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
 126static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
 127{
 128#ifdef CONFIG_SLUB_CPU_PARTIAL
 129        return !kmem_cache_debug(s);
 130#else
 131        return false;
 132#endif
 133}
 134
 135/*
 136 * Issues still to be resolved:
 137 *
 138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
 139 *
 140 * - Variable sizing of the per node arrays
 141 */
 142
 143/* Enable to test recovery from slab corruption on boot */
 144#undef SLUB_RESILIENCY_TEST
 145
 146/* Enable to log cmpxchg failures */
 147#undef SLUB_DEBUG_CMPXCHG
 148
 149/*
 150 * Mininum number of partial slabs. These will be left on the partial
 151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
 152 */
 153#define MIN_PARTIAL 5
 154
 155/*
 156 * Maximum number of desirable partial slabs.
 157 * The existence of more partial slabs makes kmem_cache_shrink
 158 * sort the partial list by the number of objects in the.
 159 */
 160#define MAX_PARTIAL 10
 161
 162#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
 163                                SLAB_POISON | SLAB_STORE_USER)
 164
 165/*
 166 * Debugging flags that require metadata to be stored in the slab.  These get
 167 * disabled when slub_debug=O is used and a cache's min order increases with
 168 * metadata.
 169 */
 170#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
 171
 172/*
 173 * Set of flags that will prevent slab merging
 174 */
 175#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
 176                SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
 177                SLAB_FAILSLAB)
 178
 179#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
 180                SLAB_CACHE_DMA | SLAB_NOTRACK)
 181
 182#define OO_SHIFT        16
 183#define OO_MASK         ((1 << OO_SHIFT) - 1)
 184#define MAX_OBJS_PER_PAGE       32767 /* since page.objects is u15 */
 185
 186/* Internal SLUB flags */
 187#define __OBJECT_POISON         0x80000000UL /* Poison object */
 188#define __CMPXCHG_DOUBLE        0x40000000UL /* Use cmpxchg_double */
 189
 190#ifdef CONFIG_SMP
 191static struct notifier_block slab_notifier;
 192#endif
 193
 194/*
 195 * Tracking user of a slab.
 196 */
 197#define TRACK_ADDRS_COUNT 16
 198struct track {
 199        unsigned long addr;     /* Called from address */
 200#ifdef CONFIG_STACKTRACE
 201        unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
 202#endif
 203        int cpu;                /* Was running on cpu */
 204        int pid;                /* Pid context */
 205        unsigned long when;     /* When did the operation occur */
 206};
 207
 208enum track_item { TRACK_ALLOC, TRACK_FREE };
 209
 210#ifdef CONFIG_SYSFS
 211static int sysfs_slab_add(struct kmem_cache *);
 212static int sysfs_slab_alias(struct kmem_cache *, const char *);
 213static void sysfs_slab_remove(struct kmem_cache *);
 214static void memcg_propagate_slab_attrs(struct kmem_cache *s);
 215#else
 216static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
 217static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
 218                                                        { return 0; }
 219static inline void sysfs_slab_remove(struct kmem_cache *s) { }
 220
 221static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
 222#endif
 223
 224static inline void stat(const struct kmem_cache *s, enum stat_item si)
 225{
 226#ifdef CONFIG_SLUB_STATS
 227        __this_cpu_inc(s->cpu_slab->stat[si]);
 228#endif
 229}
 230
 231/********************************************************************
 232 *                      Core slab cache functions
 233 *******************************************************************/
 234
 235static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
 236{
 237        return s->node[node];
 238}
 239
 240/* Verify that a pointer has an address that is valid within a slab page */
 241static inline int check_valid_pointer(struct kmem_cache *s,
 242                                struct page *page, const void *object)
 243{
 244        void *base;
 245
 246        if (!object)
 247                return 1;
 248
 249        base = page_address(page);
 250        if (object < base || object >= base + page->objects * s->size ||
 251                (object - base) % s->size) {
 252                return 0;
 253        }
 254
 255        return 1;
 256}
 257
 258static inline void *get_freepointer(struct kmem_cache *s, void *object)
 259{
 260        return *(void **)(object + s->offset);
 261}
 262
 263static void prefetch_freepointer(const struct kmem_cache *s, void *object)
 264{
 265        prefetch(object + s->offset);
 266}
 267
 268static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
 269{
 270        void *p;
 271
 272#ifdef CONFIG_DEBUG_PAGEALLOC
 273        probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
 274#else
 275        p = get_freepointer(s, object);
 276#endif
 277        return p;
 278}
 279
 280static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
 281{
 282        *(void **)(object + s->offset) = fp;
 283}
 284
 285/* Loop over all objects in a slab */
 286#define for_each_object(__p, __s, __addr, __objects) \
 287        for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
 288                        __p += (__s)->size)
 289
 290/* Determine object index from a given position */
 291static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
 292{
 293        return (p - addr) / s->size;
 294}
 295
 296static inline size_t slab_ksize(const struct kmem_cache *s)
 297{
 298#ifdef CONFIG_SLUB_DEBUG
 299        /*
 300         * Debugging requires use of the padding between object
 301         * and whatever may come after it.
 302         */
 303        if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
 304                return s->object_size;
 305
 306#endif
 307        /*
 308         * If we have the need to store the freelist pointer
 309         * back there or track user information then we can
 310         * only use the space before that information.
 311         */
 312        if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
 313                return s->inuse;
 314        /*
 315         * Else we can use all the padding etc for the allocation
 316         */
 317        return s->size;
 318}
 319
 320static inline int order_objects(int order, unsigned long size, int reserved)
 321{
 322        return ((PAGE_SIZE << order) - reserved) / size;
 323}
 324
 325static inline struct kmem_cache_order_objects oo_make(int order,
 326                unsigned long size, int reserved)
 327{
 328        struct kmem_cache_order_objects x = {
 329                (order << OO_SHIFT) + order_objects(order, size, reserved)
 330        };
 331
 332        return x;
 333}
 334
 335static inline int oo_order(struct kmem_cache_order_objects x)
 336{
 337        return x.x >> OO_SHIFT;
 338}
 339
 340static inline int oo_objects(struct kmem_cache_order_objects x)
 341{
 342        return x.x & OO_MASK;
 343}
 344
 345/*
 346 * Per slab locking using the pagelock
 347 */
 348static __always_inline void slab_lock(struct page *page)
 349{
 350        bit_spin_lock(PG_locked, &page->flags);
 351}
 352
 353static __always_inline void slab_unlock(struct page *page)
 354{
 355        __bit_spin_unlock(PG_locked, &page->flags);
 356}
 357
 358/* Interrupts must be disabled (for the fallback code to work right) */
 359static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
 360                void *freelist_old, unsigned long counters_old,
 361                void *freelist_new, unsigned long counters_new,
 362                const char *n)
 363{
 364        VM_BUG_ON(!irqs_disabled());
 365#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
 366    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
 367        if (s->flags & __CMPXCHG_DOUBLE) {
 368                if (cmpxchg_double(&page->freelist, &page->counters,
 369                        freelist_old, counters_old,
 370                        freelist_new, counters_new))
 371                return 1;
 372        } else
 373#endif
 374        {
 375                slab_lock(page);
 376                if (page->freelist == freelist_old && page->counters == counters_old) {
 377                        page->freelist = freelist_new;
 378                        page->counters = counters_new;
 379                        slab_unlock(page);
 380                        return 1;
 381                }
 382                slab_unlock(page);
 383        }
 384
 385        cpu_relax();
 386        stat(s, CMPXCHG_DOUBLE_FAIL);
 387
 388#ifdef SLUB_DEBUG_CMPXCHG
 389        printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
 390#endif
 391
 392        return 0;
 393}
 394
 395static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
 396                void *freelist_old, unsigned long counters_old,
 397                void *freelist_new, unsigned long counters_new,
 398                const char *n)
 399{
 400#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
 401    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
 402        if (s->flags & __CMPXCHG_DOUBLE) {
 403                if (cmpxchg_double(&page->freelist, &page->counters,
 404                        freelist_old, counters_old,
 405                        freelist_new, counters_new))
 406                return 1;
 407        } else
 408#endif
 409        {
 410                unsigned long flags;
 411
 412                local_irq_save(flags);
 413                slab_lock(page);
 414                if (page->freelist == freelist_old && page->counters == counters_old) {
 415                        page->freelist = freelist_new;
 416                        page->counters = counters_new;
 417                        slab_unlock(page);
 418                        local_irq_restore(flags);
 419                        return 1;
 420                }
 421                slab_unlock(page);
 422                local_irq_restore(flags);
 423        }
 424
 425        cpu_relax();
 426        stat(s, CMPXCHG_DOUBLE_FAIL);
 427
 428#ifdef SLUB_DEBUG_CMPXCHG
 429        printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
 430#endif
 431
 432        return 0;
 433}
 434
 435#ifdef CONFIG_SLUB_DEBUG
 436/*
 437 * Determine a map of object in use on a page.
 438 *
 439 * Node listlock must be held to guarantee that the page does
 440 * not vanish from under us.
 441 */
 442static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
 443{
 444        void *p;
 445        void *addr = page_address(page);
 446
 447        for (p = page->freelist; p; p = get_freepointer(s, p))
 448                set_bit(slab_index(p, s, addr), map);
 449}
 450
 451/*
 452 * Debug settings:
 453 */
 454#ifdef CONFIG_SLUB_DEBUG_ON
 455static int slub_debug = DEBUG_DEFAULT_FLAGS;
 456#else
 457static int slub_debug;
 458#endif
 459
 460static char *slub_debug_slabs;
 461static int disable_higher_order_debug;
 462
 463/*
 464 * Object debugging
 465 */
 466static void print_section(char *text, u8 *addr, unsigned int length)
 467{
 468        print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
 469                        length, 1);
 470}
 471
 472static struct track *get_track(struct kmem_cache *s, void *object,
 473        enum track_item alloc)
 474{
 475        struct track *p;
 476
 477        if (s->offset)
 478                p = object + s->offset + sizeof(void *);
 479        else
 480                p = object + s->inuse;
 481
 482        return p + alloc;
 483}
 484
 485static void set_track(struct kmem_cache *s, void *object,
 486                        enum track_item alloc, unsigned long addr)
 487{
 488        struct track *p = get_track(s, object, alloc);
 489
 490        if (addr) {
 491#ifdef CONFIG_STACKTRACE
 492                struct stack_trace trace;
 493                int i;
 494
 495                trace.nr_entries = 0;
 496                trace.max_entries = TRACK_ADDRS_COUNT;
 497                trace.entries = p->addrs;
 498                trace.skip = 3;
 499                save_stack_trace(&trace);
 500
 501                /* See rant in lockdep.c */
 502                if (trace.nr_entries != 0 &&
 503                    trace.entries[trace.nr_entries - 1] == ULONG_MAX)
 504                        trace.nr_entries--;
 505
 506                for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
 507                        p->addrs[i] = 0;
 508#endif
 509                p->addr = addr;
 510                p->cpu = smp_processor_id();
 511                p->pid = current->pid;
 512                p->when = jiffies;
 513        } else
 514                memset(p, 0, sizeof(struct track));
 515}
 516
 517static void init_tracking(struct kmem_cache *s, void *object)
 518{
 519        if (!(s->flags & SLAB_STORE_USER))
 520                return;
 521
 522        set_track(s, object, TRACK_FREE, 0UL);
 523        set_track(s, object, TRACK_ALLOC, 0UL);
 524}
 525
 526static void print_track(const char *s, struct track *t)
 527{
 528        if (!t->addr)
 529                return;
 530
 531        printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
 532                s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
 533#ifdef CONFIG_STACKTRACE
 534        {
 535                int i;
 536                for (i = 0; i < TRACK_ADDRS_COUNT; i++)
 537                        if (t->addrs[i])
 538                                printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
 539                        else
 540                                break;
 541        }
 542#endif
 543}
 544
 545static void print_tracking(struct kmem_cache *s, void *object)
 546{
 547        if (!(s->flags & SLAB_STORE_USER))
 548                return;
 549
 550        print_track("Allocated", get_track(s, object, TRACK_ALLOC));
 551        print_track("Freed", get_track(s, object, TRACK_FREE));
 552}
 553
 554static void print_page_info(struct page *page)
 555{
 556        printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
 557                page, page->objects, page->inuse, page->freelist, page->flags);
 558
 559}
 560
 561static void slab_bug(struct kmem_cache *s, char *fmt, ...)
 562{
 563        va_list args;
 564        char buf[100];
 565
 566        va_start(args, fmt);
 567        vsnprintf(buf, sizeof(buf), fmt, args);
 568        va_end(args);
 569        printk(KERN_ERR "========================================"
 570                        "=====================================\n");
 571        printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
 572        printk(KERN_ERR "----------------------------------------"
 573                        "-------------------------------------\n\n");
 574
 575        add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 576}
 577
 578static void slab_fix(struct kmem_cache *s, char *fmt, ...)
 579{
 580        va_list args;
 581        char buf[100];
 582
 583        va_start(args, fmt);
 584        vsnprintf(buf, sizeof(buf), fmt, args);
 585        va_end(args);
 586        printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
 587}
 588
 589static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
 590{
 591        unsigned int off;       /* Offset of last byte */
 592        u8 *addr = page_address(page);
 593
 594        print_tracking(s, p);
 595
 596        print_page_info(page);
 597
 598        printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
 599                        p, p - addr, get_freepointer(s, p));
 600
 601        if (p > addr + 16)
 602                print_section("Bytes b4 ", p - 16, 16);
 603
 604        print_section("Object ", p, min_t(unsigned long, s->object_size,
 605                                PAGE_SIZE));
 606        if (s->flags & SLAB_RED_ZONE)
 607                print_section("Redzone ", p + s->object_size,
 608                        s->inuse - s->object_size);
 609
 610        if (s->offset)
 611                off = s->offset + sizeof(void *);
 612        else
 613                off = s->inuse;
 614
 615        if (s->flags & SLAB_STORE_USER)
 616                off += 2 * sizeof(struct track);
 617
 618        if (off != s->size)
 619                /* Beginning of the filler is the free pointer */
 620                print_section("Padding ", p + off, s->size - off);
 621
 622        dump_stack();
 623}
 624
 625static void object_err(struct kmem_cache *s, struct page *page,
 626                        u8 *object, char *reason)
 627{
 628        slab_bug(s, "%s", reason);
 629        print_trailer(s, page, object);
 630}
 631
 632static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
 633{
 634        va_list args;
 635        char buf[100];
 636
 637        va_start(args, fmt);
 638        vsnprintf(buf, sizeof(buf), fmt, args);
 639        va_end(args);
 640        slab_bug(s, "%s", buf);
 641        print_page_info(page);
 642        dump_stack();
 643}
 644
 645static void init_object(struct kmem_cache *s, void *object, u8 val)
 646{
 647        u8 *p = object;
 648
 649        if (s->flags & __OBJECT_POISON) {
 650                memset(p, POISON_FREE, s->object_size - 1);
 651                p[s->object_size - 1] = POISON_END;
 652        }
 653
 654        if (s->flags & SLAB_RED_ZONE)
 655                memset(p + s->object_size, val, s->inuse - s->object_size);
 656}
 657
 658static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
 659                                                void *from, void *to)
 660{
 661        slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
 662        memset(from, data, to - from);
 663}
 664
 665static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
 666                        u8 *object, char *what,
 667                        u8 *start, unsigned int value, unsigned int bytes)
 668{
 669        u8 *fault;
 670        u8 *end;
 671
 672        fault = memchr_inv(start, value, bytes);
 673        if (!fault)
 674                return 1;
 675
 676        end = start + bytes;
 677        while (end > fault && end[-1] == value)
 678                end--;
 679
 680        slab_bug(s, "%s overwritten", what);
 681        printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
 682                                        fault, end - 1, fault[0], value);
 683        print_trailer(s, page, object);
 684
 685        restore_bytes(s, what, value, fault, end);
 686        return 0;
 687}
 688
 689/*
 690 * Object layout:
 691 *
 692 * object address
 693 *      Bytes of the object to be managed.
 694 *      If the freepointer may overlay the object then the free
 695 *      pointer is the first word of the object.
 696 *
 697 *      Poisoning uses 0x6b (POISON_FREE) and the last byte is
 698 *      0xa5 (POISON_END)
 699 *
 700 * object + s->object_size
 701 *      Padding to reach word boundary. This is also used for Redzoning.
 702 *      Padding is extended by another word if Redzoning is enabled and
 703 *      object_size == inuse.
 704 *
 705 *      We fill with 0xbb (RED_INACTIVE) for inactive objects and with
 706 *      0xcc (RED_ACTIVE) for objects in use.
 707 *
 708 * object + s->inuse
 709 *      Meta data starts here.
 710 *
 711 *      A. Free pointer (if we cannot overwrite object on free)
 712 *      B. Tracking data for SLAB_STORE_USER
 713 *      C. Padding to reach required alignment boundary or at mininum
 714 *              one word if debugging is on to be able to detect writes
 715 *              before the word boundary.
 716 *
 717 *      Padding is done using 0x5a (POISON_INUSE)
 718 *
 719 * object + s->size
 720 *      Nothing is used beyond s->size.
 721 *
 722 * If slabcaches are merged then the object_size and inuse boundaries are mostly
 723 * ignored. And therefore no slab options that rely on these boundaries
 724 * may be used with merged slabcaches.
 725 */
 726
 727static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
 728{
 729        unsigned long off = s->inuse;   /* The end of info */
 730
 731        if (s->offset)
 732                /* Freepointer is placed after the object. */
 733                off += sizeof(void *);
 734
 735        if (s->flags & SLAB_STORE_USER)
 736                /* We also have user information there */
 737                off += 2 * sizeof(struct track);
 738
 739        if (s->size == off)
 740                return 1;
 741
 742        return check_bytes_and_report(s, page, p, "Object padding",
 743                                p + off, POISON_INUSE, s->size - off);
 744}
 745
 746/* Check the pad bytes at the end of a slab page */
 747static int slab_pad_check(struct kmem_cache *s, struct page *page)
 748{
 749        u8 *start;
 750        u8 *fault;
 751        u8 *end;
 752        int length;
 753        int remainder;
 754
 755        if (!(s->flags & SLAB_POISON))
 756                return 1;
 757
 758        start = page_address(page);
 759        length = (PAGE_SIZE << compound_order(page)) - s->reserved;
 760        end = start + length;
 761        remainder = length % s->size;
 762        if (!remainder)
 763                return 1;
 764
 765        fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
 766        if (!fault)
 767                return 1;
 768        while (end > fault && end[-1] == POISON_INUSE)
 769                end--;
 770
 771        slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
 772        print_section("Padding ", end - remainder, remainder);
 773
 774        restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
 775        return 0;
 776}
 777
 778static int check_object(struct kmem_cache *s, struct page *page,
 779                                        void *object, u8 val)
 780{
 781        u8 *p = object;
 782        u8 *endobject = object + s->object_size;
 783
 784        if (s->flags & SLAB_RED_ZONE) {
 785                if (!check_bytes_and_report(s, page, object, "Redzone",
 786                        endobject, val, s->inuse - s->object_size))
 787                        return 0;
 788        } else {
 789                if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
 790                        check_bytes_and_report(s, page, p, "Alignment padding",
 791                                endobject, POISON_INUSE, s->inuse - s->object_size);
 792                }
 793        }
 794
 795        if (s->flags & SLAB_POISON) {
 796                if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
 797                        (!check_bytes_and_report(s, page, p, "Poison", p,
 798                                        POISON_FREE, s->object_size - 1) ||
 799                         !check_bytes_and_report(s, page, p, "Poison",
 800                                p + s->object_size - 1, POISON_END, 1)))
 801                        return 0;
 802                /*
 803                 * check_pad_bytes cleans up on its own.
 804                 */
 805                check_pad_bytes(s, page, p);
 806        }
 807
 808        if (!s->offset && val == SLUB_RED_ACTIVE)
 809                /*
 810                 * Object and freepointer overlap. Cannot check
 811                 * freepointer while object is allocated.
 812                 */
 813                return 1;
 814
 815        /* Check free pointer validity */
 816        if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
 817                object_err(s, page, p, "Freepointer corrupt");
 818                /*
 819                 * No choice but to zap it and thus lose the remainder
 820                 * of the free objects in this slab. May cause
 821                 * another error because the object count is now wrong.
 822                 */
 823                set_freepointer(s, p, NULL);
 824                return 0;
 825        }
 826        return 1;
 827}
 828
 829static int check_slab(struct kmem_cache *s, struct page *page)
 830{
 831        int maxobj;
 832
 833        VM_BUG_ON(!irqs_disabled());
 834
 835        if (!PageSlab(page)) {
 836                slab_err(s, page, "Not a valid slab page");
 837                return 0;
 838        }
 839
 840        maxobj = order_objects(compound_order(page), s->size, s->reserved);
 841        if (page->objects > maxobj) {
 842                slab_err(s, page, "objects %u > max %u",
 843                        s->name, page->objects, maxobj);
 844                return 0;
 845        }
 846        if (page->inuse > page->objects) {
 847                slab_err(s, page, "inuse %u > max %u",
 848                        s->name, page->inuse, page->objects);
 849                return 0;
 850        }
 851        /* Slab_pad_check fixes things up after itself */
 852        slab_pad_check(s, page);
 853        return 1;
 854}
 855
 856/*
 857 * Determine if a certain object on a page is on the freelist. Must hold the
 858 * slab lock to guarantee that the chains are in a consistent state.
 859 */
 860static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
 861{
 862        int nr = 0;
 863        void *fp;
 864        void *object = NULL;
 865        unsigned long max_objects;
 866
 867        fp = page->freelist;
 868        while (fp && nr <= page->objects) {
 869                if (fp == search)
 870                        return 1;
 871                if (!check_valid_pointer(s, page, fp)) {
 872                        if (object) {
 873                                object_err(s, page, object,
 874                                        "Freechain corrupt");
 875                                set_freepointer(s, object, NULL);
 876                                break;
 877                        } else {
 878                                slab_err(s, page, "Freepointer corrupt");
 879                                page->freelist = NULL;
 880                                page->inuse = page->objects;
 881                                slab_fix(s, "Freelist cleared");
 882                                return 0;
 883                        }
 884                        break;
 885                }
 886                object = fp;
 887                fp = get_freepointer(s, object);
 888                nr++;
 889        }
 890
 891        max_objects = order_objects(compound_order(page), s->size, s->reserved);
 892        if (max_objects > MAX_OBJS_PER_PAGE)
 893                max_objects = MAX_OBJS_PER_PAGE;
 894
 895        if (page->objects != max_objects) {
 896                slab_err(s, page, "Wrong number of objects. Found %d but "
 897                        "should be %d", page->objects, max_objects);
 898                page->objects = max_objects;
 899                slab_fix(s, "Number of objects adjusted.");
 900        }
 901        if (page->inuse != page->objects - nr) {
 902                slab_err(s, page, "Wrong object count. Counter is %d but "
 903                        "counted were %d", page->inuse, page->objects - nr);
 904                page->inuse = page->objects - nr;
 905                slab_fix(s, "Object count adjusted.");
 906        }
 907        return search == NULL;
 908}
 909
 910static void trace(struct kmem_cache *s, struct page *page, void *object,
 911                                                                int alloc)
 912{
 913        if (s->flags & SLAB_TRACE) {
 914                printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
 915                        s->name,
 916                        alloc ? "alloc" : "free",
 917                        object, page->inuse,
 918                        page->freelist);
 919
 920                if (!alloc)
 921                        print_section("Object ", (void *)object, s->object_size);
 922
 923                dump_stack();
 924        }
 925}
 926
 927/*
 928 * Hooks for other subsystems that check memory allocations. In a typical
 929 * production configuration these hooks all should produce no code at all.
 930 */
 931static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
 932{
 933        flags &= gfp_allowed_mask;
 934        lockdep_trace_alloc(flags);
 935        might_sleep_if(flags & __GFP_WAIT);
 936
 937        return should_failslab(s->object_size, flags, s->flags);
 938}
 939
 940static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
 941{
 942        flags &= gfp_allowed_mask;
 943        kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
 944        kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
 945}
 946
 947static inline void slab_free_hook(struct kmem_cache *s, void *x)
 948{
 949        kmemleak_free_recursive(x, s->flags);
 950
 951        /*
 952         * Trouble is that we may no longer disable interupts in the fast path
 953         * So in order to make the debug calls that expect irqs to be
 954         * disabled we need to disable interrupts temporarily.
 955         */
 956#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
 957        {
 958                unsigned long flags;
 959
 960                local_irq_save(flags);
 961                kmemcheck_slab_free(s, x, s->object_size);
 962                debug_check_no_locks_freed(x, s->object_size);
 963                local_irq_restore(flags);
 964        }
 965#endif
 966        if (!(s->flags & SLAB_DEBUG_OBJECTS))
 967                debug_check_no_obj_freed(x, s->object_size);
 968}
 969
 970/*
 971 * Tracking of fully allocated slabs for debugging purposes.
 972 *
 973 * list_lock must be held.
 974 */
 975static void add_full(struct kmem_cache *s,
 976        struct kmem_cache_node *n, struct page *page)
 977{
 978        if (!(s->flags & SLAB_STORE_USER))
 979                return;
 980
 981        list_add(&page->lru, &n->full);
 982}
 983
 984/*
 985 * list_lock must be held.
 986 */
 987static void remove_full(struct kmem_cache *s, struct page *page)
 988{
 989        if (!(s->flags & SLAB_STORE_USER))
 990                return;
 991
 992        list_del(&page->lru);
 993}
 994
 995/* Tracking of the number of slabs for debugging purposes */
 996static inline unsigned long slabs_node(struct kmem_cache *s, int node)
 997{
 998        struct kmem_cache_node *n = get_node(s, node);
 999
1000        return atomic_long_read(&n->nr_slabs);
1001}
1002
1003static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1004{
1005        return atomic_long_read(&n->nr_slabs);
1006}
1007
1008static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1009{
1010        struct kmem_cache_node *n = get_node(s, node);
1011
1012        /*
1013         * May be called early in order to allocate a slab for the
1014         * kmem_cache_node structure. Solve the chicken-egg
1015         * dilemma by deferring the increment of the count during
1016         * bootstrap (see early_kmem_cache_node_alloc).
1017         */
1018        if (likely(n)) {
1019                atomic_long_inc(&n->nr_slabs);
1020                atomic_long_add(objects, &n->total_objects);
1021        }
1022}
1023static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1024{
1025        struct kmem_cache_node *n = get_node(s, node);
1026
1027        atomic_long_dec(&n->nr_slabs);
1028        atomic_long_sub(objects, &n->total_objects);
1029}
1030
1031/* Object debug checks for alloc/free paths */
1032static void setup_object_debug(struct kmem_cache *s, struct page *page,
1033                                                                void *object)
1034{
1035        if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1036                return;
1037
1038        init_object(s, object, SLUB_RED_INACTIVE);
1039        init_tracking(s, object);
1040}
1041
1042static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1043                                        void *object, unsigned long addr)
1044{
1045        if (!check_slab(s, page))
1046                goto bad;
1047
1048        if (!check_valid_pointer(s, page, object)) {
1049                object_err(s, page, object, "Freelist Pointer check fails");
1050                goto bad;
1051        }
1052
1053        if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1054                goto bad;
1055
1056        /* Success perform special debug activities for allocs */
1057        if (s->flags & SLAB_STORE_USER)
1058                set_track(s, object, TRACK_ALLOC, addr);
1059        trace(s, page, object, 1);
1060        init_object(s, object, SLUB_RED_ACTIVE);
1061        return 1;
1062
1063bad:
1064        if (PageSlab(page)) {
1065                /*
1066                 * If this is a slab page then lets do the best we can
1067                 * to avoid issues in the future. Marking all objects
1068                 * as used avoids touching the remaining objects.
1069                 */
1070                slab_fix(s, "Marking all objects used");
1071                page->inuse = page->objects;
1072                page->freelist = NULL;
1073        }
1074        return 0;
1075}
1076
1077static noinline struct kmem_cache_node *free_debug_processing(
1078        struct kmem_cache *s, struct page *page, void *object,
1079        unsigned long addr, unsigned long *flags)
1080{
1081        struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1082
1083        spin_lock_irqsave(&n->list_lock, *flags);
1084        slab_lock(page);
1085
1086        if (!check_slab(s, page))
1087                goto fail;
1088
1089        if (!check_valid_pointer(s, page, object)) {
1090                slab_err(s, page, "Invalid object pointer 0x%p", object);
1091                goto fail;
1092        }
1093
1094        if (on_freelist(s, page, object)) {
1095                object_err(s, page, object, "Object already free");
1096                goto fail;
1097        }
1098
1099        if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1100                goto out;
1101
1102        if (unlikely(s != page->slab_cache)) {
1103                if (!PageSlab(page)) {
1104                        slab_err(s, page, "Attempt to free object(0x%p) "
1105                                "outside of slab", object);
1106                } else if (!page->slab_cache) {
1107                        printk(KERN_ERR
1108                                "SLUB <none>: no slab for object 0x%p.\n",
1109                                                object);
1110                        dump_stack();
1111                } else
1112                        object_err(s, page, object,
1113                                        "page slab pointer corrupt.");
1114                goto fail;
1115        }
1116
1117        if (s->flags & SLAB_STORE_USER)
1118                set_track(s, object, TRACK_FREE, addr);
1119        trace(s, page, object, 0);
1120        init_object(s, object, SLUB_RED_INACTIVE);
1121out:
1122        slab_unlock(page);
1123        /*
1124         * Keep node_lock to preserve integrity
1125         * until the object is actually freed
1126         */
1127        return n;
1128
1129fail:
1130        slab_unlock(page);
1131        spin_unlock_irqrestore(&n->list_lock, *flags);
1132        slab_fix(s, "Object at 0x%p not freed", object);
1133        return NULL;
1134}
1135
1136static int __init setup_slub_debug(char *str)
1137{
1138        slub_debug = DEBUG_DEFAULT_FLAGS;
1139        if (*str++ != '=' || !*str)
1140                /*
1141                 * No options specified. Switch on full debugging.
1142                 */
1143                goto out;
1144
1145        if (*str == ',')
1146                /*
1147                 * No options but restriction on slabs. This means full
1148                 * debugging for slabs matching a pattern.
1149                 */
1150                goto check_slabs;
1151
1152        if (tolower(*str) == 'o') {
1153                /*
1154                 * Avoid enabling debugging on caches if its minimum order
1155                 * would increase as a result.
1156                 */
1157                disable_higher_order_debug = 1;
1158                goto out;
1159        }
1160
1161        slub_debug = 0;
1162        if (*str == '-')
1163                /*
1164                 * Switch off all debugging measures.
1165                 */
1166                goto out;
1167
1168        /*
1169         * Determine which debug features should be switched on
1170         */
1171        for (; *str && *str != ','; str++) {
1172                switch (tolower(*str)) {
1173                case 'f':
1174                        slub_debug |= SLAB_DEBUG_FREE;
1175                        break;
1176                case 'z':
1177                        slub_debug |= SLAB_RED_ZONE;
1178                        break;
1179                case 'p':
1180                        slub_debug |= SLAB_POISON;
1181                        break;
1182                case 'u':
1183                        slub_debug |= SLAB_STORE_USER;
1184                        break;
1185                case 't':
1186                        slub_debug |= SLAB_TRACE;
1187                        break;
1188                case 'a':
1189                        slub_debug |= SLAB_FAILSLAB;
1190                        break;
1191                default:
1192                        printk(KERN_ERR "slub_debug option '%c' "
1193                                "unknown. skipped\n", *str);
1194                }
1195        }
1196
1197check_slabs:
1198        if (*str == ',')
1199                slub_debug_slabs = str + 1;
1200out:
1201        return 1;
1202}
1203
1204__setup("slub_debug", setup_slub_debug);
1205
1206static unsigned long kmem_cache_flags(unsigned long object_size,
1207        unsigned long flags, const char *name,
1208        void (*ctor)(void *))
1209{
1210        /*
1211         * Enable debugging if selected on the kernel commandline.
1212         */
1213        if (slub_debug && (!slub_debug_slabs ||
1214                !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1215                flags |= slub_debug;
1216
1217        return flags;
1218}
1219#else
1220static inline void setup_object_debug(struct kmem_cache *s,
1221                        struct page *page, void *object) {}
1222
1223static inline int alloc_debug_processing(struct kmem_cache *s,
1224        struct page *page, void *object, unsigned long addr) { return 0; }
1225
1226static inline struct kmem_cache_node *free_debug_processing(
1227        struct kmem_cache *s, struct page *page, void *object,
1228        unsigned long addr, unsigned long *flags) { return NULL; }
1229
1230static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1231                        { return 1; }
1232static inline int check_object(struct kmem_cache *s, struct page *page,
1233                        void *object, u8 val) { return 1; }
1234static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1235                                        struct page *page) {}
1236static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1237static inline unsigned long kmem_cache_flags(unsigned long object_size,
1238        unsigned long flags, const char *name,
1239        void (*ctor)(void *))
1240{
1241        return flags;
1242}
1243#define slub_debug 0
1244
1245#define disable_higher_order_debug 0
1246
1247static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1248                                                        { return 0; }
1249static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1250                                                        { return 0; }
1251static inline void inc_slabs_node(struct kmem_cache *s, int node,
1252                                                        int objects) {}
1253static inline void dec_slabs_node(struct kmem_cache *s, int node,
1254                                                        int objects) {}
1255
1256static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1257                                                        { return 0; }
1258
1259static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1260                void *object) {}
1261
1262static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1263
1264#endif /* CONFIG_SLUB_DEBUG */
1265
1266/*
1267 * Slab allocation and freeing
1268 */
1269static inline struct page *alloc_slab_page(gfp_t flags, int node,
1270                                        struct kmem_cache_order_objects oo)
1271{
1272        int order = oo_order(oo);
1273
1274        flags |= __GFP_NOTRACK;
1275
1276        if (node == NUMA_NO_NODE)
1277                return alloc_pages(flags, order);
1278        else
1279                return alloc_pages_exact_node(node, flags, order);
1280}
1281
1282static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1283{
1284        struct page *page;
1285        struct kmem_cache_order_objects oo = s->oo;
1286        gfp_t alloc_gfp;
1287
1288        flags &= gfp_allowed_mask;
1289
1290        if (flags & __GFP_WAIT)
1291                local_irq_enable();
1292
1293        flags |= s->allocflags;
1294
1295        /*
1296         * Let the initial higher-order allocation fail under memory pressure
1297         * so we fall-back to the minimum order allocation.
1298         */
1299        alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1300
1301        page = alloc_slab_page(alloc_gfp, node, oo);
1302        if (unlikely(!page)) {
1303                oo = s->min;
1304                /*
1305                 * Allocation may have failed due to fragmentation.
1306                 * Try a lower order alloc if possible
1307                 */
1308                page = alloc_slab_page(flags, node, oo);
1309
1310                if (page)
1311                        stat(s, ORDER_FALLBACK);
1312        }
1313
1314        if (kmemcheck_enabled && page
1315                && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1316                int pages = 1 << oo_order(oo);
1317
1318                kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1319
1320                /*
1321                 * Objects from caches that have a constructor don't get
1322                 * cleared when they're allocated, so we need to do it here.
1323                 */
1324                if (s->ctor)
1325                        kmemcheck_mark_uninitialized_pages(page, pages);
1326                else
1327                        kmemcheck_mark_unallocated_pages(page, pages);
1328        }
1329
1330        if (flags & __GFP_WAIT)
1331                local_irq_disable();
1332        if (!page)
1333                return NULL;
1334
1335        page->objects = oo_objects(oo);
1336        mod_zone_page_state(page_zone(page),
1337                (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1338                NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1339                1 << oo_order(oo));
1340
1341        return page;
1342}
1343
1344static void setup_object(struct kmem_cache *s, struct page *page,
1345                                void *object)
1346{
1347        setup_object_debug(s, page, object);
1348        if (unlikely(s->ctor))
1349                s->ctor(object);
1350}
1351
1352static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1353{
1354        struct page *page;
1355        void *start;
1356        void *last;
1357        void *p;
1358        int order;
1359
1360        BUG_ON(flags & GFP_SLAB_BUG_MASK);
1361
1362        page = allocate_slab(s,
1363                flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1364        if (!page)
1365                goto out;
1366
1367        order = compound_order(page);
1368        inc_slabs_node(s, page_to_nid(page), page->objects);
1369        memcg_bind_pages(s, order);
1370        page->slab_cache = s;
1371        __SetPageSlab(page);
1372        if (page->pfmemalloc)
1373                SetPageSlabPfmemalloc(page);
1374
1375        start = page_address(page);
1376
1377        if (unlikely(s->flags & SLAB_POISON))
1378                memset(start, POISON_INUSE, PAGE_SIZE << order);
1379
1380        last = start;
1381        for_each_object(p, s, start, page->objects) {
1382                setup_object(s, page, last);
1383                set_freepointer(s, last, p);
1384                last = p;
1385        }
1386        setup_object(s, page, last);
1387        set_freepointer(s, last, NULL);
1388
1389        page->freelist = start;
1390        page->inuse = page->objects;
1391        page->frozen = 1;
1392out:
1393        return page;
1394}
1395
1396static void __free_slab(struct kmem_cache *s, struct page *page)
1397{
1398        int order = compound_order(page);
1399        int pages = 1 << order;
1400
1401        if (kmem_cache_debug(s)) {
1402                void *p;
1403
1404                slab_pad_check(s, page);
1405                for_each_object(p, s, page_address(page),
1406                                                page->objects)
1407                        check_object(s, page, p, SLUB_RED_INACTIVE);
1408        }
1409
1410        kmemcheck_free_shadow(page, compound_order(page));
1411
1412        mod_zone_page_state(page_zone(page),
1413                (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1414                NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1415                -pages);
1416
1417        __ClearPageSlabPfmemalloc(page);
1418        __ClearPageSlab(page);
1419
1420        memcg_release_pages(s, order);
1421        page_mapcount_reset(page);
1422        if (current->reclaim_state)
1423                current->reclaim_state->reclaimed_slab += pages;
1424        __free_memcg_kmem_pages(page, order);
1425}
1426
1427#define need_reserve_slab_rcu                                           \
1428        (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1429
1430static void rcu_free_slab(struct rcu_head *h)
1431{
1432        struct page *page;
1433
1434        if (need_reserve_slab_rcu)
1435                page = virt_to_head_page(h);
1436        else
1437                page = container_of((struct list_head *)h, struct page, lru);
1438
1439        __free_slab(page->slab_cache, page);
1440}
1441
1442static void free_slab(struct kmem_cache *s, struct page *page)
1443{
1444        if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1445                struct rcu_head *head;
1446
1447                if (need_reserve_slab_rcu) {
1448                        int order = compound_order(page);
1449                        int offset = (PAGE_SIZE << order) - s->reserved;
1450
1451                        VM_BUG_ON(s->reserved != sizeof(*head));
1452                        head = page_address(page) + offset;
1453                } else {
1454                        /*
1455                         * RCU free overloads the RCU head over the LRU
1456                         */
1457                        head = (void *)&page->lru;
1458                }
1459
1460                call_rcu(head, rcu_free_slab);
1461        } else
1462                __free_slab(s, page);
1463}
1464
1465static void discard_slab(struct kmem_cache *s, struct page *page)
1466{
1467        dec_slabs_node(s, page_to_nid(page), page->objects);
1468        free_slab(s, page);
1469}
1470
1471/*
1472 * Management of partially allocated slabs.
1473 *
1474 * list_lock must be held.
1475 */
1476static inline void add_partial(struct kmem_cache_node *n,
1477                                struct page *page, int tail)
1478{
1479        n->nr_partial++;
1480        if (tail == DEACTIVATE_TO_TAIL)
1481                list_add_tail(&page->lru, &n->partial);
1482        else
1483                list_add(&page->lru, &n->partial);
1484}
1485
1486/*
1487 * list_lock must be held.
1488 */
1489static inline void remove_partial(struct kmem_cache_node *n,
1490                                        struct page *page)
1491{
1492        list_del(&page->lru);
1493        n->nr_partial--;
1494}
1495
1496/*
1497 * Remove slab from the partial list, freeze it and
1498 * return the pointer to the freelist.
1499 *
1500 * Returns a list of objects or NULL if it fails.
1501 *
1502 * Must hold list_lock since we modify the partial list.
1503 */
1504static inline void *acquire_slab(struct kmem_cache *s,
1505                struct kmem_cache_node *n, struct page *page,
1506                int mode, int *objects)
1507{
1508        void *freelist;
1509        unsigned long counters;
1510        struct page new;
1511
1512        /*
1513         * Zap the freelist and set the frozen bit.
1514         * The old freelist is the list of objects for the
1515         * per cpu allocation list.
1516         */
1517        freelist = page->freelist;
1518        counters = page->counters;
1519        new.counters = counters;
1520        *objects = new.objects - new.inuse;
1521        if (mode) {
1522                new.inuse = page->objects;
1523                new.freelist = NULL;
1524        } else {
1525                new.freelist = freelist;
1526        }
1527
1528        VM_BUG_ON(new.frozen);
1529        new.frozen = 1;
1530
1531        if (!__cmpxchg_double_slab(s, page,
1532                        freelist, counters,
1533                        new.freelist, new.counters,
1534                        "acquire_slab"))
1535                return NULL;
1536
1537        remove_partial(n, page);
1538        WARN_ON(!freelist);
1539        return freelist;
1540}
1541
1542static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1543static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1544
1545/*
1546 * Try to allocate a partial slab from a specific node.
1547 */
1548static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1549                                struct kmem_cache_cpu *c, gfp_t flags)
1550{
1551        struct page *page, *page2;
1552        void *object = NULL;
1553        int available = 0;
1554        int objects;
1555
1556        /*
1557         * Racy check. If we mistakenly see no partial slabs then we
1558         * just allocate an empty slab. If we mistakenly try to get a
1559         * partial slab and there is none available then get_partials()
1560         * will return NULL.
1561         */
1562        if (!n || !n->nr_partial)
1563                return NULL;
1564
1565        spin_lock(&n->list_lock);
1566        list_for_each_entry_safe(page, page2, &n->partial, lru) {
1567                void *t;
1568
1569                if (!pfmemalloc_match(page, flags))
1570                        continue;
1571
1572                t = acquire_slab(s, n, page, object == NULL, &objects);
1573                if (!t)
1574                        break;
1575
1576                available += objects;
1577                if (!object) {
1578                        c->page = page;
1579                        stat(s, ALLOC_FROM_PARTIAL);
1580                        object = t;
1581                } else {
1582                        put_cpu_partial(s, page, 0);
1583                        stat(s, CPU_PARTIAL_NODE);
1584                }
1585                if (!kmem_cache_has_cpu_partial(s)
1586                        || available > s->cpu_partial / 2)
1587                        break;
1588
1589        }
1590        spin_unlock(&n->list_lock);
1591        return object;
1592}
1593
1594/*
1595 * Get a page from somewhere. Search in increasing NUMA distances.
1596 */
1597static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1598                struct kmem_cache_cpu *c)
1599{
1600#ifdef CONFIG_NUMA
1601        struct zonelist *zonelist;
1602        struct zoneref *z;
1603        struct zone *zone;
1604        enum zone_type high_zoneidx = gfp_zone(flags);
1605        void *object;
1606        unsigned int cpuset_mems_cookie;
1607
1608        /*
1609         * The defrag ratio allows a configuration of the tradeoffs between
1610         * inter node defragmentation and node local allocations. A lower
1611         * defrag_ratio increases the tendency to do local allocations
1612         * instead of attempting to obtain partial slabs from other nodes.
1613         *
1614         * If the defrag_ratio is set to 0 then kmalloc() always
1615         * returns node local objects. If the ratio is higher then kmalloc()
1616         * may return off node objects because partial slabs are obtained
1617         * from other nodes and filled up.
1618         *
1619         * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1620         * defrag_ratio = 1000) then every (well almost) allocation will
1621         * first attempt to defrag slab caches on other nodes. This means
1622         * scanning over all nodes to look for partial slabs which may be
1623         * expensive if we do it every time we are trying to find a slab
1624         * with available objects.
1625         */
1626        if (!s->remote_node_defrag_ratio ||
1627                        get_cycles() % 1024 > s->remote_node_defrag_ratio)
1628                return NULL;
1629
1630        do {
1631                cpuset_mems_cookie = get_mems_allowed();
1632                zonelist = node_zonelist(slab_node(), flags);
1633                for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1634                        struct kmem_cache_node *n;
1635
1636                        n = get_node(s, zone_to_nid(zone));
1637
1638                        if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1639                                        n->nr_partial > s->min_partial) {
1640                                object = get_partial_node(s, n, c, flags);
1641                                if (object) {
1642                                        /*
1643                                         * Return the object even if
1644                                         * put_mems_allowed indicated that
1645                                         * the cpuset mems_allowed was
1646                                         * updated in parallel. It's a
1647                                         * harmless race between the alloc
1648                                         * and the cpuset update.
1649                                         */
1650                                        put_mems_allowed(cpuset_mems_cookie);
1651                                        return object;
1652                                }
1653                        }
1654                }
1655        } while (!put_mems_allowed(cpuset_mems_cookie));
1656#endif
1657        return NULL;
1658}
1659
1660/*
1661 * Get a partial page, lock it and return it.
1662 */
1663static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1664                struct kmem_cache_cpu *c)
1665{
1666        void *object;
1667        int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1668
1669        object = get_partial_node(s, get_node(s, searchnode), c, flags);
1670        if (object || node != NUMA_NO_NODE)
1671                return object;
1672
1673        return get_any_partial(s, flags, c);
1674}
1675
1676#ifdef CONFIG_PREEMPT
1677/*
1678 * Calculate the next globally unique transaction for disambiguiation
1679 * during cmpxchg. The transactions start with the cpu number and are then
1680 * incremented by CONFIG_NR_CPUS.
1681 */
1682#define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
1683#else
1684/*
1685 * No preemption supported therefore also no need to check for
1686 * different cpus.
1687 */
1688#define TID_STEP 1
1689#endif
1690
1691static inline unsigned long next_tid(unsigned long tid)
1692{
1693        return tid + TID_STEP;
1694}
1695
1696static inline unsigned int tid_to_cpu(unsigned long tid)
1697{
1698        return tid % TID_STEP;
1699}
1700
1701static inline unsigned long tid_to_event(unsigned long tid)
1702{
1703        return tid / TID_STEP;
1704}
1705
1706static inline unsigned int init_tid(int cpu)
1707{
1708        return cpu;
1709}
1710
1711static inline void note_cmpxchg_failure(const char *n,
1712                const struct kmem_cache *s, unsigned long tid)
1713{
1714#ifdef SLUB_DEBUG_CMPXCHG
1715        unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1716
1717        printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1718
1719#ifdef CONFIG_PREEMPT
1720        if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1721                printk("due to cpu change %d -> %d\n",
1722                        tid_to_cpu(tid), tid_to_cpu(actual_tid));
1723        else
1724#endif
1725        if (tid_to_event(tid) != tid_to_event(actual_tid))
1726                printk("due to cpu running other code. Event %ld->%ld\n",
1727                        tid_to_event(tid), tid_to_event(actual_tid));
1728        else
1729                printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1730                        actual_tid, tid, next_tid(tid));
1731#endif
1732        stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1733}
1734
1735static void init_kmem_cache_cpus(struct kmem_cache *s)
1736{
1737        int cpu;
1738
1739        for_each_possible_cpu(cpu)
1740                per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1741}
1742
1743/*
1744 * Remove the cpu slab
1745 */
1746static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1747{
1748        enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1749        struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1750        int lock = 0;
1751        enum slab_modes l = M_NONE, m = M_NONE;
1752        void *nextfree;
1753        int tail = DEACTIVATE_TO_HEAD;
1754        struct page new;
1755        struct page old;
1756
1757        if (page->freelist) {
1758                stat(s, DEACTIVATE_REMOTE_FREES);
1759                tail = DEACTIVATE_TO_TAIL;
1760        }
1761
1762        /*
1763         * Stage one: Free all available per cpu objects back
1764         * to the page freelist while it is still frozen. Leave the
1765         * last one.
1766         *
1767         * There is no need to take the list->lock because the page
1768         * is still frozen.
1769         */
1770        while (freelist && (nextfree = get_freepointer(s, freelist))) {
1771                void *prior;
1772                unsigned long counters;
1773
1774                do {
1775                        prior = page->freelist;
1776                        counters = page->counters;
1777                        set_freepointer(s, freelist, prior);
1778                        new.counters = counters;
1779                        new.inuse--;
1780                        VM_BUG_ON(!new.frozen);
1781
1782                } while (!__cmpxchg_double_slab(s, page,
1783                        prior, counters,
1784                        freelist, new.counters,
1785                        "drain percpu freelist"));
1786
1787                freelist = nextfree;
1788        }
1789
1790        /*
1791         * Stage two: Ensure that the page is unfrozen while the
1792         * list presence reflects the actual number of objects
1793         * during unfreeze.
1794         *
1795         * We setup the list membership and then perform a cmpxchg
1796         * with the count. If there is a mismatch then the page
1797         * is not unfrozen but the page is on the wrong list.
1798         *
1799         * Then we restart the process which may have to remove
1800         * the page from the list that we just put it on again
1801         * because the number of objects in the slab may have
1802         * changed.
1803         */
1804redo:
1805
1806        old.freelist = page->freelist;
1807        old.counters = page->counters;
1808        VM_BUG_ON(!old.frozen);
1809
1810        /* Determine target state of the slab */
1811        new.counters = old.counters;
1812        if (freelist) {
1813                new.inuse--;
1814                set_freepointer(s, freelist, old.freelist);
1815                new.freelist = freelist;
1816        } else
1817                new.freelist = old.freelist;
1818
1819        new.frozen = 0;
1820
1821        if (!new.inuse && n->nr_partial > s->min_partial)
1822                m = M_FREE;
1823        else if (new.freelist) {
1824                m = M_PARTIAL;
1825                if (!lock) {
1826                        lock = 1;
1827                        /*
1828                         * Taking the spinlock removes the possiblity
1829                         * that acquire_slab() will see a slab page that
1830                         * is frozen
1831                         */
1832                        spin_lock(&n->list_lock);
1833                }
1834        } else {
1835                m = M_FULL;
1836                if (kmem_cache_debug(s) && !lock) {
1837                        lock = 1;
1838                        /*
1839                         * This also ensures that the scanning of full
1840                         * slabs from diagnostic functions will not see
1841                         * any frozen slabs.
1842                         */
1843                        spin_lock(&n->list_lock);
1844                }
1845        }
1846
1847        if (l != m) {
1848
1849                if (l == M_PARTIAL)
1850
1851                        remove_partial(n, page);
1852
1853                else if (l == M_FULL)
1854
1855                        remove_full(s, page);
1856
1857                if (m == M_PARTIAL) {
1858
1859                        add_partial(n, page, tail);
1860                        stat(s, tail);
1861
1862                } else if (m == M_FULL) {
1863
1864                        stat(s, DEACTIVATE_FULL);
1865                        add_full(s, n, page);
1866
1867                }
1868        }
1869
1870        l = m;
1871        if (!__cmpxchg_double_slab(s, page,
1872                                old.freelist, old.counters,
1873                                new.freelist, new.counters,
1874                                "unfreezing slab"))
1875                goto redo;
1876
1877        if (lock)
1878                spin_unlock(&n->list_lock);
1879
1880        if (m == M_FREE) {
1881                stat(s, DEACTIVATE_EMPTY);
1882                discard_slab(s, page);
1883                stat(s, FREE_SLAB);
1884        }
1885}
1886
1887/*
1888 * Unfreeze all the cpu partial slabs.
1889 *
1890 * This function must be called with interrupts disabled
1891 * for the cpu using c (or some other guarantee must be there
1892 * to guarantee no concurrent accesses).
1893 */
1894static void unfreeze_partials(struct kmem_cache *s,
1895                struct kmem_cache_cpu *c)
1896{
1897#ifdef CONFIG_SLUB_CPU_PARTIAL
1898        struct kmem_cache_node *n = NULL, *n2 = NULL;
1899        struct page *page, *discard_page = NULL;
1900
1901        while ((page = c->partial)) {
1902                struct page new;
1903                struct page old;
1904
1905                c->partial = page->next;
1906
1907                n2 = get_node(s, page_to_nid(page));
1908                if (n != n2) {
1909                        if (n)
1910                                spin_unlock(&n->list_lock);
1911
1912                        n = n2;
1913                        spin_lock(&n->list_lock);
1914                }
1915
1916                do {
1917
1918                        old.freelist = page->freelist;
1919                        old.counters = page->counters;
1920                        VM_BUG_ON(!old.frozen);
1921
1922                        new.counters = old.counters;
1923                        new.freelist = old.freelist;
1924
1925                        new.frozen = 0;
1926
1927                } while (!__cmpxchg_double_slab(s, page,
1928                                old.freelist, old.counters,
1929                                new.freelist, new.counters,
1930                                "unfreezing slab"));
1931
1932                if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1933                        page->next = discard_page;
1934                        discard_page = page;
1935                } else {
1936                        add_partial(n, page, DEACTIVATE_TO_TAIL);
1937                        stat(s, FREE_ADD_PARTIAL);
1938                }
1939        }
1940
1941        if (n)
1942                spin_unlock(&n->list_lock);
1943
1944        while (discard_page) {
1945                page = discard_page;
1946                discard_page = discard_page->next;
1947
1948                stat(s, DEACTIVATE_EMPTY);
1949                discard_slab(s, page);
1950                stat(s, FREE_SLAB);
1951        }
1952#endif
1953}
1954
1955/*
1956 * Put a page that was just frozen (in __slab_free) into a partial page
1957 * slot if available. This is done without interrupts disabled and without
1958 * preemption disabled. The cmpxchg is racy and may put the partial page
1959 * onto a random cpus partial slot.
1960 *
1961 * If we did not find a slot then simply move all the partials to the
1962 * per node partial list.
1963 */
1964static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1965{
1966#ifdef CONFIG_SLUB_CPU_PARTIAL
1967        struct page *oldpage;
1968        int pages;
1969        int pobjects;
1970
1971        do {
1972                pages = 0;
1973                pobjects = 0;
1974                oldpage = this_cpu_read(s->cpu_slab->partial);
1975
1976                if (oldpage) {
1977                        pobjects = oldpage->pobjects;
1978                        pages = oldpage->pages;
1979                        if (drain && pobjects > s->cpu_partial) {
1980                                unsigned long flags;
1981                                /*
1982                                 * partial array is full. Move the existing
1983                                 * set to the per node partial list.
1984                                 */
1985                                local_irq_save(flags);
1986                                unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
1987                                local_irq_restore(flags);
1988                                oldpage = NULL;
1989                                pobjects = 0;
1990                                pages = 0;
1991                                stat(s, CPU_PARTIAL_DRAIN);
1992                        }
1993                }
1994
1995                pages++;
1996                pobjects += page->objects - page->inuse;
1997
1998                page->pages = pages;
1999                page->pobjects = pobjects;
2000                page->next = oldpage;
2001
2002        } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
2003#endif
2004}
2005
2006static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2007{
2008        stat(s, CPUSLAB_FLUSH);
2009        deactivate_slab(s, c->page, c->freelist);
2010
2011        c->tid = next_tid(c->tid);
2012        c->page = NULL;
2013        c->freelist = NULL;
2014}
2015
2016/*
2017 * Flush cpu slab.
2018 *
2019 * Called from IPI handler with interrupts disabled.
2020 */
2021static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2022{
2023        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2024
2025        if (likely(c)) {
2026                if (c->page)
2027                        flush_slab(s, c);
2028
2029                unfreeze_partials(s, c);
2030        }
2031}
2032
2033static void flush_cpu_slab(void *d)
2034{
2035        struct kmem_cache *s = d;
2036
2037        __flush_cpu_slab(s, smp_processor_id());
2038}
2039
2040static bool has_cpu_slab(int cpu, void *info)
2041{
2042        struct kmem_cache *s = info;
2043        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2044
2045        return c->page || c->partial;
2046}
2047
2048static void flush_all(struct kmem_cache *s)
2049{
2050        on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2051}
2052
2053/*
2054 * Check if the objects in a per cpu structure fit numa
2055 * locality expectations.
2056 */
2057static inline int node_match(struct page *page, int node)
2058{
2059#ifdef CONFIG_NUMA
2060        if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2061                return 0;
2062#endif
2063        return 1;
2064}
2065
2066static int count_free(struct page *page)
2067{
2068        return page->objects - page->inuse;
2069}
2070
2071static unsigned long count_partial(struct kmem_cache_node *n,
2072                                        int (*get_count)(struct page *))
2073{
2074        unsigned long flags;
2075        unsigned long x = 0;
2076        struct page *page;
2077
2078        spin_lock_irqsave(&n->list_lock, flags);
2079        list_for_each_entry(page, &n->partial, lru)
2080                x += get_count(page);
2081        spin_unlock_irqrestore(&n->list_lock, flags);
2082        return x;
2083}
2084
2085static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2086{
2087#ifdef CONFIG_SLUB_DEBUG
2088        return atomic_long_read(&n->total_objects);
2089#else
2090        return 0;
2091#endif
2092}
2093
2094static noinline void
2095slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2096{
2097        int node;
2098
2099        printk(KERN_WARNING
2100                "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2101                nid, gfpflags);
2102        printk(KERN_WARNING "  cache: %s, object size: %d, buffer size: %d, "
2103                "default order: %d, min order: %d\n", s->name, s->object_size,
2104                s->size, oo_order(s->oo), oo_order(s->min));
2105
2106        if (oo_order(s->min) > get_order(s->object_size))
2107                printk(KERN_WARNING "  %s debugging increased min order, use "
2108                       "slub_debug=O to disable.\n", s->name);
2109
2110        for_each_online_node(node) {
2111                struct kmem_cache_node *n = get_node(s, node);
2112                unsigned long nr_slabs;
2113                unsigned long nr_objs;
2114                unsigned long nr_free;
2115
2116                if (!n)
2117                        continue;
2118
2119                nr_free  = count_partial(n, count_free);
2120                nr_slabs = node_nr_slabs(n);
2121                nr_objs  = node_nr_objs(n);
2122
2123                printk(KERN_WARNING
2124                        "  node %d: slabs: %ld, objs: %ld, free: %ld\n",
2125                        node, nr_slabs, nr_objs, nr_free);
2126        }
2127}
2128
2129static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2130                        int node, struct kmem_cache_cpu **pc)
2131{
2132        void *freelist;
2133        struct kmem_cache_cpu *c = *pc;
2134        struct page *page;
2135
2136        freelist = get_partial(s, flags, node, c);
2137
2138        if (freelist)
2139                return freelist;
2140
2141        page = new_slab(s, flags, node);
2142        if (page) {
2143                c = __this_cpu_ptr(s->cpu_slab);
2144                if (c->page)
2145                        flush_slab(s, c);
2146
2147                /*
2148                 * No other reference to the page yet so we can
2149                 * muck around with it freely without cmpxchg
2150                 */
2151                freelist = page->freelist;
2152                page->freelist = NULL;
2153
2154                stat(s, ALLOC_SLAB);
2155                c->page = page;
2156                *pc = c;
2157        } else
2158                freelist = NULL;
2159
2160        return freelist;
2161}
2162
2163static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2164{
2165        if (unlikely(PageSlabPfmemalloc(page)))
2166                return gfp_pfmemalloc_allowed(gfpflags);
2167
2168        return true;
2169}
2170
2171/*
2172 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2173 * or deactivate the page.
2174 *
2175 * The page is still frozen if the return value is not NULL.
2176 *
2177 * If this function returns NULL then the page has been unfrozen.
2178 *
2179 * This function must be called with interrupt disabled.
2180 */
2181static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2182{
2183        struct page new;
2184        unsigned long counters;
2185        void *freelist;
2186
2187        do {
2188                freelist = page->freelist;
2189                counters = page->counters;
2190
2191                new.counters = counters;
2192                VM_BUG_ON(!new.frozen);
2193
2194                new.inuse = page->objects;
2195                new.frozen = freelist != NULL;
2196
2197        } while (!__cmpxchg_double_slab(s, page,
2198                freelist, counters,
2199                NULL, new.counters,
2200                "get_freelist"));
2201
2202        return freelist;
2203}
2204
2205/*
2206 * Slow path. The lockless freelist is empty or we need to perform
2207 * debugging duties.
2208 *
2209 * Processing is still very fast if new objects have been freed to the
2210 * regular freelist. In that case we simply take over the regular freelist
2211 * as the lockless freelist and zap the regular freelist.
2212 *
2213 * If that is not working then we fall back to the partial lists. We take the
2214 * first element of the freelist as the object to allocate now and move the
2215 * rest of the freelist to the lockless freelist.
2216 *
2217 * And if we were unable to get a new slab from the partial slab lists then
2218 * we need to allocate a new slab. This is the slowest path since it involves
2219 * a call to the page allocator and the setup of a new slab.
2220 */
2221static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2222                          unsigned long addr, struct kmem_cache_cpu *c)
2223{
2224        void *freelist;
2225        struct page *page;
2226        unsigned long flags;
2227
2228        local_irq_save(flags);
2229#ifdef CONFIG_PREEMPT
2230        /*
2231         * We may have been preempted and rescheduled on a different
2232         * cpu before disabling interrupts. Need to reload cpu area
2233         * pointer.
2234         */
2235        c = this_cpu_ptr(s->cpu_slab);
2236#endif
2237
2238        page = c->page;
2239        if (!page)
2240                goto new_slab;
2241redo:
2242
2243        if (unlikely(!node_match(page, node))) {
2244                stat(s, ALLOC_NODE_MISMATCH);
2245                deactivate_slab(s, page, c->freelist);
2246                c->page = NULL;
2247                c->freelist = NULL;
2248                goto new_slab;
2249        }
2250
2251        /*
2252         * By rights, we should be searching for a slab page that was
2253         * PFMEMALLOC but right now, we are losing the pfmemalloc
2254         * information when the page leaves the per-cpu allocator
2255         */
2256        if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2257                deactivate_slab(s, page, c->freelist);
2258                c->page = NULL;
2259                c->freelist = NULL;
2260                goto new_slab;
2261        }
2262
2263        /* must check again c->freelist in case of cpu migration or IRQ */
2264        freelist = c->freelist;
2265        if (freelist)
2266                goto load_freelist;
2267
2268        stat(s, ALLOC_SLOWPATH);
2269
2270        freelist = get_freelist(s, page);
2271
2272        if (!freelist) {
2273                c->page = NULL;
2274                stat(s, DEACTIVATE_BYPASS);
2275                goto new_slab;
2276        }
2277
2278        stat(s, ALLOC_REFILL);
2279
2280load_freelist:
2281        /*
2282         * freelist is pointing to the list of objects to be used.
2283         * page is pointing to the page from which the objects are obtained.
2284         * That page must be frozen for per cpu allocations to work.
2285         */
2286        VM_BUG_ON(!c->page->frozen);
2287        c->freelist = get_freepointer(s, freelist);
2288        c->tid = next_tid(c->tid);
2289        local_irq_restore(flags);
2290        return freelist;
2291
2292new_slab:
2293
2294        if (c->partial) {
2295                page = c->page = c->partial;
2296                c->partial = page->next;
2297                stat(s, CPU_PARTIAL_ALLOC);
2298                c->freelist = NULL;
2299                goto redo;
2300        }
2301
2302        freelist = new_slab_objects(s, gfpflags, node, &c);
2303
2304        if (unlikely(!freelist)) {
2305                if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2306                        slab_out_of_memory(s, gfpflags, node);
2307
2308                local_irq_restore(flags);
2309                return NULL;
2310        }
2311
2312        page = c->page;
2313        if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2314                goto load_freelist;
2315
2316        /* Only entered in the debug case */
2317        if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2318                goto new_slab;  /* Slab failed checks. Next slab needed */
2319
2320        deactivate_slab(s, page, get_freepointer(s, freelist));
2321        c->page = NULL;
2322        c->freelist = NULL;
2323        local_irq_restore(flags);
2324        return freelist;
2325}
2326
2327/*
2328 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2329 * have the fastpath folded into their functions. So no function call
2330 * overhead for requests that can be satisfied on the fastpath.
2331 *
2332 * The fastpath works by first checking if the lockless freelist can be used.
2333 * If not then __slab_alloc is called for slow processing.
2334 *
2335 * Otherwise we can simply pick the next object from the lockless free list.
2336 */
2337static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2338                gfp_t gfpflags, int node, unsigned long addr)
2339{
2340        void **object;
2341        struct kmem_cache_cpu *c;
2342        struct page *page;
2343        unsigned long tid;
2344
2345        if (slab_pre_alloc_hook(s, gfpflags))
2346                return NULL;
2347
2348        s = memcg_kmem_get_cache(s, gfpflags);
2349redo:
2350        /*
2351         * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2352         * enabled. We may switch back and forth between cpus while
2353         * reading from one cpu area. That does not matter as long
2354         * as we end up on the original cpu again when doing the cmpxchg.
2355         *
2356         * Preemption is disabled for the retrieval of the tid because that
2357         * must occur from the current processor. We cannot allow rescheduling
2358         * on a different processor between the determination of the pointer
2359         * and the retrieval of the tid.
2360         */
2361        preempt_disable();
2362        c = __this_cpu_ptr(s->cpu_slab);
2363
2364        /*
2365         * The transaction ids are globally unique per cpu and per operation on
2366         * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2367         * occurs on the right processor and that there was no operation on the
2368         * linked list in between.
2369         */
2370        tid = c->tid;
2371        preempt_enable();
2372
2373        object = c->freelist;
2374        page = c->page;
2375        if (unlikely(!object || !page || !node_match(page, node)))
2376                object = __slab_alloc(s, gfpflags, node, addr, c);
2377
2378        else {
2379                void *next_object = get_freepointer_safe(s, object);
2380
2381                /*
2382                 * The cmpxchg will only match if there was no additional
2383                 * operation and if we are on the right processor.
2384                 *
2385                 * The cmpxchg does the following atomically (without lock semantics!)
2386                 * 1. Relocate first pointer to the current per cpu area.
2387                 * 2. Verify that tid and freelist have not been changed
2388                 * 3. If they were not changed replace tid and freelist
2389                 *
2390                 * Since this is without lock semantics the protection is only against
2391                 * code executing on this cpu *not* from access by other cpus.
2392                 */
2393                if (unlikely(!this_cpu_cmpxchg_double(
2394                                s->cpu_slab->freelist, s->cpu_slab->tid,
2395                                object, tid,
2396                                next_object, next_tid(tid)))) {
2397
2398                        note_cmpxchg_failure("slab_alloc", s, tid);
2399                        goto redo;
2400                }
2401                prefetch_freepointer(s, next_object);
2402                stat(s, ALLOC_FASTPATH);
2403        }
2404
2405        if (unlikely(gfpflags & __GFP_ZERO) && object)
2406                memset(object, 0, s->object_size);
2407
2408        slab_post_alloc_hook(s, gfpflags, object);
2409
2410        return object;
2411}
2412
2413static __always_inline void *slab_alloc(struct kmem_cache *s,
2414                gfp_t gfpflags, unsigned long addr)
2415{
2416        return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2417}
2418
2419void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2420{
2421        void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2422
2423        trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2424
2425        return ret;
2426}
2427EXPORT_SYMBOL(kmem_cache_alloc);
2428
2429#ifdef CONFIG_TRACING
2430void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2431{
2432        void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2433        trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2434        return ret;
2435}
2436EXPORT_SYMBOL(kmem_cache_alloc_trace);
2437
2438void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2439{
2440        void *ret = kmalloc_order(size, flags, order);
2441        trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2442        return ret;
2443}
2444EXPORT_SYMBOL(kmalloc_order_trace);
2445#endif
2446
2447#ifdef CONFIG_NUMA
2448void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2449{
2450        void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2451
2452        trace_kmem_cache_alloc_node(_RET_IP_, ret,
2453                                    s->object_size, s->size, gfpflags, node);
2454
2455        return ret;
2456}
2457EXPORT_SYMBOL(kmem_cache_alloc_node);
2458
2459#ifdef CONFIG_TRACING
2460void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2461                                    gfp_t gfpflags,
2462                                    int node, size_t size)
2463{
2464        void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2465
2466        trace_kmalloc_node(_RET_IP_, ret,
2467                           size, s->size, gfpflags, node);
2468        return ret;
2469}
2470EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2471#endif
2472#endif
2473
2474/*
2475 * Slow patch handling. This may still be called frequently since objects
2476 * have a longer lifetime than the cpu slabs in most processing loads.
2477 *
2478 * So we still attempt to reduce cache line usage. Just take the slab
2479 * lock and free the item. If there is no additional partial page
2480 * handling required then we can return immediately.
2481 */
2482static void __slab_free(struct kmem_cache *s, struct page *page,
2483                        void *x, unsigned long addr)
2484{
2485        void *prior;
2486        void **object = (void *)x;
2487        int was_frozen;
2488        struct page new;
2489        unsigned long counters;
2490        struct kmem_cache_node *n = NULL;
2491        unsigned long uninitialized_var(flags);
2492
2493        stat(s, FREE_SLOWPATH);
2494
2495        if (kmem_cache_debug(s) &&
2496                !(n = free_debug_processing(s, page, x, addr, &flags)))
2497                return;
2498
2499        do {
2500                if (unlikely(n)) {
2501                        spin_unlock_irqrestore(&n->list_lock, flags);
2502                        n = NULL;
2503                }
2504                prior = page->freelist;
2505                counters = page->counters;
2506                set_freepointer(s, object, prior);
2507                new.counters = counters;
2508                was_frozen = new.frozen;
2509                new.inuse--;
2510                if ((!new.inuse || !prior) && !was_frozen) {
2511
2512                        if (kmem_cache_has_cpu_partial(s) && !prior)
2513
2514                                /*
2515                                 * Slab was on no list before and will be partially empty
2516                                 * We can defer the list move and instead freeze it.
2517                                 */
2518                                new.frozen = 1;
2519
2520                        else { /* Needs to be taken off a list */
2521
2522                                n = get_node(s, page_to_nid(page));
2523                                /*
2524                                 * Speculatively acquire the list_lock.
2525                                 * If the cmpxchg does not succeed then we may
2526                                 * drop the list_lock without any processing.
2527                                 *
2528                                 * Otherwise the list_lock will synchronize with
2529                                 * other processors updating the list of slabs.
2530                                 */
2531                                spin_lock_irqsave(&n->list_lock, flags);
2532
2533                        }
2534                }
2535
2536        } while (!cmpxchg_double_slab(s, page,
2537                prior, counters,
2538                object, new.counters,
2539                "__slab_free"));
2540
2541        if (likely(!n)) {
2542
2543                /*
2544                 * If we just froze the page then put it onto the
2545                 * per cpu partial list.
2546                 */
2547                if (new.frozen && !was_frozen) {
2548                        put_cpu_partial(s, page, 1);
2549                        stat(s, CPU_PARTIAL_FREE);
2550                }
2551                /*
2552                 * The list lock was not taken therefore no list
2553                 * activity can be necessary.
2554                 */
2555                if (was_frozen)
2556                        stat(s, FREE_FROZEN);
2557                return;
2558        }
2559
2560        if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2561                goto slab_empty;
2562
2563        /*
2564         * Objects left in the slab. If it was not on the partial list before
2565         * then add it.
2566         */
2567        if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2568                if (kmem_cache_debug(s))
2569                        remove_full(s, page);
2570                add_partial(n, page, DEACTIVATE_TO_TAIL);
2571                stat(s, FREE_ADD_PARTIAL);
2572        }
2573        spin_unlock_irqrestore(&n->list_lock, flags);
2574        return;
2575
2576slab_empty:
2577        if (prior) {
2578                /*
2579                 * Slab on the partial list.
2580                 */
2581                remove_partial(n, page);
2582                stat(s, FREE_REMOVE_PARTIAL);
2583        } else
2584                /* Slab must be on the full list */
2585                remove_full(s, page);
2586
2587        spin_unlock_irqrestore(&n->list_lock, flags);
2588        stat(s, FREE_SLAB);
2589        discard_slab(s, page);
2590}
2591
2592/*
2593 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2594 * can perform fastpath freeing without additional function calls.
2595 *
2596 * The fastpath is only possible if we are freeing to the current cpu slab
2597 * of this processor. This typically the case if we have just allocated
2598 * the item before.
2599 *
2600 * If fastpath is not possible then fall back to __slab_free where we deal
2601 * with all sorts of special processing.
2602 */
2603static __always_inline void slab_free(struct kmem_cache *s,
2604                        struct page *page, void *x, unsigned long addr)
2605{
2606        void **object = (void *)x;
2607        struct kmem_cache_cpu *c;
2608        unsigned long tid;
2609
2610        slab_free_hook(s, x);
2611
2612redo:
2613        /*
2614         * Determine the currently cpus per cpu slab.
2615         * The cpu may change afterward. However that does not matter since
2616         * data is retrieved via this pointer. If we are on the same cpu
2617         * during the cmpxchg then the free will succedd.
2618         */
2619        preempt_disable();
2620        c = __this_cpu_ptr(s->cpu_slab);
2621
2622        tid = c->tid;
2623        preempt_enable();
2624
2625        if (likely(page == c->page)) {
2626                set_freepointer(s, object, c->freelist);
2627
2628                if (unlikely(!this_cpu_cmpxchg_double(
2629                                s->cpu_slab->freelist, s->cpu_slab->tid,
2630                                c->freelist, tid,
2631                                object, next_tid(tid)))) {
2632
2633                        note_cmpxchg_failure("slab_free", s, tid);
2634                        goto redo;
2635                }
2636                stat(s, FREE_FASTPATH);
2637        } else
2638                __slab_free(s, page, x, addr);
2639
2640}
2641
2642void kmem_cache_free(struct kmem_cache *s, void *x)
2643{
2644        s = cache_from_obj(s, x);
2645        if (!s)
2646                return;
2647        slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2648        trace_kmem_cache_free(_RET_IP_, x);
2649}
2650EXPORT_SYMBOL(kmem_cache_free);
2651
2652/*
2653 * Object placement in a slab is made very easy because we always start at
2654 * offset 0. If we tune the size of the object to the alignment then we can
2655 * get the required alignment by putting one properly sized object after
2656 * another.
2657 *
2658 * Notice that the allocation order determines the sizes of the per cpu
2659 * caches. Each processor has always one slab available for allocations.
2660 * Increasing the allocation order reduces the number of times that slabs
2661 * must be moved on and off the partial lists and is therefore a factor in
2662 * locking overhead.
2663 */
2664
2665/*
2666 * Mininum / Maximum order of slab pages. This influences locking overhead
2667 * and slab fragmentation. A higher order reduces the number of partial slabs
2668 * and increases the number of allocations possible without having to
2669 * take the list_lock.
2670 */
2671static int slub_min_order;
2672static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2673static int slub_min_objects;
2674
2675/*
2676 * Merge control. If this is set then no merging of slab caches will occur.
2677 * (Could be removed. This was introduced to pacify the merge skeptics.)
2678 */
2679static int slub_nomerge;
2680
2681/*
2682 * Calculate the order of allocation given an slab object size.
2683 *
2684 * The order of allocation has significant impact on performance and other
2685 * system components. Generally order 0 allocations should be preferred since
2686 * order 0 does not cause fragmentation in the page allocator. Larger objects
2687 * be problematic to put into order 0 slabs because there may be too much
2688 * unused space left. We go to a higher order if more than 1/16th of the slab
2689 * would be wasted.
2690 *
2691 * In order to reach satisfactory performance we must ensure that a minimum
2692 * number of objects is in one slab. Otherwise we may generate too much
2693 * activity on the partial lists which requires taking the list_lock. This is
2694 * less a concern for large slabs though which are rarely used.
2695 *
2696 * slub_max_order specifies the order where we begin to stop considering the
2697 * number of objects in a slab as critical. If we reach slub_max_order then
2698 * we try to keep the page order as low as possible. So we accept more waste
2699 * of space in favor of a small page order.
2700 *
2701 * Higher order allocations also allow the placement of more objects in a
2702 * slab and thereby reduce object handling overhead. If the user has
2703 * requested a higher mininum order then we start with that one instead of
2704 * the smallest order which will fit the object.
2705 */
2706static inline int slab_order(int size, int min_objects,
2707                                int max_order, int fract_leftover, int reserved)
2708{
2709        int order;
2710        int rem;
2711        int min_order = slub_min_order;
2712
2713        if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2714                return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2715
2716        for (order = max(min_order,
2717                                fls(min_objects * size - 1) - PAGE_SHIFT);
2718                        order <= max_order; order++) {
2719
2720                unsigned long slab_size = PAGE_SIZE << order;
2721
2722                if (slab_size < min_objects * size + reserved)
2723                        continue;
2724
2725                rem = (slab_size - reserved) % size;
2726
2727                if (rem <= slab_size / fract_leftover)
2728                        break;
2729
2730        }
2731
2732        return order;
2733}
2734
2735static inline int calculate_order(int size, int reserved)
2736{
2737        int order;
2738        int min_objects;
2739        int fraction;
2740        int max_objects;
2741
2742        /*
2743         * Attempt to find best configuration for a slab. This
2744         * works by first attempting to generate a layout with
2745         * the best configuration and backing off gradually.
2746         *
2747         * First we reduce the acceptable waste in a slab. Then
2748         * we reduce the minimum objects required in a slab.
2749         */
2750        min_objects = slub_min_objects;
2751        if (!min_objects)
2752                min_objects = 4 * (fls(nr_cpu_ids) + 1);
2753        max_objects = order_objects(slub_max_order, size, reserved);
2754        min_objects = min(min_objects, max_objects);
2755
2756        while (min_objects > 1) {
2757                fraction = 16;
2758                while (fraction >= 4) {
2759                        order = slab_order(size, min_objects,
2760                                        slub_max_order, fraction, reserved);
2761                        if (order <= slub_max_order)
2762                                return order;
2763                        fraction /= 2;
2764                }
2765                min_objects--;
2766        }
2767
2768        /*
2769         * We were unable to place multiple objects in a slab. Now
2770         * lets see if we can place a single object there.
2771         */
2772        order = slab_order(size, 1, slub_max_order, 1, reserved);
2773        if (order <= slub_max_order)
2774                return order;
2775
2776        /*
2777         * Doh this slab cannot be placed using slub_max_order.
2778         */
2779        order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2780        if (order < MAX_ORDER)
2781                return order;
2782        return -ENOSYS;
2783}
2784
2785static void
2786init_kmem_cache_node(struct kmem_cache_node *n)
2787{
2788        n->nr_partial = 0;
2789        spin_lock_init(&n->list_lock);
2790        INIT_LIST_HEAD(&n->partial);
2791#ifdef CONFIG_SLUB_DEBUG
2792        atomic_long_set(&n->nr_slabs, 0);
2793        atomic_long_set(&n->total_objects, 0);
2794        INIT_LIST_HEAD(&n->full);
2795#endif
2796}
2797
2798static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2799{
2800        BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2801                        KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2802
2803        /*
2804         * Must align to double word boundary for the double cmpxchg
2805         * instructions to work; see __pcpu_double_call_return_bool().
2806         */
2807        s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2808                                     2 * sizeof(void *));
2809
2810        if (!s->cpu_slab)
2811                return 0;
2812
2813        init_kmem_cache_cpus(s);
2814
2815        return 1;
2816}
2817
2818static struct kmem_cache *kmem_cache_node;
2819
2820/*
2821 * No kmalloc_node yet so do it by hand. We know that this is the first
2822 * slab on the node for this slabcache. There are no concurrent accesses
2823 * possible.
2824 *
2825 * Note that this function only works on the kmalloc_node_cache
2826 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2827 * memory on a fresh node that has no slab structures yet.
2828 */
2829static void early_kmem_cache_node_alloc(int node)
2830{
2831        struct page *page;
2832        struct kmem_cache_node *n;
2833
2834        BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2835
2836        page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2837
2838        BUG_ON(!page);
2839        if (page_to_nid(page) != node) {
2840                printk(KERN_ERR "SLUB: Unable to allocate memory from "
2841                                "node %d\n", node);
2842                printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2843                                "in order to be able to continue\n");
2844        }
2845
2846        n = page->freelist;
2847        BUG_ON(!n);
2848        page->freelist = get_freepointer(kmem_cache_node, n);
2849        page->inuse = 1;
2850        page->frozen = 0;
2851        kmem_cache_node->node[node] = n;
2852#ifdef CONFIG_SLUB_DEBUG
2853        init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2854        init_tracking(kmem_cache_node, n);
2855#endif
2856        init_kmem_cache_node(n);
2857        inc_slabs_node(kmem_cache_node, node, page->objects);
2858
2859        add_partial(n, page, DEACTIVATE_TO_HEAD);
2860}
2861
2862static void free_kmem_cache_nodes(struct kmem_cache *s)
2863{
2864        int node;
2865
2866        for_each_node_state(node, N_NORMAL_MEMORY) {
2867                struct kmem_cache_node *n = s->node[node];
2868
2869                if (n)
2870                        kmem_cache_free(kmem_cache_node, n);
2871
2872                s->node[node] = NULL;
2873        }
2874}
2875
2876static int init_kmem_cache_nodes(struct kmem_cache *s)
2877{
2878        int node;
2879
2880        for_each_node_state(node, N_NORMAL_MEMORY) {
2881                struct kmem_cache_node *n;
2882
2883                if (slab_state == DOWN) {
2884                        early_kmem_cache_node_alloc(node);
2885                        continue;
2886                }
2887                n = kmem_cache_alloc_node(kmem_cache_node,
2888                                                GFP_KERNEL, node);
2889
2890                if (!n) {
2891                        free_kmem_cache_nodes(s);
2892                        return 0;
2893                }
2894
2895                s->node[node] = n;
2896                init_kmem_cache_node(n);
2897        }
2898        return 1;
2899}
2900
2901static void set_min_partial(struct kmem_cache *s, unsigned long min)
2902{
2903        if (min < MIN_PARTIAL)
2904                min = MIN_PARTIAL;
2905        else if (min > MAX_PARTIAL)
2906                min = MAX_PARTIAL;
2907        s->min_partial = min;
2908}
2909
2910/*
2911 * calculate_sizes() determines the order and the distribution of data within
2912 * a slab object.
2913 */
2914static int calculate_sizes(struct kmem_cache *s, int forced_order)
2915{
2916        unsigned long flags = s->flags;
2917        unsigned long size = s->object_size;
2918        int order;
2919
2920        /*
2921         * Round up object size to the next word boundary. We can only
2922         * place the free pointer at word boundaries and this determines
2923         * the possible location of the free pointer.
2924         */
2925        size = ALIGN(size, sizeof(void *));
2926
2927#ifdef CONFIG_SLUB_DEBUG
2928        /*
2929         * Determine if we can poison the object itself. If the user of
2930         * the slab may touch the object after free or before allocation
2931         * then we should never poison the object itself.
2932         */
2933        if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2934                        !s->ctor)
2935                s->flags |= __OBJECT_POISON;
2936        else
2937                s->flags &= ~__OBJECT_POISON;
2938
2939
2940        /*
2941         * If we are Redzoning then check if there is some space between the
2942         * end of the object and the free pointer. If not then add an
2943         * additional word to have some bytes to store Redzone information.
2944         */
2945        if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2946                size += sizeof(void *);
2947#endif
2948
2949        /*
2950         * With that we have determined the number of bytes in actual use
2951         * by the object. This is the potential offset to the free pointer.
2952         */
2953        s->inuse = size;
2954
2955        if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2956                s->ctor)) {
2957                /*
2958                 * Relocate free pointer after the object if it is not
2959                 * permitted to overwrite the first word of the object on
2960                 * kmem_cache_free.
2961                 *
2962                 * This is the case if we do RCU, have a constructor or
2963                 * destructor or are poisoning the objects.
2964                 */
2965                s->offset = size;
2966                size += sizeof(void *);
2967        }
2968
2969#ifdef CONFIG_SLUB_DEBUG
2970        if (flags & SLAB_STORE_USER)
2971                /*
2972                 * Need to store information about allocs and frees after
2973                 * the object.
2974                 */
2975                size += 2 * sizeof(struct track);
2976
2977        if (flags & SLAB_RED_ZONE)
2978                /*
2979                 * Add some empty padding so that we can catch
2980                 * overwrites from earlier objects rather than let
2981                 * tracking information or the free pointer be
2982                 * corrupted if a user writes before the start
2983                 * of the object.
2984                 */
2985                size += sizeof(void *);
2986#endif
2987
2988        /*
2989         * SLUB stores one object immediately after another beginning from
2990         * offset 0. In order to align the objects we have to simply size
2991         * each object to conform to the alignment.
2992         */
2993        size = ALIGN(size, s->align);
2994        s->size = size;
2995        if (forced_order >= 0)
2996                order = forced_order;
2997        else
2998                order = calculate_order(size, s->reserved);
2999
3000        if (order < 0)
3001                return 0;
3002
3003        s->allocflags = 0;
3004        if (order)
3005                s->allocflags |= __GFP_COMP;
3006
3007        if (s->flags & SLAB_CACHE_DMA)
3008                s->allocflags |= GFP_DMA;
3009
3010        if (s->flags & SLAB_RECLAIM_ACCOUNT)
3011                s->allocflags |= __GFP_RECLAIMABLE;
3012
3013        /*
3014         * Determine the number of objects per slab
3015         */
3016        s->oo = oo_make(order, size, s->reserved);
3017        s->min = oo_make(get_order(size), size, s->reserved);
3018        if (oo_objects(s->oo) > oo_objects(s->max))
3019                s->max = s->oo;
3020
3021        return !!oo_objects(s->oo);
3022}
3023
3024static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3025{
3026        s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3027        s->reserved = 0;
3028
3029        if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3030                s->reserved = sizeof(struct rcu_head);
3031
3032        if (!calculate_sizes(s, -1))
3033                goto error;
3034        if (disable_higher_order_debug) {
3035                /*
3036                 * Disable debugging flags that store metadata if the min slab
3037                 * order increased.
3038                 */
3039                if (get_order(s->size) > get_order(s->object_size)) {
3040                        s->flags &= ~DEBUG_METADATA_FLAGS;
3041                        s->offset = 0;
3042                        if (!calculate_sizes(s, -1))
3043                                goto error;
3044                }
3045        }
3046
3047#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3048    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3049        if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3050                /* Enable fast mode */
3051                s->flags |= __CMPXCHG_DOUBLE;
3052#endif
3053
3054        /*
3055         * The larger the object size is, the more pages we want on the partial
3056         * list to avoid pounding the page allocator excessively.
3057         */
3058        set_min_partial(s, ilog2(s->size) / 2);
3059
3060        /*
3061         * cpu_partial determined the maximum number of objects kept in the
3062         * per cpu partial lists of a processor.
3063         *
3064         * Per cpu partial lists mainly contain slabs that just have one
3065         * object freed. If they are used for allocation then they can be
3066         * filled up again with minimal effort. The slab will never hit the
3067         * per node partial lists and therefore no locking will be required.
3068         *
3069         * This setting also determines
3070         *
3071         * A) The number of objects from per cpu partial slabs dumped to the
3072         *    per node list when we reach the limit.
3073         * B) The number of objects in cpu partial slabs to extract from the
3074         *    per node list when we run out of per cpu objects. We only fetch 50%
3075         *    to keep some capacity around for frees.
3076         */
3077        if (!kmem_cache_has_cpu_partial(s))
3078                s->cpu_partial = 0;
3079        else if (s->size >= PAGE_SIZE)
3080                s->cpu_partial = 2;
3081        else if (s->size >= 1024)
3082                s->cpu_partial = 6;
3083        else if (s->size >= 256)
3084                s->cpu_partial = 13;
3085        else
3086                s->cpu_partial = 30;
3087
3088#ifdef CONFIG_NUMA
3089        s->remote_node_defrag_ratio = 1000;
3090#endif
3091        if (!init_kmem_cache_nodes(s))
3092                goto error;
3093
3094        if (alloc_kmem_cache_cpus(s))
3095                return 0;
3096
3097        free_kmem_cache_nodes(s);
3098error:
3099        if (flags & SLAB_PANIC)
3100                panic("Cannot create slab %s size=%lu realsize=%u "
3101                        "order=%u offset=%u flags=%lx\n",
3102                        s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3103                        s->offset, flags);
3104        return -EINVAL;
3105}
3106
3107static void list_slab_objects(struct kmem_cache *s, struct page *page,
3108                                                        const char *text)
3109{
3110#ifdef CONFIG_SLUB_DEBUG
3111        void *addr = page_address(page);
3112        void *p;
3113        unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3114                                     sizeof(long), GFP_ATOMIC);
3115        if (!map)
3116                return;
3117        slab_err(s, page, text, s->name);
3118        slab_lock(page);
3119
3120        get_map(s, page, map);
3121        for_each_object(p, s, addr, page->objects) {
3122
3123                if (!test_bit(slab_index(p, s, addr), map)) {
3124                        printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3125                                                        p, p - addr);
3126                        print_tracking(s, p);
3127                }
3128        }
3129        slab_unlock(page);
3130        kfree(map);
3131#endif
3132}
3133
3134/*
3135 * Attempt to free all partial slabs on a node.
3136 * This is called from kmem_cache_close(). We must be the last thread
3137 * using the cache and therefore we do not need to lock anymore.
3138 */
3139static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3140{
3141        struct page *page, *h;
3142
3143        list_for_each_entry_safe(page, h, &n->partial, lru) {
3144                if (!page->inuse) {
3145                        remove_partial(n, page);
3146                        discard_slab(s, page);
3147                } else {
3148                        list_slab_objects(s, page,
3149                        "Objects remaining in %s on kmem_cache_close()");
3150                }
3151        }
3152}
3153
3154/*
3155 * Release all resources used by a slab cache.
3156 */
3157static inline int kmem_cache_close(struct kmem_cache *s)
3158{
3159        int node;
3160
3161        flush_all(s);
3162        /* Attempt to free all objects */
3163        for_each_node_state(node, N_NORMAL_MEMORY) {
3164                struct kmem_cache_node *n = get_node(s, node);
3165
3166                free_partial(s, n);
3167                if (n->nr_partial || slabs_node(s, node))
3168                        return 1;
3169        }
3170        free_percpu(s->cpu_slab);
3171        free_kmem_cache_nodes(s);
3172        return 0;
3173}
3174
3175int __kmem_cache_shutdown(struct kmem_cache *s)
3176{
3177        int rc = kmem_cache_close(s);
3178
3179        if (!rc) {
3180                /*
3181                 * We do the same lock strategy around sysfs_slab_add, see
3182                 * __kmem_cache_create. Because this is pretty much the last
3183                 * operation we do and the lock will be released shortly after
3184                 * that in slab_common.c, we could just move sysfs_slab_remove
3185                 * to a later point in common code. We should do that when we
3186                 * have a common sysfs framework for all allocators.
3187                 */
3188                mutex_unlock(&slab_mutex);
3189                sysfs_slab_remove(s);
3190                mutex_lock(&slab_mutex);
3191        }
3192
3193        return rc;
3194}
3195
3196/********************************************************************
3197 *              Kmalloc subsystem
3198 *******************************************************************/
3199
3200static int __init setup_slub_min_order(char *str)
3201{
3202        get_option(&str, &slub_min_order);
3203
3204        return 1;
3205}
3206
3207__setup("slub_min_order=", setup_slub_min_order);
3208
3209static int __init setup_slub_max_order(char *str)
3210{
3211        get_option(&str, &slub_max_order);
3212        slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3213
3214        return 1;
3215}
3216
3217__setup("slub_max_order=", setup_slub_max_order);
3218
3219static int __init setup_slub_min_objects(char *str)
3220{
3221        get_option(&str, &slub_min_objects);
3222
3223        return 1;
3224}
3225
3226__setup("slub_min_objects=", setup_slub_min_objects);
3227
3228static int __init setup_slub_nomerge(char *str)
3229{
3230        slub_nomerge = 1;
3231        return 1;
3232}
3233
3234__setup("slub_nomerge", setup_slub_nomerge);
3235
3236void *__kmalloc(size_t size, gfp_t flags)
3237{
3238        struct kmem_cache *s;
3239        void *ret;
3240
3241        if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3242                return kmalloc_large(size, flags);
3243
3244        s = kmalloc_slab(size, flags);
3245
3246        if (unlikely(ZERO_OR_NULL_PTR(s)))
3247                return s;
3248
3249        ret = slab_alloc(s, flags, _RET_IP_);
3250
3251        trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3252
3253        return ret;
3254}
3255EXPORT_SYMBOL(__kmalloc);
3256
3257#ifdef CONFIG_NUMA
3258static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3259{
3260        struct page *page;
3261        void *ptr = NULL;
3262
3263        flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3264        page = alloc_pages_node(node, flags, get_order(size));
3265        if (page)
3266                ptr = page_address(page);
3267
3268        kmemleak_alloc(ptr, size, 1, flags);
3269        return ptr;
3270}
3271
3272void *__kmalloc_node(size_t size, gfp_t flags, int node)
3273{
3274        struct kmem_cache *s;
3275        void *ret;
3276
3277        if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3278                ret = kmalloc_large_node(size, flags, node);
3279
3280                trace_kmalloc_node(_RET_IP_, ret,
3281                                   size, PAGE_SIZE << get_order(size),
3282                                   flags, node);
3283
3284                return ret;
3285        }
3286
3287        s = kmalloc_slab(size, flags);
3288
3289        if (unlikely(ZERO_OR_NULL_PTR(s)))
3290                return s;
3291
3292        ret = slab_alloc_node(s, flags, node, _RET_IP_);
3293
3294        trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3295
3296        return ret;
3297}
3298EXPORT_SYMBOL(__kmalloc_node);
3299#endif
3300
3301size_t ksize(const void *object)
3302{
3303        struct page *page;
3304
3305        if (unlikely(object == ZERO_SIZE_PTR))
3306                return 0;
3307
3308        page = virt_to_head_page(object);
3309
3310        if (unlikely(!PageSlab(page))) {
3311                WARN_ON(!PageCompound(page));
3312                return PAGE_SIZE << compound_order(page);
3313        }
3314
3315        return slab_ksize(page->slab_cache);
3316}
3317EXPORT_SYMBOL(ksize);
3318
3319#ifdef CONFIG_SLUB_DEBUG
3320bool verify_mem_not_deleted(const void *x)
3321{
3322        struct page *page;
3323        void *object = (void *)x;
3324        unsigned long flags;
3325        bool rv;
3326
3327        if (unlikely(ZERO_OR_NULL_PTR(x)))
3328                return false;
3329
3330        local_irq_save(flags);
3331
3332        page = virt_to_head_page(x);
3333        if (unlikely(!PageSlab(page))) {
3334                /* maybe it was from stack? */
3335                rv = true;
3336                goto out_unlock;
3337        }
3338
3339        slab_lock(page);
3340        if (on_freelist(page->slab_cache, page, object)) {
3341                object_err(page->slab_cache, page, object, "Object is on free-list");
3342                rv = false;
3343        } else {
3344                rv = true;
3345        }
3346        slab_unlock(page);
3347
3348out_unlock:
3349        local_irq_restore(flags);
3350        return rv;
3351}
3352EXPORT_SYMBOL(verify_mem_not_deleted);
3353#endif
3354
3355void kfree(const void *x)
3356{
3357        struct page *page;
3358        void *object = (void *)x;
3359
3360        trace_kfree(_RET_IP_, x);
3361
3362        if (unlikely(ZERO_OR_NULL_PTR(x)))
3363                return;
3364
3365        page = virt_to_head_page(x);
3366        if (unlikely(!PageSlab(page))) {
3367                BUG_ON(!PageCompound(page));
3368                kmemleak_free(x);
3369                __free_memcg_kmem_pages(page, compound_order(page));
3370                return;
3371        }
3372        slab_free(page->slab_cache, page, object, _RET_IP_);
3373}
3374EXPORT_SYMBOL(kfree);
3375
3376/*
3377 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3378 * the remaining slabs by the number of items in use. The slabs with the
3379 * most items in use come first. New allocations will then fill those up
3380 * and thus they can be removed from the partial lists.
3381 *
3382 * The slabs with the least items are placed last. This results in them
3383 * being allocated from last increasing the chance that the last objects
3384 * are freed in them.
3385 */
3386int kmem_cache_shrink(struct kmem_cache *s)
3387{
3388        int node;
3389        int i;
3390        struct kmem_cache_node *n;
3391        struct page *page;
3392        struct page *t;
3393        int objects = oo_objects(s->max);
3394        struct list_head *slabs_by_inuse =
3395                kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3396        unsigned long flags;
3397
3398        if (!slabs_by_inuse)
3399                return -ENOMEM;
3400
3401        flush_all(s);
3402        for_each_node_state(node, N_NORMAL_MEMORY) {
3403                n = get_node(s, node);
3404
3405                if (!n->nr_partial)
3406                        continue;
3407
3408                for (i = 0; i < objects; i++)
3409                        INIT_LIST_HEAD(slabs_by_inuse + i);
3410
3411                spin_lock_irqsave(&n->list_lock, flags);
3412
3413                /*
3414                 * Build lists indexed by the items in use in each slab.
3415                 *
3416                 * Note that concurrent frees may occur while we hold the
3417                 * list_lock. page->inuse here is the upper limit.
3418                 */
3419                list_for_each_entry_safe(page, t, &n->partial, lru) {
3420                        list_move(&page->lru, slabs_by_inuse + page->inuse);
3421                        if (!page->inuse)
3422                                n->nr_partial--;
3423                }
3424
3425                /*
3426                 * Rebuild the partial list with the slabs filled up most
3427                 * first and the least used slabs at the end.
3428                 */
3429                for (i = objects - 1; i > 0; i--)
3430                        list_splice(slabs_by_inuse + i, n->partial.prev);
3431
3432                spin_unlock_irqrestore(&n->list_lock, flags);
3433
3434                /* Release empty slabs */
3435                list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3436                        discard_slab(s, page);
3437        }
3438
3439        kfree(slabs_by_inuse);
3440        return 0;
3441}
3442EXPORT_SYMBOL(kmem_cache_shrink);
3443
3444static int slab_mem_going_offline_callback(void *arg)
3445{
3446        struct kmem_cache *s;
3447
3448        mutex_lock(&slab_mutex);
3449        list_for_each_entry(s, &slab_caches, list)
3450                kmem_cache_shrink(s);
3451        mutex_unlock(&slab_mutex);
3452
3453        return 0;
3454}
3455
3456static void slab_mem_offline_callback(void *arg)
3457{
3458        struct kmem_cache_node *n;
3459        struct kmem_cache *s;
3460        struct memory_notify *marg = arg;
3461        int offline_node;
3462
3463        offline_node = marg->status_change_nid_normal;
3464
3465        /*
3466         * If the node still has available memory. we need kmem_cache_node
3467         * for it yet.
3468         */
3469        if (offline_node < 0)
3470                return;
3471
3472        mutex_lock(&slab_mutex);
3473        list_for_each_entry(s, &slab_caches, list) {
3474                n = get_node(s, offline_node);
3475                if (n) {
3476                        /*
3477                         * if n->nr_slabs > 0, slabs still exist on the node
3478                         * that is going down. We were unable to free them,
3479                         * and offline_pages() function shouldn't call this
3480                         * callback. So, we must fail.
3481                         */
3482                        BUG_ON(slabs_node(s, offline_node));
3483
3484                        s->node[offline_node] = NULL;
3485                        kmem_cache_free(kmem_cache_node, n);
3486                }
3487        }
3488        mutex_unlock(&slab_mutex);
3489}
3490
3491static int slab_mem_going_online_callback(void *arg)
3492{
3493        struct kmem_cache_node *n;
3494        struct kmem_cache *s;
3495        struct memory_notify *marg = arg;
3496        int nid = marg->status_change_nid_normal;
3497        int ret = 0;
3498
3499        /*
3500         * If the node's memory is already available, then kmem_cache_node is
3501         * already created. Nothing to do.
3502         */
3503        if (nid < 0)
3504                return 0;
3505
3506        /*
3507         * We are bringing a node online. No memory is available yet. We must
3508         * allocate a kmem_cache_node structure in order to bring the node
3509         * online.
3510         */
3511        mutex_lock(&slab_mutex);
3512        list_for_each_entry(s, &slab_caches, list) {
3513                /*
3514                 * XXX: kmem_cache_alloc_node will fallback to other nodes
3515                 *      since memory is not yet available from the node that
3516                 *      is brought up.
3517                 */
3518                n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3519                if (!n) {
3520                        ret = -ENOMEM;
3521                        goto out;
3522                }
3523                init_kmem_cache_node(n);
3524                s->node[nid] = n;
3525        }
3526out:
3527        mutex_unlock(&slab_mutex);
3528        return ret;
3529}
3530
3531static int slab_memory_callback(struct notifier_block *self,
3532                                unsigned long action, void *arg)
3533{
3534        int ret = 0;
3535
3536        switch (action) {
3537        case MEM_GOING_ONLINE:
3538                ret = slab_mem_going_online_callback(arg);
3539                break;
3540        case MEM_GOING_OFFLINE:
3541                ret = slab_mem_going_offline_callback(arg);
3542                break;
3543        case MEM_OFFLINE:
3544        case MEM_CANCEL_ONLINE:
3545                slab_mem_offline_callback(arg);
3546                break;
3547        case MEM_ONLINE:
3548        case MEM_CANCEL_OFFLINE:
3549                break;
3550        }
3551        if (ret)
3552                ret = notifier_from_errno(ret);
3553        else
3554                ret = NOTIFY_OK;
3555        return ret;
3556}
3557
3558static struct notifier_block slab_memory_callback_nb = {
3559        .notifier_call = slab_memory_callback,
3560        .priority = SLAB_CALLBACK_PRI,
3561};
3562
3563/********************************************************************
3564 *                      Basic setup of slabs
3565 *******************************************************************/
3566
3567/*
3568 * Used for early kmem_cache structures that were allocated using
3569 * the page allocator. Allocate them properly then fix up the pointers
3570 * that may be pointing to the wrong kmem_cache structure.
3571 */
3572
3573static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3574{
3575        int node;
3576        struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3577
3578        memcpy(s, static_cache, kmem_cache->object_size);
3579
3580        /*
3581         * This runs very early, and only the boot processor is supposed to be
3582         * up.  Even if it weren't true, IRQs are not up so we couldn't fire
3583         * IPIs around.
3584         */
3585        __flush_cpu_slab(s, smp_processor_id());
3586        for_each_node_state(node, N_NORMAL_MEMORY) {
3587                struct kmem_cache_node *n = get_node(s, node);
3588                struct page *p;
3589
3590                if (n) {
3591                        list_for_each_entry(p, &n->partial, lru)
3592                                p->slab_cache = s;
3593
3594#ifdef CONFIG_SLUB_DEBUG
3595                        list_for_each_entry(p, &n->full, lru)
3596                                p->slab_cache = s;
3597#endif
3598                }
3599        }
3600        list_add(&s->list, &slab_caches);
3601        return s;
3602}
3603
3604void __init kmem_cache_init(void)
3605{
3606        static __initdata struct kmem_cache boot_kmem_cache,
3607                boot_kmem_cache_node;
3608
3609        if (debug_guardpage_minorder())
3610                slub_max_order = 0;
3611
3612        kmem_cache_node = &boot_kmem_cache_node;
3613        kmem_cache = &boot_kmem_cache;
3614
3615        create_boot_cache(kmem_cache_node, "kmem_cache_node",
3616                sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3617
3618        register_hotmemory_notifier(&slab_memory_callback_nb);
3619
3620        /* Able to allocate the per node structures */
3621        slab_state = PARTIAL;
3622
3623        create_boot_cache(kmem_cache, "kmem_cache",
3624                        offsetof(struct kmem_cache, node) +
3625                                nr_node_ids * sizeof(struct kmem_cache_node *),
3626                       SLAB_HWCACHE_ALIGN);
3627
3628        kmem_cache = bootstrap(&boot_kmem_cache);
3629
3630        /*
3631         * Allocate kmem_cache_node properly from the kmem_cache slab.
3632         * kmem_cache_node is separately allocated so no need to
3633         * update any list pointers.
3634         */
3635        kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3636
3637        /* Now we can use the kmem_cache to allocate kmalloc slabs */
3638        create_kmalloc_caches(0);
3639
3640#ifdef CONFIG_SMP
3641        register_cpu_notifier(&slab_notifier);
3642#endif
3643
3644        printk(KERN_INFO
3645                "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3646                " CPUs=%d, Nodes=%d\n",
3647                cache_line_size(),
3648                slub_min_order, slub_max_order, slub_min_objects,
3649                nr_cpu_ids, nr_node_ids);
3650}
3651
3652void __init kmem_cache_init_late(void)
3653{
3654}
3655
3656/*
3657 * Find a mergeable slab cache
3658 */
3659static int slab_unmergeable(struct kmem_cache *s)
3660{
3661        if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3662                return 1;
3663
3664        if (s->ctor)
3665                return 1;
3666
3667        /*
3668         * We may have set a slab to be unmergeable during bootstrap.
3669         */
3670        if (s->refcount < 0)
3671                return 1;
3672
3673        return 0;
3674}
3675
3676static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
3677                size_t align, unsigned long flags, const char *name,
3678                void (*ctor)(void *))
3679{
3680        struct kmem_cache *s;
3681
3682        if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3683                return NULL;
3684
3685        if (ctor)
3686                return NULL;
3687
3688        size = ALIGN(size, sizeof(void *));
3689        align = calculate_alignment(flags, align, size);
3690        size = ALIGN(size, align);
3691        flags = kmem_cache_flags(size, flags, name, NULL);
3692
3693        list_for_each_entry(s, &slab_caches, list) {
3694                if (slab_unmergeable(s))
3695                        continue;
3696
3697                if (size > s->size)
3698                        continue;
3699
3700                if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3701                                continue;
3702                /*
3703                 * Check if alignment is compatible.
3704                 * Courtesy of Adrian Drzewiecki
3705                 */
3706                if ((s->size & ~(align - 1)) != s->size)
3707                        continue;
3708
3709                if (s->size - size >= sizeof(void *))
3710                        continue;
3711
3712                if (!cache_match_memcg(s, memcg))
3713                        continue;
3714
3715                return s;
3716        }
3717        return NULL;
3718}
3719
3720struct kmem_cache *
3721__kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3722                   size_t align, unsigned long flags, void (*ctor)(void *))
3723{
3724        struct kmem_cache *s;
3725
3726        s = find_mergeable(memcg, size, align, flags, name, ctor);
3727        if (s) {
3728                s->refcount++;
3729                /*
3730                 * Adjust the object sizes so that we clear
3731                 * the complete object on kzalloc.
3732                 */
3733                s->object_size = max(s->object_size, (int)size);
3734                s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3735
3736                if (sysfs_slab_alias(s, name)) {
3737                        s->refcount--;
3738                        s = NULL;
3739                }
3740        }
3741
3742        return s;
3743}
3744
3745int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3746{
3747        int err;
3748
3749        err = kmem_cache_open(s, flags);
3750        if (err)
3751                return err;
3752
3753        /* Mutex is not taken during early boot */
3754        if (slab_state <= UP)
3755                return 0;
3756
3757        memcg_propagate_slab_attrs(s);
3758        mutex_unlock(&slab_mutex);
3759        err = sysfs_slab_add(s);
3760        mutex_lock(&slab_mutex);
3761
3762        if (err)
3763                kmem_cache_close(s);
3764
3765        return err;
3766}
3767
3768#ifdef CONFIG_SMP
3769/*
3770 * Use the cpu notifier to insure that the cpu slabs are flushed when
3771 * necessary.
3772 */
3773static int slab_cpuup_callback(struct notifier_block *nfb,
3774                unsigned long action, void *hcpu)
3775{
3776        long cpu = (long)hcpu;
3777        struct kmem_cache *s;
3778        unsigned long flags;
3779
3780        switch (action) {
3781        case CPU_UP_CANCELED:
3782        case CPU_UP_CANCELED_FROZEN:
3783        case CPU_DEAD:
3784        case CPU_DEAD_FROZEN:
3785                mutex_lock(&slab_mutex);
3786                list_for_each_entry(s, &slab_caches, list) {
3787                        local_irq_save(flags);
3788                        __flush_cpu_slab(s, cpu);
3789                        local_irq_restore(flags);
3790                }
3791                mutex_unlock(&slab_mutex);
3792                break;
3793        default:
3794                break;
3795        }
3796        return NOTIFY_OK;
3797}
3798
3799static struct notifier_block slab_notifier = {
3800        .notifier_call = slab_cpuup_callback
3801};
3802
3803#endif
3804
3805void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3806{
3807        struct kmem_cache *s;
3808        void *ret;
3809
3810        if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3811                return kmalloc_large(size, gfpflags);
3812
3813        s = kmalloc_slab(size, gfpflags);
3814
3815        if (unlikely(ZERO_OR_NULL_PTR(s)))
3816                return s;
3817
3818        ret = slab_alloc(s, gfpflags, caller);
3819
3820        /* Honor the call site pointer we received. */
3821        trace_kmalloc(caller, ret, size, s->size, gfpflags);
3822
3823        return ret;
3824}
3825
3826#ifdef CONFIG_NUMA
3827void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3828                                        int node, unsigned long caller)
3829{
3830        struct kmem_cache *s;
3831        void *ret;
3832
3833        if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3834                ret = kmalloc_large_node(size, gfpflags, node);
3835
3836                trace_kmalloc_node(caller, ret,
3837                                   size, PAGE_SIZE << get_order(size),
3838                                   gfpflags, node);
3839
3840                return ret;
3841        }
3842
3843        s = kmalloc_slab(size, gfpflags);
3844
3845        if (unlikely(ZERO_OR_NULL_PTR(s)))
3846                return s;
3847
3848        ret = slab_alloc_node(s, gfpflags, node, caller);
3849
3850        /* Honor the call site pointer we received. */
3851        trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3852
3853        return ret;
3854}
3855#endif
3856
3857#ifdef CONFIG_SYSFS
3858static int count_inuse(struct page *page)
3859{
3860        return page->inuse;
3861}
3862
3863static int count_total(struct page *page)
3864{
3865        return page->objects;
3866}
3867#endif
3868
3869#ifdef CONFIG_SLUB_DEBUG
3870static int validate_slab(struct kmem_cache *s, struct page *page,
3871                                                unsigned long *map)
3872{
3873        void *p;
3874        void *addr = page_address(page);
3875
3876        if (!check_slab(s, page) ||
3877                        !on_freelist(s, page, NULL))
3878                return 0;
3879
3880        /* Now we know that a valid freelist exists */
3881        bitmap_zero(map, page->objects);
3882
3883        get_map(s, page, map);
3884        for_each_object(p, s, addr, page->objects) {
3885                if (test_bit(slab_index(p, s, addr), map))
3886                        if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3887                                return 0;
3888        }
3889
3890        for_each_object(p, s, addr, page->objects)
3891                if (!test_bit(slab_index(p, s, addr), map))
3892                        if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3893                                return 0;
3894        return 1;
3895}
3896
3897static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3898                                                unsigned long *map)
3899{
3900        slab_lock(page);
3901        validate_slab(s, page, map);
3902        slab_unlock(page);
3903}
3904
3905static int validate_slab_node(struct kmem_cache *s,
3906                struct kmem_cache_node *n, unsigned long *map)
3907{
3908        unsigned long count = 0;
3909        struct page *page;
3910        unsigned long flags;
3911
3912        spin_lock_irqsave(&n->list_lock, flags);
3913
3914        list_for_each_entry(page, &n->partial, lru) {
3915                validate_slab_slab(s, page, map);
3916                count++;
3917        }
3918        if (count != n->nr_partial)
3919                printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3920                        "counter=%ld\n", s->name, count, n->nr_partial);
3921
3922        if (!(s->flags & SLAB_STORE_USER))
3923                goto out;
3924
3925        list_for_each_entry(page, &n->full, lru) {
3926                validate_slab_slab(s, page, map);
3927                count++;
3928        }
3929        if (count != atomic_long_read(&n->nr_slabs))
3930                printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3931                        "counter=%ld\n", s->name, count,
3932                        atomic_long_read(&n->nr_slabs));
3933
3934out:
3935        spin_unlock_irqrestore(&n->list_lock, flags);
3936        return count;
3937}
3938
3939static long validate_slab_cache(struct kmem_cache *s)
3940{
3941        int node;
3942        unsigned long count = 0;
3943        unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3944                                sizeof(unsigned long), GFP_KERNEL);
3945
3946        if (!map)
3947                return -ENOMEM;
3948
3949        flush_all(s);
3950        for_each_node_state(node, N_NORMAL_MEMORY) {
3951                struct kmem_cache_node *n = get_node(s, node);
3952
3953                count += validate_slab_node(s, n, map);
3954        }
3955        kfree(map);
3956        return count;
3957}
3958/*
3959 * Generate lists of code addresses where slabcache objects are allocated
3960 * and freed.
3961 */
3962
3963struct location {
3964        unsigned long count;
3965        unsigned long addr;
3966        long long sum_time;
3967        long min_time;
3968        long max_time;
3969        long min_pid;
3970        long max_pid;
3971        DECLARE_BITMAP(cpus, NR_CPUS);
3972        nodemask_t nodes;
3973};
3974
3975struct loc_track {
3976        unsigned long max;
3977        unsigned long count;
3978        struct location *loc;
3979};
3980
3981static void free_loc_track(struct loc_track *t)
3982{
3983        if (t->max)
3984                free_pages((unsigned long)t->loc,
3985                        get_order(sizeof(struct location) * t->max));
3986}
3987
3988static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3989{
3990        struct location *l;
3991        int order;
3992
3993        order = get_order(sizeof(struct location) * max);
3994
3995        l = (void *)__get_free_pages(flags, order);
3996        if (!l)
3997                return 0;
3998
3999        if (t->count) {
4000                memcpy(l, t->loc, sizeof(struct location) * t->count);
4001                free_loc_track(t);
4002        }
4003        t->max = max;
4004        t->loc = l;
4005        return 1;
4006}
4007
4008static int add_location(struct loc_track *t, struct kmem_cache *s,
4009                                const struct track *track)
4010{
4011        long start, end, pos;
4012        struct location *l;
4013        unsigned long caddr;
4014        unsigned long age = jiffies - track->when;
4015
4016        start = -1;
4017        end = t->count;
4018
4019        for ( ; ; ) {
4020                pos = start + (end - start + 1) / 2;
4021
4022                /*
4023                 * There is nothing at "end". If we end up there
4024                 * we need to add something to before end.
4025                 */
4026                if (pos == end)
4027                        break;
4028
4029                caddr = t->loc[pos].addr;
4030                if (track->addr == caddr) {
4031
4032                        l = &t->loc[pos];
4033                        l->count++;
4034                        if (track->when) {
4035                                l->sum_time += age;
4036                                if (age < l->min_time)
4037                                        l->min_time = age;
4038                                if (age > l->max_time)
4039                                        l->max_time = age;
4040
4041                                if (track->pid < l->min_pid)
4042                                        l->min_pid = track->pid;
4043                                if (track->pid > l->max_pid)
4044                                        l->max_pid = track->pid;
4045
4046                                cpumask_set_cpu(track->cpu,
4047                                                to_cpumask(l->cpus));
4048                        }
4049                        node_set(page_to_nid(virt_to_page(track)), l->nodes);
4050                        return 1;
4051                }
4052
4053                if (track->addr < caddr)
4054                        end = pos;
4055                else
4056                        start = pos;
4057        }
4058
4059        /*
4060         * Not found. Insert new tracking element.
4061         */
4062        if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4063                return 0;
4064
4065        l = t->loc + pos;
4066        if (pos < t->count)
4067                memmove(l + 1, l,
4068                        (t->count - pos) * sizeof(struct location));
4069        t->count++;
4070        l->count = 1;
4071        l->addr = track->addr;
4072        l->sum_time = age;
4073        l->min_time = age;
4074        l->max_time = age;
4075        l->min_pid = track->pid;
4076        l->max_pid = track->pid;
4077        cpumask_clear(to_cpumask(l->cpus));
4078        cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4079        nodes_clear(l->nodes);
4080        node_set(page_to_nid(virt_to_page(track)), l->nodes);
4081        return 1;
4082}
4083
4084static void process_slab(struct loc_track *t, struct kmem_cache *s,
4085                struct page *page, enum track_item alloc,
4086                unsigned long *map)
4087{
4088        void *addr = page_address(page);
4089        void *p;
4090
4091        bitmap_zero(map, page->objects);
4092        get_map(s, page, map);
4093
4094        for_each_object(p, s, addr, page->objects)
4095                if (!test_bit(slab_index(p, s, addr), map))
4096                        add_location(t, s, get_track(s, p, alloc));
4097}
4098
4099static int list_locations(struct kmem_cache *s, char *buf,
4100                                        enum track_item alloc)
4101{
4102        int len = 0;
4103        unsigned long i;
4104        struct loc_track t = { 0, 0, NULL };
4105        int node;
4106        unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4107                                     sizeof(unsigned long), GFP_KERNEL);
4108
4109        if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4110                                     GFP_TEMPORARY)) {
4111                kfree(map);
4112                return sprintf(buf, "Out of memory\n");
4113        }
4114        /* Push back cpu slabs */
4115        flush_all(s);
4116
4117        for_each_node_state(node, N_NORMAL_MEMORY) {
4118                struct kmem_cache_node *n = get_node(s, node);
4119                unsigned long flags;
4120                struct page *page;
4121
4122                if (!atomic_long_read(&n->nr_slabs))
4123                        continue;
4124
4125                spin_lock_irqsave(&n->list_lock, flags);
4126                list_for_each_entry(page, &n->partial, lru)
4127                        process_slab(&t, s, page, alloc, map);
4128                list_for_each_entry(page, &n->full, lru)
4129                        process_slab(&t, s, page, alloc, map);
4130                spin_unlock_irqrestore(&n->list_lock, flags);
4131        }
4132
4133        for (i = 0; i < t.count; i++) {
4134                struct location *l = &t.loc[i];
4135
4136                if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4137                        break;
4138                len += sprintf(buf + len, "%7ld ", l->count);
4139
4140                if (l->addr)
4141                        len += sprintf(buf + len, "%pS", (void *)l->addr);
4142                else
4143                        len += sprintf(buf + len, "<not-available>");
4144
4145                if (l->sum_time != l->min_time) {
4146                        len += sprintf(buf + len, " age=%ld/%ld/%ld",
4147                                l->min_time,
4148                                (long)div_u64(l->sum_time, l->count),
4149                                l->max_time);
4150                } else
4151                        len += sprintf(buf + len, " age=%ld",
4152                                l->min_time);
4153
4154                if (l->min_pid != l->max_pid)
4155                        len += sprintf(buf + len, " pid=%ld-%ld",
4156                                l->min_pid, l->max_pid);
4157                else
4158                        len += sprintf(buf + len, " pid=%ld",
4159                                l->min_pid);
4160
4161                if (num_online_cpus() > 1 &&
4162                                !cpumask_empty(to_cpumask(l->cpus)) &&
4163                                len < PAGE_SIZE - 60) {
4164                        len += sprintf(buf + len, " cpus=");
4165                        len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4166                                                 to_cpumask(l->cpus));
4167                }
4168
4169                if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4170                                len < PAGE_SIZE - 60) {
4171                        len += sprintf(buf + len, " nodes=");
4172                        len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4173                                        l->nodes);
4174                }
4175
4176                len += sprintf(buf + len, "\n");
4177        }
4178
4179        free_loc_track(&t);
4180        kfree(map);
4181        if (!t.count)
4182                len += sprintf(buf, "No data\n");
4183        return len;
4184}
4185#endif
4186
4187#ifdef SLUB_RESILIENCY_TEST
4188static void resiliency_test(void)
4189{
4190        u8 *p;
4191
4192        BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4193
4194        printk(KERN_ERR "SLUB resiliency testing\n");
4195        printk(KERN_ERR "-----------------------\n");
4196        printk(KERN_ERR "A. Corruption after allocation\n");
4197
4198        p = kzalloc(16, GFP_KERNEL);
4199        p[16] = 0x12;
4200        printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4201                        " 0x12->0x%p\n\n", p + 16);
4202
4203        validate_slab_cache(kmalloc_caches[4]);
4204
4205        /* Hmmm... The next two are dangerous */
4206        p = kzalloc(32, GFP_KERNEL);
4207        p[32 + sizeof(void *)] = 0x34;
4208        printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4209                        " 0x34 -> -0x%p\n", p);
4210        printk(KERN_ERR
4211                "If allocated object is overwritten then not detectable\n\n");
4212
4213        validate_slab_cache(kmalloc_caches[5]);
4214        p = kzalloc(64, GFP_KERNEL);
4215        p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4216        *p = 0x56;
4217        printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4218                                                                        p);
4219        printk(KERN_ERR
4220                "If allocated object is overwritten then not detectable\n\n");
4221        validate_slab_cache(kmalloc_caches[6]);
4222
4223        printk(KERN_ERR "\nB. Corruption after free\n");
4224        p = kzalloc(128, GFP_KERNEL);
4225        kfree(p);
4226        *p = 0x78;
4227        printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4228        validate_slab_cache(kmalloc_caches[7]);
4229
4230        p = kzalloc(256, GFP_KERNEL);
4231        kfree(p);
4232        p[50] = 0x9a;
4233        printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4234                        p);
4235        validate_slab_cache(kmalloc_caches[8]);
4236
4237        p = kzalloc(512, GFP_KERNEL);
4238        kfree(p);
4239        p[512] = 0xab;
4240        printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4241        validate_slab_cache(kmalloc_caches[9]);
4242}
4243#else
4244#ifdef CONFIG_SYSFS
4245static void resiliency_test(void) {};
4246#endif
4247#endif
4248
4249#ifdef CONFIG_SYSFS
4250enum slab_stat_type {
4251        SL_ALL,                 /* All slabs */
4252        SL_PARTIAL,             /* Only partially allocated slabs */
4253        SL_CPU,                 /* Only slabs used for cpu caches */
4254        SL_OBJECTS,             /* Determine allocated objects not slabs */
4255        SL_TOTAL                /* Determine object capacity not slabs */
4256};
4257
4258#define SO_ALL          (1 << SL_ALL)
4259#define SO_PARTIAL      (1 << SL_PARTIAL)
4260#define SO_CPU          (1 << SL_CPU)
4261#define SO_OBJECTS      (1 << SL_OBJECTS)
4262#define SO_TOTAL        (1 << SL_TOTAL)
4263
4264static ssize_t show_slab_objects(struct kmem_cache *s,
4265                            char *buf, unsigned long flags)
4266{
4267        unsigned long total = 0;
4268        int node;
4269        int x;
4270        unsigned long *nodes;
4271        unsigned long *per_cpu;
4272
4273        nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4274        if (!nodes)
4275                return -ENOMEM;
4276        per_cpu = nodes + nr_node_ids;
4277
4278        if (flags & SO_CPU) {
4279                int cpu;
4280
4281                for_each_possible_cpu(cpu) {
4282                        struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4283                        int node;
4284                        struct page *page;
4285
4286                        page = ACCESS_ONCE(c->page);
4287                        if (!page)
4288                                continue;
4289
4290                        node = page_to_nid(page);
4291                        if (flags & SO_TOTAL)
4292                                x = page->objects;
4293                        else if (flags & SO_OBJECTS)
4294                                x = page->inuse;
4295                        else
4296                                x = 1;
4297
4298                        total += x;
4299                        nodes[node] += x;
4300
4301                        page = ACCESS_ONCE(c->partial);
4302                        if (page) {
4303                                x = page->pobjects;
4304                                total += x;
4305                                nodes[node] += x;
4306                        }
4307
4308                        per_cpu[node]++;
4309                }
4310        }
4311
4312        lock_memory_hotplug();
4313#ifdef CONFIG_SLUB_DEBUG
4314        if (flags & SO_ALL) {
4315                for_each_node_state(node, N_NORMAL_MEMORY) {
4316                        struct kmem_cache_node *n = get_node(s, node);
4317
4318                if (flags & SO_TOTAL)
4319                        x = atomic_long_read(&n->total_objects);
4320                else if (flags & SO_OBJECTS)
4321                        x = atomic_long_read(&n->total_objects) -
4322                                count_partial(n, count_free);
4323
4324                        else
4325                                x = atomic_long_read(&n->nr_slabs);
4326                        total += x;
4327                        nodes[node] += x;
4328                }
4329
4330        } else
4331#endif
4332        if (flags & SO_PARTIAL) {
4333                for_each_node_state(node, N_NORMAL_MEMORY) {
4334                        struct kmem_cache_node *n = get_node(s, node);
4335
4336                        if (flags & SO_TOTAL)
4337                                x = count_partial(n, count_total);
4338                        else if (flags & SO_OBJECTS)
4339                                x = count_partial(n, count_inuse);
4340                        else
4341                                x = n->nr_partial;
4342                        total += x;
4343                        nodes[node] += x;
4344                }
4345        }
4346        x = sprintf(buf, "%lu", total);
4347#ifdef CONFIG_NUMA
4348        for_each_node_state(node, N_NORMAL_MEMORY)
4349                if (nodes[node])
4350                        x += sprintf(buf + x, " N%d=%lu",
4351                                        node, nodes[node]);
4352#endif
4353        unlock_memory_hotplug();
4354        kfree(nodes);
4355        return x + sprintf(buf + x, "\n");
4356}
4357
4358#ifdef CONFIG_SLUB_DEBUG
4359static int any_slab_objects(struct kmem_cache *s)
4360{
4361        int node;
4362
4363        for_each_online_node(node) {
4364                struct kmem_cache_node *n = get_node(s, node);
4365
4366                if (!n)
4367                        continue;
4368
4369                if (atomic_long_read(&n->total_objects))
4370                        return 1;
4371        }
4372        return 0;
4373}
4374#endif
4375
4376#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4377#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4378
4379struct slab_attribute {
4380        struct attribute attr;
4381        ssize_t (*show)(struct kmem_cache *s, char *buf);
4382        ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4383};
4384
4385#define SLAB_ATTR_RO(_name) \
4386        static struct slab_attribute _name##_attr = \
4387        __ATTR(_name, 0400, _name##_show, NULL)
4388
4389#define SLAB_ATTR(_name) \
4390        static struct slab_attribute _name##_attr =  \
4391        __ATTR(_name, 0600, _name##_show, _name##_store)
4392
4393static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4394{
4395        return sprintf(buf, "%d\n", s->size);
4396}
4397SLAB_ATTR_RO(slab_size);
4398
4399static ssize_t align_show(struct kmem_cache *s, char *buf)
4400{
4401        return sprintf(buf, "%d\n", s->align);
4402}
4403SLAB_ATTR_RO(align);
4404
4405static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4406{
4407        return sprintf(buf, "%d\n", s->object_size);
4408}
4409SLAB_ATTR_RO(object_size);
4410
4411static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4412{
4413        return sprintf(buf, "%d\n", oo_objects(s->oo));
4414}
4415SLAB_ATTR_RO(objs_per_slab);
4416
4417static ssize_t order_store(struct kmem_cache *s,
4418                                const char *buf, size_t length)
4419{
4420        unsigned long order;
4421        int err;
4422
4423        err = strict_strtoul(buf, 10, &order);
4424        if (err)
4425                return err;
4426
4427        if (order > slub_max_order || order < slub_min_order)
4428                return -EINVAL;
4429
4430        calculate_sizes(s, order);
4431        return length;
4432}
4433
4434static ssize_t order_show(struct kmem_cache *s, char *buf)
4435{
4436        return sprintf(buf, "%d\n", oo_order(s->oo));
4437}
4438SLAB_ATTR(order);
4439
4440static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4441{
4442        return sprintf(buf, "%lu\n", s->min_partial);
4443}
4444
4445static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4446                                 size_t length)
4447{
4448        unsigned long min;
4449        int err;
4450
4451        err = strict_strtoul(buf, 10, &min);
4452        if (err)
4453                return err;
4454
4455        set_min_partial(s, min);
4456        return length;
4457}
4458SLAB_ATTR(min_partial);
4459
4460static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4461{
4462        return sprintf(buf, "%u\n", s->cpu_partial);
4463}
4464
4465static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4466                                 size_t length)
4467{
4468        unsigned long objects;
4469        int err;
4470
4471        err = strict_strtoul(buf, 10, &objects);
4472        if (err)
4473                return err;
4474        if (objects && !kmem_cache_has_cpu_partial(s))
4475                return -EINVAL;
4476
4477        s->cpu_partial = objects;
4478        flush_all(s);
4479        return length;
4480}
4481SLAB_ATTR(cpu_partial);
4482
4483static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4484{
4485        if (!s->ctor)
4486                return 0;
4487        return sprintf(buf, "%pS\n", s->ctor);
4488}
4489SLAB_ATTR_RO(ctor);
4490
4491static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4492{
4493        return sprintf(buf, "%d\n", s->refcount - 1);
4494}
4495SLAB_ATTR_RO(aliases);
4496
4497static ssize_t partial_show(struct kmem_cache *s, char *buf)
4498{
4499        return show_slab_objects(s, buf, SO_PARTIAL);
4500}
4501SLAB_ATTR_RO(partial);
4502
4503static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4504{
4505        return show_slab_objects(s, buf, SO_CPU);
4506}
4507SLAB_ATTR_RO(cpu_slabs);
4508
4509static ssize_t objects_show(struct kmem_cache *s, char *buf)
4510{
4511        return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4512}
4513SLAB_ATTR_RO(objects);
4514
4515static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4516{
4517        return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4518}
4519SLAB_ATTR_RO(objects_partial);
4520
4521static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4522{
4523        int objects = 0;
4524        int pages = 0;
4525        int cpu;
4526        int len;
4527
4528        for_each_online_cpu(cpu) {
4529                struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4530
4531                if (page) {
4532                        pages += page->pages;
4533                        objects += page->pobjects;
4534                }
4535        }
4536
4537        len = sprintf(buf, "%d(%d)", objects, pages);
4538
4539#ifdef CONFIG_SMP
4540        for_each_online_cpu(cpu) {
4541                struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4542
4543                if (page && len < PAGE_SIZE - 20)
4544                        len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4545                                page->pobjects, page->pages);
4546        }
4547#endif
4548        return len + sprintf(buf + len, "\n");
4549}
4550SLAB_ATTR_RO(slabs_cpu_partial);
4551
4552static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4553{
4554        return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4555}
4556
4557static ssize_t reclaim_account_store(struct kmem_cache *s,
4558                                const char *buf, size_t length)
4559{
4560        s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4561        if (buf[0] == '1')
4562                s->flags |= SLAB_RECLAIM_ACCOUNT;
4563        return length;
4564}
4565SLAB_ATTR(reclaim_account);
4566
4567static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4568{
4569        return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4570}
4571SLAB_ATTR_RO(hwcache_align);
4572
4573#ifdef CONFIG_ZONE_DMA
4574static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4575{
4576        return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4577}
4578SLAB_ATTR_RO(cache_dma);
4579#endif
4580
4581static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4582{
4583        return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4584}
4585SLAB_ATTR_RO(destroy_by_rcu);
4586
4587static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4588{
4589        return sprintf(buf, "%d\n", s->reserved);
4590}
4591SLAB_ATTR_RO(reserved);
4592
4593#ifdef CONFIG_SLUB_DEBUG
4594static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4595{
4596        return show_slab_objects(s, buf, SO_ALL);
4597}
4598SLAB_ATTR_RO(slabs);
4599
4600static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4601{
4602        return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4603}
4604SLAB_ATTR_RO(total_objects);
4605
4606static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4607{
4608        return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4609}
4610
4611static ssize_t sanity_checks_store(struct kmem_cache *s,
4612                                const char *buf, size_t length)
4613{
4614        s->flags &= ~SLAB_DEBUG_FREE;
4615        if (buf[0] == '1') {
4616                s->flags &= ~__CMPXCHG_DOUBLE;
4617                s->flags |= SLAB_DEBUG_FREE;
4618        }
4619        return length;
4620}
4621SLAB_ATTR(sanity_checks);
4622
4623static ssize_t trace_show(struct kmem_cache *s, char *buf)
4624{
4625        return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4626}
4627
4628static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4629                                                        size_t length)
4630{
4631        s->flags &= ~SLAB_TRACE;
4632        if (buf[0] == '1') {
4633                s->flags &= ~__CMPXCHG_DOUBLE;
4634                s->flags |= SLAB_TRACE;
4635        }
4636        return length;
4637}
4638SLAB_ATTR(trace);
4639
4640static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4641{
4642        return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4643}
4644
4645static ssize_t red_zone_store(struct kmem_cache *s,
4646                                const char *buf, size_t length)
4647{
4648        if (any_slab_objects(s))
4649                return -EBUSY;
4650
4651        s->flags &= ~SLAB_RED_ZONE;
4652        if (buf[0] == '1') {
4653                s->flags &= ~__CMPXCHG_DOUBLE;
4654                s->flags |= SLAB_RED_ZONE;
4655        }
4656        calculate_sizes(s, -1);
4657        return length;
4658}
4659SLAB_ATTR(red_zone);
4660
4661static ssize_t poison_show(struct kmem_cache *s, char *buf)
4662{
4663        return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4664}
4665
4666static ssize_t poison_store(struct kmem_cache *s,
4667                                const char *buf, size_t length)
4668{
4669        if (any_slab_objects(s))
4670                return -EBUSY;
4671
4672        s->flags &= ~SLAB_POISON;
4673        if (buf[0] == '1') {
4674                s->flags &= ~__CMPXCHG_DOUBLE;
4675                s->flags |= SLAB_POISON;
4676        }
4677        calculate_sizes(s, -1);
4678        return length;
4679}
4680SLAB_ATTR(poison);
4681
4682static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4683{
4684        return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4685}
4686
4687static ssize_t store_user_store(struct kmem_cache *s,
4688                                const char *buf, size_t length)
4689{
4690        if (any_slab_objects(s))
4691                return -EBUSY;
4692
4693        s->flags &= ~SLAB_STORE_USER;
4694        if (buf[0] == '1') {
4695                s->flags &= ~__CMPXCHG_DOUBLE;
4696                s->flags |= SLAB_STORE_USER;
4697        }
4698        calculate_sizes(s, -1);
4699        return length;
4700}
4701SLAB_ATTR(store_user);
4702
4703static ssize_t validate_show(struct kmem_cache *s, char *buf)
4704{
4705        return 0;
4706}
4707
4708static ssize_t validate_store(struct kmem_cache *s,
4709                        const char *buf, size_t length)
4710{
4711        int ret = -EINVAL;
4712
4713        if (buf[0] == '1') {
4714                ret = validate_slab_cache(s);
4715                if (ret >= 0)
4716                        ret = length;
4717        }
4718        return ret;
4719}
4720SLAB_ATTR(validate);
4721
4722static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4723{
4724        if (!(s->flags & SLAB_STORE_USER))
4725                return -ENOSYS;
4726        return list_locations(s, buf, TRACK_ALLOC);
4727}
4728SLAB_ATTR_RO(alloc_calls);
4729
4730static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4731{
4732        if (!(s->flags & SLAB_STORE_USER))
4733                return -ENOSYS;
4734        return list_locations(s, buf, TRACK_FREE);
4735}
4736SLAB_ATTR_RO(free_calls);
4737#endif /* CONFIG_SLUB_DEBUG */
4738
4739#ifdef CONFIG_FAILSLAB
4740static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4741{
4742        return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4743}
4744
4745static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4746                                                        size_t length)
4747{
4748        s->flags &= ~SLAB_FAILSLAB;
4749        if (buf[0] == '1')
4750                s->flags |= SLAB_FAILSLAB;
4751        return length;
4752}
4753SLAB_ATTR(failslab);
4754#endif
4755
4756static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4757{
4758        return 0;
4759}
4760
4761static ssize_t shrink_store(struct kmem_cache *s,
4762                        const char *buf, size_t length)
4763{
4764        if (buf[0] == '1') {
4765                int rc = kmem_cache_shrink(s);
4766
4767                if (rc)
4768                        return rc;
4769        } else
4770                return -EINVAL;
4771        return length;
4772}
4773SLAB_ATTR(shrink);
4774
4775#ifdef CONFIG_NUMA
4776static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4777{
4778        return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4779}
4780
4781static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4782                                const char *buf, size_t length)
4783{
4784        unsigned long ratio;
4785        int err;
4786
4787        err = strict_strtoul(buf, 10, &ratio);
4788        if (err)
4789                return err;
4790
4791        if (ratio <= 100)
4792                s->remote_node_defrag_ratio = ratio * 10;
4793
4794        return length;
4795}
4796SLAB_ATTR(remote_node_defrag_ratio);
4797#endif
4798
4799#ifdef CONFIG_SLUB_STATS
4800static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4801{
4802        unsigned long sum  = 0;
4803        int cpu;
4804        int len;
4805        int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4806
4807        if (!data)
4808                return -ENOMEM;
4809
4810        for_each_online_cpu(cpu) {
4811                unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4812
4813                data[cpu] = x;
4814                sum += x;
4815        }
4816
4817        len = sprintf(buf, "%lu", sum);
4818
4819#ifdef CONFIG_SMP
4820        for_each_online_cpu(cpu) {
4821                if (data[cpu] && len < PAGE_SIZE - 20)
4822                        len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4823        }
4824#endif
4825        kfree(data);
4826        return len + sprintf(buf + len, "\n");
4827}
4828
4829static void clear_stat(struct kmem_cache *s, enum stat_item si)
4830{
4831        int cpu;
4832
4833        for_each_online_cpu(cpu)
4834                per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4835}
4836
4837#define STAT_ATTR(si, text)                                     \
4838static ssize_t text##_show(struct kmem_cache *s, char *buf)     \
4839{                                                               \
4840        return show_stat(s, buf, si);                           \
4841}                                                               \
4842static ssize_t text##_store(struct kmem_cache *s,               \
4843                                const char *buf, size_t length) \
4844{                                                               \
4845        if (buf[0] != '0')                                      \
4846                return -EINVAL;                                 \
4847        clear_stat(s, si);                                      \
4848        return length;                                          \
4849}                                                               \
4850SLAB_ATTR(text);                                                \
4851
4852STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4853STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4854STAT_ATTR(FREE_FASTPATH, free_fastpath);
4855STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4856STAT_ATTR(FREE_FROZEN, free_frozen);
4857STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4858STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4859STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4860STAT_ATTR(ALLOC_SLAB, alloc_slab);
4861STAT_ATTR(ALLOC_REFILL, alloc_refill);
4862STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4863STAT_ATTR(FREE_SLAB, free_slab);
4864STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4865STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4866STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4867STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4868STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4869STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4870STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4871STAT_ATTR(ORDER_FALLBACK, order_fallback);
4872STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4873STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4874STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4875STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4876STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4877STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4878#endif
4879
4880static struct attribute *slab_attrs[] = {
4881        &slab_size_attr.attr,
4882        &object_size_attr.attr,
4883        &objs_per_slab_attr.attr,
4884        &order_attr.attr,
4885        &min_partial_attr.attr,
4886        &cpu_partial_attr.attr,
4887        &objects_attr.attr,
4888        &objects_partial_attr.attr,
4889        &partial_attr.attr,
4890        &cpu_slabs_attr.attr,
4891        &ctor_attr.attr,
4892        &aliases_attr.attr,
4893        &align_attr.attr,
4894        &hwcache_align_attr.attr,
4895        &reclaim_account_attr.attr,
4896        &destroy_by_rcu_attr.attr,
4897        &shrink_attr.attr,
4898        &reserved_attr.attr,
4899        &slabs_cpu_partial_attr.attr,
4900#ifdef CONFIG_SLUB_DEBUG
4901        &total_objects_attr.attr,
4902        &slabs_attr.attr,
4903        &sanity_checks_attr.attr,
4904        &trace_attr.attr,
4905        &red_zone_attr.attr,
4906        &poison_attr.attr,
4907        &store_user_attr.attr,
4908        &validate_attr.attr,
4909        &alloc_calls_attr.attr,
4910        &free_calls_attr.attr,
4911#endif
4912#ifdef CONFIG_ZONE_DMA
4913        &cache_dma_attr.attr,
4914#endif
4915#ifdef CONFIG_NUMA
4916        &remote_node_defrag_ratio_attr.attr,
4917#endif
4918#ifdef CONFIG_SLUB_STATS
4919        &alloc_fastpath_attr.attr,
4920        &alloc_slowpath_attr.attr,
4921        &free_fastpath_attr.attr,
4922        &free_slowpath_attr.attr,
4923        &free_frozen_attr.attr,
4924        &free_add_partial_attr.attr,
4925        &free_remove_partial_attr.attr,
4926        &alloc_from_partial_attr.attr,
4927        &alloc_slab_attr.attr,
4928        &alloc_refill_attr.attr,
4929        &alloc_node_mismatch_attr.attr,
4930        &free_slab_attr.attr,
4931        &cpuslab_flush_attr.attr,
4932        &deactivate_full_attr.attr,
4933        &deactivate_empty_attr.attr,
4934        &deactivate_to_head_attr.attr,
4935        &deactivate_to_tail_attr.attr,
4936        &deactivate_remote_frees_attr.attr,
4937        &deactivate_bypass_attr.attr,
4938        &order_fallback_attr.attr,
4939        &cmpxchg_double_fail_attr.attr,
4940        &cmpxchg_double_cpu_fail_attr.attr,
4941        &cpu_partial_alloc_attr.attr,
4942        &cpu_partial_free_attr.attr,
4943        &cpu_partial_node_attr.attr,
4944        &cpu_partial_drain_attr.attr,
4945#endif
4946#ifdef CONFIG_FAILSLAB
4947        &failslab_attr.attr,
4948#endif
4949
4950        NULL
4951};
4952
4953static struct attribute_group slab_attr_group = {
4954        .attrs = slab_attrs,
4955};
4956
4957static ssize_t slab_attr_show(struct kobject *kobj,
4958                                struct attribute *attr,
4959                                char *buf)
4960{
4961        struct slab_attribute *attribute;
4962        struct kmem_cache *s;
4963        int err;
4964
4965        attribute = to_slab_attr(attr);
4966        s = to_slab(kobj);
4967
4968        if (!attribute->show)
4969                return -EIO;
4970
4971        err = attribute->show(s, buf);
4972
4973        return err;
4974}
4975
4976static ssize_t slab_attr_store(struct kobject *kobj,
4977                                struct attribute *attr,
4978                                const char *buf, size_t len)
4979{
4980        struct slab_attribute *attribute;
4981        struct kmem_cache *s;
4982        int err;
4983
4984        attribute = to_slab_attr(attr);
4985        s = to_slab(kobj);
4986
4987        if (!attribute->store)
4988                return -EIO;
4989
4990        err = attribute->store(s, buf, len);
4991#ifdef CONFIG_MEMCG_KMEM
4992        if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4993                int i;
4994
4995                mutex_lock(&slab_mutex);
4996                if (s->max_attr_size < len)
4997                        s->max_attr_size = len;
4998
4999                /*
5000                 * This is a best effort propagation, so this function's return
5001                 * value will be determined by the parent cache only. This is
5002                 * basically because not all attributes will have a well
5003                 * defined semantics for rollbacks - most of the actions will
5004                 * have permanent effects.
5005                 *
5006                 * Returning the error value of any of the children that fail
5007                 * is not 100 % defined, in the sense that users seeing the
5008                 * error code won't be able to know anything about the state of
5009                 * the cache.
5010                 *
5011                 * Only returning the error code for the parent cache at least
5012                 * has well defined semantics. The cache being written to
5013                 * directly either failed or succeeded, in which case we loop
5014                 * through the descendants with best-effort propagation.
5015                 */
5016                for_each_memcg_cache_index(i) {
5017                        struct kmem_cache *c = cache_from_memcg(s, i);
5018                        if (c)
5019                                attribute->store(c, buf, len);
5020                }
5021                mutex_unlock(&slab_mutex);
5022        }
5023#endif
5024        return err;
5025}
5026
5027static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5028{
5029#ifdef CONFIG_MEMCG_KMEM
5030        int i;
5031        char *buffer = NULL;
5032
5033        if (!is_root_cache(s))
5034                return;
5035
5036        /*
5037         * This mean this cache had no attribute written. Therefore, no point
5038         * in copying default values around
5039         */
5040        if (!s->max_attr_size)
5041                return;
5042
5043        for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5044                char mbuf[64];
5045                char *buf;
5046                struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5047
5048                if (!attr || !attr->store || !attr->show)
5049                        continue;
5050
5051                /*
5052                 * It is really bad that we have to allocate here, so we will
5053                 * do it only as a fallback. If we actually allocate, though,
5054                 * we can just use the allocated buffer until the end.
5055                 *
5056                 * Most of the slub attributes will tend to be very small in
5057                 * size, but sysfs allows buffers up to a page, so they can
5058                 * theoretically happen.
5059                 */
5060                if (buffer)
5061                        buf = buffer;
5062                else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5063                        buf = mbuf;
5064                else {
5065                        buffer = (char *) get_zeroed_page(GFP_KERNEL);
5066                        if (WARN_ON(!buffer))
5067                                continue;
5068                        buf = buffer;
5069                }
5070
5071                attr->show(s->memcg_params->root_cache, buf);
5072                attr->store(s, buf, strlen(buf));
5073        }
5074
5075        if (buffer)
5076                free_page((unsigned long)buffer);
5077#endif
5078}
5079
5080static const struct sysfs_ops slab_sysfs_ops = {
5081        .show = slab_attr_show,
5082        .store = slab_attr_store,
5083};
5084
5085static struct kobj_type slab_ktype = {
5086        .sysfs_ops = &slab_sysfs_ops,
5087};
5088
5089static int uevent_filter(struct kset *kset, struct kobject *kobj)
5090{
5091        struct kobj_type *ktype = get_ktype(kobj);
5092
5093        if (ktype == &slab_ktype)
5094                return 1;
5095        return 0;
5096}
5097
5098static const struct kset_uevent_ops slab_uevent_ops = {
5099        .filter = uevent_filter,
5100};
5101
5102static struct kset *slab_kset;
5103
5104#define ID_STR_LENGTH 64
5105
5106/* Create a unique string id for a slab cache:
5107 *
5108 * Format       :[flags-]size
5109 */
5110static char *create_unique_id(struct kmem_cache *s)
5111{
5112        char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5113        char *p = name;
5114
5115        BUG_ON(!name);
5116
5117        *p++ = ':';
5118        /*
5119         * First flags affecting slabcache operations. We will only
5120         * get here for aliasable slabs so we do not need to support
5121         * too many flags. The flags here must cover all flags that
5122         * are matched during merging to guarantee that the id is
5123         * unique.
5124         */
5125        if (s->flags & SLAB_CACHE_DMA)
5126                *p++ = 'd';
5127        if (s->flags & SLAB_RECLAIM_ACCOUNT)
5128                *p++ = 'a';
5129        if (s->flags & SLAB_DEBUG_FREE)
5130                *p++ = 'F';
5131        if (!(s->flags & SLAB_NOTRACK))
5132                *p++ = 't';
5133        if (p != name + 1)
5134                *p++ = '-';
5135        p += sprintf(p, "%07d", s->size);
5136
5137#ifdef CONFIG_MEMCG_KMEM
5138        if (!is_root_cache(s))
5139                p += sprintf(p, "-%08d", memcg_cache_id(s->memcg_params->memcg));
5140#endif
5141
5142        BUG_ON(p > name + ID_STR_LENGTH - 1);
5143        return name;
5144}
5145
5146static int sysfs_slab_add(struct kmem_cache *s)
5147{
5148        int err;
5149        const char *name;
5150        int unmergeable = slab_unmergeable(s);
5151
5152        if (unmergeable) {
5153                /*
5154                 * Slabcache can never be merged so we can use the name proper.
5155                 * This is typically the case for debug situations. In that
5156                 * case we can catch duplicate names easily.
5157                 */
5158                sysfs_remove_link(&slab_kset->kobj, s->name);
5159                name = s->name;
5160        } else {
5161                /*
5162                 * Create a unique name for the slab as a target
5163                 * for the symlinks.
5164                 */
5165                name = create_unique_id(s);
5166        }
5167
5168        s->kobj.kset = slab_kset;
5169        err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5170        if (err) {
5171                kobject_put(&s->kobj);
5172                return err;
5173        }
5174
5175        err = sysfs_create_group(&s->kobj, &slab_attr_group);
5176        if (err) {
5177                kobject_del(&s->kobj);
5178                kobject_put(&s->kobj);
5179                return err;
5180        }
5181        kobject_uevent(&s->kobj, KOBJ_ADD);
5182        if (!unmergeable) {
5183                /* Setup first alias */
5184                sysfs_slab_alias(s, s->name);
5185                kfree(name);
5186        }
5187        return 0;
5188}
5189
5190static void sysfs_slab_remove(struct kmem_cache *s)
5191{
5192        if (slab_state < FULL)
5193                /*
5194                 * Sysfs has not been setup yet so no need to remove the
5195                 * cache from sysfs.
5196                 */
5197                return;
5198
5199        kobject_uevent(&s->kobj, KOBJ_REMOVE);
5200        kobject_del(&s->kobj);
5201        kobject_put(&s->kobj);
5202}
5203
5204/*
5205 * Need to buffer aliases during bootup until sysfs becomes
5206 * available lest we lose that information.
5207 */
5208struct saved_alias {
5209        struct kmem_cache *s;
5210        const char *name;
5211        struct saved_alias *next;
5212};
5213
5214static struct saved_alias *alias_list;
5215
5216static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5217{
5218        struct saved_alias *al;
5219
5220        if (slab_state == FULL) {
5221                /*
5222                 * If we have a leftover link then remove it.
5223                 */
5224                sysfs_remove_link(&slab_kset->kobj, name);
5225                return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5226        }
5227
5228        al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5229        if (!al)
5230                return -ENOMEM;
5231
5232        al->s = s;
5233        al->name = name;
5234        al->next = alias_list;
5235        alias_list = al;
5236        return 0;
5237}
5238
5239static int __init slab_sysfs_init(void)
5240{
5241        struct kmem_cache *s;
5242        int err;
5243
5244        mutex_lock(&slab_mutex);
5245
5246        slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5247        if (!slab_kset) {
5248                mutex_unlock(&slab_mutex);
5249                printk(KERN_ERR "Cannot register slab subsystem.\n");
5250                return -ENOSYS;
5251        }
5252
5253        slab_state = FULL;
5254
5255        list_for_each_entry(s, &slab_caches, list) {
5256                err = sysfs_slab_add(s);
5257                if (err)
5258                        printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5259                                                " to sysfs\n", s->name);
5260        }
5261
5262        while (alias_list) {
5263                struct saved_alias *al = alias_list;
5264
5265                alias_list = alias_list->next;
5266                err = sysfs_slab_alias(al->s, al->name);
5267                if (err)
5268                        printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5269                                        " %s to sysfs\n", al->name);
5270                kfree(al);
5271        }
5272
5273        mutex_unlock(&slab_mutex);
5274        resiliency_test();
5275        return 0;
5276}
5277
5278__initcall(slab_sysfs_init);
5279#endif /* CONFIG_SYSFS */
5280
5281/*
5282 * The /proc/slabinfo ABI
5283 */
5284#ifdef CONFIG_SLABINFO
5285void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5286{
5287        unsigned long nr_slabs = 0;
5288        unsigned long nr_objs = 0;
5289        unsigned long nr_free = 0;
5290        int node;
5291
5292        for_each_online_node(node) {
5293                struct kmem_cache_node *n = get_node(s, node);
5294
5295                if (!n)
5296                        continue;
5297
5298                nr_slabs += node_nr_slabs(n);
5299                nr_objs += node_nr_objs(n);
5300                nr_free += count_partial(n, count_free);
5301        }
5302
5303        sinfo->active_objs = nr_objs - nr_free;
5304        sinfo->num_objs = nr_objs;
5305        sinfo->active_slabs = nr_slabs;
5306        sinfo->num_slabs = nr_slabs;
5307        sinfo->objects_per_slab = oo_objects(s->oo);
5308        sinfo->cache_order = oo_order(s->oo);
5309}
5310
5311void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5312{
5313}
5314
5315ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5316                       size_t count, loff_t *ppos)
5317{
5318        return -EIO;
5319}
5320#endif /* CONFIG_SLABINFO */
5321
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.