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/notifier.h>
  22#include <linux/seq_file.h>
  23#include <linux/kmemcheck.h>
  24#include <linux/cpu.h>
  25#include <linux/cpuset.h>
  26#include <linux/mempolicy.h>
  27#include <linux/ctype.h>
  28#include <linux/debugobjects.h>
  29#include <linux/kallsyms.h>
  30#include <linux/memory.h>
  31#include <linux/math64.h>
  32#include <linux/fault-inject.h>
  33#include <linux/stacktrace.h>
  34#include <linux/prefetch.h>
  35#include <linux/memcontrol.h>
  36
  37#include <trace/events/kmem.h>
  38
  39#include "internal.h"
  40
  41/*
  42 * Lock order:
  43 *   1. slab_mutex (Global Mutex)
  44 *   2. node->list_lock
  45 *   3. slab_lock(page) (Only on some arches and for debugging)
  46 *
  47 *   slab_mutex
  48 *
  49 *   The role of the slab_mutex is to protect the list of all the slabs
  50 *   and to synchronize major metadata changes to slab cache structures.
  51 *
  52 *   The slab_lock is only used for debugging and on arches that do not
  53 *   have the ability to do a cmpxchg_double. It only protects the second
  54 *   double word in the page struct. Meaning
  55 *      A. page->freelist       -> List of object free in a page
  56 *      B. page->counters       -> Counters of objects
  57 *      C. page->frozen         -> frozen state
  58 *
  59 *   If a slab is frozen then it is exempt from list management. It is not
  60 *   on any list. The processor that froze the slab is the one who can
  61 *   perform list operations on the page. Other processors may put objects
  62 *   onto the freelist but the processor that froze the slab is the only
  63 *   one that can retrieve the objects from the page's freelist.
  64 *
  65 *   The list_lock protects the partial and full list on each node and
  66 *   the partial slab counter. If taken then no new slabs may be added or
  67 *   removed from the lists nor make the number of partial slabs be modified.
  68 *   (Note that the total number of slabs is an atomic value that may be
  69 *   modified without taking the list lock).
  70 *
  71 *   The list_lock is a centralized lock and thus we avoid taking it as
  72 *   much as possible. As long as SLUB does not have to handle partial
  73 *   slabs, operations can continue without any centralized lock. F.e.
  74 *   allocating a long series of objects that fill up slabs does not require
  75 *   the list lock.
  76 *   Interrupts are disabled during allocation and deallocation in order to
  77 *   make the slab allocator safe to use in the context of an irq. In addition
  78 *   interrupts are disabled to ensure that the processor does not change
  79 *   while handling per_cpu slabs, due to kernel preemption.
  80 *
  81 * SLUB assigns one slab for allocation to each processor.
  82 * Allocations only occur from these slabs called cpu slabs.
  83 *
  84 * Slabs with free elements are kept on a partial list and during regular
  85 * operations no list for full slabs is used. If an object in a full slab is
  86 * freed then the slab will show up again on the partial lists.
  87 * We track full slabs for debugging purposes though because otherwise we
  88 * cannot scan all objects.
  89 *
  90 * Slabs are freed when they become empty. Teardown and setup is
  91 * minimal so we rely on the page allocators per cpu caches for
  92 * fast frees and allocs.
  93 *
  94 * Overloading of page flags that are otherwise used for LRU management.
  95 *
  96 * PageActive           The slab is frozen and exempt from list processing.
  97 *                      This means that the slab is dedicated to a purpose
  98 *                      such as satisfying allocations for a specific
  99 *                      processor. Objects may be freed in the slab while
 100 *                      it is frozen but slab_free will then skip the usual
 101 *                      list operations. It is up to the processor holding
 102 *                      the slab to integrate the slab into the slab lists
 103 *                      when the slab is no longer needed.
 104 *
 105 *                      One use of this flag is to mark slabs that are
 106 *                      used for allocations. Then such a slab becomes a cpu
 107 *                      slab. The cpu slab may be equipped with an additional
 108 *                      freelist that allows lockless access to
 109 *                      free objects in addition to the regular freelist
 110 *                      that requires the slab lock.
 111 *
 112 * PageError            Slab requires special handling due to debug
 113 *                      options set. This moves slab handling out of
 114 *                      the fast path and disables lockless freelists.
 115 */
 116
 117static inline int kmem_cache_debug(struct kmem_cache *s)
 118{
 119#ifdef CONFIG_SLUB_DEBUG
 120        return unlikely(s->flags & SLAB_DEBUG_FLAGS);
 121#else
 122        return 0;
 123#endif
 124}
 125
 126/*
 127 * Issues still to be resolved:
 128 *
 129 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
 130 *
 131 * - Variable sizing of the per node arrays
 132 */
 133
 134/* Enable to test recovery from slab corruption on boot */
 135#undef SLUB_RESILIENCY_TEST
 136
 137/* Enable to log cmpxchg failures */
 138#undef SLUB_DEBUG_CMPXCHG
 139
 140/*
 141 * Mininum number of partial slabs. These will be left on the partial
 142 * lists even if they are empty. kmem_cache_shrink may reclaim them.
 143 */
 144#define MIN_PARTIAL 5
 145
 146/*
 147 * Maximum number of desirable partial slabs.
 148 * The existence of more partial slabs makes kmem_cache_shrink
 149 * sort the partial list by the number of objects in the.
 150 */
 151#define MAX_PARTIAL 10
 152
 153#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
 154                                SLAB_POISON | SLAB_STORE_USER)
 155
 156/*
 157 * Debugging flags that require metadata to be stored in the slab.  These get
 158 * disabled when slub_debug=O is used and a cache's min order increases with
 159 * metadata.
 160 */
 161#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
 162
 163/*
 164 * Set of flags that will prevent slab merging
 165 */
 166#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
 167                SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
 168                SLAB_FAILSLAB)
 169
 170#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
 171                SLAB_CACHE_DMA | SLAB_NOTRACK)
 172
 173#define OO_SHIFT        16
 174#define OO_MASK         ((1 << OO_SHIFT) - 1)
 175#define MAX_OBJS_PER_PAGE       32767 /* since page.objects is u15 */
 176
 177/* Internal SLUB flags */
 178#define __OBJECT_POISON         0x80000000UL /* Poison object */
 179#define __CMPXCHG_DOUBLE        0x40000000UL /* Use cmpxchg_double */
 180
 181#ifdef CONFIG_SMP
 182static struct notifier_block slab_notifier;
 183#endif
 184
 185/*
 186 * Tracking user of a slab.
 187 */
 188#define TRACK_ADDRS_COUNT 16
 189struct track {
 190        unsigned long addr;     /* Called from address */
 191#ifdef CONFIG_STACKTRACE
 192        unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
 193#endif
 194        int cpu;                /* Was running on cpu */
 195        int pid;                /* Pid context */
 196        unsigned long when;     /* When did the operation occur */
 197};
 198
 199enum track_item { TRACK_ALLOC, TRACK_FREE };
 200
 201#ifdef CONFIG_SYSFS
 202static int sysfs_slab_add(struct kmem_cache *);
 203static int sysfs_slab_alias(struct kmem_cache *, const char *);
 204static void sysfs_slab_remove(struct kmem_cache *);
 205static void memcg_propagate_slab_attrs(struct kmem_cache *s);
 206#else
 207static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
 208static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
 209                                                        { return 0; }
 210static inline void sysfs_slab_remove(struct kmem_cache *s) { }
 211
 212static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
 213#endif
 214
 215static inline void stat(const struct kmem_cache *s, enum stat_item si)
 216{
 217#ifdef CONFIG_SLUB_STATS
 218        __this_cpu_inc(s->cpu_slab->stat[si]);
 219#endif
 220}
 221
 222/********************************************************************
 223 *                      Core slab cache functions
 224 *******************************************************************/
 225
 226static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
 227{
 228        return s->node[node];
 229}
 230
 231/* Verify that a pointer has an address that is valid within a slab page */
 232static inline int check_valid_pointer(struct kmem_cache *s,
 233                                struct page *page, const void *object)
 234{
 235        void *base;
 236
 237        if (!object)
 238                return 1;
 239
 240        base = page_address(page);
 241        if (object < base || object >= base + page->objects * s->size ||
 242                (object - base) % s->size) {
 243                return 0;
 244        }
 245
 246        return 1;
 247}
 248
 249static inline void *get_freepointer(struct kmem_cache *s, void *object)
 250{
 251        return *(void **)(object + s->offset);
 252}
 253
 254static void prefetch_freepointer(const struct kmem_cache *s, void *object)
 255{
 256        prefetch(object + s->offset);
 257}
 258
 259static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
 260{
 261        void *p;
 262
 263#ifdef CONFIG_DEBUG_PAGEALLOC
 264        probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
 265#else
 266        p = get_freepointer(s, object);
 267#endif
 268        return p;
 269}
 270
 271static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
 272{
 273        *(void **)(object + s->offset) = fp;
 274}
 275
 276/* Loop over all objects in a slab */
 277#define for_each_object(__p, __s, __addr, __objects) \
 278        for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
 279                        __p += (__s)->size)
 280
 281/* Determine object index from a given position */
 282static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
 283{
 284        return (p - addr) / s->size;
 285}
 286
 287static inline size_t slab_ksize(const struct kmem_cache *s)
 288{
 289#ifdef CONFIG_SLUB_DEBUG
 290        /*
 291         * Debugging requires use of the padding between object
 292         * and whatever may come after it.
 293         */
 294        if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
 295                return s->object_size;
 296
 297#endif
 298        /*
 299         * If we have the need to store the freelist pointer
 300         * back there or track user information then we can
 301         * only use the space before that information.
 302         */
 303        if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
 304                return s->inuse;
 305        /*
 306         * Else we can use all the padding etc for the allocation
 307         */
 308        return s->size;
 309}
 310
 311static inline int order_objects(int order, unsigned long size, int reserved)
 312{
 313        return ((PAGE_SIZE << order) - reserved) / size;
 314}
 315
 316static inline struct kmem_cache_order_objects oo_make(int order,
 317                unsigned long size, int reserved)
 318{
 319        struct kmem_cache_order_objects x = {
 320                (order << OO_SHIFT) + order_objects(order, size, reserved)
 321        };
 322
 323        return x;
 324}
 325
 326static inline int oo_order(struct kmem_cache_order_objects x)
 327{
 328        return x.x >> OO_SHIFT;
 329}
 330
 331static inline int oo_objects(struct kmem_cache_order_objects x)
 332{
 333        return x.x & OO_MASK;
 334}
 335
 336/*
 337 * Per slab locking using the pagelock
 338 */
 339static __always_inline void slab_lock(struct page *page)
 340{
 341        bit_spin_lock(PG_locked, &page->flags);
 342}
 343
 344static __always_inline void slab_unlock(struct page *page)
 345{
 346        __bit_spin_unlock(PG_locked, &page->flags);
 347}
 348
 349/* Interrupts must be disabled (for the fallback code to work right) */
 350static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
 351                void *freelist_old, unsigned long counters_old,
 352                void *freelist_new, unsigned long counters_new,
 353                const char *n)
 354{
 355        VM_BUG_ON(!irqs_disabled());
 356#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
 357    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
 358        if (s->flags & __CMPXCHG_DOUBLE) {
 359                if (cmpxchg_double(&page->freelist, &page->counters,
 360                        freelist_old, counters_old,
 361                        freelist_new, counters_new))
 362                return 1;
 363        } else
 364#endif
 365        {
 366                slab_lock(page);
 367                if (page->freelist == freelist_old && page->counters == counters_old) {
 368                        page->freelist = freelist_new;
 369                        page->counters = counters_new;
 370                        slab_unlock(page);
 371                        return 1;
 372                }
 373                slab_unlock(page);
 374        }
 375
 376        cpu_relax();
 377        stat(s, CMPXCHG_DOUBLE_FAIL);
 378
 379#ifdef SLUB_DEBUG_CMPXCHG
 380        printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
 381#endif
 382
 383        return 0;
 384}
 385
 386static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
 387                void *freelist_old, unsigned long counters_old,
 388                void *freelist_new, unsigned long counters_new,
 389                const char *n)
 390{
 391#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
 392    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
 393        if (s->flags & __CMPXCHG_DOUBLE) {
 394                if (cmpxchg_double(&page->freelist, &page->counters,
 395                        freelist_old, counters_old,
 396                        freelist_new, counters_new))
 397                return 1;
 398        } else
 399#endif
 400        {
 401                unsigned long flags;
 402
 403                local_irq_save(flags);
 404                slab_lock(page);
 405                if (page->freelist == freelist_old && page->counters == counters_old) {
 406                        page->freelist = freelist_new;
 407                        page->counters = counters_new;
 408                        slab_unlock(page);
 409                        local_irq_restore(flags);
 410                        return 1;
 411                }
 412                slab_unlock(page);
 413                local_irq_restore(flags);
 414        }
 415
 416        cpu_relax();
 417        stat(s, CMPXCHG_DOUBLE_FAIL);
 418
 419#ifdef SLUB_DEBUG_CMPXCHG
 420        printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
 421#endif
 422
 423        return 0;
 424}
 425
 426#ifdef CONFIG_SLUB_DEBUG
 427/*
 428 * Determine a map of object in use on a page.
 429 *
 430 * Node listlock must be held to guarantee that the page does
 431 * not vanish from under us.
 432 */
 433static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
 434{
 435        void *p;
 436        void *addr = page_address(page);
 437
 438        for (p = page->freelist; p; p = get_freepointer(s, p))
 439                set_bit(slab_index(p, s, addr), map);
 440}
 441
 442/*
 443 * Debug settings:
 444 */
 445#ifdef CONFIG_SLUB_DEBUG_ON
 446static int slub_debug = DEBUG_DEFAULT_FLAGS;
 447#else
 448static int slub_debug;
 449#endif
 450
 451static char *slub_debug_slabs;
 452static int disable_higher_order_debug;
 453
 454/*
 455 * Object debugging
 456 */
 457static void print_section(char *text, u8 *addr, unsigned int length)
 458{
 459        print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
 460                        length, 1);
 461}
 462
 463static struct track *get_track(struct kmem_cache *s, void *object,
 464        enum track_item alloc)
 465{
 466        struct track *p;
 467
 468        if (s->offset)
 469                p = object + s->offset + sizeof(void *);
 470        else
 471                p = object + s->inuse;
 472
 473        return p + alloc;
 474}
 475
 476static void set_track(struct kmem_cache *s, void *object,
 477                        enum track_item alloc, unsigned long addr)
 478{
 479        struct track *p = get_track(s, object, alloc);
 480
 481        if (addr) {
 482#ifdef CONFIG_STACKTRACE
 483                struct stack_trace trace;
 484                int i;
 485
 486                trace.nr_entries = 0;
 487                trace.max_entries = TRACK_ADDRS_COUNT;
 488                trace.entries = p->addrs;
 489                trace.skip = 3;
 490                save_stack_trace(&trace);
 491
 492                /* See rant in lockdep.c */
 493                if (trace.nr_entries != 0 &&
 494                    trace.entries[trace.nr_entries - 1] == ULONG_MAX)
 495                        trace.nr_entries--;
 496
 497                for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
 498                        p->addrs[i] = 0;
 499#endif
 500                p->addr = addr;
 501                p->cpu = smp_processor_id();
 502                p->pid = current->pid;
 503                p->when = jiffies;
 504        } else
 505                memset(p, 0, sizeof(struct track));
 506}
 507
 508static void init_tracking(struct kmem_cache *s, void *object)
 509{
 510        if (!(s->flags & SLAB_STORE_USER))
 511                return;
 512
 513        set_track(s, object, TRACK_FREE, 0UL);
 514        set_track(s, object, TRACK_ALLOC, 0UL);
 515}
 516
 517static void print_track(const char *s, struct track *t)
 518{
 519        if (!t->addr)
 520                return;
 521
 522        printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
 523                s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
 524#ifdef CONFIG_STACKTRACE
 525        {
 526                int i;
 527                for (i = 0; i < TRACK_ADDRS_COUNT; i++)
 528                        if (t->addrs[i])
 529                                printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
 530                        else
 531                                break;
 532        }
 533#endif
 534}
 535
 536static void print_tracking(struct kmem_cache *s, void *object)
 537{
 538        if (!(s->flags & SLAB_STORE_USER))
 539                return;
 540
 541        print_track("Allocated", get_track(s, object, TRACK_ALLOC));
 542        print_track("Freed", get_track(s, object, TRACK_FREE));
 543}
 544
 545static void print_page_info(struct page *page)
 546{
 547        printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
 548                page, page->objects, page->inuse, page->freelist, page->flags);
 549
 550}
 551
 552static void slab_bug(struct kmem_cache *s, char *fmt, ...)
 553{
 554        va_list args;
 555        char buf[100];
 556
 557        va_start(args, fmt);
 558        vsnprintf(buf, sizeof(buf), fmt, args);
 559        va_end(args);
 560        printk(KERN_ERR "========================================"
 561                        "=====================================\n");
 562        printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
 563        printk(KERN_ERR "----------------------------------------"
 564                        "-------------------------------------\n\n");
 565
 566        add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 567}
 568
 569static void slab_fix(struct kmem_cache *s, char *fmt, ...)
 570{
 571        va_list args;
 572        char buf[100];
 573
 574        va_start(args, fmt);
 575        vsnprintf(buf, sizeof(buf), fmt, args);
 576        va_end(args);
 577        printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
 578}
 579
 580static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
 581{
 582        unsigned int off;       /* Offset of last byte */
 583        u8 *addr = page_address(page);
 584
 585        print_tracking(s, p);
 586
 587        print_page_info(page);
 588
 589        printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
 590                        p, p - addr, get_freepointer(s, p));
 591
 592        if (p > addr + 16)
 593                print_section("Bytes b4 ", p - 16, 16);
 594
 595        print_section("Object ", p, min_t(unsigned long, s->object_size,
 596                                PAGE_SIZE));
 597        if (s->flags & SLAB_RED_ZONE)
 598                print_section("Redzone ", p + s->object_size,
 599                        s->inuse - s->object_size);
 600
 601        if (s->offset)
 602                off = s->offset + sizeof(void *);
 603        else
 604                off = s->inuse;
 605
 606        if (s->flags & SLAB_STORE_USER)
 607                off += 2 * sizeof(struct track);
 608
 609        if (off != s->size)
 610                /* Beginning of the filler is the free pointer */
 611                print_section("Padding ", p + off, s->size - off);
 612
 613        dump_stack();
 614}
 615
 616static void object_err(struct kmem_cache *s, struct page *page,
 617                        u8 *object, char *reason)
 618{
 619        slab_bug(s, "%s", reason);
 620        print_trailer(s, page, object);
 621}
 622
 623static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
 624{
 625        va_list args;
 626        char buf[100];
 627
 628        va_start(args, fmt);
 629        vsnprintf(buf, sizeof(buf), fmt, args);
 630        va_end(args);
 631        slab_bug(s, "%s", buf);
 632        print_page_info(page);
 633        dump_stack();
 634}
 635
 636static void init_object(struct kmem_cache *s, void *object, u8 val)
 637{
 638        u8 *p = object;
 639
 640        if (s->flags & __OBJECT_POISON) {
 641                memset(p, POISON_FREE, s->object_size - 1);
 642                p[s->object_size - 1] = POISON_END;
 643        }
 644
 645        if (s->flags & SLAB_RED_ZONE)
 646                memset(p + s->object_size, val, s->inuse - s->object_size);
 647}
 648
 649static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
 650                                                void *from, void *to)
 651{
 652        slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
 653        memset(from, data, to - from);
 654}
 655
 656static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
 657                        u8 *object, char *what,
 658                        u8 *start, unsigned int value, unsigned int bytes)
 659{
 660        u8 *fault;
 661        u8 *end;
 662
 663        fault = memchr_inv(start, value, bytes);
 664        if (!fault)
 665                return 1;
 666
 667        end = start + bytes;
 668        while (end > fault && end[-1] == value)
 669                end--;
 670
 671        slab_bug(s, "%s overwritten", what);
 672        printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
 673                                        fault, end - 1, fault[0], value);
 674        print_trailer(s, page, object);
 675
 676        restore_bytes(s, what, value, fault, end);
 677        return 0;
 678}
 679
 680/*
 681 * Object layout:
 682 *
 683 * object address
 684 *      Bytes of the object to be managed.
 685 *      If the freepointer may overlay the object then the free
 686 *      pointer is the first word of the object.
 687 *
 688 *      Poisoning uses 0x6b (POISON_FREE) and the last byte is
 689 *      0xa5 (POISON_END)
 690 *
 691 * object + s->object_size
 692 *      Padding to reach word boundary. This is also used for Redzoning.
 693 *      Padding is extended by another word if Redzoning is enabled and
 694 *      object_size == inuse.
 695 *
 696 *      We fill with 0xbb (RED_INACTIVE) for inactive objects and with
 697 *      0xcc (RED_ACTIVE) for objects in use.
 698 *
 699 * object + s->inuse
 700 *      Meta data starts here.
 701 *
 702 *      A. Free pointer (if we cannot overwrite object on free)
 703 *      B. Tracking data for SLAB_STORE_USER
 704 *      C. Padding to reach required alignment boundary or at mininum
 705 *              one word if debugging is on to be able to detect writes
 706 *              before the word boundary.
 707 *
 708 *      Padding is done using 0x5a (POISON_INUSE)
 709 *
 710 * object + s->size
 711 *      Nothing is used beyond s->size.
 712 *
 713 * If slabcaches are merged then the object_size and inuse boundaries are mostly
 714 * ignored. And therefore no slab options that rely on these boundaries
 715 * may be used with merged slabcaches.
 716 */
 717
 718static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
 719{
 720        unsigned long off = s->inuse;   /* The end of info */
 721
 722        if (s->offset)
 723                /* Freepointer is placed after the object. */
 724                off += sizeof(void *);
 725
 726        if (s->flags & SLAB_STORE_USER)
 727                /* We also have user information there */
 728                off += 2 * sizeof(struct track);
 729
 730        if (s->size == off)
 731                return 1;
 732
 733        return check_bytes_and_report(s, page, p, "Object padding",
 734                                p + off, POISON_INUSE, s->size - off);
 735}
 736
 737/* Check the pad bytes at the end of a slab page */
 738static int slab_pad_check(struct kmem_cache *s, struct page *page)
 739{
 740        u8 *start;
 741        u8 *fault;
 742        u8 *end;
 743        int length;
 744        int remainder;
 745
 746        if (!(s->flags & SLAB_POISON))
 747                return 1;
 748
 749        start = page_address(page);
 750        length = (PAGE_SIZE << compound_order(page)) - s->reserved;
 751        end = start + length;
 752        remainder = length % s->size;
 753        if (!remainder)
 754                return 1;
 755
 756        fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
 757        if (!fault)
 758                return 1;
 759        while (end > fault && end[-1] == POISON_INUSE)
 760                end--;
 761
 762        slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
 763        print_section("Padding ", end - remainder, remainder);
 764
 765        restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
 766        return 0;
 767}
 768
 769static int check_object(struct kmem_cache *s, struct page *page,
 770                                        void *object, u8 val)
 771{
 772        u8 *p = object;
 773        u8 *endobject = object + s->object_size;
 774
 775        if (s->flags & SLAB_RED_ZONE) {
 776                if (!check_bytes_and_report(s, page, object, "Redzone",
 777                        endobject, val, s->inuse - s->object_size))
 778                        return 0;
 779        } else {
 780                if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
 781                        check_bytes_and_report(s, page, p, "Alignment padding",
 782                                endobject, POISON_INUSE, s->inuse - s->object_size);
 783                }
 784        }
 785
 786        if (s->flags & SLAB_POISON) {
 787                if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
 788                        (!check_bytes_and_report(s, page, p, "Poison", p,
 789                                        POISON_FREE, s->object_size - 1) ||
 790                         !check_bytes_and_report(s, page, p, "Poison",
 791                                p + s->object_size - 1, POISON_END, 1)))
 792                        return 0;
 793                /*
 794                 * check_pad_bytes cleans up on its own.
 795                 */
 796                check_pad_bytes(s, page, p);
 797        }
 798
 799        if (!s->offset && val == SLUB_RED_ACTIVE)
 800                /*
 801                 * Object and freepointer overlap. Cannot check
 802                 * freepointer while object is allocated.
 803                 */
 804                return 1;
 805
 806        /* Check free pointer validity */
 807        if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
 808                object_err(s, page, p, "Freepointer corrupt");
 809                /*
 810                 * No choice but to zap it and thus lose the remainder
 811                 * of the free objects in this slab. May cause
 812                 * another error because the object count is now wrong.
 813                 */
 814                set_freepointer(s, p, NULL);
 815                return 0;
 816        }
 817        return 1;
 818}
 819
 820static int check_slab(struct kmem_cache *s, struct page *page)
 821{
 822        int maxobj;
 823
 824        VM_BUG_ON(!irqs_disabled());
 825
 826        if (!PageSlab(page)) {
 827                slab_err(s, page, "Not a valid slab page");
 828                return 0;
 829        }
 830
 831        maxobj = order_objects(compound_order(page), s->size, s->reserved);
 832        if (page->objects > maxobj) {
 833                slab_err(s, page, "objects %u > max %u",
 834                        s->name, page->objects, maxobj);
 835                return 0;
 836        }
 837        if (page->inuse > page->objects) {
 838                slab_err(s, page, "inuse %u > max %u",
 839                        s->name, page->inuse, page->objects);
 840                return 0;
 841        }
 842        /* Slab_pad_check fixes things up after itself */
 843        slab_pad_check(s, page);
 844        return 1;
 845}
 846
 847/*
 848 * Determine if a certain object on a page is on the freelist. Must hold the
 849 * slab lock to guarantee that the chains are in a consistent state.
 850 */
 851static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
 852{
 853        int nr = 0;
 854        void *fp;
 855        void *object = NULL;
 856        unsigned long max_objects;
 857
 858        fp = page->freelist;
 859        while (fp && nr <= page->objects) {
 860                if (fp == search)
 861                        return 1;
 862                if (!check_valid_pointer(s, page, fp)) {
 863                        if (object) {
 864                                object_err(s, page, object,
 865                                        "Freechain corrupt");
 866                                set_freepointer(s, object, NULL);
 867                                break;
 868                        } else {
 869                                slab_err(s, page, "Freepointer corrupt");
 870                                page->freelist = NULL;
 871                                page->inuse = page->objects;
 872                                slab_fix(s, "Freelist cleared");
 873                                return 0;
 874                        }
 875                        break;
 876                }
 877                object = fp;
 878                fp = get_freepointer(s, object);
 879                nr++;
 880        }
 881
 882        max_objects = order_objects(compound_order(page), s->size, s->reserved);
 883        if (max_objects > MAX_OBJS_PER_PAGE)
 884                max_objects = MAX_OBJS_PER_PAGE;
 885
 886        if (page->objects != max_objects) {
 887                slab_err(s, page, "Wrong number of objects. Found %d but "
 888                        "should be %d", page->objects, max_objects);
 889                page->objects = max_objects;
 890                slab_fix(s, "Number of objects adjusted.");
 891        }
 892        if (page->inuse != page->objects - nr) {
 893                slab_err(s, page, "Wrong object count. Counter is %d but "
 894                        "counted were %d", page->inuse, page->objects - nr);
 895                page->inuse = page->objects - nr;
 896                slab_fix(s, "Object count adjusted.");
 897        }
 898        return search == NULL;
 899}
 900
 901static void trace(struct kmem_cache *s, struct page *page, void *object,
 902                                                                int alloc)
 903{
 904        if (s->flags & SLAB_TRACE) {
 905                printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
 906                        s->name,
 907                        alloc ? "alloc" : "free",
 908                        object, page->inuse,
 909                        page->freelist);
 910
 911                if (!alloc)
 912                        print_section("Object ", (void *)object, s->object_size);
 913
 914                dump_stack();
 915        }
 916}
 917
 918/*
 919 * Hooks for other subsystems that check memory allocations. In a typical
 920 * production configuration these hooks all should produce no code at all.
 921 */
 922static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
 923{
 924        flags &= gfp_allowed_mask;
 925        lockdep_trace_alloc(flags);
 926        might_sleep_if(flags & __GFP_WAIT);
 927
 928        return should_failslab(s->object_size, flags, s->flags);
 929}
 930
 931static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
 932{
 933        flags &= gfp_allowed_mask;
 934        kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
 935        kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
 936}
 937
 938static inline void slab_free_hook(struct kmem_cache *s, void *x)
 939{
 940        kmemleak_free_recursive(x, s->flags);
 941
 942        /*
 943         * Trouble is that we may no longer disable interupts in the fast path
 944         * So in order to make the debug calls that expect irqs to be
 945         * disabled we need to disable interrupts temporarily.
 946         */
 947#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
 948        {
 949                unsigned long flags;
 950
 951                local_irq_save(flags);
 952                kmemcheck_slab_free(s, x, s->object_size);
 953                debug_check_no_locks_freed(x, s->object_size);
 954                local_irq_restore(flags);
 955        }
 956#endif
 957        if (!(s->flags & SLAB_DEBUG_OBJECTS))
 958                debug_check_no_obj_freed(x, s->object_size);
 959}
 960
 961/*
 962 * Tracking of fully allocated slabs for debugging purposes.
 963 *
 964 * list_lock must be held.
 965 */
 966static void add_full(struct kmem_cache *s,
 967        struct kmem_cache_node *n, struct page *page)
 968{
 969        if (!(s->flags & SLAB_STORE_USER))
 970                return;
 971
 972        list_add(&page->lru, &n->full);
 973}
 974
 975/*
 976 * list_lock must be held.
 977 */
 978static void remove_full(struct kmem_cache *s, struct page *page)
 979{
 980        if (!(s->flags & SLAB_STORE_USER))
 981                return;
 982
 983        list_del(&page->lru);
 984}
 985
 986/* Tracking of the number of slabs for debugging purposes */
 987static inline unsigned long slabs_node(struct kmem_cache *s, int node)
 988{
 989        struct kmem_cache_node *n = get_node(s, node);
 990
 991        return atomic_long_read(&n->nr_slabs);
 992}
 993
 994static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
 995{
 996        return atomic_long_read(&n->nr_slabs);
 997}
 998
 999static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1000{
1001        struct kmem_cache_node *n = get_node(s, node);
1002
1003        /*
1004         * May be called early in order to allocate a slab for the
1005         * kmem_cache_node structure. Solve the chicken-egg
1006         * dilemma by deferring the increment of the count during
1007         * bootstrap (see early_kmem_cache_node_alloc).
1008         */
1009        if (likely(n)) {
1010                atomic_long_inc(&n->nr_slabs);
1011                atomic_long_add(objects, &n->total_objects);
1012        }
1013}
1014static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1015{
1016        struct kmem_cache_node *n = get_node(s, node);
1017
1018        atomic_long_dec(&n->nr_slabs);
1019        atomic_long_sub(objects, &n->total_objects);
1020}
1021
1022/* Object debug checks for alloc/free paths */
1023static void setup_object_debug(struct kmem_cache *s, struct page *page,
1024                                                                void *object)
1025{
1026        if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1027                return;
1028
1029        init_object(s, object, SLUB_RED_INACTIVE);
1030        init_tracking(s, object);
1031}
1032
1033static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1034                                        void *object, unsigned long addr)
1035{
1036        if (!check_slab(s, page))
1037                goto bad;
1038
1039        if (!check_valid_pointer(s, page, object)) {
1040                object_err(s, page, object, "Freelist Pointer check fails");
1041                goto bad;
1042        }
1043
1044        if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1045                goto bad;
1046
1047        /* Success perform special debug activities for allocs */
1048        if (s->flags & SLAB_STORE_USER)
1049                set_track(s, object, TRACK_ALLOC, addr);
1050        trace(s, page, object, 1);
1051        init_object(s, object, SLUB_RED_ACTIVE);
1052        return 1;
1053
1054bad:
1055        if (PageSlab(page)) {
1056                /*
1057                 * If this is a slab page then lets do the best we can
1058                 * to avoid issues in the future. Marking all objects
1059                 * as used avoids touching the remaining objects.
1060                 */
1061                slab_fix(s, "Marking all objects used");
1062                page->inuse = page->objects;
1063                page->freelist = NULL;
1064        }
1065        return 0;
1066}
1067
1068static noinline struct kmem_cache_node *free_debug_processing(
1069        struct kmem_cache *s, struct page *page, void *object,
1070        unsigned long addr, unsigned long *flags)
1071{
1072        struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1073
1074        spin_lock_irqsave(&n->list_lock, *flags);
1075        slab_lock(page);
1076
1077        if (!check_slab(s, page))
1078                goto fail;
1079
1080        if (!check_valid_pointer(s, page, object)) {
1081                slab_err(s, page, "Invalid object pointer 0x%p", object);
1082                goto fail;
1083        }
1084
1085        if (on_freelist(s, page, object)) {
1086                object_err(s, page, object, "Object already free");
1087                goto fail;
1088        }
1089
1090        if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1091                goto out;
1092
1093        if (unlikely(s != page->slab_cache)) {
1094                if (!PageSlab(page)) {
1095                        slab_err(s, page, "Attempt to free object(0x%p) "
1096                                "outside of slab", object);
1097                } else if (!page->slab_cache) {
1098                        printk(KERN_ERR
1099                                "SLUB <none>: no slab for object 0x%p.\n",
1100                                                object);
1101                        dump_stack();
1102                } else
1103                        object_err(s, page, object,
1104                                        "page slab pointer corrupt.");
1105                goto fail;
1106        }
1107
1108        if (s->flags & SLAB_STORE_USER)
1109                set_track(s, object, TRACK_FREE, addr);
1110        trace(s, page, object, 0);
1111        init_object(s, object, SLUB_RED_INACTIVE);
1112out:
1113        slab_unlock(page);
1114        /*
1115         * Keep node_lock to preserve integrity
1116         * until the object is actually freed
1117         */
1118        return n;
1119
1120fail:
1121        slab_unlock(page);
1122        spin_unlock_irqrestore(&n->list_lock, *flags);
1123        slab_fix(s, "Object at 0x%p not freed", object);
1124        return NULL;
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 struct kmem_cache_node *free_debug_processing(
1218        struct kmem_cache *s, struct page *page, void *object,
1219        unsigned long addr, unsigned long *flags) { return NULL; }
1220
1221static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1222                        { return 1; }
1223static inline int check_object(struct kmem_cache *s, struct page *page,
1224                        void *object, u8 val) { return 1; }
1225static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1226                                        struct page *page) {}
1227static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1228static inline unsigned long kmem_cache_flags(unsigned long object_size,
1229        unsigned long flags, const char *name,
1230        void (*ctor)(void *))
1231{
1232        return flags;
1233}
1234#define slub_debug 0
1235
1236#define disable_higher_order_debug 0
1237
1238static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1239                                                        { return 0; }
1240static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1241                                                        { return 0; }
1242static inline void inc_slabs_node(struct kmem_cache *s, int node,
1243                                                        int objects) {}
1244static inline void dec_slabs_node(struct kmem_cache *s, int node,
1245                                                        int objects) {}
1246
1247static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1248                                                        { return 0; }
1249
1250static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1251                void *object) {}
1252
1253static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1254
1255#endif /* CONFIG_SLUB_DEBUG */
1256
1257/*
1258 * Slab allocation and freeing
1259 */
1260static inline struct page *alloc_slab_page(gfp_t flags, int node,
1261                                        struct kmem_cache_order_objects oo)
1262{
1263        int order = oo_order(oo);
1264
1265        flags |= __GFP_NOTRACK;
1266
1267        if (node == NUMA_NO_NODE)
1268                return alloc_pages(flags, order);
1269        else
1270                return alloc_pages_exact_node(node, flags, order);
1271}
1272
1273static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1274{
1275        struct page *page;
1276        struct kmem_cache_order_objects oo = s->oo;
1277        gfp_t alloc_gfp;
1278
1279        flags &= gfp_allowed_mask;
1280
1281        if (flags & __GFP_WAIT)
1282                local_irq_enable();
1283
1284        flags |= s->allocflags;
1285
1286        /*
1287         * Let the initial higher-order allocation fail under memory pressure
1288         * so we fall-back to the minimum order allocation.
1289         */
1290        alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1291
1292        page = alloc_slab_page(alloc_gfp, node, oo);
1293        if (unlikely(!page)) {
1294                oo = s->min;
1295                /*
1296                 * Allocation may have failed due to fragmentation.
1297                 * Try a lower order alloc if possible
1298                 */
1299                page = alloc_slab_page(flags, node, oo);
1300
1301                if (page)
1302                        stat(s, ORDER_FALLBACK);
1303        }
1304
1305        if (kmemcheck_enabled && page
1306                && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1307                int pages = 1 << oo_order(oo);
1308
1309                kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1310
1311                /*
1312                 * Objects from caches that have a constructor don't get
1313                 * cleared when they're allocated, so we need to do it here.
1314                 */
1315                if (s->ctor)
1316                        kmemcheck_mark_uninitialized_pages(page, pages);
1317                else
1318                        kmemcheck_mark_unallocated_pages(page, pages);
1319        }
1320
1321        if (flags & __GFP_WAIT)
1322                local_irq_disable();
1323        if (!page)
1324                return NULL;
1325
1326        page->objects = oo_objects(oo);
1327        mod_zone_page_state(page_zone(page),
1328                (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1329                NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1330                1 << oo_order(oo));
1331
1332        return page;
1333}
1334
1335static void setup_object(struct kmem_cache *s, struct page *page,
1336                                void *object)
1337{
1338        setup_object_debug(s, page, object);
1339        if (unlikely(s->ctor))
1340                s->ctor(object);
1341}
1342
1343static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1344{
1345        struct page *page;
1346        void *start;
1347        void *last;
1348        void *p;
1349        int order;
1350
1351        BUG_ON(flags & GFP_SLAB_BUG_MASK);
1352
1353        page = allocate_slab(s,
1354                flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1355        if (!page)
1356                goto out;
1357
1358        order = compound_order(page);
1359        inc_slabs_node(s, page_to_nid(page), page->objects);
1360        memcg_bind_pages(s, order);
1361        page->slab_cache = s;
1362        __SetPageSlab(page);
1363        if (page->pfmemalloc)
1364                SetPageSlabPfmemalloc(page);
1365
1366        start = page_address(page);
1367
1368        if (unlikely(s->flags & SLAB_POISON))
1369                memset(start, POISON_INUSE, PAGE_SIZE << order);
1370
1371        last = start;
1372        for_each_object(p, s, start, page->objects) {
1373                setup_object(s, page, last);
1374                set_freepointer(s, last, p);
1375                last = p;
1376        }
1377        setup_object(s, page, last);
1378        set_freepointer(s, last, NULL);
1379
1380        page->freelist = start;
1381        page->inuse = page->objects;
1382        page->frozen = 1;
1383out:
1384        return page;
1385}
1386
1387static void __free_slab(struct kmem_cache *s, struct page *page)
1388{
1389        int order = compound_order(page);
1390        int pages = 1 << order;
1391
1392        if (kmem_cache_debug(s)) {
1393                void *p;
1394
1395                slab_pad_check(s, page);
1396                for_each_object(p, s, page_address(page),
1397                                                page->objects)
1398                        check_object(s, page, p, SLUB_RED_INACTIVE);
1399        }
1400
1401        kmemcheck_free_shadow(page, compound_order(page));
1402
1403        mod_zone_page_state(page_zone(page),
1404                (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1405                NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1406                -pages);
1407
1408        __ClearPageSlabPfmemalloc(page);
1409        __ClearPageSlab(page);
1410
1411        memcg_release_pages(s, order);
1412        page_mapcount_reset(page);
1413        if (current->reclaim_state)
1414                current->reclaim_state->reclaimed_slab += pages;
1415        __free_memcg_kmem_pages(page, order);
1416}
1417
1418#define need_reserve_slab_rcu                                           \
1419        (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1420
1421static void rcu_free_slab(struct rcu_head *h)
1422{
1423        struct page *page;
1424
1425        if (need_reserve_slab_rcu)
1426                page = virt_to_head_page(h);
1427        else
1428                page = container_of((struct list_head *)h, struct page, lru);
1429
1430        __free_slab(page->slab_cache, page);
1431}
1432
1433static void free_slab(struct kmem_cache *s, struct page *page)
1434{
1435        if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1436                struct rcu_head *head;
1437
1438                if (need_reserve_slab_rcu) {
1439                        int order = compound_order(page);
1440                        int offset = (PAGE_SIZE << order) - s->reserved;
1441
1442                        VM_BUG_ON(s->reserved != sizeof(*head));
1443                        head = page_address(page) + offset;
1444                } else {
1445                        /*
1446                         * RCU free overloads the RCU head over the LRU
1447                         */
1448                        head = (void *)&page->lru;
1449                }
1450
1451                call_rcu(head, rcu_free_slab);
1452        } else
1453                __free_slab(s, page);
1454}
1455
1456static void discard_slab(struct kmem_cache *s, struct page *page)
1457{
1458        dec_slabs_node(s, page_to_nid(page), page->objects);
1459        free_slab(s, page);
1460}
1461
1462/*
1463 * Management of partially allocated slabs.
1464 *
1465 * list_lock must be held.
1466 */
1467static inline void add_partial(struct kmem_cache_node *n,
1468                                struct page *page, int tail)
1469{
1470        n->nr_partial++;
1471        if (tail == DEACTIVATE_TO_TAIL)
1472                list_add_tail(&page->lru, &n->partial);
1473        else
1474                list_add(&page->lru, &n->partial);
1475}
1476
1477/*
1478 * list_lock must be held.
1479 */
1480static inline void remove_partial(struct kmem_cache_node *n,
1481                                        struct page *page)
1482{
1483        list_del(&page->lru);
1484        n->nr_partial--;
1485}
1486
1487/*
1488 * Remove slab from the partial list, freeze it and
1489 * return the pointer to the freelist.
1490 *
1491 * Returns a list of objects or NULL if it fails.
1492 *
1493 * Must hold list_lock since we modify the partial list.
1494 */
1495static inline void *acquire_slab(struct kmem_cache *s,
1496                struct kmem_cache_node *n, struct page *page,
1497                int mode, int *objects)
1498{
1499        void *freelist;
1500        unsigned long counters;
1501        struct page new;
1502
1503        /*
1504         * Zap the freelist and set the frozen bit.
1505         * The old freelist is the list of objects for the
1506         * per cpu allocation list.
1507         */
1508        freelist = page->freelist;
1509        counters = page->counters;
1510        new.counters = counters;
1511        *objects = new.objects - new.inuse;
1512        if (mode) {
1513                new.inuse = page->objects;
1514                new.freelist = NULL;
1515        } else {
1516                new.freelist = freelist;
1517        }
1518
1519        VM_BUG_ON(new.frozen);
1520        new.frozen = 1;
1521
1522        if (!__cmpxchg_double_slab(s, page,
1523                        freelist, counters,
1524                        new.freelist, new.counters,
1525                        "acquire_slab"))
1526                return NULL;
1527
1528        remove_partial(n, page);
1529        WARN_ON(!freelist);
1530        return freelist;
1531}
1532
1533static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1534static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1535
1536/*
1537 * Try to allocate a partial slab from a specific node.
1538 */
1539static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1540                                struct kmem_cache_cpu *c, gfp_t flags)
1541{
1542        struct page *page, *page2;
1543        void *object = NULL;
1544        int available = 0;
1545        int objects;
1546
1547        /*
1548         * Racy check. If we mistakenly see no partial slabs then we
1549         * just allocate an empty slab. If we mistakenly try to get a
1550         * partial slab and there is none available then get_partials()
1551         * will return NULL.
1552         */
1553        if (!n || !n->nr_partial)
1554                return NULL;
1555
1556        spin_lock(&n->list_lock);
1557        list_for_each_entry_safe(page, page2, &n->partial, lru) {
1558                void *t;
1559
1560                if (!pfmemalloc_match(page, flags))
1561                        continue;
1562
1563                t = acquire_slab(s, n, page, object == NULL, &objects);
1564                if (!t)
1565                        break;
1566
1567                available += objects;
1568                if (!object) {
1569                        c->page = page;
1570                        stat(s, ALLOC_FROM_PARTIAL);
1571                        object = t;
1572                } else {
1573                        put_cpu_partial(s, page, 0);
1574                        stat(s, CPU_PARTIAL_NODE);
1575                }
1576                if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1577                        break;
1578
1579        }
1580        spin_unlock(&n->list_lock);
1581        return object;
1582}
1583
1584/*
1585 * Get a page from somewhere. Search in increasing NUMA distances.
1586 */
1587static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1588                struct kmem_cache_cpu *c)
1589{
1590#ifdef CONFIG_NUMA
1591        struct zonelist *zonelist;
1592        struct zoneref *z;
1593        struct zone *zone;
1594        enum zone_type high_zoneidx = gfp_zone(flags);
1595        void *object;
1596        unsigned int cpuset_mems_cookie;
1597
1598        /*
1599         * The defrag ratio allows a configuration of the tradeoffs between
1600         * inter node defragmentation and node local allocations. A lower
1601         * defrag_ratio increases the tendency to do local allocations
1602         * instead of attempting to obtain partial slabs from other nodes.
1603         *
1604         * If the defrag_ratio is set to 0 then kmalloc() always
1605         * returns node local objects. If the ratio is higher then kmalloc()
1606         * may return off node objects because partial slabs are obtained
1607         * from other nodes and filled up.
1608         *
1609         * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1610         * defrag_ratio = 1000) then every (well almost) allocation will
1611         * first attempt to defrag slab caches on other nodes. This means
1612         * scanning over all nodes to look for partial slabs which may be
1613         * expensive if we do it every time we are trying to find a slab
1614         * with available objects.
1615         */
1616        if (!s->remote_node_defrag_ratio ||
1617                        get_cycles() % 1024 > s->remote_node_defrag_ratio)
1618                return NULL;
1619
1620        do {
1621                cpuset_mems_cookie = get_mems_allowed();
1622                zonelist = node_zonelist(slab_node(), flags);
1623                for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1624                        struct kmem_cache_node *n;
1625
1626                        n = get_node(s, zone_to_nid(zone));
1627
1628                        if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1629                                        n->nr_partial > s->min_partial) {
1630                                object = get_partial_node(s, n, c, flags);
1631                                if (object) {
1632                                        /*
1633                                         * Return the object even if
1634                                         * put_mems_allowed indicated that
1635                                         * the cpuset mems_allowed was
1636                                         * updated in parallel. It's a
1637                                         * harmless race between the alloc
1638                                         * and the cpuset update.
1639                                         */
1640                                        put_mems_allowed(cpuset_mems_cookie);
1641                                        return object;
1642                                }
1643                        }
1644                }
1645        } while (!put_mems_allowed(cpuset_mems_cookie));
1646#endif
1647        return NULL;
1648}
1649
1650/*
1651 * Get a partial page, lock it and return it.
1652 */
1653static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1654                struct kmem_cache_cpu *c)
1655{
1656        void *object;
1657        int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1658
1659        object = get_partial_node(s, get_node(s, searchnode), c, flags);
1660        if (object || node != NUMA_NO_NODE)
1661                return object;
1662
1663        return get_any_partial(s, flags, c);
1664}
1665
1666#ifdef CONFIG_PREEMPT
1667/*
1668 * Calculate the next globally unique transaction for disambiguiation
1669 * during cmpxchg. The transactions start with the cpu number and are then
1670 * incremented by CONFIG_NR_CPUS.
1671 */
1672#define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
1673#else
1674/*
1675 * No preemption supported therefore also no need to check for
1676 * different cpus.
1677 */
1678#define TID_STEP 1
1679#endif
1680
1681static inline unsigned long next_tid(unsigned long tid)
1682{
1683        return tid + TID_STEP;
1684}
1685
1686static inline unsigned int tid_to_cpu(unsigned long tid)
1687{
1688        return tid % TID_STEP;
1689}
1690
1691static inline unsigned long tid_to_event(unsigned long tid)
1692{
1693        return tid / TID_STEP;
1694}
1695
1696static inline unsigned int init_tid(int cpu)
1697{
1698        return cpu;
1699}
1700
1701static inline void note_cmpxchg_failure(const char *n,
1702                const struct kmem_cache *s, unsigned long tid)
1703{
1704#ifdef SLUB_DEBUG_CMPXCHG
1705        unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1706
1707        printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1708
1709#ifdef CONFIG_PREEMPT
1710        if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1711                printk("due to cpu change %d -> %d\n",
1712                        tid_to_cpu(tid), tid_to_cpu(actual_tid));
1713        else
1714#endif
1715        if (tid_to_event(tid) != tid_to_event(actual_tid))
1716                printk("due to cpu running other code. Event %ld->%ld\n",
1717                        tid_to_event(tid), tid_to_event(actual_tid));
1718        else
1719                printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1720                        actual_tid, tid, next_tid(tid));
1721#endif
1722        stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1723}
1724
1725static void init_kmem_cache_cpus(struct kmem_cache *s)
1726{
1727        int cpu;
1728
1729        for_each_possible_cpu(cpu)
1730                per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1731}
1732
1733/*
1734 * Remove the cpu slab
1735 */
1736static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1737{
1738        enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1739        struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1740        int lock = 0;
1741        enum slab_modes l = M_NONE, m = M_NONE;
1742        void *nextfree;
1743        int tail = DEACTIVATE_TO_HEAD;
1744        struct page new;
1745        struct page old;
1746
1747        if (page->freelist) {
1748                stat(s, DEACTIVATE_REMOTE_FREES);
1749                tail = DEACTIVATE_TO_TAIL;
1750        }
1751
1752        /*
1753         * Stage one: Free all available per cpu objects back
1754         * to the page freelist while it is still frozen. Leave the
1755         * last one.
1756         *
1757         * There is no need to take the list->lock because the page
1758         * is still frozen.
1759         */
1760        while (freelist && (nextfree = get_freepointer(s, freelist))) {
1761                void *prior;
1762                unsigned long counters;
1763
1764                do {
1765                        prior = page->freelist;
1766                        counters = page->counters;
1767                        set_freepointer(s, freelist, prior);
1768                        new.counters = counters;
1769                        new.inuse--;
1770                        VM_BUG_ON(!new.frozen);
1771
1772                } while (!__cmpxchg_double_slab(s, page,
1773                        prior, counters,
1774                        freelist, new.counters,
1775                        "drain percpu freelist"));
1776
1777                freelist = nextfree;
1778        }
1779
1780        /*
1781         * Stage two: Ensure that the page is unfrozen while the
1782         * list presence reflects the actual number of objects
1783         * during unfreeze.
1784         *
1785         * We setup the list membership and then perform a cmpxchg
1786         * with the count. If there is a mismatch then the page
1787         * is not unfrozen but the page is on the wrong list.
1788         *
1789         * Then we restart the process which may have to remove
1790         * the page from the list that we just put it on again
1791         * because the number of objects in the slab may have
1792         * changed.
1793         */
1794redo:
1795
1796        old.freelist = page->freelist;
1797        old.counters = page->counters;
1798        VM_BUG_ON(!old.frozen);
1799
1800        /* Determine target state of the slab */
1801        new.counters = old.counters;
1802        if (freelist) {
1803                new.inuse--;
1804                set_freepointer(s, freelist, old.freelist);
1805                new.freelist = freelist;
1806        } else
1807                new.freelist = old.freelist;
1808
1809        new.frozen = 0;
1810
1811        if (!new.inuse && n->nr_partial > s->min_partial)
1812                m = M_FREE;
1813        else if (new.freelist) {
1814                m = M_PARTIAL;
1815                if (!lock) {
1816                        lock = 1;
1817                        /*
1818                         * Taking the spinlock removes the possiblity
1819                         * that acquire_slab() will see a slab page that
1820                         * is frozen
1821                         */
1822                        spin_lock(&n->list_lock);
1823                }
1824        } else {
1825                m = M_FULL;
1826                if (kmem_cache_debug(s) && !lock) {
1827                        lock = 1;
1828                        /*
1829                         * This also ensures that the scanning of full
1830                         * slabs from diagnostic functions will not see
1831                         * any frozen slabs.
1832                         */
1833                        spin_lock(&n->list_lock);
1834                }
1835        }
1836
1837        if (l != m) {
1838
1839                if (l == M_PARTIAL)
1840
1841                        remove_partial(n, page);
1842
1843                else if (l == M_FULL)
1844
1845                        remove_full(s, page);
1846
1847                if (m == M_PARTIAL) {
1848
1849                        add_partial(n, page, tail);
1850                        stat(s, tail);
1851
1852                } else if (m == M_FULL) {
1853
1854                        stat(s, DEACTIVATE_FULL);
1855                        add_full(s, n, page);
1856
1857                }
1858        }
1859
1860        l = m;
1861        if (!__cmpxchg_double_slab(s, page,
1862                                old.freelist, old.counters,
1863                                new.freelist, new.counters,
1864                                "unfreezing slab"))
1865                goto redo;
1866
1867        if (lock)
1868                spin_unlock(&n->list_lock);
1869
1870        if (m == M_FREE) {
1871                stat(s, DEACTIVATE_EMPTY);
1872                discard_slab(s, page);
1873                stat(s, FREE_SLAB);
1874        }
1875}
1876
1877/*
1878 * Unfreeze all the cpu partial slabs.
1879 *
1880 * This function must be called with interrupts disabled
1881 * for the cpu using c (or some other guarantee must be there
1882 * to guarantee no concurrent accesses).
1883 */
1884static void unfreeze_partials(struct kmem_cache *s,
1885                struct kmem_cache_cpu *c)
1886{
1887        struct kmem_cache_node *n = NULL, *n2 = NULL;
1888        struct page *page, *discard_page = NULL;
1889
1890        while ((page = c->partial)) {
1891                struct page new;
1892                struct page old;
1893
1894                c->partial = page->next;
1895
1896                n2 = get_node(s, page_to_nid(page));
1897                if (n != n2) {
1898                        if (n)
1899                                spin_unlock(&n->list_lock);
1900
1901                        n = n2;
1902                        spin_lock(&n->list_lock);
1903                }
1904
1905                do {
1906
1907                        old.freelist = page->freelist;
1908                        old.counters = page->counters;
1909                        VM_BUG_ON(!old.frozen);
1910
1911                        new.counters = old.counters;
1912                        new.freelist = old.freelist;
1913
1914                        new.frozen = 0;
1915
1916                } while (!__cmpxchg_double_slab(s, page,
1917                                old.freelist, old.counters,
1918                                new.freelist, new.counters,
1919                                "unfreezing slab"));
1920
1921                if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1922                        page->next = discard_page;
1923                        discard_page = page;
1924                } else {
1925                        add_partial(n, page, DEACTIVATE_TO_TAIL);
1926                        stat(s, FREE_ADD_PARTIAL);
1927                }
1928        }
1929
1930        if (n)
1931                spin_unlock(&n->list_lock);
1932
1933        while (discard_page) {
1934                page = discard_page;
1935                discard_page = discard_page->next;
1936
1937                stat(s, DEACTIVATE_EMPTY);
1938                discard_slab(s, page);
1939                stat(s, FREE_SLAB);
1940        }
1941}
1942
1943/*
1944 * Put a page that was just frozen (in __slab_free) into a partial page
1945 * slot if available. This is done without interrupts disabled and without
1946 * preemption disabled. The cmpxchg is racy and may put the partial page
1947 * onto a random cpus partial slot.
1948 *
1949 * If we did not find a slot then simply move all the partials to the
1950 * per node partial list.
1951 */
1952static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1953{
1954        struct page *oldpage;
1955        int pages;
1956        int pobjects;
1957
1958        do {
1959                pages = 0;
1960                pobjects = 0;
1961                oldpage = this_cpu_read(s->cpu_slab->partial);
1962
1963                if (oldpage) {
1964                        pobjects = oldpage->pobjects;
1965                        pages = oldpage->pages;
1966                        if (drain && pobjects > s->cpu_partial) {
1967                                unsigned long flags;
1968                                /*
1969                                 * partial array is full. Move the existing
1970                                 * set to the per node partial list.
1971                                 */
1972                                local_irq_save(flags);
1973                                unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
1974                                local_irq_restore(flags);
1975                                oldpage = NULL;
1976                                pobjects = 0;
1977                                pages = 0;
1978                                stat(s, CPU_PARTIAL_DRAIN);
1979                        }
1980                }
1981
1982                pages++;
1983                pobjects += page->objects - page->inuse;
1984
1985                page->pages = pages;
1986                page->pobjects = pobjects;
1987                page->next = oldpage;
1988
1989        } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1990}
1991
1992static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1993{
1994        stat(s, CPUSLAB_FLUSH);
1995        deactivate_slab(s, c->page, c->freelist);
1996
1997        c->tid = next_tid(c->tid);
1998        c->page = NULL;
1999        c->freelist = NULL;
2000}
2001
2002/*
2003 * Flush cpu slab.
2004 *
2005 * Called from IPI handler with interrupts disabled.
2006 */
2007static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2008{
2009        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2010
2011        if (likely(c)) {
2012                if (c->page)
2013                        flush_slab(s, c);
2014
2015                unfreeze_partials(s, c);
2016        }
2017}
2018
2019static void flush_cpu_slab(void *d)
2020{
2021        struct kmem_cache *s = d;
2022
2023        __flush_cpu_slab(s, smp_processor_id());
2024}
2025
2026static bool has_cpu_slab(int cpu, void *info)
2027{
2028        struct kmem_cache *s = info;
2029        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2030
2031        return c->page || c->partial;
2032}
2033
2034static void flush_all(struct kmem_cache *s)
2035{
2036        on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2037}
2038
2039/*
2040 * Check if the objects in a per cpu structure fit numa
2041 * locality expectations.
2042 */
2043static inline int node_match(struct page *page, int node)
2044{
2045#ifdef CONFIG_NUMA
2046        if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2047                return 0;
2048#endif
2049        return 1;
2050}
2051
2052static int count_free(struct page *page)
2053{
2054        return page->objects - page->inuse;
2055}
2056
2057static unsigned long count_partial(struct kmem_cache_node *n,
2058                                        int (*get_count)(struct page *))
2059{
2060        unsigned long flags;
2061        unsigned long x = 0;
2062        struct page *page;
2063
2064        spin_lock_irqsave(&n->list_lock, flags);
2065        list_for_each_entry(page, &n->partial, lru)
2066                x += get_count(page);
2067        spin_unlock_irqrestore(&n->list_lock, flags);
2068        return x;
2069}
2070
2071static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2072{
2073#ifdef CONFIG_SLUB_DEBUG
2074        return atomic_long_read(&n->total_objects);
2075#else
2076        return 0;
2077#endif
2078}
2079
2080static noinline void
2081slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2082{
2083        int node;
2084
2085        printk(KERN_WARNING
2086                "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2087                nid, gfpflags);
2088        printk(KERN_WARNING "  cache: %s, object size: %d, buffer size: %d, "
2089                "default order: %d, min order: %d\n", s->name, s->object_size,
2090                s->size, oo_order(s->oo), oo_order(s->min));
2091
2092        if (oo_order(s->min) > get_order(s->object_size))
2093                printk(KERN_WARNING "  %s debugging increased min order, use "
2094                       "slub_debug=O to disable.\n", s->name);
2095
2096        for_each_online_node(node) {
2097                struct kmem_cache_node *n = get_node(s, node);
2098                unsigned long nr_slabs;
2099                unsigned long nr_objs;
2100                unsigned long nr_free;
2101
2102                if (!n)
2103                        continue;
2104
2105                nr_free  = count_partial(n, count_free);
2106                nr_slabs = node_nr_slabs(n);
2107                nr_objs  = node_nr_objs(n);
2108
2109                printk(KERN_WARNING
2110                        "  node %d: slabs: %ld, objs: %ld, free: %ld\n",
2111                        node, nr_slabs, nr_objs, nr_free);
2112        }
2113}
2114
2115static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2116                        int node, struct kmem_cache_cpu **pc)
2117{
2118        void *freelist;
2119        struct kmem_cache_cpu *c = *pc;
2120        struct page *page;
2121
2122        freelist = get_partial(s, flags, node, c);
2123
2124        if (freelist)
2125                return freelist;
2126
2127        page = new_slab(s, flags, node);
2128        if (page) {
2129                c = __this_cpu_ptr(s->cpu_slab);
2130                if (c->page)
2131                        flush_slab(s, c);
2132
2133                /*
2134                 * No other reference to the page yet so we can
2135                 * muck around with it freely without cmpxchg
2136                 */
2137                freelist = page->freelist;
2138                page->freelist = NULL;
2139
2140                stat(s, ALLOC_SLAB);
2141                c->page = page;
2142                *pc = c;
2143        } else
2144                freelist = NULL;
2145
2146        return freelist;
2147}
2148
2149static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2150{
2151        if (unlikely(PageSlabPfmemalloc(page)))
2152                return gfp_pfmemalloc_allowed(gfpflags);
2153
2154        return true;
2155}
2156
2157/*
2158 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2159 * or deactivate the page.
2160 *
2161 * The page is still frozen if the return value is not NULL.
2162 *
2163 * If this function returns NULL then the page has been unfrozen.
2164 *
2165 * This function must be called with interrupt disabled.
2166 */
2167static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2168{
2169        struct page new;
2170        unsigned long counters;
2171        void *freelist;
2172
2173        do {
2174                freelist = page->freelist;
2175                counters = page->counters;
2176
2177                new.counters = counters;
2178                VM_BUG_ON(!new.frozen);
2179
2180                new.inuse = page->objects;
2181                new.frozen = freelist != NULL;
2182
2183        } while (!__cmpxchg_double_slab(s, page,
2184                freelist, counters,
2185                NULL, new.counters,
2186                "get_freelist"));
2187
2188        return freelist;
2189}
2190
2191/*
2192 * Slow path. The lockless freelist is empty or we need to perform
2193 * debugging duties.
2194 *
2195 * Processing is still very fast if new objects have been freed to the
2196 * regular freelist. In that case we simply take over the regular freelist
2197 * as the lockless freelist and zap the regular freelist.
2198 *
2199 * If that is not working then we fall back to the partial lists. We take the
2200 * first element of the freelist as the object to allocate now and move the
2201 * rest of the freelist to the lockless freelist.
2202 *
2203 * And if we were unable to get a new slab from the partial slab lists then
2204 * we need to allocate a new slab. This is the slowest path since it involves
2205 * a call to the page allocator and the setup of a new slab.
2206 */
2207static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2208                          unsigned long addr, struct kmem_cache_cpu *c)
2209{
2210        void *freelist;
2211        struct page *page;
2212        unsigned long flags;
2213
2214        local_irq_save(flags);
2215#ifdef CONFIG_PREEMPT
2216        /*
2217         * We may have been preempted and rescheduled on a different
2218         * cpu before disabling interrupts. Need to reload cpu area
2219         * pointer.
2220         */
2221        c = this_cpu_ptr(s->cpu_slab);
2222#endif
2223
2224        page = c->page;
2225        if (!page)
2226                goto new_slab;
2227redo:
2228
2229        if (unlikely(!node_match(page, node))) {
2230                stat(s, ALLOC_NODE_MISMATCH);
2231                deactivate_slab(s, page, c->freelist);
2232                c->page = NULL;
2233                c->freelist = NULL;
2234                goto new_slab;
2235        }
2236
2237        /*
2238         * By rights, we should be searching for a slab page that was
2239         * PFMEMALLOC but right now, we are losing the pfmemalloc
2240         * information when the page leaves the per-cpu allocator
2241         */
2242        if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2243                deactivate_slab(s, page, c->freelist);
2244                c->page = NULL;
2245                c->freelist = NULL;
2246                goto new_slab;
2247        }
2248
2249        /* must check again c->freelist in case of cpu migration or IRQ */
2250        freelist = c->freelist;
2251        if (freelist)
2252                goto load_freelist;
2253
2254        stat(s, ALLOC_SLOWPATH);
2255
2256        freelist = get_freelist(s, page);
2257
2258        if (!freelist) {
2259                c->page = NULL;
2260                stat(s, DEACTIVATE_BYPASS);
2261                goto new_slab;
2262        }
2263
2264        stat(s, ALLOC_REFILL);
2265
2266load_freelist:
2267        /*
2268         * freelist is pointing to the list of objects to be used.
2269         * page is pointing to the page from which the objects are obtained.
2270         * That page must be frozen for per cpu allocations to work.
2271         */
2272        VM_BUG_ON(!c->page->frozen);
2273        c->freelist = get_freepointer(s, freelist);
2274        c->tid = next_tid(c->tid);
2275        local_irq_restore(flags);
2276        return freelist;
2277
2278new_slab:
2279
2280        if (c->partial) {
2281                page = c->page = c->partial;
2282                c->partial = page->next;
2283                stat(s, CPU_PARTIAL_ALLOC);
2284                c->freelist = NULL;
2285                goto redo;
2286        }
2287
2288        freelist = new_slab_objects(s, gfpflags, node, &c);
2289
2290        if (unlikely(!freelist)) {
2291                if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2292                        slab_out_of_memory(s, gfpflags, node);
2293
2294                local_irq_restore(flags);
2295                return NULL;
2296        }
2297
2298        page = c->page;
2299        if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2300                goto load_freelist;
2301
2302        /* Only entered in the debug case */
2303        if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2304                goto new_slab;  /* Slab failed checks. Next slab needed */
2305
2306        deactivate_slab(s, page, get_freepointer(s, freelist));
2307        c->page = NULL;
2308        c->freelist = NULL;
2309        local_irq_restore(flags);
2310        return freelist;
2311}
2312
2313/*
2314 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2315 * have the fastpath folded into their functions. So no function call
2316 * overhead for requests that can be satisfied on the fastpath.
2317 *
2318 * The fastpath works by first checking if the lockless freelist can be used.
2319 * If not then __slab_alloc is called for slow processing.
2320 *
2321 * Otherwise we can simply pick the next object from the lockless free list.
2322 */
2323static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2324                gfp_t gfpflags, int node, unsigned long addr)
2325{
2326        void **object;
2327        struct kmem_cache_cpu *c;
2328        struct page *page;
2329        unsigned long tid;
2330
2331        if (slab_pre_alloc_hook(s, gfpflags))
2332                return NULL;
2333
2334        s = memcg_kmem_get_cache(s, gfpflags);
2335redo:
2336        /*
2337         * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2338         * enabled. We may switch back and forth between cpus while
2339         * reading from one cpu area. That does not matter as long
2340         * as we end up on the original cpu again when doing the cmpxchg.
2341         *
2342         * Preemption is disabled for the retrieval of the tid because that
2343         * must occur from the current processor. We cannot allow rescheduling
2344         * on a different processor between the determination of the pointer
2345         * and the retrieval of the tid.
2346         */
2347        preempt_disable();
2348        c = __this_cpu_ptr(s->cpu_slab);
2349
2350        /*
2351         * The transaction ids are globally unique per cpu and per operation on
2352         * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2353         * occurs on the right processor and that there was no operation on the
2354         * linked list in between.
2355         */
2356        tid = c->tid;
2357        preempt_enable();
2358
2359        object = c->freelist;
2360        page = c->page;
2361        if (unlikely(!object || !node_match(page, node)))
2362                object = __slab_alloc(s, gfpflags, node, addr, c);
2363
2364        else {
2365                void *next_object = get_freepointer_safe(s, object);
2366
2367                /*
2368                 * The cmpxchg will only match if there was no additional
2369                 * operation and if we are on the right processor.
2370                 *
2371                 * The cmpxchg does the following atomically (without lock semantics!)
2372                 * 1. Relocate first pointer to the current per cpu area.
2373                 * 2. Verify that tid and freelist have not been changed
2374                 * 3. If they were not changed replace tid and freelist
2375                 *
2376                 * Since this is without lock semantics the protection is only against
2377                 * code executing on this cpu *not* from access by other cpus.
2378                 */
2379                if (unlikely(!this_cpu_cmpxchg_double(
2380                                s->cpu_slab->freelist, s->cpu_slab->tid,
2381                                object, tid,
2382                                next_object, next_tid(tid)))) {
2383
2384                        note_cmpxchg_failure("slab_alloc", s, tid);
2385                        goto redo;
2386                }
2387                prefetch_freepointer(s, next_object);
2388                stat(s, ALLOC_FASTPATH);
2389        }
2390
2391        if (unlikely(gfpflags & __GFP_ZERO) && object)
2392                memset(object, 0, s->object_size);
2393
2394        slab_post_alloc_hook(s, gfpflags, object);
2395
2396        return object;
2397}
2398
2399static __always_inline void *slab_alloc(struct kmem_cache *s,
2400                gfp_t gfpflags, unsigned long addr)
2401{
2402        return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2403}
2404
2405void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2406{
2407        void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2408
2409        trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2410
2411        return ret;
2412}
2413EXPORT_SYMBOL(kmem_cache_alloc);
2414
2415#ifdef CONFIG_TRACING
2416void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2417{
2418        void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2419        trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2420        return ret;
2421}
2422EXPORT_SYMBOL(kmem_cache_alloc_trace);
2423
2424void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2425{
2426        void *ret = kmalloc_order(size, flags, order);
2427        trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2428        return ret;
2429}
2430EXPORT_SYMBOL(kmalloc_order_trace);
2431#endif
2432
2433#ifdef CONFIG_NUMA
2434void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2435{
2436        void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2437
2438        trace_kmem_cache_alloc_node(_RET_IP_, ret,
2439                                    s->object_size, s->size, gfpflags, node);
2440
2441        return ret;
2442}
2443EXPORT_SYMBOL(kmem_cache_alloc_node);
2444
2445#ifdef CONFIG_TRACING
2446void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2447                                    gfp_t gfpflags,
2448                                    int node, size_t size)
2449{
2450        void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2451
2452        trace_kmalloc_node(_RET_IP_, ret,
2453                           size, s->size, gfpflags, node);
2454        return ret;
2455}
2456EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2457#endif
2458#endif
2459
2460/*
2461 * Slow patch handling. This may still be called frequently since objects
2462 * have a longer lifetime than the cpu slabs in most processing loads.
2463 *
2464 * So we still attempt to reduce cache line usage. Just take the slab
2465 * lock and free the item. If there is no additional partial page
2466 * handling required then we can return immediately.
2467 */
2468static void __slab_free(struct kmem_cache *s, struct page *page,
2469                        void *x, unsigned long addr)
2470{
2471        void *prior;
2472        void **object = (void *)x;
2473        int was_frozen;
2474        struct page new;
2475        unsigned long counters;
2476        struct kmem_cache_node *n = NULL;
2477        unsigned long uninitialized_var(flags);
2478
2479        stat(s, FREE_SLOWPATH);
2480
2481        if (kmem_cache_debug(s) &&
2482                !(n = free_debug_processing(s, page, x, addr, &flags)))
2483                return;
2484
2485        do {
2486                if (unlikely(n)) {
2487                        spin_unlock_irqrestore(&n->list_lock, flags);
2488                        n = NULL;
2489                }
2490                prior = page->freelist;
2491                counters = page->counters;
2492                set_freepointer(s, object, prior);
2493                new.counters = counters;
2494                was_frozen = new.frozen;
2495                new.inuse--;
2496                if ((!new.inuse || !prior) && !was_frozen) {
2497
2498                        if (!kmem_cache_debug(s) && !prior)
2499
2500                                /*
2501                                 * Slab was on no list before and will be partially empty
2502                                 * We can defer the list move and instead freeze it.
2503                                 */
2504                                new.frozen = 1;
2505
2506                        else { /* Needs to be taken off a list */
2507
2508                                n = get_node(s, page_to_nid(page));
2509                                /*
2510                                 * Speculatively acquire the list_lock.
2511                                 * If the cmpxchg does not succeed then we may
2512                                 * drop the list_lock without any processing.
2513                                 *
2514                                 * Otherwise the list_lock will synchronize with
2515                                 * other processors updating the list of slabs.
2516                                 */
2517                                spin_lock_irqsave(&n->list_lock, flags);
2518
2519                        }
2520                }
2521
2522        } while (!cmpxchg_double_slab(s, page,
2523                prior, counters,
2524                object, new.counters,
2525                "__slab_free"));
2526
2527        if (likely(!n)) {
2528
2529                /*
2530                 * If we just froze the page then put it onto the
2531                 * per cpu partial list.
2532                 */
2533                if (new.frozen && !was_frozen) {
2534                        put_cpu_partial(s, page, 1);
2535                        stat(s, CPU_PARTIAL_FREE);
2536                }
2537                /*
2538                 * The list lock was not taken therefore no list
2539                 * activity can be necessary.
2540                 */
2541                if (was_frozen)
2542                        stat(s, FREE_FROZEN);
2543                return;
2544        }
2545
2546        if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2547                goto slab_empty;
2548
2549        /*
2550         * Objects left in the slab. If it was not on the partial list before
2551         * then add it.
2552         */
2553        if (kmem_cache_debug(s) && unlikely(!prior)) {
2554                remove_full(s, page);
2555                add_partial(n, page, DEACTIVATE_TO_TAIL);
2556                stat(s, FREE_ADD_PARTIAL);
2557        }
2558        spin_unlock_irqrestore(&n->list_lock, flags);
2559        return;
2560
2561slab_empty:
2562        if (prior) {
2563                /*
2564                 * Slab on the partial list.
2565                 */
2566                remove_partial(n, page);
2567                stat(s, FREE_REMOVE_PARTIAL);
2568        } else
2569                /* Slab must be on the full list */
2570                remove_full(s, page);
2571
2572        spin_unlock_irqrestore(&n->list_lock, flags);
2573        stat(s, FREE_SLAB);
2574        discard_slab(s, page);
2575}
2576
2577/*
2578 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2579 * can perform fastpath freeing without additional function calls.
2580 *
2581 * The fastpath is only possible if we are freeing to the current cpu slab
2582 * of this processor. This typically the case if we have just allocated
2583 * the item before.
2584 *
2585 * If fastpath is not possible then fall back to __slab_free where we deal
2586 * with all sorts of special processing.
2587 */
2588static __always_inline void slab_free(struct kmem_cache *s,
2589                        struct page *page, void *x, unsigned long addr)
2590{
2591        void **object = (void *)x;
2592        struct kmem_cache_cpu *c;
2593        unsigned long tid;
2594
2595        slab_free_hook(s, x);
2596
2597redo:
2598        /*
2599         * Determine the currently cpus per cpu slab.
2600         * The cpu may change afterward. However that does not matter since
2601         * data is retrieved via this pointer. If we are on the same cpu
2602         * during the cmpxchg then the free will succedd.
2603         */
2604        preempt_disable();
2605        c = __this_cpu_ptr(s->cpu_slab);
2606
2607        tid = c->tid;
2608        preempt_enable();
2609
2610        if (likely(page == c->page)) {
2611                set_freepointer(s, object, c->freelist);
2612
2613                if (unlikely(!this_cpu_cmpxchg_double(
2614                                s->cpu_slab->freelist, s->cpu_slab->tid,
2615                                c->freelist, tid,
2616                                object, next_tid(tid)))) {
2617
2618                        note_cmpxchg_failure("slab_free", s, tid);
2619                        goto redo;
2620                }
2621                stat(s, FREE_FASTPATH);
2622        } else
2623                __slab_free(s, page, x, addr);
2624
2625}
2626
2627void kmem_cache_free(struct kmem_cache *s, void *x)
2628{
2629        s = cache_from_obj(s, x);
2630        if (!s)
2631                return;
2632        slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2633        trace_kmem_cache_free(_RET_IP_, x);
2634}
2635EXPORT_SYMBOL(kmem_cache_free);
2636
2637/*
2638 * Object placement in a slab is made very easy because we always start at
2639 * offset 0. If we tune the size of the object to the alignment then we can
2640 * get the required alignment by putting one properly sized object after
2641 * another.
2642 *
2643 * Notice that the allocation order determines the sizes of the per cpu
2644 * caches. Each processor has always one slab available for allocations.
2645 * Increasing the allocation order reduces the number of times that slabs
2646 * must be moved on and off the partial lists and is therefore a factor in
2647 * locking overhead.
2648 */
2649
2650/*
2651 * Mininum / Maximum order of slab pages. This influences locking overhead
2652 * and slab fragmentation. A higher order reduces the number of partial slabs
2653 * and increases the number of allocations possible without having to
2654 * take the list_lock.
2655 */
2656static int slub_min_order;
2657static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2658static int slub_min_objects;
2659
2660/*
2661 * Merge control. If this is set then no merging of slab caches will occur.
2662 * (Could be removed. This was introduced to pacify the merge skeptics.)
2663 */
2664static int slub_nomerge;
2665
2666/*
2667 * Calculate the order of allocation given an slab object size.
2668 *
2669 * The order of allocation has significant impact on performance and other
2670 * system components. Generally order 0 allocations should be preferred since
2671 * order 0 does not cause fragmentation in the page allocator. Larger objects
2672 * be problematic to put into order 0 slabs because there may be too much
2673 * unused space left. We go to a higher order if more than 1/16th of the slab
2674 * would be wasted.
2675 *
2676 * In order to reach satisfactory performance we must ensure that a minimum
2677 * number of objects is in one slab. Otherwise we may generate too much
2678 * activity on the partial lists which requires taking the list_lock. This is
2679 * less a concern for large slabs though which are rarely used.
2680 *
2681 * slub_max_order specifies the order where we begin to stop considering the
2682 * number of objects in a slab as critical. If we reach slub_max_order then
2683 * we try to keep the page order as low as possible. So we accept more waste
2684 * of space in favor of a small page order.
2685 *
2686 * Higher order allocations also allow the placement of more objects in a
2687 * slab and thereby reduce object handling overhead. If the user has
2688 * requested a higher mininum order then we start with that one instead of
2689 * the smallest order which will fit the object.
2690 */
2691static inline int slab_order(int size, int min_objects,
2692                                int max_order, int fract_leftover, int reserved)
2693{
2694        int order;
2695        int rem;
2696        int min_order = slub_min_order;
2697
2698        if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2699                return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2700
2701        for (order = max(min_order,
2702                                fls(min_objects * size - 1) - PAGE_SHIFT);
2703                        order <= max_order; order++) {
2704
2705                unsigned long slab_size = PAGE_SIZE << order;
2706
2707                if (slab_size < min_objects * size + reserved)
2708                        continue;
2709
2710                rem = (slab_size - reserved) % size;
2711
2712                if (rem <= slab_size / fract_leftover)
2713                        break;
2714
2715        }
2716
2717        return order;
2718}
2719
2720static inline int calculate_order(int size, int reserved)
2721{
2722        int order;
2723        int min_objects;
2724        int fraction;
2725        int max_objects;
2726
2727        /*
2728         * Attempt to find best configuration for a slab. This
2729         * works by first attempting to generate a layout with
2730         * the best configuration and backing off gradually.
2731         *
2732         * First we reduce the acceptable waste in a slab. Then
2733         * we reduce the minimum objects required in a slab.
2734         */
2735        min_objects = slub_min_objects;
2736        if (!min_objects)
2737                min_objects = 4 * (fls(nr_cpu_ids) + 1);
2738        max_objects = order_objects(slub_max_order, size, reserved);
2739        min_objects = min(min_objects, max_objects);
2740
2741        while (min_objects > 1) {
2742                fraction = 16;
2743                while (fraction >= 4) {
2744                        order = slab_order(size, min_objects,
2745                                        slub_max_order, fraction, reserved);
2746                        if (order <= slub_max_order)
2747                                return order;
2748                        fraction /= 2;
2749                }
2750                min_objects--;
2751        }
2752
2753        /*
2754         * We were unable to place multiple objects in a slab. Now
2755         * lets see if we can place a single object there.
2756         */
2757        order = slab_order(size, 1, slub_max_order, 1, reserved);
2758        if (order <= slub_max_order)
2759                return order;
2760
2761        /*
2762         * Doh this slab cannot be placed using slub_max_order.
2763         */
2764        order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2765        if (order < MAX_ORDER)
2766                return order;
2767        return -ENOSYS;
2768}
2769
2770static void
2771init_kmem_cache_node(struct kmem_cache_node *n)
2772{
2773        n->nr_partial = 0;
2774        spin_lock_init(&n->list_lock);
2775        INIT_LIST_HEAD(&n->partial);
2776#ifdef CONFIG_SLUB_DEBUG
2777        atomic_long_set(&n->nr_slabs, 0);
2778        atomic_long_set(&n->total_objects, 0);
2779        INIT_LIST_HEAD(&n->full);
2780#endif
2781}
2782
2783static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2784{
2785        BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2786                        KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2787
2788        /*
2789         * Must align to double word boundary for the double cmpxchg
2790         * instructions to work; see __pcpu_double_call_return_bool().
2791         */
2792        s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2793                                     2 * sizeof(void *));
2794
2795        if (!s->cpu_slab)
2796                return 0;
2797
2798        init_kmem_cache_cpus(s);
2799
2800        return 1;
2801}
2802
2803static struct kmem_cache *kmem_cache_node;
2804
2805/*
2806 * No kmalloc_node yet so do it by hand. We know that this is the first
2807 * slab on the node for this slabcache. There are no concurrent accesses
2808 * possible.
2809 *
2810 * Note that this function only works on the kmalloc_node_cache
2811 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2812 * memory on a fresh node that has no slab structures yet.
2813 */
2814static void early_kmem_cache_node_alloc(int node)
2815{
2816        struct page *page;
2817        struct kmem_cache_node *n;
2818
2819        BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2820
2821        page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2822
2823        BUG_ON(!page);
2824        if (page_to_nid(page) != node) {
2825                printk(KERN_ERR "SLUB: Unable to allocate memory from "
2826                                "node %d\n", node);
2827                printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2828                                "in order to be able to continue\n");
2829        }
2830
2831        n = page->freelist;
2832        BUG_ON(!n);
2833        page->freelist = get_freepointer(kmem_cache_node, n);
2834        page->inuse = 1;
2835        page->frozen = 0;
2836        kmem_cache_node->node[node] = n;
2837#ifdef CONFIG_SLUB_DEBUG
2838        init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2839        init_tracking(kmem_cache_node, n);
2840#endif
2841        init_kmem_cache_node(n);
2842        inc_slabs_node(kmem_cache_node, node, page->objects);
2843
2844        add_partial(n, page, DEACTIVATE_TO_HEAD);
2845}
2846
2847static void free_kmem_cache_nodes(struct kmem_cache *s)
2848{
2849        int node;
2850
2851        for_each_node_state(node, N_NORMAL_MEMORY) {
2852                struct kmem_cache_node *n = s->node[node];
2853
2854                if (n)
2855                        kmem_cache_free(kmem_cache_node, n);
2856
2857                s->node[node] = NULL;
2858        }
2859}
2860
2861static int init_kmem_cache_nodes(struct kmem_cache *s)
2862{
2863        int node;
2864
2865        for_each_node_state(node, N_NORMAL_MEMORY) {
2866                struct kmem_cache_node *n;
2867
2868                if (slab_state == DOWN) {
2869                        early_kmem_cache_node_alloc(node);
2870                        continue;
2871                }
2872                n = kmem_cache_alloc_node(kmem_cache_node,
2873                                                GFP_KERNEL, node);
2874
2875                if (!n) {
2876                        free_kmem_cache_nodes(s);
2877                        return 0;
2878                }
2879
2880                s->node[node] = n;
2881                init_kmem_cache_node(n);
2882        }
2883        return 1;
2884}
2885
2886static void set_min_partial(struct kmem_cache *s, unsigned long min)
2887{
2888        if (min < MIN_PARTIAL)
2889                min = MIN_PARTIAL;
2890        else if (min > MAX_PARTIAL)
2891                min = MAX_PARTIAL;
2892        s->min_partial = min;
2893}
2894
2895/*
2896 * calculate_sizes() determines the order and the distribution of data within
2897 * a slab object.
2898 */
2899static int calculate_sizes(struct kmem_cache *s, int forced_order)
2900{
2901        unsigned long flags = s->flags;
2902        unsigned long size = s->object_size;
2903        int order;
2904
2905        /*
2906         * Round up object size to the next word boundary. We can only
2907         * place the free pointer at word boundaries and this determines
2908         * the possible location of the free pointer.
2909         */
2910        size = ALIGN(size, sizeof(void *));
2911
2912#ifdef CONFIG_SLUB_DEBUG
2913        /*
2914         * Determine if we can poison the object itself. If the user of
2915         * the slab may touch the object after free or before allocation
2916         * then we should never poison the object itself.
2917         */
2918        if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2919                        !s->ctor)
2920                s->flags |= __OBJECT_POISON;
2921        else
2922                s->flags &= ~__OBJECT_POISON;
2923
2924
2925        /*
2926         * If we are Redzoning then check if there is some space between the
2927         * end of the object and the free pointer. If not then add an
2928         * additional word to have some bytes to store Redzone information.
2929         */
2930        if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2931                size += sizeof(void *);
2932#endif
2933
2934        /*
2935         * With that we have determined the number of bytes in actual use
2936         * by the object. This is the potential offset to the free pointer.
2937         */
2938        s->inuse = size;
2939
2940        if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2941                s->ctor)) {
2942                /*
2943                 * Relocate free pointer after the object if it is not
2944                 * permitted to overwrite the first word of the object on
2945                 * kmem_cache_free.
2946                 *
2947                 * This is the case if we do RCU, have a constructor or
2948                 * destructor or are poisoning the objects.
2949                 */
2950                s->offset = size;
2951                size += sizeof(void *);
2952        }
2953
2954#ifdef CONFIG_SLUB_DEBUG
2955        if (flags & SLAB_STORE_USER)
2956                /*
2957                 * Need to store information about allocs and frees after
2958                 * the object.
2959                 */
2960                size += 2 * sizeof(struct track);
2961
2962        if (flags & SLAB_RED_ZONE)
2963                /*
2964                 * Add some empty padding so that we can catch
2965                 * overwrites from earlier objects rather than let
2966                 * tracking information or the free pointer be
2967                 * corrupted if a user writes before the start
2968                 * of the object.
2969                 */
2970                size += sizeof(void *);
2971#endif
2972
2973        /*
2974         * SLUB stores one object immediately after another beginning from
2975         * offset 0. In order to align the objects we have to simply size
2976         * each object to conform to the alignment.
2977         */
2978        size = ALIGN(size, s->align);
2979        s->size = size;
2980        if (forced_order >= 0)
2981                order = forced_order;
2982        else
2983                order = calculate_order(size, s->reserved);
2984
2985        if (order < 0)
2986                return 0;
2987
2988        s->allocflags = 0;
2989        if (order)
2990                s->allocflags |= __GFP_COMP;
2991
2992        if (s->flags & SLAB_CACHE_DMA)
2993                s->allocflags |= GFP_DMA;
2994
2995        if (s->flags & SLAB_RECLAIM_ACCOUNT)
2996                s->allocflags |= __GFP_RECLAIMABLE;
2997
2998        /*
2999         * Determine the number of objects per slab
3000         */
3001        s->oo = oo_make(order, size, s->reserved);
3002        s->min = oo_make(get_order(size), size, s->reserved);
3003        if (oo_objects(s->oo) > oo_objects(s->max))
3004                s->max = s->oo;
3005
3006        return !!oo_objects(s->oo);
3007}
3008
3009static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3010{
3011        s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3012        s->reserved = 0;
3013
3014        if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3015                s->reserved = sizeof(struct rcu_head);
3016
3017        if (!calculate_sizes(s, -1))
3018                goto error;
3019        if (disable_higher_order_debug) {
3020                /*
3021                 * Disable debugging flags that store metadata if the min slab
3022                 * order increased.
3023                 */
3024                if (get_order(s->size) > get_order(s->object_size)) {
3025                        s->flags &= ~DEBUG_METADATA_FLAGS;
3026                        s->offset = 0;
3027                        if (!calculate_sizes(s, -1))
3028                                goto error;
3029                }
3030        }
3031
3032#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3033    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3034        if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3035                /* Enable fast mode */
3036                s->flags |= __CMPXCHG_DOUBLE;
3037#endif
3038
3039        /*
3040         * The larger the object size is, the more pages we want on the partial
3041         * list to avoid pounding the page allocator excessively.
3042         */
3043        set_min_partial(s, ilog2(s->size) / 2);
3044
3045        /*
3046         * cpu_partial determined the maximum number of objects kept in the
3047         * per cpu partial lists of a processor.
3048         *
3049         * Per cpu partial lists mainly contain slabs that just have one
3050         * object freed. If they are used for allocation then they can be
3051         * filled up again with minimal effort. The slab will never hit the
3052         * per node partial lists and therefore no locking will be required.
3053         *
3054         * This setting also determines
3055         *
3056         * A) The number of objects from per cpu partial slabs dumped to the
3057         *    per node list when we reach the limit.
3058         * B) The number of objects in cpu partial slabs to extract from the
3059         *    per node list when we run out of per cpu objects. We only fetch 50%
3060         *    to keep some capacity around for frees.
3061         */
3062        if (kmem_cache_debug(s))
3063                s->cpu_partial = 0;
3064        else if (s->size >= PAGE_SIZE)
3065                s->cpu_partial = 2;
3066        else if (s->size >= 1024)
3067                s->cpu_partial = 6;
3068        else if (s->size >= 256)
3069                s->cpu_partial = 13;
3070        else
3071                s->cpu_partial = 30;
3072
3073#ifdef CONFIG_NUMA
3074        s->remote_node_defrag_ratio = 1000;
3075#endif
3076        if (!init_kmem_cache_nodes(s))
3077                goto error;
3078
3079        if (alloc_kmem_cache_cpus(s))
3080                return 0;
3081
3082        free_kmem_cache_nodes(s);
3083error:
3084        if (flags & SLAB_PANIC)
3085                panic("Cannot create slab %s size=%lu realsize=%u "
3086                        "order=%u offset=%u flags=%lx\n",
3087                        s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3088                        s->offset, flags);
3089        return -EINVAL;
3090}
3091
3092static void list_slab_objects(struct kmem_cache *s, struct page *page,
3093                                                        const char *text)
3094{
3095#ifdef CONFIG_SLUB_DEBUG
3096        void *addr = page_address(page);
3097        void *p;
3098        unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3099                                     sizeof(long), GFP_ATOMIC);
3100        if (!map)
3101                return;
3102        slab_err(s, page, text, s->name);
3103        slab_lock(page);
3104
3105        get_map(s, page, map);
3106        for_each_object(p, s, addr, page->objects) {
3107
3108                if (!test_bit(slab_index(p, s, addr), map)) {
3109                        printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3110                                                        p, p - addr);
3111                        print_tracking(s, p);
3112                }
3113        }
3114        slab_unlock(page);
3115        kfree(map);
3116#endif
3117}
3118
3119/*
3120 * Attempt to free all partial slabs on a node.
3121 * This is called from kmem_cache_close(). We must be the last thread
3122 * using the cache and therefore we do not need to lock anymore.
3123 */
3124static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3125{
3126        struct page *page, *h;
3127
3128        list_for_each_entry_safe(page, h, &n->partial, lru) {
3129                if (!page->inuse) {
3130                        remove_partial(n, page);
3131                        discard_slab(s, page);
3132                } else {
3133                        list_slab_objects(s, page,
3134                        "Objects remaining in %s on kmem_cache_close()");
3135                }
3136        }
3137}
3138
3139/*
3140 * Release all resources used by a slab cache.
3141 */
3142static inline int kmem_cache_close(struct kmem_cache *s)
3143{
3144        int node;
3145
3146        flush_all(s);
3147        /* Attempt to free all objects */
3148        for_each_node_state(node, N_NORMAL_MEMORY) {
3149                struct kmem_cache_node *n = get_node(s, node);
3150
3151                free_partial(s, n);
3152                if (n->nr_partial || slabs_node(s, node))
3153                        return 1;
3154        }
3155        free_percpu(s->cpu_slab);
3156        free_kmem_cache_nodes(s);
3157        return 0;
3158}
3159
3160int __kmem_cache_shutdown(struct kmem_cache *s)
3161{
3162        int rc = kmem_cache_close(s);
3163
3164        if (!rc) {
3165                /*
3166                 * We do the same lock strategy around sysfs_slab_add, see
3167                 * __kmem_cache_create. Because this is pretty much the last
3168                 * operation we do and the lock will be released shortly after
3169                 * that in slab_common.c, we could just move sysfs_slab_remove
3170                 * to a later point in common code. We should do that when we
3171                 * have a common sysfs framework for all allocators.
3172                 */
3173                mutex_unlock(&slab_mutex);
3174                sysfs_slab_remove(s);
3175                mutex_lock(&slab_mutex);
3176        }
3177
3178        return rc;
3179}
3180
3181/********************************************************************
3182 *              Kmalloc subsystem
3183 *******************************************************************/
3184
3185static int __init setup_slub_min_order(char *str)
3186{
3187        get_option(&str, &slub_min_order);
3188
3189        return 1;
3190}
3191
3192__setup("slub_min_order=", setup_slub_min_order);
3193
3194static int __init setup_slub_max_order(char *str)
3195{
3196        get_option(&str, &slub_max_order);
3197        slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3198
3199        return 1;
3200}
3201
3202__setup("slub_max_order=", setup_slub_max_order);
3203
3204static int __init setup_slub_min_objects(char *str)
3205{
3206        get_option(&str, &slub_min_objects);
3207
3208        return 1;
3209}
3210
3211__setup("slub_min_objects=", setup_slub_min_objects);
3212
3213static int __init setup_slub_nomerge(char *str)
3214{
3215        slub_nomerge = 1;
3216        return 1;
3217}
3218
3219__setup("slub_nomerge", setup_slub_nomerge);
3220
3221void *__kmalloc(size_t size, gfp_t flags)
3222{
3223        struct kmem_cache *s;
3224        void *ret;
3225
3226        if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3227                return kmalloc_large(size, flags);
3228
3229        s = kmalloc_slab(size, flags);
3230
3231        if (unlikely(ZERO_OR_NULL_PTR(s)))
3232                return s;
3233
3234        ret = slab_alloc(s, flags, _RET_IP_);
3235
3236        trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3237
3238        return ret;
3239}
3240EXPORT_SYMBOL(__kmalloc);
3241
3242#ifdef CONFIG_NUMA
3243static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3244{
3245        struct page *page;
3246        void *ptr = NULL;
3247
3248        flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3249        page = alloc_pages_node(node, flags, get_order(size));
3250        if (page)
3251                ptr = page_address(page);
3252
3253        kmemleak_alloc(ptr, size, 1, flags);
3254        return ptr;
3255}
3256
3257void *__kmalloc_node(size_t size, gfp_t flags, int node)
3258{
3259        struct kmem_cache *s;
3260        void *ret;
3261
3262        if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3263                ret = kmalloc_large_node(size, flags, node);
3264
3265                trace_kmalloc_node(_RET_IP_, ret,
3266                                   size, PAGE_SIZE << get_order(size),
3267                                   flags, node);
3268
3269                return ret;
3270        }
3271
3272        s = kmalloc_slab(size, flags);
3273
3274        if (unlikely(ZERO_OR_NULL_PTR(s)))
3275                return s;
3276
3277        ret = slab_alloc_node(s, flags, node, _RET_IP_);
3278
3279        trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3280
3281        return ret;
3282}
3283EXPORT_SYMBOL(__kmalloc_node);
3284#endif
3285
3286size_t ksize(const void *object)
3287{
3288        struct page *page;
3289
3290        if (unlikely(object == ZERO_SIZE_PTR))
3291                return 0;
3292
3293        page = virt_to_head_page(object);
3294
3295        if (unlikely(!PageSlab(page))) {
3296                WARN_ON(!PageCompound(page));
3297                return PAGE_SIZE << compound_order(page);
3298        }
3299
3300        return slab_ksize(page->slab_cache);
3301}
3302EXPORT_SYMBOL(ksize);
3303
3304#ifdef CONFIG_SLUB_DEBUG
3305bool verify_mem_not_deleted(const void *x)
3306{
3307        struct page *page;
3308        void *object = (void *)x;
3309        unsigned long flags;
3310        bool rv;
3311
3312        if (unlikely(ZERO_OR_NULL_PTR(x)))
3313                return false;
3314
3315        local_irq_save(flags);
3316
3317        page = virt_to_head_page(x);
3318        if (unlikely(!PageSlab(page))) {
3319                /* maybe it was from stack? */
3320                rv = true;
3321                goto out_unlock;
3322        }
3323
3324        slab_lock(page);
3325        if (on_freelist(page->slab_cache, page, object)) {
3326                object_err(page->slab_cache, page, object, "Object is on free-list");
3327                rv = false;
3328        } else {
3329                rv = true;
3330        }
3331        slab_unlock(page);
3332
3333out_unlock:
3334        local_irq_restore(flags);
3335        return rv;
3336}
3337EXPORT_SYMBOL(verify_mem_not_deleted);
3338#endif
3339
3340void kfree(const void *x)
3341{
3342        struct page *page;
3343        void *object = (void *)x;
3344
3345        trace_kfree(_RET_IP_, x);
3346
3347        if (unlikely(ZERO_OR_NULL_PTR(x)))
3348                return;
3349
3350        page = virt_to_head_page(x);
3351        if (unlikely(!PageSlab(page))) {
3352                BUG_ON(!PageCompound(page));
3353                kmemleak_free(x);
3354                __free_memcg_kmem_pages(page, compound_order(page));
3355                return;
3356        }
3357        slab_free(page->slab_cache, page, object, _RET_IP_);
3358}
3359EXPORT_SYMBOL(kfree);
3360
3361/*
3362 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3363 * the remaining slabs by the number of items in use. The slabs with the
3364 * most items in use come first. New allocations will then fill those up
3365 * and thus they can be removed from the partial lists.
3366 *
3367 * The slabs with the least items are placed last. This results in them
3368 * being allocated from last increasing the chance that the last objects
3369 * are freed in them.
3370 */
3371int kmem_cache_shrink(struct kmem_cache *s)
3372{
3373        int node;
3374        int i;
3375        struct kmem_cache_node *n;
3376        struct page *page;
3377        struct page *t;
3378        int objects = oo_objects(s->max);
3379        struct list_head *slabs_by_inuse =
3380                kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3381        unsigned long flags;
3382
3383        if (!slabs_by_inuse)
3384                return -ENOMEM;
3385
3386        flush_all(s);
3387        for_each_node_state(node, N_NORMAL_MEMORY) {
3388                n = get_node(s, node);
3389
3390                if (!n->nr_partial)
3391                        continue;
3392
3393                for (i = 0; i < objects; i++)
3394                        INIT_LIST_HEAD(slabs_by_inuse + i);
3395
3396                spin_lock_irqsave(&n->list_lock, flags);
3397
3398                /*
3399                 * Build lists indexed by the items in use in each slab.
3400                 *
3401                 * Note that concurrent frees may occur while we hold the
3402                 * list_lock. page->inuse here is the upper limit.
3403                 */
3404                list_for_each_entry_safe(page, t, &n->partial, lru) {
3405                        list_move(&page->lru, slabs_by_inuse + page->inuse);
3406                        if (!page->inuse)
3407                                n->nr_partial--;
3408                }
3409
3410                /*
3411                 * Rebuild the partial list with the slabs filled up most
3412                 * first and the least used slabs at the end.
3413                 */
3414                for (i = objects - 1; i > 0; i--)
3415                        list_splice(slabs_by_inuse + i, n->partial.prev);
3416
3417                spin_unlock_irqrestore(&n->list_lock, flags);
3418
3419                /* Release empty slabs */
3420                list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3421                        discard_slab(s, page);
3422        }
3423
3424        kfree(slabs_by_inuse);
3425        return 0;
3426}
3427EXPORT_SYMBOL(kmem_cache_shrink);
3428
3429static int slab_mem_going_offline_callback(void *arg)
3430{
3431        struct kmem_cache *s;
3432
3433        mutex_lock(&slab_mutex);
3434        list_for_each_entry(s, &slab_caches, list)
3435                kmem_cache_shrink(s);
3436        mutex_unlock(&slab_mutex);
3437
3438        return 0;
3439}
3440
3441static void slab_mem_offline_callback(void *arg)
3442{
3443        struct kmem_cache_node *n;
3444        struct kmem_cache *s;
3445        struct memory_notify *marg = arg;
3446        int offline_node;
3447
3448        offline_node = marg->status_change_nid_normal;
3449
3450        /*
3451         * If the node still has available memory. we need kmem_cache_node
3452         * for it yet.
3453         */
3454        if (offline_node < 0)
3455                return;
3456
3457        mutex_lock(&slab_mutex);
3458        list_for_each_entry(s, &slab_caches, list) {
3459                n = get_node(s, offline_node);
3460                if (n) {
3461                        /*
3462                         * if n->nr_slabs > 0, slabs still exist on the node
3463                         * that is going down. We were unable to free them,
3464                         * and offline_pages() function shouldn't call this
3465                         * callback. So, we must fail.
3466                         */
3467                        BUG_ON(slabs_node(s, offline_node));
3468
3469                        s->node[offline_node] = NULL;
3470                        kmem_cache_free(kmem_cache_node, n);
3471                }
3472        }
3473        mutex_unlock(&slab_mutex);
3474}
3475
3476static int slab_mem_going_online_callback(void *arg)
3477{
3478        struct kmem_cache_node *n;
3479        struct kmem_cache *s;
3480        struct memory_notify *marg = arg;
3481        int nid = marg->status_change_nid_normal;
3482        int ret = 0;
3483
3484        /*
3485         * If the node's memory is already available, then kmem_cache_node is
3486         * already created. Nothing to do.
3487         */
3488        if (nid < 0)
3489                return 0;
3490
3491        /*
3492         * We are bringing a node online. No memory is available yet. We must
3493         * allocate a kmem_cache_node structure in order to bring the node
3494         * online.
3495         */
3496        mutex_lock(&slab_mutex);
3497        list_for_each_entry(s, &slab_caches, list) {
3498                /*
3499                 * XXX: kmem_cache_alloc_node will fallback to other nodes
3500                 *      since memory is not yet available from the node that
3501                 *      is brought up.
3502                 */
3503                n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3504                if (!n) {
3505                        ret = -ENOMEM;
3506                        goto out;
3507                }
3508                init_kmem_cache_node(n);
3509                s->node[nid] = n;
3510        }
3511out:
3512        mutex_unlock(&slab_mutex);
3513        return ret;
3514}
3515
3516static int slab_memory_callback(struct notifier_block *self,
3517                                unsigned long action, void *arg)
3518{
3519        int ret = 0;
3520
3521        switch (action) {
3522        case MEM_GOING_ONLINE:
3523                ret = slab_mem_going_online_callback(arg);
3524                break;
3525        case MEM_GOING_OFFLINE:
3526                ret = slab_mem_going_offline_callback(arg);
3527                break;
3528        case MEM_OFFLINE:
3529        case MEM_CANCEL_ONLINE:
3530                slab_mem_offline_callback(arg);
3531                break;
3532        case MEM_ONLINE:
3533        case MEM_CANCEL_OFFLINE:
3534                break;
3535        }
3536        if (ret)
3537                ret = notifier_from_errno(ret);
3538        else
3539                ret = NOTIFY_OK;
3540        return ret;
3541}
3542
3543static struct notifier_block slab_memory_callback_nb = {
3544        .notifier_call = slab_memory_callback,
3545        .priority = SLAB_CALLBACK_PRI,
3546};
3547
3548/********************************************************************
3549 *                      Basic setup of slabs
3550 *******************************************************************/
3551
3552/*
3553 * Used for early kmem_cache structures that were allocated using
3554 * the page allocator. Allocate them properly then fix up the pointers
3555 * that may be pointing to the wrong kmem_cache structure.
3556 */
3557
3558static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3559{
3560        int node;
3561        struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3562
3563        memcpy(s, static_cache, kmem_cache->object_size);
3564
3565        /*
3566         * This runs very early, and only the boot processor is supposed to be
3567         * up.  Even if it weren't true, IRQs are not up so we couldn't fire
3568         * IPIs around.
3569         */
3570        __flush_cpu_slab(s, smp_processor_id());
3571        for_each_node_state(node, N_NORMAL_MEMORY) {
3572                struct kmem_cache_node *n = get_node(s, node);
3573                struct page *p;
3574
3575                if (n) {
3576                        list_for_each_entry(p, &n->partial, lru)
3577                                p->slab_cache = s;
3578
3579#ifdef CONFIG_SLUB_DEBUG
3580                        list_for_each_entry(p, &n->full, lru)
3581                                p->slab_cache = s;
3582#endif
3583                }
3584        }
3585        list_add(&s->list, &slab_caches);
3586        return s;
3587}
3588
3589void __init kmem_cache_init(void)
3590{
3591        static __initdata struct kmem_cache boot_kmem_cache,
3592                boot_kmem_cache_node;
3593
3594        if (debug_guardpage_minorder())
3595                slub_max_order = 0;
3596
3597        kmem_cache_node = &boot_kmem_cache_node;
3598        kmem_cache = &boot_kmem_cache;
3599
3600        create_boot_cache(kmem_cache_node, "kmem_cache_node",
3601                sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3602
3603        register_hotmemory_notifier(&slab_memory_callback_nb);
3604
3605        /* Able to allocate the per node structures */
3606        slab_state = PARTIAL;
3607
3608        create_boot_cache(kmem_cache, "kmem_cache",
3609                        offsetof(struct kmem_cache, node) +
3610                                nr_node_ids * sizeof(struct kmem_cache_node *),
3611                       SLAB_HWCACHE_ALIGN);
3612
3613        kmem_cache = bootstrap(&boot_kmem_cache);
3614
3615        /*
3616         * Allocate kmem_cache_node properly from the kmem_cache slab.
3617         * kmem_cache_node is separately allocated so no need to
3618         * update any list pointers.
3619         */
3620        kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3621
3622        /* Now we can use the kmem_cache to allocate kmalloc slabs */
3623        create_kmalloc_caches(0);
3624
3625#ifdef CONFIG_SMP
3626        register_cpu_notifier(&slab_notifier);
3627#endif
3628
3629        printk(KERN_INFO
3630                "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3631                " CPUs=%d, Nodes=%d\n",
3632                cache_line_size(),
3633                slub_min_order, slub_max_order, slub_min_objects,
3634                nr_cpu_ids, nr_node_ids);
3635}
3636
3637void __init kmem_cache_init_late(void)
3638{
3639}
3640
3641/*
3642 * Find a mergeable slab cache
3643 */
3644static int slab_unmergeable(struct kmem_cache *s)
3645{
3646        if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3647                return 1;
3648
3649        if (s->ctor)
3650                return 1;
3651
3652        /*
3653         * We may have set a slab to be unmergeable during bootstrap.
3654         */
3655        if (s->refcount < 0)
3656                return 1;
3657
3658        return 0;
3659}
3660
3661static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
3662                size_t align, unsigned long flags, const char *name,
3663                void (*ctor)(void *))
3664{
3665        struct kmem_cache *s;
3666
3667        if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3668                return NULL;
3669
3670        if (ctor)
3671                return NULL;
3672
3673        size = ALIGN(size, sizeof(void *));
3674        align = calculate_alignment(flags, align, size);
3675        size = ALIGN(size, align);
3676        flags = kmem_cache_flags(size, flags, name, NULL);
3677
3678        list_for_each_entry(s, &slab_caches, list) {
3679                if (slab_unmergeable(s))
3680                        continue;
3681
3682                if (size > s->size)
3683                        continue;
3684
3685                if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3686                                continue;
3687                /*
3688                 * Check if alignment is compatible.
3689                 * Courtesy of Adrian Drzewiecki
3690                 */
3691                if ((s->size & ~(align - 1)) != s->size)
3692                        continue;
3693
3694                if (s->size - size >= sizeof(void *))
3695                        continue;
3696
3697                if (!cache_match_memcg(s, memcg))
3698                        continue;
3699
3700                return s;
3701        }
3702        return NULL;
3703}
3704
3705struct kmem_cache *
3706__kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3707                   size_t align, unsigned long flags, void (*ctor)(void *))
3708{
3709        struct kmem_cache *s;
3710
3711        s = find_mergeable(memcg, size, align, flags, name, ctor);
3712        if (s) {
3713                s->refcount++;
3714                /*
3715                 * Adjust the object sizes so that we clear
3716                 * the complete object on kzalloc.
3717                 */
3718                s->object_size = max(s->object_size, (int)size);
3719                s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3720
3721                if (sysfs_slab_alias(s, name)) {
3722                        s->refcount--;
3723                        s = NULL;
3724                }
3725        }
3726
3727        return s;
3728}
3729
3730int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3731{
3732        int err;
3733
3734        err = kmem_cache_open(s, flags);
3735        if (err)
3736                return err;
3737
3738        /* Mutex is not taken during early boot */
3739        if (slab_state <= UP)
3740                return 0;
3741
3742        memcg_propagate_slab_attrs(s);
3743        mutex_unlock(&slab_mutex);
3744        err = sysfs_slab_add(s);
3745        mutex_lock(&slab_mutex);
3746
3747        if (err)
3748                kmem_cache_close(s);
3749
3750        return err;
3751}
3752
3753#ifdef CONFIG_SMP
3754/*
3755 * Use the cpu notifier to insure that the cpu slabs are flushed when
3756 * necessary.
3757 */
3758static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3759                unsigned long action, void *hcpu)
3760{
3761        long cpu = (long)hcpu;
3762        struct kmem_cache *s;
3763        unsigned long flags;
3764
3765        switch (action) {
3766        case CPU_UP_CANCELED:
3767        case CPU_UP_CANCELED_FROZEN:
3768        case CPU_DEAD:
3769        case CPU_DEAD_FROZEN:
3770                mutex_lock(&slab_mutex);
3771                list_for_each_entry(s, &slab_caches, list) {
3772                        local_irq_save(flags);
3773                        __flush_cpu_slab(s, cpu);
3774                        local_irq_restore(flags);
3775                }
3776                mutex_unlock(&slab_mutex);
3777                break;
3778        default:
3779                break;
3780        }
3781        return NOTIFY_OK;
3782}
3783
3784static struct notifier_block __cpuinitdata slab_notifier = {
3785        .notifier_call = slab_cpuup_callback
3786};
3787
3788#endif
3789
3790void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3791{
3792        struct kmem_cache *s;
3793        void *ret;
3794
3795        if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3796                return kmalloc_large(size, gfpflags);
3797
3798        s = kmalloc_slab(size, gfpflags);
3799
3800        if (unlikely(ZERO_OR_NULL_PTR(s)))
3801                return s;
3802
3803        ret = slab_alloc(s, gfpflags, caller);
3804
3805        /* Honor the call site pointer we received. */
3806        trace_kmalloc(caller, ret, size, s->size, gfpflags);
3807
3808        return ret;
3809}
3810
3811#ifdef CONFIG_NUMA
3812void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3813                                        int node, unsigned long caller)
3814{
3815        struct kmem_cache *s;
3816        void *ret;
3817
3818        if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3819                ret = kmalloc_large_node(size, gfpflags, node);
3820
3821                trace_kmalloc_node(caller, ret,
3822                                   size, PAGE_SIZE << get_order(size),
3823                                   gfpflags, node);
3824
3825                return ret;
3826        }
3827
3828        s = kmalloc_slab(size, gfpflags);
3829
3830        if (unlikely(ZERO_OR_NULL_PTR(s)))
3831                return s;
3832
3833        ret = slab_alloc_node(s, gfpflags, node, caller);
3834
3835        /* Honor the call site pointer we received. */
3836        trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3837
3838        return ret;
3839}
3840#endif
3841
3842#ifdef CONFIG_SYSFS
3843static int count_inuse(struct page *page)
3844{
3845        return page->inuse;
3846}
3847
3848static int count_total(struct page *page)
3849{
3850        return page->objects;
3851}
3852#endif
3853
3854#ifdef CONFIG_SLUB_DEBUG
3855static int validate_slab(struct kmem_cache *s, struct page *page,
3856                                                unsigned long *map)
3857{
3858        void *p;
3859        void *addr = page_address(page);
3860
3861        if (!check_slab(s, page) ||
3862                        !on_freelist(s, page, NULL))
3863                return 0;
3864
3865        /* Now we know that a valid freelist exists */
3866        bitmap_zero(map, page->objects);
3867
3868        get_map(s, page, map);
3869        for_each_object(p, s, addr, page->objects) {
3870                if (test_bit(slab_index(p, s, addr), map))
3871                        if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3872                                return 0;
3873        }
3874
3875        for_each_object(p, s, addr, page->objects)
3876                if (!test_bit(slab_index(p, s, addr), map))
3877                        if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3878                                return 0;
3879        return 1;
3880}
3881
3882static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3883                                                unsigned long *map)
3884{
3885        slab_lock(page);
3886        validate_slab(s, page, map);
3887        slab_unlock(page);
3888}
3889
3890static int validate_slab_node(struct kmem_cache *s,
3891                struct kmem_cache_node *n, unsigned long *map)
3892{
3893        unsigned long count = 0;
3894        struct page *page;
3895        unsigned long flags;
3896
3897        spin_lock_irqsave(&n->list_lock, flags);
3898
3899        list_for_each_entry(page, &n->partial, lru) {
3900                validate_slab_slab(s, page, map);
3901                count++;
3902        }
3903        if (count != n->nr_partial)
3904                printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3905                        "counter=%ld\n", s->name, count, n->nr_partial);
3906
3907        if (!(s->flags & SLAB_STORE_USER))
3908                goto out;
3909
3910        list_for_each_entry(page, &n->full, lru) {
3911                validate_slab_slab(s, page, map);
3912                count++;
3913        }
3914        if (count != atomic_long_read(&n->nr_slabs))
3915                printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3916                        "counter=%ld\n", s->name, count,
3917                        atomic_long_read(&n->nr_slabs));
3918
3919out:
3920        spin_unlock_irqrestore(&n->list_lock, flags);
3921        return count;
3922}
3923
3924static long validate_slab_cache(struct kmem_cache *s)
3925{
3926        int node;
3927        unsigned long count = 0;
3928        unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3929                                sizeof(unsigned long), GFP_KERNEL);
3930
3931        if (!map)
3932                return -ENOMEM;
3933
3934        flush_all(s);
3935        for_each_node_state(node, N_NORMAL_MEMORY) {
3936                struct kmem_cache_node *n = get_node(s, node);
3937
3938                count += validate_slab_node(s, n, map);
3939        }
3940        kfree(map);
3941        return count;
3942}
3943/*
3944 * Generate lists of code addresses where slabcache objects are allocated
3945 * and freed.
3946 */
3947
3948struct location {
3949        unsigned long count;
3950        unsigned long addr;
3951        long long sum_time;
3952        long min_time;
3953        long max_time;
3954        long min_pid;
3955        long max_pid;
3956        DECLARE_BITMAP(cpus, NR_CPUS);
3957        nodemask_t nodes;
3958};
3959
3960struct loc_track {
3961        unsigned long max;
3962        unsigned long count;
3963        struct location *loc;
3964};
3965
3966static void free_loc_track(struct loc_track *t)
3967{
3968        if (t->max)
3969                free_pages((unsigned long)t->loc,
3970                        get_order(sizeof(struct location) * t->max));
3971}
3972
3973static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3974{
3975        struct location *l;
3976        int order;
3977
3978        order = get_order(sizeof(struct location) * max);
3979
3980        l = (void *)__get_free_pages(flags, order);
3981        if (!l)
3982                return 0;
3983
3984        if (t->count) {
3985                memcpy(l, t->loc, sizeof(struct location) * t->count);
3986                free_loc_track(t);
3987        }
3988        t->max = max;
3989        t->loc = l;
3990        return 1;
3991}
3992
3993static int add_location(struct loc_track *t, struct kmem_cache *s,
3994                                const struct track *track)
3995{
3996        long start, end, pos;
3997        struct location *l;
3998        unsigned long caddr;
3999        unsigned long age = jiffies - track->when;
4000
4001        start = -1;
4002        end = t->count;
4003
4004        for ( ; ; ) {
4005                pos = start + (end - start + 1) / 2;
4006
4007                /*
4008                 * There is nothing at "end". If we end up there
4009                 * we need to add something to before end.
4010                 */
4011                if (pos == end)
4012                        break;
4013
4014                caddr = t->loc[pos].addr;
4015                if (track->addr == caddr) {
4016
4017                        l = &t->loc[pos];
4018                        l->count++;
4019                        if (track->when) {
4020                                l->sum_time += age;
4021                                if (age < l->min_time)
4022                                        l->min_time = age;
4023                                if (age > l->max_time)
4024                                        l->max_time = age;
4025
4026                                if (track->pid < l->min_pid)
4027                                        l->min_pid = track->pid;
4028                                if (track->pid > l->max_pid)
4029                                        l->max_pid = track->pid;
4030
4031                                cpumask_set_cpu(track->cpu,
4032                                                to_cpumask(l->cpus));
4033                        }
4034                        node_set(page_to_nid(virt_to_page(track)), l->nodes);
4035                        return 1;
4036                }
4037
4038                if (track->addr < caddr)
4039                        end = pos;
4040                else
4041                        start = pos;
4042        }
4043
4044        /*
4045         * Not found. Insert new tracking element.
4046         */
4047        if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4048                return 0;
4049
4050        l = t->loc + pos;
4051        if (pos < t->count)
4052                memmove(l + 1, l,
4053                        (t->count - pos) * sizeof(struct location));
4054        t->count++;
4055        l->count = 1;
4056        l->addr = track->addr;
4057        l->sum_time = age;
4058        l->min_time = age;
4059        l->max_time = age;
4060        l->min_pid = track->pid;
4061        l->max_pid = track->pid;
4062        cpumask_clear(to_cpumask(l->cpus));
4063        cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4064        nodes_clear(l->nodes);
4065        node_set(page_to_nid(virt_to_page(track)), l->nodes);
4066        return 1;
4067}
4068
4069static void process_slab(struct loc_track *t, struct kmem_cache *s,
4070                struct page *page, enum track_item alloc,
4071                unsigned long *map)
4072{
4073        void *addr = page_address(page);
4074        void *p;
4075
4076        bitmap_zero(map, page->objects);
4077        get_map(s, page, map);
4078
4079        for_each_object(p, s, addr, page->objects)
4080                if (!test_bit(slab_index(p, s, addr), map))
4081                        add_location(t, s, get_track(s, p, alloc));
4082}
4083
4084static int list_locations(struct kmem_cache *s, char *buf,
4085                                        enum track_item alloc)
4086{
4087        int len = 0;
4088        unsigned long i;
4089        struct loc_track t = { 0, 0, NULL };
4090        int node;
4091        unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4092                                     sizeof(unsigned long), GFP_KERNEL);
4093
4094        if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4095                                     GFP_TEMPORARY)) {
4096                kfree(map);
4097                return sprintf(buf, "Out of memory\n");
4098        }
4099        /* Push back cpu slabs */
4100        flush_all(s);
4101
4102        for_each_node_state(node, N_NORMAL_MEMORY) {
4103                struct kmem_cache_node *n = get_node(s, node);
4104                unsigned long flags;
4105                struct page *page;
4106
4107                if (!atomic_long_read(&n->nr_slabs))
4108                        continue;
4109
4110                spin_lock_irqsave(&n->list_lock, flags);
4111                list_for_each_entry(page, &n->partial, lru)
4112                        process_slab(&t, s, page, alloc, map);
4113                list_for_each_entry(page, &n->full, lru)
4114                        process_slab(&t, s, page, alloc, map);
4115                spin_unlock_irqrestore(&n->list_lock, flags);
4116        }
4117
4118        for (i = 0; i < t.count; i++) {
4119                struct location *l = &t.loc[i];
4120
4121                if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4122                        break;
4123                len += sprintf(buf + len, "%7ld ", l->count);
4124
4125                if (l->addr)
4126                        len += sprintf(buf + len, "%pS", (void *)l->addr);
4127                else
4128                        len += sprintf(buf + len, "<not-available>");
4129
4130                if (l->sum_time != l->min_time) {
4131                        len += sprintf(buf + len, " age=%ld/%ld/%ld",
4132                                l->min_time,
4133                                (long)div_u64(l->sum_time, l->count),
4134                                l->max_time);
4135                } else
4136                        len += sprintf(buf + len, " age=%ld",
4137                                l->min_time);
4138
4139                if (l->min_pid != l->max_pid)
4140                        len += sprintf(buf + len, " pid=%ld-%ld",
4141                                l->min_pid, l->max_pid);
4142                else
4143                        len += sprintf(buf + len, " pid=%ld",
4144                                l->min_pid);
4145
4146                if (num_online_cpus() > 1 &&
4147                                !cpumask_empty(to_cpumask(l->cpus)) &&
4148                                len < PAGE_SIZE - 60) {
4149                        len += sprintf(buf + len, " cpus=");
4150                        len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4151                                                 to_cpumask(l->cpus));
4152                }
4153
4154                if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4155                                len < PAGE_SIZE - 60) {
4156                        len += sprintf(buf + len, " nodes=");
4157                        len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4158                                        l->nodes);
4159                }
4160
4161                len += sprintf(buf + len, "\n");
4162        }
4163
4164        free_loc_track(&t);
4165        kfree(map);
4166        if (!t.count)
4167                len += sprintf(buf, "No data\n");
4168        return len;
4169}
4170#endif
4171
4172#ifdef SLUB_RESILIENCY_TEST
4173static void resiliency_test(void)
4174{
4175        u8 *p;
4176
4177        BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4178
4179        printk(KERN_ERR "SLUB resiliency testing\n");
4180        printk(KERN_ERR "-----------------------\n");
4181        printk(KERN_ERR "A. Corruption after allocation\n");
4182
4183        p = kzalloc(16, GFP_KERNEL);
4184        p[16] = 0x12;
4185        printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4186                        " 0x12->0x%p\n\n", p + 16);
4187
4188        validate_slab_cache(kmalloc_caches[4]);
4189
4190        /* Hmmm... The next two are dangerous */
4191        p = kzalloc(32, GFP_KERNEL);
4192        p[32 + sizeof(void *)] = 0x34;
4193        printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4194                        " 0x34 -> -0x%p\n", p);
4195        printk(KERN_ERR
4196                "If allocated object is overwritten then not detectable\n\n");
4197
4198        validate_slab_cache(kmalloc_caches[5]);
4199        p = kzalloc(64, GFP_KERNEL);
4200        p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4201        *p = 0x56;
4202        printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4203                                                                        p);
4204        printk(KERN_ERR
4205                "If allocated object is overwritten then not detectable\n\n");
4206        validate_slab_cache(kmalloc_caches[6]);
4207
4208        printk(KERN_ERR "\nB. Corruption after free\n");
4209        p = kzalloc(128, GFP_KERNEL);
4210        kfree(p);
4211        *p = 0x78;
4212        printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4213        validate_slab_cache(kmalloc_caches[7]);
4214
4215        p = kzalloc(256, GFP_KERNEL);
4216        kfree(p);
4217        p[50] = 0x9a;
4218        printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4219                        p);
4220        validate_slab_cache(kmalloc_caches[8]);
4221
4222        p = kzalloc(512, GFP_KERNEL);
4223        kfree(p);
4224        p[512] = 0xab;
4225        printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4226        validate_slab_cache(kmalloc_caches[9]);
4227}
4228#else
4229#ifdef CONFIG_SYSFS
4230static void resiliency_test(void) {};
4231#endif
4232#endif
4233
4234#ifdef CONFIG_SYSFS
4235enum slab_stat_type {
4236        SL_ALL,                 /* All slabs */
4237        SL_PARTIAL,             /* Only partially allocated slabs */
4238        SL_CPU,                 /* Only slabs used for cpu caches */
4239        SL_OBJECTS,             /* Determine allocated objects not slabs */
4240        SL_TOTAL                /* Determine object capacity not slabs */
4241};
4242
4243#define SO_ALL          (1 << SL_ALL)
4244#define SO_PARTIAL      (1 << SL_PARTIAL)
4245#define SO_CPU          (1 << SL_CPU)
4246#define SO_OBJECTS      (1 << SL_OBJECTS)
4247#define SO_TOTAL        (1 << SL_TOTAL)
4248
4249static ssize_t show_slab_objects(struct kmem_cache *s,
4250                            char *buf, unsigned long flags)
4251{
4252        unsigned long total = 0;
4253        int node;
4254        int x;
4255        unsigned long *nodes;
4256        unsigned long *per_cpu;
4257
4258        nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4259        if (!nodes)
4260                return -ENOMEM;
4261        per_cpu = nodes + nr_node_ids;
4262
4263        if (flags & SO_CPU) {
4264                int cpu;
4265
4266                for_each_possible_cpu(cpu) {
4267                        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4268                        int node;
4269                        struct page *page;
4270
4271                        page = ACCESS_ONCE(c->page);
4272                        if (!page)
4273                                continue;
4274
4275                        node = page_to_nid(page);
4276                        if (flags & SO_TOTAL)
4277                                x = page->objects;
4278                        else if (flags & SO_OBJECTS)
4279                                x = page->inuse;
4280                        else
4281                                x = 1;
4282
4283                        total += x;
4284                        nodes[node] += x;
4285
4286                        page = ACCESS_ONCE(c->partial);
4287                        if (page) {
4288                                x = page->pobjects;
4289                                total += x;
4290                                nodes[node] += x;
4291                        }
4292
4293                        per_cpu[node]++;
4294                }
4295        }
4296
4297        lock_memory_hotplug();
4298#ifdef CONFIG_SLUB_DEBUG
4299        if (flags & SO_ALL) {
4300                for_each_node_state(node, N_NORMAL_MEMORY) {
4301                        struct kmem_cache_node *n = get_node(s, node);
4302
4303                if (flags & SO_TOTAL)
4304                        x = atomic_long_read(&n->total_objects);
4305                else if (flags & SO_OBJECTS)
4306                        x = atomic_long_read(&n->total_objects) -
4307                                count_partial(n, count_free);
4308
4309                        else
4310                                x = atomic_long_read(&n->nr_slabs);
4311                        total += x;
4312                        nodes[node] += x;
4313                }
4314
4315        } else
4316#endif
4317        if (flags & SO_PARTIAL) {
4318                for_each_node_state(node, N_NORMAL_MEMORY) {
4319                        struct kmem_cache_node *n = get_node(s, node);
4320
4321                        if (flags & SO_TOTAL)
4322                                x = count_partial(n, count_total);
4323                        else if (flags & SO_OBJECTS)
4324                                x = count_partial(n, count_inuse);
4325                        else
4326                                x = n->nr_partial;
4327                        total += x;
4328                        nodes[node] += x;
4329                }
4330        }
4331        x = sprintf(buf, "%lu", total);
4332#ifdef CONFIG_NUMA
4333        for_each_node_state(node, N_NORMAL_MEMORY)
4334                if (nodes[node])
4335                        x += sprintf(buf + x, " N%d=%lu",
4336                                        node, nodes[node]);
4337#endif
4338        unlock_memory_hotplug();
4339        kfree(nodes);
4340        return x + sprintf(buf + x, "\n");
4341}
4342
4343#ifdef CONFIG_SLUB_DEBUG
4344static int any_slab_objects(struct kmem_cache *s)
4345{
4346        int node;
4347
4348        for_each_online_node(node) {
4349                struct kmem_cache_node *n = get_node(s, node);
4350
4351                if (!n)
4352                        continue;
4353
4354                if (atomic_long_read(&n->total_objects))
4355                        return 1;
4356        }
4357        return 0;
4358}
4359#endif
4360
4361#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4362#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4363
4364struct slab_attribute {
4365        struct attribute attr;
4366        ssize_t (*show)(struct kmem_cache *s, char *buf);
4367        ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4368};
4369
4370#define SLAB_ATTR_RO(_name) \
4371        static struct slab_attribute _name##_attr = \
4372        __ATTR(_name, 0400, _name##_show, NULL)
4373
4374#define SLAB_ATTR(_name) \
4375        static struct slab_attribute _name##_attr =  \
4376        __ATTR(_name, 0600, _name##_show, _name##_store)
4377
4378static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4379{
4380        return sprintf(buf, "%d\n", s->size);
4381}
4382SLAB_ATTR_RO(slab_size);
4383
4384static ssize_t align_show(struct kmem_cache *s, char *buf)
4385{
4386        return sprintf(buf, "%d\n", s->align);
4387}
4388SLAB_ATTR_RO(align);
4389
4390static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4391{
4392        return sprintf(buf, "%d\n", s->object_size);
4393}
4394SLAB_ATTR_RO(object_size);
4395
4396static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4397{
4398        return sprintf(buf, "%d\n", oo_objects(s->oo));
4399}
4400SLAB_ATTR_RO(objs_per_slab);
4401
4402static ssize_t order_store(struct kmem_cache *s,
4403                                const char *buf, size_t length)
4404{
4405        unsigned long order;
4406        int err;
4407
4408        err = strict_strtoul(buf, 10, &order);
4409        if (err)
4410                return err;
4411
4412        if (order > slub_max_order || order < slub_min_order)
4413                return -EINVAL;
4414
4415        calculate_sizes(s, order);
4416        return length;
4417}
4418
4419static ssize_t order_show(struct kmem_cache *s, char *buf)
4420{
4421        return sprintf(buf, "%d\n", oo_order(s->oo));
4422}
4423SLAB_ATTR(order);
4424
4425static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4426{
4427        return sprintf(buf, "%lu\n", s->min_partial);
4428}
4429
4430static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4431                                 size_t length)
4432{
4433        unsigned long min;
4434        int err;
4435
4436        err = strict_strtoul(buf, 10, &min);
4437        if (err)
4438                return err;
4439
4440        set_min_partial(s, min);
4441        return length;
4442}
4443SLAB_ATTR(min_partial);
4444
4445static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4446{
4447        return sprintf(buf, "%u\n", s->cpu_partial);
4448}
4449
4450static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4451                                 size_t length)
4452{
4453        unsigned long objects;
4454        int err;
4455
4456        err = strict_strtoul(buf, 10, &objects);
4457        if (err)
4458                return err;
4459        if (objects && kmem_cache_debug(s))
4460                return -EINVAL;
4461
4462        s->cpu_partial = objects;
4463        flush_all(s);
4464        return length;
4465}
4466SLAB_ATTR(cpu_partial);
4467
4468static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4469{
4470        if (!s->ctor)
4471                return 0;
4472        return sprintf(buf, "%pS\n", s->ctor);
4473}
4474SLAB_ATTR_RO(ctor);
4475
4476static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4477{
4478        return sprintf(buf, "%d\n", s->refcount - 1);
4479}
4480SLAB_ATTR_RO(aliases);
4481
4482static ssize_t partial_show(struct kmem_cache *s, char *buf)
4483{
4484        return show_slab_objects(s, buf, SO_PARTIAL);
4485}
4486SLAB_ATTR_RO(partial);
4487
4488static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4489{
4490        return show_slab_objects(s, buf, SO_CPU);
4491}
4492SLAB_ATTR_RO(cpu_slabs);
4493
4494static ssize_t objects_show(struct kmem_cache *s, char *buf)
4495{
4496        return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4497}
4498SLAB_ATTR_RO(objects);
4499
4500static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4501{
4502        return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4503}
4504SLAB_ATTR_RO(objects_partial);
4505
4506static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4507{
4508        int objects = 0;
4509        int pages = 0;
4510        int cpu;
4511        int len;
4512
4513        for_each_online_cpu(cpu) {
4514                struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4515
4516                if (page) {
4517                        pages += page->pages;
4518                        objects += page->pobjects;
4519                }
4520        }
4521
4522        len = sprintf(buf, "%d(%d)", objects, pages);
4523
4524#ifdef CONFIG_SMP
4525        for_each_online_cpu(cpu) {
4526                struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4527
4528                if (page && len < PAGE_SIZE - 20)
4529                        len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4530                                page->pobjects, page->pages);
4531        }
4532#endif
4533        return len + sprintf(buf + len, "\n");
4534}
4535SLAB_ATTR_RO(slabs_cpu_partial);
4536
4537static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4538{
4539        return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4540}
4541
4542static ssize_t reclaim_account_store(struct kmem_cache *s,
4543                                const char *buf, size_t length)
4544{
4545        s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4546        if (buf[0] == '1')
4547                s->flags |= SLAB_RECLAIM_ACCOUNT;
4548        return length;
4549}
4550SLAB_ATTR(reclaim_account);
4551
4552static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4553{
4554        return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4555}
4556SLAB_ATTR_RO(hwcache_align);
4557
4558#ifdef CONFIG_ZONE_DMA
4559static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4560{
4561        return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4562}
4563SLAB_ATTR_RO(cache_dma);
4564#endif
4565
4566static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4567{
4568        return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4569}
4570SLAB_ATTR_RO(destroy_by_rcu);
4571
4572static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4573{
4574        return sprintf(buf, "%d\n", s->reserved);
4575}
4576SLAB_ATTR_RO(reserved);
4577
4578#ifdef CONFIG_SLUB_DEBUG
4579static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4580{
4581        return show_slab_objects(s, buf, SO_ALL);
4582}
4583SLAB_ATTR_RO(slabs);
4584
4585static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4586{
4587        return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4588}
4589SLAB_ATTR_RO(total_objects);
4590
4591static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4592{
4593        return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4594}
4595
4596static ssize_t sanity_checks_store(struct kmem_cache *s,
4597                                const char *buf, size_t length)
4598{
4599        s->flags &= ~SLAB_DEBUG_FREE;
4600        if (buf[0] == '1') {
4601                s->flags &= ~__CMPXCHG_DOUBLE;
4602                s->flags |= SLAB_DEBUG_FREE;
4603        }
4604        return length;
4605}
4606SLAB_ATTR(sanity_checks);
4607
4608static ssize_t trace_show(struct kmem_cache *s, char *buf)
4609{
4610        return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4611}
4612
4613static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4614                                                        size_t length)
4615{
4616        s->flags &= ~SLAB_TRACE;
4617        if (buf[0] == '1') {
4618                s->flags &= ~__CMPXCHG_DOUBLE;
4619                s->flags |= SLAB_TRACE;
4620        }
4621        return length;
4622}
4623SLAB_ATTR(trace);
4624
4625static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4626{
4627        return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4628}
4629
4630static ssize_t red_zone_store(struct kmem_cache *s,
4631                                const char *buf, size_t length)
4632{
4633        if (any_slab_objects(s))
4634                return -EBUSY;
4635
4636        s->flags &= ~SLAB_RED_ZONE;
4637        if (buf[0] == '1') {
4638                s->flags &= ~__CMPXCHG_DOUBLE;
4639                s->flags |= SLAB_RED_ZONE;
4640        }
4641        calculate_sizes(s, -1);
4642        return length;
4643}
4644SLAB_ATTR(red_zone);
4645
4646static ssize_t poison_show(struct kmem_cache *s, char *buf)
4647{
4648        return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4649}
4650
4651static ssize_t poison_store(struct kmem_cache *s,
4652                                const char *buf, size_t length)
4653{
4654        if (any_slab_objects(s))
4655                return -EBUSY;
4656
4657        s->flags &= ~SLAB_POISON;
4658        if (buf[0] == '1') {
4659                s->flags &= ~__CMPXCHG_DOUBLE;
4660                s->flags |= SLAB_POISON;
4661        }
4662        calculate_sizes(s, -1);
4663        return length;
4664}
4665SLAB_ATTR(poison);
4666
4667static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4668{
4669        return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4670}
4671
4672static ssize_t store_user_store(struct kmem_cache *s,
4673                                const char *buf, size_t length)
4674{
4675        if (any_slab_objects(s))
4676                return -EBUSY;
4677
4678        s->flags &= ~SLAB_STORE_USER;
4679        if (buf[0] == '1') {
4680                s->flags &= ~__CMPXCHG_DOUBLE;
4681                s->flags |= SLAB_STORE_USER;
4682        }
4683        calculate_sizes(s, -1);
4684        return length;
4685}
4686SLAB_ATTR(store_user);
4687
4688static ssize_t validate_show(struct kmem_cache *s, char *buf)
4689{
4690        return 0;
4691}
4692
4693static ssize_t validate_store(struct kmem_cache *s,
4694                        const char *buf, size_t length)
4695{
4696        int ret = -EINVAL;
4697
4698        if (buf[0] == '1') {
4699                ret = validate_slab_cache(s);
4700                if (ret >= 0)
4701                        ret = length;
4702        }
4703        return ret;
4704}
4705SLAB_ATTR(validate);
4706
4707static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4708{
4709        if (!(s->flags & SLAB_STORE_USER))
4710                return -ENOSYS;
4711        return list_locations(s, buf, TRACK_ALLOC);
4712}
4713SLAB_ATTR_RO(alloc_calls);
4714
4715static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4716{
4717        if (!(s->flags & SLAB_STORE_USER))
4718                return -ENOSYS;
4719        return list_locations(s, buf, TRACK_FREE);
4720}
4721SLAB_ATTR_RO(free_calls);
4722#endif /* CONFIG_SLUB_DEBUG */
4723
4724#ifdef CONFIG_FAILSLAB
4725static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4726{
4727        return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4728}
4729
4730static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4731                                                        size_t length)
4732{
4733        s->flags &= ~SLAB_FAILSLAB;
4734        if (buf[0] == '1')
4735                s->flags |= SLAB_FAILSLAB;
4736        return length;
4737}
4738SLAB_ATTR(failslab);
4739#endif
4740
4741static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4742{
4743        return 0;
4744}
4745
4746static ssize_t shrink_store(struct kmem_cache *s,
4747                        const char *buf, size_t length)
4748{
4749        if (buf[0] == '1') {
4750                int rc = kmem_cache_shrink(s);
4751
4752                if (rc)
4753                        return rc;
4754        } else
4755                return -EINVAL;
4756        return length;
4757}
4758SLAB_ATTR(shrink);
4759
4760#ifdef CONFIG_NUMA
4761static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4762{
4763        return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4764}
4765
4766static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4767                                const char *buf, size_t length)
4768{
4769        unsigned long ratio;
4770        int err;
4771
4772        err = strict_strtoul(buf, 10, &ratio);
4773        if (err)
4774                return err;
4775
4776        if (ratio <= 100)
4777                s->remote_node_defrag_ratio = ratio * 10;
4778
4779        return length;
4780}
4781SLAB_ATTR(remote_node_defrag_ratio);
4782#endif
4783
4784#ifdef CONFIG_SLUB_STATS
4785static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4786{
4787        unsigned long sum  = 0;
4788        int cpu;
4789        int len;
4790        int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4791
4792        if (!data)
4793                return -ENOMEM;
4794
4795        for_each_online_cpu(cpu) {
4796                unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4797
4798                data[cpu] = x;
4799                sum += x;
4800        }
4801
4802        len = sprintf(buf, "%lu", sum);
4803
4804#ifdef CONFIG_SMP
4805        for_each_online_cpu(cpu) {
4806                if (data[cpu] && len < PAGE_SIZE - 20)
4807                        len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4808        }
4809#endif
4810        kfree(data);
4811        return len + sprintf(buf + len, "\n");
4812}
4813
4814static void clear_stat(struct kmem_cache *s, enum stat_item si)
4815{
4816        int cpu;
4817
4818        for_each_online_cpu(cpu)
4819                per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4820}
4821
4822#define STAT_ATTR(si, text)                                     \
4823static ssize_t text##_show(struct kmem_cache *s, char *buf)     \
4824{                                                               \
4825        return show_stat(s, buf, si);                           \
4826}                                                               \
4827static ssize_t text##_store(struct kmem_cache *s,               \
4828                                const char *buf, size_t length) \
4829{                                                               \
4830        if (buf[0] != '0')                                      \
4831                return -EINVAL;                                 \
4832        clear_stat(s, si);                                      \
4833        return length;                                          \
4834}                                                               \
4835SLAB_ATTR(text);                                                \
4836
4837STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4838STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4839STAT_ATTR(FREE_FASTPATH, free_fastpath);
4840STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4841STAT_ATTR(FREE_FROZEN, free_frozen);
4842STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4843STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4844STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4845STAT_ATTR(ALLOC_SLAB, alloc_slab);
4846STAT_ATTR(ALLOC_REFILL, alloc_refill);
4847STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4848STAT_ATTR(FREE_SLAB, free_slab);
4849STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4850STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4851STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4852STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4853STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4854STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4855STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4856STAT_ATTR(ORDER_FALLBACK, order_fallback);
4857STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4858STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4859STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4860STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4861STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4862STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4863#endif
4864
4865static struct attribute *slab_attrs[] = {
4866        &slab_size_attr.attr,
4867        &object_size_attr.attr,
4868        &objs_per_slab_attr.attr,
4869        &order_attr.attr,
4870        &min_partial_attr.attr,
4871        &cpu_partial_attr.attr,
4872        &objects_attr.attr,
4873        &objects_partial_attr.attr,
4874        &partial_attr.attr,
4875        &cpu_slabs_attr.attr,
4876        &ctor_attr.attr,
4877        &aliases_attr.attr,
4878        &align_attr.attr,
4879        &hwcache_align_attr.attr,
4880        &reclaim_account_attr.attr,
4881        &destroy_by_rcu_attr.attr,
4882        &shrink_attr.attr,
4883        &reserved_attr.attr,
4884        &slabs_cpu_partial_attr.attr,
4885#ifdef CONFIG_SLUB_DEBUG
4886        &total_objects_attr.attr,
4887        &slabs_attr.attr,
4888        &sanity_checks_attr.attr,
4889        &trace_attr.attr,
4890        &red_zone_attr.attr,
4891        &poison_attr.attr,
4892        &store_user_attr.attr,
4893        &validate_attr.attr,
4894        &alloc_calls_attr.attr,
4895        &free_calls_attr.attr,
4896#endif
4897#ifdef CONFIG_ZONE_DMA
4898        &cache_dma_attr.attr,
4899#endif
4900#ifdef CONFIG_NUMA
4901        &remote_node_defrag_ratio_attr.attr,
4902#endif
4903#ifdef CONFIG_SLUB_STATS
4904        &alloc_fastpath_attr.attr,
4905        &alloc_slowpath_attr.attr,
4906        &free_fastpath_attr.attr,
4907        &free_slowpath_attr.attr,
4908        &free_frozen_attr.attr,
4909        &free_add_partial_attr.attr,
4910        &free_remove_partial_attr.attr,
4911        &alloc_from_partial_attr.attr,
4912        &alloc_slab_attr.attr,
4913        &alloc_refill_attr.attr,
4914        &alloc_node_mismatch_attr.attr,
4915        &free_slab_attr.attr,
4916        &cpuslab_flush_attr.attr,
4917        &deactivate_full_attr.attr,
4918        &deactivate_empty_attr.attr,
4919        &deactivate_to_head_attr.attr,
4920        &deactivate_to_tail_attr.attr,
4921        &deactivate_remote_frees_attr.attr,
4922        &deactivate_bypass_attr.attr,
4923        &order_fallback_attr.attr,
4924        &cmpxchg_double_fail_attr.attr,
4925        &cmpxchg_double_cpu_fail_attr.attr,
4926        &cpu_partial_alloc_attr.attr,
4927        &cpu_partial_free_attr.attr,
4928        &cpu_partial_node_attr.attr,
4929        &cpu_partial_drain_attr.attr,
4930#endif
4931#ifdef CONFIG_FAILSLAB
4932        &failslab_attr.attr,
4933#endif
4934
4935        NULL
4936};
4937
4938static struct attribute_group slab_attr_group = {
4939        .attrs = slab_attrs,
4940};
4941
4942static ssize_t slab_attr_show(struct kobject *kobj,
4943                                struct attribute *attr,
4944                                char *buf)
4945{
4946        struct slab_attribute *attribute;
4947        struct kmem_cache *s;
4948        int err;
4949
4950        attribute = to_slab_attr(attr);
4951        s = to_slab(kobj);
4952
4953        if (!attribute->show)
4954                return -EIO;
4955
4956        err = attribute->show(s, buf);
4957
4958        return err;
4959}
4960
4961static ssize_t slab_attr_store(struct kobject *kobj,
4962                                struct attribute *attr,
4963                                const char *buf, size_t len)
4964{
4965        struct slab_attribute *attribute;
4966        struct kmem_cache *s;
4967        int err;
4968
4969        attribute = to_slab_attr(attr);
4970        s = to_slab(kobj);
4971
4972        if (!attribute->store)
4973                return -EIO;
4974
4975        err = attribute->store(s, buf, len);
4976#ifdef CONFIG_MEMCG_KMEM
4977        if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4978                int i;
4979
4980                mutex_lock(&slab_mutex);
4981                if (s->max_attr_size < len)
4982                        s->max_attr_size = len;
4983
4984                /*
4985                 * This is a best effort propagation, so this function's return
4986                 * value will be determined by the parent cache only. This is
4987                 * basically because not all attributes will have a well
4988                 * defined semantics for rollbacks - most of the actions will
4989                 * have permanent effects.
4990                 *
4991                 * Returning the error value of any of the children that fail
4992                 * is not 100 % defined, in the sense that users seeing the
4993                 * error code won't be able to know anything about the state of
4994                 * the cache.
4995                 *
4996                 * Only returning the error code for the parent cache at least
4997                 * has well defined semantics. The cache being written to
4998                 * directly either failed or succeeded, in which case we loop
4999                 * through the descendants with best-effort propagation.
5000                 */
5001                for_each_memcg_cache_index(i) {
5002                        struct kmem_cache *c = cache_from_memcg(s, i);
5003                        if (c)
5004                                attribute->store(c, buf, len);
5005                }
5006                mutex_unlock(&slab_mutex);
5007        }
5008#endif
5009        return err;
5010}
5011
5012static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5013{
5014#ifdef CONFIG_MEMCG_KMEM
5015        int i;
5016        char *buffer = NULL;
5017
5018        if (!is_root_cache(s))
5019                return;
5020
5021        /*
5022         * This mean this cache had no attribute written. Therefore, no point
5023         * in copying default values around
5024         */
5025        if (!s->max_attr_size)
5026                return;
5027
5028        for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5029                char mbuf[64];
5030                char *buf;
5031                struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5032
5033                if (!attr || !attr->store || !attr->show)
5034                        continue;
5035
5036                /*
5037                 * It is really bad that we have to allocate here, so we will
5038                 * do it only as a fallback. If we actually allocate, though,
5039                 * we can just use the allocated buffer until the end.
5040                 *
5041                 * Most of the slub attributes will tend to be very small in
5042                 * size, but sysfs allows buffers up to a page, so they can
5043                 * theoretically happen.
5044                 */
5045                if (buffer)
5046                        buf = buffer;
5047                else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5048                        buf = mbuf;
5049                else {
5050                        buffer = (char *) get_zeroed_page(GFP_KERNEL);
5051                        if (WARN_ON(!buffer))
5052                                continue;
5053                        buf = buffer;
5054                }
5055
5056                attr->show(s->memcg_params->root_cache, buf);
5057                attr->store(s, buf, strlen(buf));
5058        }
5059
5060        if (buffer)
5061                free_page((unsigned long)buffer);
5062#endif
5063}
5064
5065static const struct sysfs_ops slab_sysfs_ops = {
5066        .show = slab_attr_show,
5067        .store = slab_attr_store,
5068};
5069
5070static struct kobj_type slab_ktype = {
5071        .sysfs_ops = &slab_sysfs_ops,
5072};
5073
5074static int uevent_filter(struct kset *kset, struct kobject *kobj)
5075{
5076        struct kobj_type *ktype = get_ktype(kobj);
5077
5078        if (ktype == &slab_ktype)
5079                return 1;
5080        return 0;
5081}
5082
5083static const struct kset_uevent_ops slab_uevent_ops = {
5084        .filter = uevent_filter,
5085};
5086
5087static struct kset *slab_kset;
5088
5089#define ID_STR_LENGTH 64
5090
5091/* Create a unique string id for a slab cache:
5092 *
5093 * Format       :[flags-]size
5094 */
5095static char *create_unique_id(struct kmem_cache *s)
5096{
5097        char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5098        char *p = name;
5099
5100        BUG_ON(!name);
5101
5102        *p++ = ':';
5103        /*
5104         * First flags affecting slabcache operations. We will only
5105         * get here for aliasable slabs so we do not need to support
5106         * too many flags. The flags here must cover all flags that
5107         * are matched during merging to guarantee that the id is
5108         * unique.
5109         */
5110        if (s->flags & SLAB_CACHE_DMA)
5111                *p++ = 'd';
5112        if (s->flags & SLAB_RECLAIM_ACCOUNT)
5113                *p++ = 'a';
5114        if (s->flags & SLAB_DEBUG_FREE)
5115                *p++ = 'F';
5116        if (!(s->flags & SLAB_NOTRACK))
5117                *p++ = 't';
5118        if (p != name + 1)
5119                *p++ = '-';
5120        p += sprintf(p, "%07d", s->size);
5121
5122#ifdef CONFIG_MEMCG_KMEM
5123        if (!is_root_cache(s))
5124                p += sprintf(p, "-%08d", memcg_cache_id(s->memcg_params->memcg));
5125#endif
5126
5127        BUG_ON(p > name + ID_STR_LENGTH - 1);
5128        return name;
5129}
5130
5131static int sysfs_slab_add(struct kmem_cache *s)
5132{
5133        int err;
5134        const char *name;
5135        int unmergeable = slab_unmergeable(s);
5136
5137        if (unmergeable) {
5138                /*
5139                 * Slabcache can never be merged so we can use the name proper.
5140                 * This is typically the case for debug situations. In that
5141                 * case we can catch duplicate names easily.
5142                 */
5143                sysfs_remove_link(&slab_kset->kobj, s->name);
5144                name = s->name;
5145        } else {
5146                /*
5147                 * Create a unique name for the slab as a target
5148                 * for the symlinks.
5149                 */
5150                name = create_unique_id(s);
5151        }
5152
5153        s->kobj.kset = slab_kset;
5154        err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5155        if (err) {
5156                kobject_put(&s->kobj);
5157                return err;
5158        }
5159
5160        err = sysfs_create_group(&s->kobj, &slab_attr_group);
5161        if (err) {
5162                kobject_del(&s->kobj);
5163                kobject_put(&s->kobj);
5164                return err;
5165        }
5166        kobject_uevent(&s->kobj, KOBJ_ADD);
5167        if (!unmergeable) {
5168                /* Setup first alias */
5169                sysfs_slab_alias(s, s->name);
5170                kfree(name);
5171        }
5172        return 0;
5173}
5174
5175static void sysfs_slab_remove(struct kmem_cache *s)
5176{
5177        if (slab_state < FULL)
5178                /*
5179                 * Sysfs has not been setup yet so no need to remove the
5180                 * cache from sysfs.
5181                 */
5182                return;
5183
5184        kobject_uevent(&s->kobj, KOBJ_REMOVE);
5185        kobject_del(&s->kobj);
5186        kobject_put(&s->kobj);
5187}
5188
5189/*
5190 * Need to buffer aliases during bootup until sysfs becomes
5191 * available lest we lose that information.
5192 */
5193struct saved_alias {
5194        struct kmem_cache *s;
5195        const char *name;
5196        struct saved_alias *next;
5197};
5198
5199static struct saved_alias *alias_list;
5200
5201static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5202{
5203        struct saved_alias *al;
5204
5205        if (slab_state == FULL) {
5206                /*
5207                 * If we have a leftover link then remove it.
5208                 */
5209                sysfs_remove_link(&slab_kset->kobj, name);
5210                return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5211        }
5212
5213        al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5214        if (!al)
5215                return -ENOMEM;
5216
5217        al->s = s;
5218        al->name = name;
5219        al->next = alias_list;
5220        alias_list = al;
5221        return 0;
5222}
5223
5224static int __init slab_sysfs_init(void)
5225{
5226        struct kmem_cache *s;
5227        int err;
5228
5229        mutex_lock(&slab_mutex);
5230
5231        slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5232        if (!slab_kset) {
5233                mutex_unlock(&slab_mutex);
5234                printk(KERN_ERR "Cannot register slab subsystem.\n");
5235                return -ENOSYS;
5236        }
5237
5238        slab_state = FULL;
5239
5240        list_for_each_entry(s, &slab_caches, list) {
5241                err = sysfs_slab_add(s);
5242                if (err)
5243                        printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5244                                                " to sysfs\n", s->name);
5245        }
5246
5247        while (alias_list) {
5248                struct saved_alias *al = alias_list;
5249
5250                alias_list = alias_list->next;
5251                err = sysfs_slab_alias(al->s, al->name);
5252                if (err)
5253                        printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5254                                        " %s to sysfs\n", al->name);
5255                kfree(al);
5256        }
5257
5258        mutex_unlock(&slab_mutex);
5259        resiliency_test();
5260        return 0;
5261}
5262
5263__initcall(slab_sysfs_init);
5264#endif /* CONFIG_SYSFS */
5265
5266/*
5267 * The /proc/slabinfo ABI
5268 */
5269#ifdef CONFIG_SLABINFO
5270void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5271{
5272        unsigned long nr_partials = 0;
5273        unsigned long nr_slabs = 0;
5274        unsigned long nr_objs = 0;
5275        unsigned long nr_free = 0;
5276        int node;
5277
5278        for_each_online_node(node) {
5279                struct kmem_cache_node *n = get_node(s, node);
5280
5281                if (!n)
5282                        continue;
5283
5284                nr_partials += n->nr_partial;
5285                nr_slabs += atomic_long_read(&n->nr_slabs);
5286                nr_objs += atomic_long_read(&n->total_objects);
5287                nr_free += count_partial(n, count_free);
5288        }
5289
5290        sinfo->active_objs = nr_objs - nr_free;
5291        sinfo->num_objs = nr_objs;
5292        sinfo->active_slabs = nr_slabs;
5293        sinfo->num_slabs = nr_slabs;
5294        sinfo->objects_per_slab = oo_objects(s->oo);
5295        sinfo->cache_order = oo_order(s->oo);
5296}
5297
5298void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5299{
5300}
5301
5302ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5303                       size_t count, loff_t *ppos)
5304{
5305        return -EIO;
5306}
5307#endif /* CONFIG_SLABINFO */
5308
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.