linux/mm/slub.c
<<
>>
Prefs
   1/*
   2 * SLUB: A slab allocator that limits cache line use instead of queuing
   3 * objects in per cpu and per node lists.
   4 *
   5 * The allocator synchronizes using per slab locks or atomic operatios
   6 * and only uses a centralized lock to manage a pool of partial slabs.
   7 *
   8 * (C) 2007 SGI, Christoph Lameter
   9 * (C) 2011 Linux Foundation, Christoph Lameter
  10 */
  11
  12#include <linux/mm.h>
  13#include <linux/swap.h> /* struct reclaim_state */
  14#include <linux/module.h>
  15#include <linux/bit_spinlock.h>
  16#include <linux/interrupt.h>
  17#include <linux/bitops.h>
  18#include <linux/slab.h>
  19#include <linux/proc_fs.h>
  20#include <linux/seq_file.h>
  21#include <linux/kmemcheck.h>
  22#include <linux/cpu.h>
  23#include <linux/cpuset.h>
  24#include <linux/mempolicy.h>
  25#include <linux/ctype.h>
  26#include <linux/debugobjects.h>
  27#include <linux/kallsyms.h>
  28#include <linux/memory.h>
  29#include <linux/math64.h>
  30#include <linux/fault-inject.h>
  31#include <linux/stacktrace.h>
  32#include <linux/prefetch.h>
  33
  34#include <trace/events/kmem.h>
  35
  36/*
  37 * Lock order:
  38 *   1. slub_lock (Global Semaphore)
  39 *   2. node->list_lock
  40 *   3. slab_lock(page) (Only on some arches and for debugging)
  41 *
  42 *   slub_lock
  43 *
  44 *   The role of the slub_lock is to protect the list of all the slabs
  45 *   and to synchronize major metadata changes to slab cache structures.
  46 *
  47 *   The slab_lock is only used for debugging and on arches that do not
  48 *   have the ability to do a cmpxchg_double. It only protects the second
  49 *   double word in the page struct. Meaning
  50 *      A. page->freelist       -> List of object free in a page
  51 *      B. page->counters       -> Counters of objects
  52 *      C. page->frozen         -> frozen state
  53 *
  54 *   If a slab is frozen then it is exempt from list management. It is not
  55 *   on any list. The processor that froze the slab is the one who can
  56 *   perform list operations on the page. Other processors may put objects
  57 *   onto the freelist but the processor that froze the slab is the only
  58 *   one that can retrieve the objects from the page's freelist.
  59 *
  60 *   The list_lock protects the partial and full list on each node and
  61 *   the partial slab counter. If taken then no new slabs may be added or
  62 *   removed from the lists nor make the number of partial slabs be modified.
  63 *   (Note that the total number of slabs is an atomic value that may be
  64 *   modified without taking the list lock).
  65 *
  66 *   The list_lock is a centralized lock and thus we avoid taking it as
  67 *   much as possible. As long as SLUB does not have to handle partial
  68 *   slabs, operations can continue without any centralized lock. F.e.
  69 *   allocating a long series of objects that fill up slabs does not require
  70 *   the list lock.
  71 *   Interrupts are disabled during allocation and deallocation in order to
  72 *   make the slab allocator safe to use in the context of an irq. In addition
  73 *   interrupts are disabled to ensure that the processor does not change
  74 *   while handling per_cpu slabs, due to kernel preemption.
  75 *
  76 * SLUB assigns one slab for allocation to each processor.
  77 * Allocations only occur from these slabs called cpu slabs.
  78 *
  79 * Slabs with free elements are kept on a partial list and during regular
  80 * operations no list for full slabs is used. If an object in a full slab is
  81 * freed then the slab will show up again on the partial lists.
  82 * We track full slabs for debugging purposes though because otherwise we
  83 * cannot scan all objects.
  84 *
  85 * Slabs are freed when they become empty. Teardown and setup is
  86 * minimal so we rely on the page allocators per cpu caches for
  87 * fast frees and allocs.
  88 *
  89 * Overloading of page flags that are otherwise used for LRU management.
  90 *
  91 * PageActive           The slab is frozen and exempt from list processing.
  92 *                      This means that the slab is dedicated to a purpose
  93 *                      such as satisfying allocations for a specific
  94 *                      processor. Objects may be freed in the slab while
  95 *                      it is frozen but slab_free will then skip the usual
  96 *                      list operations. It is up to the processor holding
  97 *                      the slab to integrate the slab into the slab lists
  98 *                      when the slab is no longer needed.
  99 *
 100 *                      One use of this flag is to mark slabs that are
 101 *                      used for allocations. Then such a slab becomes a cpu
 102 *                      slab. The cpu slab may be equipped with an additional
 103 *                      freelist that allows lockless access to
 104 *                      free objects in addition to the regular freelist
 105 *                      that requires the slab lock.
 106 *
 107 * PageError            Slab requires special handling due to debug
 108 *                      options set. This moves slab handling out of
 109 *                      the fast path and disables lockless freelists.
 110 */
 111
 112#define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
 113                SLAB_TRACE | SLAB_DEBUG_FREE)
 114
 115static inline int kmem_cache_debug(struct kmem_cache *s)
 116{
 117#ifdef CONFIG_SLUB_DEBUG
 118        return unlikely(s->flags & SLAB_DEBUG_FLAGS);
 119#else
 120        return 0;
 121#endif
 122}
 123
 124/*
 125 * Issues still to be resolved:
 126 *
 127 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
 128 *
 129 * - Variable sizing of the per node arrays
 130 */
 131
 132/* Enable to test recovery from slab corruption on boot */
 133#undef SLUB_RESILIENCY_TEST
 134
 135/* Enable to log cmpxchg failures */
 136#undef SLUB_DEBUG_CMPXCHG
 137
 138/*
 139 * Mininum number of partial slabs. These will be left on the partial
 140 * lists even if they are empty. kmem_cache_shrink may reclaim them.
 141 */
 142#define MIN_PARTIAL 5
 143
 144/*
 145 * Maximum number of desirable partial slabs.
 146 * The existence of more partial slabs makes kmem_cache_shrink
 147 * sort the partial list by the number of objects in the.
 148 */
 149#define MAX_PARTIAL 10
 150
 151#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
 152                                SLAB_POISON | SLAB_STORE_USER)
 153
 154/*
 155 * Debugging flags that require metadata to be stored in the slab.  These get
 156 * disabled when slub_debug=O is used and a cache's min order increases with
 157 * metadata.
 158 */
 159#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
 160
 161/*
 162 * Set of flags that will prevent slab merging
 163 */
 164#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
 165                SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
 166                SLAB_FAILSLAB)
 167
 168#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
 169                SLAB_CACHE_DMA | SLAB_NOTRACK)
 170
 171#define OO_SHIFT        16
 172#define OO_MASK         ((1 << OO_SHIFT) - 1)
 173#define MAX_OBJS_PER_PAGE       32767 /* since page.objects is u15 */
 174
 175/* Internal SLUB flags */
 176#define __OBJECT_POISON         0x80000000UL /* Poison object */
 177#define __CMPXCHG_DOUBLE        0x40000000UL /* Use cmpxchg_double */
 178
 179static int kmem_size = sizeof(struct kmem_cache);
 180
 181#ifdef CONFIG_SMP
 182static struct notifier_block slab_notifier;
 183#endif
 184
 185static enum {
 186        DOWN,           /* No slab functionality available */
 187        PARTIAL,        /* Kmem_cache_node works */
 188        UP,             /* Everything works but does not show up in sysfs */
 189        SYSFS           /* Sysfs up */
 190} slab_state = DOWN;
 191
 192/* A list of all slab caches on the system */
 193static DECLARE_RWSEM(slub_lock);
 194static LIST_HEAD(slab_caches);
 195
 196/*
 197 * Tracking user of a slab.
 198 */
 199#define TRACK_ADDRS_COUNT 16
 200struct track {
 201        unsigned long addr;     /* Called from address */
 202#ifdef CONFIG_STACKTRACE
 203        unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
 204#endif
 205        int cpu;                /* Was running on cpu */
 206        int pid;                /* Pid context */
 207        unsigned long when;     /* When did the operation occur */
 208};
 209
 210enum track_item { TRACK_ALLOC, TRACK_FREE };
 211
 212#ifdef CONFIG_SYSFS
 213static int sysfs_slab_add(struct kmem_cache *);
 214static int sysfs_slab_alias(struct kmem_cache *, const char *);
 215static void sysfs_slab_remove(struct kmem_cache *);
 216
 217#else
 218static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
 219static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
 220                                                        { return 0; }
 221static inline void sysfs_slab_remove(struct kmem_cache *s)
 222{
 223        kfree(s->name);
 224        kfree(s);
 225}
 226
 227#endif
 228
 229static inline void stat(const struct kmem_cache *s, enum stat_item si)
 230{
 231#ifdef CONFIG_SLUB_STATS
 232        __this_cpu_inc(s->cpu_slab->stat[si]);
 233#endif
 234}
 235
 236/********************************************************************
 237 *                      Core slab cache functions
 238 *******************************************************************/
 239
 240int slab_is_available(void)
 241{
 242        return slab_state >= UP;
 243}
 244
 245static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
 246{
 247        return s->node[node];
 248}
 249
 250/* Verify that a pointer has an address that is valid within a slab page */
 251static inline int check_valid_pointer(struct kmem_cache *s,
 252                                struct page *page, const void *object)
 253{
 254        void *base;
 255
 256        if (!object)
 257                return 1;
 258
 259        base = page_address(page);
 260        if (object < base || object >= base + page->objects * s->size ||
 261                (object - base) % s->size) {
 262                return 0;
 263        }
 264
 265        return 1;
 266}
 267
 268static inline void *get_freepointer(struct kmem_cache *s, void *object)
 269{
 270        return *(void **)(object + s->offset);
 271}
 272
 273static void prefetch_freepointer(const struct kmem_cache *s, void *object)
 274{
 275        prefetch(object + s->offset);
 276}
 277
 278static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
 279{
 280        void *p;
 281
 282#ifdef CONFIG_DEBUG_PAGEALLOC
 283        probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
 284#else
 285        p = get_freepointer(s, object);
 286#endif
 287        return p;
 288}
 289
 290static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
 291{
 292        *(void **)(object + s->offset) = fp;
 293}
 294
 295/* Loop over all objects in a slab */
 296#define for_each_object(__p, __s, __addr, __objects) \
 297        for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
 298                        __p += (__s)->size)
 299
 300/* Determine object index from a given position */
 301static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
 302{
 303        return (p - addr) / s->size;
 304}
 305
 306static inline size_t slab_ksize(const struct kmem_cache *s)
 307{
 308#ifdef CONFIG_SLUB_DEBUG
 309        /*
 310         * Debugging requires use of the padding between object
 311         * and whatever may come after it.
 312         */
 313        if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
 314                return s->objsize;
 315
 316#endif
 317        /*
 318         * If we have the need to store the freelist pointer
 319         * back there or track user information then we can
 320         * only use the space before that information.
 321         */
 322        if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
 323                return s->inuse;
 324        /*
 325         * Else we can use all the padding etc for the allocation
 326         */
 327        return s->size;
 328}
 329
 330static inline int order_objects(int order, unsigned long size, int reserved)
 331{
 332        return ((PAGE_SIZE << order) - reserved) / size;
 333}
 334
 335static inline struct kmem_cache_order_objects oo_make(int order,
 336                unsigned long size, int reserved)
 337{
 338        struct kmem_cache_order_objects x = {
 339                (order << OO_SHIFT) + order_objects(order, size, reserved)
 340        };
 341
 342        return x;
 343}
 344
 345static inline int oo_order(struct kmem_cache_order_objects x)
 346{
 347        return x.x >> OO_SHIFT;
 348}
 349
 350static inline int oo_objects(struct kmem_cache_order_objects x)
 351{
 352        return x.x & OO_MASK;
 353}
 354
 355/*
 356 * Per slab locking using the pagelock
 357 */
 358static __always_inline void slab_lock(struct page *page)
 359{
 360        bit_spin_lock(PG_locked, &page->flags);
 361}
 362
 363static __always_inline void slab_unlock(struct page *page)
 364{
 365        __bit_spin_unlock(PG_locked, &page->flags);
 366}
 367
 368/* Interrupts must be disabled (for the fallback code to work right) */
 369static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
 370                void *freelist_old, unsigned long counters_old,
 371                void *freelist_new, unsigned long counters_new,
 372                const char *n)
 373{
 374        VM_BUG_ON(!irqs_disabled());
 375#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
 376    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
 377        if (s->flags & __CMPXCHG_DOUBLE) {
 378                if (cmpxchg_double(&page->freelist, &page->counters,
 379                        freelist_old, counters_old,
 380                        freelist_new, counters_new))
 381                return 1;
 382        } else
 383#endif
 384        {
 385                slab_lock(page);
 386                if (page->freelist == freelist_old && page->counters == counters_old) {
 387                        page->freelist = freelist_new;
 388                        page->counters = counters_new;
 389                        slab_unlock(page);
 390                        return 1;
 391                }
 392                slab_unlock(page);
 393        }
 394
 395        cpu_relax();
 396        stat(s, CMPXCHG_DOUBLE_FAIL);
 397
 398#ifdef SLUB_DEBUG_CMPXCHG
 399        printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
 400#endif
 401
 402        return 0;
 403}
 404
 405static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
 406                void *freelist_old, unsigned long counters_old,
 407                void *freelist_new, unsigned long counters_new,
 408                const char *n)
 409{
 410#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
 411    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
 412        if (s->flags & __CMPXCHG_DOUBLE) {
 413                if (cmpxchg_double(&page->freelist, &page->counters,
 414                        freelist_old, counters_old,
 415                        freelist_new, counters_new))
 416                return 1;
 417        } else
 418#endif
 419        {
 420                unsigned long flags;
 421
 422                local_irq_save(flags);
 423                slab_lock(page);
 424                if (page->freelist == freelist_old && page->counters == counters_old) {
 425                        page->freelist = freelist_new;
 426                        page->counters = counters_new;
 427                        slab_unlock(page);
 428                        local_irq_restore(flags);
 429                        return 1;
 430                }
 431                slab_unlock(page);
 432                local_irq_restore(flags);
 433        }
 434
 435        cpu_relax();
 436        stat(s, CMPXCHG_DOUBLE_FAIL);
 437
 438#ifdef SLUB_DEBUG_CMPXCHG
 439        printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
 440#endif
 441
 442        return 0;
 443}
 444
 445#ifdef CONFIG_SLUB_DEBUG
 446/*
 447 * Determine a map of object in use on a page.
 448 *
 449 * Node listlock must be held to guarantee that the page does
 450 * not vanish from under us.
 451 */
 452static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
 453{
 454        void *p;
 455        void *addr = page_address(page);
 456
 457        for (p = page->freelist; p; p = get_freepointer(s, p))
 458                set_bit(slab_index(p, s, addr), map);
 459}
 460
 461/*
 462 * Debug settings:
 463 */
 464#ifdef CONFIG_SLUB_DEBUG_ON
 465static int slub_debug = DEBUG_DEFAULT_FLAGS;
 466#else
 467static int slub_debug;
 468#endif
 469
 470static char *slub_debug_slabs;
 471static int disable_higher_order_debug;
 472
 473/*
 474 * Object debugging
 475 */
 476static void print_section(char *text, u8 *addr, unsigned int length)
 477{
 478        print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
 479                        length, 1);
 480}
 481
 482static struct track *get_track(struct kmem_cache *s, void *object,
 483        enum track_item alloc)
 484{
 485        struct track *p;
 486
 487        if (s->offset)
 488                p = object + s->offset + sizeof(void *);
 489        else
 490                p = object + s->inuse;
 491
 492        return p + alloc;
 493}
 494
 495static void set_track(struct kmem_cache *s, void *object,
 496                        enum track_item alloc, unsigned long addr)
 497{
 498        struct track *p = get_track(s, object, alloc);
 499
 500        if (addr) {
 501#ifdef CONFIG_STACKTRACE
 502                struct stack_trace trace;
 503                int i;
 504
 505                trace.nr_entries = 0;
 506                trace.max_entries = TRACK_ADDRS_COUNT;
 507                trace.entries = p->addrs;
 508                trace.skip = 3;
 509                save_stack_trace(&trace);
 510
 511                /* See rant in lockdep.c */
 512                if (trace.nr_entries != 0 &&
 513                    trace.entries[trace.nr_entries - 1] == ULONG_MAX)
 514                        trace.nr_entries--;
 515
 516                for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
 517                        p->addrs[i] = 0;
 518#endif
 519                p->addr = addr;
 520                p->cpu = smp_processor_id();
 521                p->pid = current->pid;
 522                p->when = jiffies;
 523        } else
 524                memset(p, 0, sizeof(struct track));
 525}
 526
 527static void init_tracking(struct kmem_cache *s, void *object)
 528{
 529        if (!(s->flags & SLAB_STORE_USER))
 530                return;
 531
 532        set_track(s, object, TRACK_FREE, 0UL);
 533        set_track(s, object, TRACK_ALLOC, 0UL);
 534}
 535
 536static void print_track(const char *s, struct track *t)
 537{
 538        if (!t->addr)
 539                return;
 540
 541        printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
 542                s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
 543#ifdef CONFIG_STACKTRACE
 544        {
 545                int i;
 546                for (i = 0; i < TRACK_ADDRS_COUNT; i++)
 547                        if (t->addrs[i])
 548                                printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
 549                        else
 550                                break;
 551        }
 552#endif
 553}
 554
 555static void print_tracking(struct kmem_cache *s, void *object)
 556{
 557        if (!(s->flags & SLAB_STORE_USER))
 558                return;
 559
 560        print_track("Allocated", get_track(s, object, TRACK_ALLOC));
 561        print_track("Freed", get_track(s, object, TRACK_FREE));
 562}
 563
 564static void print_page_info(struct page *page)
 565{
 566        printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
 567                page, page->objects, page->inuse, page->freelist, page->flags);
 568
 569}
 570
 571static void slab_bug(struct kmem_cache *s, char *fmt, ...)
 572{
 573        va_list args;
 574        char buf[100];
 575
 576        va_start(args, fmt);
 577        vsnprintf(buf, sizeof(buf), fmt, args);
 578        va_end(args);
 579        printk(KERN_ERR "========================================"
 580                        "=====================================\n");
 581        printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
 582        printk(KERN_ERR "----------------------------------------"
 583                        "-------------------------------------\n\n");
 584}
 585
 586static void slab_fix(struct kmem_cache *s, char *fmt, ...)
 587{
 588        va_list args;
 589        char buf[100];
 590
 591        va_start(args, fmt);
 592        vsnprintf(buf, sizeof(buf), fmt, args);
 593        va_end(args);
 594        printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
 595}
 596
 597static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
 598{
 599        unsigned int off;       /* Offset of last byte */
 600        u8 *addr = page_address(page);
 601
 602        print_tracking(s, p);
 603
 604        print_page_info(page);
 605
 606        printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
 607                        p, p - addr, get_freepointer(s, p));
 608
 609        if (p > addr + 16)
 610                print_section("Bytes b4 ", p - 16, 16);
 611
 612        print_section("Object ", p, min_t(unsigned long, s->objsize,
 613                                PAGE_SIZE));
 614        if (s->flags & SLAB_RED_ZONE)
 615                print_section("Redzone ", p + s->objsize,
 616                        s->inuse - s->objsize);
 617
 618        if (s->offset)
 619                off = s->offset + sizeof(void *);
 620        else
 621                off = s->inuse;
 622
 623        if (s->flags & SLAB_STORE_USER)
 624                off += 2 * sizeof(struct track);
 625
 626        if (off != s->size)
 627                /* Beginning of the filler is the free pointer */
 628                print_section("Padding ", p + off, s->size - off);
 629
 630        dump_stack();
 631}
 632
 633static void object_err(struct kmem_cache *s, struct page *page,
 634                        u8 *object, char *reason)
 635{
 636        slab_bug(s, "%s", reason);
 637        print_trailer(s, page, object);
 638}
 639
 640static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
 641{
 642        va_list args;
 643        char buf[100];
 644
 645        va_start(args, fmt);
 646        vsnprintf(buf, sizeof(buf), fmt, args);
 647        va_end(args);
 648        slab_bug(s, "%s", buf);
 649        print_page_info(page);
 650        dump_stack();
 651}
 652
 653static void init_object(struct kmem_cache *s, void *object, u8 val)
 654{
 655        u8 *p = object;
 656
 657        if (s->flags & __OBJECT_POISON) {
 658                memset(p, POISON_FREE, s->objsize - 1);
 659                p[s->objsize - 1] = POISON_END;
 660        }
 661
 662        if (s->flags & SLAB_RED_ZONE)
 663                memset(p + s->objsize, val, s->inuse - s->objsize);
 664}
 665
 666static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
 667                                                void *from, void *to)
 668{
 669        slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
 670        memset(from, data, to - from);
 671}
 672
 673static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
 674                        u8 *object, char *what,
 675                        u8 *start, unsigned int value, unsigned int bytes)
 676{
 677        u8 *fault;
 678        u8 *end;
 679
 680        fault = memchr_inv(start, value, bytes);
 681        if (!fault)
 682                return 1;
 683
 684        end = start + bytes;
 685        while (end > fault && end[-1] == value)
 686                end--;
 687
 688        slab_bug(s, "%s overwritten", what);
 689        printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
 690                                        fault, end - 1, fault[0], value);
 691        print_trailer(s, page, object);
 692
 693        restore_bytes(s, what, value, fault, end);
 694        return 0;
 695}
 696
 697/*
 698 * Object layout:
 699 *
 700 * object address
 701 *      Bytes of the object to be managed.
 702 *      If the freepointer may overlay the object then the free
 703 *      pointer is the first word of the object.
 704 *
 705 *      Poisoning uses 0x6b (POISON_FREE) and the last byte is
 706 *      0xa5 (POISON_END)
 707 *
 708 * object + s->objsize
 709 *      Padding to reach word boundary. This is also used for Redzoning.
 710 *      Padding is extended by another word if Redzoning is enabled and
 711 *      objsize == inuse.
 712 *
 713 *      We fill with 0xbb (RED_INACTIVE) for inactive objects and with
 714 *      0xcc (RED_ACTIVE) for objects in use.
 715 *
 716 * object + s->inuse
 717 *      Meta data starts here.
 718 *
 719 *      A. Free pointer (if we cannot overwrite object on free)
 720 *      B. Tracking data for SLAB_STORE_USER
 721 *      C. Padding to reach required alignment boundary or at mininum
 722 *              one word if debugging is on to be able to detect writes
 723 *              before the word boundary.
 724 *
 725 *      Padding is done using 0x5a (POISON_INUSE)
 726 *
 727 * object + s->size
 728 *      Nothing is used beyond s->size.
 729 *
 730 * If slabcaches are merged then the objsize and inuse boundaries are mostly
 731 * ignored. And therefore no slab options that rely on these boundaries
 732 * may be used with merged slabcaches.
 733 */
 734
 735static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
 736{
 737        unsigned long off = s->inuse;   /* The end of info */
 738
 739        if (s->offset)
 740                /* Freepointer is placed after the object. */
 741                off += sizeof(void *);
 742
 743        if (s->flags & SLAB_STORE_USER)
 744                /* We also have user information there */
 745                off += 2 * sizeof(struct track);
 746
 747        if (s->size == off)
 748                return 1;
 749
 750        return check_bytes_and_report(s, page, p, "Object padding",
 751                                p + off, POISON_INUSE, s->size - off);
 752}
 753
 754/* Check the pad bytes at the end of a slab page */
 755static int slab_pad_check(struct kmem_cache *s, struct page *page)
 756{
 757        u8 *start;
 758        u8 *fault;
 759        u8 *end;
 760        int length;
 761        int remainder;
 762
 763        if (!(s->flags & SLAB_POISON))
 764                return 1;
 765
 766        start = page_address(page);
 767        length = (PAGE_SIZE << compound_order(page)) - s->reserved;
 768        end = start + length;
 769        remainder = length % s->size;
 770        if (!remainder)
 771                return 1;
 772
 773        fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
 774        if (!fault)
 775                return 1;
 776        while (end > fault && end[-1] == POISON_INUSE)
 777                end--;
 778
 779        slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
 780        print_section("Padding ", end - remainder, remainder);
 781
 782        restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
 783        return 0;
 784}
 785
 786static int check_object(struct kmem_cache *s, struct page *page,
 787                                        void *object, u8 val)
 788{
 789        u8 *p = object;
 790        u8 *endobject = object + s->objsize;
 791
 792        if (s->flags & SLAB_RED_ZONE) {
 793                if (!check_bytes_and_report(s, page, object, "Redzone",
 794                        endobject, val, s->inuse - s->objsize))
 795                        return 0;
 796        } else {
 797                if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
 798                        check_bytes_and_report(s, page, p, "Alignment padding",
 799                                endobject, POISON_INUSE, s->inuse - s->objsize);
 800                }
 801        }
 802
 803        if (s->flags & SLAB_POISON) {
 804                if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
 805                        (!check_bytes_and_report(s, page, p, "Poison", p,
 806                                        POISON_FREE, s->objsize - 1) ||
 807                         !check_bytes_and_report(s, page, p, "Poison",
 808                                p + s->objsize - 1, POISON_END, 1)))
 809                        return 0;
 810                /*
 811                 * check_pad_bytes cleans up on its own.
 812                 */
 813                check_pad_bytes(s, page, p);
 814        }
 815
 816        if (!s->offset && val == SLUB_RED_ACTIVE)
 817                /*
 818                 * Object and freepointer overlap. Cannot check
 819                 * freepointer while object is allocated.
 820                 */
 821                return 1;
 822
 823        /* Check free pointer validity */
 824        if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
 825                object_err(s, page, p, "Freepointer corrupt");
 826                /*
 827                 * No choice but to zap it and thus lose the remainder
 828                 * of the free objects in this slab. May cause
 829                 * another error because the object count is now wrong.
 830                 */
 831                set_freepointer(s, p, NULL);
 832                return 0;
 833        }
 834        return 1;
 835}
 836
 837static int check_slab(struct kmem_cache *s, struct page *page)
 838{
 839        int maxobj;
 840
 841        VM_BUG_ON(!irqs_disabled());
 842
 843        if (!PageSlab(page)) {
 844                slab_err(s, page, "Not a valid slab page");
 845                return 0;
 846        }
 847
 848        maxobj = order_objects(compound_order(page), s->size, s->reserved);
 849        if (page->objects > maxobj) {
 850                slab_err(s, page, "objects %u > max %u",
 851                        s->name, page->objects, maxobj);
 852                return 0;
 853        }
 854        if (page->inuse > page->objects) {
 855                slab_err(s, page, "inuse %u > max %u",
 856                        s->name, page->inuse, page->objects);
 857                return 0;
 858        }
 859        /* Slab_pad_check fixes things up after itself */
 860        slab_pad_check(s, page);
 861        return 1;
 862}
 863
 864/*
 865 * Determine if a certain object on a page is on the freelist. Must hold the
 866 * slab lock to guarantee that the chains are in a consistent state.
 867 */
 868static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
 869{
 870        int nr = 0;
 871        void *fp;
 872        void *object = NULL;
 873        unsigned long max_objects;
 874
 875        fp = page->freelist;
 876        while (fp && nr <= page->objects) {
 877                if (fp == search)
 878                        return 1;
 879                if (!check_valid_pointer(s, page, fp)) {
 880                        if (object) {
 881                                object_err(s, page, object,
 882                                        "Freechain corrupt");
 883                                set_freepointer(s, object, NULL);
 884                                break;
 885                        } else {
 886                                slab_err(s, page, "Freepointer corrupt");
 887                                page->freelist = NULL;
 888                                page->inuse = page->objects;
 889                                slab_fix(s, "Freelist cleared");
 890                                return 0;
 891                        }
 892                        break;
 893                }
 894                object = fp;
 895                fp = get_freepointer(s, object);
 896                nr++;
 897        }
 898
 899        max_objects = order_objects(compound_order(page), s->size, s->reserved);
 900        if (max_objects > MAX_OBJS_PER_PAGE)
 901                max_objects = MAX_OBJS_PER_PAGE;
 902
 903        if (page->objects != max_objects) {
 904                slab_err(s, page, "Wrong number of objects. Found %d but "
 905                        "should be %d", page->objects, max_objects);
 906                page->objects = max_objects;
 907                slab_fix(s, "Number of objects adjusted.");
 908        }
 909        if (page->inuse != page->objects - nr) {
 910                slab_err(s, page, "Wrong object count. Counter is %d but "
 911                        "counted were %d", page->inuse, page->objects - nr);
 912                page->inuse = page->objects - nr;
 913                slab_fix(s, "Object count adjusted.");
 914        }
 915        return search == NULL;
 916}
 917
 918static void trace(struct kmem_cache *s, struct page *page, void *object,
 919                                                                int alloc)
 920{
 921        if (s->flags & SLAB_TRACE) {
 922                printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
 923                        s->name,
 924                        alloc ? "alloc" : "free",
 925                        object, page->inuse,
 926                        page->freelist);
 927
 928                if (!alloc)
 929                        print_section("Object ", (void *)object, s->objsize);
 930
 931                dump_stack();
 932        }
 933}
 934
 935/*
 936 * Hooks for other subsystems that check memory allocations. In a typical
 937 * production configuration these hooks all should produce no code at all.
 938 */
 939static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
 940{
 941        flags &= gfp_allowed_mask;
 942        lockdep_trace_alloc(flags);
 943        might_sleep_if(flags & __GFP_WAIT);
 944
 945        return should_failslab(s->objsize, flags, s->flags);
 946}
 947
 948static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
 949{
 950        flags &= gfp_allowed_mask;
 951        kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
 952        kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
 953}
 954
 955static inline void slab_free_hook(struct kmem_cache *s, void *x)
 956{
 957        kmemleak_free_recursive(x, s->flags);
 958
 959        /*
 960         * Trouble is that we may no longer disable interupts in the fast path
 961         * So in order to make the debug calls that expect irqs to be
 962         * disabled we need to disable interrupts temporarily.
 963         */
 964#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
 965        {
 966                unsigned long flags;
 967
 968                local_irq_save(flags);
 969                kmemcheck_slab_free(s, x, s->objsize);
 970                debug_check_no_locks_freed(x, s->objsize);
 971                local_irq_restore(flags);
 972        }
 973#endif
 974        if (!(s->flags & SLAB_DEBUG_OBJECTS))
 975                debug_check_no_obj_freed(x, s->objsize);
 976}
 977
 978/*
 979 * Tracking of fully allocated slabs for debugging purposes.
 980 *
 981 * list_lock must be held.
 982 */
 983static void add_full(struct kmem_cache *s,
 984        struct kmem_cache_node *n, struct page *page)
 985{
 986        if (!(s->flags & SLAB_STORE_USER))
 987                return;
 988
 989        list_add(&page->lru, &n->full);
 990}
 991
 992/*
 993 * list_lock must be held.
 994 */
 995static void remove_full(struct kmem_cache *s, struct page *page)
 996{
 997        if (!(s->flags & SLAB_STORE_USER))
 998                return;
 999
1000        list_del(&page->lru);
1001}
1002
1003/* Tracking of the number of slabs for debugging purposes */
1004static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1005{
1006        struct kmem_cache_node *n = get_node(s, node);
1007
1008        return atomic_long_read(&n->nr_slabs);
1009}
1010
1011static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1012{
1013        return atomic_long_read(&n->nr_slabs);
1014}
1015
1016static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1017{
1018        struct kmem_cache_node *n = get_node(s, node);
1019
1020        /*
1021         * May be called early in order to allocate a slab for the
1022         * kmem_cache_node structure. Solve the chicken-egg
1023         * dilemma by deferring the increment of the count during
1024         * bootstrap (see early_kmem_cache_node_alloc).
1025         */
1026        if (n) {
1027                atomic_long_inc(&n->nr_slabs);
1028                atomic_long_add(objects, &n->total_objects);
1029        }
1030}
1031static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1032{
1033        struct kmem_cache_node *n = get_node(s, node);
1034
1035        atomic_long_dec(&n->nr_slabs);
1036        atomic_long_sub(objects, &n->total_objects);
1037}
1038
1039/* Object debug checks for alloc/free paths */
1040static void setup_object_debug(struct kmem_cache *s, struct page *page,
1041                                                                void *object)
1042{
1043        if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1044                return;
1045
1046        init_object(s, object, SLUB_RED_INACTIVE);
1047        init_tracking(s, object);
1048}
1049
1050static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1051                                        void *object, unsigned long addr)
1052{
1053        if (!check_slab(s, page))
1054                goto bad;
1055
1056        if (!check_valid_pointer(s, page, object)) {
1057                object_err(s, page, object, "Freelist Pointer check fails");
1058                goto bad;
1059        }
1060
1061        if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1062                goto bad;
1063
1064        /* Success perform special debug activities for allocs */
1065        if (s->flags & SLAB_STORE_USER)
1066                set_track(s, object, TRACK_ALLOC, addr);
1067        trace(s, page, object, 1);
1068        init_object(s, object, SLUB_RED_ACTIVE);
1069        return 1;
1070
1071bad:
1072        if (PageSlab(page)) {
1073                /*
1074                 * If this is a slab page then lets do the best we can
1075                 * to avoid issues in the future. Marking all objects
1076                 * as used avoids touching the remaining objects.
1077                 */
1078                slab_fix(s, "Marking all objects used");
1079                page->inuse = page->objects;
1080                page->freelist = NULL;
1081        }
1082        return 0;
1083}
1084
1085static noinline int free_debug_processing(struct kmem_cache *s,
1086                 struct page *page, void *object, unsigned long addr)
1087{
1088        unsigned long flags;
1089        int rc = 0;
1090
1091        local_irq_save(flags);
1092        slab_lock(page);
1093
1094        if (!check_slab(s, page))
1095                goto fail;
1096
1097        if (!check_valid_pointer(s, page, object)) {
1098                slab_err(s, page, "Invalid object pointer 0x%p", object);
1099                goto fail;
1100        }
1101
1102        if (on_freelist(s, page, object)) {
1103                object_err(s, page, object, "Object already free");
1104                goto fail;
1105        }
1106
1107        if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1108                goto out;
1109
1110        if (unlikely(s != page->slab)) {
1111                if (!PageSlab(page)) {
1112                        slab_err(s, page, "Attempt to free object(0x%p) "
1113                                "outside of slab", object);
1114                } else if (!page->slab) {
1115                        printk(KERN_ERR
1116                                "SLUB <none>: no slab for object 0x%p.\n",
1117                                                object);
1118                        dump_stack();
1119                } else
1120                        object_err(s, page, object,
1121                                        "page slab pointer corrupt.");
1122                goto fail;
1123        }
1124
1125        if (s->flags & SLAB_STORE_USER)
1126                set_track(s, object, TRACK_FREE, addr);
1127        trace(s, page, object, 0);
1128        init_object(s, object, SLUB_RED_INACTIVE);
1129        rc = 1;
1130out:
1131        slab_unlock(page);
1132        local_irq_restore(flags);
1133        return rc;
1134
1135fail:
1136        slab_fix(s, "Object at 0x%p not freed", object);
1137        goto out;
1138}
1139
1140static int __init setup_slub_debug(char *str)
1141{
1142        slub_debug = DEBUG_DEFAULT_FLAGS;
1143        if (*str++ != '=' || !*str)
1144                /*
1145                 * No options specified. Switch on full debugging.
1146                 */
1147                goto out;
1148
1149        if (*str == ',')
1150                /*
1151                 * No options but restriction on slabs. This means full
1152                 * debugging for slabs matching a pattern.
1153                 */
1154                goto check_slabs;
1155
1156        if (tolower(*str) == 'o') {
1157                /*
1158                 * Avoid enabling debugging on caches if its minimum order
1159                 * would increase as a result.
1160                 */
1161                disable_higher_order_debug = 1;
1162                goto out;
1163        }
1164
1165        slub_debug = 0;
1166        if (*str == '-')
1167                /*
1168                 * Switch off all debugging measures.
1169                 */
1170                goto out;
1171
1172        /*
1173         * Determine which debug features should be switched on
1174         */
1175        for (; *str && *str != ','; str++) {
1176                switch (tolower(*str)) {
1177                case 'f':
1178                        slub_debug |= SLAB_DEBUG_FREE;
1179                        break;
1180                case 'z':
1181                        slub_debug |= SLAB_RED_ZONE;
1182                        break;
1183                case 'p':
1184                        slub_debug |= SLAB_POISON;
1185                        break;
1186                case 'u':
1187                        slub_debug |= SLAB_STORE_USER;
1188                        break;
1189                case 't':
1190                        slub_debug |= SLAB_TRACE;
1191                        break;
1192                case 'a':
1193                        slub_debug |= SLAB_FAILSLAB;
1194                        break;
1195                default:
1196                        printk(KERN_ERR "slub_debug option '%c' "
1197                                "unknown. skipped\n", *str);
1198                }
1199        }
1200
1201check_slabs:
1202        if (*str == ',')
1203                slub_debug_slabs = str + 1;
1204out:
1205        return 1;
1206}
1207
1208__setup("slub_debug", setup_slub_debug);
1209
1210static unsigned long kmem_cache_flags(unsigned long objsize,
1211        unsigned long flags, const char *name,
1212        void (*ctor)(void *))
1213{
1214        /*
1215         * Enable debugging if selected on the kernel commandline.
1216         */
1217        if (slub_debug && (!slub_debug_slabs ||
1218                !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1219                flags |= slub_debug;
1220
1221        return flags;
1222}
1223#else
1224static inline void setup_object_debug(struct kmem_cache *s,
1225                        struct page *page, void *object) {}
1226
1227static inline int alloc_debug_processing(struct kmem_cache *s,
1228        struct page *page, void *object, unsigned long addr) { return 0; }
1229
1230static inline int free_debug_processing(struct kmem_cache *s,
1231        struct page *page, void *object, unsigned long addr) { return 0; }
1232
1233static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1234                        { return 1; }
1235static inline int check_object(struct kmem_cache *s, struct page *page,
1236                        void *object, u8 val) { return 1; }
1237static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1238                                        struct page *page) {}
1239static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1240static inline unsigned long kmem_cache_flags(unsigned long objsize,
1241        unsigned long flags, const char *name,
1242        void (*ctor)(void *))
1243{
1244        return flags;
1245}
1246#define slub_debug 0
1247
1248#define disable_higher_order_debug 0
1249
1250static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1251                                                        { return 0; }
1252static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1253                                                        { return 0; }
1254static inline void inc_slabs_node(struct kmem_cache *s, int node,
1255                                                        int objects) {}
1256static inline void dec_slabs_node(struct kmem_cache *s, int node,
1257                                                        int objects) {}
1258
1259static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1260                                                        { return 0; }
1261
1262static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1263                void *object) {}
1264
1265static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1266
1267#endif /* CONFIG_SLUB_DEBUG */
1268
1269/*
1270 * Slab allocation and freeing
1271 */
1272static inline struct page *alloc_slab_page(gfp_t flags, int node,
1273                                        struct kmem_cache_order_objects oo)
1274{
1275        int order = oo_order(oo);
1276
1277        flags |= __GFP_NOTRACK;
1278
1279        if (node == NUMA_NO_NODE)
1280                return alloc_pages(flags, order);
1281        else
1282                return alloc_pages_exact_node(node, flags, order);
1283}
1284
1285static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1286{
1287        struct page *page;
1288        struct kmem_cache_order_objects oo = s->oo;
1289        gfp_t alloc_gfp;
1290
1291        flags &= gfp_allowed_mask;
1292
1293        if (flags & __GFP_WAIT)
1294                local_irq_enable();
1295
1296        flags |= s->allocflags;
1297
1298        /*
1299         * Let the initial higher-order allocation fail under memory pressure
1300         * so we fall-back to the minimum order allocation.
1301         */
1302        alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1303
1304        page = alloc_slab_page(alloc_gfp, node, oo);
1305        if (unlikely(!page)) {
1306                oo = s->min;
1307                /*
1308                 * Allocation may have failed due to fragmentation.
1309                 * Try a lower order alloc if possible
1310                 */
1311                page = alloc_slab_page(flags, node, oo);
1312
1313                if (page)
1314                        stat(s, ORDER_FALLBACK);
1315        }
1316
1317        if (flags & __GFP_WAIT)
1318                local_irq_disable();
1319
1320        if (!page)
1321                return NULL;
1322
1323        if (kmemcheck_enabled
1324                && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1325                int pages = 1 << oo_order(oo);
1326
1327                kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1328
1329                /*
1330                 * Objects from caches that have a constructor don't get
1331                 * cleared when they're allocated, so we need to do it here.
1332                 */
1333                if (s->ctor)
1334                        kmemcheck_mark_uninitialized_pages(page, pages);
1335                else
1336                        kmemcheck_mark_unallocated_pages(page, pages);
1337        }
1338
1339        page->objects = oo_objects(oo);
1340        mod_zone_page_state(page_zone(page),
1341                (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1342                NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1343                1 << oo_order(oo));
1344
1345        return page;
1346}
1347
1348static void setup_object(struct kmem_cache *s, struct page *page,
1349                                void *object)
1350{
1351        setup_object_debug(s, page, object);
1352        if (unlikely(s->ctor))
1353                s->ctor(object);
1354}
1355
1356static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1357{
1358        struct page *page;
1359        void *start;
1360        void *last;
1361        void *p;
1362
1363        BUG_ON(flags & GFP_SLAB_BUG_MASK);
1364
1365        page = allocate_slab(s,
1366                flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1367        if (!page)
1368                goto out;
1369
1370        inc_slabs_node(s, page_to_nid(page), page->objects);
1371        page->slab = s;
1372        __SetPageSlab(page);
1373
1374        start = page_address(page);
1375
1376        if (unlikely(s->flags & SLAB_POISON))
1377                memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1378
1379        last = start;
1380        for_each_object(p, s, start, page->objects) {
1381                setup_object(s, page, last);
1382                set_freepointer(s, last, p);
1383                last = p;
1384        }
1385        setup_object(s, page, last);
1386        set_freepointer(s, last, NULL);
1387
1388        page->freelist = start;
1389        page->inuse = page->objects;
1390        page->frozen = 1;
1391out:
1392        return page;
1393}
1394
1395static void __free_slab(struct kmem_cache *s, struct page *page)
1396{
1397        int order = compound_order(page);
1398        int pages = 1 << order;
1399
1400        if (kmem_cache_debug(s)) {
1401                void *p;
1402
1403                slab_pad_check(s, page);
1404                for_each_object(p, s, page_address(page),
1405                                                page->objects)
1406                        check_object(s, page, p, SLUB_RED_INACTIVE);
1407        }
1408
1409        kmemcheck_free_shadow(page, compound_order(page));
1410
1411        mod_zone_page_state(page_zone(page),
1412                (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1413                NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1414                -pages);
1415
1416        __ClearPageSlab(page);
1417        reset_page_mapcount(page);
1418        if (current->reclaim_state)
1419                current->reclaim_state->reclaimed_slab += pages;
1420        __free_pages(page, order);
1421}
1422
1423#define need_reserve_slab_rcu                                           \
1424        (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1425
1426static void rcu_free_slab(struct rcu_head *h)
1427{
1428        struct page *page;
1429
1430        if (need_reserve_slab_rcu)
1431                page = virt_to_head_page(h);
1432        else
1433                page = container_of((struct list_head *)h, struct page, lru);
1434
1435        __free_slab(page->slab, page);
1436}
1437
1438static void free_slab(struct kmem_cache *s, struct page *page)
1439{
1440        if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1441                struct rcu_head *head;
1442
1443                if (need_reserve_slab_rcu) {
1444                        int order = compound_order(page);
1445                        int offset = (PAGE_SIZE << order) - s->reserved;
1446
1447                        VM_BUG_ON(s->reserved != sizeof(*head));
1448                        head = page_address(page) + offset;
1449                } else {
1450                        /*
1451                         * RCU free overloads the RCU head over the LRU
1452                         */
1453                        head = (void *)&page->lru;
1454                }
1455
1456                call_rcu(head, rcu_free_slab);
1457        } else
1458                __free_slab(s, page);
1459}
1460
1461static void discard_slab(struct kmem_cache *s, struct page *page)
1462{
1463        dec_slabs_node(s, page_to_nid(page), page->objects);
1464        free_slab(s, page);
1465}
1466
1467/*
1468 * Management of partially allocated slabs.
1469 *
1470 * list_lock must be held.
1471 */
1472static inline void add_partial(struct kmem_cache_node *n,
1473                                struct page *page, int tail)
1474{
1475        n->nr_partial++;
1476        if (tail == DEACTIVATE_TO_TAIL)
1477                list_add_tail(&page->lru, &n->partial);
1478        else
1479                list_add(&page->lru, &n->partial);
1480}
1481
1482/*
1483 * list_lock must be held.
1484 */
1485static inline void remove_partial(struct kmem_cache_node *n,
1486                                        struct page *page)
1487{
1488        list_del(&page->lru);
1489        n->nr_partial--;
1490}
1491
1492/*
1493 * Lock slab, remove from the partial list and put the object into the
1494 * per cpu freelist.
1495 *
1496 * Returns a list of objects or NULL if it fails.
1497 *
1498 * Must hold list_lock.
1499 */
1500static inline void *acquire_slab(struct kmem_cache *s,
1501                struct kmem_cache_node *n, struct page *page,
1502                int mode)
1503{
1504        void *freelist;
1505        unsigned long counters;
1506        struct page new;
1507
1508        /*
1509         * Zap the freelist and set the frozen bit.
1510         * The old freelist is the list of objects for the
1511         * per cpu allocation list.
1512         */
1513        do {
1514                freelist = page->freelist;
1515                counters = page->counters;
1516                new.counters = counters;
1517                if (mode) {
1518                        new.inuse = page->objects;
1519                        new.freelist = NULL;
1520                } else {
1521                        new.freelist = freelist;
1522                }
1523
1524                VM_BUG_ON(new.frozen);
1525                new.frozen = 1;
1526
1527        } while (!__cmpxchg_double_slab(s, page,
1528                        freelist, counters,
1529                        new.freelist, new.counters,
1530                        "lock and freeze"));
1531
1532        remove_partial(n, page);
1533        return freelist;
1534}
1535
1536static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1537
1538/*
1539 * Try to allocate a partial slab from a specific node.
1540 */
1541static void *get_partial_node(struct kmem_cache *s,
1542                struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1543{
1544        struct page *page, *page2;
1545        void *object = NULL;
1546
1547        /*
1548         * Racy check. If we mistakenly see no partial slabs then we
1549         * just allocate an empty slab. If we mistakenly try to get a
1550         * partial slab and there is none available then get_partials()
1551         * will return NULL.
1552         */
1553        if (!n || !n->nr_partial)
1554                return NULL;
1555
1556        spin_lock(&n->list_lock);
1557        list_for_each_entry_safe(page, page2, &n->partial, lru) {
1558                void *t = acquire_slab(s, n, page, object == NULL);
1559                int available;
1560
1561                if (!t)
1562                        break;
1563
1564                if (!object) {
1565                        c->page = page;
1566                        c->node = page_to_nid(page);
1567                        stat(s, ALLOC_FROM_PARTIAL);
1568                        object = t;
1569                        available =  page->objects - page->inuse;
1570                } else {
1571                        available = put_cpu_partial(s, page, 0);
1572                        stat(s, CPU_PARTIAL_NODE);
1573                }
1574                if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1575                        break;
1576
1577        }
1578        spin_unlock(&n->list_lock);
1579        return object;
1580}
1581
1582/*
1583 * Get a page from somewhere. Search in increasing NUMA distances.
1584 */
1585static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1586                struct kmem_cache_cpu *c)
1587{
1588#ifdef CONFIG_NUMA
1589        struct zonelist *zonelist;
1590        struct zoneref *z;
1591        struct zone *zone;
1592        enum zone_type high_zoneidx = gfp_zone(flags);
1593        void *object;
1594        unsigned int cpuset_mems_cookie;
1595
1596        /*
1597         * The defrag ratio allows a configuration of the tradeoffs between
1598         * inter node defragmentation and node local allocations. A lower
1599         * defrag_ratio increases the tendency to do local allocations
1600         * instead of attempting to obtain partial slabs from other nodes.
1601         *
1602         * If the defrag_ratio is set to 0 then kmalloc() always
1603         * returns node local objects. If the ratio is higher then kmalloc()
1604         * may return off node objects because partial slabs are obtained
1605         * from other nodes and filled up.
1606         *
1607         * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1608         * defrag_ratio = 1000) then every (well almost) allocation will
1609         * first attempt to defrag slab caches on other nodes. This means
1610         * scanning over all nodes to look for partial slabs which may be
1611         * expensive if we do it every time we are trying to find a slab
1612         * with available objects.
1613         */
1614        if (!s->remote_node_defrag_ratio ||
1615                        get_cycles() % 1024 > s->remote_node_defrag_ratio)
1616                return NULL;
1617
1618        do {
1619                cpuset_mems_cookie = get_mems_allowed();
1620                zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1621                for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1622                        struct kmem_cache_node *n;
1623
1624                        n = get_node(s, zone_to_nid(zone));
1625
1626                        if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1627                                        n->nr_partial > s->min_partial) {
1628                                object = get_partial_node(s, n, c);
1629                                if (object) {
1630                                        /*
1631                                         * Return the object even if
1632                                         * put_mems_allowed indicated that
1633                                         * the cpuset mems_allowed was
1634                                         * updated in parallel. It's a
1635                                         * harmless race between the alloc
1636                                         * and the cpuset update.
1637                                         */
1638                                        put_mems_allowed(cpuset_mems_cookie);
1639                                        return object;
1640                                }
1641                        }
1642                }
1643        } while (!put_mems_allowed(cpuset_mems_cookie));
1644#endif
1645        return NULL;
1646}
1647
1648/*
1649 * Get a partial page, lock it and return it.
1650 */
1651static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1652                struct kmem_cache_cpu *c)
1653{
1654        void *object;
1655        int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1656
1657        object = get_partial_node(s, get_node(s, searchnode), c);
1658        if (object || node != NUMA_NO_NODE)
1659                return object;
1660
1661        return get_any_partial(s, flags, c);
1662}
1663
1664#ifdef CONFIG_PREEMPT
1665/*
1666 * Calculate the next globally unique transaction for disambiguiation
1667 * during cmpxchg. The transactions start with the cpu number and are then
1668 * incremented by CONFIG_NR_CPUS.
1669 */
1670#define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
1671#else
1672/*
1673 * No preemption supported therefore also no need to check for
1674 * different cpus.
1675 */
1676#define TID_STEP 1
1677#endif
1678
1679static inline unsigned long next_tid(unsigned long tid)
1680{
1681        return tid + TID_STEP;
1682}
1683
1684static inline unsigned int tid_to_cpu(unsigned long tid)
1685{
1686        return tid % TID_STEP;
1687}
1688
1689static inline unsigned long tid_to_event(unsigned long tid)
1690{
1691        return tid / TID_STEP;
1692}
1693
1694static inline unsigned int init_tid(int cpu)
1695{
1696        return cpu;
1697}
1698
1699static inline void note_cmpxchg_failure(const char *n,
1700                const struct kmem_cache *s, unsigned long tid)
1701{
1702#ifdef SLUB_DEBUG_CMPXCHG
1703        unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1704
1705        printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1706
1707#ifdef CONFIG_PREEMPT
1708        if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1709                printk("due to cpu change %d -> %d\n",
1710                        tid_to_cpu(tid), tid_to_cpu(actual_tid));
1711        else
1712#endif
1713        if (tid_to_event(tid) != tid_to_event(actual_tid))
1714                printk("due to cpu running other code. Event %ld->%ld\n",
1715                        tid_to_event(tid), tid_to_event(actual_tid));
1716        else
1717                printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1718                        actual_tid, tid, next_tid(tid));
1719#endif
1720        stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1721}
1722
1723void init_kmem_cache_cpus(struct kmem_cache *s)
1724{
1725        int cpu;
1726
1727        for_each_possible_cpu(cpu)
1728                per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1729}
1730
1731/*
1732 * Remove the cpu slab
1733 */
1734static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1735{
1736        enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1737        struct page *page = c->page;
1738        struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1739        int lock = 0;
1740        enum slab_modes l = M_NONE, m = M_NONE;
1741        void *freelist;
1742        void *nextfree;
1743        int tail = DEACTIVATE_TO_HEAD;
1744        struct page new;
1745        struct page old;
1746
1747        if (page->freelist) {
1748                stat(s, DEACTIVATE_REMOTE_FREES);
1749                tail = DEACTIVATE_TO_TAIL;
1750        }
1751
1752        c->tid = next_tid(c->tid);
1753        c->page = NULL;
1754        freelist = c->freelist;
1755        c->freelist = NULL;
1756
1757        /*
1758         * Stage one: Free all available per cpu objects back
1759         * to the page freelist while it is still frozen. Leave the
1760         * last one.
1761         *
1762         * There is no need to take the list->lock because the page
1763         * is still frozen.
1764         */
1765        while (freelist && (nextfree = get_freepointer(s, freelist))) {
1766                void *prior;
1767                unsigned long counters;
1768
1769                do {
1770                        prior = page->freelist;
1771                        counters = page->counters;
1772                        set_freepointer(s, freelist, prior);
1773                        new.counters = counters;
1774                        new.inuse--;
1775                        VM_BUG_ON(!new.frozen);
1776
1777                } while (!__cmpxchg_double_slab(s, page,
1778                        prior, counters,
1779                        freelist, new.counters,
1780                        "drain percpu freelist"));
1781
1782                freelist = nextfree;
1783        }
1784
1785        /*
1786         * Stage two: Ensure that the page is unfrozen while the
1787         * list presence reflects the actual number of objects
1788         * during unfreeze.
1789         *
1790         * We setup the list membership and then perform a cmpxchg
1791         * with the count. If there is a mismatch then the page
1792         * is not unfrozen but the page is on the wrong list.
1793         *
1794         * Then we restart the process which may have to remove
1795         * the page from the list that we just put it on again
1796         * because the number of objects in the slab may have
1797         * changed.
1798         */
1799redo:
1800
1801        old.freelist = page->freelist;
1802        old.counters = page->counters;
1803        VM_BUG_ON(!old.frozen);
1804
1805        /* Determine target state of the slab */
1806        new.counters = old.counters;
1807        if (freelist) {
1808                new.inuse--;
1809                set_freepointer(s, freelist, old.freelist);
1810                new.freelist = freelist;
1811        } else
1812                new.freelist = old.freelist;
1813
1814        new.frozen = 0;
1815
1816        if (!new.inuse && n->nr_partial > s->min_partial)
1817                m = M_FREE;
1818        else if (new.freelist) {
1819                m = M_PARTIAL;
1820                if (!lock) {
1821                        lock = 1;
1822                        /*
1823                         * Taking the spinlock removes the possiblity
1824                         * that acquire_slab() will see a slab page that
1825                         * is frozen
1826                         */
1827                        spin_lock(&n->list_lock);
1828                }
1829        } else {
1830                m = M_FULL;
1831                if (kmem_cache_debug(s) && !lock) {
1832                        lock = 1;
1833                        /*
1834                         * This also ensures that the scanning of full
1835                         * slabs from diagnostic functions will not see
1836                         * any frozen slabs.
1837                         */
1838                        spin_lock(&n->list_lock);
1839                }
1840        }
1841
1842        if (l != m) {
1843
1844                if (l == M_PARTIAL)
1845
1846                        remove_partial(n, page);
1847
1848                else if (l == M_FULL)
1849
1850                        remove_full(s, page);
1851
1852                if (m == M_PARTIAL) {
1853
1854                        add_partial(n, page, tail);
1855                        stat(s, tail);
1856
1857                } else if (m == M_FULL) {
1858
1859                        stat(s, DEACTIVATE_FULL);
1860                        add_full(s, n, page);
1861
1862                }
1863        }
1864
1865        l = m;
1866        if (!__cmpxchg_double_slab(s, page,
1867                                old.freelist, old.counters,
1868                                new.freelist, new.counters,
1869                                "unfreezing slab"))
1870                goto redo;
1871
1872        if (lock)
1873                spin_unlock(&n->list_lock);
1874
1875        if (m == M_FREE) {
1876                stat(s, DEACTIVATE_EMPTY);
1877                discard_slab(s, page);
1878                stat(s, FREE_SLAB);
1879        }
1880}
1881
1882/* Unfreeze all the cpu partial slabs */
1883static void unfreeze_partials(struct kmem_cache *s)
1884{
1885        struct kmem_cache_node *n = NULL;
1886        struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1887        struct page *page, *discard_page = NULL;
1888
1889        while ((page = c->partial)) {
1890                enum slab_modes { M_PARTIAL, M_FREE };
1891                enum slab_modes l, m;
1892                struct page new;
1893                struct page old;
1894
1895                c->partial = page->next;
1896                l = M_FREE;
1897
1898                do {
1899
1900                        old.freelist = page->freelist;
1901                        old.counters = page->counters;
1902                        VM_BUG_ON(!old.frozen);
1903
1904                        new.counters = old.counters;
1905                        new.freelist = old.freelist;
1906
1907                        new.frozen = 0;
1908
1909                        if (!new.inuse && (!n || n->nr_partial > s->min_partial))
1910                                m = M_FREE;
1911                        else {
1912                                struct kmem_cache_node *n2 = get_node(s,
1913                                                        page_to_nid(page));
1914
1915                                m = M_PARTIAL;
1916                                if (n != n2) {
1917                                        if (n)
1918                                                spin_unlock(&n->list_lock);
1919
1920                                        n = n2;
1921                                        spin_lock(&n->list_lock);
1922                                }
1923                        }
1924
1925                        if (l != m) {
1926                                if (l == M_PARTIAL) {
1927                                        remove_partial(n, page);
1928                                        stat(s, FREE_REMOVE_PARTIAL);
1929                                } else {
1930                                        add_partial(n, page,
1931                                                DEACTIVATE_TO_TAIL);
1932                                        stat(s, FREE_ADD_PARTIAL);
1933                                }
1934
1935                                l = m;
1936                        }
1937
1938                } while (!cmpxchg_double_slab(s, page,
1939                                old.freelist, old.counters,
1940                                new.freelist, new.counters,
1941                                "unfreezing slab"));
1942
1943                if (m == M_FREE) {
1944                        page->next = discard_page;
1945                        discard_page = page;
1946                }
1947        }
1948
1949        if (n)
1950                spin_unlock(&n->list_lock);
1951
1952        while (discard_page) {
1953                page = discard_page;
1954                discard_page = discard_page->next;
1955
1956                stat(s, DEACTIVATE_EMPTY);
1957                discard_slab(s, page);
1958                stat(s, FREE_SLAB);
1959        }
1960}
1961
1962/*
1963 * Put a page that was just frozen (in __slab_free) into a partial page
1964 * slot if available. This is done without interrupts disabled and without
1965 * preemption disabled. The cmpxchg is racy and may put the partial page
1966 * onto a random cpus partial slot.
1967 *
1968 * If we did not find a slot then simply move all the partials to the
1969 * per node partial list.
1970 */
1971int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1972{
1973        struct page *oldpage;
1974        int pages;
1975        int pobjects;
1976
1977        do {
1978                pages = 0;
1979                pobjects = 0;
1980                oldpage = this_cpu_read(s->cpu_slab->partial);
1981
1982                if (oldpage) {
1983                        pobjects = oldpage->pobjects;
1984                        pages = oldpage->pages;
1985                        if (drain && pobjects > s->cpu_partial) {
1986                                unsigned long flags;
1987                                /*
1988                                 * partial array is full. Move the existing
1989                                 * set to the per node partial list.
1990                                 */
1991                                local_irq_save(flags);
1992                                unfreeze_partials(s);
1993                                local_irq_restore(flags);
1994                                pobjects = 0;
1995                                pages = 0;
1996                                stat(s, CPU_PARTIAL_DRAIN);
1997                        }
1998                }
1999
2000                pages++;
2001                pobjects += page->objects - page->inuse;
2002
2003                page->pages = pages;
2004                page->pobjects = pobjects;
2005                page->next = oldpage;
2006
2007        } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
2008        return pobjects;
2009}
2010
2011static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2012{
2013        stat(s, CPUSLAB_FLUSH);
2014        deactivate_slab(s, c);
2015}
2016
2017/*
2018 * Flush cpu slab.
2019 *
2020 * Called from IPI handler with interrupts disabled.
2021 */
2022static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2023{
2024        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2025
2026        if (likely(c)) {
2027                if (c->page)
2028                        flush_slab(s, c);
2029
2030                unfreeze_partials(s);
2031        }
2032}
2033
2034static void flush_cpu_slab(void *d)
2035{
2036        struct kmem_cache *s = d;
2037
2038        __flush_cpu_slab(s, smp_processor_id());
2039}
2040
2041static bool has_cpu_slab(int cpu, void *info)
2042{
2043        struct kmem_cache *s = info;
2044        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2045
2046        return c->page || c->partial;
2047}
2048
2049static void flush_all(struct kmem_cache *s)
2050{
2051        on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2052}
2053
2054/*
2055 * Check if the objects in a per cpu structure fit numa
2056 * locality expectations.
2057 */
2058static inline int node_match(struct kmem_cache_cpu *c, int node)
2059{
2060#ifdef CONFIG_NUMA
2061        if (node != NUMA_NO_NODE && c->node != node)
2062                return 0;
2063#endif
2064        return 1;
2065}
2066
2067static int count_free(struct page *page)
2068{
2069        return page->objects - page->inuse;
2070}
2071
2072static unsigned long count_partial(struct kmem_cache_node *n,
2073                                        int (*get_count)(struct page *))
2074{
2075        unsigned long flags;
2076        unsigned long x = 0;
2077        struct page *page;
2078
2079        spin_lock_irqsave(&n->list_lock, flags);
2080        list_for_each_entry(page, &n->partial, lru)
2081                x += get_count(page);
2082        spin_unlock_irqrestore(&n->list_lock, flags);
2083        return x;
2084}
2085
2086static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2087{
2088#ifdef CONFIG_SLUB_DEBUG
2089        return atomic_long_read(&n->total_objects);
2090#else
2091        return 0;
2092#endif
2093}
2094
2095static noinline void
2096slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2097{
2098        int node;
2099
2100        printk(KERN_WARNING
2101                "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2102                nid, gfpflags);
2103        printk(KERN_WARNING "  cache: %s, object size: %d, buffer size: %d, "
2104                "default order: %d, min order: %d\n", s->name, s->objsize,
2105                s->size, oo_order(s->oo), oo_order(s->min));
2106
2107        if (oo_order(s->min) > get_order(s->objsize))
2108                printk(KERN_WARNING "  %s debugging increased min order, use "
2109                       "slub_debug=O to disable.\n", s->name);
2110
2111        for_each_online_node(node) {
2112                struct kmem_cache_node *n = get_node(s, node);
2113                unsigned long nr_slabs;
2114                unsigned long nr_objs;
2115                unsigned long nr_free;
2116
2117                if (!n)
2118                        continue;
2119
2120                nr_free  = count_partial(n, count_free);
2121                nr_slabs = node_nr_slabs(n);
2122                nr_objs  = node_nr_objs(n);
2123
2124                printk(KERN_WARNING
2125                        "  node %d: slabs: %ld, objs: %ld, free: %ld\n",
2126                        node, nr_slabs, nr_objs, nr_free);
2127        }
2128}
2129
2130static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2131                        int node, struct kmem_cache_cpu **pc)
2132{
2133        void *object;
2134        struct kmem_cache_cpu *c;
2135        struct page *page = new_slab(s, flags, node);
2136
2137        if (page) {
2138                c = __this_cpu_ptr(s->cpu_slab);
2139                if (c->page)
2140                        flush_slab(s, c);
2141
2142                /*
2143                 * No other reference to the page yet so we can
2144                 * muck around with it freely without cmpxchg
2145                 */
2146                object = page->freelist;
2147                page->freelist = NULL;
2148
2149                stat(s, ALLOC_SLAB);
2150                c->node = page_to_nid(page);
2151                c->page = page;
2152                *pc = c;
2153        } else
2154                object = NULL;
2155
2156        return object;
2157}
2158
2159/*
2160 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2161 * or deactivate the page.
2162 *
2163 * The page is still frozen if the return value is not NULL.
2164 *
2165 * If this function returns NULL then the page has been unfrozen.
2166 */
2167static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2168{
2169        struct page new;
2170        unsigned long counters;
2171        void *freelist;
2172
2173        do {
2174                freelist = page->freelist;
2175                counters = page->counters;
2176                new.counters = counters;
2177                VM_BUG_ON(!new.frozen);
2178
2179                new.inuse = page->objects;
2180                new.frozen = freelist != NULL;
2181
2182        } while (!cmpxchg_double_slab(s, page,
2183                freelist, counters,
2184                NULL, new.counters,
2185                "get_freelist"));
2186
2187        return freelist;
2188}
2189
2190/*
2191 * Slow path. The lockless freelist is empty or we need to perform
2192 * debugging duties.
2193 *
2194 * Processing is still very fast if new objects have been freed to the
2195 * regular freelist. In that case we simply take over the regular freelist
2196 * as the lockless freelist and zap the regular freelist.
2197 *
2198 * If that is not working then we fall back to the partial lists. We take the
2199 * first element of the freelist as the object to allocate now and move the
2200 * rest of the freelist to the lockless freelist.
2201 *
2202 * And if we were unable to get a new slab from the partial slab lists then
2203 * we need to allocate a new slab. This is the slowest path since it involves
2204 * a call to the page allocator and the setup of a new slab.
2205 */
2206static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2207                          unsigned long addr, struct kmem_cache_cpu *c)
2208{
2209        void **object;
2210        unsigned long flags;
2211
2212        local_irq_save(flags);
2213#ifdef CONFIG_PREEMPT
2214        /*
2215         * We may have been preempted and rescheduled on a different
2216         * cpu before disabling interrupts. Need to reload cpu area
2217         * pointer.
2218         */
2219        c = this_cpu_ptr(s->cpu_slab);
2220#endif
2221
2222        if (!c->page)
2223                goto new_slab;
2224redo:
2225        if (unlikely(!node_match(c, node))) {
2226                stat(s, ALLOC_NODE_MISMATCH);
2227                deactivate_slab(s, c);
2228                goto new_slab;
2229        }
2230
2231        /* must check again c->freelist in case of cpu migration or IRQ */
2232        object = c->freelist;
2233        if (object)
2234                goto load_freelist;
2235
2236        stat(s, ALLOC_SLOWPATH);
2237
2238        object = get_freelist(s, c->page);
2239
2240        if (!object) {
2241                c->page = NULL;
2242                stat(s, DEACTIVATE_BYPASS);
2243                goto new_slab;
2244        }
2245
2246        stat(s, ALLOC_REFILL);
2247
2248load_freelist:
2249        c->freelist = get_freepointer(s, object);
2250        c->tid = next_tid(c->tid);
2251        local_irq_restore(flags);
2252        return object;
2253
2254new_slab:
2255
2256        if (c->partial) {
2257                c->page = c->partial;
2258                c->partial = c->page->next;
2259                c->node = page_to_nid(c->page);
2260                stat(s, CPU_PARTIAL_ALLOC);
2261                c->freelist = NULL;
2262                goto redo;
2263        }
2264
2265        /* Then do expensive stuff like retrieving pages from the partial lists */
2266        object = get_partial(s, gfpflags, node, c);
2267
2268        if (unlikely(!object)) {
2269
2270                object = new_slab_objects(s, gfpflags, node, &c);
2271
2272                if (unlikely(!object)) {
2273                        if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2274                                slab_out_of_memory(s, gfpflags, node);
2275
2276                        local_irq_restore(flags);
2277                        return NULL;
2278                }
2279        }
2280
2281        if (likely(!kmem_cache_debug(s)))
2282                goto load_freelist;
2283
2284        /* Only entered in the debug case */
2285        if (!alloc_debug_processing(s, c->page, object, addr))
2286                goto new_slab;  /* Slab failed checks. Next slab needed */
2287
2288        c->freelist = get_freepointer(s, object);
2289        deactivate_slab(s, c);
2290        c->node = NUMA_NO_NODE;
2291        local_irq_restore(flags);
2292        return object;
2293}
2294
2295/*
2296 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2297 * have the fastpath folded into their functions. So no function call
2298 * overhead for requests that can be satisfied on the fastpath.
2299 *
2300 * The fastpath works by first checking if the lockless freelist can be used.
2301 * If not then __slab_alloc is called for slow processing.
2302 *
2303 * Otherwise we can simply pick the next object from the lockless free list.
2304 */
2305static __always_inline void *slab_alloc(struct kmem_cache *s,
2306                gfp_t gfpflags, int node, unsigned long addr)
2307{
2308        void **object;
2309        struct kmem_cache_cpu *c;
2310        unsigned long tid;
2311
2312        if (slab_pre_alloc_hook(s, gfpflags))
2313                return NULL;
2314
2315redo:
2316
2317        /*
2318         * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2319         * enabled. We may switch back and forth between cpus while
2320         * reading from one cpu area. That does not matter as long
2321         * as we end up on the original cpu again when doing the cmpxchg.
2322         */
2323        c = __this_cpu_ptr(s->cpu_slab);
2324
2325        /*
2326         * The transaction ids are globally unique per cpu and per operation on
2327         * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2328         * occurs on the right processor and that there was no operation on the
2329         * linked list in between.
2330         */
2331        tid = c->tid;
2332        barrier();
2333
2334        object = c->freelist;
2335        if (unlikely(!object || !node_match(c, node)))
2336
2337                object = __slab_alloc(s, gfpflags, node, addr, c);
2338
2339        else {
2340                void *next_object = get_freepointer_safe(s, object);
2341
2342                /*
2343                 * The cmpxchg will only match if there was no additional
2344                 * operation and if we are on the right processor.
2345                 *
2346                 * The cmpxchg does the following atomically (without lock semantics!)
2347                 * 1. Relocate first pointer to the current per cpu area.
2348                 * 2. Verify that tid and freelist have not been changed
2349                 * 3. If they were not changed replace tid and freelist
2350                 *
2351                 * Since this is without lock semantics the protection is only against
2352                 * code executing on this cpu *not* from access by other cpus.
2353                 */
2354                if (unlikely(!this_cpu_cmpxchg_double(
2355                                s->cpu_slab->freelist, s->cpu_slab->tid,
2356                                object, tid,
2357                                next_object, next_tid(tid)))) {
2358
2359                        note_cmpxchg_failure("slab_alloc", s, tid);
2360                        goto redo;
2361                }
2362                prefetch_freepointer(s, next_object);
2363                stat(s, ALLOC_FASTPATH);
2364        }
2365
2366        if (unlikely(gfpflags & __GFP_ZERO) && object)
2367                memset(object, 0, s->objsize);
2368
2369        slab_post_alloc_hook(s, gfpflags, object);
2370
2371        return object;
2372}
2373
2374void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2375{
2376        void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2377
2378        trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2379
2380        return ret;
2381}
2382EXPORT_SYMBOL(kmem_cache_alloc);
2383
2384#ifdef CONFIG_TRACING
2385void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2386{
2387        void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2388        trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2389        return ret;
2390}
2391EXPORT_SYMBOL(kmem_cache_alloc_trace);
2392
2393void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2394{
2395        void *ret = kmalloc_order(size, flags, order);
2396        trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2397        return ret;
2398}
2399EXPORT_SYMBOL(kmalloc_order_trace);
2400#endif
2401
2402#ifdef CONFIG_NUMA
2403void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2404{
2405        void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2406
2407        trace_kmem_cache_alloc_node(_RET_IP_, ret,
2408                                    s->objsize, s->size, gfpflags, node);
2409
2410        return ret;
2411}
2412EXPORT_SYMBOL(kmem_cache_alloc_node);
2413
2414#ifdef CONFIG_TRACING
2415void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2416                                    gfp_t gfpflags,
2417                                    int node, size_t size)
2418{
2419        void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2420
2421        trace_kmalloc_node(_RET_IP_, ret,
2422                           size, s->size, gfpflags, node);
2423        return ret;
2424}
2425EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2426#endif
2427#endif
2428
2429/*
2430 * Slow patch handling. This may still be called frequently since objects
2431 * have a longer lifetime than the cpu slabs in most processing loads.
2432 *
2433 * So we still attempt to reduce cache line usage. Just take the slab
2434 * lock and free the item. If there is no additional partial page
2435 * handling required then we can return immediately.
2436 */
2437static void __slab_free(struct kmem_cache *s, struct page *page,
2438                        void *x, unsigned long addr)
2439{
2440        void *prior;
2441        void **object = (void *)x;
2442        int was_frozen;
2443        int inuse;
2444        struct page new;
2445        unsigned long counters;
2446        struct kmem_cache_node *n = NULL;
2447        unsigned long uninitialized_var(flags);
2448
2449        stat(s, FREE_SLOWPATH);
2450
2451        if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2452                return;
2453
2454        do {
2455                prior = page->freelist;
2456                counters = page->counters;
2457                set_freepointer(s, object, prior);
2458                new.counters = counters;
2459                was_frozen = new.frozen;
2460                new.inuse--;
2461                if ((!new.inuse || !prior) && !was_frozen && !n) {
2462
2463                        if (!kmem_cache_debug(s) && !prior)
2464
2465                                /*
2466                                 * Slab was on no list before and will be partially empty
2467                                 * We can defer the list move and instead freeze it.
2468                                 */
2469                                new.frozen = 1;
2470
2471                        else { /* Needs to be taken off a list */
2472
2473                                n = get_node(s, page_to_nid(page));
2474                                /*
2475                                 * Speculatively acquire the list_lock.
2476                                 * If the cmpxchg does not succeed then we may
2477                                 * drop the list_lock without any processing.
2478                                 *
2479                                 * Otherwise the list_lock will synchronize with
2480                                 * other processors updating the list of slabs.
2481                                 */
2482                                spin_lock_irqsave(&n->list_lock, flags);
2483
2484                        }
2485                }
2486                inuse = new.inuse;
2487
2488        } while (!cmpxchg_double_slab(s, page,
2489                prior, counters,
2490                object, new.counters,
2491                "__slab_free"));
2492
2493        if (likely(!n)) {
2494
2495                /*
2496                 * If we just froze the page then put it onto the
2497                 * per cpu partial list.
2498                 */
2499                if (new.frozen && !was_frozen) {
2500                        put_cpu_partial(s, page, 1);
2501                        stat(s, CPU_PARTIAL_FREE);
2502                }
2503                /*
2504                 * The list lock was not taken therefore no list
2505                 * activity can be necessary.
2506                 */
2507                if (was_frozen)
2508                        stat(s, FREE_FROZEN);
2509                return;
2510        }
2511
2512        /*
2513         * was_frozen may have been set after we acquired the list_lock in
2514         * an earlier loop. So we need to check it here again.
2515         */
2516        if (was_frozen)
2517                stat(s, FREE_FROZEN);
2518        else {
2519                if (unlikely(!inuse && n->nr_partial > s->min_partial))
2520                        goto slab_empty;
2521
2522                /*
2523                 * Objects left in the slab. If it was not on the partial list before
2524                 * then add it.
2525                 */
2526                if (unlikely(!prior)) {
2527                        remove_full(s, page);
2528                        add_partial(n, page, DEACTIVATE_TO_TAIL);
2529                        stat(s, FREE_ADD_PARTIAL);
2530                }
2531        }
2532        spin_unlock_irqrestore(&n->list_lock, flags);
2533        return;
2534
2535slab_empty:
2536        if (prior) {
2537                /*
2538                 * Slab on the partial list.
2539                 */
2540                remove_partial(n, page);
2541                stat(s, FREE_REMOVE_PARTIAL);
2542        } else
2543                /* Slab must be on the full list */
2544                remove_full(s, page);
2545
2546        spin_unlock_irqrestore(&n->list_lock, flags);
2547        stat(s, FREE_SLAB);
2548        discard_slab(s, page);
2549}
2550
2551/*
2552 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2553 * can perform fastpath freeing without additional function calls.
2554 *
2555 * The fastpath is only possible if we are freeing to the current cpu slab
2556 * of this processor. This typically the case if we have just allocated
2557 * the item before.
2558 *
2559 * If fastpath is not possible then fall back to __slab_free where we deal
2560 * with all sorts of special processing.
2561 */
2562static __always_inline void slab_free(struct kmem_cache *s,
2563                        struct page *page, void *x, unsigned long addr)
2564{
2565        void **object = (void *)x;
2566        struct kmem_cache_cpu *c;
2567        unsigned long tid;
2568
2569        slab_free_hook(s, x);
2570
2571redo:
2572        /*
2573         * Determine the currently cpus per cpu slab.
2574         * The cpu may change afterward. However that does not matter since
2575         * data is retrieved via this pointer. If we are on the same cpu
2576         * during the cmpxchg then the free will succedd.
2577         */
2578        c = __this_cpu_ptr(s->cpu_slab);
2579
2580        tid = c->tid;
2581        barrier();
2582
2583        if (likely(page == c->page)) {
2584                set_freepointer(s, object, c->freelist);
2585
2586                if (unlikely(!this_cpu_cmpxchg_double(
2587                                s->cpu_slab->freelist, s->cpu_slab->tid,
2588                                c->freelist, tid,
2589                                object, next_tid(tid)))) {
2590
2591                        note_cmpxchg_failure("slab_free", s, tid);
2592                        goto redo;
2593                }
2594                stat(s, FREE_FASTPATH);
2595        } else
2596                __slab_free(s, page, x, addr);
2597
2598}
2599
2600void kmem_cache_free(struct kmem_cache *s, void *x)
2601{
2602        struct page *page;
2603
2604        page = virt_to_head_page(x);
2605
2606        slab_free(s, page, x, _RET_IP_);
2607
2608        trace_kmem_cache_free(_RET_IP_, x);
2609}
2610EXPORT_SYMBOL(kmem_cache_free);
2611
2612/*
2613 * Object placement in a slab is made very easy because we always start at
2614 * offset 0. If we tune the size of the object to the alignment then we can
2615 * get the required alignment by putting one properly sized object after
2616 * another.
2617 *
2618 * Notice that the allocation order determines the sizes of the per cpu
2619 * caches. Each processor has always one slab available for allocations.
2620 * Increasing the allocation order reduces the number of times that slabs
2621 * must be moved on and off the partial lists and is therefore a factor in
2622 * locking overhead.
2623 */
2624
2625/*
2626 * Mininum / Maximum order of slab pages. This influences locking overhead
2627 * and slab fragmentation. A higher order reduces the number of partial slabs
2628 * and increases the number of allocations possible without having to
2629 * take the list_lock.
2630 */
2631static int slub_min_order;
2632static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2633static int slub_min_objects;
2634
2635/*
2636 * Merge control. If this is set then no merging of slab caches will occur.
2637 * (Could be removed. This was introduced to pacify the merge skeptics.)
2638 */
2639static int slub_nomerge;
2640
2641/*
2642 * Calculate the order of allocation given an slab object size.
2643 *
2644 * The order of allocation has significant impact on performance and other
2645 * system components. Generally order 0 allocations should be preferred since
2646 * order 0 does not cause fragmentation in the page allocator. Larger objects
2647 * be problematic to put into order 0 slabs because there may be too much
2648 * unused space left. We go to a higher order if more than 1/16th of the slab
2649 * would be wasted.
2650 *
2651 * In order to reach satisfactory performance we must ensure that a minimum
2652 * number of objects is in one slab. Otherwise we may generate too much
2653 * activity on the partial lists which requires taking the list_lock. This is
2654 * less a concern for large slabs though which are rarely used.
2655 *
2656 * slub_max_order specifies the order where we begin to stop considering the
2657 * number of objects in a slab as critical. If we reach slub_max_order then
2658 * we try to keep the page order as low as possible. So we accept more waste
2659 * of space in favor of a small page order.
2660 *
2661 * Higher order allocations also allow the placement of more objects in a
2662 * slab and thereby reduce object handling overhead. If the user has
2663 * requested a higher mininum order then we start with that one instead of
2664 * the smallest order which will fit the object.
2665 */
2666static inline int slab_order(int size, int min_objects,
2667                                int max_order, int fract_leftover, int reserved)
2668{
2669        int order;
2670        int rem;
2671        int min_order = slub_min_order;
2672
2673        if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2674                return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2675
2676        for (order = max(min_order,
2677                                fls(min_objects * size - 1) - PAGE_SHIFT);
2678                        order <= max_order; order++) {
2679
2680                unsigned long slab_size = PAGE_SIZE << order;
2681
2682                if (slab_size < min_objects * size + reserved)
2683                        continue;
2684
2685                rem = (slab_size - reserved) % size;
2686
2687                if (rem <= slab_size / fract_leftover)
2688                        break;
2689
2690        }
2691
2692        return order;
2693}
2694
2695static inline int calculate_order(int size, int reserved)
2696{
2697        int order;
2698        int min_objects;
2699        int fraction;
2700        int max_objects;
2701
2702        /*
2703         * Attempt to find best configuration for a slab. This
2704         * works by first attempting to generate a layout with
2705         * the best configuration and backing off gradually.
2706         *
2707         * First we reduce the acceptable waste in a slab. Then
2708         * we reduce the minimum objects required in a slab.
2709         */
2710        min_objects = slub_min_objects;
2711        if (!min_objects)
2712                min_objects = 4 * (fls(nr_cpu_ids) + 1);
2713        max_objects = order_objects(slub_max_order, size, reserved);
2714        min_objects = min(min_objects, max_objects);
2715
2716        while (min_objects > 1) {
2717                fraction = 16;
2718                while (fraction >= 4) {
2719                        order = slab_order(size, min_objects,
2720                                        slub_max_order, fraction, reserved);
2721                        if (order <= slub_max_order)
2722                                return order;
2723                        fraction /= 2;
2724                }
2725                min_objects--;
2726        }
2727
2728        /*
2729         * We were unable to place multiple objects in a slab. Now
2730         * lets see if we can place a single object there.
2731         */
2732        order = slab_order(size, 1, slub_max_order, 1, reserved);
2733        if (order <= slub_max_order)
2734                return order;
2735
2736        /*
2737         * Doh this slab cannot be placed using slub_max_order.
2738         */
2739        order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2740        if (order < MAX_ORDER)
2741                return order;
2742        return -ENOSYS;
2743}
2744
2745/*
2746 * Figure out what the alignment of the objects will be.
2747 */
2748static unsigned long calculate_alignment(unsigned long flags,
2749                unsigned long align, unsigned long size)
2750{
2751        /*
2752         * If the user wants hardware cache aligned objects then follow that
2753         * suggestion if the object is sufficiently large.
2754         *
2755         * The hardware cache alignment cannot override the specified
2756         * alignment though. If that is greater then use it.
2757         */
2758        if (flags & SLAB_HWCACHE_ALIGN) {
2759                unsigned long ralign = cache_line_size();
2760                while (size <= ralign / 2)
2761                        ralign /= 2;
2762                align = max(align, ralign);
2763        }
2764
2765        if (align < ARCH_SLAB_MINALIGN)
2766                align = ARCH_SLAB_MINALIGN;
2767
2768        return ALIGN(align, sizeof(void *));
2769}
2770
2771static void
2772init_kmem_cache_node(struct kmem_cache_node *n)
2773{
2774        n->nr_partial = 0;
2775        spin_lock_init(&n->list_lock);
2776        INIT_LIST_HEAD(&n->partial);
2777#ifdef CONFIG_SLUB_DEBUG
2778        atomic_long_set(&n->nr_slabs, 0);
2779        atomic_long_set(&n->total_objects, 0);
2780        INIT_LIST_HEAD(&n->full);
2781#endif
2782}
2783
2784static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2785{
2786        BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2787                        SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2788
2789        /*
2790         * Must align to double word boundary for the double cmpxchg
2791         * instructions to work; see __pcpu_double_call_return_bool().
2792         */
2793        s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2794                                     2 * sizeof(void *));
2795
2796        if (!s->cpu_slab)
2797                return 0;
2798
2799        init_kmem_cache_cpus(s);
2800
2801        return 1;
2802}
2803
2804static struct kmem_cache *kmem_cache_node;
2805
2806/*
2807 * No kmalloc_node yet so do it by hand. We know that this is the first
2808 * slab on the node for this slabcache. There are no concurrent accesses
2809 * possible.
2810 *
2811 * Note that this function only works on the kmalloc_node_cache
2812 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2813 * memory on a fresh node that has no slab structures yet.
2814 */
2815static void early_kmem_cache_node_alloc(int node)
2816{
2817        struct page *page;
2818        struct kmem_cache_node *n;
2819
2820        BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2821
2822        page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2823
2824        BUG_ON(!page);
2825        if (page_to_nid(page) != node) {
2826                printk(KERN_ERR "SLUB: Unable to allocate memory from "
2827                                "node %d\n", node);
2828                printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2829                                "in order to be able to continue\n");
2830        }
2831
2832        n = page->freelist;
2833        BUG_ON(!n);
2834        page->freelist = get_freepointer(kmem_cache_node, n);
2835        page->inuse = 1;
2836        page->frozen = 0;
2837        kmem_cache_node->node[node] = n;
2838#ifdef CONFIG_SLUB_DEBUG
2839        init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2840        init_tracking(kmem_cache_node, n);
2841#endif
2842        init_kmem_cache_node(n);
2843        inc_slabs_node(kmem_cache_node, node, page->objects);
2844
2845        add_partial(n, page, DEACTIVATE_TO_HEAD);
2846}
2847
2848static void free_kmem_cache_nodes(struct kmem_cache *s)
2849{
2850        int node;
2851
2852        for_each_node_state(node, N_NORMAL_MEMORY) {
2853                struct kmem_cache_node *n = s->node[node];
2854
2855                if (n)
2856                        kmem_cache_free(kmem_cache_node, n);
2857
2858                s->node[node] = NULL;
2859        }
2860}
2861
2862static int init_kmem_cache_nodes(struct kmem_cache *s)
2863{
2864        int node;
2865
2866        for_each_node_state(node, N_NORMAL_MEMORY) {
2867                struct kmem_cache_node *n;
2868
2869                if (slab_state == DOWN) {
2870                        early_kmem_cache_node_alloc(node);
2871                        continue;
2872                }
2873                n = kmem_cache_alloc_node(kmem_cache_node,
2874                                                GFP_KERNEL, node);
2875
2876                if (!n) {
2877                        free_kmem_cache_nodes(s);
2878                        return 0;
2879                }
2880
2881                s->node[node] = n;
2882                init_kmem_cache_node(n);
2883        }
2884        return 1;
2885}
2886
2887static void set_min_partial(struct kmem_cache *s, unsigned long min)
2888{
2889        if (min < MIN_PARTIAL)
2890                min = MIN_PARTIAL;
2891        else if (min > MAX_PARTIAL)
2892                min = MAX_PARTIAL;
2893        s->min_partial = min;
2894}
2895
2896/*
2897 * calculate_sizes() determines the order and the distribution of data within
2898 * a slab object.
2899 */
2900static int calculate_sizes(struct kmem_cache *s, int forced_order)
2901{
2902        unsigned long flags = s->flags;
2903        unsigned long size = s->objsize;
2904        unsigned long align = s->align;
2905        int order;
2906
2907        /*
2908         * Round up object size to the next word boundary. We can only
2909         * place the free pointer at word boundaries and this determines
2910         * the possible location of the free pointer.
2911         */
2912        size = ALIGN(size, sizeof(void *));
2913
2914#ifdef CONFIG_SLUB_DEBUG
2915        /*
2916         * Determine if we can poison the object itself. If the user of
2917         * the slab may touch the object after free or before allocation
2918         * then we should never poison the object itself.
2919         */
2920        if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2921                        !s->ctor)
2922                s->flags |= __OBJECT_POISON;
2923        else
2924                s->flags &= ~__OBJECT_POISON;
2925
2926
2927        /*
2928         * If we are Redzoning then check if there is some space between the
2929         * end of the object and the free pointer. If not then add an
2930         * additional word to have some bytes to store Redzone information.
2931         */
2932        if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2933                size += sizeof(void *);
2934#endif
2935
2936        /*
2937         * With that we have determined the number of bytes in actual use
2938         * by the object. This is the potential offset to the free pointer.
2939         */
2940        s->inuse = size;
2941
2942        if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2943                s->ctor)) {
2944                /*
2945                 * Relocate free pointer after the object if it is not
2946                 * permitted to overwrite the first word of the object on
2947                 * kmem_cache_free.
2948                 *
2949                 * This is the case if we do RCU, have a constructor or
2950                 * destructor or are poisoning the objects.
2951                 */
2952                s->offset = size;
2953                size += sizeof(void *);
2954        }
2955
2956#ifdef CONFIG_SLUB_DEBUG
2957        if (flags & SLAB_STORE_USER)
2958                /*
2959                 * Need to store information about allocs and frees after
2960                 * the object.
2961                 */
2962                size += 2 * sizeof(struct track);
2963
2964        if (flags & SLAB_RED_ZONE)
2965                /*
2966                 * Add some empty padding so that we can catch
2967                 * overwrites from earlier objects rather than let
2968                 * tracking information or the free pointer be
2969                 * corrupted if a user writes before the start
2970                 * of the object.
2971                 */
2972                size += sizeof(void *);
2973#endif
2974
2975        /*
2976         * Determine the alignment based on various parameters that the
2977         * user specified and the dynamic determination of cache line size
2978         * on bootup.
2979         */
2980        align = calculate_alignment(flags, align, s->objsize);
2981        s->align = align;
2982
2983        /*
2984         * SLUB stores one object immediately after another beginning from
2985         * offset 0. In order to align the objects we have to simply size
2986         * each object to conform to the alignment.
2987         */
2988        size = ALIGN(size, align);
2989        s->size = size;
2990        if (forced_order >= 0)
2991                order = forced_order;
2992        else
2993                order = calculate_order(size, s->reserved);
2994
2995        if (order < 0)
2996                return 0;
2997
2998        s->allocflags = 0;
2999        if (order)
3000                s->allocflags |= __GFP_COMP;
3001
3002        if (s->flags & SLAB_CACHE_DMA)
3003                s->allocflags |= SLUB_DMA;
3004
3005        if (s->flags & SLAB_RECLAIM_ACCOUNT)
3006                s->allocflags |= __GFP_RECLAIMABLE;
3007
3008        /*
3009         * Determine the number of objects per slab
3010         */
3011        s->oo = oo_make(order, size, s->reserved);
3012        s->min = oo_make(get_order(size), size, s->reserved);
3013        if (oo_objects(s->oo) > oo_objects(s->max))
3014                s->max = s->oo;
3015
3016        return !!oo_objects(s->oo);
3017
3018}
3019
3020static int kmem_cache_open(struct kmem_cache *s,
3021                const char *name, size_t size,
3022                size_t align, unsigned long flags,
3023                void (*ctor)(void *))
3024{
3025        memset(s, 0, kmem_size);
3026        s->name = name;
3027        s->ctor = ctor;
3028        s->objsize = size;
3029        s->align = align;
3030        s->flags = kmem_cache_flags(size, flags, name, ctor);
3031        s->reserved = 0;
3032
3033        if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3034                s->reserved = sizeof(struct rcu_head);
3035
3036        if (!calculate_sizes(s, -1))
3037                goto error;
3038        if (disable_higher_order_debug) {
3039                /*
3040                 * Disable debugging flags that store metadata if the min slab
3041                 * order increased.
3042                 */
3043                if (get_order(s->size) > get_order(s->objsize)) {
3044                        s->flags &= ~DEBUG_METADATA_FLAGS;
3045                        s->offset = 0;
3046                        if (!calculate_sizes(s, -1))
3047                                goto error;
3048                }
3049        }
3050
3051#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3052    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3053        if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3054                /* Enable fast mode */
3055                s->flags |= __CMPXCHG_DOUBLE;
3056#endif
3057
3058        /*
3059         * The larger the object size is, the more pages we want on the partial
3060         * list to avoid pounding the page allocator excessively.
3061         */
3062        set_min_partial(s, ilog2(s->size) / 2);
3063
3064        /*
3065         * cpu_partial determined the maximum number of objects kept in the
3066         * per cpu partial lists of a processor.
3067         *
3068         * Per cpu partial lists mainly contain slabs that just have one
3069         * object freed. If they are used for allocation then they can be
3070         * filled up again with minimal effort. The slab will never hit the
3071         * per node partial lists and therefore no locking will be required.
3072         *
3073         * This setting also determines
3074         *
3075         * A) The number of objects from per cpu partial slabs dumped to the
3076         *    per node list when we reach the limit.
3077         * B) The number of objects in cpu partial slabs to extract from the
3078         *    per node list when we run out of per cpu objects. We only fetch 50%
3079         *    to keep some capacity around for frees.
3080         */
3081        if (kmem_cache_debug(s))
3082                s->cpu_partial = 0;
3083        else if (s->size >= PAGE_SIZE)
3084                s->cpu_partial = 2;
3085        else if (s->size >= 1024)
3086                s->cpu_partial = 6;
3087        else if (s->size >= 256)
3088                s->cpu_partial = 13;
3089        else
3090                s->cpu_partial = 30;
3091
3092        s->refcount = 1;
3093#ifdef CONFIG_NUMA
3094        s->remote_node_defrag_ratio = 1000;
3095#endif
3096        if (!init_kmem_cache_nodes(s))
3097                goto error;
3098
3099        if (alloc_kmem_cache_cpus(s))
3100                return 1;
3101
3102        free_kmem_cache_nodes(s);
3103error:
3104        if (flags & SLAB_PANIC)
3105                panic("Cannot create slab %s size=%lu realsize=%u "
3106                        "order=%u offset=%u flags=%lx\n",
3107                        s->name, (unsigned long)size, s->size, oo_order(s->oo),
3108                        s->offset, flags);
3109        return 0;
3110}
3111
3112/*
3113 * Determine the size of a slab object
3114 */
3115unsigned int kmem_cache_size(struct kmem_cache *s)
3116{
3117        return s->objsize;
3118}
3119EXPORT_SYMBOL(kmem_cache_size);
3120
3121static void list_slab_objects(struct kmem_cache *s, struct page *page,
3122                                                        const char *text)
3123{
3124#ifdef CONFIG_SLUB_DEBUG
3125        void *addr = page_address(page);
3126        void *p;
3127        unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3128                                     sizeof(long), GFP_ATOMIC);
3129        if (!map)
3130                return;
3131        slab_err(s, page, "%s", text);
3132        slab_lock(page);
3133
3134        get_map(s, page, map);
3135        for_each_object(p, s, addr, page->objects) {
3136
3137                if (!test_bit(slab_index(p, s, addr), map)) {
3138                        printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3139                                                        p, p - addr);
3140                        print_tracking(s, p);
3141                }
3142        }
3143        slab_unlock(page);
3144        kfree(map);
3145#endif
3146}
3147
3148/*
3149 * Attempt to free all partial slabs on a node.
3150 * This is called from kmem_cache_close(). We must be the last thread
3151 * using the cache and therefore we do not need to lock anymore.
3152 */
3153static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3154{
3155        struct page *page, *h;
3156
3157        list_for_each_entry_safe(page, h, &n->partial, lru) {
3158                if (!page->inuse) {
3159                        remove_partial(n, page);
3160                        discard_slab(s, page);
3161                } else {
3162                        list_slab_objects(s, page,
3163                                "Objects remaining on kmem_cache_close()");
3164                }
3165        }
3166}
3167
3168/*
3169 * Release all resources used by a slab cache.
3170 */
3171static inline int kmem_cache_close(struct kmem_cache *s)
3172{
3173        int node;
3174
3175        flush_all(s);
3176        free_percpu(s->cpu_slab);
3177        /* Attempt to free all objects */
3178        for_each_node_state(node, N_NORMAL_MEMORY) {
3179                struct kmem_cache_node *n = get_node(s, node);
3180
3181                free_partial(s, n);
3182                if (n->nr_partial || slabs_node(s, node))
3183                        return 1;
3184        }
3185        free_kmem_cache_nodes(s);
3186        return 0;
3187}
3188
3189/*
3190 * Close a cache and release the kmem_cache structure
3191 * (must be used for caches created using kmem_cache_create)
3192 */
3193void kmem_cache_destroy(struct kmem_cache *s)
3194{
3195        down_write(&slub_lock);
3196        s->refcount--;
3197        if (!s->refcount) {
3198                list_del(&s->list);
3199                up_write(&slub_lock);
3200                if (kmem_cache_close(s)) {
3201                        printk(KERN_ERR "SLUB %s: %s called for cache that "
3202                                "still has objects.\n", s->name, __func__);
3203                        dump_stack();
3204                }
3205                if (s->flags & SLAB_DESTROY_BY_RCU)
3206                        rcu_barrier();
3207                sysfs_slab_remove(s);
3208        } else
3209                up_write(&slub_lock);
3210}
3211EXPORT_SYMBOL(kmem_cache_destroy);
3212
3213/********************************************************************
3214 *              Kmalloc subsystem
3215 *******************************************************************/
3216
3217struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3218EXPORT_SYMBOL(kmalloc_caches);
3219
3220static struct kmem_cache *kmem_cache;
3221
3222#ifdef CONFIG_ZONE_DMA
3223static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3224#endif
3225
3226static int __init setup_slub_min_order(char *str)
3227{
3228        get_option(&str, &slub_min_order);
3229
3230        return 1;
3231}
3232
3233__setup("slub_min_order=", setup_slub_min_order);
3234
3235static int __init setup_slub_max_order(char *str)
3236{
3237        get_option(&str, &slub_max_order);
3238        slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3239
3240        return 1;
3241}
3242
3243__setup("slub_max_order=", setup_slub_max_order);
3244
3245static int __init setup_slub_min_objects(char *str)
3246{
3247        get_option(&str, &slub_min_objects);
3248
3249        return 1;
3250}
3251
3252__setup("slub_min_objects=", setup_slub_min_objects);
3253
3254static int __init setup_slub_nomerge(char *str)
3255{
3256        slub_nomerge = 1;
3257        return 1;
3258}
3259
3260__setup("slub_nomerge", setup_slub_nomerge);
3261
3262static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3263                                                int size, unsigned int flags)
3264{
3265        struct kmem_cache *s;
3266
3267        s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3268
3269        /*
3270         * This function is called with IRQs disabled during early-boot on
3271         * single CPU so there's no need to take slub_lock here.
3272         */
3273        if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3274                                                                flags, NULL))
3275                goto panic;
3276
3277        list_add(&s->list, &slab_caches);
3278        return s;
3279
3280panic:
3281        panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3282        return NULL;
3283}
3284
3285/*
3286 * Conversion table for small slabs sizes / 8 to the index in the
3287 * kmalloc array. This is necessary for slabs < 192 since we have non power
3288 * of two cache sizes there. The size of larger slabs can be determined using
3289 * fls.
3290 */
3291static s8 size_index[24] = {
3292        3,      /* 8 */
3293        4,      /* 16 */
3294        5,      /* 24 */
3295        5,      /* 32 */
3296        6,      /* 40 */
3297        6,      /* 48 */
3298        6,      /* 56 */
3299        6,      /* 64 */
3300        1,      /* 72 */
3301        1,      /* 80 */
3302        1,      /* 88 */
3303        1,      /* 96 */
3304        7,      /* 104 */
3305        7,      /* 112 */
3306        7,      /* 120 */
3307        7,      /* 128 */
3308        2,      /* 136 */
3309        2,      /* 144 */
3310        2,      /* 152 */
3311        2,      /* 160 */
3312        2,      /* 168 */
3313        2,      /* 176 */
3314        2,      /* 184 */
3315        2       /* 192 */
3316};
3317
3318static inline int size_index_elem(size_t bytes)
3319{
3320        return (bytes - 1) / 8;
3321}
3322
3323static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3324{
3325        int index;
3326
3327        if (size <= 192) {
3328                if (!size)
3329                        return ZERO_SIZE_PTR;
3330
3331                index = size_index[size_index_elem(size)];
3332        } else
3333                index = fls(size - 1);
3334
3335#ifdef CONFIG_ZONE_DMA
3336        if (unlikely((flags & SLUB_DMA)))
3337                return kmalloc_dma_caches[index];
3338
3339#endif
3340        return kmalloc_caches[index];
3341}
3342
3343void *__kmalloc(size_t size, gfp_t flags)
3344{
3345        struct kmem_cache *s;
3346        void *ret;
3347
3348        if (unlikely(size > SLUB_MAX_SIZE))
3349                return kmalloc_large(size, flags);
3350
3351        s = get_slab(size, flags);
3352
3353        if (unlikely(ZERO_OR_NULL_PTR(s)))
3354                return s;
3355
3356        ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3357
3358        trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3359
3360        return ret;
3361}
3362EXPORT_SYMBOL(__kmalloc);
3363
3364#ifdef CONFIG_NUMA
3365static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3366{
3367        struct page *page;
3368        void *ptr = NULL;
3369
3370        flags |= __GFP_COMP | __GFP_NOTRACK;
3371        page = alloc_pages_node(node, flags, get_order(size));
3372        if (page)
3373                ptr = page_address(page);
3374
3375        kmemleak_alloc(ptr, size, 1, flags);
3376        return ptr;
3377}
3378
3379void *__kmalloc_node(size_t size, gfp_t flags, int node)
3380{
3381        struct kmem_cache *s;
3382        void *ret;
3383
3384        if (unlikely(size > SLUB_MAX_SIZE)) {
3385                ret = kmalloc_large_node(size, flags, node);
3386
3387                trace_kmalloc_node(_RET_IP_, ret,
3388                                   size, PAGE_SIZE << get_order(size),
3389                                   flags, node);
3390
3391                return ret;
3392        }
3393
3394        s = get_slab(size, flags);
3395
3396        if (unlikely(ZERO_OR_NULL_PTR(s)))
3397                return s;
3398
3399        ret = slab_alloc(s, flags, node, _RET_IP_);
3400
3401        trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3402
3403        return ret;
3404}
3405EXPORT_SYMBOL(__kmalloc_node);
3406#endif
3407
3408size_t ksize(const void *object)
3409{
3410        struct page *page;
3411
3412        if (unlikely(object == ZERO_SIZE_PTR))
3413                return 0;
3414
3415        page = virt_to_head_page(object);
3416
3417        if (unlikely(!PageSlab(page))) {
3418                WARN_ON(!PageCompound(page));
3419                return PAGE_SIZE << compound_order(page);
3420        }
3421
3422        return slab_ksize(page->slab);
3423}
3424EXPORT_SYMBOL(ksize);
3425
3426#ifdef CONFIG_SLUB_DEBUG
3427bool verify_mem_not_deleted(const void *x)
3428{
3429        struct page *page;
3430        void *object = (void *)x;
3431        unsigned long flags;
3432        bool rv;
3433
3434        if (unlikely(ZERO_OR_NULL_PTR(x)))
3435                return false;
3436
3437        local_irq_save(flags);
3438
3439        page = virt_to_head_page(x);
3440        if (unlikely(!PageSlab(page))) {
3441                /* maybe it was from stack? */
3442                rv = true;
3443                goto out_unlock;
3444        }
3445
3446        slab_lock(page);
3447        if (on_freelist(page->slab, page, object)) {
3448                object_err(page->slab, page, object, "Object is on free-list");
3449                rv = false;
3450        } else {
3451                rv = true;
3452        }
3453        slab_unlock(page);
3454
3455out_unlock:
3456        local_irq_restore(flags);
3457        return rv;
3458}
3459EXPORT_SYMBOL(verify_mem_not_deleted);
3460#endif
3461
3462void kfree(const void *x)
3463{
3464        struct page *page;
3465        void *object = (void *)x;
3466
3467        trace_kfree(_RET_IP_, x);
3468
3469        if (unlikely(ZERO_OR_NULL_PTR(x)))
3470                return;
3471
3472        page = virt_to_head_page(x);
3473        if (unlikely(!PageSlab(page))) {
3474                BUG_ON(!PageCompound(page));
3475                kmemleak_free(x);
3476                put_page(page);
3477                return;
3478        }
3479        slab_free(page->slab, page, object, _RET_IP_);
3480}
3481EXPORT_SYMBOL(kfree);
3482
3483/*
3484 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3485 * the remaining slabs by the number of items in use. The slabs with the
3486 * most items in use come first. New allocations will then fill those up
3487 * and thus they can be removed from the partial lists.
3488 *
3489 * The slabs with the least items are placed last. This results in them
3490 * being allocated from last increasing the chance that the last objects
3491 * are freed in them.
3492 */
3493int kmem_cache_shrink(struct kmem_cache *s)
3494{
3495        int node;
3496        int i;
3497        struct kmem_cache_node *n;
3498        struct page *page;
3499        struct page *t;
3500        int objects = oo_objects(s->max);
3501        struct list_head *slabs_by_inuse =
3502                kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3503        unsigned long flags;
3504
3505        if (!slabs_by_inuse)
3506                return -ENOMEM;
3507
3508        flush_all(s);
3509        for_each_node_state(node, N_NORMAL_MEMORY) {
3510                n = get_node(s, node);
3511
3512                if (!n->nr_partial)
3513                        continue;
3514
3515                for (i = 0; i < objects; i++)
3516                        INIT_LIST_HEAD(slabs_by_inuse + i);
3517
3518                spin_lock_irqsave(&n->list_lock, flags);
3519
3520                /*
3521                 * Build lists indexed by the items in use in each slab.
3522                 *
3523                 * Note that concurrent frees may occur while we hold the
3524                 * list_lock. page->inuse here is the upper limit.
3525                 */
3526                list_for_each_entry_safe(page, t, &n->partial, lru) {
3527                        list_move(&page->lru, slabs_by_inuse + page->inuse);
3528                        if (!page->inuse)
3529                                n->nr_partial--;
3530                }
3531
3532                /*
3533                 * Rebuild the partial list with the slabs filled up most
3534                 * first and the least used slabs at the end.
3535                 */
3536                for (i = objects - 1; i > 0; i--)
3537                        list_splice(slabs_by_inuse + i, n->partial.prev);
3538
3539                spin_unlock_irqrestore(&n->list_lock, flags);
3540
3541                /* Release empty slabs */
3542                list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3543                        discard_slab(s, page);
3544        }
3545
3546        kfree(slabs_by_inuse);
3547        return 0;
3548}
3549EXPORT_SYMBOL(kmem_cache_shrink);
3550
3551#if defined(CONFIG_MEMORY_HOTPLUG)
3552static int slab_mem_going_offline_callback(void *arg)
3553{
3554        struct kmem_cache *s;
3555
3556        down_read(&slub_lock);
3557        list_for_each_entry(s, &slab_caches, list)
3558                kmem_cache_shrink(s);
3559        up_read(&slub_lock);
3560
3561        return 0;
3562}
3563
3564static void slab_mem_offline_callback(void *arg)
3565{
3566        struct kmem_cache_node *n;
3567        struct kmem_cache *s;
3568        struct memory_notify *marg = arg;
3569        int offline_node;
3570
3571        offline_node = marg->status_change_nid;
3572
3573        /*
3574         * If the node still has available memory. we need kmem_cache_node
3575         * for it yet.
3576         */
3577        if (offline_node < 0)
3578                return;
3579
3580        down_read(&slub_lock);
3581        list_for_each_entry(s, &slab_caches, list) {
3582                n = get_node(s, offline_node);
3583                if (n) {
3584                        /*
3585                         * if n->nr_slabs > 0, slabs still exist on the node
3586                         * that is going down. We were unable to free them,
3587                         * and offline_pages() function shouldn't call this
3588                         * callback. So, we must fail.
3589                         */
3590                        BUG_ON(slabs_node(s, offline_node));
3591
3592                        s->node[offline_node] = NULL;
3593                        kmem_cache_free(kmem_cache_node, n);
3594                }
3595        }
3596        up_read(&slub_lock);
3597}
3598
3599static int slab_mem_going_online_callback(void *arg)
3600{
3601        struct kmem_cache_node *n;
3602        struct kmem_cache *s;
3603        struct memory_notify *marg = arg;
3604        int nid = marg->status_change_nid;
3605        int ret = 0;
3606
3607        /*
3608         * If the node's memory is already available, then kmem_cache_node is
3609         * already created. Nothing to do.
3610         */
3611        if (nid < 0)
3612                return 0;
3613
3614        /*
3615         * We are bringing a node online. No memory is available yet. We must
3616         * allocate a kmem_cache_node structure in order to bring the node
3617         * online.
3618         */
3619        down_read(&slub_lock);
3620        list_for_each_entry(s, &slab_caches, list) {
3621                /*
3622                 * XXX: kmem_cache_alloc_node will fallback to other nodes
3623                 *      since memory is not yet available from the node that
3624                 *      is brought up.
3625                 */
3626                n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3627                if (!n) {
3628                        ret = -ENOMEM;
3629                        goto out;
3630                }
3631                init_kmem_cache_node(n);
3632                s->node[nid] = n;
3633        }
3634out:
3635        up_read(&slub_lock);
3636        return ret;
3637}
3638
3639static int slab_memory_callback(struct notifier_block *self,
3640                                unsigned long action, void *arg)
3641{
3642        int ret = 0;
3643
3644        switch (action) {
3645        case MEM_GOING_ONLINE:
3646                ret = slab_mem_going_online_callback(arg);
3647                break;
3648        case MEM_GOING_OFFLINE:
3649                ret = slab_mem_going_offline_callback(arg);
3650                break;
3651        case MEM_OFFLINE:
3652        case MEM_CANCEL_ONLINE:
3653                slab_mem_offline_callback(arg);
3654                break;
3655        case MEM_ONLINE:
3656        case MEM_CANCEL_OFFLINE:
3657                break;
3658        }
3659        if (ret)
3660                ret = notifier_from_errno(ret);
3661        else
3662                ret = NOTIFY_OK;
3663        return ret;
3664}
3665
3666#endif /* CONFIG_MEMORY_HOTPLUG */
3667
3668/********************************************************************
3669 *                      Basic setup of slabs
3670 *******************************************************************/
3671
3672/*
3673 * Used for early kmem_cache structures that were allocated using
3674 * the page allocator
3675 */
3676
3677static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3678{
3679        int node;
3680
3681        list_add(&s->list, &slab_caches);
3682        s->refcount = -1;
3683
3684        for_each_node_state(node, N_NORMAL_MEMORY) {
3685                struct kmem_cache_node *n = get_node(s, node);
3686                struct page *p;
3687
3688                if (n) {
3689                        list_for_each_entry(p, &n->partial, lru)
3690                                p->slab = s;
3691
3692#ifdef CONFIG_SLUB_DEBUG
3693                        list_for_each_entry(p, &n->full, lru)
3694                                p->slab = s;
3695#endif
3696                }
3697        }
3698}
3699
3700void __init kmem_cache_init(void)
3701{
3702        int i;
3703        int caches = 0;
3704        struct kmem_cache *temp_kmem_cache;
3705        int order;
3706        struct kmem_cache *temp_kmem_cache_node;
3707        unsigned long kmalloc_size;
3708
3709        if (debug_guardpage_minorder())
3710                slub_max_order = 0;
3711
3712        kmem_size = offsetof(struct kmem_cache, node) +
3713                                nr_node_ids * sizeof(struct kmem_cache_node *);
3714
3715        /* Allocate two kmem_caches from the page allocator */
3716        kmalloc_size = ALIGN(kmem_size, cache_line_size());
3717        order = get_order(2 * kmalloc_size);
3718        kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3719
3720        /*
3721         * Must first have the slab cache available for the allocations of the
3722         * struct kmem_cache_node's. There is special bootstrap code in
3723         * kmem_cache_open for slab_state == DOWN.
3724         */
3725        kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3726
3727        kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3728                sizeof(struct kmem_cache_node),
3729                0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3730
3731        hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3732
3733        /* Able to allocate the per node structures */
3734        slab_state = PARTIAL;
3735
3736        temp_kmem_cache = kmem_cache;
3737        kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3738                0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3739        kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3740        memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3741
3742        /*
3743         * Allocate kmem_cache_node properly from the kmem_cache slab.
3744         * kmem_cache_node is separately allocated so no need to
3745         * update any list pointers.
3746         */
3747        temp_kmem_cache_node = kmem_cache_node;
3748
3749        kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3750        memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3751
3752        kmem_cache_bootstrap_fixup(kmem_cache_node);
3753
3754        caches++;
3755        kmem_cache_bootstrap_fixup(kmem_cache);
3756        caches++;
3757        /* Free temporary boot structure */
3758        free_pages((unsigned long)temp_kmem_cache, order);
3759
3760        /* Now we can use the kmem_cache to allocate kmalloc slabs */
3761
3762        /*
3763         * Patch up the size_index table if we have strange large alignment
3764         * requirements for the kmalloc array. This is only the case for
3765         * MIPS it seems. The standard arches will not generate any code here.
3766         *
3767         * Largest permitted alignment is 256 bytes due to the way we
3768         * handle the index determination for the smaller caches.
3769         *
3770         * Make sure that nothing crazy happens if someone starts tinkering
3771         * around with ARCH_KMALLOC_MINALIGN
3772         */
3773        BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3774                (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3775
3776        for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3777                int elem = size_index_elem(i);
3778                if (elem >= ARRAY_SIZE(size_index))
3779                        break;
3780                size_index[elem] = KMALLOC_SHIFT_LOW;
3781        }
3782
3783        if (KMALLOC_MIN_SIZE == 64) {
3784                /*
3785                 * The 96 byte size cache is not used if the alignment
3786                 * is 64 byte.
3787                 */
3788                for (i = 64 + 8; i <= 96; i += 8)
3789                        size_index[size_index_elem(i)] = 7;
3790        } else if (KMALLOC_MIN_SIZE == 128) {
3791                /*
3792                 * The 192 byte sized cache is not used if the alignment
3793                 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3794                 * instead.
3795                 */
3796                for (i = 128 + 8; i <= 192; i += 8)
3797                        size_index[size_index_elem(i)] = 8;
3798        }
3799
3800        /* Caches that are not of the two-to-the-power-of size */
3801        if (KMALLOC_MIN_SIZE <= 32) {
3802                kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3803                caches++;
3804        }
3805
3806        if (KMALLOC_MIN_SIZE <= 64) {
3807                kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3808                caches++;
3809        }
3810
3811        for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3812                kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3813                caches++;
3814        }
3815
3816        slab_state = UP;
3817
3818        /* Provide the correct kmalloc names now that the caches are up */
3819        if (KMALLOC_MIN_SIZE <= 32) {
3820                kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3821                BUG_ON(!kmalloc_caches[1]->name);
3822        }
3823
3824        if (KMALLOC_MIN_SIZE <= 64) {
3825                kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3826                BUG_ON(!kmalloc_caches[2]->name);
3827        }
3828
3829        for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3830                char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3831
3832                BUG_ON(!s);
3833                kmalloc_caches[i]->name = s;
3834        }
3835
3836#ifdef CONFIG_SMP
3837        register_cpu_notifier(&slab_notifier);
3838#endif
3839
3840#ifdef CONFIG_ZONE_DMA
3841        for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3842                struct kmem_cache *s = kmalloc_caches[i];
3843
3844                if (s && s->size) {
3845                        char *name = kasprintf(GFP_NOWAIT,
3846                                 "dma-kmalloc-%d", s->objsize);
3847
3848                        BUG_ON(!name);
3849                        kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3850                                s->objsize, SLAB_CACHE_DMA);
3851                }
3852        }
3853#endif
3854        printk(KERN_INFO
3855                "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3856                " CPUs=%d, Nodes=%d\n",
3857                caches, cache_line_size(),
3858                slub_min_order, slub_max_order, slub_min_objects,
3859                nr_cpu_ids, nr_node_ids);
3860}
3861
3862void __init kmem_cache_init_late(void)
3863{
3864}
3865
3866/*
3867 * Find a mergeable slab cache
3868 */
3869static int slab_unmergeable(struct kmem_cache *s)
3870{
3871        if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3872                return 1;
3873
3874        if (s->ctor)
3875                return 1;
3876
3877        /*
3878         * We may have set a slab to be unmergeable during bootstrap.
3879         */
3880        if (s->refcount < 0)
3881                return 1;
3882
3883        return 0;
3884}
3885
3886static struct kmem_cache *find_mergeable(size_t size,
3887                size_t align, unsigned long flags, const char *name,
3888                void (*ctor)(void *))
3889{
3890        struct kmem_cache *s;
3891
3892        if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3893                return NULL;
3894
3895        if (ctor)
3896                return NULL;
3897
3898        size = ALIGN(size, sizeof(void *));
3899        align = calculate_alignment(flags, align, size);
3900        size = ALIGN(size, align);
3901        flags = kmem_cache_flags(size, flags, name, NULL);
3902
3903        list_for_each_entry(s, &slab_caches, list) {
3904                if (slab_unmergeable(s))
3905                        continue;
3906
3907                if (size > s->size)
3908                        continue;
3909
3910                if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3911                                continue;
3912                /*
3913                 * Check if alignment is compatible.
3914                 * Courtesy of Adrian Drzewiecki
3915                 */
3916                if ((s->size & ~(align - 1)) != s->size)
3917                        continue;
3918
3919                if (s->size - size >= sizeof(void *))
3920                        continue;
3921
3922                return s;
3923        }
3924        return NULL;
3925}
3926
3927struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3928                size_t align, unsigned long flags, void (*ctor)(void *))
3929{
3930        struct kmem_cache *s;
3931        char *n;
3932
3933        if (WARN_ON(!name))
3934                return NULL;
3935
3936        down_write(&slub_lock);
3937        s = find_mergeable(size, align, flags, name, ctor);
3938        if (s) {
3939                s->refcount++;
3940                /*
3941                 * Adjust the object sizes so that we clear
3942                 * the complete object on kzalloc.
3943                 */
3944                s->objsize = max(s->objsize, (int)size);
3945                s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3946
3947                if (sysfs_slab_alias(s, name)) {
3948                        s->refcount--;
3949                        goto err;
3950                }
3951                up_write(&slub_lock);
3952                return s;
3953        }
3954
3955        n = kstrdup(name, GFP_KERNEL);
3956        if (!n)
3957                goto err;
3958
3959        s = kmalloc(kmem_size, GFP_KERNEL);
3960        if (s) {
3961                if (kmem_cache_open(s, n,
3962                                size, align, flags, ctor)) {
3963                        list_add(&s->list, &slab_caches);
3964                        up_write(&slub_lock);
3965                        if (sysfs_slab_add(s)) {
3966                                down_write(&slub_lock);
3967                                list_del(&s->list);
3968                                kfree(n);
3969                                kfree(s);
3970                                goto err;
3971                        }
3972                        return s;
3973                }
3974                kfree(s);
3975        }
3976        kfree(n);
3977err:
3978        up_write(&slub_lock);
3979
3980        if (flags & SLAB_PANIC)
3981                panic("Cannot create slabcache %s\n", name);
3982        else
3983                s = NULL;
3984        return s;
3985}
3986EXPORT_SYMBOL(kmem_cache_create);
3987
3988#ifdef CONFIG_SMP
3989/*
3990 * Use the cpu notifier to insure that the cpu slabs are flushed when
3991 * necessary.
3992 */
3993static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3994                unsigned long action, void *hcpu)
3995{
3996        long cpu = (long)hcpu;
3997        struct kmem_cache *s;
3998        unsigned long flags;
3999
4000        switch (action) {
4001        case CPU_UP_CANCELED:
4002        case CPU_UP_CANCELED_FROZEN:
4003        case CPU_DEAD:
4004        case CPU_DEAD_FROZEN:
4005                down_read(&slub_lock);
4006                list_for_each_entry(s, &slab_caches, list) {
4007                        local_irq_save(flags);
4008                        __flush_cpu_slab(s, cpu);
4009                        local_irq_restore(flags);
4010                }
4011                up_read(&slub_lock);
4012                break;
4013        default:
4014                break;
4015        }
4016        return NOTIFY_OK;
4017}
4018
4019static struct notifier_block __cpuinitdata slab_notifier = {
4020        .notifier_call = slab_cpuup_callback
4021};
4022
4023#endif
4024
4025void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4026{
4027        struct kmem_cache *s;
4028        void *ret;
4029
4030        if (unlikely(size > SLUB_MAX_SIZE))
4031                return kmalloc_large(size, gfpflags);
4032
4033        s = get_slab(size, gfpflags);
4034
4035        if (unlikely(ZERO_OR_NULL_PTR(s)))
4036                return s;
4037
4038        ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
4039
4040        /* Honor the call site pointer we received. */
4041        trace_kmalloc(caller, ret, size, s->size, gfpflags);
4042
4043        return ret;
4044}
4045
4046#ifdef CONFIG_NUMA
4047void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4048                                        int node, unsigned long caller)
4049{
4050        struct kmem_cache *s;
4051        void *ret;
4052
4053        if (unlikely(size > SLUB_MAX_SIZE)) {
4054                ret = kmalloc_large_node(size, gfpflags, node);
4055
4056                trace_kmalloc_node(caller, ret,
4057                                   size, PAGE_SIZE << get_order(size),
4058                                   gfpflags, node);
4059
4060                return ret;
4061        }
4062
4063        s = get_slab(size, gfpflags);
4064
4065        if (unlikely(ZERO_OR_NULL_PTR(s)))
4066                return s;
4067
4068        ret = slab_alloc(s, gfpflags, node, caller);
4069
4070        /* Honor the call site pointer we received. */
4071        trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4072
4073        return ret;
4074}
4075#endif
4076
4077#ifdef CONFIG_SYSFS
4078static int count_inuse(struct page *page)
4079{
4080        return page->inuse;
4081}
4082
4083static int count_total(struct page *page)
4084{
4085        return page->objects;
4086}
4087#endif
4088
4089#ifdef CONFIG_SLUB_DEBUG
4090static int validate_slab(struct kmem_cache *s, struct page *page,
4091                                                unsigned long *map)
4092{
4093        void *p;
4094        void *addr = page_address(page);
4095
4096        if (!check_slab(s, page) ||
4097                        !on_freelist(s, page, NULL))
4098                return 0;
4099
4100        /* Now we know that a valid freelist exists */
4101        bitmap_zero(map, page->objects);
4102
4103        get_map(s, page, map);
4104        for_each_object(p, s, addr, page->objects) {
4105                if (test_bit(slab_index(p, s, addr), map))
4106                        if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4107                                return 0;
4108        }
4109
4110        for_each_object(p, s, addr, page->objects)
4111                if (!test_bit(slab_index(p, s, addr), map))
4112                        if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4113                                return 0;
4114        return 1;
4115}
4116
4117static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4118                                                unsigned long *map)
4119{
4120        slab_lock(page);
4121        validate_slab(s, page, map);
4122        slab_unlock(page);
4123}
4124
4125static int validate_slab_node(struct kmem_cache *s,
4126                struct kmem_cache_node *n, unsigned long *map)
4127{
4128        unsigned long count = 0;
4129        struct page *page;
4130        unsigned long flags;
4131
4132        spin_lock_irqsave(&n->list_lock, flags);
4133
4134        list_for_each_entry(page, &n->partial, lru) {
4135                validate_slab_slab(s, page, map);
4136                count++;
4137        }
4138        if (count != n->nr_partial)
4139                printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4140                        "counter=%ld\n", s->name, count, n->nr_partial);
4141
4142        if (!(s->flags & SLAB_STORE_USER))
4143                goto out;
4144
4145        list_for_each_entry(page, &n->full, lru) {
4146                validate_slab_slab(s, page, map);
4147                count++;
4148        }
4149        if (count != atomic_long_read(&n->nr_slabs))
4150                printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4151                        "counter=%ld\n", s->name, count,
4152                        atomic_long_read(&n->nr_slabs));
4153
4154out:
4155        spin_unlock_irqrestore(&n->list_lock, flags);
4156        return count;
4157}
4158
4159static long validate_slab_cache(struct kmem_cache *s)
4160{
4161        int node;
4162        unsigned long count = 0;
4163        unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4164                                sizeof(unsigned long), GFP_KERNEL);
4165
4166        if (!map)
4167                return -ENOMEM;
4168
4169        flush_all(s);
4170        for_each_node_state(node, N_NORMAL_MEMORY) {
4171                struct kmem_cache_node *n = get_node(s, node);
4172
4173                count += validate_slab_node(s, n, map);
4174        }
4175        kfree(map);
4176        return count;
4177}
4178/*
4179 * Generate lists of code addresses where slabcache objects are allocated
4180 * and freed.
4181 */
4182
4183struct location {
4184        unsigned long count;
4185        unsigned long addr;
4186        long long sum_time;
4187        long min_time;
4188        long max_time;
4189        long min_pid;
4190        long max_pid;
4191        DECLARE_BITMAP(cpus, NR_CPUS);
4192        nodemask_t nodes;
4193};
4194
4195struct loc_track {
4196        unsigned long max;
4197        unsigned long count;
4198        struct location *loc;
4199};
4200
4201static void free_loc_track(struct loc_track *t)
4202{
4203        if (t->max)
4204                free_pages((unsigned long)t->loc,
4205                        get_order(sizeof(struct location) * t->max));
4206}
4207
4208static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4209{
4210        struct location *l;
4211        int order;
4212
4213        order = get_order(sizeof(struct location) * max);
4214
4215        l = (void *)__get_free_pages(flags, order);
4216        if (!l)
4217                return 0;
4218
4219        if (t->count) {
4220                memcpy(l, t->loc, sizeof(struct location) * t->count);
4221                free_loc_track(t);
4222        }
4223        t->max = max;
4224        t->loc = l;
4225        return 1;
4226}
4227
4228static int add_location(struct loc_track *t, struct kmem_cache *s,
4229                                const struct track *track)
4230{
4231        long start, end, pos;
4232        struct location *l;
4233        unsigned long caddr;
4234        unsigned long age = jiffies - track->when;
4235
4236        start = -1;
4237        end = t->count;
4238
4239        for ( ; ; ) {
4240                pos = start + (end - start + 1) / 2;
4241
4242                /*
4243                 * There is nothing at "end". If we end up there
4244                 * we need to add something to before end.
4245                 */
4246                if (pos == end)
4247                        break;
4248
4249                caddr = t->loc[pos].addr;
4250                if (track->addr == caddr) {
4251
4252                        l = &t->loc[pos];
4253                        l->count++;
4254                        if (track->when) {
4255                                l->sum_time += age;
4256                                if (age < l->min_time)
4257                                        l->min_time = age;
4258                                if (age > l->max_time)
4259                                        l->max_time = age;
4260
4261                                if (track->pid < l->min_pid)
4262                                        l->min_pid = track->pid;
4263                                if (track->pid > l->max_pid)
4264                                        l->max_pid = track->pid;
4265
4266                                cpumask_set_cpu(track->cpu,
4267                                                to_cpumask(l->cpus));
4268                        }
4269                        node_set(page_to_nid(virt_to_page(track)), l->nodes);
4270                        return 1;
4271                }
4272
4273                if (track->addr < caddr)
4274                        end = pos;
4275                else
4276                        start = pos;
4277        }
4278
4279        /*
4280         * Not found. Insert new tracking element.
4281         */
4282        if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4283                return 0;
4284
4285        l = t->loc + pos;
4286        if (pos < t->count)
4287                memmove(l + 1, l,
4288                        (t->count - pos) * sizeof(struct location));
4289        t->count++;
4290        l->count = 1;
4291        l->addr = track->addr;
4292        l->sum_time = age;
4293        l->min_time = age;
4294        l->max_time = age;
4295        l->min_pid = track->pid;
4296        l->max_pid = track->pid;
4297        cpumask_clear(to_cpumask(l->cpus));
4298        cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4299        nodes_clear(l->nodes);
4300        node_set(page_to_nid(virt_to_page(track)), l->nodes);
4301        return 1;
4302}
4303
4304static void process_slab(struct loc_track *t, struct kmem_cache *s,
4305                struct page *page, enum track_item alloc,
4306                unsigned long *map)
4307{
4308        void *addr = page_address(page);
4309        void *p;
4310
4311        bitmap_zero(map, page->objects);
4312        get_map(s, page, map);
4313
4314        for_each_object(p, s, addr, page->objects)
4315                if (!test_bit(slab_index(p, s, addr), map))
4316                        add_location(t, s, get_track(s, p, alloc));
4317}
4318
4319static int list_locations(struct kmem_cache *s, char *buf,
4320                                        enum track_item alloc)
4321{
4322        int len = 0;
4323        unsigned long i;
4324        struct loc_track t = { 0, 0, NULL };
4325        int node;
4326        unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4327                                     sizeof(unsigned long), GFP_KERNEL);
4328
4329        if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4330                                     GFP_TEMPORARY)) {
4331                kfree(map);
4332                return sprintf(buf, "Out of memory\n");
4333        }
4334        /* Push back cpu slabs */
4335        flush_all(s);
4336
4337        for_each_node_state(node, N_NORMAL_MEMORY) {
4338                struct kmem_cache_node *n = get_node(s, node);
4339                unsigned long flags;
4340                struct page *page;
4341
4342                if (!atomic_long_read(&n->nr_slabs))
4343                        continue;
4344
4345                spin_lock_irqsave(&n->list_lock, flags);
4346                list_for_each_entry(page, &n->partial, lru)
4347                        process_slab(&t, s, page, alloc, map);
4348                list_for_each_entry(page, &n->full, lru)
4349                        process_slab(&t, s, page, alloc, map);
4350                spin_unlock_irqrestore(&n->list_lock, flags);
4351        }
4352
4353        for (i = 0; i < t.count; i++) {
4354                struct location *l = &t.loc[i];
4355
4356                if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4357                        break;
4358                len += sprintf(buf + len, "%7ld ", l->count);
4359
4360                if (l->addr)
4361                        len += sprintf(buf + len, "%pS", (void *)l->addr);
4362                else
4363                        len += sprintf(buf + len, "<not-available>");
4364
4365                if (l->sum_time != l->min_time) {
4366                        len += sprintf(buf + len, " age=%ld/%ld/%ld",
4367                                l->min_time,
4368                                (long)div_u64(l->sum_time, l->count),
4369                                l->max_time);
4370                } else
4371                        len += sprintf(buf + len, " age=%ld",
4372                                l->min_time);
4373
4374                if (l->min_pid != l->max_pid)
4375                        len += sprintf(buf + len, " pid=%ld-%ld",
4376                                l->min_pid, l->max_pid);
4377                else
4378                        len += sprintf(buf + len, " pid=%ld",
4379                                l->min_pid);
4380
4381                if (num_online_cpus() > 1 &&
4382                                !cpumask_empty(to_cpumask(l->cpus)) &&
4383                                len < PAGE_SIZE - 60) {
4384                        len += sprintf(buf + len, " cpus=");
4385                        len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4386                                                 to_cpumask(l->cpus));
4387                }
4388
4389                if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4390                                len < PAGE_SIZE - 60) {
4391                        len += sprintf(buf + len, " nodes=");
4392                        len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4393                                        l->nodes);
4394                }
4395
4396                len += sprintf(buf + len, "\n");
4397        }
4398
4399        free_loc_track(&t);
4400        kfree(map);
4401        if (!t.count)
4402                len += sprintf(buf, "No data\n");
4403        return len;
4404}
4405#endif
4406
4407#ifdef SLUB_RESILIENCY_TEST
4408static void resiliency_test(void)
4409{
4410        u8 *p;
4411
4412        BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4413
4414        printk(KERN_ERR "SLUB resiliency testing\n");
4415        printk(KERN_ERR "-----------------------\n");
4416        printk(KERN_ERR "A. Corruption after allocation\n");
4417
4418        p = kzalloc(16, GFP_KERNEL);
4419        p[16] = 0x12;
4420        printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4421                        " 0x12->0x%p\n\n", p + 16);
4422
4423        validate_slab_cache(kmalloc_caches[4]);
4424
4425        /* Hmmm... The next two are dangerous */
4426        p = kzalloc(32, GFP_KERNEL);
4427        p[32 + sizeof(void *)] = 0x34;
4428        printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4429                        " 0x34 -> -0x%p\n", p);
4430        printk(KERN_ERR
4431                "If allocated object is overwritten then not detectable\n\n");
4432
4433        validate_slab_cache(kmalloc_caches[5]);
4434        p = kzalloc(64, GFP_KERNEL);
4435        p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4436        *p = 0x56;
4437        printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4438                                                                        p);
4439        printk(KERN_ERR
4440                "If allocated object is overwritten then not detectable\n\n");
4441        validate_slab_cache(kmalloc_caches[6]);
4442
4443        printk(KERN_ERR "\nB. Corruption after free\n");
4444        p = kzalloc(128, GFP_KERNEL);
4445        kfree(p);
4446        *p = 0x78;
4447        printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4448        validate_slab_cache(kmalloc_caches[7]);
4449
4450        p = kzalloc(256, GFP_KERNEL);
4451        kfree(p);
4452        p[50] = 0x9a;
4453        printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4454                        p);
4455        validate_slab_cache(kmalloc_caches[8]);
4456
4457        p = kzalloc(512, GFP_KERNEL);
4458        kfree(p);
4459        p[512] = 0xab;
4460        printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4461        validate_slab_cache(kmalloc_caches[9]);
4462}
4463#else
4464#ifdef CONFIG_SYSFS
4465static void resiliency_test(void) {};
4466#endif
4467#endif
4468
4469#ifdef CONFIG_SYSFS
4470enum slab_stat_type {
4471        SL_ALL,                 /* All slabs */
4472        SL_PARTIAL,             /* Only partially allocated slabs */
4473        SL_CPU,                 /* Only slabs used for cpu caches */
4474        SL_OBJECTS,             /* Determine allocated objects not slabs */
4475        SL_TOTAL                /* Determine object capacity not slabs */
4476};
4477
4478#define SO_ALL          (1 << SL_ALL)
4479#define SO_PARTIAL      (1 << SL_PARTIAL)
4480#define SO_CPU          (1 << SL_CPU)
4481#define SO_OBJECTS      (1 << SL_OBJECTS)
4482#define SO_TOTAL        (1 << SL_TOTAL)
4483
4484static ssize_t show_slab_objects(struct kmem_cache *s,
4485                            char *buf, unsigned long flags)
4486{
4487        unsigned long total = 0;
4488        int node;
4489        int x;
4490        unsigned long *nodes;
4491        unsigned long *per_cpu;
4492
4493        nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4494        if (!nodes)
4495                return -ENOMEM;
4496        per_cpu = nodes + nr_node_ids;
4497
4498        if (flags & SO_CPU) {
4499                int cpu;
4500
4501                for_each_possible_cpu(cpu) {
4502                        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4503                        int node = ACCESS_ONCE(c->node);
4504                        struct page *page;
4505
4506                        if (node < 0)
4507                                continue;
4508                        page = ACCESS_ONCE(c->page);
4509                        if (page) {
4510                                if (flags & SO_TOTAL)
4511                                        x = page->objects;
4512                                else if (flags & SO_OBJECTS)
4513                                        x = page->inuse;
4514                                else
4515                                        x = 1;
4516
4517                                total += x;
4518                                nodes[node] += x;
4519                        }
4520                        page = c->partial;
4521
4522                        if (page) {
4523                                x = page->pobjects;
4524                                total += x;
4525                                nodes[node] += x;
4526                        }
4527                        per_cpu[node]++;
4528                }
4529        }
4530
4531        lock_memory_hotplug();
4532#ifdef CONFIG_SLUB_DEBUG
4533        if (flags & SO_ALL) {
4534                for_each_node_state(node, N_NORMAL_MEMORY) {
4535                        struct kmem_cache_node *n = get_node(s, node);
4536
4537                if (flags & SO_TOTAL)
4538                        x = atomic_long_read(&n->total_objects);
4539                else if (flags & SO_OBJECTS)
4540                        x = atomic_long_read(&n->total_objects) -
4541                                count_partial(n, count_free);
4542
4543                        else
4544                                x = atomic_long_read(&n->nr_slabs);
4545                        total += x;
4546                        nodes[node] += x;
4547                }
4548
4549        } else
4550#endif
4551        if (flags & SO_PARTIAL) {
4552                for_each_node_state(node, N_NORMAL_MEMORY) {
4553                        struct kmem_cache_node *n = get_node(s, node);
4554
4555                        if (flags & SO_TOTAL)
4556                                x = count_partial(n, count_total);
4557                        else if (flags & SO_OBJECTS)
4558                                x = count_partial(n, count_inuse);
4559                        else
4560                                x = n->nr_partial;
4561                        total += x;
4562                        nodes[node] += x;
4563                }
4564        }
4565        x = sprintf(buf, "%lu", total);
4566#ifdef CONFIG_NUMA
4567        for_each_node_state(node, N_NORMAL_MEMORY)
4568                if (nodes[node])
4569                        x += sprintf(buf + x, " N%d=%lu",
4570                                        node, nodes[node]);
4571#endif
4572        unlock_memory_hotplug();
4573        kfree(nodes);
4574        return x + sprintf(buf + x, "\n");
4575}
4576
4577#ifdef CONFIG_SLUB_DEBUG
4578static int any_slab_objects(struct kmem_cache *s)
4579{
4580        int node;
4581
4582        for_each_online_node(node) {
4583                struct kmem_cache_node *n = get_node(s, node);
4584
4585                if (!n)
4586                        continue;
4587
4588                if (atomic_long_read(&n->total_objects))
4589                        return 1;
4590        }
4591        return 0;
4592}
4593#endif
4594
4595#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4596#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4597
4598struct slab_attribute {
4599        struct attribute attr;
4600        ssize_t (*show)(struct kmem_cache *s, char *buf);
4601        ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4602};
4603
4604#define SLAB_ATTR_RO(_name) \
4605        static struct slab_attribute _name##_attr = \
4606        __ATTR(_name, 0400, _name##_show, NULL)
4607
4608#define SLAB_ATTR(_name) \
4609        static struct slab_attribute _name##_attr =  \
4610        __ATTR(_name, 0600, _name##_show, _name##_store)
4611
4612static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4613{
4614        return sprintf(buf, "%d\n", s->size);
4615}
4616SLAB_ATTR_RO(slab_size);
4617
4618static ssize_t align_show(struct kmem_cache *s, char *buf)
4619{
4620        return sprintf(buf, "%d\n", s->align);
4621}
4622SLAB_ATTR_RO(align);
4623
4624static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4625{
4626        return sprintf(buf, "%d\n", s->objsize);
4627}
4628SLAB_ATTR_RO(object_size);
4629
4630static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4631{
4632        return sprintf(buf, "%d\n", oo_objects(s->oo));
4633}
4634SLAB_ATTR_RO(objs_per_slab);
4635
4636static ssize_t order_store(struct kmem_cache *s,
4637                                const char *buf, size_t length)
4638{
4639        unsigned long order;
4640        int err;
4641
4642        err = strict_strtoul(buf, 10, &order);
4643        if (err)
4644                return err;
4645
4646        if (order > slub_max_order || order < slub_min_order)
4647                return -EINVAL;
4648
4649        calculate_sizes(s, order);
4650        return length;
4651}
4652
4653static ssize_t order_show(struct kmem_cache *s, char *buf)
4654{
4655        return sprintf(buf, "%d\n", oo_order(s->oo));
4656}
4657SLAB_ATTR(order);
4658
4659static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4660{
4661        return sprintf(buf, "%lu\n", s->min_partial);
4662}
4663
4664static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4665                                 size_t length)
4666{
4667        unsigned long min;
4668        int err;
4669
4670        err = strict_strtoul(buf, 10, &min);
4671        if (err)
4672                return err;
4673
4674        set_min_partial(s, min);
4675        return length;
4676}
4677SLAB_ATTR(min_partial);
4678
4679static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4680{
4681        return sprintf(buf, "%u\n", s->cpu_partial);
4682}
4683
4684static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4685                                 size_t length)
4686{
4687        unsigned long objects;
4688        int err;
4689
4690        err = strict_strtoul(buf, 10, &objects);
4691        if (err)
4692                return err;
4693        if (objects && kmem_cache_debug(s))
4694                return -EINVAL;
4695
4696        s->cpu_partial = objects;
4697        flush_all(s);
4698        return length;
4699}
4700SLAB_ATTR(cpu_partial);
4701
4702static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4703{
4704        if (!s->ctor)
4705                return 0;
4706        return sprintf(buf, "%pS\n", s->ctor);
4707}
4708SLAB_ATTR_RO(ctor);
4709
4710static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4711{
4712        return sprintf(buf, "%d\n", s->refcount - 1);
4713}
4714SLAB_ATTR_RO(aliases);
4715
4716static ssize_t partial_show(struct kmem_cache *s, char *buf)
4717{
4718        return show_slab_objects(s, buf, SO_PARTIAL);
4719}
4720SLAB_ATTR_RO(partial);
4721
4722static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4723{
4724        return show_slab_objects(s, buf, SO_CPU);
4725}
4726SLAB_ATTR_RO(cpu_slabs);
4727
4728static ssize_t objects_show(struct kmem_cache *s, char *buf)
4729{
4730        return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4731}
4732SLAB_ATTR_RO(objects);
4733
4734static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4735{
4736        return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4737}
4738SLAB_ATTR_RO(objects_partial);
4739
4740static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4741{
4742        int objects = 0;
4743        int pages = 0;
4744        int cpu;
4745        int len;
4746
4747        for_each_online_cpu(cpu) {
4748                struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4749
4750                if (page) {
4751                        pages += page->pages;
4752                        objects += page->pobjects;
4753                }
4754        }
4755
4756        len = sprintf(buf, "%d(%d)", objects, pages);
4757
4758#ifdef CONFIG_SMP
4759        for_each_online_cpu(cpu) {
4760                struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4761
4762                if (page && len < PAGE_SIZE - 20)
4763                        len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4764                                page->pobjects, page->pages);
4765        }
4766#endif
4767        return len + sprintf(buf + len, "\n");
4768}
4769SLAB_ATTR_RO(slabs_cpu_partial);
4770
4771static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4772{
4773        return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4774}
4775
4776static ssize_t reclaim_account_store(struct kmem_cache *s,
4777                                const char *buf, size_t length)
4778{
4779        s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4780        if (buf[0] == '1')
4781                s->flags |= SLAB_RECLAIM_ACCOUNT;
4782        return length;
4783}
4784SLAB_ATTR(reclaim_account);
4785
4786static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4787{
4788        return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4789}
4790SLAB_ATTR_RO(hwcache_align);
4791
4792#ifdef CONFIG_ZONE_DMA
4793static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4794{
4795        return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4796}
4797SLAB_ATTR_RO(cache_dma);
4798#endif
4799
4800static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4801{
4802        return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4803}
4804SLAB_ATTR_RO(destroy_by_rcu);
4805
4806static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4807{
4808        return sprintf(buf, "%d\n", s->reserved);
4809}
4810SLAB_ATTR_RO(reserved);
4811
4812#ifdef CONFIG_SLUB_DEBUG
4813static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4814{
4815        return show_slab_objects(s, buf, SO_ALL);
4816}
4817SLAB_ATTR_RO(slabs);
4818
4819static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4820{
4821        return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4822}
4823SLAB_ATTR_RO(total_objects);
4824
4825static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4826{
4827        return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4828}
4829
4830static ssize_t sanity_checks_store(struct kmem_cache *s,
4831                                const char *buf, size_t length)
4832{
4833        s->flags &= ~SLAB_DEBUG_FREE;
4834        if (buf[0] == '1') {
4835                s->flags &= ~__CMPXCHG_DOUBLE;
4836                s->flags |= SLAB_DEBUG_FREE;
4837        }
4838        return length;
4839}
4840SLAB_ATTR(sanity_checks);
4841
4842static ssize_t trace_show(struct kmem_cache *s, char *buf)
4843{
4844        return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4845}
4846
4847static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4848                                                        size_t length)
4849{
4850        s->flags &= ~SLAB_TRACE;
4851        if (buf[0] == '1') {
4852                s->flags &= ~__CMPXCHG_DOUBLE;
4853                s->flags |= SLAB_TRACE;
4854        }
4855        return length;
4856}
4857SLAB_ATTR(trace);
4858
4859static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4860{
4861        return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4862}
4863
4864static ssize_t red_zone_store(struct kmem_cache *s,
4865                                const char *buf, size_t length)
4866{
4867        if (any_slab_objects(s))
4868                return -EBUSY;
4869
4870        s->flags &= ~SLAB_RED_ZONE;
4871        if (buf[0] == '1') {
4872                s->flags &= ~__CMPXCHG_DOUBLE;
4873                s->flags |= SLAB_RED_ZONE;
4874        }
4875        calculate_sizes(s, -1);
4876        return length;
4877}
4878SLAB_ATTR(red_zone);
4879
4880static ssize_t poison_show(struct kmem_cache *s, char *buf)
4881{
4882        return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4883}
4884
4885static ssize_t poison_store(struct kmem_cache *s,
4886                                const char *buf, size_t length)
4887{
4888        if (any_slab_objects(s))
4889                return -EBUSY;
4890
4891        s->flags &= ~SLAB_POISON;
4892        if (buf[0] == '1') {
4893                s->flags &= ~__CMPXCHG_DOUBLE;
4894                s->flags |= SLAB_POISON;
4895        }
4896        calculate_sizes(s, -1);
4897        return length;
4898}
4899SLAB_ATTR(poison);
4900
4901static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4902{
4903        return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4904}
4905
4906static ssize_t store_user_store(struct kmem_cache *s,
4907                                const char *buf, size_t length)
4908{
4909        if (any_slab_objects(s))
4910                return -EBUSY;
4911
4912        s->flags &= ~SLAB_STORE_USER;
4913        if (buf[0] == '1') {
4914                s->flags &= ~__CMPXCHG_DOUBLE;
4915                s->flags |= SLAB_STORE_USER;
4916        }
4917        calculate_sizes(s, -1);
4918        return length;
4919}
4920SLAB_ATTR(store_user);
4921
4922static ssize_t validate_show(struct kmem_cache *s, char *buf)
4923{
4924        return 0;
4925}
4926
4927static ssize_t validate_store(struct kmem_cache *s,
4928                        const char *buf, size_t length)
4929{
4930        int ret = -EINVAL;
4931
4932        if (buf[0] == '1') {
4933                ret = validate_slab_cache(s);
4934                if (ret >= 0)
4935                        ret = length;
4936        }
4937        return ret;
4938}
4939SLAB_ATTR(validate);
4940
4941static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4942{
4943        if (!(s->flags & SLAB_STORE_USER))
4944                return -ENOSYS;
4945        return list_locations(s, buf, TRACK_ALLOC);
4946}
4947SLAB_ATTR_RO(alloc_calls);
4948
4949static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4950{
4951        if (!(s->flags & SLAB_STORE_USER))
4952                return -ENOSYS;
4953        return list_locations(s, buf, TRACK_FREE);
4954}
4955SLAB_ATTR_RO(free_calls);
4956#endif /* CONFIG_SLUB_DEBUG */
4957
4958#ifdef CONFIG_FAILSLAB
4959static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4960{
4961        return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4962}
4963
4964static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4965                                                        size_t length)
4966{
4967        s->flags &= ~SLAB_FAILSLAB;
4968        if (buf[0] == '1')
4969                s->flags |= SLAB_FAILSLAB;
4970        return length;
4971}
4972SLAB_ATTR(failslab);
4973#endif
4974
4975static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4976{
4977        return 0;
4978}
4979
4980static ssize_t shrink_store(struct kmem_cache *s,
4981                        const char *buf, size_t length)
4982{
4983        if (buf[0] == '1') {
4984                int rc = kmem_cache_shrink(s);
4985
4986                if (rc)
4987                        return rc;
4988        } else
4989                return -EINVAL;
4990        return length;
4991}
4992SLAB_ATTR(shrink);
4993
4994#ifdef CONFIG_NUMA
4995static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4996{
4997        return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4998}
4999
5000static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5001                                const char *buf, size_t length)
5002{
5003        unsigned long ratio;
5004        int err;
5005
5006        err = strict_strtoul(buf, 10, &ratio);
5007        if (err)
5008                return err;
5009
5010        if (ratio <= 100)
5011                s->remote_node_defrag_ratio = ratio * 10;
5012
5013        return length;
5014}
5015SLAB_ATTR(remote_node_defrag_ratio);
5016#endif
5017
5018#ifdef CONFIG_SLUB_STATS
5019static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5020{
5021        unsigned long sum  = 0;
5022        int cpu;
5023        int len;
5024        int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5025
5026        if (!data)
5027                return -ENOMEM;
5028
5029        for_each_online_cpu(cpu) {
5030                unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5031
5032                data[cpu] = x;
5033                sum += x;
5034        }
5035
5036        len = sprintf(buf, "%lu", sum);
5037
5038#ifdef CONFIG_SMP
5039        for_each_online_cpu(cpu) {
5040                if (data[cpu] && len < PAGE_SIZE - 20)
5041                        len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5042        }
5043#endif
5044        kfree(data);
5045        return len + sprintf(buf + len, "\n");
5046}
5047
5048static void clear_stat(struct kmem_cache *s, enum stat_item si)
5049{
5050        int cpu;
5051
5052        for_each_online_cpu(cpu)
5053                per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5054}
5055
5056#define STAT_ATTR(si, text)                                     \
5057static ssize_t text##_show(struct kmem_cache *s, char *buf)     \
5058{                                                               \
5059        return show_stat(s, buf, si);                           \
5060}                                                               \
5061static ssize_t text##_store(struct kmem_cache *s,               \
5062                                const char *buf, size_t length) \
5063{                                                               \
5064        if (buf[0] != '0')                                      \
5065                return -EINVAL;                                 \
5066        clear_stat(s, si);                                      \
5067        return length;                                          \
5068}                                                               \
5069SLAB_ATTR(text);                                                \
5070
5071STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5072STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5073STAT_ATTR(FREE_FASTPATH, free_fastpath);
5074STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5075STAT_ATTR(FREE_FROZEN, free_frozen);
5076STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5077STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5078STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5079STAT_ATTR(ALLOC_SLAB, alloc_slab);
5080STAT_ATTR(ALLOC_REFILL, alloc_refill);
5081STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5082STAT_ATTR(FREE_SLAB, free_slab);
5083STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5084STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5085STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5086STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5087STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5088STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5089STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5090STAT_ATTR(ORDER_FALLBACK, order_fallback);
5091STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5092STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5093STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5094STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5095STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5096STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5097#endif
5098
5099static struct attribute *slab_attrs[] = {
5100        &slab_size_attr.attr,
5101        &object_size_attr.attr,
5102        &objs_per_slab_attr.attr,
5103        &order_attr.attr,
5104        &min_partial_attr.attr,
5105        &cpu_partial_attr.attr,
5106        &objects_attr.attr,
5107        &objects_partial_attr.attr,
5108        &partial_attr.attr,
5109        &cpu_slabs_attr.attr,
5110        &ctor_attr.attr,
5111        &aliases_attr.attr,
5112        &align_attr.attr,
5113        &hwcache_align_attr.attr,
5114        &reclaim_account_attr.attr,
5115        &destroy_by_rcu_attr.attr,
5116        &shrink_attr.attr,
5117        &reserved_attr.attr,
5118        &slabs_cpu_partial_attr.attr,
5119#ifdef CONFIG_SLUB_DEBUG
5120        &total_objects_attr.attr,
5121        &slabs_attr.attr,
5122        &sanity_checks_attr.attr,
5123        &trace_attr.attr,
5124        &red_zone_attr.attr,
5125        &poison_attr.attr,
5126        &store_user_attr.attr,
5127        &validate_attr.attr,
5128        &alloc_calls_attr.attr,
5129        &free_calls_attr.attr,
5130#endif
5131#ifdef CONFIG_ZONE_DMA
5132        &cache_dma_attr.attr,
5133#endif
5134#ifdef CONFIG_NUMA
5135        &remote_node_defrag_ratio_attr.attr,
5136#endif
5137#ifdef CONFIG_SLUB_STATS
5138        &alloc_fastpath_attr.attr,
5139        &alloc_slowpath_attr.attr,
5140        &free_fastpath_attr.attr,
5141        &free_slowpath_attr.attr,
5142        &free_frozen_attr.attr,
5143        &free_add_partial_attr.attr,
5144        &free_remove_partial_attr.attr,
5145        &alloc_from_partial_attr.attr,
5146        &alloc_slab_attr.attr,
5147        &alloc_refill_attr.attr,
5148        &alloc_node_mismatch_attr.attr,
5149        &free_slab_attr.attr,
5150        &cpuslab_flush_attr.attr,
5151        &deactivate_full_attr.attr,
5152        &deactivate_empty_attr.attr,
5153        &deactivate_to_head_attr.attr,
5154        &deactivate_to_tail_attr.attr,
5155        &deactivate_remote_frees_attr.attr,
5156        &deactivate_bypass_attr.attr,
5157        &order_fallback_attr.attr,
5158        &cmpxchg_double_fail_attr.attr,
5159        &cmpxchg_double_cpu_fail_attr.attr,
5160        &cpu_partial_alloc_attr.attr,
5161        &cpu_partial_free_attr.attr,
5162        &cpu_partial_node_attr.attr,
5163        &cpu_partial_drain_attr.attr,
5164#endif
5165#ifdef CONFIG_FAILSLAB
5166        &failslab_attr.attr,
5167#endif
5168
5169        NULL
5170};
5171
5172static struct attribute_group slab_attr_group = {
5173        .attrs = slab_attrs,
5174};
5175
5176static ssize_t slab_attr_show(struct kobject *kobj,
5177                                struct attribute *attr,
5178                                char *buf)
5179{
5180        struct slab_attribute *attribute;
5181        struct kmem_cache *s;
5182        int err;
5183
5184        attribute = to_slab_attr(attr);
5185        s = to_slab(kobj);
5186
5187        if (!attribute->show)
5188                return -EIO;
5189
5190        err = attribute->show(s, buf);
5191
5192        return err;
5193}
5194
5195static ssize_t slab_attr_store(struct kobject *kobj,
5196                                struct attribute *attr,
5197                                const char *buf, size_t len)
5198{
5199        struct slab_attribute *attribute;
5200        struct kmem_cache *s;
5201        int err;
5202
5203        attribute = to_slab_attr(attr);
5204        s = to_slab(kobj);
5205
5206        if (!attribute->store)
5207                return -EIO;
5208
5209        err = attribute->store(s, buf, len);
5210
5211        return err;
5212}
5213
5214static void kmem_cache_release(struct kobject *kobj)
5215{
5216        struct kmem_cache *s = to_slab(kobj);
5217
5218        kfree(s->name);
5219        kfree(s);
5220}
5221
5222static const struct sysfs_ops slab_sysfs_ops = {
5223        .show = slab_attr_show,
5224        .store = slab_attr_store,
5225};
5226
5227static struct kobj_type slab_ktype = {
5228        .sysfs_ops = &slab_sysfs_ops,
5229        .release = kmem_cache_release
5230};
5231
5232static int uevent_filter(struct kset *kset, struct kobject *kobj)
5233{
5234        struct kobj_type *ktype = get_ktype(kobj);
5235
5236        if (ktype == &slab_ktype)
5237                return 1;
5238        return 0;
5239}
5240
5241static const struct kset_uevent_ops slab_uevent_ops = {
5242        .filter = uevent_filter,
5243};
5244
5245static struct kset *slab_kset;
5246
5247#define ID_STR_LENGTH 64
5248
5249/* Create a unique string id for a slab cache:
5250 *
5251 * Format       :[flags-]size
5252 */
5253static char *create_unique_id(struct kmem_cache *s)
5254{
5255        char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5256        char *p = name;
5257
5258        BUG_ON(!name);
5259
5260        *p++ = ':';
5261        /*
5262         * First flags affecting slabcache operations. We will only
5263         * get here for aliasable slabs so we do not need to support
5264         * too many flags. The flags here must cover all flags that
5265         * are matched during merging to guarantee that the id is
5266         * unique.
5267         */
5268        if (s->flags & SLAB_CACHE_DMA)
5269                *p++ = 'd';
5270        if (s->flags & SLAB_RECLAIM_ACCOUNT)
5271                *p++ = 'a';
5272        if (s->flags & SLAB_DEBUG_FREE)
5273                *p++ = 'F';
5274        if (!(s->flags & SLAB_NOTRACK))
5275                *p++ = 't';
5276        if (p != name + 1)
5277                *p++ = '-';
5278        p += sprintf(p, "%07d", s->size);
5279        BUG_ON(p > name + ID_STR_LENGTH - 1);
5280        return name;
5281}
5282
5283static int sysfs_slab_add(struct kmem_cache *s)
5284{
5285        int err;
5286        const char *name;
5287        int unmergeable;
5288
5289        if (slab_state < SYSFS)
5290                /* Defer until later */
5291                return 0;
5292
5293        unmergeable = slab_unmergeable(s);
5294        if (unmergeable) {
5295                /*
5296                 * Slabcache can never be merged so we can use the name proper.
5297                 * This is typically the case for debug situations. In that
5298                 * case we can catch duplicate names easily.
5299                 */
5300                sysfs_remove_link(&slab_kset->kobj, s->name);
5301                name = s->name;
5302        } else {
5303                /*
5304                 * Create a unique name for the slab as a target
5305                 * for the symlinks.
5306                 */
5307                name = create_unique_id(s);
5308        }
5309
5310        s->kobj.kset = slab_kset;
5311        err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5312        if (err) {
5313                kobject_put(&s->kobj);
5314                return err;
5315        }
5316
5317        err = sysfs_create_group(&s->kobj, &slab_attr_group);
5318        if (err) {
5319                kobject_del(&s->kobj);
5320                kobject_put(&s->kobj);
5321                return err;
5322        }
5323        kobject_uevent(&s->kobj, KOBJ_ADD);
5324        if (!unmergeable) {
5325                /* Setup first alias */
5326                sysfs_slab_alias(s, s->name);
5327                kfree(name);
5328        }
5329        return 0;
5330}
5331
5332static void sysfs_slab_remove(struct kmem_cache *s)
5333{
5334        if (slab_state < SYSFS)
5335                /*
5336                 * Sysfs has not been setup yet so no need to remove the
5337                 * cache from sysfs.
5338                 */
5339                return;
5340
5341        kobject_uevent(&s->kobj, KOBJ_REMOVE);
5342        kobject_del(&s->kobj);
5343        kobject_put(&s->kobj);
5344}
5345
5346/*
5347 * Need to buffer aliases during bootup until sysfs becomes
5348 * available lest we lose that information.
5349 */
5350struct saved_alias {
5351        struct kmem_cache *s;
5352        const char *name;
5353        struct saved_alias *next;
5354};
5355
5356static struct saved_alias *alias_list;
5357
5358static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5359{
5360        struct saved_alias *al;
5361
5362        if (slab_state == SYSFS) {
5363                /*
5364                 * If we have a leftover link then remove it.
5365                 */
5366                sysfs_remove_link(&slab_kset->kobj, name);
5367                return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5368        }
5369
5370        al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5371        if (!al)
5372                return -ENOMEM;
5373
5374        al->s = s;
5375        al->name = name;
5376        al->next = alias_list;
5377        alias_list = al;
5378        return 0;
5379}
5380
5381static int __init slab_sysfs_init(void)
5382{
5383        struct kmem_cache *s;
5384        int err;
5385
5386        down_write(&slub_lock);
5387
5388        slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5389        if (!slab_kset) {
5390                up_write(&slub_lock);
5391                printk(KERN_ERR "Cannot register slab subsystem.\n");
5392                return -ENOSYS;
5393        }
5394
5395        slab_state = SYSFS;
5396
5397        list_for_each_entry(s, &slab_caches, list) {
5398                err = sysfs_slab_add(s);
5399                if (err)
5400                        printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5401                                                " to sysfs\n", s->name);
5402        }
5403
5404        while (alias_list) {
5405                struct saved_alias *al = alias_list;
5406
5407                alias_list = alias_list->next;
5408                err = sysfs_slab_alias(al->s, al->name);
5409                if (err)
5410                        printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5411                                        " %s to sysfs\n", s->name);
5412                kfree(al);
5413        }
5414
5415        up_write(&slub_lock);
5416        resiliency_test();
5417        return 0;
5418}
5419
5420__initcall(slab_sysfs_init);
5421#endif /* CONFIG_SYSFS */
5422
5423/*
5424 * The /proc/slabinfo ABI
5425 */
5426#ifdef CONFIG_SLABINFO
5427static void print_slabinfo_header(struct seq_file *m)
5428{
5429        seq_puts(m, "slabinfo - version: 2.1\n");
5430        seq_puts(m, "# name            <active_objs> <num_objs> <objsize> "
5431                 "<objperslab> <pagesperslab>");
5432        seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5433        seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5434        seq_putc(m, '\n');
5435}
5436
5437static void *s_start(struct seq_file *m, loff_t *pos)
5438{
5439        loff_t n = *pos;
5440
5441        down_read(&slub_lock);
5442        if (!n)
5443                print_slabinfo_header(m);
5444
5445        return seq_list_start(&slab_caches, *pos);
5446}
5447
5448static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5449{
5450        return seq_list_next(p, &slab_caches, pos);
5451}
5452
5453static void s_stop(struct seq_file *m, void *p)
5454{
5455        up_read(&slub_lock);
5456}
5457
5458static int s_show(struct seq_file *m, void *p)
5459{
5460        unsigned long nr_partials = 0;
5461        unsigned long nr_slabs = 0;
5462        unsigned long nr_inuse = 0;
5463        unsigned long nr_objs = 0;
5464        unsigned long nr_free = 0;
5465        struct kmem_cache *s;
5466        int node;
5467
5468        s = list_entry(p, struct kmem_cache, list);
5469
5470        for_each_online_node(node) {
5471                struct kmem_cache_node *n = get_node(s, node);
5472
5473                if (!n)
5474                        continue;
5475
5476                nr_partials += n->nr_partial;
5477                nr_slabs += atomic_long_read(&n->nr_slabs);
5478                nr_objs += atomic_long_read(&n->total_objects);
5479                nr_free += count_partial(n, count_free);
5480        }
5481
5482        nr_inuse = nr_objs - nr_free;
5483
5484        seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5485                   nr_objs, s->size, oo_objects(s->oo),
5486                   (1 << oo_order(s->oo)));
5487        seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5488        seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5489                   0UL);
5490        seq_putc(m, '\n');
5491        return 0;
5492}
5493
5494static const struct seq_operations slabinfo_op = {
5495        .start = s_start,
5496        .next = s_next,
5497        .stop = s_stop,
5498        .show = s_show,
5499};
5500
5501static int slabinfo_open(struct inode *inode, struct file *file)
5502{
5503        return seq_open(file, &slabinfo_op);
5504}
5505
5506static const struct file_operations proc_slabinfo_operations = {
5507        .open           = slabinfo_open,
5508        .read           = seq_read,
5509        .llseek         = seq_lseek,
5510        .release        = seq_release,
5511};
5512
5513static int __init slab_proc_init(void)
5514{
5515        proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5516        return 0;
5517}
5518module_init(slab_proc_init);
5519#endif /* CONFIG_SLABINFO */
5520
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.