gfp_t flags, bool force_refill)
 870{
 871        int i;
 872        void *objp = ac->entry[--ac->avail];
 873
 874        /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
 875        if (unlikely(is_obj_pfmemalloc(objp))) {
 876                struct kmem_cache_node *n;
 877
 878                if (gfp_pfmemalloc_allowed(flags)) {
 879                        clear_obj_pfmemalloc(&objp);
 880                        return objp;
 881                }
 882
 883                /* The caller cannot use PFMEMALLOC objects, find another one */
 884                for (i = 0; i < ac->avail; i++) {
 885                        /* If a !PFMEMALLOC object is found, swap them */
 886                        if (!is_obj_pfmemalloc(ac->entry[i])) {
 887                                objp = ac->entry[i];
 888                                ac->entry[i] = ac->entry[ac->avail];
 889                                ac->entry[ac->avail] = objp;
 890                                return objp;
 891                        }
 892                }
 893
 894                /*
 895                 * If there are empty slabs on the slabs_free list and we are
 896                 * being forced to refill the cache, mark this one !pfmemalloc.
 897                 */
 898                n = cachep->node[numa_mem_id()];
 899                if (!list_empty(&n->slabs_free) && force_refill) {
 900                        struct slab *slabp = virt_to_slab(objp);
 901                        ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
 902                        clear_obj_pfmemalloc(&objp);
 903                        recheck_pfmemalloc_active(cachep, ac);
 904                        return objp;
 905                }
 906
 907                /* No !PFMEMALLOC objects available */
 908                ac->avail++;
 909                objp = NULL;
 910        }
 911
 912        return objp;
 913}
 914
 915static inline void *ac_get_obj(struct kmem_cache *cachep,
 916                        struct array_cache *ac, gfp_t flags, bool force_refill)
 917{
 918        void *objp;
 919
 920        if (unlikely(sk_memalloc_socks()))
 921                objp = __ac_get_obj(cachep, ac, flags, force_refill);
 922        else
 923                objp = ac->entry[--ac->avail];
 924
 925        return objp;
 926}
 927
 928static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
 929                                                                void *objp)
 930{
 931        if (unlikely(pfmemalloc_active)) {
 932                /* Some pfmemalloc slabs exist, check if this is one */
 933                struct page *page = virt_to_head_page(objp);
 934                if (PageSlabPfmemalloc(page))
 935                        set_obj_pfmemalloc(&objp);
 936        }
 937
 938        return objp;
 939}
 940
 941static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
 942                                                                void *objp)
 943{
 944        if (unlikely(sk_memalloc_socks()))
 945                objp = __ac_put_obj(cachep, ac, objp);
 946
 947        ac->entry[ac->avail++] = objp;
 948}
 949
 950/*
 951 * Transfer objects in one arraycache to another.
 952 * Locking must be handled by the caller.
 953 *
 954 * Return the number of entries transferred.
 955 */
 956static int transfer_objects(struct array_cache *to,
 957                struct array_cache *from, unsigned int max)
 958{
 959        /* Figure out how many entries to transfer */
 960        int nr = min3(from->avail, max, to->limit - to->avail);
 961
 962        if (!nr)
 963                return 0;
 964
 965        memcpy(to->entry + to->avail, from->entry + from->avail -nr,
 966                        sizeof(void *) *nr);
 967
 968        from->avail -= nr;
 969        to->avail += nr;
 970        return nr;
 971}
 972
 973#ifndef CONFIG_NUMA
 974
 975#define drain_alien_cache(cachep, alien) do { } while (0)
 976#define reap_alien(cachep, n) do { } while (0)
 977
 978static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
 979{
 980        return (struct array_cache **)BAD_ALIEN_MAGIC;
 981}
 982
 983static inline void free_alien_cache(struct array_cache **ac_ptr)
 984{
 985}
 986
 987static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
 988{
 989        return 0;
 990}
 991
 992static inline void *alternate_node_alloc(struct kmem_cache *cachep,
 993                gfp_t flags)
 994{
 995        return NULL;
 996}
 997
 998static inline void *____cache_alloc_node(struct kmem_cache *cachep,
 999                 gfp_t flags, int nodeid)
1000{
1001        return NULL;
1002}
1003
1004#else   /* CONFIG_NUMA */
1005
1006static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1007static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1008
1009static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1010{
1011        struct array_cache **ac_ptr;
1012        int memsize = sizeof(void *) * nr_node_ids;
1013        int i;
1014
1015        if (limit > 1)
1016                limit = 12;
1017        ac_ptr = kzalloc_node(memsize, gfp, node);
1018        if (ac_ptr) {
1019                for_each_node(i) {
1020                        if (i == node || !node_online(i))
1021                                continue;
1022                        ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1023                        if (!ac_ptr[i]) {
1024                                for (i--; i >= 0; i--)
1025                                        kfree(ac_ptr[i]);
1026                                kfree(ac_ptr);
1027                                return NULL;
1028                        }
1029                }
1030        }
1031        return ac_ptr;
1032}
1033
1034static void free_alien_cache(struct array_cache **ac_ptr)
1035{
1036        int i;
1037
1038        if (!ac_ptr)
1039                return;
1040        for_each_node(i)
1041            kfree(ac_ptr[i]);
1042        kfree(ac_ptr);
1043}
1044
1045static void __drain_alien_cache(struct kmem_cache *cachep,
1046                                struct array_cache *ac, int node)
1047{
1048        struct kmem_cache_node *n = cachep->node[node];
1049
1050        if (ac->avail) {
1051                spin_lock(&n->list_lock);
1052                /*
1053                 * Stuff objects into the remote nodes shared array first.
1054                 * That way we could avoid the overhead of putting the objects
1055                 * into the free lists and getting them back later.
1056                 */
1057                if (n->shared)
1058                        transfer_objects(n->shared, ac, ac->limit);
1059
1060                free_block(cachep, ac->entry, ac->avail, node);
1061                ac->avail = 0;
1062                spin_unlock(&n->list_lock);
1063        }
1064}
1065
1066/*
1067 * Called from cache_reap() to regularly drain alien caches round robin.
1068 */
1069static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
1070{
1071        int node = __this_cpu_read(slab_reap_node);
1072
1073        if (n->alien) {
1074                struct array_cache *ac = n->alien[node];
1075
1076                if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1077                        __drain_alien_cache(cachep, ac, node);
1078                        spin_unlock_irq(&ac->lock);
1079                }
1080        }
1081}
1082
1083static void drain_alien_cache(struct kmem_cache *cachep,
1084                                struct array_cache **alien)
1085{
1086        int i = 0;
1087        struct array_cache *ac;
1088        unsigned long flags;
1089
1090        for_each_online_node(i) {
1091                ac = alien[i];
1092                if (ac) {
1093                        spin_lock_irqsave(&ac->lock, flags);
1094                        __drain_alien_cache(cachep, ac, i);
1095                        spin_unlock_irqrestore(&ac->lock, flags);
1096                }
1097        }
1098}
1099
1100static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1101{
1102        struct slab *slabp = virt_to_slab(objp);
1103        int nodeid = slabp->nodeid;
1104        struct kmem_cache_node *n;
1105        struct array_cache *alien = NULL;
1106        int node;
1107
1108        node = numa_mem_id();
1109
1110        /*
1111         * Make sure we are not freeing a object from another node to the array
1112         * cache on this cpu.
1113         */
1114        if (likely(slabp->nodeid == node))
1115                return 0;
1116
1117        n = cachep->node[node];
1118        STATS_INC_NODEFREES(cachep);
1119        if (n->alien && n->alien[nodeid]) {
1120                alien = n->alien[nodeid];
1121                spin_lock(&alien->lock);
1122                if (unlikely(alien->avail == alien->limit)) {
1123                        STATS_INC_ACOVERFLOW(cachep);
1124                        __drain_alien_cache(cachep, alien, nodeid);
1125                }
1126                ac_put_obj(cachep, alien, objp);
1127                spin_unlock(&alien->lock);
1128        } else {
1129                spin_lock(&(cachep->node[nodeid])->list_lock);
1130                free_block(cachep, &objp, 1, nodeid);
1131                spin_unlock(&(cachep->node[nodeid])->list_lock);
1132        }
1133        return 1;
1134}
1135#endif
1136
1137/*
1138 * Allocates and initializes node for a node on each slab cache, used for
1139 * either memory or cpu hotplug.  If memory is being hot-added, the kmem_cache_node
1140 * will be allocated off-node since memory is not yet online for the new node.
1141 * When hotplugging memory or a cpu, existing node are not replaced if
1142 * already in use.
1143 *
1144 * Must hold slab_mutex.
1145 */
1146static int init_cache_node_node(int node)
1147{
1148        struct kmem_cache *cachep;
1149        struct kmem_cache_node *n;
1150        const int memsize = sizeof(struct kmem_cache_node);
1151
1152        list_for_each_entry(cachep, &slab_caches, list) {
1153                /*
1154                 * Set up the size64 kmemlist for cpu before we can
1155                 * begin anything. Make sure some other cpu on this
1156                 * node has not already allocated this
1157                 */
1158                if (!cachep->node[node]) {
1159                        n = kmalloc_node(memsize, GFP_KERNEL, node);
1160                        if (!n)
1161                                return -ENOMEM;
1162                        kmem_cache_node_init(n);
1163                        n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1164                            ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1165
1166                        /*
1167                         * The l3s don't come and go as CPUs come and
1168                         * go.  slab_mutex is sufficient
1169                         * protection here.
1170                         */
1171                        cachep->node[node] = n;
1172                }
1173
1174                spin_lock_irq(&cachep->node[node]->list_lock);
1175                cachep->node[node]->free_limit =
1176                        (1 + nr_cpus_node(node)) *
1177                        cachep->batchcount + cachep->num;
1178                spin_unlock_irq(&cachep->node[node]->list_lock);
1179        }
1180        return 0;
1181}
1182
1183static void __cpuinit cpuup_canceled(long cpu)
1184{
1185        struct kmem_cache *cachep;
1186        struct kmem_cache_node *n = NULL;
1187        int node = cpu_to_mem(cpu);
1188        const struct cpumask *mask = cpumask_of_node(node);
1189
1190        list_for_each_entry(cachep, &slab_caches, list) {
1191                struct array_cache *nc;
1192                struct array_cache *shared;
1193                struct array_cache **alien;
1194
1195                /* cpu is dead; no one can alloc from it. */
1196                nc = cachep->array[cpu];
1197                cachep->array[cpu] = NULL;
1198                n = cachep->node[node];
1199
1200                if (!n)
1201                        goto free_array_cache;
1202
1203                spin_lock_irq(&n->list_lock);
1204
1205                /* Free limit for this kmem_cache_node */
1206                n->free_limit -= cachep->batchcount;
1207                if (nc)
1208                        free_block(cachep, nc->entry, nc->avail, node);
1209
1210                if (!cpumask_empty(mask)) {
1211                        spin_unlock_irq(&n->list_lock);
1212                        goto free_array_cache;
1213                }
1214
1215                shared = n->shared;
1216                if (shared) {
1217                        free_block(cachep, shared->entry,
1218                                   shared->avail, node);
1219                        n->shared = NULL;
1220                }
1221
1222                alien = n->alien;
1223                n->alien = NULL;
1224
1225                spin_unlock_irq(&n->list_lock);
1226
1227                kfree(shared);
1228                if (alien) {
1229                        drain_alien_cache(cachep, alien);
1230                        free_alien_cache(alien);
1231                }
1232free_array_cache:
1233                kfree(nc);
1234        }
1235        /*
1236         * In the previous loop, all the objects were freed to
1237         * the respective cache's slabs,  now we can go ahead and
1238         * shrink each nodelist to its limit.
1239         */
1240        list_for_each_entry(cachep, &slab_caches, list) {
1241                n = cachep->node[node];
1242                if (!n)
1243                        continue;
1244                drain_freelist(cachep, n, n->free_objects);
1245        }
1246}
1247
1248static int __cpuinit cpuup_prepare(long cpu)
1249{
1250        struct kmem_cache *cachep;
1251        struct kmem_cache_node *n = NULL;
1252        int node = cpu_to_mem(cpu);
1253        int err;
1254
1255        /*
1256         * We need to do this right in the beginning since
1257         * alloc_arraycache's are going to use this list.
1258         * kmalloc_node allows us to add the slab to the right
1259         * kmem_cache_node and not this cpu's kmem_cache_node
1260         */
1261        err = init_cache_node_node(node);
1262        if (err < 0)
1263                goto bad;
1264
1265        /*
1266         * Now we can go ahead with allocating the shared arrays and
1267         * array caches
1268         */
1269        list_for_each_entry(cachep, &slab_caches, list) {
1270                struct array_cache *nc;
1271                struct array_cache *shared = NULL;
1272                struct array_cache **alien = NULL;
1273
1274                nc = alloc_arraycache(node, cachep->limit,
1275                                        cachep->batchcount, GFP_KERNEL);
1276                if (!nc)
1277                        goto bad;
1278                if (cachep->shared) {
1279                        shared = alloc_arraycache(node,
1280                                cachep->shared * cachep->batchcount,
1281                                0xbaadf00d, GFP_KERNEL);
1282                        if (!shared) {
1283                                kfree(nc);
1284                                goto bad;
1285                        }
1286                }
1287                if (use_alien_caches) {
1288                        alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1289                        if (!alien) {
1290                                kfree(shared);
1291                                kfree(nc);
1292                                goto bad;
1293                        }
1294                }
1295                cachep->array[cpu] = nc;
1296                n = cachep->node[node];
1297                BUG_ON(!n);
1298
1299                spin_lock_irq(&n->list_lock);
1300                if (!n->shared) {
1301                        /*
1302                         * We are serialised from CPU_DEAD or
1303                         * CPU_UP_CANCELLED by the cpucontrol lock
1304                         */
1305                        n->shared = shared;
1306                        shared = NULL;
1307                }
1308#ifdef CONFIG_NUMA
1309                if (!n->alien) {
1310                        n->alien = alien;
1311                        alien = NULL;
1312                }
1313#endif
1314                spin_unlock_irq(&n->list_lock);
1315                kfree(shared);
1316                free_alien_cache(alien);
1317                if (cachep->flags & SLAB_DEBUG_OBJECTS)
1318                        slab_set_debugobj_lock_classes_node(cachep, node);
1319                else if (!OFF_SLAB(cachep) &&
1320                         !(cachep->flags & SLAB_DESTROY_BY_RCU))
1321                        on_slab_lock_classes_node(cachep, node);
1322        }
1323        init_node_lock_keys(node);
1324
1325        return 0;
1326bad:
1327        cpuup_canceled(cpu);
1328        return -ENOMEM;
1329}
1330
1331static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1332                                    unsigned long action, void *hcpu)
1333{
1334        long cpu = (long)hcpu;
1335        int err = 0;
1336
1337        switch (action) {
1338        case CPU_UP_PREPARE:
1339        case CPU_UP_PREPARE_FROZEN:
1340                mutex_lock(&slab_mutex);
1341                err = cpuup_prepare(cpu);
1342                mutex_unlock(&slab_mutex);
1343                break;
1344        case CPU_ONLINE:
1345        case CPU_ONLINE_FROZEN:
1346                start_cpu_timer(cpu);
1347                break;
1348#ifdef CONFIG_HOTPLUG_CPU
1349        case CPU_DOWN_PREPARE:
1350        case CPU_DOWN_PREPARE_FROZEN:
1351                /*
1352                 * Shutdown cache reaper. Note that the slab_mutex is
1353                 * held so that if cache_reap() is invoked it cannot do
1354                 * anything expensive but will only modify reap_work
1355                 * and reschedule the timer.
1356                */
1357                cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1358                /* Now the cache_reaper is guaranteed to be not running. */
1359                per_cpu(slab_reap_work, cpu).work.func = NULL;
1360                break;
1361        case CPU_DOWN_FAILED:
1362        case CPU_DOWN_FAILED_FROZEN:
1363                start_cpu_timer(cpu);
1364                break;
1365        case CPU_DEAD:
1366        case CPU_DEAD_FROZEN:
1367                /*
1368                 * Even if all the cpus of a node are down, we don't free the
1369                 * kmem_cache_node of any cache. This to avoid a race between
1370                 * cpu_down, and a kmalloc allocation from another cpu for
1371                 * memory from the node of the cpu going down.  The node
1372                 * structure is usually allocated from kmem_cache_create() and
1373                 * gets destroyed at kmem_cache_destroy().
1374                 */
1375                /* fall through */
1376#endif
1377        case CPU_UP_CANCELED:
1378        case CPU_UP_CANCELED_FROZEN:
1379                mutex_lock(&slab_mutex);
1380                cpuup_canceled(cpu);
1381                mutex_unlock(&slab_mutex);
1382                break;
1383        }
1384        return notifier_from_errno(err);
1385}
1386
1387static struct notifier_block __cpuinitdata cpucache_notifier = {
1388        &cpuup_callback, NULL, 0
1389};
1390
1391#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1392/*
1393 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1394 * Returns -EBUSY if all objects cannot be drained so that the node is not
1395 * removed.
1396 *
1397 * Must hold slab_mutex.
1398 */
1399static int __meminit drain_cache_node_node(int node)
1400{
1401        struct kmem_cache *cachep;
1402        int ret = 0;
1403
1404        list_for_each_entry(cachep, &slab_caches, list) {
1405                struct kmem_cache_node *n;
1406
1407                n = cachep->node[node];
1408                if (!n)
1409                        continue;
1410
1411                drain_freelist(cachep, n, n->free_objects);
1412
1413                if (!list_empty(&n->slabs_full) ||
1414                    !list_empty(&n->slabs_partial)) {
1415                        ret = -EBUSY;
1416                        break;
1417                }
1418        }
1419        return ret;
1420}
1421
1422static int __meminit slab_memory_callback(struct notifier_block *self,
1423                                        unsigned long action, void *arg)
1424{
1425        struct memory_notify *mnb = arg;
1426        int ret = 0;
1427        int nid;
1428
1429        nid = mnb->status_change_nid;
1430        if (nid < 0)
1431                goto out;
1432
1433        switch (action) {
1434        case MEM_GOING_ONLINE:
1435                mutex_lock(&slab_mutex);
1436                ret = init_cache_node_node(nid);
1437                mutex_unlock(&slab_mutex);
1438                break;
1439        case MEM_GOING_OFFLINE:
1440                mutex_lock(&slab_mutex);
1441                ret = drain_cache_node_node(nid);
1442                mutex_unlock(&slab_mutex);
1443                break;
1444        case MEM_ONLINE:
1445        case MEM_OFFLINE:
1446        case MEM_CANCEL_ONLINE:
1447        case MEM_CANCEL_OFFLINE:
1448                break;
1449        }
1450out:
1451        return notifier_from_errno(ret);
1452}
1453#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1454
1455/*
1456 * swap the static kmem_cache_node with kmalloced memory
1457 */
1458static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1459                                int nodeid)
1460{
1461        struct kmem_cache_node *ptr;
1462
1463        ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1464        BUG_ON(!ptr);
1465
1466        memcpy(ptr, list, sizeof(struct kmem_cache_node));
1467        /*
1468         * Do not assume that spinlocks can be initialized via memcpy:
1469         */
1470        spin_lock_init(&ptr->list_lock);
1471
1472        MAKE_ALL_LISTS(cachep, ptr, nodeid);
1473        cachep->node[nodeid] = ptr;
1474}
1475
1476/*
1477 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1478 * size of kmem_cache_node.
1479 */
1480static void __init set_up_node(struct kmem_cache *cachep, int index)
1481{
1482        int node;
1483
1484        for_each_online_node(node) {
1485                cachep->node[node] = &init_kmem_cache_node[index + node];
1486                cachep->node[node]->next_reap = jiffies +
1487                    REAPTIMEOUT_LIST3 +
1488                    ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1489        }
1490}
1491
1492/*
1493 * The memory after the last cpu cache pointer is used for the
1494 * the node pointer.
1495 */
1496static void setup_node_pointer(struct kmem_cache *cachep)
1497{
1498        cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1499}
1500
1501/*
1502 * Initialisation.  Called after the page allocator have been initialised and
1503 * before smp_init().
1504 */
1505void __init kmem_cache_init(void)
1506{
1507        int i;
1508
1509        kmem_cache = &kmem_cache_boot;
1510        setup_node_pointer(kmem_cache);
1511
1512        if (num_possible_nodes() == 1)
1513                use_alien_caches = 0;
1514
1515        for (i = 0; i < NUM_INIT_LISTS; i++)
1516                kmem_cache_node_init(&init_kmem_cache_node[i]);
1517
1518        set_up_node(kmem_cache, CACHE_CACHE);
1519
1520        /*
1521         * Fragmentation resistance on low memory - only use bigger
1522         * page orders on machines with more than 32MB of memory if
1523         * not overridden on the command line.
1524         */
1525        if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1526                slab_max_order = SLAB_MAX_ORDER_HI;
1527
1528        /* Bootstrap is tricky, because several objects are allocated
1529         * from caches that do not exist yet:
1530         * 1) initialize the kmem_cache cache: it contains the struct
1531         *    kmem_cache structures of all caches, except kmem_cache itself:
1532         *    kmem_cache is statically allocated.
1533         *    Initially an __init data area is used for the head array and the
1534         *    kmem_cache_node structures, it's replaced with a kmalloc allocated
1535         *    array at the end of the bootstrap.
1536         * 2) Create the first kmalloc cache.
1537         *    The struct kmem_cache for the new cache is allocated normally.
1538         *    An __init data area is used for the head array.
1539         * 3) Create the remaining kmalloc caches, with minimally sized
1540         *    head arrays.
1541         * 4) Replace the __init data head arrays for kmem_cache and the first
1542         *    kmalloc cache with kmalloc allocated arrays.
1543         * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1544         *    the other cache's with kmalloc allocated memory.
1545         * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1546         */
1547
1548        /* 1) create the kmem_cache */
1549
1550        /*
1551         * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1552         */
1553        create_boot_cache(kmem_cache, "kmem_cache",
1554                offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1555                                  nr_node_ids * sizeof(struct kmem_cache_node *),
1556                                  SLAB_HWCACHE_ALIGN);
1557        list_add(&kmem_cache->list, &slab_caches);
1558
1559        /* 2+3) create the kmalloc caches */
1560
1561        /*
1562         * Initialize the caches that provide memory for the array cache and the
1563         * kmem_cache_node structures first.  Without this, further allocations will
1564         * bug.
1565         */
1566
1567        kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1568                                        kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1569
1570        if (INDEX_AC != INDEX_NODE)
1571                kmalloc_caches[INDEX_NODE] =
1572                        create_kmalloc_cache("kmalloc-node",
1573                                kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1574
1575        slab_early_init = 0;
1576
1577        /* 4) Replace the bootstrap head arrays */
1578        {
1579                struct array_cache *ptr;
1580
1581                ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1582
1583                memcpy(ptr, cpu_cache_get(kmem_cache),
1584                       sizeof(struct arraycache_init));
1585                /*
1586                 * Do not assume that spinlocks can be initialized via memcpy:
1587                 */
1588                spin_lock_init(&ptr->lock);
1589
1590                kmem_cache->array[smp_processor_id()] = ptr;
1591
1592                ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1593
1594                BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1595                       != &initarray_generic.cache);
1596                memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1597                       sizeof(struct arraycache_init));
1598                /*
1599                 * Do not assume that spinlocks can be initialized via memcpy:
1600                 */
1601                spin_lock_init(&ptr->lock);
1602
1603                kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1604        }
1605        /* 5) Replace the bootstrap kmem_cache_node */
1606        {
1607                int nid;
1608
1609                for_each_online_node(nid) {
1610                        init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1611
1612                        init_list(kmalloc_caches[INDEX_AC],
1613                                  &init_kmem_cache_node[SIZE_AC + nid], nid);
1614
1615                        if (INDEX_AC != INDEX_NODE) {
1616                                init_list(kmalloc_caches[INDEX_NODE],
1617                                          &init_kmem_cache_node[SIZE_NODE + nid], nid);
1618                        }
1619                }
1620        }
1621
1622        create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1623}
1624
1625void __init kmem_cache_init_late(void)
1626{
1627        struct kmem_cache *cachep;
1628
1629        slab_state = UP;
1630
1631        /* 6) resize the head arrays to their final sizes */
1632        mutex_lock(&slab_mutex);
1633        list_for_each_entry(cachep, &slab_caches, list)
1634                if (enable_cpucache(cachep, GFP_NOWAIT))
1635                        BUG();
1636        mutex_unlock(&slab_mutex);
1637
1638        /* Annotate slab for lockdep -- annotate the malloc caches */
1639        init_lock_keys();
1640
1641        /* Done! */
1642        slab_state = FULL;
1643
1644        /*
1645         * Register a cpu startup notifier callback that initializes
1646         * cpu_cache_get for all new cpus
1647         */
1648        register_cpu_notifier(&cpucache_notifier);
1649
1650#ifdef CONFIG_NUMA
1651        /*
1652         * Register a memory hotplug callback that initializes and frees
1653         * node.
1654         */
1655        hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1656#endif
1657
1658        /*
1659         * The reap timers are started later, with a module init call: That part
1660         * of the kernel is not yet operational.
1661         */
1662}
1663
1664static int __init cpucache_init(void)
1665{
1666        int cpu;
1667
1668        /*
1669         * Register the timers that return unneeded pages to the page allocator
1670         */
1671        for_each_online_cpu(cpu)
1672                start_cpu_timer(cpu);
1673
1674        /* Done! */
1675        slab_state = FULL;
1676        return 0;
1677}
1678__initcall(cpucache_init);
1679
1680static noinline void
1681slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1682{
1683        struct kmem_cache_node *n;
1684        struct slab *slabp;
1685        unsigned long flags;
1686        int node;
1687
1688        printk(KERN_WARNING
1689                "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1690                nodeid, gfpflags);
1691        printk(KERN_WARNING "  cache: %s, object size: %d, order: %d\n",
1692                cachep->name, cachep->size, cachep->gfporder);
1693
1694        for_each_online_node(node) {
1695                unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1696                unsigned long active_slabs = 0, num_slabs = 0;
1697
1698                n = cachep->node[node];
1699                if (!n)
1700                        continue;
1701
1702                spin_lock_irqsave(&n->list_lock, flags);
1703                list_for_each_entry(slabp, &n->slabs_full, list) {
1704                        active_objs += cachep->num;
1705                        active_slabs++;
1706                }
1707                list_for_each_entry(slabp, &n->slabs_partial, list) {
1708                        active_objs += slabp->inuse;
1709                        active_slabs++;
1710                }
1711                list_for_each_entry(slabp, &n->slabs_free, list)
1712                        num_slabs++;
1713
1714                free_objects += n->free_objects;
1715                spin_unlock_irqrestore(&n->list_lock, flags);
1716
1717                num_slabs += active_slabs;
1718                num_objs = num_slabs * cachep->num;
1719                printk(KERN_WARNING
1720                        "  node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1721                        node, active_slabs, num_slabs, active_objs, num_objs,
1722                        free_objects);
1723        }
1724}
1725
1726/*
1727 * Interface to system's page allocator. No need to hold the cache-lock.
1728 *
1729 * If we requested dmaable memory, we will get it. Even if we
1730 * did not request dmaable memory, we might get it, but that
1731 * would be relatively rare and ignorable.
1732 */
1733static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1734{
1735        struct page *page;
1736        int nr_pages;
1737        int i;
1738
1739#ifndef CONFIG_MMU
1740        /*
1741         * Nommu uses slab's for process anonymous memory allocations, and thus
1742         * requires __GFP_COMP to properly refcount higher order allocations
1743         */
1744        flags |= __GFP_COMP;
1745#endif
1746
1747        flags |= cachep->allocflags;
1748        if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1749                flags |= __GFP_RECLAIMABLE;
1750
1751        page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1752        if (!page) {
1753                if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1754                        slab_out_of_memory(cachep, flags, nodeid);
1755                return NULL;
1756        }
1757
1758        /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1759        if (unlikely(page->pfmemalloc))
1760                pfmemalloc_active = true;
1761
1762        nr_pages = (1 << cachep->gfporder);
1763        if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1764                add_zone_page_state(page_zone(page),
1765                        NR_SLAB_RECLAIMABLE, nr_pages);
1766        else
1767                add_zone_page_state(page_zone(page),
1768                        NR_SLAB_UNRECLAIMABLE, nr_pages);
1769        for (i = 0; i < nr_pages; i++) {
1770                __SetPageSlab(page + i);
1771
1772                if (page->pfmemalloc)
1773                        SetPageSlabPfmemalloc(page + i);
1774        }
1775        memcg_bind_pages(cachep, cachep->gfporder);
1776
1777        if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1778                kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1779
1780                if (cachep->ctor)
1781                        kmemcheck_mark_uninitialized_pages(page, nr_pages);
1782                else
1783                        kmemcheck_mark_unallocated_pages(page, nr_pages);
1784        }
1785
1786        return page_address(page);
1787}
1788
1789/*
1790 * Interface to system's page release.
1791 */
1792static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1793{
1794        unsigned long i = (1 << cachep->gfporder);
1795        struct page *page = virt_to_page(addr);
1796        const unsigned long nr_freed = i;
1797
1798        kmemcheck_free_shadow(page, cachep->gfporder);
1799
1800        if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1801                sub_zone_page_state(page_zone(page),
1802                                NR_SLAB_RECLAIMABLE, nr_freed);
1803        else
1804                sub_zone_page_state(page_zone(page),
1805                                NR_SLAB_UNRECLAIMABLE, nr_freed);
1806        while (i--) {
1807                BUG_ON(!PageSlab(page));
1808                __ClearPageSlabPfmemalloc(page);
1809                __ClearPageSlab(page);
1810                page++;
1811        }
1812
1813        memcg_release_pages(cachep, cachep->gfporder);
1814        if (current->reclaim_state)
1815                current->reclaim_state->reclaimed_slab += nr_freed;
1816        free_memcg_kmem_pages((unsigned long)addr, cachep->gfporder);
1817}
1818
1819static void kmem_rcu_free(struct rcu_head *head)
1820{
1821        struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1822        struct kmem_cache *cachep = slab_rcu->cachep;
1823
1824        kmem_freepages(cachep, slab_rcu->addr);
1825        if (OFF_SLAB(cachep))
1826                kmem_cache_free(cachep->slabp_cache, slab_rcu);
1827}
1828
1829#if DEBUG
1830
1831#ifdef CONFIG_DEBUG_PAGEALLOC
1832static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1833                            unsigned long caller)
1834{
1835        int size = cachep->object_size;
1836
1837        addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1838
1839        if (size < 5 * sizeof(unsigned long))
1840                return;
1841
1842        *addr++ = 0x12345678;
1843        *addr++ = caller;
1844        *addr++ = smp_processor_id();
1845        size -= 3 * sizeof(unsigned long);
1846        {
1847                unsigned long *sptr = &caller;
1848                unsigned long svalue;
1849
1850                while (!kstack_end(sptr)) {
1851                        svalue = *sptr++;
1852                        if (kernel_text_address(svalue)) {
1853                                *addr++ = svalue;
1854                                size -= sizeof(unsigned long);
1855                                if (size <= sizeof(unsigned long))
1856                                        break;
1857                        }
1858                }
1859
1860        }
1861        *addr++ = 0x87654321;
1862}
1863#endif
1864
1865static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1866{
1867        int size = cachep->object_size;
1868        addr = &((char *)addr)[obj_offset(cachep)];
1869
1870        memset(addr, val, size);
1871        *(unsigned char *)(addr + size - 1) = POISON_END;
1872}
1873
1874static void dump_line(char *data, int offset, int limit)
1875{
1876        int i;
1877        unsigned char error = 0;
1878        int bad_count = 0;
1879
1880        printk(KERN_ERR "%03x: ", offset);
1881        for (i = 0; i < limit; i++) {
1882                if (data[offset + i] != POISON_FREE) {
1883                        error = data[offset + i];
1884                        bad_count++;
1885                }
1886        }
1887        print_hex_dump(KERN_CONT, "", 0, 16, 1,
1888                        &data[offset], limit, 1);
1889
1890        if (bad_count == 1) {
1891                error ^= POISON_FREE;
1892                if (!(error & (error - 1))) {
1893                        printk(KERN_ERR "Single bit error detected. Probably "
1894                                        "bad RAM.\n");
1895#ifdef CONFIG_X86
1896                        printk(KERN_ERR "Run memtest86+ or a similar memory "
1897                                        "test tool.\n");
1898#else
1899                        printk(KERN_ERR "Run a memory test tool.\n");
1900#endif
1901                }
1902        }
1903}
1904#endif
1905
1906#if DEBUG
1907
1908static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1909{
1910        int i, size;
1911        char *realobj;
1912
1913        if (cachep->flags & SLAB_RED_ZONE) {
1914                printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1915                        *dbg_redzone1(cachep, objp),
1916                        *dbg_redzone2(cachep, objp));
1917        }
1918
1919        if (cachep->flags & SLAB_STORE_USER) {
1920                printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1921                       *dbg_userword(cachep, objp),
1922                       *dbg_userword(cachep, objp));
1923        }
1924        realobj = (char *)objp + obj_offset(cachep);
1925        size = cachep->object_size;
1926        for (i = 0; i < size && lines; i += 16, lines--) {
1927                int limit;
1928                limit = 16;
1929                if (i + limit > size)
1930                        limit = size - i;
1931                dump_line(realobj, i, limit);
1932        }
1933}
1934
1935static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1936{
1937        char *realobj;
1938        int size, i;
1939        int lines = 0;
1940
1941        realobj = (char *)objp + obj_offset(cachep);
1942        size = cachep->object_size;
1943
1944        for (i = 0; i < size; i++) {
1945                char exp = POISON_FREE;
1946                if (i == size - 1)
1947                        exp = POISON_END;
1948                if (realobj[i] != exp) {
1949                        int limit;
1950                        /* Mismatch ! */
1951                        /* Print header */
1952                        if (lines == 0) {
1953                                printk(KERN_ERR
1954                                        "Slab corruption (%s): %s start=%p, len=%d\n",
1955                                        print_tainted(), cachep->name, realobj, size);
1956                                print_objinfo(cachep, objp, 0);
1957                        }
1958                        /* Hexdump the affected line */
1959                        i = (i / 16) * 16;
1960                        limit = 16;
1961                        if (i + limit > size)
1962                                limit = size - i;
1963                        dump_line(realobj, i, limit);
1964                        i += 16;
1965                        lines++;
1966                        /* Limit to 5 lines */
1967                        if (lines > 5)
1968                                break;
1969                }
1970        }
1971        if (lines != 0) {
1972                /* Print some data about the neighboring objects, if they
1973                 * exist:
1974                 */
1975                struct slab *slabp = virt_to_slab(objp);
1976                unsigned int objnr;
1977
1978                objnr = obj_to_index(cachep, slabp, objp);
1979                if (objnr) {
1980                        objp = index_to_obj(cachep, slabp, objnr - 1);
1981                        realobj = (char *)objp + obj_offset(cachep);
1982                        printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1983                               realobj, size);
1984                        print_objinfo(cachep, objp, 2);
1985                }
1986                if (objnr + 1 < cachep->num) {
1987                        objp = index_to_obj(cachep, slabp, objnr + 1);
1988                        realobj = (char *)objp + obj_offset(cachep);
1989                        printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1990                               realobj, size);
1991                        print_objinfo(cachep, objp, 2);
1992                }
1993        }
1994}
1995#endif
1996
1997#if DEBUG
1998static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1999{
2000        int i;
2001        for (i = 0; i < cachep->num; i++) {
2002                void *objp = index_to_obj(cachep, slabp, i);
2003
2004                if (cachep->flags & SLAB_POISON) {
2005#ifdef CONFIG_DEBUG_PAGEALLOC
2006                        if (cachep->size % PAGE_SIZE == 0 &&
2007                                        OFF_SLAB(cachep))
2008                                kernel_map_pages(virt_to_page(objp),
2009                                        cachep->size / PAGE_SIZE, 1);
2010                        else
2011                                check_poison_obj(cachep, objp);
2012#else
2013                        check_poison_obj(cachep, objp);
2014#endif
2015                }
2016                if (cachep->flags & SLAB_RED_ZONE) {
2017                        if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2018                                slab_error(cachep, "start of a freed object "
2019                                           "was overwritten");
2020                        if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2021                                slab_error(cachep, "end of a freed object "
2022                                           "was overwritten");
2023                }
2024        }
2025}
2026#else
2027static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2028{
2029}
2030#endif
2031
2032/**
2033 * slab_destroy - destroy and release all objects in a slab
2034 * @cachep: cache pointer being destroyed
2035 * @slabp: slab pointer being destroyed
2036 *
2037 * Destroy all the objs in a slab, and release the mem back to the system.
2038 * Before calling the slab must have been unlinked from the cache.  The
2039 * cache-lock is not held/needed.
2040 */
2041static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2042{
2043        void *addr = slabp->s_mem - slabp->colouroff;
2044
2045        slab_destroy_debugcheck(cachep, slabp);
2046        if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2047                struct slab_rcu *slab_rcu;
2048
2049                slab_rcu = (struct slab_rcu *)slabp;
2050                slab_rcu->cachep = cachep;
2051                slab_rcu->addr = addr;
2052                call_rcu(&slab_rcu->head, kmem_rcu_free);
2053        } else {
2054                kmem_freepages(cachep, addr);
2055                if (OFF_SLAB(cachep))
2056                        kmem_cache_free(cachep->slabp_cache, slabp);
2057        }
2058}
2059
2060/**
2061 * calculate_slab_order - calculate size (page order) of slabs
2062 * @cachep: pointer to the cache that is being created
2063 * @size: size of objects to be created in this cache.
2064 * @align: required alignment for the objects.
2065 * @flags: slab allocation flags
2066 *
2067 * Also calculates the number of objects per slab.
2068 *
2069 * This could be made much more intelligent.  For now, try to avoid using
2070 * high order pages for slabs.  When the gfp() functions are more friendly
2071 * towards high-order requests, this should be changed.
2072 */
2073static size_t calculate_slab_order(struct kmem_cache *cachep,
2074                        size_t size, size_t align, unsigned long flags)
2075{
2076        unsigned long offslab_limit;
2077        size_t left_over = 0;
2078        int gfporder;
2079
2080        for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2081                unsigned int num;
2082                size_t remainder;
2083
2084                cache_estimate(gfporder, size, align, flags, &remainder, &num);
2085                if (!num)
2086                        continue;
2087
2088                if (flags & CFLGS_OFF_SLAB) {
2089                        /*
2090                         * Max number of objs-per-slab for caches which
2091                         * use off-slab slabs. Needed to avoid a possible
2092                         * looping condition in cache_grow().
2093                         */
2094                        offslab_limit = size - sizeof(struct slab);
2095                        offslab_limit /= sizeof(kmem_bufctl_t);
2096
2097                        if (num > offslab_limit)
2098                                break;
2099                }
2100
2101                /* Found something acceptable - save it away */
2102                cachep->num = num;
2103                cachep->gfporder = gfporder;
2104                left_over = remainder;
2105
2106                /*
2107                 * A VFS-reclaimable slab tends to have most allocations
2108                 * as GFP_NOFS and we really don't want to have to be allocating
2109                 * higher-order pages when we are unable to shrink dcache.
2110                 */
2111                if (flags & SLAB_RECLAIM_ACCOUNT)
2112                        break;
2113
2114                /*
2115                 * Large number of objects is good, but very large slabs are
2116                 * currently bad for the gfp()s.
2117                 */
2118                if (gfporder >= slab_max_order)
2119                        break;
2120
2121                /*
2122                 * Acceptable internal fragmentation?
2123                 */
2124                if (left_over * 8 <= (PAGE_SIZE << gfporder))
2125                        break;
2126        }
2127        return left_over;
2128}
2129
2130static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2131{
2132        if (slab_state >= FULL)
2133                return enable_cpucache(cachep, gfp);
2134
2135        if (slab_state == DOWN) {
2136                /*
2137                 * Note: Creation of first cache (kmem_cache).
2138                 * The setup_node is taken care
2139                 * of by the caller of __kmem_cache_create
2140                 */
2141                cachep->array[smp_processor_id()] = &initarray_generic.cache;
2142                slab_state = PARTIAL;
2143        } else if (slab_state == PARTIAL) {
2144                /*
2145                 * Note: the second kmem_cache_create must create the cache
2146                 * that's used by kmalloc(24), otherwise the creation of
2147                 * further caches will BUG().
2148                 */
2149                cachep->array[smp_processor_id()] = &initarray_generic.cache;
2150
2151                /*
2152                 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2153                 * the second cache, then we need to set up all its node/,
2154                 * otherwise the creation of further caches will BUG().
2155                 */
2156                set_up_node(cachep, SIZE_AC);
2157                if (INDEX_AC == INDEX_NODE)
2158                        slab_state = PARTIAL_NODE;
2159                else
2160                        slab_state = PARTIAL_ARRAYCACHE;
2161        } else {
2162                /* Remaining boot caches */
2163                cachep->array[smp_processor_id()] =
2164                        kmalloc(sizeof(struct arraycache_init), gfp);
2165
2166                if (slab_state == PARTIAL_ARRAYCACHE) {
2167                        set_up_node(cachep, SIZE_NODE);
2168                        slab_state = PARTIAL_NODE;
2169                } else {
2170                        int node;
2171                        for_each_online_node(node) {
2172                                cachep->node[node] =
2173                                    kmalloc_node(sizeof(struct kmem_cache_node),
2174                                                gfp, node);
2175                                BUG_ON(!cachep->node[node]);
2176                                kmem_cache_node_init(cachep->node[node]);
2177                        }
2178                }
2179        }
2180        cachep->node[numa_mem_id()]->next_reap =
2181                        jiffies + REAPTIMEOUT_LIST3 +
2182                        ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2183
2184        cpu_cache_get(cachep)->avail = 0;
2185        cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2186        cpu_cache_get(cachep)->batchcount = 1;
2187        cpu_cache_get(cachep)->touched = 0;
2188        cachep->batchcount = 1;
2189        cachep->limit = BOOT_CPUCACHE_ENTRIES;
2190        return 0;
2191}
2192
2193/**
2194 * __kmem_cache_create - Create a cache.
2195 * @cachep: cache management descriptor
2196 * @flags: SLAB flags
2197 *
2198 * Returns a ptr to the cache on success, NULL on failure.
2199 * Cannot be called within a int, but can be interrupted.
2200 * The @ctor is run when new pages are allocated by the cache.
2201 *
2202 * The flags are
2203 *
2204 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2205 * to catch references to uninitialised memory.
2206 *
2207 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2208 * for buffer overruns.
2209 *
2210 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2211 * cacheline.  This can be beneficial if you're counting cycles as closely
2212 * as davem.
2213 */
2214int
2215__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2216{
2217        size_t left_over, slab_size, ralign;
2218        gfp_t gfp;
2219        int err;
2220        size_t size = cachep->size;
2221
2222#if DEBUG
2223#if FORCED_DEBUG
2224        /*
2225         * Enable redzoning and last user accounting, except for caches with
2226         * large objects, if the increased size would increase the object size
2227         * above the next power of two: caches with object sizes just above a
2228         * power of two have a significant amount of internal fragmentation.
2229         */
2230        if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2231                                                2 * sizeof(unsigned long long)))
2232                flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2233        if (!(flags & SLAB_DESTROY_BY_RCU))
2234                flags |= SLAB_POISON;
2235#endif
2236        if (flags & SLAB_DESTROY_BY_RCU)
2237                BUG_ON(flags & SLAB_POISON);
2238#endif
2239
2240        /*
2241         * Check that size is in terms of words.  This is needed to avoid
2242         * unaligned accesses for some archs when redzoning is used, and makes
2243         * sure any on-slab bufctl's are also correctly aligned.
2244         */
2245        if (size & (BYTES_PER_WORD - 1)) {
2246                size += (BYTES_PER_WORD - 1);
2247                size &= ~(BYTES_PER_WORD - 1);
2248        }
2249
2250        /*
2251         * Redzoning and user store require word alignment or possibly larger.
2252         * Note this will be overridden by architecture or caller mandated
2253         * alignment if either is greater than BYTES_PER_WORD.
2254         */
2255        if (flags & SLAB_STORE_USER)
2256                ralign = BYTES_PER_WORD;
2257
2258        if (flags & SLAB_RED_ZONE) {
2259                ralign = REDZONE_ALIGN;
2260                /* If redzoning, ensure that the second redzone is suitably
2261                 * aligned, by adjusting the object size accordingly. */
2262                size += REDZONE_ALIGN - 1;
2263                size &= ~(REDZONE_ALIGN - 1);
2264        }
2265
2266        /* 3) caller mandated alignment */
2267        if (ralign < cachep->align) {
2268                ralign = cachep->align;
2269        }
2270        /* disable debug if necessary */
2271        if (ralign > __alignof__(unsigned long long))
2272                flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2273        /*
2274         * 4) Store it.
2275         */
2276        cachep->align = ralign;
2277
2278        if (slab_is_available())
2279                gfp = GFP_KERNEL;
2280        else
2281                gfp = GFP_NOWAIT;
2282
2283        setup_node_pointer(cachep);
2284#if DEBUG
2285
2286        /*
2287         * Both debugging options require word-alignment which is calculated
2288         * into align above.
2289         */
2290        if (flags & SLAB_RED_ZONE) {
2291                /* add space for red zone words */
2292                cachep->obj_offset += sizeof(unsigned long long);
2293                size += 2 * sizeof(unsigned long long);
2294        }
2295        if (flags & SLAB_STORE_USER) {
2296                /* user store requires one word storage behind the end of
2297                 * the real object. But if the second red zone needs to be
2298                 * aligned to 64 bits, we must allow that much space.
2299                 */
2300                if (flags & SLAB_RED_ZONE)
2301                        size += REDZONE_ALIGN;
2302                else
2303                        size += BYTES_PER_WORD;
2304        }
2305#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2306        if (size >= kmalloc_size(INDEX_NODE + 1)
2307            && cachep->object_size > cache_line_size()
2308            && ALIGN(size, cachep->align) < PAGE_SIZE) {
2309                cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2310                size = PAGE_SIZE;
2311        }
2312#endif
2313#endif
2314
2315        /*
2316         * Determine if the slab management is 'on' or 'off' slab.
2317         * (bootstrapping cannot cope with offslab caches so don't do
2318         * it too early on. Always use on-slab management when
2319         * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2320         */
2321        if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2322            !(flags & SLAB_NOLEAKTRACE))
2323                /*
2324                 * Size is large, assume best to place the slab management obj
2325                 * off-slab (should allow better packing of objs).
2326                 */
2327                flags |= CFLGS_OFF_SLAB;
2328
2329        size = ALIGN(size, cachep->align);
2330
2331        left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2332
2333        if (!cachep->num)
2334                return -E2BIG;
2335
2336        slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2337                          + sizeof(struct slab), cachep->align);
2338
2339        /*
2340         * If the slab has been placed off-slab, and we have enough space then
2341         * move it on-slab. This is at the expense of any extra colouring.
2342         */
2343        if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2344                flags &= ~CFLGS_OFF_SLAB;
2345                left_over -= slab_size;
2346        }
2347
2348        if (flags & CFLGS_OFF_SLAB) {
2349                /* really off slab. No need for manual alignment */
2350                slab_size =
2351                    cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2352
2353#ifdef CONFIG_PAGE_POISONING
2354                /* If we're going to use the generic kernel_map_pages()
2355                 * poisoning, then it's going to smash the contents of
2356                 * the redzone and userword anyhow, so switch them off.
2357                 */
2358                if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2359                        flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2360#endif
2361        }
2362
2363        cachep->colour_off = cache_line_size();
2364        /* Offset must be a multiple of the alignment. */
2365        if (cachep->colour_off < cachep->align)
2366                cachep->colour_off = cachep->align;
2367        cachep->colour = left_over / cachep->colour_off;
2368        cachep->slab_size = slab_size;
2369        cachep->flags = flags;
2370        cachep->allocflags = 0;
2371        if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2372                cachep->allocflags |= GFP_DMA;
2373        cachep->size = size;
2374        cachep->reciprocal_buffer_size = reciprocal_value(size);
2375
2376        if (flags & CFLGS_OFF_SLAB) {
2377                cachep->slabp_cache = kmalloc_slab(slab_size, 0u);
2378                /*
2379                 * This is a possibility for one of the malloc_sizes caches.
2380                 * But since we go off slab only for object size greater than
2381                 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2382                 * this should not happen at all.
2383                 * But leave a BUG_ON for some lucky dude.
2384                 */
2385                BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2386        }
2387
2388        err = setup_cpu_cache(cachep, gfp);
2389        if (err) {
2390                __kmem_cache_shutdown(cachep);
2391                return err;
2392        }
2393
2394        if (flags & SLAB_DEBUG_OBJECTS) {
2395                /*
2396                 * Would deadlock through slab_destroy()->call_rcu()->
2397                 * debug_object_activate()->kmem_cache_alloc().
2398                 */
2399                WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2400
2401                slab_set_debugobj_lock_classes(cachep);
2402        } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2403                on_slab_lock_classes(cachep);
2404
2405        return 0;
2406}
2407
2408#if DEBUG
2409static void check_irq_off(void)
2410{
2411        BUG_ON(!irqs_disabled());
2412}
2413
2414static void check_irq_on(void)
2415{
2416        BUG_ON(irqs_disabled());
2417}
2418
2419static void check_spinlock_acquired(struct kmem_cache *cachep)
2420{
2421#ifdef CONFIG_SMP
2422        check_irq_off();
2423        assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
2424#endif
2425}
2426
2427static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2428{
2429#ifdef CONFIG_SMP
2430        check_irq_off();
2431        assert_spin_locked(&cachep->node[node]->list_lock);
2432#endif
2433}
2434
2435#else
2436#define check_irq_off() do { } while(0)
2437#define check_irq_on()  do { } while(0)
2438#define check_spinlock_acquired(x) do { } while(0)
2439#define check_spinlock_acquired_node(x, y) do { } while(0)
2440#endif
2441
2442static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2443                        struct array_cache *ac,
2444                        int force, int node);
2445
2446static void do_drain(void *arg)
2447{
2448        struct kmem_cache *cachep = arg;
2449        struct array_cache *ac;
2450        int node = numa_mem_id();
2451
2452        check_irq_off();
2453        ac = cpu_cache_get(cachep);
2454        spin_lock(&cachep->node[node]->list_lock);
2455        free_block(cachep, ac->entry, ac->avail, node);
2456        spin_unlock(&cachep->node[node]->list_lock);
2457        ac->avail = 0;
2458}
2459
2460static void drain_cpu_caches(struct kmem_cache *cachep)
2461{
2462        struct kmem_cache_node *n;
2463        int node;
2464
2465        on_each_cpu(do_drain, cachep, 1);
2466        check_irq_on();
2467        for_each_online_node(node) {
2468                n = cachep->node[node];
2469                if (n && n->alien)
2470                        drain_alien_cache(cachep, n->alien);
2471        }
2472
2473        for_each_online_node(node) {
2474                n = cachep->node[node];
2475                if (n)
2476                        drain_array(cachep, n, n->shared, 1, node);
2477        }
2478}
2479
2480/*
2481 * Remove slabs from the list of free slabs.
2482 * Specify the number of slabs to drain in tofree.
2483 *
2484 * Returns the actual number of slabs released.
2485 */
2486static int drain_freelist(struct kmem_cache *cache,
2487                        struct kmem_cache_node *n, int tofree)
2488{
2489        struct list_head *p;
2490        int nr_freed;
2491        struct slab *slabp;
2492
2493        nr_freed = 0;
2494        while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2495
2496                spin_lock_irq(&n->list_lock);
2497                p = n->slabs_free.prev;
2498                if (p == &n->slabs_free) {
2499                        spin_unlock_irq(&n->list_lock);
2500                        goto out;
2501                }
2502
2503                slabp = list_entry(p, struct slab, list);
2504#if DEBUG
2505                BUG_ON(slabp->inuse);
2506#endif
2507                list_del(&slabp->list);
2508                /*
2509                 * Safe to drop the lock. The slab is no longer linked
2510                 * to the cache.
2511                 */
2512                n->free_objects -= cache->num;
2513                spin_unlock_irq(&n->list_lock);
2514                slab_destroy(cache, slabp);
2515                nr_freed++;
2516        }
2517out:
2518        return nr_freed;
2519}
2520
2521/* Called with slab_mutex held to protect against cpu hotplug */
2522static int __cache_shrink(struct kmem_cache *cachep)
2523{
2524        int ret = 0, i = 0;
2525        struct kmem_cache_node *n;
2526
2527        drain_cpu_caches(cachep);
2528
2529        check_irq_on();
2530        for_each_online_node(i) {
2531                n = cachep->node[i];
2532                if (!n)
2533                        continue;
2534
2535                drain_freelist(cachep, n, n->free_objects);
2536
2537                ret += !list_empty(&n->slabs_full) ||
2538                        !list_empty(&n->slabs_partial);
2539        }
2540        return (ret ? 1 : 0);
2541}
2542
2543/**
2544 * kmem_cache_shrink - Shrink a cache.
2545 * @cachep: The cache to shrink.
2546 *
2547 * Releases as many slabs as possible for a cache.
2548 * To help debugging, a zero exit status indicates all slabs were released.
2549 */
2550int kmem_cache_shrink(struct kmem_cache *cachep)
2551{
2552        int ret;
2553        BUG_ON(!cachep || in_interrupt());
2554
2555        get_online_cpus();
2556        mutex_lock(&slab_mutex);
2557        ret = __cache_shrink(cachep);
2558        mutex_unlock(&slab_mutex);
2559        put_online_cpus();
2560        return ret;
2561}
2562EXPORT_SYMBOL(kmem_cache_shrink);
2563
2564int __kmem_cache_shutdown(struct kmem_cache *cachep)
2565{
2566        int i;
2567        struct kmem_cache_node *n;
2568        int rc = __cache_shrink(cachep);
2569
2570        if (rc)
2571                return rc;
2572
2573        for_each_online_cpu(i)
2574            kfree(cachep->array[i]);
2575
2576        /* NUMA: free the node structures */
2577        for_each_online_node(i) {
2578                n = cachep->node[i];
2579                if (n) {
2580                        kfree(n->shared);
2581                        free_alien_cache(n->alien);
2582                        kfree(n);
2583                }
2584        }
2585        return 0;
2586}
2587
2588/*
2589 * Get the memory for a slab management obj.
2590 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2591 * always come from malloc_sizes caches.  The slab descriptor cannot
2592 * come from the same cache which is getting created because,
2593 * when we are searching for an appropriate cache for these
2594 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2595 * If we are creating a malloc_sizes cache here it would not be visible to
2596 * kmem_find_general_cachep till the initialization is complete.
2597 * Hence we cannot have slabp_cache same as the original cache.
2598 */
2599static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2600                                   int colour_off, gfp_t local_flags,
2601                                   int nodeid)
2602{
2603        struct slab *slabp;
2604
2605        if (OFF_SLAB(cachep)) {
2606                /* Slab management obj is off-slab. */
2607                slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2608                                              local_flags, nodeid);
2609                /*
2610                 * If the first object in the slab is leaked (it's allocated
2611                 * but no one has a reference to it), we want to make sure
2612                 * kmemleak does not treat the ->s_mem pointer as a reference
2613                 * to the object. Otherwise we will not report the leak.
2614                 */
2615                kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2616                                   local_flags);
2617                if (!slabp)
2618                        return NULL;
2619        } else {
2620                slabp = objp + colour_off;
2621                colour_off += cachep->slab_size;
2622        }
2623        slabp->inuse = 0;
2624        slabp->colouroff = colour_off;
2625        slabp->s_mem = objp + colour_off;
2626        slabp->nodeid = nodeid;
2627        slabp->free = 0;
2628        return slabp;
2629}
2630
2631static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2632{
2633        return (kmem_bufctl_t *) (slabp + 1);
2634}
2635
2636static void cache_init_objs(struct kmem_cache *cachep,
2637                            struct slab *slabp)
2638{
2639        int i;
2640
2641        for (i = 0; i < cachep->num; i++) {
2642                void *objp = index_to_obj(cachep, slabp, i);
2643#if DEBUG
2644                /* need to poison the objs? */
2645                if (cachep->flags & SLAB_POISON)
2646                        poison_obj(cachep, objp, POISON_FREE);
2647                if (cachep->flags & SLAB_STORE_USER)
2648                        *dbg_userword(cachep, objp) = NULL;
2649
2650                if (cachep->flags & SLAB_RED_ZONE) {
2651                        *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2652                        *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2653                }
2654                /*
2655                 * Constructors are not allowed to allocate memory from the same
2656                 * cache which they are a constructor for.  Otherwise, deadlock.
2657                 * They must also be threaded.
2658                 */
2659                if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2660                        cachep->ctor(objp + obj_offset(cachep));
2661
2662                if (cachep->flags & SLAB_RED_ZONE) {
2663                        if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2664                                slab_error(cachep, "constructor overwrote the"
2665                                           " end of an object");
2666                        if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2667                                slab_error(cachep, "constructor overwrote the"
2668                                           " start of an object");
2669                }
2670                if ((cachep->size % PAGE_SIZE) == 0 &&
2671                            OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2672                        kernel_map_pages(virt_to_page(objp),
2673                                         cachep->size / PAGE_SIZE, 0);
2674#else
2675                if (cachep->ctor)
2676                        cachep->ctor(objp);
2677#endif
2678                slab_bufctl(slabp)[i] = i + 1;
2679        }
2680        slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2681}
2682
2683static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2684{
2685        if (CONFIG_ZONE_DMA_FLAG) {
2686                if (flags & GFP_DMA)
2687                        BUG_ON(!(cachep->allocflags & GFP_DMA));
2688                else
2689                        BUG_ON(cachep->allocflags & GFP_DMA);
2690        }
2691}
2692
2693static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2694                                int nodeid)
2695{
2696        void *objp = index_to_obj(cachep, slabp, slabp->free);
2697        kmem_bufctl_t next;
2698
2699        slabp->inuse++;
2700        next = slab_bufctl(slabp)[slabp->free];
2701#if DEBUG
2702        slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2703        WARN_ON(slabp->nodeid != nodeid);
2704#endif
2705        slabp->free = next;
2706
2707        return objp;
2708}
2709
2710static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2711                                void *objp, int nodeid)
2712{
2713        unsigned int objnr = obj_to_index(cachep, slabp, objp);
2714
2715#if DEBUG
2716        /* Verify that the slab belongs to the intended node */
2717        WARN_ON(slabp->nodeid != nodeid);
2718
2719        if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2720                printk(KERN_ERR "slab: double free detected in cache "
2721                                "'%s', objp %p\n", cachep->name, objp);
2722                BUG();
2723        }
2724#endif
2725        slab_bufctl(slabp)[objnr] = slabp->free;
2726        slabp->free = objnr;
2727        slabp->inuse--;
2728}
2729
2730/*
2731 * Map pages beginning at addr to the given cache and slab. This is required
2732 * for the slab allocator to be able to lookup the cache and slab of a
2733 * virtual address for kfree, ksize, and slab debugging.
2734 */
2735static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2736                           void *addr)
2737{
2738        int nr_pages;
2739        struct page *page;
2740
2741        page = virt_to_page(addr);
2742
2743        nr_pages = 1;
2744        if (likely(!PageCompound(page)))
2745                nr_pages <<= cache->gfporder;
2746
2747        do {
2748                page->slab_cache = cache;
2749                page->slab_page = slab;
2750                page++;
2751        } while (--nr_pages);
2752}
2753
2754/*
2755 * Grow (by 1) the number of slabs within a cache.  This is called by
2756 * kmem_cache_alloc() when there are no active objs left in a cache.
2757 */
2758static int cache_grow(struct kmem_cache *cachep,
2759                gfp_t flags, int nodeid, void *objp)
2760{
2761        struct slab *slabp;
2762        size_t offset;
2763        gfp_t local_flags;
2764        struct kmem_cache_node *n;
2765
2766        /*
2767         * Be lazy and only check for valid flags here,  keeping it out of the
2768         * critical path in kmem_cache_alloc().
2769         */
2770        BUG_ON(flags & GFP_SLAB_BUG_MASK);
2771        local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2772
2773        /* Take the node list lock to change the colour_next on this node */
2774        check_irq_off();
2775        n = cachep->node[nodeid];
2776        spin_lock(&n->list_lock);
2777
2778        /* Get colour for the slab, and cal the next value. */
2779        offset = n->colour_next;
2780        n->colour_next++;
2781        if (n->colour_next >= cachep->colour)
2782                n->colour_next = 0;
2783        spin_unlock(&n->list_lock);
2784
2785        offset *= cachep->colour_off;
2786
2787        if (local_flags & __GFP_WAIT)
2788                local_irq_enable();
2789
2790        /*
2791         * The test for missing atomic flag is performed here, rather than
2792         * the more obvious place, simply to reduce the critical path length
2793         * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2794         * will eventually be caught here (where it matters).
2795         */
2796        kmem_flagcheck(cachep, flags);
2797
2798        /*
2799         * Get mem for the objs.  Attempt to allocate a physical page from
2800         * 'nodeid'.
2801         */
2802        if (!objp)
2803                objp = kmem_getpages(cachep, local_flags, nodeid);
2804        if (!objp)
2805                goto failed;
2806
2807        /* Get slab management. */
2808        slabp = alloc_slabmgmt(cachep, objp, offset,
2809                        local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2810        if (!slabp)
2811                goto opps1;
2812
2813        slab_map_pages(cachep, slabp, objp);
2814
2815        cache_init_objs(cachep, slabp);
2816
2817        if (local_flags & __GFP_WAIT)
2818                local_irq_disable();
2819        check_irq_off();
2820        spin_lock(&n->list_lock);
2821
2822        /* Make slab active. */
2823        list_add_tail(&slabp->list, &(n->slabs_free));
2824        STATS_INC_GROWN(cachep);
2825        n->free_objects += cachep->num;
2826        spin_unlock(&n->list_lock);
2827        return 1;
2828opps1:
2829        kmem_freepages(cachep, objp);
2830failed:
2831        if (local_flags & __GFP_WAIT)
2832                local_irq_disable();
2833        return 0;
2834}
2835
2836#if DEBUG
2837
2838/*
2839 * Perform extra freeing checks:
2840 * - detect bad pointers.
2841 * - POISON/RED_ZONE checking
2842 */
2843static void kfree_debugcheck(const void *objp)
2844{
2845        if (!virt_addr_valid(objp)) {
2846                printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2847                       (unsigned long)objp);
2848                BUG();
2849        }
2850}
2851
2852static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2853{
2854        unsigned long long redzone1, redzone2;
2855
2856        redzone1 = *dbg_redzone1(cache, obj);
2857        redzone2 = *dbg_redzone2(cache, obj);
2858
2859        /*
2860         * Redzone is ok.
2861         */
2862        if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2863                return;
2864
2865        if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2866                slab_error(cache, "double free detected");
2867        else
2868                slab_error(cache, "memory outside object was overwritten");
2869
2870        printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2871                        obj, redzone1, redzone2);
2872}
2873
2874static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2875                                   unsigned long caller)
2876{
2877        struct page *page;
2878        unsigned int objnr;
2879        struct slab *slabp;
2880
2881        BUG_ON(virt_to_cache(objp) != cachep);
2882
2883        objp -= obj_offset(cachep);
2884        kfree_debugcheck(objp);
2885        page = virt_to_head_page(objp);
2886
2887        slabp = page->slab_page;
2888
2889        if (cachep->flags & SLAB_RED_ZONE) {
2890                verify_redzone_free(cachep, objp);
2891                *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2892                *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2893        }
2894        if (cachep->flags & SLAB_STORE_USER)
2895                *dbg_userword(cachep, objp) = (void *)caller;
2896
2897        objnr = obj_to_index(cachep, slabp, objp);
2898
2899        BUG_ON(objnr >= cachep->num);
2900        BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2901
2902#ifdef CONFIG_DEBUG_SLAB_LEAK
2903        slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2904#endif
2905        if (cachep->flags & SLAB_POISON) {
2906#ifdef CONFIG_DEBUG_PAGEALLOC
2907                if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2908                        store_stackinfo(cachep, objp, caller);
2909                        kernel_map_pages(virt_to_page(objp),
2910                                         cachep->size / PAGE_SIZE, 0);
2911                } else {
2912                        poison_obj(cachep, objp, POISON_FREE);
2913                }
2914#else
2915                poison_obj(cachep, objp, POISON_FREE);
2916#endif
2917        }
2918        return objp;
2919}
2920
2921static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2922{
2923        kmem_bufctl_t i;
2924        int entries = 0;
2925
2926        /* Check slab's freelist to see if this obj is there. */
2927        for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2928                entries++;
2929                if (entries > cachep->num || i >= cachep->num)
2930                        goto bad;
2931        }
2932        if (entries != cachep->num - slabp->inuse) {
2933bad:
2934                printk(KERN_ERR "slab: Internal list corruption detected in "
2935                        "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
2936                        cachep->name, cachep->num, slabp, slabp->inuse,
2937                        print_tainted());
2938                print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
2939                        sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
2940                        1);
2941                BUG();
2942        }
2943}
2944#else
2945#define kfree_debugcheck(x) do { } while(0)
2946#define cache_free_debugcheck(x,objp,z) (objp)
2947#define check_slabp(x,y) do { } while(0)
2948#endif
2949
2950static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2951                                                        bool force_refill)
2952{
2953        int batchcount;
2954        struct kmem_cache_node *n;
2955        struct array_cache *ac;
2956        int node;
2957
2958        check_irq_off();
2959        node = numa_mem_id();
2960        if (unlikely(force_refill))
2961                goto force_grow;
2962retry:
2963        ac = cpu_cache_get(cachep);
2964        batchcount = ac->batchcount;
2965        if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2966                /*
2967                 * If there was little recent activity on this cache, then
2968                 * perform only a partial refill.  Otherwise we could generate
2969                 * refill bouncing.
2970                 */
2971                batchcount = BATCHREFILL_LIMIT;
2972        }
2973        n = cachep->node[node];
2974
2975        BUG_ON(ac->avail > 0 || !n);
2976        spin_lock(&n->list_lock);
2977
2978        /* See if we can refill from the shared array */
2979        if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2980                n->shared->touched = 1;
2981                goto alloc_done;
2982        }
2983
2984        while (batchcount > 0) {
2985                struct list_head *entry;
2986                struct slab *slabp;
2987                /* Get slab alloc is to come from. */
2988                entry = n->slabs_partial.next;
2989                if (entry == &n->slabs_partial) {
2990                        n->free_touched = 1;
2991                        entry = n->slabs_free.next;
2992                        if (entry == &n->slabs_free)
2993                                goto must_grow;
2994                }
2995
2996                slabp = list_entry(entry, struct slab, list);
2997                check_slabp(cachep, slabp);
2998                check_spinlock_acquired(cachep);
2999
3000                /*
3001                 * The slab was either on partial or free list so
3002                 * there must be at least one object available for
3003                 * allocation.
3004                 */
3005                BUG_ON(slabp->inuse >= cachep->num);
3006
3007                while (slabp->inuse < cachep->num && batchcount--) {
3008                        STATS_INC_ALLOCED(cachep);
3009                        STATS_INC_ACTIVE(cachep);
3010                        STATS_SET_HIGH(cachep);
3011
3012                        ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3013                                                                        node));
3014                }
3015                check_slabp(cachep, slabp);
3016
3017                /* move slabp to correct slabp list: */
3018                list_del(&slabp->list);
3019                if (slabp->free == BUFCTL_END)
3020                        list_add(&slabp->list, &n->slabs_full);
3021                else
3022                        list_add(&slabp->list, &n->slabs_partial);
3023        }
3024
3025must_grow:
3026        n->free_objects -= ac->avail;
3027alloc_done:
3028        spin_unlock(&n->list_lock);
3029
3030        if (unlikely(!ac->avail)) {
3031                int x;
3032force_grow:
3033                x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3034
3035                /* cache_grow can reenable interrupts, then ac could change. */
3036                ac = cpu_cache_get(cachep);
3037                node = numa_mem_id();
3038
3039                /* no objects in sight? abort */
3040                if (!x && (ac->avail == 0 || force_refill))
3041                        return NULL;
3042
3043                if (!ac->avail)         /* objects refilled by interrupt? */
3044                        goto retry;
3045        }
3046        ac->touched = 1;
3047
3048        return ac_get_obj(cachep, ac, flags, force_refill);
3049}
3050
3051static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3052                                                gfp_t flags)
3053{
3054        might_sleep_if(flags & __GFP_WAIT);
3055#if DEBUG
3056        kmem_flagcheck(cachep, flags);
3057#endif
3058}
3059
3060#if DEBUG
3061static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3062                                gfp_t flags, void *objp, unsigned long caller)
3063{
3064        if (!objp)
3065                return objp;
3066        if (cachep->flags & SLAB_POISON) {
3067#ifdef CONFIG_DEBUG_PAGEALLOC
3068                if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3069                        kernel_map_pages(virt_to_page(objp),
3070                                         cachep->size / PAGE_SIZE, 1);
3071                else
3072                        check_poison_obj(cachep, objp);
3073#else
3074                check_poison_obj(cachep, objp);
3075#endif
3076                poison_obj(cachep, objp, POISON_INUSE);
3077        }
3078        if (cachep->flags & SLAB_STORE_USER)
3079                *dbg_userword(cachep, objp) = (void *)caller;
3080
3081        if (cachep->flags & SLAB_RED_ZONE) {
3082                if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3083                                *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3084                        slab_error(cachep, "double free, or memory outside"
3085                                                " object was overwritten");
3086                        printk(KERN_ERR
3087                                "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3088                                objp, *dbg_redzone1(cachep, objp),
3089                                *dbg_redzone2(cachep, objp));
3090                }
3091                *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3092                *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3093        }
3094#ifdef CONFIG_DEBUG_SLAB_LEAK
3095        {
3096                struct slab *slabp;
3097                unsigned objnr;
3098
3099                slabp = virt_to_head_page(objp)->slab_page;
3100                objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3101                slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3102        }
3103#endif
3104        objp += obj_offset(cachep);
3105        if (cachep->ctor && cachep->flags & SLAB_POISON)
3106                cachep->ctor(objp);
3107        if (ARCH_SLAB_MINALIGN &&
3108            ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3109                printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3110                       objp, (int)ARCH_SLAB_MINALIGN);
3111        }
3112        return objp;
3113}
3114#else
3115#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3116#endif
3117
3118static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3119{
3120        if (cachep == kmem_cache)
3121                return false;
3122
3123        return should_failslab(cachep->object_size, flags, cachep->flags);
3124}
3125
3126static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3127{
3128        void *objp;
3129        struct array_cache *ac;
3130        bool force_refill = false;
3131
3132        check_irq_off();
3133
3134        ac = cpu_cache_get(cachep);
3135        if (likely(ac->avail)) {
3136                ac->touched = 1;
3137                objp = ac_get_obj(cachep, ac, flags, false);
3138
3139                /*
3140                 * Allow for the possibility all avail objects are not allowed
3141                 * by the current flags
3142                 */
3143                if (objp) {
3144                        STATS_INC_ALLOCHIT(cachep);
3145                        goto out;
3146                }
3147                force_refill = true;
3148        }
3149
3150        STATS_INC_ALLOCMISS(cachep);
3151        objp = cache_alloc_refill(cachep, flags, force_refill);
3152        /*
3153         * the 'ac' may be updated by cache_alloc_refill(),
3154         * and kmemleak_erase() requires its correct value.
3155         */
3156        ac = cpu_cache_get(cachep);
3157
3158out:
3159        /*
3160         * To avoid a false negative, if an object that is in one of the
3161         * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3162         * treat the array pointers as a reference to the object.
3163         */
3164        if (objp)
3165                kmemleak_erase(&ac->entry[ac->avail]);
3166        return objp;
3167}
3168
3169#ifdef CONFIG_NUMA
3170/*
3171 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3172 *
3173 * If we are in_interrupt, then process context, including cpusets and
3174 * mempolicy, may not apply and should not be used for allocation policy.
3175 */
3176static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3177{
3178        int nid_alloc, nid_here;
3179
3180        if (in_interrupt() || (flags & __GFP_THISNODE))
3181                return NULL;
3182        nid_alloc = nid_here = numa_mem_id();
3183        if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3184                nid_alloc = cpuset_slab_spread_node();
3185        else if (current->mempolicy)
3186                nid_alloc = slab_node();
3187        if (nid_alloc != nid_here)
3188                return ____cache_alloc_node(cachep, flags, nid_alloc);
3189        return NULL;
3190}
3191
3192/*
3193 * Fallback function if there was no memory available and no objects on a
3194 * certain node and fall back is permitted. First we scan all the
3195 * available node for available objects. If that fails then we
3196 * perform an allocation without specifying a node. This allows the page
3197 * allocator to do its reclaim / fallback magic. We then insert the
3198 * slab into the proper nodelist and then allocate from it.
3199 */
3200static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3201{
3202        struct zonelist *zonelist;
3203        gfp_t local_flags;
3204        struct zoneref *z;
3205        struct zone *zone;
3206        enum zone_type high_zoneidx = gfp_zone(flags);
3207        void *obj = NULL;
3208        int nid;
3209        unsigned int cpuset_mems_cookie;
3210
3211        if (flags & __GFP_THISNODE)
3212                return NULL;
3213
3214        local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3215
3216retry_cpuset:
3217        cpuset_mems_cookie = get_mems_allowed();
3218        zonelist = node_zonelist(slab_node(), flags);
3219
3220retry:
3221        /*
3222         * Look through allowed nodes for objects available
3223         * from existing per node queues.
3224         */
3225        for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3226                nid = zone_to_nid(zone);
3227
3228                if (cpuset_zone_allowed_hardwall(zone, flags) &&
3229                        cache->node[nid] &&
3230                        cache->node[nid]->free_objects) {
3231                                obj = ____cache_alloc_node(cache,
3232                                        flags | GFP_THISNODE, nid);
3233                                if (obj)
3234                                        break;
3235                }
3236        }
3237
3238        if (!obj) {
3239                /*
3240                 * This allocation will be performed within the constraints
3241                 * of the current cpuset / memory policy requirements.
3242                 * We may trigger various forms of reclaim on the allowed
3243                 * set and go into memory reserves if necessary.
3244                 */
3245                if (local_flags & __GFP_WAIT)
3246                        local_irq_enable();
3247                kmem_flagcheck(cache, flags);
3248                obj = kmem_getpages(cache, local_flags, numa_mem_id());
3249                if (local_flags & __GFP_WAIT)
3250                        local_irq_disable();
3251                if (obj) {
3252                        /*
3253                         * Insert into the appropriate per node queues
3254                         */
3255                        nid = page_to_nid(virt_to_page(obj));
3256                        if (cache_grow(cache, flags, nid, obj)) {
3257                                obj = ____cache_alloc_node(cache,
3258                                        flags | GFP_THISNODE, nid);
3259                                if (!obj)
3260                                        /*
3261                                         * Another processor may allocate the
3262                                         * objects in the slab since we are
3263                                         * not holding any locks.
3264                                         */
3265                                        goto retry;
3266                        } else {
3267                                /* cache_grow already freed obj */
3268                                obj = NULL;
3269                        }
3270                }
3271        }
3272
3273        if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3274                goto retry_cpuset;
3275        return obj;
3276}
3277
3278/*
3279 * A interface to enable slab creation on nodeid
3280 */
3281static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3282                                int nodeid)
3283{
3284        struct list_head *entry;
3285        struct slab *slabp;
3286        struct kmem_cache_node *n;
3287        void *obj;
3288        int x;
3289
3290        VM_BUG_ON(nodeid > num_online_nodes());
3291        n = cachep->node[nodeid];
3292        BUG_ON(!n);
3293
3294retry:
3295        check_irq_off();
3296        spin_lock(&n->list_lock);
3297        entry = n->slabs_partial.next;
3298        if (entry == &n->slabs_partial) {
3299                n->free_touched = 1;
3300                entry = n->slabs_free.next;
3301                if (entry == &n->slabs_free)
3302                        goto must_grow;
3303        }
3304
3305        slabp = list_entry(entry, struct slab, list);
3306        check_spinlock_acquired_node(cachep, nodeid);
3307        check_slabp(cachep, slabp);
3308
3309        STATS_INC_NODEALLOCS(cachep);
3310        STATS_INC_ACTIVE(cachep);
3311        STATS_SET_HIGH(cachep);
3312
3313        BUG_ON(slabp->inuse == cachep->num);
3314
3315        obj = slab_get_obj(cachep, slabp, nodeid);
3316        check_slabp(cachep, slabp);
3317        n->free_objects--;
3318        /* move slabp to correct slabp list: */
3319        list_del(&slabp->list);
3320
3321        if (slabp->free == BUFCTL_END)
3322                list_add(&slabp->list, &n->slabs_full);
3323        else
3324                list_add(&slabp->list, &n->slabs_partial);
3325
3326        spin_unlock(&n->list_lock);
3327        goto done;
3328
3329must_grow:
3330        spin_unlock(&n->list_lock);
3331        x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3332        if (x)
3333                goto retry;
3334
3335        return fallback_alloc(cachep, flags);
3336
3337done:
3338        return obj;
3339}
3340
3341/**
3342 * kmem_cache_alloc_node - Allocate an object on the specified node
3343 * @cachep: The cache to allocate from.
3344 * @flags: See kmalloc().
3345 * @nodeid: node number of the target node.
3346 * @caller: return address of caller, used for debug information
3347 *
3348 * Identical to kmem_cache_alloc but it will allocate memory on the given
3349 * node, which can improve the performance for cpu bound structures.
3350 *
3351 * Fallback to other node is possible if __GFP_THISNODE is not set.
3352 */
3353static __always_inline void *
3354slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3355                   unsigned long caller)
3356{
3357        unsigned long save_flags;
3358        void *ptr;
3359        int slab_node = numa_mem_id();
3360
3361        flags &= gfp_allowed_mask;
3362
3363        lockdep_trace_alloc(flags);
3364
3365        if (slab_should_failslab(cachep, flags))
3366                return NULL;
3367
3368        cachep = memcg_kmem_get_cache(cachep, flags);
3369
3370        cache_alloc_debugcheck_before(cachep, flags);
3371        local_irq_save(save_flags);
3372
3373        if (nodeid == NUMA_NO_NODE)
3374                nodeid = slab_node;
3375
3376        if (unlikely(!cachep->node[nodeid])) {
3377                /* Node not bootstrapped yet */
3378                ptr = fallback_alloc(cachep, flags);
3379                goto out;
3380        }
3381
3382        if (nodeid == slab_node) {
3383                /*
3384                 * Use the locally cached objects if possible.
3385                 * However ____cache_alloc does not allow fallback
3386                 * to other nodes. It may fail while we still have
3387                 * objects on other nodes available.
3388                 */
3389                ptr = ____cache_alloc(cachep, flags);
3390                if (ptr)
3391                        goto out;
3392        }
3393        /* ___cache_alloc_node can fall back to other nodes */
3394        ptr = ____cache_alloc_node(cachep, flags, nodeid);
3395  out:
3396        local_irq_restore(save_flags);
3397        ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3398        kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3399                                 flags);
3400
3401        if (likely(ptr))
3402                kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3403
3404        if (unlikely((flags & __GFP_ZERO) && ptr))
3405                memset(ptr, 0, cachep->object_size);
3406
3407        return ptr;
3408}
3409
3410static __always_inline void *
3411__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3412{
3413        void *objp;
3414
3415        if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3416                objp = alternate_node_alloc(cache, flags);
3417                if (objp)
3418                        goto out;
3419        }
3420        objp = ____cache_alloc(cache, flags);
3421
3422        /*
3423         * We may just have run out of memory on the local node.
3424         * ____cache_alloc_node() knows how to locate memory on other nodes
3425         */
3426        if (!objp)
3427                objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3428
3429  out:
3430        return objp;
3431}
3432#else
3433
3434static __always_inline void *
3435__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3436{
3437        return ____cache_alloc(cachep, flags);
3438}
3439
3440#endif /* CONFIG_NUMA */
3441
3442static __always_inline void *
3443slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3444{
3445        unsigned long save_flags;
3446        void *objp;
3447
3448        flags &= gfp_allowed_mask;
3449
3450        lockdep_trace_alloc(flags);
3451
3452        if (slab_should_failslab(cachep, flags))
3453                return NULL;
3454
3455        cachep = memcg_kmem_get_cache(cachep, flags);
3456
3457        cache_alloc_debugcheck_before(cachep, flags);
3458        local_irq_save(save_flags);
3459        objp = __do_cache_alloc(cachep, flags);
3460        local_irq_restore(save_flags);
3461        objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3462        kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3463                                 flags);
3464        prefetchw(objp);
3465
3466        if (likely(objp))
3467                kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3468
3469        if (unlikely((flags & __GFP_ZERO) && objp))
3470                memset(objp, 0, cachep->object_size);
3471
3472        return objp;
3473}
3474
3475/*
3476 * Caller needs to acquire correct kmem_list's list_lock
3477 */
3478static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3479                       int node)
3480{
3481        int i;
3482        struct kmem_cache_node *n;
3483
3484        for (i = 0; i < nr_objects; i++) {
3485                void *objp;
3486                struct slab *slabp;
3487
3488                clear_obj_pfmemalloc(&objpp[i]);
3489                objp = objpp[i];
3490
3491                slabp = virt_to_slab(objp);
3492                n = cachep->node[node];
3493                list_del(&slabp->list);
3494                check_spinlock_acquired_node(cachep, node);
3495                check_slabp(cachep, slabp);
3496                slab_put_obj(cachep, slabp, objp, node);
3497                STATS_DEC_ACTIVE(cachep);
3498                n->free_objects++;
3499                check_slabp(cachep, slabp);
3500
3501                /* fixup slab chains */
3502                if (slabp->inuse == 0) {
3503                        if (n->free_objects > n->free_limit) {
3504                                n->free_objects -= cachep->num;
3505                                /* No need to drop any previously held
3506                                 * lock here, even if we have a off-slab slab
3507                                 * descriptor it is guaranteed to come from
3508                                 * a different cache, refer to comments before
3509                                 * alloc_slabmgmt.
3510                                 */
3511                                slab_destroy(cachep, slabp);
3512                        } else {
3513                                list_add(&slabp->list, &n->slabs_free);
3514                        }
3515                } else {
3516                        /* Unconditionally move a slab to the end of the
3517                         * partial list on free - maximum time for the
3518                         * other objects to be freed, too.
3519                         */
3520                        list_add_tail(&slabp->list, &n->slabs_partial);
3521                }
3522        }
3523}
3524
3525static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3526{
3527        int batchcount;
3528        struct kmem_cache_node *n;
3529        int node = numa_mem_id();
3530
3531        batchcount = ac->batchcount;
3532#if DEBUG
3533        BUG_ON(!batchcount || batchcount > ac->avail);
3534#endif
3535        check_irq_off();
3536        n = cachep->node[node];
3537        spin_lock(&n->list_lock);
3538        if (n->shared) {
3539                struct array_cache *shared_array = n->shared;
3540                int max = shared_array->limit - shared_array->avail;
3541                if (max) {
3542                        if (batchcount > max)
3543                                batchcount = max;
3544                        memcpy(&(shared_array->entry[shared_array->avail]),
3545                               ac->entry, sizeof(void *) * batchcount);
3546                        shared_array->avail += batchcount;
3547                        goto free_done;
3548                }
3549        }
3550
3551        free_block(cachep, ac->entry, batchcount, node);
3552free_done:
3553#if STATS
3554        {
3555                int i = 0;
3556                struct list_head *p;
3557
3558                p = n->slabs_free.next;
3559                while (p != &(n->slabs_free)) {
3560                        struct slab *slabp;
3561
3562                        slabp = list_entry(p, struct slab, list);
3563                        BUG_ON(slabp->inuse);
3564
3565                        i++;
3566                        p = p->next;
3567                }
3568                STATS_SET_FREEABLE(cachep, i);
3569        }
3570#endif
3571        spin_unlock(&n->list_lock);
3572        ac->avail -= batchcount;
3573        memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3574}
3575
3576/*
3577 * Release an obj back to its cache. If the obj has a constructed state, it must
3578 * be in this state _before_ it is released.  Called with disabled ints.
3579 */
3580static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3581                                unsigned long caller)
3582{
3583        struct array_cache *ac = cpu_cache_get(cachep);
3584
3585        check_irq_off();
3586        kmemleak_free_recursive(objp, cachep->flags);
3587        objp = cache_free_debugcheck(cachep, objp, caller);
3588
3589        kmemcheck_slab_free(cachep, objp, cachep->object_size);
3590
3591        /*
3592         * Skip calling cache_free_alien() when the platform is not numa.
3593         * This will avoid cache misses that happen while accessing slabp (which
3594         * is per page memory  reference) to get nodeid. Instead use a global
3595         * variable to skip the call, which is mostly likely to be present in
3596         * the cache.
3597         */
3598        if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3599                return;
3600
3601        if (likely(ac->avail < ac->limit)) {
3602                STATS_INC_FREEHIT(cachep);
3603        } else {
3604                STATS_INC_FREEMISS(cachep);
3605                cache_flusharray(cachep, ac);
3606        }
3607
3608        ac_put_obj(cachep, ac, objp);
3609}
3610
3611/**
3612 * kmem_cache_alloc - Allocate an object
3613 * @cachep: The cache to allocate from.
3614 * @flags: See kmalloc().
3615 *
3616 * Allocate an object from this cache.  The flags are only relevant
3617 * if the cache has no available objects.
3618 */
3619void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3620{
3621        void *ret = slab_alloc(cachep, flags, _RET_IP_);
3622
3623        trace_kmem_cache_alloc(_RET_IP_, ret,
3624                               cachep->object_size, cachep->size, flags);
3625
3626        return ret;
3627}
3628EXPORT_SYMBOL(kmem_cache_alloc);
3629
3630#ifdef CONFIG_TRACING
3631void *
3632kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3633{
3634        void *ret;
3635
3636        ret = slab_alloc(cachep, flags, _RET_IP_);
3637
3638        trace_kmalloc(_RET_IP_, ret,
3639                      size, cachep->size, flags);
3640        return ret;
3641}
3642EXPORT_SYMBOL(kmem_cache_alloc_trace);
3643#endif
3644
3645#ifdef CONFIG_NUMA
3646void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3647{
3648        void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3649
3650        trace_kmem_cache_alloc_node(_RET_IP_, ret,
3651                                    cachep->object_size, cachep->size,
3652                                    flags, nodeid);
3653
3654        return ret;
3655}
3656EXPORT_SYMBOL(kmem_cache_alloc_node);
3657
3658#ifdef CONFIG_TRACING
3659void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3660                                  gfp_t flags,
3661                                  int nodeid,
3662                                  size_t size)
3663{
3664        void *ret;
3665
3666        ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3667
3668        trace_kmalloc_node(_RET_IP_, ret,
3669                           size, cachep->size,
3670                           flags, nodeid);
3671        return ret;
3672}
3673EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3674#endif
3675
3676static __always_inline void *
3677__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3678{
3679        struct kmem_cache *cachep;
3680
3681        cachep = kmalloc_slab(size, flags);
3682        if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3683                return cachep;
3684        return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3685}
3686
3687#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3688void *__kmalloc_node(size_t size, gfp_t flags, int node)
3689{
3690        return __do_kmalloc_node(size, flags, node, _RET_IP_);
3691}
3692EXPORT_SYMBOL(__kmalloc_node);
3693
3694void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3695                int node, unsigned long caller)
3696{
3697        return __do_kmalloc_node(size, flags, node, caller);
3698}
3699EXPORT_SYMBOL(__kmalloc_node_track_caller);
3700#else
3701void *__kmalloc_node(size_t size, gfp_t flags, int node)
3702{
3703        return __do_kmalloc_node(size, flags, node, 0);
3704}
3705EXPORT_SYMBOL(__kmalloc_node);
3706#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3707#endif /* CONFIG_NUMA */
3708
3709/**
3710 * __do_kmalloc - allocate memory
3711 * @size: how many bytes of memory are required.
3712 * @flags: the type of memory to allocate (see kmalloc).
3713 * @caller: function caller for debug tracking of the caller
3714 */
3715static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3716                                          unsigned long caller)
3717{
3718        struct kmem_cache *cachep;
3719        void *ret;
3720
3721        /* If you want to save a few bytes .text space: replace
3722         * __ with kmem_.
3723         * Then kmalloc uses the uninlined functions instead of the inline
3724         * functions.
3725         */
3726        cachep = kmalloc_slab(size, flags);
3727        if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3728                return cachep;
3729        ret = slab_alloc(cachep, flags, caller);
3730
3731        trace_kmalloc(caller, ret,
3732                      size, cachep->size, flags);
3733
3734        return ret;
3735}
3736
3737
3738#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3739void *__kmalloc(size_t size, gfp_t flags)
3740{
3741        return __do_kmalloc(size, flags, _RET_IP_);
3742}
3743EXPORT_SYMBOL(__kmalloc);
3744
3745void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3746{
3747        return __do_kmalloc(size, flags, caller);
3748}
3749EXPORT_SYMBOL(__kmalloc_track_caller);
3750
3751#else
3752void *__kmalloc(size_t size, gfp_t flags)
3753{
3754        return __do_kmalloc(size, flags, 0);
3755}
3756EXPORT_SYMBOL(__kmalloc);
3757#endif
3758
3759/**
3760 * kmem_cache_free - Deallocate an object
3761 * @cachep: The cache the allocation was from.
3762 * @objp: The previously allocated object.
3763 *
3764 * Free an object which was previously allocated from this
3765 * cache.
3766 */
3767void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3768{
3769        unsigned long flags;
3770        cachep = cache_from_obj(cachep, objp);
3771        if (!cachep)
3772                return;
3773
3774        local_irq_save(flags);
3775        debug_check_no_locks_freed(objp, cachep->object_size);
3776        if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3777                debug_check_no_obj_freed(objp, cachep->object_size);
3778        __cache_free(cachep, objp, _RET_IP_);
3779        local_irq_restore(flags);
3780
3781        trace_kmem_cache_free(_RET_IP_, objp);
3782}
3783EXPORT_SYMBOL(kmem_cache_free);
3784
3785/**
3786 * kfree - free previously allocated memory
3787 * @objp: pointer returned by kmalloc.
3788 *
3789 * If @objp is NULL, no operation is performed.
3790 *
3791 * Don't free memory not originally allocated by kmalloc()
3792 * or you will run into trouble.
3793 */
3794void kfree(const void *objp)
3795{
3796        struct kmem_cache *c;
3797        unsigned long flags;
3798
3799        trace_kfree(_RET_IP_, objp);
3800
3801        if (unlikely(ZERO_OR_NULL_PTR(objp)))
3802                return;
3803        local_irq_save(flags);
3804        kfree_debugcheck(objp);
3805        c = virt_to_cache(objp);
3806        debug_check_no_locks_freed(objp, c->object_size);
3807
3808        debug_check_no_obj_freed(objp, c->object_size);
3809        __cache_free(c, (void *)objp, _RET_IP_);
3810        local_irq_restore(flags);
3811}
3812EXPORT_SYMBOL(kfree);
3813
3814/*
3815 * This initializes kmem_cache_node or resizes various caches for all nodes.
3816 */
3817static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3818{
3819        int node;
3820        struct kmem_cache_node *n;
3821        struct array_cache *new_shared;
3822        struct array_cache **new_alien = NULL;
3823
3824        for_each_online_node(node) {
3825
3826                if (use_alien_caches) {
3827                        new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3828                        if (!new_alien)
3829                                goto fail;
3830                }
3831
3832                new_shared = NULL;
3833                if (cachep->shared) {
3834                        new_shared = alloc_arraycache(node,
3835                                cachep->shared*cachep->batchcount,
3836                                        0xbaadf00d, gfp);
3837                        if (!new_shared) {
3838                                free_alien_cache(new_alien);
3839                                goto fail;
3840                        }
3841                }
3842
3843                n = cachep->node[node];
3844                if (n) {
3845                        struct array_cache *shared = n->shared;
3846
3847                        spin_lock_irq(&n->list_lock);
3848
3849                        if (shared)
3850                                free_block(cachep, shared->entry,
3851                                                shared->avail, node);
3852
3853                        n->shared = new_shared;
3854                        if (!n->alien) {
3855                                n->alien = new_alien;
3856                                new_alien = NULL;
3857                        }
3858                        n->free_limit = (1 + nr_cpus_node(node)) *
3859                                        cachep->batchcount + cachep->num;
3860                        spin_unlock_irq(&n->list_lock);
3861                        kfree(shared);
3862                        free_alien_cache(new_alien);
3863                        continue;
3864                }
3865                n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3866                if (!n) {
3867                        free_alien_cache(new_alien);
3868                        kfree(new_shared);
3869                        goto fail;
3870                }
3871
3872                kmem_cache_node_init(n);
3873                n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3874                                ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3875                n->shared = new_shared;
3876                n->alien = new_alien;
3877                n->free_limit = (1 + nr_cpus_node(node)) *
3878                                        cachep->batchcount + cachep->num;
3879                cachep->node[node] = n;
3880        }
3881        return 0;
3882
3883fail:
3884        if (!cachep->list.next) {
3885                /* Cache is not active yet. Roll back what we did */
3886                node--;
3887                while (node >= 0) {
3888                        if (cachep->node[node]) {
3889                                n = cachep->node[node];
3890
3891                                kfree(n->shared);
3892                                free_alien_cache(n->alien);
3893                                kfree(n);
3894                                cachep->node[node] = NULL;
3895                        }
3896                        node--;
3897                }
3898        }
3899        return -ENOMEM;
3900}
3901
3902struct ccupdate_struct {
3903        struct kmem_cache *cachep;
3904        struct array_cache *new[0];
3905};
3906
3907static void do_ccupdate_local(void *info)
3908{
3909        struct ccupdate_struct *new = info;
3910        struct array_cache *old;
3911
3912        check_irq_off();
3913        old = cpu_cache_get(new->cachep);
3914
3915        new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3916        new->new[smp_processor_id()] = old;
3917}
3918
3919/* Always called with the slab_mutex held */
3920static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3921                                int batchcount, int shared, gfp_t gfp)
3922{
3923        struct ccupdate_struct *new;
3924        int i;
3925
3926        new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3927                      gfp);
3928        if (!new)
3929                return -ENOMEM;
3930
3931        for_each_online_cpu(i) {
3932                new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3933                                                batchcount, gfp);
3934                if (!new->new[i]) {
3935                        for (i--; i >= 0; i--)
3936                                kfree(new->new[i]);
3937                        kfree(new);
3938                        return -ENOMEM;
3939                }
3940        }
3941        new->cachep = cachep;
3942
3943        on_each_cpu(do_ccupdate_local, (void *)new, 1);
3944
3945        check_irq_on();
3946        cachep->batchcount = batchcount;
3947        cachep->limit = limit;
3948        cachep->shared = shared;
3949
3950        for_each_online_cpu(i) {
3951                struct array_cache *ccold = new->new[i];
3952                if (!ccold)
3953                        continue;
3954                spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3955                free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3956                spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3957                kfree(ccold);
3958        }
3959        kfree(new);
3960        return alloc_kmemlist(cachep, gfp);
3961}
3962
3963static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3964                                int batchcount, int shared, gfp_t gfp)
3965{
3966        int ret;
3967        struct kmem_cache *c = NULL;
3968        int i = 0;
3969
3970        ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3971
3972        if (slab_state < FULL)
3973                return ret;
3974
3975        if ((ret < 0) || !is_root_cache(cachep))
3976                return ret;
3977
3978        VM_BUG_ON(!mutex_is_locked(&slab_mutex));
3979        for_each_memcg_cache_index(i) {
3980                c = cache_from_memcg(cachep, i);
3981                if (c)
3982                        /* return value determined by the parent cache only */
3983                        __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3984        }
3985
3986        return ret;
3987}
3988
3989/* Called with slab_mutex held always */
3990static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3991{
3992        int err;
3993        int limit = 0;
3994        int shared = 0;
3995        int batchcount = 0;
3996
3997        if (!is_root_cache(cachep)) {
3998                struct kmem_cache *root = memcg_root_cache(cachep);
3999                limit = root->limit;
4000                shared = root->shared;
4001                batchcount = root->batchcount;
4002        }
4003
4004        if (limit && shared && batchcount)
4005                goto skip_setup;
4006        /*
4007         * The head array serves three purposes:
4008         * - create a LIFO ordering, i.e. return objects that are cache-warm
4009         * - reduce the number of spinlock operations.
4010         * - reduce the number of linked list operations on the slab and
4011         *   bufctl chains: array operations are cheaper.
4012         * The numbers are guessed, we should auto-tune as described by
4013         * Bonwick.
4014         */
4015        if (cachep->size > 131072)
4016                limit = 1;
4017        else if (cachep->size > PAGE_SIZE)
4018                limit = 8;
4019        else if (cachep->size > 1024)
4020                limit = 24;
4021        else if (cachep->size > 256)
4022                limit = 54;
4023        else
4024                limit = 120;
4025
4026        /*
4027         * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4028         * allocation behaviour: Most allocs on one cpu, most free operations
4029         * on another cpu. For these cases, an efficient object passing between
4030         * cpus is necessary. This is provided by a shared array. The array
4031         * replaces Bonwick's magazine layer.
4032         * On uniprocessor, it's functionally equivalent (but less efficient)
4033         * to a larger limit. Thus disabled by default.
4034         */
4035        shared = 0;
4036        if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4037                shared = 8;
4038
4039#if DEBUG
4040        /*
4041         * With debugging enabled, large batchcount lead to excessively long
4042         * periods with disabled local interrupts. Limit the batchcount
4043         */
4044        if (limit > 32)
4045                limit = 32;
4046#endif
4047        batchcount = (limit + 1) / 2;
4048skip_setup:
4049        err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4050        if (err)
4051                printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4052                       cachep->name, -err);
4053        return err;
4054}
4055
4056/*
4057 * Drain an array if it contains any elements taking the node lock only if
4058 * necessary. Note that the node listlock also protects the array_cache
4059 * if drain_array() is used on the shared array.
4060 */
4061static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4062                         struct array_cache *ac, int force, int node)
4063{
4064        int tofree;
4065
4066        if (!ac || !ac->avail)
4067                return;
4068        if (ac->touched && !force) {
4069                ac->touched = 0;
4070        } else {
4071                spin_lock_irq(&n->list_lock);
4072                if (ac->avail) {
4073                        tofree = force ? ac->avail : (ac->limit + 4) / 5;
4074                        if (tofree > ac->avail)
4075                                tofree = (ac->avail + 1) / 2;
4076                        free_block(cachep, ac->entry, tofree, node);
4077                        ac->avail -= tofree;
4078                        memmove(ac->entry, &(ac->entry[tofree]),
4079                                sizeof(void *) * ac->avail);
4080                }
4081                spin_unlock_irq(&n->list_lock);
4082        }
4083}
4084
4085/**
4086 * cache_reap - Reclaim memory from caches.
4087 * @w: work descriptor
4088 *
4089 * Called from workqueue/eventd every few seconds.
4090 * Purpose:
4091 * - clear the per-cpu caches for this CPU.
4092 * - return freeable pages to the main free memory pool.
4093 *
4094 * If we cannot acquire the cache chain mutex then just give up - we'll try
4095 * again on the next iteration.
4096 */
4097static void cache_reap(struct work_struct *w)
4098{
4099        struct kmem_cache *searchp;
4100        struct kmem_cache_node *n;
4101        int node = numa_mem_id();
4102        struct delayed_work *work = to_delayed_work(w);
4103
4104        if (!mutex_trylock(&slab_mutex))
4105                /* Give up. Setup the next iteration. */
4106                goto out;
4107
4108        list_for_each_entry(searchp, &slab_caches, list) {
4109                check_irq_on();
4110
4111                /*
4112                 * We only take the node lock if absolutely necessary and we
4113                 * have established with reasonable certainty that
4114                 * we can do some work if the lock was obtained.
4115                 */
4116                n = searchp->node[node];
4117
4118                reap_alien(searchp, n);
4119
4120                drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
4121
4122                /*
4123                 * These are racy checks but it does not matter
4124                 * if we skip one check or scan twice.
4125                 */
4126                if (time_after(n->next_reap, jiffies))
4127                        goto next;
4128
4129                n->next_reap = jiffies + REAPTIMEOUT_LIST3;
4130
4131                drain_array(searchp, n, n->shared, 0, node);
4132
4133                if (n->free_touched)
4134                        n->free_touched = 0;
4135                else {
4136                        int freed;
4137
4138                        freed = drain_freelist(searchp, n, (n->free_limit +
4139                                5 * searchp->num - 1) / (5 * searchp->num));
4140                        STATS_ADD_REAPED(searchp, freed);
4141                }
4142next:
4143                cond_resched();
4144        }
4145        check_irq_on();
4146        mutex_unlock(&slab_mutex);
4147        next_reap_node();
4148out:
4149        /* Set up the next iteration */
4150        schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4151}
4152
4153#ifdef CONFIG_SLABINFO
4154void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4155{
4156        struct slab *slabp;
4157        unsigned long active_objs;
4158        unsigned long num_objs;
4159        unsigned long active_slabs = 0;
4160        unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4161        const char *name;
4162        char *error = NULL;
4163        int node;
4164        struct kmem_cache_node *n;
4165
4166        active_objs = 0;
4167        num_slabs = 0;
4168        for_each_online_node(node) {
4169                n = cachep->node[node];
4170                if (!n)
4171                        continue;
4172
4173                check_irq_on();
4174                spin_lock_irq(&n->list_lock);
4175
4176                list_for_each_entry(slabp, &n->slabs_full, list) {
4177                        if (slabp->inuse != cachep->num && !error)
4178                                error = "slabs_full accounting error";
4179                        active_objs += cachep->num;
4180                        active_slabs++;
4181                }
4182                list_for_each_entry(slabp, &n->slabs_partial, list) {
4183                        if (slabp->inuse == cachep->num && !error)
4184                                error = "slabs_partial inuse accounting error";
4185                        if (!slabp->inuse && !error)
4186                                error = "slabs_partial/inuse accounting error";
4187                        active_objs += slabp->inuse;
4188                        active_slabs++;
4189                }
4190                list_for_each_entry(slabp, &n->slabs_free, list) {
4191                        if (slabp->inuse && !error)
4192                                error = "slabs_free/inuse accounting error";
4193                        num_slabs++;
4194                }
4195                free_objects += n->free_objects;
4196                if (n->shared)
4197                        shared_avail += n->shared->avail;
4198
4199                spin_unlock_irq(&n->list_lock);
4200        }
4201        num_slabs += active_slabs;
4202        num_objs = num_slabs * cachep->num;
4203        if (num_objs - active_objs != free_objects && !error)
4204                error = "free_objects accounting error";
4205
4206        name = cachep->name;
4207        if (error)
4208                printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4209
4210        sinfo->active_objs = active_objs;
4211        sinfo->num_objs = num_objs;
4212        sinfo->active_slabs = active_slabs;
4213        sinfo->num_slabs = num_slabs;
4214        sinfo->shared_avail = shared_avail;
4215        sinfo->limit = cachep->limit;
4216        sinfo->batchcount = cachep->batchcount;
4217        sinfo->shared = cachep->shared;
4218        sinfo->objects_per_slab = cachep->num;
4219        sinfo->cache_order = cachep->gfporder;
4220}
4221
4222void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4223{
4224#if STATS
4225        {                       /* node stats */
4226                unsigned long high = cachep->high_mark;
4227                unsigned long allocs = cachep->num_allocations;
4228                unsigned long grown = cachep->grown;
4229                unsigned long reaped = cachep->reaped;
4230                unsigned long errors = cachep->errors;
4231                unsigned long max_freeable = cachep->max_freeable;
4232                unsigned long node_allocs = cachep->node_allocs;
4233                unsigned long node_frees = cachep->node_frees;
4234                unsigned long overflows = cachep->node_overflow;
4235
4236                seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4237                           "%4lu %4lu %4lu %4lu %4lu",
4238                           allocs, high, grown,
4239                           reaped, errors, max_freeable, node_allocs,
4240                           node_frees, overflows);
4241        }
4242        /* cpu stats */
4243        {
4244                unsigned long allochit = atomic_read(&cachep->allochit);
4245                unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4246                unsigned long freehit = atomic_read(&cachep->freehit);
4247                unsigned long freemiss = atomic_read(&cachep->freemiss);
4248
4249                seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4250                           allochit, allocmiss, freehit, freemiss);
4251        }
4252#endif
4253}
4254
4255#define MAX_SLABINFO_WRITE 128
4256/**
4257 * slabinfo_write - Tuning for the slab allocator
4258 * @file: unused
4259 * @buffer: user buffer
4260 * @count: data length
4261 * @ppos: unused
4262 */
4263ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4264                       size_t count, loff_t *ppos)
4265{
4266        char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4267        int limit, batchcount, shared, res;
4268        struct kmem_cache *cachep;
4269
4270        if (count > MAX_SLABINFO_WRITE)
4271                return -EINVAL;
4272        if (copy_from_user(&kbuf, buffer, count))
4273                return -EFAULT;
4274        kbuf[MAX_SLABINFO_WRITE] = '\0';
4275
4276        tmp = strchr(kbuf, ' ');
4277        if (!tmp)
4278                return -EINVAL;
4279        *tmp = '\0';
4280        tmp++;
4281        if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4282                return -EINVAL;
4283
4284        /* Find the cache in the chain of caches. */
4285        mutex_lock(&slab_mutex);
4286        res = -EINVAL;
4287        list_for_each_entry(cachep, &slab_caches, list) {
4288                if (!strcmp(cachep->name, kbuf)) {
4289                        if (limit < 1 || batchcount < 1 ||
4290                                        batchcount > limit || shared < 0) {
4291                                res = 0;
4292                        } else {
4293                                res = do_tune_cpucache(cachep, limit,
4294                                                       batchcount, shared,
4295                                                       GFP_KERNEL);
4296                        }
4297                        break;
4298                }
4299        }
4300        mutex_unlock(&slab_mutex);
4301        if (res >= 0)
4302                res = count;
4303        return res;
4304}
4305
4306#ifdef CONFIG_DEBUG_SLAB_LEAK
4307
4308static void *leaks_start(struct seq_file *m, loff_t *pos)
4309{
4310        mutex_lock(&slab_mutex);
4311        return seq_list_start(&slab_caches, *pos);
4312}
4313
4314static inline int add_caller(unsigned long *n, unsigned long v)
4315{
4316        unsigned long *p;
4317        int l;
4318        if (!v)
4319                return 1;
4320        l = n[1];
4321        p = n + 2;
4322        while (l) {
4323                int i = l/2;
4324                unsigned long *q = p + 2 * i;
4325                if (*q == v) {
4326                        q[1]++;
4327                        return 1;
4328                }
4329                if (*q > v) {
4330                        l = i;
4331                } else {
4332                        p = q + 2;
4333                        l -= i + 1;
4334                }
4335        }
4336        if (++n[1] == n[0])
4337                return 0;
4338        memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4339        p[0] = v;
4340        p[1] = 1;
4341        return 1;
4342}
4343
4344static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4345{
4346        void *p;
4347        int i;
4348        if (n[0] == n[1])
4349                return;
4350        for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4351                if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4352                        continue;
4353                if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4354                        return;
4355        }
4356}
4357
4358static void show_symbol(struct seq_file *m, unsigned long address)
4359{
4360#ifdef CONFIG_KALLSYMS
4361        unsigned long offset, size;
4362        char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4363
4364        if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4365                seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4366                if (modname[0])
4367                        seq_printf(m, " [%s]", modname);
4368                return;
4369        }
4370#endif
4371        seq_printf(m, "%p", (void *)address);
4372}
4373
4374static int leaks_show(struct seq_file *m, void *p)
4375{
4376        struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4377        struct slab *slabp;
4378        struct kmem_cache_node *n;
4379        const char *name;
4380        unsigned long *x = m->private;
4381        int node;
4382        int i;
4383
4384        if (!(cachep->flags & SLAB_STORE_USER))
4385                return 0;
4386        if (!(cachep->flags & SLAB_RED_ZONE))
4387                return 0;
4388
4389        /* OK, we can do it */
4390
4391        x[1] = 0;
4392
4393        for_each_online_node(node) {
4394                n = cachep->node[node];
4395                if (!n)
4396                        continue;
4397
4398                check_irq_on();
4399                spin_lock_irq(&n->list_lock);
4400
4401                list_for_each_entry(slabp, &n->slabs_full, list)
4402                        handle_slab(x, cachep, slabp);
4403                list_for_each_entry(slabp, &n->slabs_partial, list)
4404                        handle_slab(x, cachep, slabp);
4405                spin_unlock_irq(&n->list_lock);
4406        }
4407        name = cachep->name;
4408        if (x[0] == x[1]) {
4409                /* Increase the buffer size */
4410                mutex_unlock(&slab_mutex);
4411                m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4412                if (!m->private) {
4413                        /* Too bad, we are really out */
4414                        m->private = x;
4415                        mutex_lock(&slab_mutex);
4416                        return -ENOMEM;
4417                }
4418                *(unsigned long *)m->private = x[0] * 2;
4419                kfree(x);
4420                mutex_lock(&slab_mutex);
4421                /* Now make sure this entry will be retried */
4422                m->count = m->size;
4423                return 0;
4424        }
4425        for (i = 0; i < x[1]; i++) {
4426                seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4427                show_symbol(m, x[2*i+2]);
4428                seq_putc(m, '\n');
4429        }
4430
4431        return 0;
4432}
4433
4434static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4435{
4436        return seq_list_next(p, &slab_caches, pos);
4437}
4438
4439static void s_stop(struct seq_file *m, void *p)
4440{
4441        mutex_unlock(&slab_mutex);
4442}
4443
4444static const struct seq_operations slabstats_op = {
4445        .start = leaks_start,
4446        .next = s_next,
4447        .stop = s_stop,
4448        .show = leaks_show,
4449};
4450
4451static int slabstats_open(struct inode *inode, struct file *file)
4452{
4453        unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4454        int ret = -ENOMEM;
4455        if (n) {
4456                ret = seq_open(file, &slabstats_op);
4457                if (!ret) {
4458                        struct seq_file *m = file->private_data;
4459                        *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4460                        m->private = n;
4461                        n = NULL;
4462                }
4463                kfree(n);
4464        }
4465        return ret;
4466}
4467
4468static const struct file_operations proc_slabstats_operations = {
4469        .open           = slabstats_open,
4470        .read           = seq_read,
4471        .llseek         = seq_lseek,
4472        .release        = seq_release_private,
4473};
4474#endif
4475
4476static int __init slab_proc_init(void)
4477{
4478#ifdef CONFIG_DEBUG_SLAB_LEAK
4479        proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4480#endif
4481        return 0;
4482}
4483module_init(slab_proc_init);
4484#endif
4485
4486/**
4487 * ksize - get the actual amount of memory allocated for a given object
4488 * @objp: Pointer to the object
4489 *
4490 * kmalloc may internally round up allocations and return more memory
4491 * than requested. ksize() can be used to determine the actual amount of
4492 * memory allocated. The caller may use this additional memory, even though
4493 * a smaller amount of memory was initially specified with the kmalloc call.
4494 * The caller must guarantee that objp points to a valid object previously
4495 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4496 * must not be freed during the duration of the call.
4497 */
4498size_t ksize(const void *objp)
4499{
4500        BUG_ON(!objp);
4501        if (unlikely(objp == ZERO_SIZE_PTR))
4502                return 0;
4503
4504        return virt_to_cache(objp)->object_size;
4505}
4506EXPORT_SYMBOL(ksize);
4507
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.