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#include <linux/memcontrol.h>
  35
  36#include <trace/events/kmem.h>
  37
  38#include "internal.h"
  39
  40/*
  41 * Lock order:
  42 *   1. slab_mutex (Global Mutex)
  43 *   2. node->list_lock
  44 *   3. slab_lock(page) (Only on some arches and for debugging)
  45 *
  46 *   slab_mutex
  47 *
  48 *   The role of the slab_mutex is to protect the list of all the slabs
  49 *   and to synchronize major metadata changes to slab cache structures.
  50 *
  51 *   The slab_lock is only used for debugging and on arches that do not
  52 *   have the ability to do a cmpxchg_double. It only protects the second
  53 *   double word in the page struct. Meaning
  54 *      A. page->freelist       -> List of object free in a page
  55 *      B. page->counters       -> Counters of objects
  56 *      C. page->frozen         -> frozen state
  57 *
  58 *   If a slab is frozen then it is exempt from list management. It is not
  59 *   on any list. The processor that froze the slab is the one who can
  60 *   perform list operations on the page. Other processors may put objects
  61 *   onto the freelist but the processor that froze the slab is the only
  62 *   one that can retrieve the objects from the page's freelist.
  63 *
  64 *   The list_lock protects the partial and full list on each node and
  65 *   the partial slab counter. If taken then no new slabs may be added or
  66 *   removed from the lists nor make the number of partial slabs be modified.
  67 *   (Note that the total number of slabs is an atomic value that may be
  68 *   modified without taking the list lock).
  69 *
  70 *   The list_lock is a centralized lock and thus we avoid taking it as
  71 *   much as possible. As long as SLUB does not have to handle partial
  72 *   slabs, operations can continue without any centralized lock. F.e.
  73 *   allocating a long series of objects that fill up slabs does not require
  74 *   the list lock.
  75 *   Interrupts are disabled during allocation and deallocation in order to
  76 *   make the slab allocator safe to use in the context of an irq. In addition
  77 *   interrupts are disabled to ensure that the processor does not change
  78 *   while handling per_cpu slabs, due to kernel preemption.
  79 *
  80 * SLUB assigns one slab for allocation to each processor.
  81 * Allocations only occur from these slabs called cpu slabs.
  82 *
  83 * Slabs with free elements are kept on a partial list and during regular
  84 * operations no list for full slabs is used. If an object in a full slab is
  85 * freed then the slab will show up again on the partial lists.
  86 * We track full slabs for debugging purposes though because otherwise we
  87 * cannot scan all objects.
  88 *
  89 * Slabs are freed when they become empty. Teardown and setup is
  90 * minimal so we rely on the page allocators per cpu caches for
  91 * fast frees and allocs.
  92 *
  93 * Overloading of page flags that are otherwise used for LRU management.
  94 *
  95 * PageActive           The slab is frozen and exempt from list processing.
  96 *                      This means that the slab is dedicated to a purpose
  97 *                      such as satisfying allocations for a specific
  98 *                      processor. Objects may be freed in the slab while
  99 *                      it is frozen but slab_free will then skip the usual
 100 *                      list operations. It is up to the processor holding
 101 *                      the slab to integrate the slab into the slab lists
 102 *                      when the slab is no longer needed.
 103 *
 104 *                      One use of this flag is to mark slabs that are
 105 *                      used for allocations. Then such a slab becomes a cpu
 106 *                      slab. The cpu slab may be equipped with an additional
 107 *                      freelist that allows lockless access to
 108 *                      free objects in addition to the regular freelist
 109 *                      that requires the slab lock.
 110 *
 111 * PageError            Slab requires special handling due to debug
 112 *                      options set. This moves slab handling out of
 113 *                      the fast path and disables lockless freelists.
 114 */
 115
 116static inline int kmem_cache_debug(struct kmem_cache *s)
 117{
 118#ifdef CONFIG_SLUB_DEBUG
 119        return unlikely(s->flags & SLAB_DEBUG_FLAGS);
 120#else
 121        return 0;
 122#endif
 123}
 124
 125/*
 126 * Issues still to be resolved:
 127 *
 128 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
 129 *
 130 * - Variable sizing of the per node arrays
 131 */
 132
 133/* Enable to test recovery from slab corruption on boot */
 134#undef SLUB_RESILIENCY_TEST
 135
 136/* Enable to log cmpxchg failures */
 137#undef SLUB_DEBUG_CMPXCHG
 138
 139/*
 140 * Mininum number of partial slabs. These will be left on the partial
 141 * lists even if they are empty. kmem_cache_shrink may reclaim them.
 142 */
 143#define MIN_PARTIAL 5
 144
 145/*
 146 * Maximum number of desirable partial slabs.
 147 * The existence of more partial slabs makes kmem_cache_shrink
 148 * sort the partial list by the number of objects in the.
 149 */
 150#define MAX_PARTIAL 10
 151
 152#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
 153                                SLAB_POISON | SLAB_STORE_USER)
 154
 155/*
 156 * Debugging flags that require metadata to be stored in the slab.  These get
 157 * disabled when slub_debug=O is used and a cache's min order increases with
 158 * metadata.
 159 */
 160#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
 161
 162/*
 163 * Set of flags that will prevent slab merging
 164 */
 165#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
 166                SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
 167                SLAB_FAILSLAB)
 168
 169#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
 170                SLAB_CACHE_DMA | SLAB_NOTRACK)
 171
 172#define OO_SHIFT        16
 173#define OO_MASK         ((1 << OO_SHIFT) - 1)
 174#define MAX_OBJS_PER_PAGE       32767 /* since page.objects is u15 */
 175
 176/* Internal SLUB flags */
 177#define __OBJECT_POISON         0x80000000UL /* Poison object */
 178#define __CMPXCHG_DOUBLE        0x40000000UL /* Use cmpxchg_double */
 179
 180#ifdef CONFIG_SMP
 181static struct notifier_block slab_notifier;
 182#endif
 183
 184/*
 185 * Tracking user of a slab.
 186 */
 187#define TRACK_ADDRS_COUNT 16
 188struct track {
 189        unsigned long addr;     /* Called from address */
 190#ifdef CONFIG_STACKTRACE
 191        unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
 192#endif
 193        int cpu;                /* Was running on cpu */
 194        int pid;                /* Pid context */
 195        unsigned long when;     /* When did the operation occur */
 196};
 197
 198enum track_item { TRACK_ALLOC, TRACK_FREE };
 199
 200#ifdef CONFIG_SYSFS
 201static int sysfs_slab_add(struct kmem_cache *);
 202static int sysfs_slab_alias(struct kmem_cache *, const char *);
 203static void sysfs_slab_remove(struct kmem_cache *);
 204static void memcg_propagate_slab_attrs(struct kmem_cache *s);
 205#else
 206static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
 207static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
 208                                                        { return 0; }
 209static inline void sysfs_slab_remove(struct kmem_cache *s) { }
 210
 211static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
 212#endif
 213
 214static inline void stat(const struct kmem_cache *s, enum stat_item si)
 215{
 216#ifdef CONFIG_SLUB_STATS
 217        __this_cpu_inc(s->cpu_slab->stat[si]);
 218#endif
 219}
 220
 221/********************************************************************
 222 *                      Core slab cache functions
 223 *******************************************************************/
 224
 225static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
 226{
 227        return s->node[node];
 228}
 229
 230/* Verify that a pointer has an address that is valid within a slab page */
 231static inline int check_valid_pointer(struct kmem_cache *s,
 232                                struct page *page, const void *object)
 233{
 234        void *base;
 235
 236        if (!object)
 237                return 1;
 238
 239        base = page_address(page);
 240        if (object < base || object >= base + page->objects * s->size ||
 241                (object - base) % s->size) {
 242                return 0;
 243        }
 244
 245        return 1;
 246}
 247
 248static inline void *get_freepointer(struct kmem_cache *s, void *object)
 249{
 250        return *(void **)(object + s->offset);
 251}
 252
 253static void prefetch_freepointer(const struct kmem_cache *s, void *object)
 254{
 255        prefetch(object + s->offset);
 256}
 257
 258static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
 259{
 260        void *p;
 261
 262#ifdef CONFIG_DEBUG_PAGEALLOC
 263        probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
 264#else
 265        p = get_freepointer(s, object);
 266#endif
 267        return p;
 268}
 269
 270static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
 271{
 272        *(void **)(object + s->offset) = fp;
 273}
 274
 275/* Loop over all objects in a slab */
 276#define for_each_object(__p, __s, __addr, __objects) \
 277        for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
 278                        __p += (__s)->size)
 279
 280/* Determine object index from a given position */
 281static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
 282{
 283        return (p - addr) / s->size;
 284}
 285
 286static inline size_t slab_ksize(const struct kmem_cache *s)
 287{
 288#ifdef CONFIG_SLUB_DEBUG
 289        /*
 290         * Debugging requires use of the padding between object
 291         * and whatever may come after it.
 292         */
 293        if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
 294                return s->object_size;
 295
 296#endif
 297        /*
 298         * If we have the need to store the freelist pointer
 299         * back there or track user information then we can
 300         * only use the space before that information.
 301         */
 302        if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
 303                return s->inuse;
 304        /*
 305         * Else we can use all the padding etc for the allocation
 306         */
 307        return s->size;
 308}
 309
 310static inline int order_objects(int order, unsigned long size, int reserved)
 311{
 312        return ((PAGE_SIZE << order) - reserved) / size;
 313}
 314
 315static inline struct kmem_cache_order_objects oo_make(int order,
 316                unsigned long size, int reserved)
 317{
 318        struct kmem_cache_order_objects x = {
 319                (order << OO_SHIFT) + order_objects(order, size, reserved)
 320        };
 321
 322        return x;
 323}
 324
 325static inline int oo_order(struct kmem_cache_order_objects x)
 326{
 327        return x.x >> OO_SHIFT;
 328}
 329
 330static inline int oo_objects(struct kmem_cache_order_objects x)
 331{
 332        return x.x & OO_MASK;
 333}
 334
 335/*
 336 * Per slab locking using the pagelock
 337 */
 338static __always_inline void slab_lock(struct page *page)
 339{
 340        bit_spin_lock(PG_locked, &page->flags);
 341}
 342
 343static __always_inline void slab_unlock(struct page *page)
 344{
 345        __bit_spin_unlock(PG_locked, &page->flags);
 346}
 347
 348/* Interrupts must be disabled (for the fallback code to work right) */
 349static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
 350                void *freelist_old, unsigned long counters_old,
 351                void *freelist_new, unsigned long counters_new,
 352                const char *n)
 353{
 354        VM_BUG_ON(!irqs_disabled());
 355#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
 356    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
 357        if (s->flags & __CMPXCHG_DOUBLE) {
 358                if (cmpxchg_double(&page->freelist, &page->counters,
 359                        freelist_old, counters_old,
 360                        freelist_new, counters_new))
 361                return 1;
 362        } else
 363#endif
 364        {
 365                slab_lock(page);
 366                if (page->freelist == freelist_old && page->counters == counters_old) {
 367                        page->freelist = freelist_new;
 368                        page->counters = counters_new;
 369                        slab_unlock(page);
 370                        return 1;
 371                }
 372                slab_unlock(page);
 373        }
 374
 375        cpu_relax();
 376        stat(s, CMPXCHG_DOUBLE_FAIL);
 377
 378#ifdef SLUB_DEBUG_CMPXCHG
 379        printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
 380#endif
 381
 382        return 0;
 383}
 384
 385static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
 386                void *freelist_old, unsigned long counters_old,
 387                void *freelist_new, unsigned long counters_new,
 388                const char *n)
 389{
 390#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
 391    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
 392        if (s->flags & __CMPXCHG_DOUBLE) {
 393                if (cmpxchg_double(&page->freelist, &page->counters,
 394                        freelist_old, counters_old,
 395                        freelist_new, counters_new))
 396                return 1;
 397        } else
 398#endif
 399        {
 400                unsigned long flags;
 401
 402                local_irq_save(flags);
 403                slab_lock(page);
 404                if (page->freelist == freelist_old && page->counters == counters_old) {
 405                        page->freelist = freelist_new;
 406                        page->counters = counters_new;
 407                        slab_unlock(page);
 408                        local_irq_restore(flags);
 409                        return 1;
 410                }
 411                slab_unlock(page);
 412                local_irq_restore(flags);
 413        }
 414
 415        cpu_relax();
 416        stat(s, CMPXCHG_DOUBLE_FAIL);
 417
 418#ifdef SLUB_DEBUG_CMPXCHG
 419        printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
 420#endif
 421
 422        return 0;
 423}
 424
 425#ifdef CONFIG_SLUB_DEBUG
 426/*
 427 * Determine a map of object in use on a page.
 428 *
 429 * Node listlock must be held to guarantee that the page does
 430 * not vanish from under us.
 431 */
 432static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
 433{
 434        void *p;
 435        void *addr = page_address(page);
 436
 437        for (p = page->freelist; p; p = get_freepointer(s, p))
 438                set_bit(slab_index(p, s, addr), map);
 439}
 440
 441/*
 442 * Debug settings:
 443 */
 444#ifdef CONFIG_SLUB_DEBUG_ON
 445static int slub_debug = DEBUG_DEFAULT_FLAGS;
 446#else
 447static int slub_debug;
 448#endif
 449
 450static char *slub_debug_slabs;
 451static int disable_higher_order_debug;
 452
 453/*
 454 * Object debugging
 455 */
 456static void print_section(char *text, u8 *addr, unsigned int length)
 457{
 458        print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
 459                        length, 1);
 460}
 461
 462static struct track *get_track(struct kmem_cache *s, void *object,
 463        enum track_item alloc)
 464{
 465        struct track *p;
 466
 467        if (s->offset)
 468                p = object + s->offset + sizeof(void *);
 469        else
 470                p = object + s->inuse;
 471
 472        return p + alloc;
 473}
 474
 475static void set_track(struct kmem_cache *s, void *object,
 476                        enum track_item alloc, unsigned long addr)
 477{
 478        struct track *p = get_track(s, object, alloc);
 479
 480        if (addr) {
 481#ifdef CONFIG_STACKTRACE
 482                struct stack_trace trace;
 483                int i;
 484
 485                trace.nr_entries = 0;
 486                trace.max_entries = TRACK_ADDRS_COUNT;
 487                trace.entries = p->addrs;
 488                trace.skip = 3;
 489                save_stack_trace(&trace);
 490
 491                /* See rant in lockdep.c */
 492                if (trace.nr_entries != 0 &&
 493                    trace.entries[trace.nr_entries - 1] == ULONG_MAX)
 494                        trace.nr_entries--;
 495
 496                for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
 497                        p->addrs[i] = 0;
 498#endif
 499                p->addr = addr;
 500                p->cpu = smp_processor_id();
 501                p->pid = current->pid;
 502                p->when = jiffies;
 503        } else
 504                memset(p, 0, sizeof(struct track));
 505}
 506
 507static void init_tracking(struct kmem_cache *s, void *object)
 508{
 509        if (!(s->flags & SLAB_STORE_USER))
 510                return;
 511
 512        set_track(s, object, TRACK_FREE, 0UL);
 513        set_track(s, object, TRACK_ALLOC, 0UL);
 514}
 515
 516static void print_track(const char *s, struct track *t)
 517{
 518        if (!t->addr)
 519                return;
 520
 521        printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
 522                s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
 523#ifdef CONFIG_STACKTRACE
 524        {
 525                int i;
 526                for (i = 0; i < TRACK_ADDRS_COUNT; i++)
 527                        if (t->addrs[i])
 528                                printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
 529                        else
 530                                break;
 531        }
 532#endif
 533}
 534
 535static void print_tracking(struct kmem_cache *s, void *object)
 536{
 537        if (!(s->flags & SLAB_STORE_USER))
 538                return;
 539
 540        print_track("Allocated", get_track(s, object, TRACK_ALLOC));
 541        print_track("Freed", get_track(s, object, TRACK_FREE));
 542}
 543
 544static void print_page_info(struct page *page)
 545{
 546        printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
 547                page, page->objects, page->inuse, page->freelist, page->flags);
 548
 549}
 550
 551static void slab_bug(struct kmem_cache *s, char *fmt, ...)
 552{
 553        va_list args;
 554        char buf[100];
 555
 556        va_start(args, fmt);
 557        vsnprintf(buf, sizeof(buf), fmt, args);
 558        va_end(args);
 559        printk(KERN_ERR "========================================"
 560                        "=====================================\n");
 561        printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
 562        printk(KERN_ERR "----------------------------------------"
 563                        "-------------------------------------\n\n");
 564
 565        add_taint(TAINT_BAD_PAGE);
 566}
 567
 568static void slab_fix(struct kmem_cache *s, char *fmt, ...)
 569{
 570        va_list args;
 571        char buf[100];
 572
 573        va_start(args, fmt);
 574        vsnprintf(buf, sizeof(buf), fmt, args);
 575        va_end(args);
 576        printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
 577}
 578
 579static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
 580{
 581        unsigned int off;       /* Offset of last byte */
 582        u8 *addr = page_address(page);
 583
 584        print_tracking(s, p);
 585
 586        print_page_info(page);
 587
 588        printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
 589                        p, p - addr, get_freepointer(s, p));
 590
 591        if (p > addr + 16)
 592                print_section("Bytes b4 ", p - 16, 16);
 593
 594        print_section("Object ", p, min_t(unsigned long, s->object_size,
 595                                PAGE_SIZE));
 596        if (s->flags & SLAB_RED_ZONE)
 597                print_section("Redzone ", p + s->object_size,
 598                        s->inuse - s->object_size);
 599
 600        if (s->offset)
 601                off = s->offset + sizeof(void *);
 602        else
 603                off = s->inuse;
 604
 605        if (s->flags & SLAB_STORE_USER)
 606                off += 2 * sizeof(struct track);
 607
 608        if (off != s->size)
 609                /* Beginning of the filler is the free pointer */
 610                print_section("Padding ", p + off, s->size - off);
 611
 612        dump_stack();
 613}
 614
 615static void object_err(struct kmem_cache *s, struct page *page,
 616                        u8 *object, char *reason)
 617{
 618        slab_bug(s, "%s", reason);
 619        print_trailer(s, page, object);
 620}
 621
 622static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
 623{
 624        va_list args;
 625        char buf[100];
 626
 627        va_start(args, fmt);
 628        vsnprintf(buf, sizeof(buf), fmt, args);
 629        va_end(args);
 630        slab_bug(s, "%s", buf);
 631        print_page_info(page);
 632        dump_stack();
 633}
 634
 635static void init_object(struct kmem_cache *s, void *object, u8 val)
 636{
 637        u8 *p = object;
 638
 639        if (s->flags & __OBJECT_POISON) {
 640                memset(p, POISON_FREE, s->object_size - 1);
 641                p[s->object_size - 1] = POISON_END;
 642        }
 643
 644        if (s->flags & SLAB_RED_ZONE)
 645                memset(p + s->object_size, val, s->inuse - s->object_size);
 646}
 647
 648static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
 649                                                void *from, void *to)
 650{
 651        slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
 652        memset(from, data, to - from);
 653}
 654
 655static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
 656                        u8 *object, char *what,
 657                        u8 *start, unsigned int value, unsigned int bytes)
 658{
 659        u8 *fault;
 660        u8 *end;
 661
 662        fault = memchr_inv(start, value, bytes);
 663        if (!fault)
 664                return 1;
 665
 666        end = start + bytes;
 667        while (end > fault && end[-1] == value)
 668                end--;
 669
 670        slab_bug(s, "%s overwritten", what);
 671        printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
 672                                        fault, end - 1, fault[0], value);
 673        print_trailer(s, page, object);
 674
 675        restore_bytes(s, what, value, fault, end);
 676        return 0;
 677}
 678
 679/*
 680 * Object layout:
 681 *
 682 * object address
 683 *      Bytes of the object to be managed.
 684 *      If the freepointer may overlay the object then the free
 685 *      pointer is the first word of the object.
 686 *
 687 *      Poisoning uses 0x6b (POISON_FREE) and the last byte is
 688 *      0xa5 (POISON_END)
 689 *
 690 * object + s->object_size
 691 *      Padding to reach word boundary. This is also used for Redzoning.
 692 *      Padding is extended by another word if Redzoning is enabled and
 693 *      object_size == inuse.
 694 *
 695 *      We fill with 0xbb (RED_INACTIVE) for inactive objects and with
 696 *      0xcc (RED_ACTIVE) for objects in use.
 697 *
 698 * object + s->inuse
 699 *      Meta data starts here.
 700 *
 701 *      A. Free pointer (if we cannot overwrite object on free)
 702 *      B. Tracking data for SLAB_STORE_USER
 703 *      C. Padding to reach required alignment boundary or at mininum
 704 *              one word if debugging is on to be able to detect writes
 705 *              before the word boundary.
 706 *
 707 *      Padding is done using 0x5a (POISON_INUSE)
 708 *
 709 * object + s->size
 710 *      Nothing is used beyond s->size.
 711 *
 712 * If slabcaches are merged then the object_size and inuse boundaries are mostly
 713 * ignored. And therefore no slab options that rely on these boundaries
 714 * may be used with merged slabcaches.
 715 */
 716
 717static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
 718{
 719        unsigned long off = s->inuse;   /* The end of info */
 720
 721        if (s->offset)
 722                /* Freepointer is placed after the object. */
 723                off += sizeof(void *);
 724
 725        if (s->flags & SLAB_STORE_USER)
 726                /* We also have user information there */
 727                off += 2 * sizeof(struct track);
 728
 729        if (s->size == off)
 730                return 1;
 731
 732        return check_bytes_and_report(s, page, p, "Object padding",
 733                                p + off, POISON_INUSE, s->size - off);
 734}
 735
 736/* Check the pad bytes at the end of a slab page */
 737static int slab_pad_check(struct kmem_cache *s, struct page *page)
 738{
 739        u8 *start;
 740        u8 *fault;
 741        u8 *end;
 742        int length;
 743        int remainder;
 744
 745        if (!(s->flags & SLAB_POISON))
 746                return 1;
 747
 748        start = page_address(page);
 749        length = (PAGE_SIZE << compound_order(page)) - s->reserved;
 750        end = start + length;
 751        remainder = length % s->size;
 752        if (!remainder)
 753                return 1;
 754
 755        fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
 756        if (!fault)
 757                return 1;
 758        while (end > fault && end[-1] == POISON_INUSE)
 759                end--;
 760
 761        slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
 762        print_section("Padding ", end - remainder, remainder);
 763
 764        restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
 765        return 0;
 766}
 767
 768static int check_object(struct kmem_cache *s, struct page *page,
 769                                        void *object, u8 val)
 770{
 771        u8 *p = object;
 772        u8 *endobject = object + s->object_size;
 773
 774        if (s->flags & SLAB_RED_ZONE) {
 775                if (!check_bytes_and_report(s, page, object, "Redzone",
 776                        endobject, val, s->inuse - s->object_size))
 777                        return 0;
 778        } else {
 779                if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
 780                        check_bytes_and_report(s, page, p, "Alignment padding",
 781                                endobject, POISON_INUSE, s->inuse - s->object_size);
 782                }
 783        }
 784
 785        if (s->flags & SLAB_POISON) {
 786                if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
 787                        (!check_bytes_and_report(s, page, p, "Poison", p,
 788                                        POISON_FREE, s->object_size - 1) ||
 789                         !check_bytes_and_report(s, page, p, "Poison",
 790                                p + s->object_size - 1, POISON_END, 1)))
 791                        return 0;
 792                /*
 793                 * check_pad_bytes cleans up on its own.
 794                 */
 795                check_pad_bytes(s, page, p);
 796        }
 797
 798        if (!s->offset && val == SLUB_RED_ACTIVE)
 799                /*
 800                 * Object and freepointer overlap. Cannot check
 801                 * freepointer while object is allocated.
 802                 */
 803                return 1;
 804
 805        /* Check free pointer validity */
 806        if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
 807                object_err(s, page, p, "Freepointer corrupt");
 808                /*
 809                 * No choice but to zap it and thus lose the remainder
 810                 * of the free objects in this slab. May cause
 811                 * another error because the object count is now wrong.
 812                 */
 813                set_freepointer(s, p, NULL);
 814                return 0;
 815        }
 816        return 1;
 817}
 818
 819static int check_slab(struct kmem_cache *s, struct page *page)
 820{
 821        int maxobj;
 822
 823        VM_BUG_ON(!irqs_disabled());
 824
 825        if (!PageSlab(page)) {
 826                slab_err(s, page, "Not a valid slab page");
 827                return 0;
 828        }
 829
 830        maxobj = order_objects(compound_order(page), s->size, s->reserved);
 831        if (page->objects > maxobj) {
 832                slab_err(s, page, "objects %u > max %u",
 833                        s->name, page->objects, maxobj);
 834                return 0;
 835        }
 836        if (page->inuse > page->objects) {
 837                slab_err(s, page, "inuse %u > max %u",
 838                        s->name, page->inuse, page->objects);
 839                return 0;
 840        }
 841        /* Slab_pad_check fixes things up after itself */
 842        slab_pad_check(s, page);
 843        return 1;
 844}
 845
 846/*
 847 * Determine if a certain object on a page is on the freelist. Must hold the
 848 * slab lock to guarantee that the chains are in a consistent state.
 849 */
 850static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
 851{
 852        int nr = 0;
 853        void *fp;
 854        void *object = NULL;
 855        unsigned long max_objects;
 856
 857        fp = page->freelist;
 858        while (fp && nr <= page->objects) {
 859                if (fp == search)
 860                        return 1;
 861                if (!check_valid_pointer(s, page, fp)) {
 862                        if (object) {
 863                                object_err(s, page, object,
 864                                        "Freechain corrupt");
 865                                set_freepointer(s, object, NULL);
 866                                break;
 867                        } else {
 868                                slab_err(s, page, "Freepointer corrupt");
 869                                page->freelist = NULL;
 870                                page->inuse = page->objects;
 871                                slab_fix(s, "Freelist cleared");
 872                                return 0;
 873                        }
 874                        break;
 875                }
 876                object = fp;
 877                fp = get_freepointer(s, object);
 878                nr++;
 879        }
 880
 881        max_objects = order_objects(compound_order(page), s->size, s->reserved);
 882        if (max_objects > MAX_OBJS_PER_PAGE)
 883                max_objects = MAX_OBJS_PER_PAGE;
 884
 885        if (page->objects != max_objects) {
 886                slab_err(s, page, "Wrong number of objects. Found %d but "
 887                        "should be %d", page->objects, max_objects);
 888                page->objects = max_objects;
 889                slab_fix(s, "Number of objects adjusted.");
 890        }
 891        if (page->inuse != page->objects - nr) {
 892                slab_err(s, page, "Wrong object count. Counter is %d but "
 893                        "counted were %d", page->inuse, page->objects - nr);
 894                page->inuse = page->objects - nr;
 895                slab_fix(s, "Object count adjusted.");
 896        }
 897        return search == NULL;
 898}
 899
 900static void trace(struct kmem_cache *s, struct page *page, void *object,
 901                                                                int alloc)
 902{
 903        if (s->flags & SLAB_TRACE) {
 904                printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
 905                        s->name,
 906                        alloc ? "alloc" : "free",
 907                        object, page->inuse,
 908                        page->freelist);
 909
 910                if (!alloc)
 911                        print_section("Object ", (void *)object, s->object_size);
 912
 913                dump_stack();
 914        }
 915}
 916
 917/*
 918 * Hooks for other subsystems that check memory allocations. In a typical
 919 * production configuration these hooks all should produce no code at all.
 920 */
 921static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
 922{
 923        flags &= gfp_allowed_mask;
 924        lockdep_trace_alloc(flags);
 925        might_sleep_if(flags & __GFP_WAIT);
 926
 927        return should_failslab(s->object_size, flags, s->flags);
 928}
 929
 930static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
 931{
 932        flags &= gfp_allowed_mask;
 933        kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
 934        kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
 935}
 936
 937static inline void slab_free_hook(struct kmem_cache *s, void *x)
 938{
 939        kmemleak_free_recursive(x, s->flags);
 940
 941        /*
 942         * Trouble is that we may no longer disable interupts in the fast path
 943         * So in order to make the debug calls that expect irqs to be
 944         * disabled we need to disable interrupts temporarily.
 945         */
 946#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
 947        {
 948                unsigned long flags;
 949
 950                local_irq_save(flags);
 951                kmemcheck_slab_free(s, x, s->object_size);
 952                debug_check_no_locks_freed(x, s->object_size);
 953                local_irq_restore(flags);
 954        }
 955#endif
 956        if (!(s->flags & SLAB_DEBUG_OBJECTS))
 957                debug_check_no_obj_freed(x, s->object_size);
 958}
 959
 960/*
 961 * Tracking of fully allocated slabs for debugging purposes.
 962 *
 963 * list_lock must be held.
 964 */
 965static void add_full(struct kmem_cache *s,
 966        struct kmem_cache_node *n, struct page *page)
 967{
 968        if (!(s->flags & SLAB_STORE_USER))
 969                return;
 970
 971        list_add(&page->lru, &n->full);
 972}
 973
 974/*
 975 * list_lock must be held.
 976 */
 977static void remove_full(struct kmem_cache *s, struct page *page)
 978{
 979        if (!(s->flags & SLAB_STORE_USER))
 980                return;
 981
 982        list_del(&page->lru);
 983}
 984
 985/* Tracking of the number of slabs for debugging purposes */
 986static inline unsigned long slabs_node(struct kmem_cache *s, int node)
 987{
 988        struct kmem_cache_node *n = get_node(s, node);
 989
 990        return atomic_long_read(&n->nr_slabs);
 991}
 992
 993static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
 994{
 995        return atomic_long_read(&n->nr_slabs);
 996}
 997
 998static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
 999{
1000        struct kmem_cache_node *n = get_node(s, node);
1001
1002        /*
1003         * May be called early in order to allocate a slab for the
1004         * kmem_cache_node structure. Solve the chicken-egg
1005         * dilemma by deferring the increment of the count during
1006         * bootstrap (see early_kmem_cache_node_alloc).
1007         */
1008        if (n) {
1009                atomic_long_inc(&n->nr_slabs);
1010                atomic_long_add(objects, &n->total_objects);
1011        }
1012}
1013static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1014{
1015        struct kmem_cache_node *n = get_node(s, node);
1016
1017        atomic_long_dec(&n->nr_slabs);
1018        atomic_long_sub(objects, &n->total_objects);
1019}
1020
1021/* Object debug checks for alloc/free paths */
1022static void setup_object_debug(struct kmem_cache *s, struct page *page,
1023                                                                void *object)
1024{
1025        if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1026                return;
1027
1028        init_object(s, object, SLUB_RED_INACTIVE);
1029        init_tracking(s, object);
1030}
1031
1032static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1033                                        void *object, unsigned long addr)
1034{
1035        if (!check_slab(s, page))
1036                goto bad;
1037
1038        if (!check_valid_pointer(s, page, object)) {
1039                object_err(s, page, object, "Freelist Pointer check fails");
1040                goto bad;
1041        }
1042
1043        if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1044                goto bad;
1045
1046        /* Success perform special debug activities for allocs */
1047        if (s->flags & SLAB_STORE_USER)
1048                set_track(s, object, TRACK_ALLOC, addr);
1049        trace(s, page, object, 1);
1050        init_object(s, object, SLUB_RED_ACTIVE);
1051        return 1;
1052
1053bad:
1054        if (PageSlab(page)) {
1055                /*
1056                 * If this is a slab page then lets do the best we can
1057                 * to avoid issues in the future. Marking all objects
1058                 * as used avoids touching the remaining objects.
1059                 */
1060                slab_fix(s, "Marking all objects used");
1061                page->inuse = page->objects;
1062                page->freelist = NULL;
1063        }
1064        return 0;
1065}
1066
1067static noinline struct kmem_cache_node *free_debug_processing(
1068        struct kmem_cache *s, struct page *page, void *object,
1069        unsigned long addr, unsigned long *flags)
1070{
1071        struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1072
1073        spin_lock_irqsave(&n->list_lock, *flags);
1074        slab_lock(page);
1075
1076        if (!check_slab(s, page))
1077                goto fail;
1078
1079        if (!check_valid_pointer(s, page, object)) {
1080                slab_err(s, page, "Invalid object pointer 0x%p", object);
1081                goto fail;
1082        }
1083
1084        if (on_freelist(s, page, object)) {
1085                object_err(s, page, object, "Object already free");
1086                goto fail;
1087        }
1088
1089        if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1090                goto out;
1091
1092        if (unlikely(s != page->slab_cache)) {
1093                if (!PageSlab(page)) {
1094                        slab_err(s, page, "Attempt to free object(0x%p) "
1095                                "outside of slab", object);
1096                } else if (!page->slab_cache) {
1097                        printk(KERN_ERR
1098                                "SLUB <none>: no slab for object 0x%p.\n",
1099                                                object);
1100                        dump_stack();
1101                } else
1102                        object_err(s, page, object,
1103                                        "page slab pointer corrupt.");
1104                goto fail;
1105        }
1106
1107        if (s->flags & SLAB_STORE_USER)
1108                set_track(s, object, TRACK_FREE, addr);
1109        trace(s, page, object, 0);
1110        init_object(s, object, SLUB_RED_INACTIVE);
1111out:
1112        slab_unlock(page);
1113        /*
1114         * Keep node_lock to preserve integrity
1115         * until the object is actually freed
1116         */
1117        return n;
1118
1119fail:
1120        slab_unlock(page);
1121        spin_unlock_irqrestore(&n->list_lock, *flags);
1122        slab_fix(s, "Object at 0x%p not freed", object);
1123        return NULL;
1124}
1125
1126static int __init setup_slub_debug(char *str)
1127{
1128        slub_debug = DEBUG_DEFAULT_FLAGS;
1129        if (*str++ != '=' || !*str)
1130                /*
1131                 * No options specified. Switch on full debugging.
1132                 */
1133                goto out;
1134
1135        if (*str == ',')
1136                /*
1137                 * No options but restriction on slabs. This means full
1138                 * debugging for slabs matching a pattern.
1139                 */
1140                goto check_slabs;
1141
1142        if (tolower(*str) == 'o') {
1143                /*
1144                 * Avoid enabling debugging on caches if its minimum order
1145                 * would increase as a result.
1146                 */
1147                disable_higher_order_debug = 1;
1148                goto out;
1149        }
1150
1151        slub_debug = 0;
1152        if (*str == '-')
1153                /*
1154                 * Switch off all debugging measures.
1155                 */
1156                goto out;
1157
1158        /*
1159         * Determine which debug features should be switched on
1160         */
1161        for (; *str && *str != ','; str++) {
1162                switch (tolower(*str)) {
1163                case 'f':
1164                        slub_debug |= SLAB_DEBUG_FREE;
1165                        break;
1166                case 'z':
1167                        slub_debug |= SLAB_RED_ZONE;
1168                        break;
1169                case 'p':
1170                        slub_debug |= SLAB_POISON;
1171                        break;
1172                case 'u':
1173                        slub_debug |= SLAB_STORE_USER;
1174                        break;
1175                case 't':
1176                        slub_debug |= SLAB_TRACE;
1177                        break;
1178                case 'a':
1179                        slub_debug |= SLAB_FAILSLAB;
1180                        break;
1181                default:
1182                        printk(KERN_ERR "slub_debug option '%c' "
1183                                "unknown. skipped\n", *str);
1184                }
1185        }
1186
1187check_slabs:
1188        if (*str == ',')
1189                slub_debug_slabs = str + 1;
1190out:
1191        return 1;
1192}
1193
1194__setup("slub_debug", setup_slub_debug);
1195
1196static unsigned long kmem_cache_flags(unsigned long object_size,
1197        unsigned long flags, const char *name,
1198        void (*ctor)(void *))
1199{
1200        /*
1201         * Enable debugging if selected on the kernel commandline.
1202         */
1203        if (slub_debug && (!slub_debug_slabs ||
1204                !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1205                flags |= slub_debug;
1206
1207        return flags;
1208}
1209#else
1210static inline void setup_object_debug(struct kmem_cache *s,
1211                        struct page *page, void *object) {}
1212
1213static inline int alloc_debug_processing(struct kmem_cache *s,
1214        struct page *page, void *object, unsigned long addr) { return 0; }
1215
1216static inline struct kmem_cache_node *free_debug_processing(
1217        struct kmem_cache *s, struct page *page, void *object,
1218        unsigned long addr, unsigned long *flags) { return NULL; }
1219
1220static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1221                        { return 1; }
1222static inline int check_object(struct kmem_cache *s, struct page *page,
1223                        void *object, u8 val) { return 1; }
1224static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1225                                        struct page *page) {}
1226static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1227static inline unsigned long kmem_cache_flags(unsigned long object_size,
1228        unsigned long flags, const char *name,
1229        void (*ctor)(void *))
1230{
1231        return flags;
1232}
1233#define slub_debug 0
1234
1235#define disable_higher_order_debug 0
1236
1237static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1238                                                        { return 0; }
1239static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1240                                                        { return 0; }
1241static inline void inc_slabs_node(struct kmem_cache *s, int node,
1242                                                        int objects) {}
1243static inline void dec_slabs_node(struct kmem_cache *s, int node,
1244                                                        int objects) {}
1245
1246static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1247                                                        { return 0; }
1248
1249static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1250                void *object) {}
1251
1252static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1253
1254#endif /* CONFIG_SLUB_DEBUG */
1255
1256/*
1257 * Slab allocation and freeing
1258 */
1259static inline struct page *alloc_slab_page(gfp_t flags, int node,
1260                                        struct kmem_cache_order_objects oo)
1261{
1262        int order = oo_order(oo);
1263
1264        flags |= __GFP_NOTRACK;
1265
1266        if (node == NUMA_NO_NODE)
1267                return alloc_pages(flags, order);
1268        else
1269                return alloc_pages_exact_node(node, flags, order);
1270}
1271
1272static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1273{
1274        struct page *page;
1275        struct kmem_cache_order_objects oo = s->oo;
1276        gfp_t alloc_gfp;
1277
1278        flags &= gfp_allowed_mask;
1279
1280        if (flags & __GFP_WAIT)
1281                local_irq_enable();
1282
1283        flags |= s->allocflags;
1284
1285        /*
1286         * Let the initial higher-order allocation fail under memory pressure
1287         * so we fall-back to the minimum order allocation.
1288         */
1289        alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1290
1291        page = alloc_slab_page(alloc_gfp, node, oo);
1292        if (unlikely(!page)) {
1293                oo = s->min;
1294                /*
1295                 * Allocation may have failed due to fragmentation.
1296                 * Try a lower order alloc if possible
1297                 */
1298                page = alloc_slab_page(flags, node, oo);
1299
1300                if (page)
1301                        stat(s, ORDER_FALLBACK);
1302        }
1303
1304        if (kmemcheck_enabled && page
1305                && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1306                int pages = 1 << oo_order(oo);
1307
1308                kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1309
1310                /*
1311                 * Objects from caches that have a constructor don't get
1312                 * cleared when they're allocated, so we need to do it here.
1313                 */
1314                if (s->ctor)
1315                        kmemcheck_mark_uninitialized_pages(page, pages);
1316                else
1317                        kmemcheck_mark_unallocated_pages(page, pages);
1318        }
1319
1320        if (flags & __GFP_WAIT)
1321                local_irq_disable();
1322        if (!page)
1323                return NULL;
1324
1325        page->objects = oo_objects(oo);
1326        mod_zone_page_state(page_zone(page),
1327                (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1328                NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1329                1 << oo_order(oo));
1330
1331        return page;
1332}
1333
1334static void setup_object(struct kmem_cache *s, struct page *page,
1335                                void *object)
1336{
1337        setup_object_debug(s, page, object);
1338        if (unlikely(s->ctor))
1339                s->ctor(object);
1340}
1341
1342static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1343{
1344        struct page *page;
1345        void *start;
1346        void *last;
1347        void *p;
1348        int order;
1349
1350        BUG_ON(flags & GFP_SLAB_BUG_MASK);
1351
1352        page = allocate_slab(s,
1353                flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1354        if (!page)
1355                goto out;
1356
1357        order = compound_order(page);
1358        inc_slabs_node(s, page_to_nid(page), page->objects);
1359        memcg_bind_pages(s, order);
1360        page->slab_cache = 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 << order);
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
1410        memcg_release_pages(s, order);
1411        reset_page_mapcount(page);
1412        if (current->reclaim_state)
1413                current->reclaim_state->reclaimed_slab += pages;
1414        __free_memcg_kmem_pages(page, order);
1415}
1416
1417#define need_reserve_slab_rcu                                           \
1418        (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1419
1420static void rcu_free_slab(struct rcu_head *h)
1421{
1422        struct page *page;
1423
1424        if (need_reserve_slab_rcu)
1425                page = virt_to_head_page(h);
1426        else
1427                page = container_of((struct list_head *)h, struct page, lru);
1428
1429        __free_slab(page->slab_cache, page);
1430}
1431
1432static void free_slab(struct kmem_cache *s, struct page *page)
1433{
1434        if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1435                struct rcu_head *head;
1436
1437                if (need_reserve_slab_rcu) {
1438                        int order = compound_order(page);
1439                        int offset = (PAGE_SIZE << order) - s->reserved;
1440
1441                        VM_BUG_ON(s->reserved != sizeof(*head));
1442                        head = page_address(page) + offset;
1443                } else {
1444                        /*
1445                         * RCU free overloads the RCU head over the LRU
1446                         */
1447                        head = (void *)&page->lru;
1448                }
1449
1450                call_rcu(head, rcu_free_slab);
1451        } else
1452                __free_slab(s, page);
1453}
1454
1455static void discard_slab(struct kmem_cache *s, struct page *page)
1456{
1457        dec_slabs_node(s, page_to_nid(page), page->objects);
1458        free_slab(s, page);
1459}
1460
1461/*
1462 * Management of partially allocated slabs.
1463 *
1464 * list_lock must be held.
1465 */
1466static inline void add_partial(struct kmem_cache_node *n,
1467                                struct page *page, int tail)
1468{
1469        n->nr_partial++;
1470        if (tail == DEACTIVATE_TO_TAIL)
1471                list_add_tail(&page->lru, &n->partial);
1472        else
1473                list_add(&page->lru, &n->partial);
1474}
1475
1476/*
1477 * list_lock must be held.
1478 */
1479static inline void remove_partial(struct kmem_cache_node *n,
1480                                        struct page *page)
1481{
1482        list_del(&page->lru);
1483        n->nr_partial--;
1484}
1485
1486/*
1487 * Remove slab from the partial list, freeze it and
1488 * return the pointer to the freelist.
1489 *
1490 * Returns a list of objects or NULL if it fails.
1491 *
1492 * Must hold list_lock since we modify the partial list.
1493 */
1494static inline void *acquire_slab(struct kmem_cache *s,
1495                struct kmem_cache_node *n, struct page *page,
1496                int mode)
1497{
1498        void *freelist;
1499        unsigned long counters;
1500        struct page new;
1501
1502        /*
1503         * Zap the freelist and set the frozen bit.
1504         * The old freelist is the list of objects for the
1505         * per cpu allocation list.
1506         */
1507        freelist = page->freelist;
1508        counters = page->counters;
1509        new.counters = counters;
1510        if (mode) {
1511                new.inuse = page->objects;
1512                new.freelist = NULL;
1513        } else {
1514                new.freelist = freelist;
1515        }
1516
1517        VM_BUG_ON(new.frozen);
1518        new.frozen = 1;
1519
1520        if (!__cmpxchg_double_slab(s, page,
1521                        freelist, counters,
1522                        new.freelist, new.counters,
1523                        "acquire_slab"))
1524                return NULL;
1525
1526        remove_partial(n, page);
1527        WARN_ON(!freelist);
1528        return freelist;
1529}
1530
1531static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1532static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1533
1534/*
1535 * Try to allocate a partial slab from a specific node.
1536 */
1537static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1538                                struct kmem_cache_cpu *c, gfp_t flags)
1539{
1540        struct page *page, *page2;
1541        void *object = NULL;
1542
1543        /*
1544         * Racy check. If we mistakenly see no partial slabs then we
1545         * just allocate an empty slab. If we mistakenly try to get a
1546         * partial slab and there is none available then get_partials()
1547         * will return NULL.
1548         */
1549        if (!n || !n->nr_partial)
1550                return NULL;
1551
1552        spin_lock(&n->list_lock);
1553        list_for_each_entry_safe(page, page2, &n->partial, lru) {
1554                void *t;
1555                int available;
1556
1557                if (!pfmemalloc_match(page, flags))
1558                        continue;
1559
1560                t = acquire_slab(s, n, page, object == NULL);
1561                if (!t)
1562                        break;
1563
1564                if (!object) {
1565                        c->page = page;
1566                        stat(s, ALLOC_FROM_PARTIAL);
1567                        object = t;
1568                        available =  page->objects - page->inuse;
1569                } else {
1570                        available = put_cpu_partial(s, page, 0);
1571                        stat(s, CPU_PARTIAL_NODE);
1572                }
1573                if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1574                        break;
1575
1576        }
1577        spin_unlock(&n->list_lock);
1578        return object;
1579}
1580
1581/*
1582 * Get a page from somewhere. Search in increasing NUMA distances.
1583 */
1584static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1585                struct kmem_cache_cpu *c)
1586{
1587#ifdef CONFIG_NUMA
1588        struct zonelist *zonelist;
1589        struct zoneref *z;
1590        struct zone *zone;
1591        enum zone_type high_zoneidx = gfp_zone(flags);
1592        void *object;
1593        unsigned int cpuset_mems_cookie;
1594
1595        /*
1596         * The defrag ratio allows a configuration of the tradeoffs between
1597         * inter node defragmentation and node local allocations. A lower
1598         * defrag_ratio increases the tendency to do local allocations
1599         * instead of attempting to obtain partial slabs from other nodes.
1600         *
1601         * If the defrag_ratio is set to 0 then kmalloc() always
1602         * returns node local objects. If the ratio is higher then kmalloc()
1603         * may return off node objects because partial slabs are obtained
1604         * from other nodes and filled up.
1605         *
1606         * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1607         * defrag_ratio = 1000) then every (well almost) allocation will
1608         * first attempt to defrag slab caches on other nodes. This means
1609         * scanning over all nodes to look for partial slabs which may be
1610         * expensive if we do it every time we are trying to find a slab
1611         * with available objects.
1612         */
1613        if (!s->remote_node_defrag_ratio ||
1614                        get_cycles() % 1024 > s->remote_node_defrag_ratio)
1615                return NULL;
1616
1617        do {
1618                cpuset_mems_cookie = get_mems_allowed();
1619                zonelist = node_zonelist(slab_node(), flags);
1620                for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1621                        struct kmem_cache_node *n;
1622
1623                        n = get_node(s, zone_to_nid(zone));
1624
1625                        if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1626                                        n->nr_partial > s->min_partial) {
1627                                object = get_partial_node(s, n, c, flags);
1628                                if (object) {
1629                                        /*
1630                                         * Return the object even if
1631                                         * put_mems_allowed indicated that
1632                                         * the cpuset mems_allowed was
1633                                         * updated in parallel. It's a
1634                                         * harmless race between the alloc
1635                                         * and the cpuset update.
1636                                         */
1637                                        put_mems_allowed(cpuset_mems_cookie);
1638                                        return object;
1639                                }
1640                        }
1641                }
1642        } while (!put_mems_allowed(cpuset_mems_cookie));
1643#endif
1644        return NULL;
1645}
1646
1647/*
1648 * Get a partial page, lock it and return it.
1649 */
1650static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1651                struct kmem_cache_cpu *c)
1652{
1653        void *object;
1654        int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1655
1656        object = get_partial_node(s, get_node(s, searchnode), c, flags);
1657        if (object || node != NUMA_NO_NODE)
1658                return object;
1659
1660        return get_any_partial(s, flags, c);
1661}
1662
1663#ifdef CONFIG_PREEMPT
1664/*
1665 * Calculate the next globally unique transaction for disambiguiation
1666 * during cmpxchg. The transactions start with the cpu number and are then
1667 * incremented by CONFIG_NR_CPUS.
1668 */
1669#define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
1670#else
1671/*
1672 * No preemption supported therefore also no need to check for
1673 * different cpus.
1674 */
1675#define TID_STEP 1
1676#endif
1677
1678static inline unsigned long next_tid(unsigned long tid)
1679{
1680        return tid + TID_STEP;
1681}
1682
1683static inline unsigned int tid_to_cpu(unsigned long tid)
1684{
1685        return tid % TID_STEP;
1686}
1687
1688static inline unsigned long tid_to_event(unsigned long tid)
1689{
1690        return tid / TID_STEP;
1691}
1692
1693static inline unsigned int init_tid(int cpu)
1694{
1695        return cpu;
1696}
1697
1698static inline void note_cmpxchg_failure(const char *n,
1699                const struct kmem_cache *s, unsigned long tid)
1700{
1701#ifdef SLUB_DEBUG_CMPXCHG
1702        unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1703
1704        printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1705
1706#ifdef CONFIG_PREEMPT
1707        if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1708                printk("due to cpu change %d -> %d\n",
1709                        tid_to_cpu(tid), tid_to_cpu(actual_tid));
1710        else
1711#endif
1712        if (tid_to_event(tid) != tid_to_event(actual_tid))
1713                printk("due to cpu running other code. Event %ld->%ld\n",
1714                        tid_to_event(tid), tid_to_event(actual_tid));
1715        else
1716                printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1717                        actual_tid, tid, next_tid(tid));
1718#endif
1719        stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1720}
1721
1722static void init_kmem_cache_cpus(struct kmem_cache *s)
1723{
1724        int cpu;
1725
1726        for_each_possible_cpu(cpu)
1727                per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1728}
1729
1730/*
1731 * Remove the cpu slab
1732 */
1733static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1734{
1735        enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1736        struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1737        int lock = 0;
1738        enum slab_modes l = M_NONE, m = M_NONE;
1739        void *nextfree;
1740        int tail = DEACTIVATE_TO_HEAD;
1741        struct page new;
1742        struct page old;
1743
1744        if (page->freelist) {
1745                stat(s, DEACTIVATE_REMOTE_FREES);
1746                tail = DEACTIVATE_TO_TAIL;
1747        }
1748
1749        /*
1750         * Stage one: Free all available per cpu objects back
1751         * to the page freelist while it is still frozen. Leave the
1752         * last one.
1753         *
1754         * There is no need to take the list->lock because the page
1755         * is still frozen.
1756         */
1757        while (freelist && (nextfree = get_freepointer(s, freelist))) {
1758                void *prior;
1759                unsigned long counters;
1760
1761                do {
1762                        prior = page->freelist;
1763                        counters = page->counters;
1764                        set_freepointer(s, freelist, prior);
1765                        new.counters = counters;
1766                        new.inuse--;
1767                        VM_BUG_ON(!new.frozen);
1768
1769                } while (!__cmpxchg_double_slab(s, page,
1770                        prior, counters,
1771                        freelist, new.counters,
1772                        "drain percpu freelist"));
1773
1774                freelist = nextfree;
1775        }
1776
1777        /*
1778         * Stage two: Ensure that the page is unfrozen while the
1779         * list presence reflects the actual number of objects
1780         * during unfreeze.
1781         *
1782         * We setup the list membership and then perform a cmpxchg
1783         * with the count. If there is a mismatch then the page
1784         * is not unfrozen but the page is on the wrong list.
1785         *
1786         * Then we restart the process which may have to remove
1787         * the page from the list that we just put it on again
1788         * because the number of objects in the slab may have
1789         * changed.
1790         */
1791redo:
1792
1793        old.freelist = page->freelist;
1794        old.counters = page->counters;
1795        VM_BUG_ON(!old.frozen);
1796
1797        /* Determine target state of the slab */
1798        new.counters = old.counters;
1799        if (freelist) {
1800                new.inuse--;
1801                set_freepointer(s, freelist, old.freelist);
1802                new.freelist = freelist;
1803        } else
1804                new.freelist = old.freelist;
1805
1806        new.frozen = 0;
1807
1808        if (!new.inuse && n->nr_partial > s->min_partial)
1809                m = M_FREE;
1810        else if (new.freelist) {
1811                m = M_PARTIAL;
1812                if (!lock) {
1813                        lock = 1;
1814                        /*
1815                         * Taking the spinlock removes the possiblity
1816                         * that acquire_slab() will see a slab page that
1817                         * is frozen
1818                         */
1819                        spin_lock(&n->list_lock);
1820                }
1821        } else {
1822                m = M_FULL;
1823                if (kmem_cache_debug(s) && !lock) {
1824                        lock = 1;
1825                        /*
1826                         * This also ensures that the scanning of full
1827                         * slabs from diagnostic functions will not see
1828                         * any frozen slabs.
1829                         */
1830                        spin_lock(&n->list_lock);
1831                }
1832        }
1833
1834        if (l != m) {
1835
1836                if (l == M_PARTIAL)
1837
1838                        remove_partial(n, page);
1839
1840                else if (l == M_FULL)
1841
1842                        remove_full(s, page);
1843
1844                if (m == M_PARTIAL) {
1845
1846                        add_partial(n, page, tail);
1847                        stat(s, tail);
1848
1849                } else if (m == M_FULL) {
1850
1851                        stat(s, DEACTIVATE_FULL);
1852                        add_full(s, n, page);
1853
1854                }
1855        }
1856
1857        l = m;
1858        if (!__cmpxchg_double_slab(s, page,
1859                                old.freelist, old.counters,
1860                                new.freelist, new.counters,
1861                                "unfreezing slab"))
1862                goto redo;
1863
1864        if (lock)
1865                spin_unlock(&n->list_lock);
1866
1867        if (m == M_FREE) {
1868                stat(s, DEACTIVATE_EMPTY);
1869                discard_slab(s, page);
1870                stat(s, FREE_SLAB);
1871        }
1872}
1873
1874/*
1875 * Unfreeze all the cpu partial slabs.
1876 *
1877 * This function must be called with interrupts disabled
1878 * for the cpu using c (or some other guarantee must be there
1879 * to guarantee no concurrent accesses).
1880 */
1881static void unfreeze_partials(struct kmem_cache *s,
1882                struct kmem_cache_cpu *c)
1883{
1884        struct kmem_cache_node *n = NULL, *n2 = NULL;
1885        struct page *page, *discard_page = NULL;
1886
1887        while ((page = c->partial)) {
1888                struct page new;
1889                struct page old;
1890
1891                c->partial = page->next;
1892
1893                n2 = get_node(s, page_to_nid(page));
1894                if (n != n2) {
1895                        if (n)
1896                                spin_unlock(&n->list_lock);
1897
1898                        n = n2;
1899                        spin_lock(&n->list_lock);
1900                }
1901
1902                do {
1903
1904                        old.freelist = page->freelist;
1905                        old.counters = page->counters;
1906                        VM_BUG_ON(!old.frozen);
1907
1908                        new.counters = old.counters;
1909                        new.freelist = old.freelist;
1910
1911                        new.frozen = 0;
1912
1913                } while (!__cmpxchg_double_slab(s, page,
1914                                old.freelist, old.counters,
1915                                new.freelist, new.counters,
1916                                "unfreezing slab"));
1917
1918                if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1919                        page->next = discard_page;
1920                        discard_page = page;
1921                } else {
1922                        add_partial(n, page, DEACTIVATE_TO_TAIL);
1923                        stat(s, FREE_ADD_PARTIAL);
1924                }
1925        }
1926
1927        if (n)
1928                spin_unlock(&n->list_lock);
1929
1930        while (discard_page) {
1931                page = discard_page;
1932                discard_page = discard_page->next;
1933
1934                stat(s, DEACTIVATE_EMPTY);
1935                discard_slab(s, page);
1936                stat(s, FREE_SLAB);
1937        }
1938}
1939
1940/*
1941 * Put a page that was just frozen (in __slab_free) into a partial page
1942 * slot if available. This is done without interrupts disabled and without
1943 * preemption disabled. The cmpxchg is racy and may put the partial page
1944 * onto a random cpus partial slot.
1945 *
1946 * If we did not find a slot then simply move all the partials to the
1947 * per node partial list.
1948 */
1949static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1950{
1951        struct page *oldpage;
1952        int pages;
1953        int pobjects;
1954
1955        do {
1956                pages = 0;
1957                pobjects = 0;
1958                oldpage = this_cpu_read(s->cpu_slab->partial);
1959
1960                if (oldpage) {
1961                        pobjects = oldpage->pobjects;
1962                        pages = oldpage->pages;
1963                        if (drain && pobjects > s->cpu_partial) {
1964                                unsigned long flags;
1965                                /*
1966                                 * partial array is full. Move the existing
1967                                 * set to the per node partial list.
1968                                 */
1969                                local_irq_save(flags);
1970                                unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
1971                                local_irq_restore(flags);
1972                                oldpage = NULL;
1973                                pobjects = 0;
1974                                pages = 0;
1975                                stat(s, CPU_PARTIAL_DRAIN);
1976                        }
1977                }
1978
1979                pages++;
1980                pobjects += page->objects - page->inuse;
1981
1982                page->pages = pages;
1983                page->pobjects = pobjects;
1984                page->next = oldpage;
1985
1986        } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1987        return pobjects;
1988}
1989
1990static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1991{
1992        stat(s, CPUSLAB_FLUSH);
1993        deactivate_slab(s, c->page, c->freelist);
1994
1995        c->tid = next_tid(c->tid);
1996        c->page = NULL;
1997        c->freelist = NULL;
1998}
1999
2000/*
2001 * Flush cpu slab.
2002 *
2003 * Called from IPI handler with interrupts disabled.
2004 */
2005static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2006{
2007        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2008
2009        if (likely(c)) {
2010                if (c->page)
2011                        flush_slab(s, c);
2012
2013                unfreeze_partials(s, c);
2014        }
2015}
2016
2017static void flush_cpu_slab(void *d)
2018{
2019        struct kmem_cache *s = d;
2020
2021        __flush_cpu_slab(s, smp_processor_id());
2022}
2023
2024static bool has_cpu_slab(int cpu, void *info)
2025{
2026        struct kmem_cache *s = info;
2027        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2028
2029        return c->page || c->partial;
2030}
2031
2032static void flush_all(struct kmem_cache *s)
2033{
2034        on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2035}
2036
2037/*
2038 * Check if the objects in a per cpu structure fit numa
2039 * locality expectations.
2040 */
2041static inline int node_match(struct page *page, int node)
2042{
2043#ifdef CONFIG_NUMA
2044        if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2045                return 0;
2046#endif
2047        return 1;
2048}
2049
2050static int count_free(struct page *page)
2051{
2052        return page->objects - page->inuse;
2053}
2054
2055static unsigned long count_partial(struct kmem_cache_node *n,
2056                                        int (*get_count)(struct page *))
2057{
2058        unsigned long flags;
2059        unsigned long x = 0;
2060        struct page *page;
2061
2062        spin_lock_irqsave(&n->list_lock, flags);
2063        list_for_each_entry(page, &n->partial, lru)
2064                x += get_count(page);
2065        spin_unlock_irqrestore(&n->list_lock, flags);
2066        return x;
2067}
2068
2069static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2070{
2071#ifdef CONFIG_SLUB_DEBUG
2072        return atomic_long_read(&n->total_objects);
2073#else
2074        return 0;
2075#endif
2076}
2077
2078static noinline void
2079slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2080{
2081        int node;
2082
2083        printk(KERN_WARNING
2084                "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2085                nid, gfpflags);
2086        printk(KERN_WARNING "  cache: %s, object size: %d, buffer size: %d, "
2087                "default order: %d, min order: %d\n", s->name, s->object_size,
2088                s->size, oo_order(s->oo), oo_order(s->min));
2089
2090        if (oo_order(s->min) > get_order(s->object_size))
2091                printk(KERN_WARNING "  %s debugging increased min order, use "
2092                       "slub_debug=O to disable.\n", s->name);
2093
2094        for_each_online_node(node) {
2095                struct kmem_cache_node *n = get_node(s, node);
2096                unsigned long nr_slabs;
2097                unsigned long nr_objs;
2098                unsigned long nr_free;
2099
2100                if (!n)
2101                        continue;
2102
2103                nr_free  = count_partial(n, count_free);
2104                nr_slabs = node_nr_slabs(n);
2105                nr_objs  = node_nr_objs(n);
2106
2107                printk(KERN_WARNING
2108                        "  node %d: slabs: %ld, objs: %ld, free: %ld\n",
2109                        node, nr_slabs, nr_objs, nr_free);
2110        }
2111}
2112
2113static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2114                        int node, struct kmem_cache_cpu **pc)
2115{
2116        void *freelist;
2117        struct kmem_cache_cpu *c = *pc;
2118        struct page *page;
2119
2120        freelist = get_partial(s, flags, node, c);
2121
2122        if (freelist)
2123                return freelist;
2124
2125        page = new_slab(s, flags, node);
2126        if (page) {
2127                c = __this_cpu_ptr(s->cpu_slab);
2128                if (c->page)
2129                        flush_slab(s, c);
2130
2131                /*
2132                 * No other reference to the page yet so we can
2133                 * muck around with it freely without cmpxchg
2134                 */
2135                freelist = page->freelist;
2136                page->freelist = NULL;
2137
2138                stat(s, ALLOC_SLAB);
2139                c->page = page;
2140                *pc = c;
2141        } else
2142                freelist = NULL;
2143
2144        return freelist;
2145}
2146
2147static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2148{
2149        if (unlikely(PageSlabPfmemalloc(page)))
2150                return gfp_pfmemalloc_allowed(gfpflags);
2151
2152        return true;
2153}
2154
2155/*
2156 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2157 * or deactivate the page.
2158 *
2159 * The page is still frozen if the return value is not NULL.
2160 *
2161 * If this function returns NULL then the page has been unfrozen.
2162 *
2163 * This function must be called with interrupt disabled.
2164 */
2165static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2166{
2167        struct page new;
2168        unsigned long counters;
2169        void *freelist;
2170
2171        do {
2172                freelist = page->freelist;
2173                counters = page->counters;
2174
2175                new.counters = counters;
2176                VM_BUG_ON(!new.frozen);
2177
2178                new.inuse = page->objects;
2179                new.frozen = freelist != NULL;
2180
2181        } while (!__cmpxchg_double_slab(s, page,
2182                freelist, counters,
2183                NULL, new.counters,
2184                "get_freelist"));
2185
2186        return freelist;
2187}
2188
2189/*
2190 * Slow path. The lockless freelist is empty or we need to perform
2191 * debugging duties.
2192 *
2193 * Processing is still very fast if new objects have been freed to the
2194 * regular freelist. In that case we simply take over the regular freelist
2195 * as the lockless freelist and zap the regular freelist.
2196 *
2197 * If that is not working then we fall back to the partial lists. We take the
2198 * first element of the freelist as the object to allocate now and move the
2199 * rest of the freelist to the lockless freelist.
2200 *
2201 * And if we were unable to get a new slab from the partial slab lists then
2202 * we need to allocate a new slab. This is the slowest path since it involves
2203 * a call to the page allocator and the setup of a new slab.
2204 */
2205static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2206                          unsigned long addr, struct kmem_cache_cpu *c)
2207{
2208        void *freelist;
2209        struct page *page;
2210        unsigned long flags;
2211
2212        local_irq_save(flags);
2213#ifdef CONFIG_PREEMPT
2214        /*
2215         * We may have been preempted and rescheduled on a different
2216         * cpu before disabling interrupts. Need to reload cpu area
2217         * pointer.
2218         */
2219        c = this_cpu_ptr(s->cpu_slab);
2220#endif
2221
2222        page = c->page;
2223        if (!page)
2224                goto new_slab;
2225redo:
2226
2227        if (unlikely(!node_match(page, node))) {
2228                stat(s, ALLOC_NODE_MISMATCH);
2229                deactivate_slab(s, page, c->freelist);
2230                c->page = NULL;
2231                c->freelist = NULL;
2232                goto new_slab;
2233        }
2234
2235        /*
2236         * By rights, we should be searching for a slab page that was
2237         * PFMEMALLOC but right now, we are losing the pfmemalloc
2238         * information when the page leaves the per-cpu allocator
2239         */
2240        if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2241                deactivate_slab(s, page, c->freelist);
2242                c->page = NULL;
2243                c->freelist = NULL;
2244                goto new_slab;
2245        }
2246
2247        /* must check again c->freelist in case of cpu migration or IRQ */
2248        freelist = c->freelist;
2249        if (freelist)
2250                goto load_freelist;
2251
2252        stat(s, ALLOC_SLOWPATH);
2253
2254        freelist = get_freelist(s, page);
2255
2256        if (!freelist) {
2257                c->page = NULL;
2258                stat(s, DEACTIVATE_BYPASS);
2259                goto new_slab;
2260        }
2261
2262        stat(s, ALLOC_REFILL);
2263
2264load_freelist:
2265        /*
2266         * freelist is pointing to the list of objects to be used.
2267         * page is pointing to the page from which the objects are obtained.
2268         * That page must be frozen for per cpu allocations to work.
2269         */
2270        VM_BUG_ON(!c->page->frozen);
2271        c->freelist = get_freepointer(s, freelist);
2272        c->tid = next_tid(c->tid);
2273        local_irq_restore(flags);
2274        return freelist;
2275
2276new_slab:
2277
2278        if (c->partial) {
2279                page = c->page = c->partial;
2280                c->partial = page->next;
2281                stat(s, CPU_PARTIAL_ALLOC);
2282                c->freelist = NULL;
2283                goto redo;
2284        }
2285
2286        freelist = new_slab_objects(s, gfpflags, node, &c);
2287
2288        if (unlikely(!freelist)) {
2289                if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2290                        slab_out_of_memory(s, gfpflags, node);
2291
2292                local_irq_restore(flags);
2293                return NULL;
2294        }
2295
2296        page = c->page;
2297        if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2298                goto load_freelist;
2299
2300        /* Only entered in the debug case */
2301        if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2302                goto new_slab;  /* Slab failed checks. Next slab needed */
2303
2304        deactivate_slab(s, page, get_freepointer(s, freelist));
2305        c->page = NULL;
2306        c->freelist = NULL;
2307        local_irq_restore(flags);
2308        return freelist;
2309}
2310
2311/*
2312 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2313 * have the fastpath folded into their functions. So no function call
2314 * overhead for requests that can be satisfied on the fastpath.
2315 *
2316 * The fastpath works by first checking if the lockless freelist can be used.
2317 * If not then __slab_alloc is called for slow processing.
2318 *
2319 * Otherwise we can simply pick the next object from the lockless free list.
2320 */
2321static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2322                gfp_t gfpflags, int node, unsigned long addr)
2323{
2324        void **object;
2325        struct kmem_cache_cpu *c;
2326        struct page *page;
2327        unsigned long tid;
2328
2329        if (slab_pre_alloc_hook(s, gfpflags))
2330                return NULL;
2331
2332        s = memcg_kmem_get_cache(s, gfpflags);
2333redo:
2334
2335        /*
2336         * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2337         * enabled. We may switch back and forth between cpus while
2338         * reading from one cpu area. That does not matter as long
2339         * as we end up on the original cpu again when doing the cmpxchg.
2340         */
2341        c = __this_cpu_ptr(s->cpu_slab);
2342
2343        /*
2344         * The transaction ids are globally unique per cpu and per operation on
2345         * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2346         * occurs on the right processor and that there was no operation on the
2347         * linked list in between.
2348         */
2349        tid = c->tid;
2350        barrier();
2351
2352        object = c->freelist;
2353        page = c->page;
2354        if (unlikely(!object || !node_match(page, node)))
2355                object = __slab_alloc(s, gfpflags, node, addr, c);
2356
2357        else {
2358                void *next_object = get_freepointer_safe(s, object);
2359
2360                /*
2361                 * The cmpxchg will only match if there was no additional
2362                 * operation and if we are on the right processor.
2363                 *
2364                 * The cmpxchg does the following atomically (without lock semantics!)
2365                 * 1. Relocate first pointer to the current per cpu area.
2366                 * 2. Verify that tid and freelist have not been changed
2367                 * 3. If they were not changed replace tid and freelist
2368                 *
2369                 * Since this is without lock semantics the protection is only against
2370                 * code executing on this cpu *not* from access by other cpus.
2371                 */
2372                if (unlikely(!this_cpu_cmpxchg_double(
2373                                s->cpu_slab->freelist, s->cpu_slab->tid,
2374                                object, tid,
2375                                next_object, next_tid(tid)))) {
2376
2377                        note_cmpxchg_failure("slab_alloc", s, tid);
2378                        goto redo;
2379                }
2380                prefetch_freepointer(s, next_object);
2381                stat(s, ALLOC_FASTPATH);
2382        }
2383
2384        if (unlikely(gfpflags & __GFP_ZERO) && object)
2385                memset(object, 0, s->object_size);
2386
2387        slab_post_alloc_hook(s, gfpflags, object);
2388
2389        return object;
2390}
2391
2392static __always_inline void *slab_alloc(struct kmem_cache *s,
2393                gfp_t gfpflags, unsigned long addr)
2394{
2395        return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2396}
2397
2398void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2399{
2400        void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2401
2402        trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2403
2404        return ret;
2405}
2406EXPORT_SYMBOL(kmem_cache_alloc);
2407
2408#ifdef CONFIG_TRACING
2409void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2410{
2411        void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2412        trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2413        return ret;
2414}
2415EXPORT_SYMBOL(kmem_cache_alloc_trace);
2416
2417void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2418{
2419        void *ret = kmalloc_order(size, flags, order);
2420        trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2421        return ret;
2422}
2423EXPORT_SYMBOL(kmalloc_order_trace);
2424#endif
2425
2426#ifdef CONFIG_NUMA
2427void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2428{
2429        void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2430
2431        trace_kmem_cache_alloc_node(_RET_IP_, ret,
2432                                    s->object_size, s->size, gfpflags, node);
2433
2434        return ret;
2435}
2436EXPORT_SYMBOL(kmem_cache_alloc_node);
2437
2438#ifdef CONFIG_TRACING
2439void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2440                                    gfp_t gfpflags,
2441                                    int node, size_t size)
2442{
2443        void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2444
2445        trace_kmalloc_node(_RET_IP_, ret,
2446                           size, s->size, gfpflags, node);
2447        return ret;
2448}
2449EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2450#endif
2451#endif
2452
2453/*
2454 * Slow patch handling. This may still be called frequently since objects
2455 * have a longer lifetime than the cpu slabs in most processing loads.
2456 *
2457 * So we still attempt to reduce cache line usage. Just take the slab
2458 * lock and free the item. If there is no additional partial page
2459 * handling required then we can return immediately.
2460 */
2461static void __slab_free(struct kmem_cache *s, struct page *page,
2462                        void *x, unsigned long addr)
2463{
2464        void *prior;
2465        void **object = (void *)x;
2466        int was_frozen;
2467        struct page new;
2468        unsigned long counters;
2469        struct kmem_cache_node *n = NULL;
2470        unsigned long uninitialized_var(flags);
2471
2472        stat(s, FREE_SLOWPATH);
2473
2474        if (kmem_cache_debug(s) &&
2475                !(n = free_debug_processing(s, page, x, addr, &flags)))
2476                return;
2477
2478        do {
2479                if (unlikely(n)) {
2480                        spin_unlock_irqrestore(&n->list_lock, flags);
2481                        n = NULL;
2482                }
2483                prior = page->freelist;
2484                counters = page->counters;
2485                set_freepointer(s, object, prior);
2486                new.counters = counters;
2487                was_frozen = new.frozen;
2488                new.inuse--;
2489                if ((!new.inuse || !prior) && !was_frozen) {
2490
2491                        if (!kmem_cache_debug(s) && !prior)
2492
2493                                /*
2494                                 * Slab was on no list before and will be partially empty
2495                                 * We can defer the list move and instead freeze it.
2496                                 */
2497                                new.frozen = 1;
2498
2499                        else { /* Needs to be taken off a list */
2500
2501                                n = get_node(s, page_to_nid(page));
2502                                /*
2503                                 * Speculatively acquire the list_lock.
2504                                 * If the cmpxchg does not succeed then we may
2505                                 * drop the list_lock without any processing.
2506                                 *
2507                                 * Otherwise the list_lock will synchronize with
2508                                 * other processors updating the list of slabs.
2509                                 */
2510                                spin_lock_irqsave(&n->list_lock, flags);
2511
2512                        }
2513                }
2514
2515        } while (!cmpxchg_double_slab(s, page,
2516                prior, counters,
2517                object, new.counters,
2518                "__slab_free"));
2519
2520        if (likely(!n)) {
2521
2522                /*
2523                 * If we just froze the page then put it onto the
2524                 * per cpu partial list.
2525                 */
2526                if (new.frozen && !was_frozen) {
2527                        put_cpu_partial(s, page, 1);
2528                        stat(s, CPU_PARTIAL_FREE);
2529                }
2530                /*
2531                 * The list lock was not taken therefore no list
2532                 * activity can be necessary.
2533                 */
2534                if (was_frozen)
2535                        stat(s, FREE_FROZEN);
2536                return;
2537        }
2538
2539        if (unlikely(!new.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 (kmem_cache_debug(s) && unlikely(!prior)) {
2547                remove_full(s, page);
2548                add_partial(n, page, DEACTIVATE_TO_TAIL);
2549                stat(s, FREE_ADD_PARTIAL);
2550        }
2551        spin_unlock_irqrestore(&n->list_lock, flags);
2552        return;
2553
2554slab_empty:
2555        if (prior) {
2556                /*
2557                 * Slab on the partial list.
2558                 */
2559                remove_partial(n, page);
2560                stat(s, FREE_REMOVE_PARTIAL);
2561        } else
2562                /* Slab must be on the full list */
2563                remove_full(s, page);
2564
2565        spin_unlock_irqrestore(&n->list_lock, flags);
2566        stat(s, FREE_SLAB);
2567        discard_slab(s, page);
2568}
2569
2570/*
2571 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2572 * can perform fastpath freeing without additional function calls.
2573 *
2574 * The fastpath is only possible if we are freeing to the current cpu slab
2575 * of this processor. This typically the case if we have just allocated
2576 * the item before.
2577 *
2578 * If fastpath is not possible then fall back to __slab_free where we deal
2579 * with all sorts of special processing.
2580 */
2581static __always_inline void slab_free(struct kmem_cache *s,
2582                        struct page *page, void *x, unsigned long addr)
2583{
2584        void **object = (void *)x;
2585        struct kmem_cache_cpu *c;
2586        unsigned long tid;
2587
2588        slab_free_hook(s, x);
2589
2590redo:
2591        /*
2592         * Determine the currently cpus per cpu slab.
2593         * The cpu may change afterward. However that does not matter since
2594         * data is retrieved via this pointer. If we are on the same cpu
2595         * during the cmpxchg then the free will succedd.
2596         */
2597        c = __this_cpu_ptr(s->cpu_slab);
2598
2599        tid = c->tid;
2600        barrier();
2601
2602        if (likely(page == c->page)) {
2603                set_freepointer(s, object, c->freelist);
2604
2605                if (unlikely(!this_cpu_cmpxchg_double(
2606                                s->cpu_slab->freelist, s->cpu_slab->tid,
2607                                c->freelist, tid,
2608                                object, next_tid(tid)))) {
2609
2610                        note_cmpxchg_failure("slab_free", s, tid);
2611                        goto redo;
2612                }
2613                stat(s, FREE_FASTPATH);
2614        } else
2615                __slab_free(s, page, x, addr);
2616
2617}
2618
2619void kmem_cache_free(struct kmem_cache *s, void *x)
2620{
2621        s = cache_from_obj(s, x);
2622        if (!s)
2623                return;
2624        slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2625        trace_kmem_cache_free(_RET_IP_, x);
2626}
2627EXPORT_SYMBOL(kmem_cache_free);
2628
2629/*
2630 * Object placement in a slab is made very easy because we always start at
2631 * offset 0. If we tune the size of the object to the alignment then we can
2632 * get the required alignment by putting one properly sized object after
2633 * another.
2634 *
2635 * Notice that the allocation order determines the sizes of the per cpu
2636 * caches. Each processor has always one slab available for allocations.
2637 * Increasing the allocation order reduces the number of times that slabs
2638 * must be moved on and off the partial lists and is therefore a factor in
2639 * locking overhead.
2640 */
2641
2642/*
2643 * Mininum / Maximum order of slab pages. This influences locking overhead
2644 * and slab fragmentation. A higher order reduces the number of partial slabs
2645 * and increases the number of allocations possible without having to
2646 * take the list_lock.
2647 */
2648static int slub_min_order;
2649static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2650static int slub_min_objects;
2651
2652/*
2653 * Merge control. If this is set then no merging of slab caches will occur.
2654 * (Could be removed. This was introduced to pacify the merge skeptics.)
2655 */
2656static int slub_nomerge;
2657
2658/*
2659 * Calculate the order of allocation given an slab object size.
2660 *
2661 * The order of allocation has significant impact on performance and other
2662 * system components. Generally order 0 allocations should be preferred since
2663 * order 0 does not cause fragmentation in the page allocator. Larger objects
2664 * be problematic to put into order 0 slabs because there may be too much
2665 * unused space left. We go to a higher order if more than 1/16th of the slab
2666 * would be wasted.
2667 *
2668 * In order to reach satisfactory performance we must ensure that a minimum
2669 * number of objects is in one slab. Otherwise we may generate too much
2670 * activity on the partial lists which requires taking the list_lock. This is
2671 * less a concern for large slabs though which are rarely used.
2672 *
2673 * slub_max_order specifies the order where we begin to stop considering the
2674 * number of objects in a slab as critical. If we reach slub_max_order then
2675 * we try to keep the page order as low as possible. So we accept more waste
2676 * of space in favor of a small page order.
2677 *
2678 * Higher order allocations also allow the placement of more objects in a
2679 * slab and thereby reduce object handling overhead. If the user has
2680 * requested a higher mininum order then we start with that one instead of
2681 * the smallest order which will fit the object.
2682 */
2683static inline int slab_order(int size, int min_objects,
2684                                int max_order, int fract_leftover, int reserved)
2685{
2686        int order;
2687        int rem;
2688        int min_order = slub_min_order;
2689
2690        if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2691                return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2692
2693        for (order = max(min_order,
2694                                fls(min_objects * size - 1) - PAGE_SHIFT);
2695                        order <= max_order; order++) {
2696
2697                unsigned long slab_size = PAGE_SIZE << order;
2698
2699                if (slab_size < min_objects * size + reserved)
2700                        continue;
2701
2702                rem = (slab_size - reserved) % size;
2703
2704                if (rem <= slab_size / fract_leftover)
2705                        break;
2706
2707        }
2708
2709        return order;
2710}
2711
2712static inline int calculate_order(int size, int reserved)
2713{
2714        int order;
2715        int min_objects;
2716        int fraction;
2717        int max_objects;
2718
2719        /*
2720         * Attempt to find best configuration for a slab. This
2721         * works by first attempting to generate a layout with
2722         * the best configuration and backing off gradually.
2723         *
2724         * First we reduce the acceptable waste in a slab. Then
2725         * we reduce the minimum objects required in a slab.
2726         */
2727        min_objects = slub_min_objects;
2728        if (!min_objects)
2729                min_objects = 4 * (fls(nr_cpu_ids) + 1);
2730        max_objects = order_objects(slub_max_order, size, reserved);
2731        min_objects = min(min_objects, max_objects);
2732
2733        while (min_objects > 1) {
2734                fraction = 16;
2735                while (fraction >= 4) {
2736                        order = slab_order(size, min_objects,
2737                                        slub_max_order, fraction, reserved);
2738                        if (order <= slub_max_order)
2739                                return order;
2740                        fraction /= 2;
2741                }
2742                min_objects--;
2743        }
2744
2745        /*
2746         * We were unable to place multiple objects in a slab. Now
2747         * lets see if we can place a single object there.
2748         */
2749        order = slab_order(size, 1, slub_max_order, 1, reserved);
2750        if (order <= slub_max_order)
2751                return order;
2752
2753        /*
2754         * Doh this slab cannot be placed using slub_max_order.
2755         */
2756        order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2757        if (order < MAX_ORDER)
2758                return order;
2759        return -ENOSYS;
2760}
2761
2762static void
2763init_kmem_cache_node(struct kmem_cache_node *n)
2764{
2765        n->nr_partial = 0;
2766        spin_lock_init(&n->list_lock);
2767        INIT_LIST_HEAD(&n->partial);
2768#ifdef CONFIG_SLUB_DEBUG
2769        atomic_long_set(&n->nr_slabs, 0);
2770        atomic_long_set(&n->total_objects, 0);
2771        INIT_LIST_HEAD(&n->full);
2772#endif
2773}
2774
2775static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2776{
2777        BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2778                        SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2779
2780        /*
2781         * Must align to double word boundary for the double cmpxchg
2782         * instructions to work; see __pcpu_double_call_return_bool().
2783         */
2784        s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2785                                     2 * sizeof(void *));
2786
2787        if (!s->cpu_slab)
2788                return 0;
2789
2790        init_kmem_cache_cpus(s);
2791
2792        return 1;
2793}
2794
2795static struct kmem_cache *kmem_cache_node;
2796
2797/*
2798 * No kmalloc_node yet so do it by hand. We know that this is the first
2799 * slab on the node for this slabcache. There are no concurrent accesses
2800 * possible.
2801 *
2802 * Note that this function only works on the kmalloc_node_cache
2803 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2804 * memory on a fresh node that has no slab structures yet.
2805 */
2806static void early_kmem_cache_node_alloc(int node)
2807{
2808        struct page *page;
2809        struct kmem_cache_node *n;
2810
2811        BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2812
2813        page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2814
2815        BUG_ON(!page);
2816        if (page_to_nid(page) != node) {
2817                printk(KERN_ERR "SLUB: Unable to allocate memory from "
2818                                "node %d\n", node);
2819                printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2820                                "in order to be able to continue\n");
2821        }
2822
2823        n = page->freelist;
2824        BUG_ON(!n);
2825        page->freelist = get_freepointer(kmem_cache_node, n);
2826        page->inuse = 1;
2827        page->frozen = 0;
2828        kmem_cache_node->node[node] = n;
2829#ifdef CONFIG_SLUB_DEBUG
2830        init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2831        init_tracking(kmem_cache_node, n);
2832#endif
2833        init_kmem_cache_node(n);
2834        inc_slabs_node(kmem_cache_node, node, page->objects);
2835
2836        add_partial(n, page, DEACTIVATE_TO_HEAD);
2837}
2838
2839static void free_kmem_cache_nodes(struct kmem_cache *s)
2840{
2841        int node;
2842
2843        for_each_node_state(node, N_NORMAL_MEMORY) {
2844                struct kmem_cache_node *n = s->node[node];
2845
2846                if (n)
2847                        kmem_cache_free(kmem_cache_node, n);
2848
2849                s->node[node] = NULL;
2850        }
2851}
2852
2853static int init_kmem_cache_nodes(struct kmem_cache *s)
2854{
2855        int node;
2856
2857        for_each_node_state(node, N_NORMAL_MEMORY) {
2858                struct kmem_cache_node *n;
2859
2860                if (slab_state == DOWN) {
2861                        early_kmem_cache_node_alloc(node);
2862                        continue;
2863                }
2864                n = kmem_cache_alloc_node(kmem_cache_node,
2865                                                GFP_KERNEL, node);
2866
2867                if (!n) {
2868                        free_kmem_cache_nodes(s);
2869                        return 0;
2870                }
2871
2872                s->node[node] = n;
2873                init_kmem_cache_node(n);
2874        }
2875        return 1;
2876}
2877
2878static void set_min_partial(struct kmem_cache *s, unsigned long min)
2879{
2880        if (min < MIN_PARTIAL)
2881                min = MIN_PARTIAL;
2882        else if (min > MAX_PARTIAL)
2883                min = MAX_PARTIAL;
2884        s->min_partial = min;
2885}
2886
2887/*
2888 * calculate_sizes() determines the order and the distribution of data within
2889 * a slab object.
2890 */
2891static int calculate_sizes(struct kmem_cache *s, int forced_order)
2892{
2893        unsigned long flags = s->flags;
2894        unsigned long size = s->object_size;
2895        int order;
2896
2897        /*
2898         * Round up object size to the next word boundary. We can only
2899         * place the free pointer at word boundaries and this determines
2900         * the possible location of the free pointer.
2901         */
2902        size = ALIGN(size, sizeof(void *));
2903
2904#ifdef CONFIG_SLUB_DEBUG
2905        /*
2906         * Determine if we can poison the object itself. If the user of
2907         * the slab may touch the object after free or before allocation
2908         * then we should never poison the object itself.
2909         */
2910        if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2911                        !s->ctor)
2912                s->flags |= __OBJECT_POISON;
2913        else
2914                s->flags &= ~__OBJECT_POISON;
2915
2916
2917        /*
2918         * If we are Redzoning then check if there is some space between the
2919         * end of the object and the free pointer. If not then add an
2920         * additional word to have some bytes to store Redzone information.
2921         */
2922        if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2923                size += sizeof(void *);
2924#endif
2925
2926        /*
2927         * With that we have determined the number of bytes in actual use
2928         * by the object. This is the potential offset to the free pointer.
2929         */
2930        s->inuse = size;
2931
2932        if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2933                s->ctor)) {
2934                /*
2935                 * Relocate free pointer after the object if it is not
2936                 * permitted to overwrite the first word of the object on
2937                 * kmem_cache_free.
2938                 *
2939                 * This is the case if we do RCU, have a constructor or
2940                 * destructor or are poisoning the objects.
2941                 */
2942                s->offset = size;
2943                size += sizeof(void *);
2944        }
2945
2946#ifdef CONFIG_SLUB_DEBUG
2947        if (flags & SLAB_STORE_USER)
2948                /*
2949                 * Need to store information about allocs and frees after
2950                 * the object.
2951                 */
2952                size += 2 * sizeof(struct track);
2953
2954        if (flags & SLAB_RED_ZONE)
2955                /*
2956                 * Add some empty padding so that we can catch
2957                 * overwrites from earlier objects rather than let
2958                 * tracking information or the free pointer be
2959                 * corrupted if a user writes before the start
2960                 * of the object.
2961                 */
2962                size += sizeof(void *);
2963#endif
2964
2965        /*
2966         * SLUB stores one object immediately after another beginning from
2967         * offset 0. In order to align the objects we have to simply size
2968         * each object to conform to the alignment.
2969         */
2970        size = ALIGN(size, s->align);
2971        s->size = size;
2972        if (forced_order >= 0)
2973                order = forced_order;
2974        else
2975                order = calculate_order(size, s->reserved);
2976
2977        if (order < 0)
2978                return 0;
2979
2980        s->allocflags = 0;
2981        if (order)
2982                s->allocflags |= __GFP_COMP;
2983
2984        if (s->flags & SLAB_CACHE_DMA)
2985                s->allocflags |= SLUB_DMA;
2986
2987        if (s->flags & SLAB_RECLAIM_ACCOUNT)
2988                s->allocflags |= __GFP_RECLAIMABLE;
2989
2990        /*
2991         * Determine the number of objects per slab
2992         */
2993        s->oo = oo_make(order, size, s->reserved);
2994        s->min = oo_make(get_order(size), size, s->reserved);
2995        if (oo_objects(s->oo) > oo_objects(s->max))
2996                s->max = s->oo;
2997
2998        return !!oo_objects(s->oo);
2999}
3000
3001static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3002{
3003        s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3004        s->reserved = 0;
3005
3006        if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3007                s->reserved = sizeof(struct rcu_head);
3008
3009        if (!calculate_sizes(s, -1))
3010                goto error;
3011        if (disable_higher_order_debug) {
3012                /*
3013                 * Disable debugging flags that store metadata if the min slab
3014                 * order increased.
3015                 */
3016                if (get_order(s->size) > get_order(s->object_size)) {
3017                        s->flags &= ~DEBUG_METADATA_FLAGS;
3018                        s->offset = 0;
3019                        if (!calculate_sizes(s, -1))
3020                                goto error;
3021                }
3022        }
3023
3024#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3025    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3026        if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3027                /* Enable fast mode */
3028                s->flags |= __CMPXCHG_DOUBLE;
3029#endif
3030
3031        /*
3032         * The larger the object size is, the more pages we want on the partial
3033         * list to avoid pounding the page allocator excessively.
3034         */
3035        set_min_partial(s, ilog2(s->size) / 2);
3036
3037        /*
3038         * cpu_partial determined the maximum number of objects kept in the
3039         * per cpu partial lists of a processor.
3040         *
3041         * Per cpu partial lists mainly contain slabs that just have one
3042         * object freed. If they are used for allocation then they can be
3043         * filled up again with minimal effort. The slab will never hit the
3044         * per node partial lists and therefore no locking will be required.
3045         *
3046         * This setting also determines
3047         *
3048         * A) The number of objects from per cpu partial slabs dumped to the
3049         *    per node list when we reach the limit.
3050         * B) The number of objects in cpu partial slabs to extract from the
3051         *    per node list when we run out of per cpu objects. We only fetch 50%
3052         *    to keep some capacity around for frees.
3053         */
3054        if (kmem_cache_debug(s))
3055                s->cpu_partial = 0;
3056        else if (s->size >= PAGE_SIZE)
3057                s->cpu_partial = 2;
3058        else if (s->size >= 1024)
3059                s->cpu_partial = 6;
3060        else if (s->size >= 256)
3061                s->cpu_partial = 13;
3062        else
3063                s->cpu_partial = 30;
3064
3065#ifdef CONFIG_NUMA
3066        s->remote_node_defrag_ratio = 1000;
3067#endif
3068        if (!init_kmem_cache_nodes(s))
3069                goto error;
3070
3071        if (alloc_kmem_cache_cpus(s))
3072                return 0;
3073
3074        free_kmem_cache_nodes(s);
3075error:
3076        if (flags & SLAB_PANIC)
3077                panic("Cannot create slab %s size=%lu realsize=%u "
3078                        "order=%u offset=%u flags=%lx\n",
3079                        s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3080                        s->offset, flags);
3081        return -EINVAL;
3082}
3083
3084static void list_slab_objects(struct kmem_cache *s, struct page *page,
3085                                                        const char *text)
3086{
3087#ifdef CONFIG_SLUB_DEBUG
3088        void *addr = page_address(page);
3089        void *p;
3090        unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3091                                     sizeof(long), GFP_ATOMIC);
3092        if (!map)
3093                return;
3094        slab_err(s, page, text, s->name);
3095        slab_lock(page);
3096
3097        get_map(s, page, map);
3098        for_each_object(p, s, addr, page->objects) {
3099
3100                if (!test_bit(slab_index(p, s, addr), map)) {
3101                        printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3102                                                        p, p - addr);
3103                        print_tracking(s, p);
3104                }
3105        }
3106        slab_unlock(page);
3107        kfree(map);
3108#endif
3109}
3110
3111/*
3112 * Attempt to free all partial slabs on a node.
3113 * This is called from kmem_cache_close(). We must be the last thread
3114 * using the cache and therefore we do not need to lock anymore.
3115 */
3116static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3117{
3118        struct page *page, *h;
3119
3120        list_for_each_entry_safe(page, h, &n->partial, lru) {
3121                if (!page->inuse) {
3122                        remove_partial(n, page);
3123                        discard_slab(s, page);
3124                } else {
3125                        list_slab_objects(s, page,
3126                        "Objects remaining in %s on kmem_cache_close()");
3127                }
3128        }
3129}
3130
3131/*
3132 * Release all resources used by a slab cache.
3133 */
3134static inline int kmem_cache_close(struct kmem_cache *s)
3135{
3136        int node;
3137
3138        flush_all(s);
3139        /* Attempt to free all objects */
3140        for_each_node_state(node, N_NORMAL_MEMORY) {
3141                struct kmem_cache_node *n = get_node(s, node);
3142
3143                free_partial(s, n);
3144                if (n->nr_partial || slabs_node(s, node))
3145                        return 1;
3146        }
3147        free_percpu(s->cpu_slab);
3148        free_kmem_cache_nodes(s);
3149        return 0;
3150}
3151
3152int __kmem_cache_shutdown(struct kmem_cache *s)
3153{
3154        int rc = kmem_cache_close(s);
3155
3156        if (!rc) {
3157                /*
3158                 * We do the same lock strategy around sysfs_slab_add, see
3159                 * __kmem_cache_create. Because this is pretty much the last
3160                 * operation we do and the lock will be released shortly after
3161                 * that in slab_common.c, we could just move sysfs_slab_remove
3162                 * to a later point in common code. We should do that when we
3163                 * have a common sysfs framework for all allocators.
3164                 */
3165                mutex_unlock(&slab_mutex);
3166                sysfs_slab_remove(s);
3167                mutex_lock(&slab_mutex);
3168        }
3169
3170        return rc;
3171}
3172
3173/********************************************************************
3174 *              Kmalloc subsystem
3175 *******************************************************************/
3176
3177struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3178EXPORT_SYMBOL(kmalloc_caches);
3179
3180#ifdef CONFIG_ZONE_DMA
3181static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3182#endif
3183
3184static int __init setup_slub_min_order(char *str)
3185{
3186        get_option(&str, &slub_min_order);
3187
3188        return 1;
3189}
3190
3191__setup("slub_min_order=", setup_slub_min_order);
3192
3193static int __init setup_slub_max_order(char *str)
3194{
3195        get_option(&str, &slub_max_order);
3196        slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3197
3198        return 1;
3199}
3200
3201__setup("slub_max_order=", setup_slub_max_order);
3202
3203static int __init setup_slub_min_objects(char *str)
3204{
3205        get_option(&str, &slub_min_objects);
3206
3207        return 1;
3208}
3209
3210__setup("slub_min_objects=", setup_slub_min_objects);
3211
3212static int __init setup_slub_nomerge(char *str)
3213{
3214        slub_nomerge = 1;
3215        return 1;
3216}
3217
3218__setup("slub_nomerge", setup_slub_nomerge);
3219
3220/*
3221 * Conversion table for small slabs sizes / 8 to the index in the
3222 * kmalloc array. This is necessary for slabs < 192 since we have non power
3223 * of two cache sizes there. The size of larger slabs can be determined using
3224 * fls.
3225 */
3226static s8 size_index[24] = {
3227        3,      /* 8 */
3228        4,      /* 16 */
3229        5,      /* 24 */
3230        5,      /* 32 */
3231        6,      /* 40 */
3232        6,      /* 48 */
3233        6,      /* 56 */
3234        6,      /* 64 */
3235        1,      /* 72 */
3236        1,      /* 80 */
3237        1,      /* 88 */
3238        1,      /* 96 */
3239        7,      /* 104 */
3240        7,      /* 112 */
3241        7,      /* 120 */
3242        7,      /* 128 */
3243        2,      /* 136 */
3244        2,      /* 144 */
3245        2,      /* 152 */
3246        2,      /* 160 */
3247        2,      /* 168 */
3248        2,      /* 176 */
3249        2,      /* 184 */
3250        2       /* 192 */
3251};
3252
3253static inline int size_index_elem(size_t bytes)
3254{
3255        return (bytes - 1) / 8;
3256}
3257
3258static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3259{
3260        int index;
3261
3262        if (size <= 192) {
3263                if (!size)
3264                        return ZERO_SIZE_PTR;
3265
3266                index = size_index[size_index_elem(size)];
3267        } else
3268                index = fls(size - 1);
3269
3270#ifdef CONFIG_ZONE_DMA
3271        if (unlikely((flags & SLUB_DMA)))
3272                return kmalloc_dma_caches[index];
3273
3274#endif
3275        return kmalloc_caches[index];
3276}
3277
3278void *__kmalloc(size_t size, gfp_t flags)
3279{
3280        struct kmem_cache *s;
3281        void *ret;
3282
3283        if (unlikely(size > SLUB_MAX_SIZE))
3284                return kmalloc_large(size, flags);
3285
3286        s = get_slab(size, flags);
3287
3288        if (unlikely(ZERO_OR_NULL_PTR(s)))
3289                return s;
3290
3291        ret = slab_alloc(s, flags, _RET_IP_);
3292
3293        trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3294
3295        return ret;
3296}
3297EXPORT_SYMBOL(__kmalloc);
3298
3299#ifdef CONFIG_NUMA
3300static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3301{
3302        struct page *page;
3303        void *ptr = NULL;
3304
3305        flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3306        page = alloc_pages_node(node, flags, get_order(size));
3307        if (page)
3308                ptr = page_address(page);
3309
3310        kmemleak_alloc(ptr, size, 1, flags);
3311        return ptr;
3312}
3313
3314void *__kmalloc_node(size_t size, gfp_t flags, int node)
3315{
3316        struct kmem_cache *s;
3317        void *ret;
3318
3319        if (unlikely(size > SLUB_MAX_SIZE)) {
3320                ret = kmalloc_large_node(size, flags, node);
3321
3322                trace_kmalloc_node(_RET_IP_, ret,
3323                                   size, PAGE_SIZE << get_order(size),
3324                                   flags, node);
3325
3326                return ret;
3327        }
3328
3329        s = get_slab(size, flags);
3330
3331        if (unlikely(ZERO_OR_NULL_PTR(s)))
3332                return s;
3333
3334        ret = slab_alloc_node(s, flags, node, _RET_IP_);
3335
3336        trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3337
3338        return ret;
3339}
3340EXPORT_SYMBOL(__kmalloc_node);
3341#endif
3342
3343size_t ksize(const void *object)
3344{
3345        struct page *page;
3346
3347        if (unlikely(object == ZERO_SIZE_PTR))
3348                return 0;
3349
3350        page = virt_to_head_page(object);
3351
3352        if (unlikely(!PageSlab(page))) {
3353                WARN_ON(!PageCompound(page));
3354                return PAGE_SIZE << compound_order(page);
3355        }
3356
3357        return slab_ksize(page->slab_cache);
3358}
3359EXPORT_SYMBOL(ksize);
3360
3361#ifdef CONFIG_SLUB_DEBUG
3362bool verify_mem_not_deleted(const void *x)
3363{
3364        struct page *page;
3365        void *object = (void *)x;
3366        unsigned long flags;
3367        bool rv;
3368
3369        if (unlikely(ZERO_OR_NULL_PTR(x)))
3370                return false;
3371
3372        local_irq_save(flags);
3373
3374        page = virt_to_head_page(x);
3375        if (unlikely(!PageSlab(page))) {
3376                /* maybe it was from stack? */
3377                rv = true;
3378                goto out_unlock;
3379        }
3380
3381        slab_lock(page);
3382        if (on_freelist(page->slab_cache, page, object)) {
3383                object_err(page->slab_cache, page, object, "Object is on free-list");
3384                rv = false;
3385        } else {
3386                rv = true;
3387        }
3388        slab_unlock(page);
3389
3390out_unlock:
3391        local_irq_restore(flags);
3392        return rv;
3393}
3394EXPORT_SYMBOL(verify_mem_not_deleted);
3395#endif
3396
3397void kfree(const void *x)
3398{
3399        struct page *page;
3400        void *object = (void *)x;
3401
3402        trace_kfree(_RET_IP_, x);
3403
3404        if (unlikely(ZERO_OR_NULL_PTR(x)))
3405                return;
3406
3407        page = virt_to_head_page(x);
3408        if (unlikely(!PageSlab(page))) {
3409                BUG_ON(!PageCompound(page));
3410                kmemleak_free(x);
3411                __free_memcg_kmem_pages(page, compound_order(page));
3412                return;
3413        }
3414        slab_free(page->slab_cache, page, object, _RET_IP_);
3415}
3416EXPORT_SYMBOL(kfree);
3417
3418/*
3419 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3420 * the remaining slabs by the number of items in use. The slabs with the
3421 * most items in use come first. New allocations will then fill those up
3422 * and thus they can be removed from the partial lists.
3423 *
3424 * The slabs with the least items are placed last. This results in them
3425 * being allocated from last increasing the chance that the last objects
3426 * are freed in them.
3427 */
3428int kmem_cache_shrink(struct kmem_cache *s)
3429{
3430        int node;
3431        int i;
3432        struct kmem_cache_node *n;
3433        struct page *page;
3434        struct page *t;
3435        int objects = oo_objects(s->max);
3436        struct list_head *slabs_by_inuse =
3437                kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3438        unsigned long flags;
3439
3440        if (!slabs_by_inuse)
3441                return -ENOMEM;
3442
3443        flush_all(s);
3444        for_each_node_state(node, N_NORMAL_MEMORY) {
3445                n = get_node(s, node);
3446
3447                if (!n->nr_partial)
3448                        continue;
3449
3450                for (i = 0; i < objects; i++)
3451                        INIT_LIST_HEAD(slabs_by_inuse + i);
3452
3453                spin_lock_irqsave(&n->list_lock, flags);
3454
3455                /*
3456                 * Build lists indexed by the items in use in each slab.
3457                 *
3458                 * Note that concurrent frees may occur while we hold the
3459                 * list_lock. page->inuse here is the upper limit.
3460                 */
3461                list_for_each_entry_safe(page, t, &n->partial, lru) {
3462                        list_move(&page->lru, slabs_by_inuse + page->inuse);
3463                        if (!page->inuse)
3464                                n->nr_partial--;
3465                }
3466
3467                /*
3468                 * Rebuild the partial list with the slabs filled up most
3469                 * first and the least used slabs at the end.
3470                 */
3471                for (i = objects - 1; i > 0; i--)
3472                        list_splice(slabs_by_inuse + i, n->partial.prev);
3473
3474                spin_unlock_irqrestore(&n->list_lock, flags);
3475
3476                /* Release empty slabs */
3477                list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3478                        discard_slab(s, page);
3479        }
3480
3481        kfree(slabs_by_inuse);
3482        return 0;
3483}
3484EXPORT_SYMBOL(kmem_cache_shrink);
3485
3486#if defined(CONFIG_MEMORY_HOTPLUG)
3487static int slab_mem_going_offline_callback(void *arg)
3488{
3489        struct kmem_cache *s;
3490
3491        mutex_lock(&slab_mutex);
3492        list_for_each_entry(s, &slab_caches, list)
3493                kmem_cache_shrink(s);
3494        mutex_unlock(&slab_mutex);
3495
3496        return 0;
3497}
3498
3499static void slab_mem_offline_callback(void *arg)
3500{
3501        struct kmem_cache_node *n;
3502        struct kmem_cache *s;
3503        struct memory_notify *marg = arg;
3504        int offline_node;
3505
3506        offline_node = marg->status_change_nid_normal;
3507
3508        /*
3509         * If the node still has available memory. we need kmem_cache_node
3510         * for it yet.
3511         */
3512        if (offline_node < 0)
3513                return;
3514
3515        mutex_lock(&slab_mutex);
3516        list_for_each_entry(s, &slab_caches, list) {
3517                n = get_node(s, offline_node);
3518                if (n) {
3519                        /*
3520                         * if n->nr_slabs > 0, slabs still exist on the node
3521                         * that is going down. We were unable to free them,
3522                         * and offline_pages() function shouldn't call this
3523                         * callback. So, we must fail.
3524                         */
3525                        BUG_ON(slabs_node(s, offline_node));
3526
3527                        s->node[offline_node] = NULL;
3528                        kmem_cache_free(kmem_cache_node, n);
3529                }
3530        }
3531        mutex_unlock(&slab_mutex);
3532}
3533
3534static int slab_mem_going_online_callback(void *arg)
3535{
3536        struct kmem_cache_node *n;
3537        struct kmem_cache *s;
3538        struct memory_notify *marg = arg;
3539        int nid = marg->status_change_nid_normal;
3540        int ret = 0;
3541
3542        /*
3543         * If the node's memory is already available, then kmem_cache_node is
3544         * already created. Nothing to do.
3545         */
3546        if (nid < 0)
3547                return 0;
3548
3549        /*
3550         * We are bringing a node online. No memory is available yet. We must
3551         * allocate a kmem_cache_node structure in order to bring the node
3552         * online.
3553         */
3554        mutex_lock(&slab_mutex);
3555        list_for_each_entry(s, &slab_caches, list) {
3556                /*
3557                 * XXX: kmem_cache_alloc_node will fallback to other nodes
3558                 *      since memory is not yet available from the node that
3559                 *      is brought up.
3560                 */
3561                n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3562                if (!n) {
3563                        ret = -ENOMEM;
3564                        goto out;
3565                }
3566                init_kmem_cache_node(n);
3567                s->node[nid] = n;
3568        }
3569out:
3570        mutex_unlock(&slab_mutex);
3571        return ret;
3572}
3573
3574static int slab_memory_callback(struct notifier_block *self,
3575                                unsigned long action, void *arg)
3576{
3577        int ret = 0;
3578
3579        switch (action) {
3580        case MEM_GOING_ONLINE:
3581                ret = slab_mem_going_online_callback(arg);
3582                break;
3583        case MEM_GOING_OFFLINE:
3584                ret = slab_mem_going_offline_callback(arg);
3585                break;
3586        case MEM_OFFLINE:
3587        case MEM_CANCEL_ONLINE:
3588                slab_mem_offline_callback(arg);
3589                break;
3590        case MEM_ONLINE:
3591        case MEM_CANCEL_OFFLINE:
3592                break;
3593        }
3594        if (ret)
3595                ret = notifier_from_errno(ret);
3596        else
3597                ret = NOTIFY_OK;
3598        return ret;
3599}
3600
3601#endif /* CONFIG_MEMORY_HOTPLUG */
3602
3603/********************************************************************
3604 *                      Basic setup of slabs
3605 *******************************************************************/
3606
3607/*
3608 * Used for early kmem_cache structures that were allocated using
3609 * the page allocator. Allocate them properly then fix up the pointers
3610 * that may be pointing to the wrong kmem_cache structure.
3611 */
3612
3613static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3614{
3615        int node;
3616        struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3617
3618        memcpy(s, static_cache, kmem_cache->object_size);
3619
3620        for_each_node_state(node, N_NORMAL_MEMORY) {
3621                struct kmem_cache_node *n = get_node(s, node);
3622                struct page *p;
3623
3624                if (n) {
3625                        list_for_each_entry(p, &n->partial, lru)
3626                                p->slab_cache = s;
3627
3628#ifdef CONFIG_SLUB_DEBUG
3629                        list_for_each_entry(p, &n->full, lru)
3630                                p->slab_cache = s;
3631#endif
3632                }
3633        }
3634        list_add(&s->list, &slab_caches);
3635        return s;
3636}
3637
3638void __init kmem_cache_init(void)
3639{
3640        static __initdata struct kmem_cache boot_kmem_cache,
3641                boot_kmem_cache_node;
3642        int i;
3643        int caches = 2;
3644
3645        if (debug_guardpage_minorder())
3646                slub_max_order = 0;
3647
3648        kmem_cache_node = &boot_kmem_cache_node;
3649        kmem_cache = &boot_kmem_cache;
3650
3651        create_boot_cache(kmem_cache_node, "kmem_cache_node",
3652                sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3653
3654        hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3655
3656        /* Able to allocate the per node structures */
3657        slab_state = PARTIAL;
3658
3659        create_boot_cache(kmem_cache, "kmem_cache",
3660                        offsetof(struct kmem_cache, node) +
3661                                nr_node_ids * sizeof(struct kmem_cache_node *),
3662                       SLAB_HWCACHE_ALIGN);
3663
3664        kmem_cache = bootstrap(&boot_kmem_cache);
3665
3666        /*
3667         * Allocate kmem_cache_node properly from the kmem_cache slab.
3668         * kmem_cache_node is separately allocated so no need to
3669         * update any list pointers.
3670         */
3671        kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3672
3673        /* Now we can use the kmem_cache to allocate kmalloc slabs */
3674
3675        /*
3676         * Patch up the size_index table if we have strange large alignment
3677         * requirements for the kmalloc array. This is only the case for
3678         * MIPS it seems. The standard arches will not generate any code here.
3679         *
3680         * Largest permitted alignment is 256 bytes due to the way we
3681         * handle the index determination for the smaller caches.
3682         *
3683         * Make sure that nothing crazy happens if someone starts tinkering
3684         * around with ARCH_KMALLOC_MINALIGN
3685         */
3686        BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3687                (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3688
3689        for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3690                int elem = size_index_elem(i);
3691                if (elem >= ARRAY_SIZE(size_index))
3692                        break;
3693                size_index[elem] = KMALLOC_SHIFT_LOW;
3694        }
3695
3696        if (KMALLOC_MIN_SIZE == 64) {
3697                /*
3698                 * The 96 byte size cache is not used if the alignment
3699                 * is 64 byte.
3700                 */
3701                for (i = 64 + 8; i <= 96; i += 8)
3702                        size_index[size_index_elem(i)] = 7;
3703        } else if (KMALLOC_MIN_SIZE == 128) {
3704                /*
3705                 * The 192 byte sized cache is not used if the alignment
3706                 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3707                 * instead.
3708                 */
3709                for (i = 128 + 8; i <= 192; i += 8)
3710                        size_index[size_index_elem(i)] = 8;
3711        }
3712
3713        /* Caches that are not of the two-to-the-power-of size */
3714        if (KMALLOC_MIN_SIZE <= 32) {
3715                kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3716                caches++;
3717        }
3718
3719        if (KMALLOC_MIN_SIZE <= 64) {
3720                kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3721                caches++;
3722        }
3723
3724        for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3725                kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3726                caches++;
3727        }
3728
3729        slab_state = UP;
3730
3731        /* Provide the correct kmalloc names now that the caches are up */
3732        if (KMALLOC_MIN_SIZE <= 32) {
3733                kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3734                BUG_ON(!kmalloc_caches[1]->name);
3735        }
3736
3737        if (KMALLOC_MIN_SIZE <= 64) {
3738                kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3739                BUG_ON(!kmalloc_caches[2]->name);
3740        }
3741
3742        for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3743                char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3744
3745                BUG_ON(!s);
3746                kmalloc_caches[i]->name = s;
3747        }
3748
3749#ifdef CONFIG_SMP
3750        register_cpu_notifier(&slab_notifier);
3751#endif
3752
3753#ifdef CONFIG_ZONE_DMA
3754        for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3755                struct kmem_cache *s = kmalloc_caches[i];
3756
3757                if (s && s->size) {
3758                        char *name = kasprintf(GFP_NOWAIT,
3759                                 "dma-kmalloc-%d", s->object_size);
3760
3761                        BUG_ON(!name);
3762                        kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3763                                s->object_size, SLAB_CACHE_DMA);
3764                }
3765        }
3766#endif
3767        printk(KERN_INFO
3768                "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3769                " CPUs=%d, Nodes=%d\n",
3770                caches, cache_line_size(),
3771                slub_min_order, slub_max_order, slub_min_objects,
3772                nr_cpu_ids, nr_node_ids);
3773}
3774
3775void __init kmem_cache_init_late(void)
3776{
3777}
3778
3779/*
3780 * Find a mergeable slab cache
3781 */
3782static int slab_unmergeable(struct kmem_cache *s)
3783{
3784        if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3785                return 1;
3786
3787        if (s->ctor)
3788                return 1;
3789
3790        /*
3791         * We may have set a slab to be unmergeable during bootstrap.
3792         */
3793        if (s->refcount < 0)
3794                return 1;
3795
3796        return 0;
3797}
3798
3799static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
3800                size_t align, unsigned long flags, const char *name,
3801                void (*ctor)(void *))
3802{
3803        struct kmem_cache *s;
3804
3805        if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3806                return NULL;
3807
3808        if (ctor)
3809                return NULL;
3810
3811        size = ALIGN(size, sizeof(void *));
3812        align = calculate_alignment(flags, align, size);
3813        size = ALIGN(size, align);
3814        flags = kmem_cache_flags(size, flags, name, NULL);
3815
3816        list_for_each_entry(s, &slab_caches, list) {
3817                if (slab_unmergeable(s))
3818                        continue;
3819
3820                if (size > s->size)
3821                        continue;
3822
3823                if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3824                                continue;
3825                /*
3826                 * Check if alignment is compatible.
3827                 * Courtesy of Adrian Drzewiecki
3828                 */
3829                if ((s->size & ~(align - 1)) != s->size)
3830                        continue;
3831
3832                if (s->size - size >= sizeof(void *))
3833                        continue;
3834
3835                if (!cache_match_memcg(s, memcg))
3836                        continue;
3837
3838                return s;
3839        }
3840        return NULL;
3841}
3842
3843struct kmem_cache *
3844__kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3845                   size_t align, unsigned long flags, void (*ctor)(void *))
3846{
3847        struct kmem_cache *s;
3848
3849        s = find_mergeable(memcg, size, align, flags, name, ctor);
3850        if (s) {
3851                s->refcount++;
3852                /*
3853                 * Adjust the object sizes so that we clear
3854                 * the complete object on kzalloc.
3855                 */
3856                s->object_size = max(s->object_size, (int)size);
3857                s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3858
3859                if (sysfs_slab_alias(s, name)) {
3860                        s->refcount--;
3861                        s = NULL;
3862                }
3863        }
3864
3865        return s;
3866}
3867
3868int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3869{
3870        int err;
3871
3872        err = kmem_cache_open(s, flags);
3873        if (err)
3874                return err;
3875
3876        /* Mutex is not taken during early boot */
3877        if (slab_state <= UP)
3878                return 0;
3879
3880        memcg_propagate_slab_attrs(s);
3881        mutex_unlock(&slab_mutex);
3882        err = sysfs_slab_add(s);
3883        mutex_lock(&slab_mutex);
3884
3885        if (err)
3886                kmem_cache_close(s);
3887
3888        return err;
3889}
3890
3891#ifdef CONFIG_SMP
3892/*
3893 * Use the cpu notifier to insure that the cpu slabs are flushed when
3894 * necessary.
3895 */
3896static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3897                unsigned long action, void *hcpu)
3898{
3899        long cpu = (long)hcpu;
3900        struct kmem_cache *s;
3901        unsigned long flags;
3902
3903        switch (action) {
3904        case CPU_UP_CANCELED:
3905        case CPU_UP_CANCELED_FROZEN:
3906        case CPU_DEAD:
3907        case CPU_DEAD_FROZEN:
3908                mutex_lock(&slab_mutex);
3909                list_for_each_entry(s, &slab_caches, list) {
3910                        local_irq_save(flags);
3911                        __flush_cpu_slab(s, cpu);
3912                        local_irq_restore(flags);
3913                }
3914                mutex_unlock(&slab_mutex);
3915                break;
3916        default:
3917                break;
3918        }
3919        return NOTIFY_OK;
3920}
3921
3922static struct notifier_block __cpuinitdata slab_notifier = {
3923        .notifier_call = slab_cpuup_callback
3924};
3925
3926#endif
3927
3928void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3929{
3930        struct kmem_cache *s;
3931        void *ret;
3932
3933        if (unlikely(size > SLUB_MAX_SIZE))
3934                return kmalloc_large(size, gfpflags);
3935
3936        s = get_slab(size, gfpflags);
3937
3938        if (unlikely(ZERO_OR_NULL_PTR(s)))
3939                return s;
3940
3941        ret = slab_alloc(s, gfpflags, caller);
3942
3943        /* Honor the call site pointer we received. */
3944        trace_kmalloc(caller, ret, size, s->size, gfpflags);
3945
3946        return ret;
3947}
3948
3949#ifdef CONFIG_NUMA
3950void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3951                                        int node, unsigned long caller)
3952{
3953        struct kmem_cache *s;
3954        void *ret;
3955
3956        if (unlikely(size > SLUB_MAX_SIZE)) {
3957                ret = kmalloc_large_node(size, gfpflags, node);
3958
3959                trace_kmalloc_node(caller, ret,
3960                                   size, PAGE_SIZE << get_order(size),
3961                                   gfpflags, node);
3962
3963                return ret;
3964        }
3965
3966        s = get_slab(size, gfpflags);
3967
3968        if (unlikely(ZERO_OR_NULL_PTR(s)))
3969                return s;
3970
3971        ret = slab_alloc_node(s, gfpflags, node, caller);
3972
3973        /* Honor the call site pointer we received. */
3974        trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3975
3976        return ret;
3977}
3978#endif
3979
3980#ifdef CONFIG_SYSFS
3981static int count_inuse(struct page *page)
3982{
3983        return page->inuse;
3984}
3985
3986static int count_total(struct page *page)
3987{
3988        return page->objects;
3989}
3990#endif
3991
3992#ifdef CONFIG_SLUB_DEBUG
3993static int validate_slab(struct kmem_cache *s, struct page *page,
3994                                                unsigned long *map)
3995{
3996        void *p;
3997        void *addr = page_address(page);
3998
3999        if (!check_slab(s, page) ||
4000                        !on_freelist(s, page, NULL))
4001                return 0;
4002
4003        /* Now we know that a valid freelist exists */
4004        bitmap_zero(map, page->objects);
4005
4006        get_map(s, page, map);
4007        for_each_object(p, s, addr, page->objects) {
4008                if (test_bit(slab_index(p, s, addr), map))
4009                        if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4010                                return 0;
4011        }
4012
4013        for_each_object(p, s, addr, page->objects)
4014                if (!test_bit(slab_index(p, s, addr), map))
4015                        if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4016                                return 0;
4017        return 1;
4018}
4019
4020static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4021                                                unsigned long *map)
4022{
4023        slab_lock(page);
4024        validate_slab(s, page, map);
4025        slab_unlock(page);
4026}
4027
4028static int validate_slab_node(struct kmem_cache *s,
4029                struct kmem_cache_node *n, unsigned long *map)
4030{
4031        unsigned long count = 0;
4032        struct page *page;
4033        unsigned long flags;
4034
4035        spin_lock_irqsave(&n->list_lock, flags);
4036
4037        list_for_each_entry(page, &n->partial, lru) {
4038                validate_slab_slab(s, page, map);
4039                count++;
4040        }
4041        if (count != n->nr_partial)
4042                printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4043                        "counter=%ld\n", s->name, count, n->nr_partial);
4044
4045        if (!(s->flags & SLAB_STORE_USER))
4046                goto out;
4047
4048        list_for_each_entry(page, &n->full, lru) {
4049                validate_slab_slab(s, page, map);
4050                count++;
4051        }
4052        if (count != atomic_long_read(&n->nr_slabs))
4053                printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4054                        "counter=%ld\n", s->name, count,
4055                        atomic_long_read(&n->nr_slabs));
4056
4057out:
4058        spin_unlock_irqrestore(&n->list_lock, flags);
4059        return count;
4060}
4061
4062static long validate_slab_cache(struct kmem_cache *s)
4063{
4064        int node;
4065        unsigned long count = 0;
4066        unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4067                                sizeof(unsigned long), GFP_KERNEL);
4068
4069        if (!map)
4070                return -ENOMEM;
4071
4072        flush_all(s);
4073        for_each_node_state(node, N_NORMAL_MEMORY) {
4074                struct kmem_cache_node *n = get_node(s, node);
4075
4076                count += validate_slab_node(s, n, map);
4077        }
4078        kfree(map);
4079        return count;
4080}
4081/*
4082 * Generate lists of code addresses where slabcache objects are allocated
4083 * and freed.
4084 */
4085
4086struct location {
4087        unsigned long count;
4088        unsigned long addr;
4089        long long sum_time;
4090        long min_time;
4091        long max_time;
4092        long min_pid;
4093        long max_pid;
4094        DECLARE_BITMAP(cpus, NR_CPUS);
4095        nodemask_t nodes;
4096};
4097
4098struct loc_track {
4099        unsigned long max;
4100        unsigned long count;
4101        struct location *loc;
4102};
4103
4104static void free_loc_track(struct loc_track *t)
4105{
4106        if (t->max)
4107                free_pages((unsigned long)t->loc,
4108                        get_order(sizeof(struct location) * t->max));
4109}
4110
4111static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4112{
4113        struct location *l;
4114        int order;
4115
4116        order = get_order(sizeof(struct location) * max);
4117
4118        l = (void *)__get_free_pages(flags, order);
4119        if (!l)
4120                return 0;
4121
4122        if (t->count) {
4123                memcpy(l, t->loc, sizeof(struct location) * t->count);
4124                free_loc_track(t);
4125        }
4126        t->max = max;
4127        t->loc = l;
4128        return 1;
4129}
4130
4131static int add_location(struct loc_track *t, struct kmem_cache *s,
4132                                const struct track *track)
4133{
4134        long start, end, pos;
4135        struct location *l;
4136        unsigned long caddr;
4137        unsigned long age = jiffies - track->when;
4138
4139        start = -1;
4140        end = t->count;
4141
4142        for ( ; ; ) {
4143                pos = start + (end - start + 1) / 2;
4144
4145                /*
4146                 * There is nothing at "end". If we end up there
4147                 * we need to add something to before end.
4148                 */
4149                if (pos == end)
4150                        break;
4151
4152                caddr = t->loc[pos].addr;
4153                if (track->addr == caddr) {
4154
4155                        l = &t->loc[pos];
4156                        l->count++;
4157                        if (track->when) {
4158                                l->sum_time += age;
4159                                if (age < l->min_time)
4160                                        l->min_time = age;
4161                                if (age > l->max_time)
4162                                        l->max_time = age;
4163
4164                                if (track->pid < l->min_pid)
4165                                        l->min_pid = track->pid;
4166                                if (track->pid > l->max_pid)
4167                                        l->max_pid = track->pid;
4168
4169                                cpumask_set_cpu(track->cpu,
4170                                                to_cpumask(l->cpus));
4171                        }
4172                        node_set(page_to_nid(virt_to_page(track)), l->nodes);
4173                        return 1;
4174                }
4175
4176                if (track->addr < caddr)
4177                        end = pos;
4178                else
4179                        start = pos;
4180        }
4181
4182        /*
4183         * Not found. Insert new tracking element.
4184         */
4185        if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4186                return 0;
4187
4188        l = t->loc + pos;
4189        if (pos < t->count)
4190                memmove(l + 1, l,
4191                        (t->count - pos) * sizeof(struct location));
4192        t->count++;
4193        l->count = 1;
4194        l->addr = track->addr;
4195        l->sum_time = age;
4196        l->min_time = age;
4197        l->max_time = age;
4198        l->min_pid = track->pid;
4199        l->max_pid = track->pid;
4200        cpumask_clear(to_cpumask(l->cpus));
4201        cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4202        nodes_clear(l->nodes);
4203        node_set(page_to_nid(virt_to_page(track)), l->nodes);
4204        return 1;
4205}
4206
4207static void process_slab(struct loc_track *t, struct kmem_cache *s,
4208                struct page *page, enum track_item alloc,
4209                unsigned long *map)
4210{
4211        void *addr = page_address(page);
4212        void *p;
4213
4214        bitmap_zero(map, page->objects);
4215        get_map(s, page, map);
4216
4217        for_each_object(p, s, addr, page->objects)
4218                if (!test_bit(slab_index(p, s, addr), map))
4219                        add_location(t, s, get_track(s, p, alloc));
4220}
4221
4222static int list_locations(struct kmem_cache *s, char *buf,
4223                                        enum track_item alloc)
4224{
4225        int len = 0;
4226        unsigned long i;
4227        struct loc_track t = { 0, 0, NULL };
4228        int node;
4229        unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4230                                     sizeof(unsigned long), GFP_KERNEL);
4231
4232        if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4233                                     GFP_TEMPORARY)) {
4234                kfree(map);
4235                return sprintf(buf, "Out of memory\n");
4236        }
4237        /* Push back cpu slabs */
4238        flush_all(s);
4239
4240        for_each_node_state(node, N_NORMAL_MEMORY) {
4241                struct kmem_cache_node *n = get_node(s, node);
4242                unsigned long flags;
4243                struct page *page;
4244
4245                if (!atomic_long_read(&n->nr_slabs))
4246                        continue;
4247
4248                spin_lock_irqsave(&n->list_lock, flags);
4249                list_for_each_entry(page, &n->partial, lru)
4250                        process_slab(&t, s, page, alloc, map);
4251                list_for_each_entry(page, &n->full, lru)
4252                        process_slab(&t, s, page, alloc, map);
4253                spin_unlock_irqrestore(&n->list_lock, flags);
4254        }
4255
4256        for (i = 0; i < t.count; i++) {
4257                struct location *l = &t.loc[i];
4258
4259                if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4260                        break;
4261                len += sprintf(buf + len, "%7ld ", l->count);
4262
4263                if (l->addr)
4264                        len += sprintf(buf + len, "%pS", (void *)l->addr);
4265                else
4266                        len += sprintf(buf + len, "<not-available>");
4267
4268                if (l->sum_time != l->min_time) {
4269                        len += sprintf(buf + len, " age=%ld/%ld/%ld",
4270                                l->min_time,
4271                                (long)div_u64(l->sum_time, l->count),
4272                                l->max_time);
4273                } else
4274                        len += sprintf(buf + len, " age=%ld",
4275                                l->min_time);
4276
4277                if (l->min_pid != l->max_pid)
4278                        len += sprintf(buf + len, " pid=%ld-%ld",
4279                                l->min_pid, l->max_pid);
4280                else
4281                        len += sprintf(buf + len, " pid=%ld",
4282                                l->min_pid);
4283
4284                if (num_online_cpus() > 1 &&
4285                                !cpumask_empty(to_cpumask(l->cpus)) &&
4286                                len < PAGE_SIZE - 60) {
4287                        len += sprintf(buf + len, " cpus=");
4288                        len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4289                                                 to_cpumask(l->cpus));
4290                }
4291
4292                if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4293                                len < PAGE_SIZE - 60) {
4294                        len += sprintf(buf + len, " nodes=");
4295                        len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4296                                        l->nodes);
4297                }
4298
4299                len += sprintf(buf + len, "\n");
4300        }
4301
4302        free_loc_track(&t);
4303        kfree(map);
4304        if (!t.count)
4305                len += sprintf(buf, "No data\n");
4306        return len;
4307}
4308#endif
4309
4310#ifdef SLUB_RESILIENCY_TEST
4311static void resiliency_test(void)
4312{
4313        u8 *p;
4314
4315        BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4316
4317        printk(KERN_ERR "SLUB resiliency testing\n");
4318        printk(KERN_ERR "-----------------------\n");
4319        printk(KERN_ERR "A. Corruption after allocation\n");
4320
4321        p = kzalloc(16, GFP_KERNEL);
4322        p[16] = 0x12;
4323        printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4324                        " 0x12->0x%p\n\n", p + 16);
4325
4326        validate_slab_cache(kmalloc_caches[4]);
4327
4328        /* Hmmm... The next two are dangerous */
4329        p = kzalloc(32, GFP_KERNEL);
4330        p[32 + sizeof(void *)] = 0x34;
4331        printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4332                        " 0x34 -> -0x%p\n", p);
4333        printk(KERN_ERR
4334                "If allocated object is overwritten then not detectable\n\n");
4335
4336        validate_slab_cache(kmalloc_caches[5]);
4337        p = kzalloc(64, GFP_KERNEL);
4338        p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4339        *p = 0x56;
4340        printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4341                                                                        p);
4342        printk(KERN_ERR
4343                "If allocated object is overwritten then not detectable\n\n");
4344        validate_slab_cache(kmalloc_caches[6]);
4345
4346        printk(KERN_ERR "\nB. Corruption after free\n");
4347        p = kzalloc(128, GFP_KERNEL);
4348        kfree(p);
4349        *p = 0x78;
4350        printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4351        validate_slab_cache(kmalloc_caches[7]);
4352
4353        p = kzalloc(256, GFP_KERNEL);
4354        kfree(p);
4355        p[50] = 0x9a;
4356        printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4357                        p);
4358        validate_slab_cache(kmalloc_caches[8]);
4359
4360        p = kzalloc(512, GFP_KERNEL);
4361        kfree(p);
4362        p[512] = 0xab;
4363        printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4364        validate_slab_cache(kmalloc_caches[9]);
4365}
4366#else
4367#ifdef CONFIG_SYSFS
4368static void resiliency_test(void) {};
4369#endif
4370#endif
4371
4372#ifdef CONFIG_SYSFS
4373enum slab_stat_type {
4374        SL_ALL,                 /* All slabs */
4375        SL_PARTIAL,             /* Only partially allocated slabs */
4376        SL_CPU,                 /* Only slabs used for cpu caches */
4377        SL_OBJECTS,             /* Determine allocated objects not slabs */
4378        SL_TOTAL                /* Determine object capacity not slabs */
4379};
4380
4381#define SO_ALL          (1 << SL_ALL)
4382#define SO_PARTIAL      (1 << SL_PARTIAL)
4383#define SO_CPU          (1 << SL_CPU)
4384#define SO_OBJECTS      (1 << SL_OBJECTS)
4385#define SO_TOTAL        (1 << SL_TOTAL)
4386
4387static ssize_t show_slab_objects(struct kmem_cache *s,
4388                            char *buf, unsigned long flags)
4389{
4390        unsigned long total = 0;
4391        int node;
4392        int x;
4393        unsigned long *nodes;
4394        unsigned long *per_cpu;
4395
4396        nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4397        if (!nodes)
4398                return -ENOMEM;
4399        per_cpu = nodes + nr_node_ids;
4400
4401        if (flags & SO_CPU) {
4402                int cpu;
4403
4404                for_each_possible_cpu(cpu) {
4405                        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4406                        int node;
4407                        struct page *page;
4408
4409                        page = ACCESS_ONCE(c->page);
4410                        if (!page)
4411                                continue;
4412
4413                        node = page_to_nid(page);
4414                        if (flags & SO_TOTAL)
4415                                x = page->objects;
4416                        else if (flags & SO_OBJECTS)
4417                                x = page->inuse;
4418                        else
4419                                x = 1;
4420
4421                        total += x;
4422                        nodes[node] += x;
4423
4424                        page = ACCESS_ONCE(c->partial);
4425                        if (page) {
4426                                x = page->pobjects;
4427                                total += x;
4428                                nodes[node] += x;
4429                        }
4430
4431                        per_cpu[node]++;
4432                }
4433        }
4434
4435        lock_memory_hotplug();
4436#ifdef CONFIG_SLUB_DEBUG
4437        if (flags & SO_ALL) {
4438                for_each_node_state(node, N_NORMAL_MEMORY) {
4439                        struct kmem_cache_node *n = get_node(s, node);
4440
4441                if (flags & SO_TOTAL)
4442                        x = atomic_long_read(&n->total_objects);
4443                else if (flags & SO_OBJECTS)
4444                        x = atomic_long_read(&n->total_objects) -
4445                                count_partial(n, count_free);
4446
4447                        else
4448                                x = atomic_long_read(&n->nr_slabs);
4449                        total += x;
4450                        nodes[node] += x;
4451                }
4452
4453        } else
4454#endif
4455        if (flags & SO_PARTIAL) {
4456                for_each_node_state(node, N_NORMAL_MEMORY) {
4457                        struct kmem_cache_node *n = get_node(s, node);
4458
4459                        if (flags & SO_TOTAL)
4460                                x = count_partial(n, count_total);
4461                        else if (flags & SO_OBJECTS)
4462                                x = count_partial(n, count_inuse);
4463                        else
4464                                x = n->nr_partial;
4465                        total += x;
4466                        nodes[node] += x;
4467                }
4468        }
4469        x = sprintf(buf, "%lu", total);
4470#ifdef CONFIG_NUMA
4471        for_each_node_state(node, N_NORMAL_MEMORY)
4472                if (nodes[node])
4473                        x += sprintf(buf + x, " N%d=%lu",
4474                                        node, nodes[node]);
4475#endif
4476        unlock_memory_hotplug();
4477        kfree(nodes);
4478        return x + sprintf(buf + x, "\n");
4479}
4480
4481#ifdef CONFIG_SLUB_DEBUG
4482static int any_slab_objects(struct kmem_cache *s)
4483{
4484        int node;
4485
4486        for_each_online_node(node) {
4487                struct kmem_cache_node *n = get_node(s, node);
4488
4489                if (!n)
4490                        continue;
4491
4492                if (atomic_long_read(&n->total_objects))
4493                        return 1;
4494        }
4495        return 0;
4496}
4497#endif
4498
4499#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4500#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4501
4502struct slab_attribute {
4503        struct attribute attr;
4504        ssize_t (*show)(struct kmem_cache *s, char *buf);
4505        ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4506};
4507
4508#define SLAB_ATTR_RO(_name) \
4509        static struct slab_attribute _name##_attr = \
4510        __ATTR(_name, 0400, _name##_show, NULL)
4511
4512#define SLAB_ATTR(_name) \
4513        static struct slab_attribute _name##_attr =  \
4514        __ATTR(_name, 0600, _name##_show, _name##_store)
4515
4516static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4517{
4518        return sprintf(buf, "%d\n", s->size);
4519}
4520SLAB_ATTR_RO(slab_size);
4521
4522static ssize_t align_show(struct kmem_cache *s, char *buf)
4523{
4524        return sprintf(buf, "%d\n", s->align);
4525}
4526SLAB_ATTR_RO(align);
4527
4528static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4529{
4530        return sprintf(buf, "%d\n", s->object_size);
4531}
4532SLAB_ATTR_RO(object_size);
4533
4534static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4535{
4536        return sprintf(buf, "%d\n", oo_objects(s->oo));
4537}
4538SLAB_ATTR_RO(objs_per_slab);
4539
4540static ssize_t order_store(struct kmem_cache *s,
4541                                const char *buf, size_t length)
4542{
4543        unsigned long order;
4544        int err;
4545
4546        err = strict_strtoul(buf, 10, &order);
4547        if (err)
4548                return err;
4549
4550        if (order > slub_max_order || order < slub_min_order)
4551                return -EINVAL;
4552
4553        calculate_sizes(s, order);
4554        return length;
4555}
4556
4557static ssize_t order_show(struct kmem_cache *s, char *buf)
4558{
4559        return sprintf(buf, "%d\n", oo_order(s->oo));
4560}
4561SLAB_ATTR(order);
4562
4563static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4564{
4565        return sprintf(buf, "%lu\n", s->min_partial);
4566}
4567
4568static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4569                                 size_t length)
4570{
4571        unsigned long min;
4572        int err;
4573
4574        err = strict_strtoul(buf, 10, &min);
4575        if (err)
4576                return err;
4577
4578        set_min_partial(s, min);
4579        return length;
4580}
4581SLAB_ATTR(min_partial);
4582
4583static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4584{
4585        return sprintf(buf, "%u\n", s->cpu_partial);
4586}
4587
4588static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4589                                 size_t length)
4590{
4591        unsigned long objects;
4592        int err;
4593
4594        err = strict_strtoul(buf, 10, &objects);
4595        if (err)
4596                return err;
4597        if (objects && kmem_cache_debug(s))
4598                return -EINVAL;
4599
4600        s->cpu_partial = objects;
4601        flush_all(s);
4602        return length;
4603}
4604SLAB_ATTR(cpu_partial);
4605
4606static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4607{
4608        if (!s->ctor)
4609                return 0;
4610        return sprintf(buf, "%pS\n", s->ctor);
4611}
4612SLAB_ATTR_RO(ctor);
4613
4614static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4615{
4616        return sprintf(buf, "%d\n", s->refcount - 1);
4617}
4618SLAB_ATTR_RO(aliases);
4619
4620static ssize_t partial_show(struct kmem_cache *s, char *buf)
4621{
4622        return show_slab_objects(s, buf, SO_PARTIAL);
4623}
4624SLAB_ATTR_RO(partial);
4625
4626static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4627{
4628        return show_slab_objects(s, buf, SO_CPU);
4629}
4630SLAB_ATTR_RO(cpu_slabs);
4631
4632static ssize_t objects_show(struct kmem_cache *s, char *buf)
4633{
4634        return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4635}
4636SLAB_ATTR_RO(objects);
4637
4638static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4639{
4640        return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4641}
4642SLAB_ATTR_RO(objects_partial);
4643
4644static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4645{
4646        int objects = 0;
4647        int pages = 0;
4648        int cpu;
4649        int len;
4650
4651        for_each_online_cpu(cpu) {
4652                struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4653
4654                if (page) {
4655                        pages += page->pages;
4656                        objects += page->pobjects;
4657                }
4658        }
4659
4660        len = sprintf(buf, "%d(%d)", objects, pages);
4661
4662#ifdef CONFIG_SMP
4663        for_each_online_cpu(cpu) {
4664                struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4665
4666                if (page && len < PAGE_SIZE - 20)
4667                        len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4668                                page->pobjects, page->pages);
4669        }
4670#endif
4671        return len + sprintf(buf + len, "\n");
4672}
4673SLAB_ATTR_RO(slabs_cpu_partial);
4674
4675static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4676{
4677        return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4678}
4679
4680static ssize_t reclaim_account_store(struct kmem_cache *s,
4681                                const char *buf, size_t length)
4682{
4683        s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4684        if (buf[0] == '1')
4685                s->flags |= SLAB_RECLAIM_ACCOUNT;
4686        return length;
4687}
4688SLAB_ATTR(reclaim_account);
4689
4690static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4691{
4692        return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4693}
4694SLAB_ATTR_RO(hwcache_align);
4695
4696#ifdef CONFIG_ZONE_DMA
4697static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4698{
4699        return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4700}
4701SLAB_ATTR_RO(cache_dma);
4702#endif
4703
4704static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4705{
4706        return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4707}
4708SLAB_ATTR_RO(destroy_by_rcu);
4709
4710static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4711{
4712        return sprintf(buf, "%d\n", s->reserved);
4713}
4714SLAB_ATTR_RO(reserved);
4715
4716#ifdef CONFIG_SLUB_DEBUG
4717static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4718{
4719        return show_slab_objects(s, buf, SO_ALL);
4720}
4721SLAB_ATTR_RO(slabs);
4722
4723static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4724{
4725        return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4726}
4727SLAB_ATTR_RO(total_objects);
4728
4729static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4730{
4731        return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4732}
4733
4734static ssize_t sanity_checks_store(struct kmem_cache *s,
4735                                const char *buf, size_t length)
4736{
4737        s->flags &= ~SLAB_DEBUG_FREE;
4738        if (buf[0] == '1') {
4739                s->flags &= ~__CMPXCHG_DOUBLE;
4740                s->flags |= SLAB_DEBUG_FREE;
4741        }
4742        return length;
4743}
4744SLAB_ATTR(sanity_checks);
4745
4746static ssize_t trace_show(struct kmem_cache *s, char *buf)
4747{
4748        return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4749}
4750
4751static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4752                                                        size_t length)
4753{
4754        s->flags &= ~SLAB_TRACE;
4755        if (buf[0] == '1') {
4756                s->flags &= ~__CMPXCHG_DOUBLE;
4757                s->flags |= SLAB_TRACE;
4758        }
4759        return length;
4760}
4761SLAB_ATTR(trace);
4762
4763static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4764{
4765        return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4766}
4767
4768static ssize_t red_zone_store(struct kmem_cache *s,
4769                                const char *buf, size_t length)
4770{
4771        if (any_slab_objects(s))
4772                return -EBUSY;
4773
4774        s->flags &= ~SLAB_RED_ZONE;
4775        if (buf[0] == '1') {
4776                s->flags &= ~__CMPXCHG_DOUBLE;
4777                s->flags |= SLAB_RED_ZONE;
4778        }
4779        calculate_sizes(s, -1);
4780        return length;
4781}
4782SLAB_ATTR(red_zone);
4783
4784static ssize_t poison_show(struct kmem_cache *s, char *buf)
4785{
4786        return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4787}
4788
4789static ssize_t poison_store(struct kmem_cache *s,
4790                                const char *buf, size_t length)
4791{
4792        if (any_slab_objects(s))
4793                return -EBUSY;
4794
4795        s->flags &= ~SLAB_POISON;
4796        if (buf[0] == '1') {
4797                s->flags &= ~__CMPXCHG_DOUBLE;
4798                s->flags |= SLAB_POISON;
4799        }
4800        calculate_sizes(s, -1);
4801        return length;
4802}
4803SLAB_ATTR(poison);
4804
4805static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4806{
4807        return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4808}
4809
4810static ssize_t store_user_store(struct kmem_cache *s,
4811                                const char *buf, size_t length)
4812{
4813        if (any_slab_objects(s))
4814                return -EBUSY;
4815
4816        s->flags &= ~SLAB_STORE_USER;
4817        if (buf[0] == '1') {
4818                s->flags &= ~__CMPXCHG_DOUBLE;
4819                s->flags |= SLAB_STORE_USER;
4820        }
4821        calculate_sizes(s, -1);
4822        return length;
4823}
4824SLAB_ATTR(store_user);
4825
4826static ssize_t validate_show(struct kmem_cache *s, char *buf)
4827{
4828        return 0;
4829}
4830
4831static ssize_t validate_store(struct kmem_cache *s,
4832                        const char *buf, size_t length)
4833{
4834        int ret = -EINVAL;
4835
4836        if (buf[0] == '1') {
4837                ret = validate_slab_cache(s);
4838                if (ret >= 0)
4839                        ret = length;
4840        }
4841        return ret;
4842}
4843SLAB_ATTR(validate);
4844
4845static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4846{
4847        if (!(s->flags & SLAB_STORE_USER))
4848                return -ENOSYS;
4849        return list_locations(s, buf, TRACK_ALLOC);
4850}
4851SLAB_ATTR_RO(alloc_calls);
4852
4853static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4854{
4855        if (!(s->flags & SLAB_STORE_USER))
4856                return -ENOSYS;
4857        return list_locations(s, buf, TRACK_FREE);
4858}
4859SLAB_ATTR_RO(free_calls);
4860#endif /* CONFIG_SLUB_DEBUG */
4861
4862#ifdef CONFIG_FAILSLAB
4863static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4864{
4865        return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4866}
4867
4868static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4869                                                        size_t length)
4870{
4871        s->flags &= ~SLAB_FAILSLAB;
4872        if (buf[0] == '1')
4873                s->flags |= SLAB_FAILSLAB;
4874        return length;
4875}
4876SLAB_ATTR(failslab);
4877#endif
4878
4879static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4880{
4881        return 0;
4882}
4883
4884static ssize_t shrink_store(struct kmem_cache *s,
4885                        const char *buf, size_t length)
4886{
4887        if (buf[0] == '1') {
4888                int rc = kmem_cache_shrink(s);
4889
4890                if (rc)
4891                        return rc;
4892        } else
4893                return -EINVAL;
4894        return length;
4895}
4896SLAB_ATTR(shrink);
4897
4898#ifdef CONFIG_NUMA
4899static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4900{
4901        return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4902}
4903
4904static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4905                                const char *buf, size_t length)
4906{
4907        unsigned long ratio;
4908        int err;
4909
4910        err = strict_strtoul(buf, 10, &ratio);
4911        if (err)
4912                return err;
4913
4914        if (ratio <= 100)
4915                s->remote_node_defrag_ratio = ratio * 10;
4916
4917        return length;
4918}
4919SLAB_ATTR(remote_node_defrag_ratio);
4920#endif
4921
4922#ifdef CONFIG_SLUB_STATS
4923static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4924{
4925        unsigned long sum  = 0;
4926        int cpu;
4927        int len;
4928        int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4929
4930        if (!data)
4931                return -ENOMEM;
4932
4933        for_each_online_cpu(cpu) {
4934                unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4935
4936                data[cpu] = x;
4937                sum += x;
4938        }
4939
4940        len = sprintf(buf, "%lu", sum);
4941
4942#ifdef CONFIG_SMP
4943        for_each_online_cpu(cpu) {
4944                if (data[cpu] && len < PAGE_SIZE - 20)
4945                        len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4946        }
4947#endif
4948        kfree(data);
4949        return len + sprintf(buf + len, "\n");
4950}
4951
4952static void clear_stat(struct kmem_cache *s, enum stat_item si)
4953{
4954        int cpu;
4955
4956        for_each_online_cpu(cpu)
4957                per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4958}
4959
4960#define STAT_ATTR(si, text)                                     \
4961static ssize_t text##_show(struct kmem_cache *s, char *buf)     \
4962{                                                               \
4963        return show_stat(s, buf, si);                           \
4964}                                                               \
4965static ssize_t text##_store(struct kmem_cache *s,               \
4966                                const char *buf, size_t length) \
4967{                                                               \
4968        if (buf[0] != '0')                                      \
4969                return -EINVAL;                                 \
4970        clear_stat(s, si);                                      \
4971        return length;                                          \
4972}                                                               \
4973SLAB_ATTR(text);                                                \
4974
4975STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4976STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4977STAT_ATTR(FREE_FASTPATH, free_fastpath);
4978STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4979STAT_ATTR(FREE_FROZEN, free_frozen);
4980STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4981STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4982STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4983STAT_ATTR(ALLOC_SLAB, alloc_slab);
4984STAT_ATTR(ALLOC_REFILL, alloc_refill);
4985STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4986STAT_ATTR(FREE_SLAB, free_slab);
4987STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4988STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4989STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4990STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4991STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4992STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4993STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4994STAT_ATTR(ORDER_FALLBACK, order_fallback);
4995STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4996STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4997STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4998STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4999STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5000STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5001#endif
5002
5003static struct attribute *slab_attrs[] = {
5004        &slab_size_attr.attr,
5005        &object_size_attr.attr,
5006        &objs_per_slab_attr.attr,
5007        &order_attr.attr,
5008        &min_partial_attr.attr,
5009        &cpu_partial_attr.attr,
5010        &objects_attr.attr,
5011        &objects_partial_attr.attr,
5012        &partial_attr.attr,
5013        &cpu_slabs_attr.attr,
5014        &ctor_attr.attr,
5015        &aliases_attr.attr,
5016        &align_attr.attr,
5017        &hwcache_align_attr.attr,
5018        &reclaim_account_attr.attr,
5019        &destroy_by_rcu_attr.attr,
5020        &shrink_attr.attr,
5021        &reserved_attr.attr,
5022        &slabs_cpu_partial_attr.attr,
5023#ifdef CONFIG_SLUB_DEBUG
5024        &total_objects_attr.attr,
5025        &slabs_attr.attr,
5026        &sanity_checks_attr.attr,
5027        &trace_attr.attr,
5028        &red_zone_attr.attr,
5029        &poison_attr.attr,
5030        &store_user_attr.attr,
5031        &validate_attr.attr,
5032        &alloc_calls_attr.attr,
5033        &free_calls_attr.attr,
5034#endif
5035#ifdef CONFIG_ZONE_DMA
5036        &cache_dma_attr.attr,
5037#endif
5038#ifdef CONFIG_NUMA
5039        &remote_node_defrag_ratio_attr.attr,
5040#endif
5041#ifdef CONFIG_SLUB_STATS
5042        &alloc_fastpath_attr.attr,
5043        &alloc_slowpath_attr.attr,
5044        &free_fastpath_attr.attr,
5045        &free_slowpath_attr.attr,
5046        &free_frozen_attr.attr,
5047        &free_add_partial_attr.attr,
5048        &free_remove_partial_attr.attr,
5049        &alloc_from_partial_attr.attr,
5050        &alloc_slab_attr.attr,
5051        &alloc_refill_attr.attr,
5052        &alloc_node_mismatch_attr.attr,
5053        &free_slab_attr.attr,
5054        &cpuslab_flush_attr.attr,
5055        &deactivate_full_attr.attr,
5056        &deactivate_empty_attr.attr,
5057        &deactivate_to_head_attr.attr,
5058        &deactivate_to_tail_attr.attr,
5059        &deactivate_remote_frees_attr.attr,
5060        &deactivate_bypass_attr.attr,
5061        &order_fallback_attr.attr,
5062        &cmpxchg_double_fail_attr.attr,
5063        &cmpxchg_double_cpu_fail_attr.attr,
5064        &cpu_partial_alloc_attr.attr,
5065        &cpu_partial_free_attr.attr,
5066        &cpu_partial_node_attr.attr,
5067        &cpu_partial_drain_attr.attr,
5068#endif
5069#ifdef CONFIG_FAILSLAB
5070        &failslab_attr.attr,
5071#endif
5072
5073        NULL
5074};
5075
5076static struct attribute_group slab_attr_group = {
5077        .attrs = slab_attrs,
5078};
5079
5080static ssize_t slab_attr_show(struct kobject *kobj,
5081                                struct attribute *attr,
5082                                char *buf)
5083{
5084        struct slab_attribute *attribute;
5085        struct kmem_cache *s;
5086        int err;
5087
5088        attribute = to_slab_attr(attr);
5089        s = to_slab(kobj);
5090
5091        if (!attribute->show)
5092                return -EIO;
5093
5094        err = attribute->show(s, buf);
5095
5096        return err;
5097}
5098
5099static ssize_t slab_attr_store(struct kobject *kobj,
5100                                struct attribute *attr,
5101                                const char *buf, size_t len)
5102{
5103        struct slab_attribute *attribute;
5104        struct kmem_cache *s;
5105        int err;
5106
5107        attribute = to_slab_attr(attr);
5108        s = to_slab(kobj);
5109
5110        if (!attribute->store)
5111                return -EIO;
5112
5113        err = attribute->store(s, buf, len);
5114#ifdef CONFIG_MEMCG_KMEM
5115        if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5116                int i;
5117
5118                mutex_lock(&slab_mutex);
5119                if (s->max_attr_size < len)
5120                        s->max_attr_size = len;
5121
5122                /*
5123                 * This is a best effort propagation, so this function's return
5124                 * value will be determined by the parent cache only. This is
5125                 * basically because not all attributes will have a well
5126                 * defined semantics for rollbacks - most of the actions will
5127                 * have permanent effects.
5128                 *
5129                 * Returning the error value of any of the children that fail
5130                 * is not 100 % defined, in the sense that users seeing the
5131                 * error code won't be able to know anything about the state of
5132                 * the cache.
5133                 *
5134                 * Only returning the error code for the parent cache at least
5135                 * has well defined semantics. The cache being written to
5136                 * directly either failed or succeeded, in which case we loop
5137                 * through the descendants with best-effort propagation.
5138                 */
5139                for_each_memcg_cache_index(i) {
5140                        struct kmem_cache *c = cache_from_memcg(s, i);
5141                        if (c)
5142                                attribute->store(c, buf, len);
5143                }
5144                mutex_unlock(&slab_mutex);
5145        }
5146#endif
5147        return err;
5148}
5149
5150static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5151{
5152#ifdef CONFIG_MEMCG_KMEM
5153        int i;
5154        char *buffer = NULL;
5155
5156        if (!is_root_cache(s))
5157                return;
5158
5159        /*
5160         * This mean this cache had no attribute written. Therefore, no point
5161         * in copying default values around
5162         */
5163        if (!s->max_attr_size)
5164                return;
5165
5166        for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5167                char mbuf[64];
5168                char *buf;
5169                struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5170
5171                if (!attr || !attr->store || !attr->show)
5172                        continue;
5173
5174                /*
5175                 * It is really bad that we have to allocate here, so we will
5176                 * do it only as a fallback. If we actually allocate, though,
5177                 * we can just use the allocated buffer until the end.
5178                 *
5179                 * Most of the slub attributes will tend to be very small in
5180                 * size, but sysfs allows buffers up to a page, so they can
5181                 * theoretically happen.
5182                 */
5183                if (buffer)
5184                        buf = buffer;
5185                else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5186                        buf = mbuf;
5187                else {
5188                        buffer = (char *) get_zeroed_page(GFP_KERNEL);
5189                        if (WARN_ON(!buffer))
5190                                continue;
5191                        buf = buffer;
5192                }
5193
5194                attr->show(s->memcg_params->root_cache, buf);
5195                attr->store(s, buf, strlen(buf));
5196        }
5197
5198        if (buffer)
5199                free_page((unsigned long)buffer);
5200#endif
5201}
5202
5203static const struct sysfs_ops slab_sysfs_ops = {
5204        .show = slab_attr_show,
5205        .store = slab_attr_store,
5206};
5207
5208static struct kobj_type slab_ktype = {
5209        .sysfs_ops = &slab_sysfs_ops,
5210};
5211
5212static int uevent_filter(struct kset *kset, struct kobject *kobj)
5213{
5214        struct kobj_type *ktype = get_ktype(kobj);
5215
5216        if (ktype == &slab_ktype)
5217                return 1;
5218        return 0;
5219}
5220
5221static const struct kset_uevent_ops slab_uevent_ops = {
5222        .filter = uevent_filter,
5223};
5224
5225static struct kset *slab_kset;
5226
5227#define ID_STR_LENGTH 64
5228
5229/* Create a unique string id for a slab cache:
5230 *
5231 * Format       :[flags-]size
5232 */
5233static char *create_unique_id(struct kmem_cache *s)
5234{
5235        char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5236        char *p = name;
5237
5238        BUG_ON(!name);
5239
5240        *p++ = ':';
5241        /*
5242         * First flags affecting slabcache operations. We will only
5243         * get here for aliasable slabs so we do not need to support
5244         * too many flags. The flags here must cover all flags that
5245         * are matched during merging to guarantee that the id is
5246         * unique.
5247         */
5248        if (s->flags & SLAB_CACHE_DMA)
5249                *p++ = 'd';
5250        if (s->flags & SLAB_RECLAIM_ACCOUNT)
5251                *p++ = 'a';
5252        if (s->flags & SLAB_DEBUG_FREE)
5253                *p++ = 'F';
5254        if (!(s->flags & SLAB_NOTRACK))
5255                *p++ = 't';
5256        if (p != name + 1)
5257                *p++ = '-';
5258        p += sprintf(p, "%07d", s->size);
5259
5260#ifdef CONFIG_MEMCG_KMEM
5261        if (!is_root_cache(s))
5262                p += sprintf(p, "-%08d", memcg_cache_id(s->memcg_params->memcg));
5263#endif
5264
5265        BUG_ON(p > name + ID_STR_LENGTH - 1);
5266        return name;
5267}
5268
5269static int sysfs_slab_add(struct kmem_cache *s)
5270{
5271        int err;
5272        const char *name;
5273        int unmergeable = slab_unmergeable(s);
5274
5275        if (unmergeable) {
5276                /*
5277                 * Slabcache can never be merged so we can use the name proper.
5278                 * This is typically the case for debug situations. In that
5279                 * case we can catch duplicate names easily.
5280                 */
5281                sysfs_remove_link(&slab_kset->kobj, s->name);
5282                name = s->name;
5283        } else {
5284                /*
5285                 * Create a unique name for the slab as a target
5286                 * for the symlinks.
5287                 */
5288                name = create_unique_id(s);
5289        }
5290
5291        s->kobj.kset = slab_kset;
5292        err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5293        if (err) {
5294                kobject_put(&s->kobj);
5295                return err;
5296        }
5297
5298        err = sysfs_create_group(&s->kobj, &slab_attr_group);
5299        if (err) {
5300                kobject_del(&s->kobj);
5301                kobject_put(&s->kobj);
5302                return err;
5303        }
5304        kobject_uevent(&s->kobj, KOBJ_ADD);
5305        if (!unmergeable) {
5306                /* Setup first alias */
5307                sysfs_slab_alias(s, s->name);
5308                kfree(name);
5309        }
5310        return 0;
5311}
5312
5313static void sysfs_slab_remove(struct kmem_cache *s)
5314{
5315        if (slab_state < FULL)
5316                /*
5317                 * Sysfs has not been setup yet so no need to remove the
5318                 * cache from sysfs.
5319                 */
5320                return;
5321
5322        kobject_uevent(&s->kobj, KOBJ_REMOVE);
5323        kobject_del(&s->kobj);
5324        kobject_put(&s->kobj);
5325}
5326
5327/*
5328 * Need to buffer aliases during bootup until sysfs becomes
5329 * available lest we lose that information.
5330 */
5331struct saved_alias {
5332        struct kmem_cache *s;
5333        const char *name;
5334        struct saved_alias *next;
5335};
5336
5337static struct saved_alias *alias_list;
5338
5339static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5340{
5341        struct saved_alias *al;
5342
5343        if (slab_state == FULL) {
5344                /*
5345                 * If we have a leftover link then remove it.
5346                 */
5347                sysfs_remove_link(&slab_kset->kobj, name);
5348                return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5349        }
5350
5351        al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5352        if (!al)
5353                return -ENOMEM;
5354
5355        al->s = s;
5356        al->name = name;
5357        al->next = alias_list;
5358        alias_list = al;
5359        return 0;
5360}
5361
5362static int __init slab_sysfs_init(void)
5363{
5364        struct kmem_cache *s;
5365        int err;
5366
5367        mutex_lock(&slab_mutex);
5368
5369        slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5370        if (!slab_kset) {
5371                mutex_unlock(&slab_mutex);
5372                printk(KERN_ERR "Cannot register slab subsystem.\n");
5373                return -ENOSYS;
5374        }
5375
5376        slab_state = FULL;
5377
5378        list_for_each_entry(s, &slab_caches, list) {
5379                err = sysfs_slab_add(s);
5380                if (err)
5381                        printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5382                                                " to sysfs\n", s->name);
5383        }
5384
5385        while (alias_list) {
5386                struct saved_alias *al = alias_list;
5387
5388                alias_list = alias_list->next;
5389                err = sysfs_slab_alias(al->s, al->name);
5390                if (err)
5391                        printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5392                                        " %s to sysfs\n", al->name);
5393                kfree(al);
5394        }
5395
5396        mutex_unlock(&slab_mutex);
5397        resiliency_test();
5398        return 0;
5399}
5400
5401__initcall(slab_sysfs_init);
5402#endif /* CONFIG_SYSFS */
5403
5404/*
5405 * The /proc/slabinfo ABI
5406 */
5407#ifdef CONFIG_SLABINFO
5408void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5409{
5410        unsigned long nr_partials = 0;
5411        unsigned long nr_slabs = 0;
5412        unsigned long nr_objs = 0;
5413        unsigned long nr_free = 0;
5414        int node;
5415
5416        for_each_online_node(node) {
5417                struct kmem_cache_node *n = get_node(s, node);
5418
5419                if (!n)
5420                        continue;
5421
5422                nr_partials += n->nr_partial;
5423                nr_slabs += atomic_long_read(&n->nr_slabs);
5424                nr_objs += atomic_long_read(&n->total_objects);
5425                nr_free += count_partial(n, count_free);
5426        }
5427
5428        sinfo->active_objs = nr_objs - nr_free;
5429        sinfo->num_objs = nr_objs;
5430        sinfo->active_slabs = nr_slabs;
5431        sinfo->num_slabs = nr_slabs;
5432        sinfo->objects_per_slab = oo_objects(s->oo);
5433        sinfo->cache_order = oo_order(s->oo);
5434}
5435
5436void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5437{
5438}
5439
5440ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5441                       size_t count, loff_t *ppos)
5442{
5443        return -EIO;
5444}
5445#endif /* CONFIG_SLABINFO */
5446
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.