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