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