Root/mm/slab.c

1/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
30 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
87 */
88
89#include <linux/slab.h>
90#include <linux/mm.h>
91#include <linux/poison.h>
92#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
97#include <linux/cpuset.h>
98#include <linux/proc_fs.h>
99#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
105#include <linux/rcupdate.h>
106#include <linux/string.h>
107#include <linux/uaccess.h>
108#include <linux/nodemask.h>
109#include <linux/kmemleak.h>
110#include <linux/mempolicy.h>
111#include <linux/mutex.h>
112#include <linux/fault-inject.h>
113#include <linux/rtmutex.h>
114#include <linux/reciprocal_div.h>
115#include <linux/debugobjects.h>
116#include <linux/kmemcheck.h>
117#include <linux/memory.h>
118#include <linux/prefetch.h>
119
120#include <asm/cacheflush.h>
121#include <asm/tlbflush.h>
122#include <asm/page.h>
123
124/*
125 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
126 * 0 for faster, smaller code (especially in the critical paths).
127 *
128 * STATS - 1 to collect stats for /proc/slabinfo.
129 * 0 for faster, smaller code (especially in the critical paths).
130 *
131 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
132 */
133
134#ifdef CONFIG_DEBUG_SLAB
135#define DEBUG 1
136#define STATS 1
137#define FORCED_DEBUG 1
138#else
139#define DEBUG 0
140#define STATS 0
141#define FORCED_DEBUG 0
142#endif
143
144/* Shouldn't this be in a header file somewhere? */
145#define BYTES_PER_WORD sizeof(void *)
146#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
147
148#ifndef ARCH_KMALLOC_FLAGS
149#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
150#endif
151
152/* Legal flag mask for kmem_cache_create(). */
153#if DEBUG
154# define CREATE_MASK (SLAB_RED_ZONE | \
155             SLAB_POISON | SLAB_HWCACHE_ALIGN | \
156             SLAB_CACHE_DMA | \
157             SLAB_STORE_USER | \
158             SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
159             SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
160             SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
161#else
162# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
163             SLAB_CACHE_DMA | \
164             SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
165             SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
166             SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
167#endif
168
169/*
170 * kmem_bufctl_t:
171 *
172 * Bufctl's are used for linking objs within a slab
173 * linked offsets.
174 *
175 * This implementation relies on "struct page" for locating the cache &
176 * slab an object belongs to.
177 * This allows the bufctl structure to be small (one int), but limits
178 * the number of objects a slab (not a cache) can contain when off-slab
179 * bufctls are used. The limit is the size of the largest general cache
180 * that does not use off-slab slabs.
181 * For 32bit archs with 4 kB pages, is this 56.
182 * This is not serious, as it is only for large objects, when it is unwise
183 * to have too many per slab.
184 * Note: This limit can be raised by introducing a general cache whose size
185 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
186 */
187
188typedef unsigned int kmem_bufctl_t;
189#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
190#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
191#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
192#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
193
194/*
195 * struct slab_rcu
196 *
197 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
198 * arrange for kmem_freepages to be called via RCU. This is useful if
199 * we need to approach a kernel structure obliquely, from its address
200 * obtained without the usual locking. We can lock the structure to
201 * stabilize it and check it's still at the given address, only if we
202 * can be sure that the memory has not been meanwhile reused for some
203 * other kind of object (which our subsystem's lock might corrupt).
204 *
205 * rcu_read_lock before reading the address, then rcu_read_unlock after
206 * taking the spinlock within the structure expected at that address.
207 */
208struct slab_rcu {
209    struct rcu_head head;
210    struct kmem_cache *cachep;
211    void *addr;
212};
213
214/*
215 * struct slab
216 *
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 */
221struct slab {
222    union {
223        struct {
224            struct list_head list;
225            unsigned long colouroff;
226            void *s_mem; /* including colour offset */
227            unsigned int inuse; /* num of objs active in slab */
228            kmem_bufctl_t free;
229            unsigned short nodeid;
230        };
231        struct slab_rcu __slab_cover_slab_rcu;
232    };
233};
234
235/*
236 * struct array_cache
237 *
238 * Purpose:
239 * - LIFO ordering, to hand out cache-warm objects from _alloc
240 * - reduce the number of linked list operations
241 * - reduce spinlock operations
242 *
243 * The limit is stored in the per-cpu structure to reduce the data cache
244 * footprint.
245 *
246 */
247struct array_cache {
248    unsigned int avail;
249    unsigned int limit;
250    unsigned int batchcount;
251    unsigned int touched;
252    spinlock_t lock;
253    void *entry[]; /*
254             * Must have this definition in here for the proper
255             * alignment of array_cache. Also simplifies accessing
256             * the entries.
257             */
258};
259
260/*
261 * bootstrap: The caches do not work without cpuarrays anymore, but the
262 * cpuarrays are allocated from the generic caches...
263 */
264#define BOOT_CPUCACHE_ENTRIES 1
265struct arraycache_init {
266    struct array_cache cache;
267    void *entries[BOOT_CPUCACHE_ENTRIES];
268};
269
270/*
271 * The slab lists for all objects.
272 */
273struct kmem_list3 {
274    struct list_head slabs_partial; /* partial list first, better asm code */
275    struct list_head slabs_full;
276    struct list_head slabs_free;
277    unsigned long free_objects;
278    unsigned int free_limit;
279    unsigned int colour_next; /* Per-node cache coloring */
280    spinlock_t list_lock;
281    struct array_cache *shared; /* shared per node */
282    struct array_cache **alien; /* on other nodes */
283    unsigned long next_reap; /* updated without locking */
284    int free_touched; /* updated without locking */
285};
286
287/*
288 * Need this for bootstrapping a per node allocator.
289 */
290#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
291static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
292#define CACHE_CACHE 0
293#define SIZE_AC MAX_NUMNODES
294#define SIZE_L3 (2 * MAX_NUMNODES)
295
296static int drain_freelist(struct kmem_cache *cache,
297            struct kmem_list3 *l3, int tofree);
298static void free_block(struct kmem_cache *cachep, void **objpp, int len,
299            int node);
300static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
301static void cache_reap(struct work_struct *unused);
302
303/*
304 * This function must be completely optimized away if a constant is passed to
305 * it. Mostly the same as what is in linux/slab.h except it returns an index.
306 */
307static __always_inline int index_of(const size_t size)
308{
309    extern void __bad_size(void);
310
311    if (__builtin_constant_p(size)) {
312        int i = 0;
313
314#define CACHE(x) \
315    if (size <=x) \
316        return i; \
317    else \
318        i++;
319#include <linux/kmalloc_sizes.h>
320#undef CACHE
321        __bad_size();
322    } else
323        __bad_size();
324    return 0;
325}
326
327static int slab_early_init = 1;
328
329#define INDEX_AC index_of(sizeof(struct arraycache_init))
330#define INDEX_L3 index_of(sizeof(struct kmem_list3))
331
332static void kmem_list3_init(struct kmem_list3 *parent)
333{
334    INIT_LIST_HEAD(&parent->slabs_full);
335    INIT_LIST_HEAD(&parent->slabs_partial);
336    INIT_LIST_HEAD(&parent->slabs_free);
337    parent->shared = NULL;
338    parent->alien = NULL;
339    parent->colour_next = 0;
340    spin_lock_init(&parent->list_lock);
341    parent->free_objects = 0;
342    parent->free_touched = 0;
343}
344
345#define MAKE_LIST(cachep, listp, slab, nodeid) \
346    do { \
347        INIT_LIST_HEAD(listp); \
348        list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
349    } while (0)
350
351#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
352    do { \
353    MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
354    MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
355    MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
356    } while (0)
357
358#define CFLGS_OFF_SLAB (0x80000000UL)
359#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
360
361#define BATCHREFILL_LIMIT 16
362/*
363 * Optimization question: fewer reaps means less probability for unnessary
364 * cpucache drain/refill cycles.
365 *
366 * OTOH the cpuarrays can contain lots of objects,
367 * which could lock up otherwise freeable slabs.
368 */
369#define REAPTIMEOUT_CPUC (2*HZ)
370#define REAPTIMEOUT_LIST3 (4*HZ)
371
372#if STATS
373#define STATS_INC_ACTIVE(x) ((x)->num_active++)
374#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
375#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
376#define STATS_INC_GROWN(x) ((x)->grown++)
377#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
378#define STATS_SET_HIGH(x) \
379    do { \
380        if ((x)->num_active > (x)->high_mark) \
381            (x)->high_mark = (x)->num_active; \
382    } while (0)
383#define STATS_INC_ERR(x) ((x)->errors++)
384#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
385#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
386#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
387#define STATS_SET_FREEABLE(x, i) \
388    do { \
389        if ((x)->max_freeable < i) \
390            (x)->max_freeable = i; \
391    } while (0)
392#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
393#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
394#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
395#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
396#else
397#define STATS_INC_ACTIVE(x) do { } while (0)
398#define STATS_DEC_ACTIVE(x) do { } while (0)
399#define STATS_INC_ALLOCED(x) do { } while (0)
400#define STATS_INC_GROWN(x) do { } while (0)
401#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
402#define STATS_SET_HIGH(x) do { } while (0)
403#define STATS_INC_ERR(x) do { } while (0)
404#define STATS_INC_NODEALLOCS(x) do { } while (0)
405#define STATS_INC_NODEFREES(x) do { } while (0)
406#define STATS_INC_ACOVERFLOW(x) do { } while (0)
407#define STATS_SET_FREEABLE(x, i) do { } while (0)
408#define STATS_INC_ALLOCHIT(x) do { } while (0)
409#define STATS_INC_ALLOCMISS(x) do { } while (0)
410#define STATS_INC_FREEHIT(x) do { } while (0)
411#define STATS_INC_FREEMISS(x) do { } while (0)
412#endif
413
414#if DEBUG
415
416/*
417 * memory layout of objects:
418 * 0 : objp
419 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
420 * the end of an object is aligned with the end of the real
421 * allocation. Catches writes behind the end of the allocation.
422 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
423 * redzone word.
424 * cachep->obj_offset: The real object.
425 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
426 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
427 * [BYTES_PER_WORD long]
428 */
429static int obj_offset(struct kmem_cache *cachep)
430{
431    return cachep->obj_offset;
432}
433
434static int obj_size(struct kmem_cache *cachep)
435{
436    return cachep->obj_size;
437}
438
439static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
440{
441    BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
442    return (unsigned long long*) (objp + obj_offset(cachep) -
443                      sizeof(unsigned long long));
444}
445
446static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
447{
448    BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
449    if (cachep->flags & SLAB_STORE_USER)
450        return (unsigned long long *)(objp + cachep->buffer_size -
451                          sizeof(unsigned long long) -
452                          REDZONE_ALIGN);
453    return (unsigned long long *) (objp + cachep->buffer_size -
454                       sizeof(unsigned long long));
455}
456
457static void **dbg_userword(struct kmem_cache *cachep, void *objp)
458{
459    BUG_ON(!(cachep->flags & SLAB_STORE_USER));
460    return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
461}
462
463#else
464
465#define obj_offset(x) 0
466#define obj_size(cachep) (cachep->buffer_size)
467#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
468#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
469#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
470
471#endif
472
473#ifdef CONFIG_TRACING
474size_t slab_buffer_size(struct kmem_cache *cachep)
475{
476    return cachep->buffer_size;
477}
478EXPORT_SYMBOL(slab_buffer_size);
479#endif
480
481/*
482 * Do not go above this order unless 0 objects fit into the slab.
483 */
484#define BREAK_GFP_ORDER_HI 1
485#define BREAK_GFP_ORDER_LO 0
486static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
487
488/*
489 * Functions for storing/retrieving the cachep and or slab from the page
490 * allocator. These are used to find the slab an obj belongs to. With kfree(),
491 * these are used to find the cache which an obj belongs to.
492 */
493static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
494{
495    page->lru.next = (struct list_head *)cache;
496}
497
498static inline struct kmem_cache *page_get_cache(struct page *page)
499{
500    page = compound_head(page);
501    BUG_ON(!PageSlab(page));
502    return (struct kmem_cache *)page->lru.next;
503}
504
505static inline void page_set_slab(struct page *page, struct slab *slab)
506{
507    page->lru.prev = (struct list_head *)slab;
508}
509
510static inline struct slab *page_get_slab(struct page *page)
511{
512    BUG_ON(!PageSlab(page));
513    return (struct slab *)page->lru.prev;
514}
515
516static inline struct kmem_cache *virt_to_cache(const void *obj)
517{
518    struct page *page = virt_to_head_page(obj);
519    return page_get_cache(page);
520}
521
522static inline struct slab *virt_to_slab(const void *obj)
523{
524    struct page *page = virt_to_head_page(obj);
525    return page_get_slab(page);
526}
527
528static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
529                 unsigned int idx)
530{
531    return slab->s_mem + cache->buffer_size * idx;
532}
533
534/*
535 * We want to avoid an expensive divide : (offset / cache->buffer_size)
536 * Using the fact that buffer_size is a constant for a particular cache,
537 * we can replace (offset / cache->buffer_size) by
538 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
539 */
540static inline unsigned int obj_to_index(const struct kmem_cache *cache,
541                    const struct slab *slab, void *obj)
542{
543    u32 offset = (obj - slab->s_mem);
544    return reciprocal_divide(offset, cache->reciprocal_buffer_size);
545}
546
547/*
548 * These are the default caches for kmalloc. Custom caches can have other sizes.
549 */
550struct cache_sizes malloc_sizes[] = {
551#define CACHE(x) { .cs_size = (x) },
552#include <linux/kmalloc_sizes.h>
553    CACHE(ULONG_MAX)
554#undef CACHE
555};
556EXPORT_SYMBOL(malloc_sizes);
557
558/* Must match cache_sizes above. Out of line to keep cache footprint low. */
559struct cache_names {
560    char *name;
561    char *name_dma;
562};
563
564static struct cache_names __initdata cache_names[] = {
565#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
566#include <linux/kmalloc_sizes.h>
567    {NULL,}
568#undef CACHE
569};
570
571static struct arraycache_init initarray_cache __initdata =
572    { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
573static struct arraycache_init initarray_generic =
574    { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
575
576/* internal cache of cache description objs */
577static struct kmem_cache cache_cache = {
578    .batchcount = 1,
579    .limit = BOOT_CPUCACHE_ENTRIES,
580    .shared = 1,
581    .buffer_size = sizeof(struct kmem_cache),
582    .name = "kmem_cache",
583};
584
585#define BAD_ALIEN_MAGIC 0x01020304ul
586
587/*
588 * chicken and egg problem: delay the per-cpu array allocation
589 * until the general caches are up.
590 */
591static enum {
592    NONE,
593    PARTIAL_AC,
594    PARTIAL_L3,
595    EARLY,
596    FULL
597} g_cpucache_up;
598
599/*
600 * used by boot code to determine if it can use slab based allocator
601 */
602int slab_is_available(void)
603{
604    return g_cpucache_up >= EARLY;
605}
606
607#ifdef CONFIG_LOCKDEP
608
609/*
610 * Slab sometimes uses the kmalloc slabs to store the slab headers
611 * for other slabs "off slab".
612 * The locking for this is tricky in that it nests within the locks
613 * of all other slabs in a few places; to deal with this special
614 * locking we put on-slab caches into a separate lock-class.
615 *
616 * We set lock class for alien array caches which are up during init.
617 * The lock annotation will be lost if all cpus of a node goes down and
618 * then comes back up during hotplug
619 */
620static struct lock_class_key on_slab_l3_key;
621static struct lock_class_key on_slab_alc_key;
622
623static void init_node_lock_keys(int q)
624{
625    struct cache_sizes *s = malloc_sizes;
626
627    if (g_cpucache_up != FULL)
628        return;
629
630    for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
631        struct array_cache **alc;
632        struct kmem_list3 *l3;
633        int r;
634
635        l3 = s->cs_cachep->nodelists[q];
636        if (!l3 || OFF_SLAB(s->cs_cachep))
637            continue;
638        lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
639        alc = l3->alien;
640        /*
641         * FIXME: This check for BAD_ALIEN_MAGIC
642         * should go away when common slab code is taught to
643         * work even without alien caches.
644         * Currently, non NUMA code returns BAD_ALIEN_MAGIC
645         * for alloc_alien_cache,
646         */
647        if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
648            continue;
649        for_each_node(r) {
650            if (alc[r])
651                lockdep_set_class(&alc[r]->lock,
652                    &on_slab_alc_key);
653        }
654    }
655}
656
657static inline void init_lock_keys(void)
658{
659    int node;
660
661    for_each_node(node)
662        init_node_lock_keys(node);
663}
664#else
665static void init_node_lock_keys(int q)
666{
667}
668
669static inline void init_lock_keys(void)
670{
671}
672#endif
673
674/*
675 * Guard access to the cache-chain.
676 */
677static DEFINE_MUTEX(cache_chain_mutex);
678static struct list_head cache_chain;
679
680static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
681
682static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
683{
684    return cachep->array[smp_processor_id()];
685}
686
687static inline struct kmem_cache *__find_general_cachep(size_t size,
688                            gfp_t gfpflags)
689{
690    struct cache_sizes *csizep = malloc_sizes;
691
692#if DEBUG
693    /* This happens if someone tries to call
694     * kmem_cache_create(), or __kmalloc(), before
695     * the generic caches are initialized.
696     */
697    BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
698#endif
699    if (!size)
700        return ZERO_SIZE_PTR;
701
702    while (size > csizep->cs_size)
703        csizep++;
704
705    /*
706     * Really subtle: The last entry with cs->cs_size==ULONG_MAX
707     * has cs_{dma,}cachep==NULL. Thus no special case
708     * for large kmalloc calls required.
709     */
710#ifdef CONFIG_ZONE_DMA
711    if (unlikely(gfpflags & GFP_DMA))
712        return csizep->cs_dmacachep;
713#endif
714    return csizep->cs_cachep;
715}
716
717static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
718{
719    return __find_general_cachep(size, gfpflags);
720}
721
722static size_t slab_mgmt_size(size_t nr_objs, size_t align)
723{
724    return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
725}
726
727/*
728 * Calculate the number of objects and left-over bytes for a given buffer size.
729 */
730static void cache_estimate(unsigned long gfporder, size_t buffer_size,
731               size_t align, int flags, size_t *left_over,
732               unsigned int *num)
733{
734    int nr_objs;
735    size_t mgmt_size;
736    size_t slab_size = PAGE_SIZE << gfporder;
737
738    /*
739     * The slab management structure can be either off the slab or
740     * on it. For the latter case, the memory allocated for a
741     * slab is used for:
742     *
743     * - The struct slab
744     * - One kmem_bufctl_t for each object
745     * - Padding to respect alignment of @align
746     * - @buffer_size bytes for each object
747     *
748     * If the slab management structure is off the slab, then the
749     * alignment will already be calculated into the size. Because
750     * the slabs are all pages aligned, the objects will be at the
751     * correct alignment when allocated.
752     */
753    if (flags & CFLGS_OFF_SLAB) {
754        mgmt_size = 0;
755        nr_objs = slab_size / buffer_size;
756
757        if (nr_objs > SLAB_LIMIT)
758            nr_objs = SLAB_LIMIT;
759    } else {
760        /*
761         * Ignore padding for the initial guess. The padding
762         * is at most @align-1 bytes, and @buffer_size is at
763         * least @align. In the worst case, this result will
764         * be one greater than the number of objects that fit
765         * into the memory allocation when taking the padding
766         * into account.
767         */
768        nr_objs = (slab_size - sizeof(struct slab)) /
769              (buffer_size + sizeof(kmem_bufctl_t));
770
771        /*
772         * This calculated number will be either the right
773         * amount, or one greater than what we want.
774         */
775        if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
776               > slab_size)
777            nr_objs--;
778
779        if (nr_objs > SLAB_LIMIT)
780            nr_objs = SLAB_LIMIT;
781
782        mgmt_size = slab_mgmt_size(nr_objs, align);
783    }
784    *num = nr_objs;
785    *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
786}
787
788#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
789
790static void __slab_error(const char *function, struct kmem_cache *cachep,
791            char *msg)
792{
793    printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
794           function, cachep->name, msg);
795    dump_stack();
796}
797
798/*
799 * By default on NUMA we use alien caches to stage the freeing of
800 * objects allocated from other nodes. This causes massive memory
801 * inefficiencies when using fake NUMA setup to split memory into a
802 * large number of small nodes, so it can be disabled on the command
803 * line
804  */
805
806static int use_alien_caches __read_mostly = 1;
807static int __init noaliencache_setup(char *s)
808{
809    use_alien_caches = 0;
810    return 1;
811}
812__setup("noaliencache", noaliencache_setup);
813
814#ifdef CONFIG_NUMA
815/*
816 * Special reaping functions for NUMA systems called from cache_reap().
817 * These take care of doing round robin flushing of alien caches (containing
818 * objects freed on different nodes from which they were allocated) and the
819 * flushing of remote pcps by calling drain_node_pages.
820 */
821static DEFINE_PER_CPU(unsigned long, slab_reap_node);
822
823static void init_reap_node(int cpu)
824{
825    int node;
826
827    node = next_node(cpu_to_mem(cpu), node_online_map);
828    if (node == MAX_NUMNODES)
829        node = first_node(node_online_map);
830
831    per_cpu(slab_reap_node, cpu) = node;
832}
833
834static void next_reap_node(void)
835{
836    int node = __this_cpu_read(slab_reap_node);
837
838    node = next_node(node, node_online_map);
839    if (unlikely(node >= MAX_NUMNODES))
840        node = first_node(node_online_map);
841    __this_cpu_write(slab_reap_node, node);
842}
843
844#else
845#define init_reap_node(cpu) do { } while (0)
846#define next_reap_node(void) do { } while (0)
847#endif
848
849/*
850 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
851 * via the workqueue/eventd.
852 * Add the CPU number into the expiration time to minimize the possibility of
853 * the CPUs getting into lockstep and contending for the global cache chain
854 * lock.
855 */
856static void __cpuinit start_cpu_timer(int cpu)
857{
858    struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
859
860    /*
861     * When this gets called from do_initcalls via cpucache_init(),
862     * init_workqueues() has already run, so keventd will be setup
863     * at that time.
864     */
865    if (keventd_up() && reap_work->work.func == NULL) {
866        init_reap_node(cpu);
867        INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
868        schedule_delayed_work_on(cpu, reap_work,
869                    __round_jiffies_relative(HZ, cpu));
870    }
871}
872
873static struct array_cache *alloc_arraycache(int node, int entries,
874                        int batchcount, gfp_t gfp)
875{
876    int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
877    struct array_cache *nc = NULL;
878
879    nc = kmalloc_node(memsize, gfp, node);
880    /*
881     * The array_cache structures contain pointers to free object.
882     * However, when such objects are allocated or transferred to another
883     * cache the pointers are not cleared and they could be counted as
884     * valid references during a kmemleak scan. Therefore, kmemleak must
885     * not scan such objects.
886     */
887    kmemleak_no_scan(nc);
888    if (nc) {
889        nc->avail = 0;
890        nc->limit = entries;
891        nc->batchcount = batchcount;
892        nc->touched = 0;
893        spin_lock_init(&nc->lock);
894    }
895    return nc;
896}
897
898/*
899 * Transfer objects in one arraycache to another.
900 * Locking must be handled by the caller.
901 *
902 * Return the number of entries transferred.
903 */
904static int transfer_objects(struct array_cache *to,
905        struct array_cache *from, unsigned int max)
906{
907    /* Figure out how many entries to transfer */
908    int nr = min3(from->avail, max, to->limit - to->avail);
909
910    if (!nr)
911        return 0;
912
913    memcpy(to->entry + to->avail, from->entry + from->avail -nr,
914            sizeof(void *) *nr);
915
916    from->avail -= nr;
917    to->avail += nr;
918    return nr;
919}
920
921#ifndef CONFIG_NUMA
922
923#define drain_alien_cache(cachep, alien) do { } while (0)
924#define reap_alien(cachep, l3) do { } while (0)
925
926static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
927{
928    return (struct array_cache **)BAD_ALIEN_MAGIC;
929}
930
931static inline void free_alien_cache(struct array_cache **ac_ptr)
932{
933}
934
935static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
936{
937    return 0;
938}
939
940static inline void *alternate_node_alloc(struct kmem_cache *cachep,
941        gfp_t flags)
942{
943    return NULL;
944}
945
946static inline void *____cache_alloc_node(struct kmem_cache *cachep,
947         gfp_t flags, int nodeid)
948{
949    return NULL;
950}
951
952#else /* CONFIG_NUMA */
953
954static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
955static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
956
957static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
958{
959    struct array_cache **ac_ptr;
960    int memsize = sizeof(void *) * nr_node_ids;
961    int i;
962
963    if (limit > 1)
964        limit = 12;
965    ac_ptr = kzalloc_node(memsize, gfp, node);
966    if (ac_ptr) {
967        for_each_node(i) {
968            if (i == node || !node_online(i))
969                continue;
970            ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
971            if (!ac_ptr[i]) {
972                for (i--; i >= 0; i--)
973                    kfree(ac_ptr[i]);
974                kfree(ac_ptr);
975                return NULL;
976            }
977        }
978    }
979    return ac_ptr;
980}
981
982static void free_alien_cache(struct array_cache **ac_ptr)
983{
984    int i;
985
986    if (!ac_ptr)
987        return;
988    for_each_node(i)
989        kfree(ac_ptr[i]);
990    kfree(ac_ptr);
991}
992
993static void __drain_alien_cache(struct kmem_cache *cachep,
994                struct array_cache *ac, int node)
995{
996    struct kmem_list3 *rl3 = cachep->nodelists[node];
997
998    if (ac->avail) {
999        spin_lock(&rl3->list_lock);
1000        /*
1001         * Stuff objects into the remote nodes shared array first.
1002         * That way we could avoid the overhead of putting the objects
1003         * into the free lists and getting them back later.
1004         */
1005        if (rl3->shared)
1006            transfer_objects(rl3->shared, ac, ac->limit);
1007
1008        free_block(cachep, ac->entry, ac->avail, node);
1009        ac->avail = 0;
1010        spin_unlock(&rl3->list_lock);
1011    }
1012}
1013
1014/*
1015 * Called from cache_reap() to regularly drain alien caches round robin.
1016 */
1017static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1018{
1019    int node = __this_cpu_read(slab_reap_node);
1020
1021    if (l3->alien) {
1022        struct array_cache *ac = l3->alien[node];
1023
1024        if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1025            __drain_alien_cache(cachep, ac, node);
1026            spin_unlock_irq(&ac->lock);
1027        }
1028    }
1029}
1030
1031static void drain_alien_cache(struct kmem_cache *cachep,
1032                struct array_cache **alien)
1033{
1034    int i = 0;
1035    struct array_cache *ac;
1036    unsigned long flags;
1037
1038    for_each_online_node(i) {
1039        ac = alien[i];
1040        if (ac) {
1041            spin_lock_irqsave(&ac->lock, flags);
1042            __drain_alien_cache(cachep, ac, i);
1043            spin_unlock_irqrestore(&ac->lock, flags);
1044        }
1045    }
1046}
1047
1048static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1049{
1050    struct slab *slabp = virt_to_slab(objp);
1051    int nodeid = slabp->nodeid;
1052    struct kmem_list3 *l3;
1053    struct array_cache *alien = NULL;
1054    int node;
1055
1056    node = numa_mem_id();
1057
1058    /*
1059     * Make sure we are not freeing a object from another node to the array
1060     * cache on this cpu.
1061     */
1062    if (likely(slabp->nodeid == node))
1063        return 0;
1064
1065    l3 = cachep->nodelists[node];
1066    STATS_INC_NODEFREES(cachep);
1067    if (l3->alien && l3->alien[nodeid]) {
1068        alien = l3->alien[nodeid];
1069        spin_lock(&alien->lock);
1070        if (unlikely(alien->avail == alien->limit)) {
1071            STATS_INC_ACOVERFLOW(cachep);
1072            __drain_alien_cache(cachep, alien, nodeid);
1073        }
1074        alien->entry[alien->avail++] = objp;
1075        spin_unlock(&alien->lock);
1076    } else {
1077        spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1078        free_block(cachep, &objp, 1, nodeid);
1079        spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1080    }
1081    return 1;
1082}
1083#endif
1084
1085/*
1086 * Allocates and initializes nodelists for a node on each slab cache, used for
1087 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1088 * will be allocated off-node since memory is not yet online for the new node.
1089 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1090 * already in use.
1091 *
1092 * Must hold cache_chain_mutex.
1093 */
1094static int init_cache_nodelists_node(int node)
1095{
1096    struct kmem_cache *cachep;
1097    struct kmem_list3 *l3;
1098    const int memsize = sizeof(struct kmem_list3);
1099
1100    list_for_each_entry(cachep, &cache_chain, next) {
1101        /*
1102         * Set up the size64 kmemlist for cpu before we can
1103         * begin anything. Make sure some other cpu on this
1104         * node has not already allocated this
1105         */
1106        if (!cachep->nodelists[node]) {
1107            l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1108            if (!l3)
1109                return -ENOMEM;
1110            kmem_list3_init(l3);
1111            l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1112                ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1113
1114            /*
1115             * The l3s don't come and go as CPUs come and
1116             * go. cache_chain_mutex is sufficient
1117             * protection here.
1118             */
1119            cachep->nodelists[node] = l3;
1120        }
1121
1122        spin_lock_irq(&cachep->nodelists[node]->list_lock);
1123        cachep->nodelists[node]->free_limit =
1124            (1 + nr_cpus_node(node)) *
1125            cachep->batchcount + cachep->num;
1126        spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1127    }
1128    return 0;
1129}
1130
1131static void __cpuinit cpuup_canceled(long cpu)
1132{
1133    struct kmem_cache *cachep;
1134    struct kmem_list3 *l3 = NULL;
1135    int node = cpu_to_mem(cpu);
1136    const struct cpumask *mask = cpumask_of_node(node);
1137
1138    list_for_each_entry(cachep, &cache_chain, next) {
1139        struct array_cache *nc;
1140        struct array_cache *shared;
1141        struct array_cache **alien;
1142
1143        /* cpu is dead; no one can alloc from it. */
1144        nc = cachep->array[cpu];
1145        cachep->array[cpu] = NULL;
1146        l3 = cachep->nodelists[node];
1147
1148        if (!l3)
1149            goto free_array_cache;
1150
1151        spin_lock_irq(&l3->list_lock);
1152
1153        /* Free limit for this kmem_list3 */
1154        l3->free_limit -= cachep->batchcount;
1155        if (nc)
1156            free_block(cachep, nc->entry, nc->avail, node);
1157
1158        if (!cpumask_empty(mask)) {
1159            spin_unlock_irq(&l3->list_lock);
1160            goto free_array_cache;
1161        }
1162
1163        shared = l3->shared;
1164        if (shared) {
1165            free_block(cachep, shared->entry,
1166                   shared->avail, node);
1167            l3->shared = NULL;
1168        }
1169
1170        alien = l3->alien;
1171        l3->alien = NULL;
1172
1173        spin_unlock_irq(&l3->list_lock);
1174
1175        kfree(shared);
1176        if (alien) {
1177            drain_alien_cache(cachep, alien);
1178            free_alien_cache(alien);
1179        }
1180free_array_cache:
1181        kfree(nc);
1182    }
1183    /*
1184     * In the previous loop, all the objects were freed to
1185     * the respective cache's slabs, now we can go ahead and
1186     * shrink each nodelist to its limit.
1187     */
1188    list_for_each_entry(cachep, &cache_chain, next) {
1189        l3 = cachep->nodelists[node];
1190        if (!l3)
1191            continue;
1192        drain_freelist(cachep, l3, l3->free_objects);
1193    }
1194}
1195
1196static int __cpuinit cpuup_prepare(long cpu)
1197{
1198    struct kmem_cache *cachep;
1199    struct kmem_list3 *l3 = NULL;
1200    int node = cpu_to_mem(cpu);
1201    int err;
1202
1203    /*
1204     * We need to do this right in the beginning since
1205     * alloc_arraycache's are going to use this list.
1206     * kmalloc_node allows us to add the slab to the right
1207     * kmem_list3 and not this cpu's kmem_list3
1208     */
1209    err = init_cache_nodelists_node(node);
1210    if (err < 0)
1211        goto bad;
1212
1213    /*
1214     * Now we can go ahead with allocating the shared arrays and
1215     * array caches
1216     */
1217    list_for_each_entry(cachep, &cache_chain, next) {
1218        struct array_cache *nc;
1219        struct array_cache *shared = NULL;
1220        struct array_cache **alien = NULL;
1221
1222        nc = alloc_arraycache(node, cachep->limit,
1223                    cachep->batchcount, GFP_KERNEL);
1224        if (!nc)
1225            goto bad;
1226        if (cachep->shared) {
1227            shared = alloc_arraycache(node,
1228                cachep->shared * cachep->batchcount,
1229                0xbaadf00d, GFP_KERNEL);
1230            if (!shared) {
1231                kfree(nc);
1232                goto bad;
1233            }
1234        }
1235        if (use_alien_caches) {
1236            alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1237            if (!alien) {
1238                kfree(shared);
1239                kfree(nc);
1240                goto bad;
1241            }
1242        }
1243        cachep->array[cpu] = nc;
1244        l3 = cachep->nodelists[node];
1245        BUG_ON(!l3);
1246
1247        spin_lock_irq(&l3->list_lock);
1248        if (!l3->shared) {
1249            /*
1250             * We are serialised from CPU_DEAD or
1251             * CPU_UP_CANCELLED by the cpucontrol lock
1252             */
1253            l3->shared = shared;
1254            shared = NULL;
1255        }
1256#ifdef CONFIG_NUMA
1257        if (!l3->alien) {
1258            l3->alien = alien;
1259            alien = NULL;
1260        }
1261#endif
1262        spin_unlock_irq(&l3->list_lock);
1263        kfree(shared);
1264        free_alien_cache(alien);
1265    }
1266    init_node_lock_keys(node);
1267
1268    return 0;
1269bad:
1270    cpuup_canceled(cpu);
1271    return -ENOMEM;
1272}
1273
1274static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1275                    unsigned long action, void *hcpu)
1276{
1277    long cpu = (long)hcpu;
1278    int err = 0;
1279
1280    switch (action) {
1281    case CPU_UP_PREPARE:
1282    case CPU_UP_PREPARE_FROZEN:
1283        mutex_lock(&cache_chain_mutex);
1284        err = cpuup_prepare(cpu);
1285        mutex_unlock(&cache_chain_mutex);
1286        break;
1287    case CPU_ONLINE:
1288    case CPU_ONLINE_FROZEN:
1289        start_cpu_timer(cpu);
1290        break;
1291#ifdef CONFIG_HOTPLUG_CPU
1292      case CPU_DOWN_PREPARE:
1293      case CPU_DOWN_PREPARE_FROZEN:
1294        /*
1295         * Shutdown cache reaper. Note that the cache_chain_mutex is
1296         * held so that if cache_reap() is invoked it cannot do
1297         * anything expensive but will only modify reap_work
1298         * and reschedule the timer.
1299        */
1300        cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1301        /* Now the cache_reaper is guaranteed to be not running. */
1302        per_cpu(slab_reap_work, cpu).work.func = NULL;
1303          break;
1304      case CPU_DOWN_FAILED:
1305      case CPU_DOWN_FAILED_FROZEN:
1306        start_cpu_timer(cpu);
1307          break;
1308    case CPU_DEAD:
1309    case CPU_DEAD_FROZEN:
1310        /*
1311         * Even if all the cpus of a node are down, we don't free the
1312         * kmem_list3 of any cache. This to avoid a race between
1313         * cpu_down, and a kmalloc allocation from another cpu for
1314         * memory from the node of the cpu going down. The list3
1315         * structure is usually allocated from kmem_cache_create() and
1316         * gets destroyed at kmem_cache_destroy().
1317         */
1318        /* fall through */
1319#endif
1320    case CPU_UP_CANCELED:
1321    case CPU_UP_CANCELED_FROZEN:
1322        mutex_lock(&cache_chain_mutex);
1323        cpuup_canceled(cpu);
1324        mutex_unlock(&cache_chain_mutex);
1325        break;
1326    }
1327    return notifier_from_errno(err);
1328}
1329
1330static struct notifier_block __cpuinitdata cpucache_notifier = {
1331    &cpuup_callback, NULL, 0
1332};
1333
1334#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1335/*
1336 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1337 * Returns -EBUSY if all objects cannot be drained so that the node is not
1338 * removed.
1339 *
1340 * Must hold cache_chain_mutex.
1341 */
1342static int __meminit drain_cache_nodelists_node(int node)
1343{
1344    struct kmem_cache *cachep;
1345    int ret = 0;
1346
1347    list_for_each_entry(cachep, &cache_chain, next) {
1348        struct kmem_list3 *l3;
1349
1350        l3 = cachep->nodelists[node];
1351        if (!l3)
1352            continue;
1353
1354        drain_freelist(cachep, l3, l3->free_objects);
1355
1356        if (!list_empty(&l3->slabs_full) ||
1357            !list_empty(&l3->slabs_partial)) {
1358            ret = -EBUSY;
1359            break;
1360        }
1361    }
1362    return ret;
1363}
1364
1365static int __meminit slab_memory_callback(struct notifier_block *self,
1366                    unsigned long action, void *arg)
1367{
1368    struct memory_notify *mnb = arg;
1369    int ret = 0;
1370    int nid;
1371
1372    nid = mnb->status_change_nid;
1373    if (nid < 0)
1374        goto out;
1375
1376    switch (action) {
1377    case MEM_GOING_ONLINE:
1378        mutex_lock(&cache_chain_mutex);
1379        ret = init_cache_nodelists_node(nid);
1380        mutex_unlock(&cache_chain_mutex);
1381        break;
1382    case MEM_GOING_OFFLINE:
1383        mutex_lock(&cache_chain_mutex);
1384        ret = drain_cache_nodelists_node(nid);
1385        mutex_unlock(&cache_chain_mutex);
1386        break;
1387    case MEM_ONLINE:
1388    case MEM_OFFLINE:
1389    case MEM_CANCEL_ONLINE:
1390    case MEM_CANCEL_OFFLINE:
1391        break;
1392    }
1393out:
1394    return notifier_from_errno(ret);
1395}
1396#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1397
1398/*
1399 * swap the static kmem_list3 with kmalloced memory
1400 */
1401static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1402                int nodeid)
1403{
1404    struct kmem_list3 *ptr;
1405
1406    ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1407    BUG_ON(!ptr);
1408
1409    memcpy(ptr, list, sizeof(struct kmem_list3));
1410    /*
1411     * Do not assume that spinlocks can be initialized via memcpy:
1412     */
1413    spin_lock_init(&ptr->list_lock);
1414
1415    MAKE_ALL_LISTS(cachep, ptr, nodeid);
1416    cachep->nodelists[nodeid] = ptr;
1417}
1418
1419/*
1420 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1421 * size of kmem_list3.
1422 */
1423static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1424{
1425    int node;
1426
1427    for_each_online_node(node) {
1428        cachep->nodelists[node] = &initkmem_list3[index + node];
1429        cachep->nodelists[node]->next_reap = jiffies +
1430            REAPTIMEOUT_LIST3 +
1431            ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1432    }
1433}
1434
1435/*
1436 * Initialisation. Called after the page allocator have been initialised and
1437 * before smp_init().
1438 */
1439void __init kmem_cache_init(void)
1440{
1441    size_t left_over;
1442    struct cache_sizes *sizes;
1443    struct cache_names *names;
1444    int i;
1445    int order;
1446    int node;
1447
1448    if (num_possible_nodes() == 1)
1449        use_alien_caches = 0;
1450
1451    for (i = 0; i < NUM_INIT_LISTS; i++) {
1452        kmem_list3_init(&initkmem_list3[i]);
1453        if (i < MAX_NUMNODES)
1454            cache_cache.nodelists[i] = NULL;
1455    }
1456    set_up_list3s(&cache_cache, CACHE_CACHE);
1457
1458    /*
1459     * Fragmentation resistance on low memory - only use bigger
1460     * page orders on machines with more than 32MB of memory.
1461     */
1462    if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1463        slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1464
1465    /* Bootstrap is tricky, because several objects are allocated
1466     * from caches that do not exist yet:
1467     * 1) initialize the cache_cache cache: it contains the struct
1468     * kmem_cache structures of all caches, except cache_cache itself:
1469     * cache_cache is statically allocated.
1470     * Initially an __init data area is used for the head array and the
1471     * kmem_list3 structures, it's replaced with a kmalloc allocated
1472     * array at the end of the bootstrap.
1473     * 2) Create the first kmalloc cache.
1474     * The struct kmem_cache for the new cache is allocated normally.
1475     * An __init data area is used for the head array.
1476     * 3) Create the remaining kmalloc caches, with minimally sized
1477     * head arrays.
1478     * 4) Replace the __init data head arrays for cache_cache and the first
1479     * kmalloc cache with kmalloc allocated arrays.
1480     * 5) Replace the __init data for kmem_list3 for cache_cache and
1481     * the other cache's with kmalloc allocated memory.
1482     * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1483     */
1484
1485    node = numa_mem_id();
1486
1487    /* 1) create the cache_cache */
1488    INIT_LIST_HEAD(&cache_chain);
1489    list_add(&cache_cache.next, &cache_chain);
1490    cache_cache.colour_off = cache_line_size();
1491    cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1492    cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1493
1494    /*
1495     * struct kmem_cache size depends on nr_node_ids, which
1496     * can be less than MAX_NUMNODES.
1497     */
1498    cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1499                 nr_node_ids * sizeof(struct kmem_list3 *);
1500#if DEBUG
1501    cache_cache.obj_size = cache_cache.buffer_size;
1502#endif
1503    cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1504                    cache_line_size());
1505    cache_cache.reciprocal_buffer_size =
1506        reciprocal_value(cache_cache.buffer_size);
1507
1508    for (order = 0; order < MAX_ORDER; order++) {
1509        cache_estimate(order, cache_cache.buffer_size,
1510            cache_line_size(), 0, &left_over, &cache_cache.num);
1511        if (cache_cache.num)
1512            break;
1513    }
1514    BUG_ON(!cache_cache.num);
1515    cache_cache.gfporder = order;
1516    cache_cache.colour = left_over / cache_cache.colour_off;
1517    cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1518                      sizeof(struct slab), cache_line_size());
1519
1520    /* 2+3) create the kmalloc caches */
1521    sizes = malloc_sizes;
1522    names = cache_names;
1523
1524    /*
1525     * Initialize the caches that provide memory for the array cache and the
1526     * kmem_list3 structures first. Without this, further allocations will
1527     * bug.
1528     */
1529
1530    sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1531                    sizes[INDEX_AC].cs_size,
1532                    ARCH_KMALLOC_MINALIGN,
1533                    ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1534                    NULL);
1535
1536    if (INDEX_AC != INDEX_L3) {
1537        sizes[INDEX_L3].cs_cachep =
1538            kmem_cache_create(names[INDEX_L3].name,
1539                sizes[INDEX_L3].cs_size,
1540                ARCH_KMALLOC_MINALIGN,
1541                ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1542                NULL);
1543    }
1544
1545    slab_early_init = 0;
1546
1547    while (sizes->cs_size != ULONG_MAX) {
1548        /*
1549         * For performance, all the general caches are L1 aligned.
1550         * This should be particularly beneficial on SMP boxes, as it
1551         * eliminates "false sharing".
1552         * Note for systems short on memory removing the alignment will
1553         * allow tighter packing of the smaller caches.
1554         */
1555        if (!sizes->cs_cachep) {
1556            sizes->cs_cachep = kmem_cache_create(names->name,
1557                    sizes->cs_size,
1558                    ARCH_KMALLOC_MINALIGN,
1559                    ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1560                    NULL);
1561        }
1562#ifdef CONFIG_ZONE_DMA
1563        sizes->cs_dmacachep = kmem_cache_create(
1564                    names->name_dma,
1565                    sizes->cs_size,
1566                    ARCH_KMALLOC_MINALIGN,
1567                    ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1568                        SLAB_PANIC,
1569                    NULL);
1570#endif
1571        sizes++;
1572        names++;
1573    }
1574    /* 4) Replace the bootstrap head arrays */
1575    {
1576        struct array_cache *ptr;
1577
1578        ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1579
1580        BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1581        memcpy(ptr, cpu_cache_get(&cache_cache),
1582               sizeof(struct arraycache_init));
1583        /*
1584         * Do not assume that spinlocks can be initialized via memcpy:
1585         */
1586        spin_lock_init(&ptr->lock);
1587
1588        cache_cache.array[smp_processor_id()] = ptr;
1589
1590        ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1591
1592        BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1593               != &initarray_generic.cache);
1594        memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1595               sizeof(struct arraycache_init));
1596        /*
1597         * Do not assume that spinlocks can be initialized via memcpy:
1598         */
1599        spin_lock_init(&ptr->lock);
1600
1601        malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1602            ptr;
1603    }
1604    /* 5) Replace the bootstrap kmem_list3's */
1605    {
1606        int nid;
1607
1608        for_each_online_node(nid) {
1609            init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1610
1611            init_list(malloc_sizes[INDEX_AC].cs_cachep,
1612                  &initkmem_list3[SIZE_AC + nid], nid);
1613
1614            if (INDEX_AC != INDEX_L3) {
1615                init_list(malloc_sizes[INDEX_L3].cs_cachep,
1616                      &initkmem_list3[SIZE_L3 + nid], nid);
1617            }
1618        }
1619    }
1620
1621    g_cpucache_up = EARLY;
1622}
1623
1624void __init kmem_cache_init_late(void)
1625{
1626    struct kmem_cache *cachep;
1627
1628    /* 6) resize the head arrays to their final sizes */
1629    mutex_lock(&cache_chain_mutex);
1630    list_for_each_entry(cachep, &cache_chain, next)
1631        if (enable_cpucache(cachep, GFP_NOWAIT))
1632            BUG();
1633    mutex_unlock(&cache_chain_mutex);
1634
1635    /* Done! */
1636    g_cpucache_up = FULL;
1637
1638    /* Annotate slab for lockdep -- annotate the malloc caches */
1639    init_lock_keys();
1640
1641    /*
1642     * Register a cpu startup notifier callback that initializes
1643     * cpu_cache_get for all new cpus
1644     */
1645    register_cpu_notifier(&cpucache_notifier);
1646
1647#ifdef CONFIG_NUMA
1648    /*
1649     * Register a memory hotplug callback that initializes and frees
1650     * nodelists.
1651     */
1652    hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1653#endif
1654
1655    /*
1656     * The reap timers are started later, with a module init call: That part
1657     * of the kernel is not yet operational.
1658     */
1659}
1660
1661static int __init cpucache_init(void)
1662{
1663    int cpu;
1664
1665    /*
1666     * Register the timers that return unneeded pages to the page allocator
1667     */
1668    for_each_online_cpu(cpu)
1669        start_cpu_timer(cpu);
1670    return 0;
1671}
1672__initcall(cpucache_init);
1673
1674/*
1675 * Interface to system's page allocator. No need to hold the cache-lock.
1676 *
1677 * If we requested dmaable memory, we will get it. Even if we
1678 * did not request dmaable memory, we might get it, but that
1679 * would be relatively rare and ignorable.
1680 */
1681static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1682{
1683    struct page *page;
1684    int nr_pages;
1685    int i;
1686
1687#ifndef CONFIG_MMU
1688    /*
1689     * Nommu uses slab's for process anonymous memory allocations, and thus
1690     * requires __GFP_COMP to properly refcount higher order allocations
1691     */
1692    flags |= __GFP_COMP;
1693#endif
1694
1695    flags |= cachep->gfpflags;
1696    if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1697        flags |= __GFP_RECLAIMABLE;
1698
1699    page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1700    if (!page)
1701        return NULL;
1702
1703    nr_pages = (1 << cachep->gfporder);
1704    if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1705        add_zone_page_state(page_zone(page),
1706            NR_SLAB_RECLAIMABLE, nr_pages);
1707    else
1708        add_zone_page_state(page_zone(page),
1709            NR_SLAB_UNRECLAIMABLE, nr_pages);
1710    for (i = 0; i < nr_pages; i++)
1711        __SetPageSlab(page + i);
1712
1713    if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1714        kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1715
1716        if (cachep->ctor)
1717            kmemcheck_mark_uninitialized_pages(page, nr_pages);
1718        else
1719            kmemcheck_mark_unallocated_pages(page, nr_pages);
1720    }
1721
1722    return page_address(page);
1723}
1724
1725/*
1726 * Interface to system's page release.
1727 */
1728static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1729{
1730    unsigned long i = (1 << cachep->gfporder);
1731    struct page *page = virt_to_page(addr);
1732    const unsigned long nr_freed = i;
1733
1734    kmemcheck_free_shadow(page, cachep->gfporder);
1735
1736    if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1737        sub_zone_page_state(page_zone(page),
1738                NR_SLAB_RECLAIMABLE, nr_freed);
1739    else
1740        sub_zone_page_state(page_zone(page),
1741                NR_SLAB_UNRECLAIMABLE, nr_freed);
1742    while (i--) {
1743        BUG_ON(!PageSlab(page));
1744        __ClearPageSlab(page);
1745        page++;
1746    }
1747    if (current->reclaim_state)
1748        current->reclaim_state->reclaimed_slab += nr_freed;
1749    free_pages((unsigned long)addr, cachep->gfporder);
1750}
1751
1752static void kmem_rcu_free(struct rcu_head *head)
1753{
1754    struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1755    struct kmem_cache *cachep = slab_rcu->cachep;
1756
1757    kmem_freepages(cachep, slab_rcu->addr);
1758    if (OFF_SLAB(cachep))
1759        kmem_cache_free(cachep->slabp_cache, slab_rcu);
1760}
1761
1762#if DEBUG
1763
1764#ifdef CONFIG_DEBUG_PAGEALLOC
1765static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1766                unsigned long caller)
1767{
1768    int size = obj_size(cachep);
1769
1770    addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1771
1772    if (size < 5 * sizeof(unsigned long))
1773        return;
1774
1775    *addr++ = 0x12345678;
1776    *addr++ = caller;
1777    *addr++ = smp_processor_id();
1778    size -= 3 * sizeof(unsigned long);
1779    {
1780        unsigned long *sptr = &caller;
1781        unsigned long svalue;
1782
1783        while (!kstack_end(sptr)) {
1784            svalue = *sptr++;
1785            if (kernel_text_address(svalue)) {
1786                *addr++ = svalue;
1787                size -= sizeof(unsigned long);
1788                if (size <= sizeof(unsigned long))
1789                    break;
1790            }
1791        }
1792
1793    }
1794    *addr++ = 0x87654321;
1795}
1796#endif
1797
1798static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1799{
1800    int size = obj_size(cachep);
1801    addr = &((char *)addr)[obj_offset(cachep)];
1802
1803    memset(addr, val, size);
1804    *(unsigned char *)(addr + size - 1) = POISON_END;
1805}
1806
1807static void dump_line(char *data, int offset, int limit)
1808{
1809    int i;
1810    unsigned char error = 0;
1811    int bad_count = 0;
1812
1813    printk(KERN_ERR "%03x:", offset);
1814    for (i = 0; i < limit; i++) {
1815        if (data[offset + i] != POISON_FREE) {
1816            error = data[offset + i];
1817            bad_count++;
1818        }
1819        printk(" %02x", (unsigned char)data[offset + i]);
1820    }
1821    printk("\n");
1822
1823    if (bad_count == 1) {
1824        error ^= POISON_FREE;
1825        if (!(error & (error - 1))) {
1826            printk(KERN_ERR "Single bit error detected. Probably "
1827                    "bad RAM.\n");
1828#ifdef CONFIG_X86
1829            printk(KERN_ERR "Run memtest86+ or a similar memory "
1830                    "test tool.\n");
1831#else
1832            printk(KERN_ERR "Run a memory test tool.\n");
1833#endif
1834        }
1835    }
1836}
1837#endif
1838
1839#if DEBUG
1840
1841static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1842{
1843    int i, size;
1844    char *realobj;
1845
1846    if (cachep->flags & SLAB_RED_ZONE) {
1847        printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1848            *dbg_redzone1(cachep, objp),
1849            *dbg_redzone2(cachep, objp));
1850    }
1851
1852    if (cachep->flags & SLAB_STORE_USER) {
1853        printk(KERN_ERR "Last user: [<%p>]",
1854            *dbg_userword(cachep, objp));
1855        print_symbol("(%s)",
1856                (unsigned long)*dbg_userword(cachep, objp));
1857        printk("\n");
1858    }
1859    realobj = (char *)objp + obj_offset(cachep);
1860    size = obj_size(cachep);
1861    for (i = 0; i < size && lines; i += 16, lines--) {
1862        int limit;
1863        limit = 16;
1864        if (i + limit > size)
1865            limit = size - i;
1866        dump_line(realobj, i, limit);
1867    }
1868}
1869
1870static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1871{
1872    char *realobj;
1873    int size, i;
1874    int lines = 0;
1875
1876    realobj = (char *)objp + obj_offset(cachep);
1877    size = obj_size(cachep);
1878
1879    for (i = 0; i < size; i++) {
1880        char exp = POISON_FREE;
1881        if (i == size - 1)
1882            exp = POISON_END;
1883        if (realobj[i] != exp) {
1884            int limit;
1885            /* Mismatch ! */
1886            /* Print header */
1887            if (lines == 0) {
1888                printk(KERN_ERR
1889                    "Slab corruption: %s start=%p, len=%d\n",
1890                    cachep->name, realobj, size);
1891                print_objinfo(cachep, objp, 0);
1892            }
1893            /* Hexdump the affected line */
1894            i = (i / 16) * 16;
1895            limit = 16;
1896            if (i + limit > size)
1897                limit = size - i;
1898            dump_line(realobj, i, limit);
1899            i += 16;
1900            lines++;
1901            /* Limit to 5 lines */
1902            if (lines > 5)
1903                break;
1904        }
1905    }
1906    if (lines != 0) {
1907        /* Print some data about the neighboring objects, if they
1908         * exist:
1909         */
1910        struct slab *slabp = virt_to_slab(objp);
1911        unsigned int objnr;
1912
1913        objnr = obj_to_index(cachep, slabp, objp);
1914        if (objnr) {
1915            objp = index_to_obj(cachep, slabp, objnr - 1);
1916            realobj = (char *)objp + obj_offset(cachep);
1917            printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1918                   realobj, size);
1919            print_objinfo(cachep, objp, 2);
1920        }
1921        if (objnr + 1 < cachep->num) {
1922            objp = index_to_obj(cachep, slabp, objnr + 1);
1923            realobj = (char *)objp + obj_offset(cachep);
1924            printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1925                   realobj, size);
1926            print_objinfo(cachep, objp, 2);
1927        }
1928    }
1929}
1930#endif
1931
1932#if DEBUG
1933static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1934{
1935    int i;
1936    for (i = 0; i < cachep->num; i++) {
1937        void *objp = index_to_obj(cachep, slabp, i);
1938
1939        if (cachep->flags & SLAB_POISON) {
1940#ifdef CONFIG_DEBUG_PAGEALLOC
1941            if (cachep->buffer_size % PAGE_SIZE == 0 &&
1942                    OFF_SLAB(cachep))
1943                kernel_map_pages(virt_to_page(objp),
1944                    cachep->buffer_size / PAGE_SIZE, 1);
1945            else
1946                check_poison_obj(cachep, objp);
1947#else
1948            check_poison_obj(cachep, objp);
1949#endif
1950        }
1951        if (cachep->flags & SLAB_RED_ZONE) {
1952            if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1953                slab_error(cachep, "start of a freed object "
1954                       "was overwritten");
1955            if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1956                slab_error(cachep, "end of a freed object "
1957                       "was overwritten");
1958        }
1959    }
1960}
1961#else
1962static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1963{
1964}
1965#endif
1966
1967/**
1968 * slab_destroy - destroy and release all objects in a slab
1969 * @cachep: cache pointer being destroyed
1970 * @slabp: slab pointer being destroyed
1971 *
1972 * Destroy all the objs in a slab, and release the mem back to the system.
1973 * Before calling the slab must have been unlinked from the cache. The
1974 * cache-lock is not held/needed.
1975 */
1976static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1977{
1978    void *addr = slabp->s_mem - slabp->colouroff;
1979
1980    slab_destroy_debugcheck(cachep, slabp);
1981    if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1982        struct slab_rcu *slab_rcu;
1983
1984        slab_rcu = (struct slab_rcu *)slabp;
1985        slab_rcu->cachep = cachep;
1986        slab_rcu->addr = addr;
1987        call_rcu(&slab_rcu->head, kmem_rcu_free);
1988    } else {
1989        kmem_freepages(cachep, addr);
1990        if (OFF_SLAB(cachep))
1991            kmem_cache_free(cachep->slabp_cache, slabp);
1992    }
1993}
1994
1995static void __kmem_cache_destroy(struct kmem_cache *cachep)
1996{
1997    int i;
1998    struct kmem_list3 *l3;
1999
2000    for_each_online_cpu(i)
2001        kfree(cachep->array[i]);
2002
2003    /* NUMA: free the list3 structures */
2004    for_each_online_node(i) {
2005        l3 = cachep->nodelists[i];
2006        if (l3) {
2007            kfree(l3->shared);
2008            free_alien_cache(l3->alien);
2009            kfree(l3);
2010        }
2011    }
2012    kmem_cache_free(&cache_cache, cachep);
2013}
2014
2015
2016/**
2017 * calculate_slab_order - calculate size (page order) of slabs
2018 * @cachep: pointer to the cache that is being created
2019 * @size: size of objects to be created in this cache.
2020 * @align: required alignment for the objects.
2021 * @flags: slab allocation flags
2022 *
2023 * Also calculates the number of objects per slab.
2024 *
2025 * This could be made much more intelligent. For now, try to avoid using
2026 * high order pages for slabs. When the gfp() functions are more friendly
2027 * towards high-order requests, this should be changed.
2028 */
2029static size_t calculate_slab_order(struct kmem_cache *cachep,
2030            size_t size, size_t align, unsigned long flags)
2031{
2032    unsigned long offslab_limit;
2033    size_t left_over = 0;
2034    int gfporder;
2035
2036    for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2037        unsigned int num;
2038        size_t remainder;
2039
2040        cache_estimate(gfporder, size, align, flags, &remainder, &num);
2041        if (!num)
2042            continue;
2043
2044        if (flags & CFLGS_OFF_SLAB) {
2045            /*
2046             * Max number of objs-per-slab for caches which
2047             * use off-slab slabs. Needed to avoid a possible
2048             * looping condition in cache_grow().
2049             */
2050            offslab_limit = size - sizeof(struct slab);
2051            offslab_limit /= sizeof(kmem_bufctl_t);
2052
2053             if (num > offslab_limit)
2054                break;
2055        }
2056
2057        /* Found something acceptable - save it away */
2058        cachep->num = num;
2059        cachep->gfporder = gfporder;
2060        left_over = remainder;
2061
2062        /*
2063         * A VFS-reclaimable slab tends to have most allocations
2064         * as GFP_NOFS and we really don't want to have to be allocating
2065         * higher-order pages when we are unable to shrink dcache.
2066         */
2067        if (flags & SLAB_RECLAIM_ACCOUNT)
2068            break;
2069
2070        /*
2071         * Large number of objects is good, but very large slabs are
2072         * currently bad for the gfp()s.
2073         */
2074        if (gfporder >= slab_break_gfp_order)
2075            break;
2076
2077        /*
2078         * Acceptable internal fragmentation?
2079         */
2080        if (left_over * 8 <= (PAGE_SIZE << gfporder))
2081            break;
2082    }
2083    return left_over;
2084}
2085
2086static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2087{
2088    if (g_cpucache_up == FULL)
2089        return enable_cpucache(cachep, gfp);
2090
2091    if (g_cpucache_up == NONE) {
2092        /*
2093         * Note: the first kmem_cache_create must create the cache
2094         * that's used by kmalloc(24), otherwise the creation of
2095         * further caches will BUG().
2096         */
2097        cachep->array[smp_processor_id()] = &initarray_generic.cache;
2098
2099        /*
2100         * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2101         * the first cache, then we need to set up all its list3s,
2102         * otherwise the creation of further caches will BUG().
2103         */
2104        set_up_list3s(cachep, SIZE_AC);
2105        if (INDEX_AC == INDEX_L3)
2106            g_cpucache_up = PARTIAL_L3;
2107        else
2108            g_cpucache_up = PARTIAL_AC;
2109    } else {
2110        cachep->array[smp_processor_id()] =
2111            kmalloc(sizeof(struct arraycache_init), gfp);
2112
2113        if (g_cpucache_up == PARTIAL_AC) {
2114            set_up_list3s(cachep, SIZE_L3);
2115            g_cpucache_up = PARTIAL_L3;
2116        } else {
2117            int node;
2118            for_each_online_node(node) {
2119                cachep->nodelists[node] =
2120                    kmalloc_node(sizeof(struct kmem_list3),
2121                        gfp, node);
2122                BUG_ON(!cachep->nodelists[node]);
2123                kmem_list3_init(cachep->nodelists[node]);
2124            }
2125        }
2126    }
2127    cachep->nodelists[numa_mem_id()]->next_reap =
2128            jiffies + REAPTIMEOUT_LIST3 +
2129            ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2130
2131    cpu_cache_get(cachep)->avail = 0;
2132    cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2133    cpu_cache_get(cachep)->batchcount = 1;
2134    cpu_cache_get(cachep)->touched = 0;
2135    cachep->batchcount = 1;
2136    cachep->limit = BOOT_CPUCACHE_ENTRIES;
2137    return 0;
2138}
2139
2140/**
2141 * kmem_cache_create - Create a cache.
2142 * @name: A string which is used in /proc/slabinfo to identify this cache.
2143 * @size: The size of objects to be created in this cache.
2144 * @align: The required alignment for the objects.
2145 * @flags: SLAB flags
2146 * @ctor: A constructor for the objects.
2147 *
2148 * Returns a ptr to the cache on success, NULL on failure.
2149 * Cannot be called within a int, but can be interrupted.
2150 * The @ctor is run when new pages are allocated by the cache.
2151 *
2152 * @name must be valid until the cache is destroyed. This implies that
2153 * the module calling this has to destroy the cache before getting unloaded.
2154 *
2155 * The flags are
2156 *
2157 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2158 * to catch references to uninitialised memory.
2159 *
2160 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2161 * for buffer overruns.
2162 *
2163 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2164 * cacheline. This can be beneficial if you're counting cycles as closely
2165 * as davem.
2166 */
2167struct kmem_cache *
2168kmem_cache_create (const char *name, size_t size, size_t align,
2169    unsigned long flags, void (*ctor)(void *))
2170{
2171    size_t left_over, slab_size, ralign;
2172    struct kmem_cache *cachep = NULL, *pc;
2173    gfp_t gfp;
2174
2175    /*
2176     * Sanity checks... these are all serious usage bugs.
2177     */
2178    if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2179        size > KMALLOC_MAX_SIZE) {
2180        printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2181                name);
2182        BUG();
2183    }
2184
2185    /*
2186     * We use cache_chain_mutex to ensure a consistent view of
2187     * cpu_online_mask as well. Please see cpuup_callback
2188     */
2189    if (slab_is_available()) {
2190        get_online_cpus();
2191        mutex_lock(&cache_chain_mutex);
2192    }
2193
2194    list_for_each_entry(pc, &cache_chain, next) {
2195        char tmp;
2196        int res;
2197
2198        /*
2199         * This happens when the module gets unloaded and doesn't
2200         * destroy its slab cache and no-one else reuses the vmalloc
2201         * area of the module. Print a warning.
2202         */
2203        res = probe_kernel_address(pc->name, tmp);
2204        if (res) {
2205            printk(KERN_ERR
2206                   "SLAB: cache with size %d has lost its name\n",
2207                   pc->buffer_size);
2208            continue;
2209        }
2210
2211        if (!strcmp(pc->name, name)) {
2212            printk(KERN_ERR
2213                   "kmem_cache_create: duplicate cache %s\n", name);
2214            dump_stack();
2215            goto oops;
2216        }
2217    }
2218
2219#if DEBUG
2220    WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2221#if FORCED_DEBUG
2222    /*
2223     * Enable redzoning and last user accounting, except for caches with
2224     * large objects, if the increased size would increase the object size
2225     * above the next power of two: caches with object sizes just above a
2226     * power of two have a significant amount of internal fragmentation.
2227     */
2228    if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2229                        2 * sizeof(unsigned long long)))
2230        flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2231    if (!(flags & SLAB_DESTROY_BY_RCU))
2232        flags |= SLAB_POISON;
2233#endif
2234    if (flags & SLAB_DESTROY_BY_RCU)
2235        BUG_ON(flags & SLAB_POISON);
2236#endif
2237    /*
2238     * Always checks flags, a caller might be expecting debug support which
2239     * isn't available.
2240     */
2241    BUG_ON(flags & ~CREATE_MASK);
2242
2243    /*
2244     * Check that size is in terms of words. This is needed to avoid
2245     * unaligned accesses for some archs when redzoning is used, and makes
2246     * sure any on-slab bufctl's are also correctly aligned.
2247     */
2248    if (size & (BYTES_PER_WORD - 1)) {
2249        size += (BYTES_PER_WORD - 1);
2250        size &= ~(BYTES_PER_WORD - 1);
2251    }
2252
2253    /* calculate the final buffer alignment: */
2254
2255    /* 1) arch recommendation: can be overridden for debug */
2256    if (flags & SLAB_HWCACHE_ALIGN) {
2257        /*
2258         * Default alignment: as specified by the arch code. Except if
2259         * an object is really small, then squeeze multiple objects into
2260         * one cacheline.
2261         */
2262        ralign = cache_line_size();
2263        while (size <= ralign / 2)
2264            ralign /= 2;
2265    } else {
2266        ralign = BYTES_PER_WORD;
2267    }
2268
2269    /*
2270     * Redzoning and user store require word alignment or possibly larger.
2271     * Note this will be overridden by architecture or caller mandated
2272     * alignment if either is greater than BYTES_PER_WORD.
2273     */
2274    if (flags & SLAB_STORE_USER)
2275        ralign = BYTES_PER_WORD;
2276
2277    if (flags & SLAB_RED_ZONE) {
2278        ralign = REDZONE_ALIGN;
2279        /* If redzoning, ensure that the second redzone is suitably
2280         * aligned, by adjusting the object size accordingly. */
2281        size += REDZONE_ALIGN - 1;
2282        size &= ~(REDZONE_ALIGN - 1);
2283    }
2284
2285    /* 2) arch mandated alignment */
2286    if (ralign < ARCH_SLAB_MINALIGN) {
2287        ralign = ARCH_SLAB_MINALIGN;
2288    }
2289    /* 3) caller mandated alignment */
2290    if (ralign < align) {
2291        ralign = align;
2292    }
2293    /* disable debug if necessary */
2294    if (ralign > __alignof__(unsigned long long))
2295        flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2296    /*
2297     * 4) Store it.
2298     */
2299    align = ralign;
2300
2301    if (slab_is_available())
2302        gfp = GFP_KERNEL;
2303    else
2304        gfp = GFP_NOWAIT;
2305
2306    /* Get cache's description obj. */
2307    cachep = kmem_cache_zalloc(&cache_cache, gfp);
2308    if (!cachep)
2309        goto oops;
2310
2311#if DEBUG
2312    cachep->obj_size = size;
2313
2314    /*
2315     * Both debugging options require word-alignment which is calculated
2316     * into align above.
2317     */
2318    if (flags & SLAB_RED_ZONE) {
2319        /* add space for red zone words */
2320        cachep->obj_offset += sizeof(unsigned long long);
2321        size += 2 * sizeof(unsigned long long);
2322    }
2323    if (flags & SLAB_STORE_USER) {
2324        /* user store requires one word storage behind the end of
2325         * the real object. But if the second red zone needs to be
2326         * aligned to 64 bits, we must allow that much space.
2327         */
2328        if (flags & SLAB_RED_ZONE)
2329            size += REDZONE_ALIGN;
2330        else
2331            size += BYTES_PER_WORD;
2332    }
2333#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2334    if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2335        && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2336        cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2337        size = PAGE_SIZE;
2338    }
2339#endif
2340#endif
2341
2342    /*
2343     * Determine if the slab management is 'on' or 'off' slab.
2344     * (bootstrapping cannot cope with offslab caches so don't do
2345     * it too early on. Always use on-slab management when
2346     * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2347     */
2348    if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2349        !(flags & SLAB_NOLEAKTRACE))
2350        /*
2351         * Size is large, assume best to place the slab management obj
2352         * off-slab (should allow better packing of objs).
2353         */
2354        flags |= CFLGS_OFF_SLAB;
2355
2356    size = ALIGN(size, align);
2357
2358    left_over = calculate_slab_order(cachep, size, align, flags);
2359
2360    if (!cachep->num) {
2361        printk(KERN_ERR
2362               "kmem_cache_create: couldn't create cache %s.\n", name);
2363        kmem_cache_free(&cache_cache, cachep);
2364        cachep = NULL;
2365        goto oops;
2366    }
2367    slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2368              + sizeof(struct slab), align);
2369
2370    /*
2371     * If the slab has been placed off-slab, and we have enough space then
2372     * move it on-slab. This is at the expense of any extra colouring.
2373     */
2374    if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2375        flags &= ~CFLGS_OFF_SLAB;
2376        left_over -= slab_size;
2377    }
2378
2379    if (flags & CFLGS_OFF_SLAB) {
2380        /* really off slab. No need for manual alignment */
2381        slab_size =
2382            cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2383
2384#ifdef CONFIG_PAGE_POISONING
2385        /* If we're going to use the generic kernel_map_pages()
2386         * poisoning, then it's going to smash the contents of
2387         * the redzone and userword anyhow, so switch them off.
2388         */
2389        if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2390            flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2391#endif
2392    }
2393
2394    cachep->colour_off = cache_line_size();
2395    /* Offset must be a multiple of the alignment. */
2396    if (cachep->colour_off < align)
2397        cachep->colour_off = align;
2398    cachep->colour = left_over / cachep->colour_off;
2399    cachep->slab_size = slab_size;
2400    cachep->flags = flags;
2401    cachep->gfpflags = 0;
2402    if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2403        cachep->gfpflags |= GFP_DMA;
2404    cachep->buffer_size = size;
2405    cachep->reciprocal_buffer_size = reciprocal_value(size);
2406
2407    if (flags & CFLGS_OFF_SLAB) {
2408        cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2409        /*
2410         * This is a possibility for one of the malloc_sizes caches.
2411         * But since we go off slab only for object size greater than
2412         * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2413         * this should not happen at all.
2414         * But leave a BUG_ON for some lucky dude.
2415         */
2416        BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2417    }
2418    cachep->ctor = ctor;
2419    cachep->name = name;
2420
2421    if (setup_cpu_cache(cachep, gfp)) {
2422        __kmem_cache_destroy(cachep);
2423        cachep = NULL;
2424        goto oops;
2425    }
2426
2427    /* cache setup completed, link it into the list */
2428    list_add(&cachep->next, &cache_chain);
2429oops:
2430    if (!cachep && (flags & SLAB_PANIC))
2431        panic("kmem_cache_create(): failed to create slab `%s'\n",
2432              name);
2433    if (slab_is_available()) {
2434        mutex_unlock(&cache_chain_mutex);
2435        put_online_cpus();
2436    }
2437    return cachep;
2438}
2439EXPORT_SYMBOL(kmem_cache_create);
2440
2441#if DEBUG
2442static void check_irq_off(void)
2443{
2444    BUG_ON(!irqs_disabled());
2445}
2446
2447static void check_irq_on(void)
2448{
2449    BUG_ON(irqs_disabled());
2450}
2451
2452static void check_spinlock_acquired(struct kmem_cache *cachep)
2453{
2454#ifdef CONFIG_SMP
2455    check_irq_off();
2456    assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2457#endif
2458}
2459
2460static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2461{
2462#ifdef CONFIG_SMP
2463    check_irq_off();
2464    assert_spin_locked(&cachep->nodelists[node]->list_lock);
2465#endif
2466}
2467
2468#else
2469#define check_irq_off() do { } while(0)
2470#define check_irq_on() do { } while(0)
2471#define check_spinlock_acquired(x) do { } while(0)
2472#define check_spinlock_acquired_node(x, y) do { } while(0)
2473#endif
2474
2475static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2476            struct array_cache *ac,
2477            int force, int node);
2478
2479static void do_drain(void *arg)
2480{
2481    struct kmem_cache *cachep = arg;
2482    struct array_cache *ac;
2483    int node = numa_mem_id();
2484
2485    check_irq_off();
2486    ac = cpu_cache_get(cachep);
2487    spin_lock(&cachep->nodelists[node]->list_lock);
2488    free_block(cachep, ac->entry, ac->avail, node);
2489    spin_unlock(&cachep->nodelists[node]->list_lock);
2490    ac->avail = 0;
2491}
2492
2493static void drain_cpu_caches(struct kmem_cache *cachep)
2494{
2495    struct kmem_list3 *l3;
2496    int node;
2497
2498    on_each_cpu(do_drain, cachep, 1);
2499    check_irq_on();
2500    for_each_online_node(node) {
2501        l3 = cachep->nodelists[node];
2502        if (l3 && l3->alien)
2503            drain_alien_cache(cachep, l3->alien);
2504    }
2505
2506    for_each_online_node(node) {
2507        l3 = cachep->nodelists[node];
2508        if (l3)
2509            drain_array(cachep, l3, l3->shared, 1, node);
2510    }
2511}
2512
2513/*
2514 * Remove slabs from the list of free slabs.
2515 * Specify the number of slabs to drain in tofree.
2516 *
2517 * Returns the actual number of slabs released.
2518 */
2519static int drain_freelist(struct kmem_cache *cache,
2520            struct kmem_list3 *l3, int tofree)
2521{
2522    struct list_head *p;
2523    int nr_freed;
2524    struct slab *slabp;
2525
2526    nr_freed = 0;
2527    while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2528
2529        spin_lock_irq(&l3->list_lock);
2530        p = l3->slabs_free.prev;
2531        if (p == &l3->slabs_free) {
2532            spin_unlock_irq(&l3->list_lock);
2533            goto out;
2534        }
2535
2536        slabp = list_entry(p, struct slab, list);
2537#if DEBUG
2538        BUG_ON(slabp->inuse);
2539#endif
2540        list_del(&slabp->list);
2541        /*
2542         * Safe to drop the lock. The slab is no longer linked
2543         * to the cache.
2544         */
2545        l3->free_objects -= cache->num;
2546        spin_unlock_irq(&l3->list_lock);
2547        slab_destroy(cache, slabp);
2548        nr_freed++;
2549    }
2550out:
2551    return nr_freed;
2552}
2553
2554/* Called with cache_chain_mutex held to protect against cpu hotplug */
2555static int __cache_shrink(struct kmem_cache *cachep)
2556{
2557    int ret = 0, i = 0;
2558    struct kmem_list3 *l3;
2559
2560    drain_cpu_caches(cachep);
2561
2562    check_irq_on();
2563    for_each_online_node(i) {
2564        l3 = cachep->nodelists[i];
2565        if (!l3)
2566            continue;
2567
2568        drain_freelist(cachep, l3, l3->free_objects);
2569
2570        ret += !list_empty(&l3->slabs_full) ||
2571            !list_empty(&l3->slabs_partial);
2572    }
2573    return (ret ? 1 : 0);
2574}
2575
2576/**
2577 * kmem_cache_shrink - Shrink a cache.
2578 * @cachep: The cache to shrink.
2579 *
2580 * Releases as many slabs as possible for a cache.
2581 * To help debugging, a zero exit status indicates all slabs were released.
2582 */
2583int kmem_cache_shrink(struct kmem_cache *cachep)
2584{
2585    int ret;
2586    BUG_ON(!cachep || in_interrupt());
2587
2588    get_online_cpus();
2589    mutex_lock(&cache_chain_mutex);
2590    ret = __cache_shrink(cachep);
2591    mutex_unlock(&cache_chain_mutex);
2592    put_online_cpus();
2593    return ret;
2594}
2595EXPORT_SYMBOL(kmem_cache_shrink);
2596
2597/**
2598 * kmem_cache_destroy - delete a cache
2599 * @cachep: the cache to destroy
2600 *
2601 * Remove a &struct kmem_cache object from the slab cache.
2602 *
2603 * It is expected this function will be called by a module when it is
2604 * unloaded. This will remove the cache completely, and avoid a duplicate
2605 * cache being allocated each time a module is loaded and unloaded, if the
2606 * module doesn't have persistent in-kernel storage across loads and unloads.
2607 *
2608 * The cache must be empty before calling this function.
2609 *
2610 * The caller must guarantee that no one will allocate memory from the cache
2611 * during the kmem_cache_destroy().
2612 */
2613void kmem_cache_destroy(struct kmem_cache *cachep)
2614{
2615    BUG_ON(!cachep || in_interrupt());
2616
2617    /* Find the cache in the chain of caches. */
2618    get_online_cpus();
2619    mutex_lock(&cache_chain_mutex);
2620    /*
2621     * the chain is never empty, cache_cache is never destroyed
2622     */
2623    list_del(&cachep->next);
2624    if (__cache_shrink(cachep)) {
2625        slab_error(cachep, "Can't free all objects");
2626        list_add(&cachep->next, &cache_chain);
2627        mutex_unlock(&cache_chain_mutex);
2628        put_online_cpus();
2629        return;
2630    }
2631
2632    if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2633        rcu_barrier();
2634
2635    __kmem_cache_destroy(cachep);
2636    mutex_unlock(&cache_chain_mutex);
2637    put_online_cpus();
2638}
2639EXPORT_SYMBOL(kmem_cache_destroy);
2640
2641/*
2642 * Get the memory for a slab management obj.
2643 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2644 * always come from malloc_sizes caches. The slab descriptor cannot
2645 * come from the same cache which is getting created because,
2646 * when we are searching for an appropriate cache for these
2647 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2648 * If we are creating a malloc_sizes cache here it would not be visible to
2649 * kmem_find_general_cachep till the initialization is complete.
2650 * Hence we cannot have slabp_cache same as the original cache.
2651 */
2652static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2653                   int colour_off, gfp_t local_flags,
2654                   int nodeid)
2655{
2656    struct slab *slabp;
2657
2658    if (OFF_SLAB(cachep)) {
2659        /* Slab management obj is off-slab. */
2660        slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2661                          local_flags, nodeid);
2662        /*
2663         * If the first object in the slab is leaked (it's allocated
2664         * but no one has a reference to it), we want to make sure
2665         * kmemleak does not treat the ->s_mem pointer as a reference
2666         * to the object. Otherwise we will not report the leak.
2667         */
2668        kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2669                   local_flags);
2670        if (!slabp)
2671            return NULL;
2672    } else {
2673        slabp = objp + colour_off;
2674        colour_off += cachep->slab_size;
2675    }
2676    slabp->inuse = 0;
2677    slabp->colouroff = colour_off;
2678    slabp->s_mem = objp + colour_off;
2679    slabp->nodeid = nodeid;
2680    slabp->free = 0;
2681    return slabp;
2682}
2683
2684static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2685{
2686    return (kmem_bufctl_t *) (slabp + 1);
2687}
2688
2689static void cache_init_objs(struct kmem_cache *cachep,
2690                struct slab *slabp)
2691{
2692    int i;
2693
2694    for (i = 0; i < cachep->num; i++) {
2695        void *objp = index_to_obj(cachep, slabp, i);
2696#if DEBUG
2697        /* need to poison the objs? */
2698        if (cachep->flags & SLAB_POISON)
2699            poison_obj(cachep, objp, POISON_FREE);
2700        if (cachep->flags & SLAB_STORE_USER)
2701            *dbg_userword(cachep, objp) = NULL;
2702
2703        if (cachep->flags & SLAB_RED_ZONE) {
2704            *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2705            *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2706        }
2707        /*
2708         * Constructors are not allowed to allocate memory from the same
2709         * cache which they are a constructor for. Otherwise, deadlock.
2710         * They must also be threaded.
2711         */
2712        if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2713            cachep->ctor(objp + obj_offset(cachep));
2714
2715        if (cachep->flags & SLAB_RED_ZONE) {
2716            if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2717                slab_error(cachep, "constructor overwrote the"
2718                       " end of an object");
2719            if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2720                slab_error(cachep, "constructor overwrote the"
2721                       " start of an object");
2722        }
2723        if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2724                OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2725            kernel_map_pages(virt_to_page(objp),
2726                     cachep->buffer_size / PAGE_SIZE, 0);
2727#else
2728        if (cachep->ctor)
2729            cachep->ctor(objp);
2730#endif
2731        slab_bufctl(slabp)[i] = i + 1;
2732    }
2733    slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2734}
2735
2736static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2737{
2738    if (CONFIG_ZONE_DMA_FLAG) {
2739        if (flags & GFP_DMA)
2740            BUG_ON(!(cachep->gfpflags & GFP_DMA));
2741        else
2742            BUG_ON(cachep->gfpflags & GFP_DMA);
2743    }
2744}
2745
2746static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2747                int nodeid)
2748{
2749    void *objp = index_to_obj(cachep, slabp, slabp->free);
2750    kmem_bufctl_t next;
2751
2752    slabp->inuse++;
2753    next = slab_bufctl(slabp)[slabp->free];
2754#if DEBUG
2755    slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2756    WARN_ON(slabp->nodeid != nodeid);
2757#endif
2758    slabp->free = next;
2759
2760    return objp;
2761}
2762
2763static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2764                void *objp, int nodeid)
2765{
2766    unsigned int objnr = obj_to_index(cachep, slabp, objp);
2767
2768#if DEBUG
2769    /* Verify that the slab belongs to the intended node */
2770    WARN_ON(slabp->nodeid != nodeid);
2771
2772    if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2773        printk(KERN_ERR "slab: double free detected in cache "
2774                "'%s', objp %p\n", cachep->name, objp);
2775        BUG();
2776    }
2777#endif
2778    slab_bufctl(slabp)[objnr] = slabp->free;
2779    slabp->free = objnr;
2780    slabp->inuse--;
2781}
2782
2783/*
2784 * Map pages beginning at addr to the given cache and slab. This is required
2785 * for the slab allocator to be able to lookup the cache and slab of a
2786 * virtual address for kfree, ksize, and slab debugging.
2787 */
2788static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2789               void *addr)
2790{
2791    int nr_pages;
2792    struct page *page;
2793
2794    page = virt_to_page(addr);
2795
2796    nr_pages = 1;
2797    if (likely(!PageCompound(page)))
2798        nr_pages <<= cache->gfporder;
2799
2800    do {
2801        page_set_cache(page, cache);
2802        page_set_slab(page, slab);
2803        page++;
2804    } while (--nr_pages);
2805}
2806
2807/*
2808 * Grow (by 1) the number of slabs within a cache. This is called by
2809 * kmem_cache_alloc() when there are no active objs left in a cache.
2810 */
2811static int cache_grow(struct kmem_cache *cachep,
2812        gfp_t flags, int nodeid, void *objp)
2813{
2814    struct slab *slabp;
2815    size_t offset;
2816    gfp_t local_flags;
2817    struct kmem_list3 *l3;
2818
2819    /*
2820     * Be lazy and only check for valid flags here, keeping it out of the
2821     * critical path in kmem_cache_alloc().
2822     */
2823    BUG_ON(flags & GFP_SLAB_BUG_MASK);
2824    local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2825
2826    /* Take the l3 list lock to change the colour_next on this node */
2827    check_irq_off();
2828    l3 = cachep->nodelists[nodeid];
2829    spin_lock(&l3->list_lock);
2830
2831    /* Get colour for the slab, and cal the next value. */
2832    offset = l3->colour_next;
2833    l3->colour_next++;
2834    if (l3->colour_next >= cachep->colour)
2835        l3->colour_next = 0;
2836    spin_unlock(&l3->list_lock);
2837
2838    offset *= cachep->colour_off;
2839
2840    if (local_flags & __GFP_WAIT)
2841        local_irq_enable();
2842
2843    /*
2844     * The test for missing atomic flag is performed here, rather than
2845     * the more obvious place, simply to reduce the critical path length
2846     * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2847     * will eventually be caught here (where it matters).
2848     */
2849    kmem_flagcheck(cachep, flags);
2850
2851    /*
2852     * Get mem for the objs. Attempt to allocate a physical page from
2853     * 'nodeid'.
2854     */
2855    if (!objp)
2856        objp = kmem_getpages(cachep, local_flags, nodeid);
2857    if (!objp)
2858        goto failed;
2859
2860    /* Get slab management. */
2861    slabp = alloc_slabmgmt(cachep, objp, offset,
2862            local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2863    if (!slabp)
2864        goto opps1;
2865
2866    slab_map_pages(cachep, slabp, objp);
2867
2868    cache_init_objs(cachep, slabp);
2869
2870    if (local_flags & __GFP_WAIT)
2871        local_irq_disable();
2872    check_irq_off();
2873    spin_lock(&l3->list_lock);
2874
2875    /* Make slab active. */
2876    list_add_tail(&slabp->list, &(l3->slabs_free));
2877    STATS_INC_GROWN(cachep);
2878    l3->free_objects += cachep->num;
2879    spin_unlock(&l3->list_lock);
2880    return 1;
2881opps1:
2882    kmem_freepages(cachep, objp);
2883failed:
2884    if (local_flags & __GFP_WAIT)
2885        local_irq_disable();
2886    return 0;
2887}
2888
2889#if DEBUG
2890
2891/*
2892 * Perform extra freeing checks:
2893 * - detect bad pointers.
2894 * - POISON/RED_ZONE checking
2895 */
2896static void kfree_debugcheck(const void *objp)
2897{
2898    if (!virt_addr_valid(objp)) {
2899        printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2900               (unsigned long)objp);
2901        BUG();
2902    }
2903}
2904
2905static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2906{
2907    unsigned long long redzone1, redzone2;
2908
2909    redzone1 = *dbg_redzone1(cache, obj);
2910    redzone2 = *dbg_redzone2(cache, obj);
2911
2912    /*
2913     * Redzone is ok.
2914     */
2915    if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2916        return;
2917
2918    if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2919        slab_error(cache, "double free detected");
2920    else
2921        slab_error(cache, "memory outside object was overwritten");
2922
2923    printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2924            obj, redzone1, redzone2);
2925}
2926
2927static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2928                   void *caller)
2929{
2930    struct page *page;
2931    unsigned int objnr;
2932    struct slab *slabp;
2933
2934    BUG_ON(virt_to_cache(objp) != cachep);
2935
2936    objp -= obj_offset(cachep);
2937    kfree_debugcheck(objp);
2938    page = virt_to_head_page(objp);
2939
2940    slabp = page_get_slab(page);
2941
2942    if (cachep->flags & SLAB_RED_ZONE) {
2943        verify_redzone_free(cachep, objp);
2944        *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2945        *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2946    }
2947    if (cachep->flags & SLAB_STORE_USER)
2948        *dbg_userword(cachep, objp) = caller;
2949
2950    objnr = obj_to_index(cachep, slabp, objp);
2951
2952    BUG_ON(objnr >= cachep->num);
2953    BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2954
2955#ifdef CONFIG_DEBUG_SLAB_LEAK
2956    slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2957#endif
2958    if (cachep->flags & SLAB_POISON) {
2959#ifdef CONFIG_DEBUG_PAGEALLOC
2960        if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2961            store_stackinfo(cachep, objp, (unsigned long)caller);
2962            kernel_map_pages(virt_to_page(objp),
2963                     cachep->buffer_size / PAGE_SIZE, 0);
2964        } else {
2965            poison_obj(cachep, objp, POISON_FREE);
2966        }
2967#else
2968        poison_obj(cachep, objp, POISON_FREE);
2969#endif
2970    }
2971    return objp;
2972}
2973
2974static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2975{
2976    kmem_bufctl_t i;
2977    int entries = 0;
2978
2979    /* Check slab's freelist to see if this obj is there. */
2980    for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2981        entries++;
2982        if (entries > cachep->num || i >= cachep->num)
2983            goto bad;
2984    }
2985    if (entries != cachep->num - slabp->inuse) {
2986bad:
2987        printk(KERN_ERR "slab: Internal list corruption detected in "
2988                "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2989            cachep->name, cachep->num, slabp, slabp->inuse);
2990        for (i = 0;
2991             i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2992             i++) {
2993            if (i % 16 == 0)
2994                printk("\n%03x:", i);
2995            printk(" %02x", ((unsigned char *)slabp)[i]);
2996        }
2997        printk("\n");
2998        BUG();
2999    }
3000}
3001#else
3002#define kfree_debugcheck(x) do { } while(0)
3003#define cache_free_debugcheck(x,objp,z) (objp)
3004#define check_slabp(x,y) do { } while(0)
3005#endif
3006
3007static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3008{
3009    int batchcount;
3010    struct kmem_list3 *l3;
3011    struct array_cache *ac;
3012    int node;
3013
3014retry:
3015    check_irq_off();
3016    node = numa_mem_id();
3017    ac = cpu_cache_get(cachep);
3018    batchcount = ac->batchcount;
3019    if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3020        /*
3021         * If there was little recent activity on this cache, then
3022         * perform only a partial refill. Otherwise we could generate
3023         * refill bouncing.
3024         */
3025        batchcount = BATCHREFILL_LIMIT;
3026    }
3027    l3 = cachep->nodelists[node];
3028
3029    BUG_ON(ac->avail > 0 || !l3);
3030    spin_lock(&l3->list_lock);
3031
3032    /* See if we can refill from the shared array */
3033    if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3034        l3->shared->touched = 1;
3035        goto alloc_done;
3036    }
3037
3038    while (batchcount > 0) {
3039        struct list_head *entry;
3040        struct slab *slabp;
3041        /* Get slab alloc is to come from. */
3042        entry = l3->slabs_partial.next;
3043        if (entry == &l3->slabs_partial) {
3044            l3->free_touched = 1;
3045            entry = l3->slabs_free.next;
3046            if (entry == &l3->slabs_free)
3047                goto must_grow;
3048        }
3049
3050        slabp = list_entry(entry, struct slab, list);
3051        check_slabp(cachep, slabp);
3052        check_spinlock_acquired(cachep);
3053
3054        /*
3055         * The slab was either on partial or free list so
3056         * there must be at least one object available for
3057         * allocation.
3058         */
3059        BUG_ON(slabp->inuse >= cachep->num);
3060
3061        while (slabp->inuse < cachep->num && batchcount--) {
3062            STATS_INC_ALLOCED(cachep);
3063            STATS_INC_ACTIVE(cachep);
3064            STATS_SET_HIGH(cachep);
3065
3066            ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3067                                node);
3068        }
3069        check_slabp(cachep, slabp);
3070
3071        /* move slabp to correct slabp list: */
3072        list_del(&slabp->list);
3073        if (slabp->free == BUFCTL_END)
3074            list_add(&slabp->list, &l3->slabs_full);
3075        else
3076            list_add(&slabp->list, &l3->slabs_partial);
3077    }
3078
3079must_grow:
3080    l3->free_objects -= ac->avail;
3081alloc_done:
3082    spin_unlock(&l3->list_lock);
3083
3084    if (unlikely(!ac->avail)) {
3085        int x;
3086        x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3087
3088        /* cache_grow can reenable interrupts, then ac could change. */
3089        ac = cpu_cache_get(cachep);
3090        if (!x && ac->avail == 0) /* no objects in sight? abort */
3091            return NULL;
3092
3093        if (!ac->avail) /* objects refilled by interrupt? */
3094            goto retry;
3095    }
3096    ac->touched = 1;
3097    return ac->entry[--ac->avail];
3098}
3099
3100static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3101                        gfp_t flags)
3102{
3103    might_sleep_if(flags & __GFP_WAIT);
3104#if DEBUG
3105    kmem_flagcheck(cachep, flags);
3106#endif
3107}
3108
3109#if DEBUG
3110static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3111                gfp_t flags, void *objp, void *caller)
3112{
3113    if (!objp)
3114        return objp;
3115    if (cachep->flags & SLAB_POISON) {
3116#ifdef CONFIG_DEBUG_PAGEALLOC
3117        if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3118            kernel_map_pages(virt_to_page(objp),
3119                     cachep->buffer_size / PAGE_SIZE, 1);
3120        else
3121            check_poison_obj(cachep, objp);
3122#else
3123        check_poison_obj(cachep, objp);
3124#endif
3125        poison_obj(cachep, objp, POISON_INUSE);
3126    }
3127    if (cachep->flags & SLAB_STORE_USER)
3128        *dbg_userword(cachep, objp) = caller;
3129
3130    if (cachep->flags & SLAB_RED_ZONE) {
3131        if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3132                *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3133            slab_error(cachep, "double free, or memory outside"
3134                        " object was overwritten");
3135            printk(KERN_ERR
3136                "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3137                objp, *dbg_redzone1(cachep, objp),
3138                *dbg_redzone2(cachep, objp));
3139        }
3140        *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3141        *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3142    }
3143#ifdef CONFIG_DEBUG_SLAB_LEAK
3144    {
3145        struct slab *slabp;
3146        unsigned objnr;
3147
3148        slabp = page_get_slab(virt_to_head_page(objp));
3149        objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3150        slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3151    }
3152#endif
3153    objp += obj_offset(cachep);
3154    if (cachep->ctor && cachep->flags & SLAB_POISON)
3155        cachep->ctor(objp);
3156#if ARCH_SLAB_MINALIGN
3157    if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3158        printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3159               objp, ARCH_SLAB_MINALIGN);
3160    }
3161#endif
3162    return objp;
3163}
3164#else
3165#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3166#endif
3167
3168static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3169{
3170    if (cachep == &cache_cache)
3171        return false;
3172
3173    return should_failslab(obj_size(cachep), flags, cachep->flags);
3174}
3175
3176static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3177{
3178    void *objp;
3179    struct array_cache *ac;
3180
3181    check_irq_off();
3182
3183    ac = cpu_cache_get(cachep);
3184    if (likely(ac->avail)) {
3185        STATS_INC_ALLOCHIT(cachep);
3186        ac->touched = 1;
3187        objp = ac->entry[--ac->avail];
3188    } else {
3189        STATS_INC_ALLOCMISS(cachep);
3190        objp = cache_alloc_refill(cachep, flags);
3191        /*
3192         * the 'ac' may be updated by cache_alloc_refill(),
3193         * and kmemleak_erase() requires its correct value.
3194         */
3195        ac = cpu_cache_get(cachep);
3196    }
3197    /*
3198     * To avoid a false negative, if an object that is in one of the
3199     * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3200     * treat the array pointers as a reference to the object.
3201     */
3202    if (objp)
3203        kmemleak_erase(&ac->entry[ac->avail]);
3204    return objp;
3205}
3206
3207#ifdef CONFIG_NUMA
3208/*
3209 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3210 *
3211 * If we are in_interrupt, then process context, including cpusets and
3212 * mempolicy, may not apply and should not be used for allocation policy.
3213 */
3214static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3215{
3216    int nid_alloc, nid_here;
3217
3218    if (in_interrupt() || (flags & __GFP_THISNODE))
3219        return NULL;
3220    nid_alloc = nid_here = numa_mem_id();
3221    get_mems_allowed();
3222    if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3223        nid_alloc = cpuset_slab_spread_node();
3224    else if (current->mempolicy)
3225        nid_alloc = slab_node(current->mempolicy);
3226    put_mems_allowed();
3227    if (nid_alloc != nid_here)
3228        return ____cache_alloc_node(cachep, flags, nid_alloc);
3229    return NULL;
3230}
3231
3232/*
3233 * Fallback function if there was no memory available and no objects on a
3234 * certain node and fall back is permitted. First we scan all the
3235 * available nodelists for available objects. If that fails then we
3236 * perform an allocation without specifying a node. This allows the page
3237 * allocator to do its reclaim / fallback magic. We then insert the
3238 * slab into the proper nodelist and then allocate from it.
3239 */
3240static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3241{
3242    struct zonelist *zonelist;
3243    gfp_t local_flags;
3244    struct zoneref *z;
3245    struct zone *zone;
3246    enum zone_type high_zoneidx = gfp_zone(flags);
3247    void *obj = NULL;
3248    int nid;
3249
3250    if (flags & __GFP_THISNODE)
3251        return NULL;
3252
3253    get_mems_allowed();
3254    zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3255    local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3256
3257retry:
3258    /*
3259     * Look through allowed nodes for objects available
3260     * from existing per node queues.
3261     */
3262    for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3263        nid = zone_to_nid(zone);
3264
3265        if (cpuset_zone_allowed_hardwall(zone, flags) &&
3266            cache->nodelists[nid] &&
3267            cache->nodelists[nid]->free_objects) {
3268                obj = ____cache_alloc_node(cache,
3269                    flags | GFP_THISNODE, nid);
3270                if (obj)
3271                    break;
3272        }
3273    }
3274
3275    if (!obj) {
3276        /*
3277         * This allocation will be performed within the constraints
3278         * of the current cpuset / memory policy requirements.
3279         * We may trigger various forms of reclaim on the allowed
3280         * set and go into memory reserves if necessary.
3281         */
3282        if (local_flags & __GFP_WAIT)
3283            local_irq_enable();
3284        kmem_flagcheck(cache, flags);
3285        obj = kmem_getpages(cache, local_flags, numa_mem_id());
3286        if (local_flags & __GFP_WAIT)
3287            local_irq_disable();
3288        if (obj) {
3289            /*
3290             * Insert into the appropriate per node queues
3291             */
3292            nid = page_to_nid(virt_to_page(obj));
3293            if (cache_grow(cache, flags, nid, obj)) {
3294                obj = ____cache_alloc_node(cache,
3295                    flags | GFP_THISNODE, nid);
3296                if (!obj)
3297                    /*
3298                     * Another processor may allocate the
3299                     * objects in the slab since we are
3300                     * not holding any locks.
3301                     */
3302                    goto retry;
3303            } else {
3304                /* cache_grow already freed obj */
3305                obj = NULL;
3306            }
3307        }
3308    }
3309    put_mems_allowed();
3310    return obj;
3311}
3312
3313/*
3314 * A interface to enable slab creation on nodeid
3315 */
3316static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3317                int nodeid)
3318{
3319    struct list_head *entry;
3320    struct slab *slabp;
3321    struct kmem_list3 *l3;
3322    void *obj;
3323    int x;
3324
3325    l3 = cachep->nodelists[nodeid];
3326    BUG_ON(!l3);
3327
3328retry:
3329    check_irq_off();
3330    spin_lock(&l3->list_lock);
3331    entry = l3->slabs_partial.next;
3332    if (entry == &l3->slabs_partial) {
3333        l3->free_touched = 1;
3334        entry = l3->slabs_free.next;
3335        if (entry == &l3->slabs_free)
3336            goto must_grow;
3337    }
3338
3339    slabp = list_entry(entry, struct slab, list);
3340    check_spinlock_acquired_node(cachep, nodeid);
3341    check_slabp(cachep, slabp);
3342
3343    STATS_INC_NODEALLOCS(cachep);
3344    STATS_INC_ACTIVE(cachep);
3345    STATS_SET_HIGH(cachep);
3346
3347    BUG_ON(slabp->inuse == cachep->num);
3348
3349    obj = slab_get_obj(cachep, slabp, nodeid);
3350    check_slabp(cachep, slabp);
3351    l3->free_objects--;
3352    /* move slabp to correct slabp list: */
3353    list_del(&slabp->list);
3354
3355    if (slabp->free == BUFCTL_END)
3356        list_add(&slabp->list, &l3->slabs_full);
3357    else
3358        list_add(&slabp->list, &l3->slabs_partial);
3359
3360    spin_unlock(&l3->list_lock);
3361    goto done;
3362
3363must_grow:
3364    spin_unlock(&l3->list_lock);
3365    x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3366    if (x)
3367        goto retry;
3368
3369    return fallback_alloc(cachep, flags);
3370
3371done:
3372    return obj;
3373}
3374
3375/**
3376 * kmem_cache_alloc_node - Allocate an object on the specified node
3377 * @cachep: The cache to allocate from.
3378 * @flags: See kmalloc().
3379 * @nodeid: node number of the target node.
3380 * @caller: return address of caller, used for debug information
3381 *
3382 * Identical to kmem_cache_alloc but it will allocate memory on the given
3383 * node, which can improve the performance for cpu bound structures.
3384 *
3385 * Fallback to other node is possible if __GFP_THISNODE is not set.
3386 */
3387static __always_inline void *
3388__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3389           void *caller)
3390{
3391    unsigned long save_flags;
3392    void *ptr;
3393    int slab_node = numa_mem_id();
3394
3395    flags &= gfp_allowed_mask;
3396
3397    lockdep_trace_alloc(flags);
3398
3399    if (slab_should_failslab(cachep, flags))
3400        return NULL;
3401
3402    cache_alloc_debugcheck_before(cachep, flags);
3403    local_irq_save(save_flags);
3404
3405    if (nodeid == -1)
3406        nodeid = slab_node;
3407
3408    if (unlikely(!cachep->nodelists[nodeid])) {
3409        /* Node not bootstrapped yet */
3410        ptr = fallback_alloc(cachep, flags);
3411        goto out;
3412    }
3413
3414    if (nodeid == slab_node) {
3415        /*
3416         * Use the locally cached objects if possible.
3417         * However ____cache_alloc does not allow fallback
3418         * to other nodes. It may fail while we still have
3419         * objects on other nodes available.
3420         */
3421        ptr = ____cache_alloc(cachep, flags);
3422        if (ptr)
3423            goto out;
3424    }
3425    /* ___cache_alloc_node can fall back to other nodes */
3426    ptr = ____cache_alloc_node(cachep, flags, nodeid);
3427  out:
3428    local_irq_restore(save_flags);
3429    ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3430    kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3431                 flags);
3432
3433    if (likely(ptr))
3434        kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3435
3436    if (unlikely((flags & __GFP_ZERO) && ptr))
3437        memset(ptr, 0, obj_size(cachep));
3438
3439    return ptr;
3440}
3441
3442static __always_inline void *
3443__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3444{
3445    void *objp;
3446
3447    if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3448        objp = alternate_node_alloc(cache, flags);
3449        if (objp)
3450            goto out;
3451    }
3452    objp = ____cache_alloc(cache, flags);
3453
3454    /*
3455     * We may just have run out of memory on the local node.
3456     * ____cache_alloc_node() knows how to locate memory on other nodes
3457     */
3458    if (!objp)
3459        objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3460
3461  out:
3462    return objp;
3463}
3464#else
3465
3466static __always_inline void *
3467__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3468{
3469    return ____cache_alloc(cachep, flags);
3470}
3471
3472#endif /* CONFIG_NUMA */
3473
3474static __always_inline void *
3475__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3476{
3477    unsigned long save_flags;
3478    void *objp;
3479
3480    flags &= gfp_allowed_mask;
3481
3482    lockdep_trace_alloc(flags);
3483
3484    if (slab_should_failslab(cachep, flags))
3485        return NULL;
3486
3487    cache_alloc_debugcheck_before(cachep, flags);
3488    local_irq_save(save_flags);
3489    objp = __do_cache_alloc(cachep, flags);
3490    local_irq_restore(save_flags);
3491    objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3492    kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3493                 flags);
3494    prefetchw(objp);
3495
3496    if (likely(objp))
3497        kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3498
3499    if (unlikely((flags & __GFP_ZERO) && objp))
3500        memset(objp, 0, obj_size(cachep));
3501
3502    return objp;
3503}
3504
3505/*
3506 * Caller needs to acquire correct kmem_list's list_lock
3507 */
3508static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3509               int node)
3510{
3511    int i;
3512    struct kmem_list3 *l3;
3513
3514    for (i = 0; i < nr_objects; i++) {
3515        void *objp = objpp[i];
3516        struct slab *slabp;
3517
3518        slabp = virt_to_slab(objp);
3519        l3 = cachep->nodelists[node];
3520        list_del(&slabp->list);
3521        check_spinlock_acquired_node(cachep, node);
3522        check_slabp(cachep, slabp);
3523        slab_put_obj(cachep, slabp, objp, node);
3524        STATS_DEC_ACTIVE(cachep);
3525        l3->free_objects++;
3526        check_slabp(cachep, slabp);
3527
3528        /* fixup slab chains */
3529        if (slabp->inuse == 0) {
3530            if (l3->free_objects > l3->free_limit) {
3531                l3->free_objects -= cachep->num;
3532                /* No need to drop any previously held
3533                 * lock here, even if we have a off-slab slab
3534                 * descriptor it is guaranteed to come from
3535                 * a different cache, refer to comments before
3536                 * alloc_slabmgmt.
3537                 */
3538                slab_destroy(cachep, slabp);
3539            } else {
3540                list_add(&slabp->list, &l3->slabs_free);
3541            }
3542        } else {
3543            /* Unconditionally move a slab to the end of the
3544             * partial list on free - maximum time for the
3545             * other objects to be freed, too.
3546             */
3547            list_add_tail(&slabp->list, &l3->slabs_partial);
3548        }
3549    }
3550}
3551
3552static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3553{
3554    int batchcount;
3555    struct kmem_list3 *l3;
3556    int node = numa_mem_id();
3557
3558    batchcount = ac->batchcount;
3559#if DEBUG
3560    BUG_ON(!batchcount || batchcount > ac->avail);
3561#endif
3562    check_irq_off();
3563    l3 = cachep->nodelists[node];
3564    spin_lock(&l3->list_lock);
3565    if (l3->shared) {
3566        struct array_cache *shared_array = l3->shared;
3567        int max = shared_array->limit - shared_array->avail;
3568        if (max) {
3569            if (batchcount > max)
3570                batchcount = max;
3571            memcpy(&(shared_array->entry[shared_array->avail]),
3572                   ac->entry, sizeof(void *) * batchcount);
3573            shared_array->avail += batchcount;
3574            goto free_done;
3575        }
3576    }
3577
3578    free_block(cachep, ac->entry, batchcount, node);
3579free_done:
3580#if STATS
3581    {
3582        int i = 0;
3583        struct list_head *p;
3584
3585        p = l3->slabs_free.next;
3586        while (p != &(l3->slabs_free)) {
3587            struct slab *slabp;
3588
3589            slabp = list_entry(p, struct slab, list);
3590            BUG_ON(slabp->inuse);
3591
3592            i++;
3593            p = p->next;
3594        }
3595        STATS_SET_FREEABLE(cachep, i);
3596    }
3597#endif
3598    spin_unlock(&l3->list_lock);
3599    ac->avail -= batchcount;
3600    memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3601}
3602
3603/*
3604 * Release an obj back to its cache. If the obj has a constructed state, it must
3605 * be in this state _before_ it is released. Called with disabled ints.
3606 */
3607static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3608    void *caller)
3609{
3610    struct array_cache *ac = cpu_cache_get(cachep);
3611
3612    check_irq_off();
3613    kmemleak_free_recursive(objp, cachep->flags);
3614    objp = cache_free_debugcheck(cachep, objp, caller);
3615
3616    kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3617
3618    /*
3619     * Skip calling cache_free_alien() when the platform is not numa.
3620     * This will avoid cache misses that happen while accessing slabp (which
3621     * is per page memory reference) to get nodeid. Instead use a global
3622     * variable to skip the call, which is mostly likely to be present in
3623     * the cache.
3624     */
3625    if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3626        return;
3627
3628    if (likely(ac->avail < ac->limit)) {
3629        STATS_INC_FREEHIT(cachep);
3630        ac->entry[ac->avail++] = objp;
3631        return;
3632    } else {
3633        STATS_INC_FREEMISS(cachep);
3634        cache_flusharray(cachep, ac);
3635        ac->entry[ac->avail++] = objp;
3636    }
3637}
3638
3639/**
3640 * kmem_cache_alloc - Allocate an object
3641 * @cachep: The cache to allocate from.
3642 * @flags: See kmalloc().
3643 *
3644 * Allocate an object from this cache. The flags are only relevant
3645 * if the cache has no available objects.
3646 */
3647void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3648{
3649    void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3650
3651    trace_kmem_cache_alloc(_RET_IP_, ret,
3652                   obj_size(cachep), cachep->buffer_size, flags);
3653
3654    return ret;
3655}
3656EXPORT_SYMBOL(kmem_cache_alloc);
3657
3658#ifdef CONFIG_TRACING
3659void *
3660kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3661{
3662    void *ret;
3663
3664    ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3665
3666    trace_kmalloc(_RET_IP_, ret,
3667              size, slab_buffer_size(cachep), flags);
3668    return ret;
3669}
3670EXPORT_SYMBOL(kmem_cache_alloc_trace);
3671#endif
3672
3673#ifdef CONFIG_NUMA
3674void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3675{
3676    void *ret = __cache_alloc_node(cachep, flags, nodeid,
3677                       __builtin_return_address(0));
3678
3679    trace_kmem_cache_alloc_node(_RET_IP_, ret,
3680                    obj_size(cachep), cachep->buffer_size,
3681                    flags, nodeid);
3682
3683    return ret;
3684}
3685EXPORT_SYMBOL(kmem_cache_alloc_node);
3686
3687#ifdef CONFIG_TRACING
3688void *kmem_cache_alloc_node_trace(size_t size,
3689                  struct kmem_cache *cachep,
3690                  gfp_t flags,
3691                  int nodeid)
3692{
3693    void *ret;
3694
3695    ret = __cache_alloc_node(cachep, flags, nodeid,
3696                  __builtin_return_address(0));
3697    trace_kmalloc_node(_RET_IP_, ret,
3698               size, slab_buffer_size(cachep),
3699               flags, nodeid);
3700    return ret;
3701}
3702EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3703#endif
3704
3705static __always_inline void *
3706__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3707{
3708    struct kmem_cache *cachep;
3709
3710    cachep = kmem_find_general_cachep(size, flags);
3711    if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3712        return cachep;
3713    return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3714}
3715
3716#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3717void *__kmalloc_node(size_t size, gfp_t flags, int node)
3718{
3719    return __do_kmalloc_node(size, flags, node,
3720            __builtin_return_address(0));
3721}
3722EXPORT_SYMBOL(__kmalloc_node);
3723
3724void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3725        int node, unsigned long caller)
3726{
3727    return __do_kmalloc_node(size, flags, node, (void *)caller);
3728}
3729EXPORT_SYMBOL(__kmalloc_node_track_caller);
3730#else
3731void *__kmalloc_node(size_t size, gfp_t flags, int node)
3732{
3733    return __do_kmalloc_node(size, flags, node, NULL);
3734}
3735EXPORT_SYMBOL(__kmalloc_node);
3736#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3737#endif /* CONFIG_NUMA */
3738
3739/**
3740 * __do_kmalloc - allocate memory
3741 * @size: how many bytes of memory are required.
3742 * @flags: the type of memory to allocate (see kmalloc).
3743 * @caller: function caller for debug tracking of the caller
3744 */
3745static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3746                      void *caller)
3747{
3748    struct kmem_cache *cachep;
3749    void *ret;
3750
3751    /* If you want to save a few bytes .text space: replace
3752     * __ with kmem_.
3753     * Then kmalloc uses the uninlined functions instead of the inline
3754     * functions.
3755     */
3756    cachep = __find_general_cachep(size, flags);
3757    if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3758        return cachep;
3759    ret = __cache_alloc(cachep, flags, caller);
3760
3761    trace_kmalloc((unsigned long) caller, ret,
3762              size, cachep->buffer_size, flags);
3763
3764    return ret;
3765}
3766
3767
3768#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3769void *__kmalloc(size_t size, gfp_t flags)
3770{
3771    return __do_kmalloc(size, flags, __builtin_return_address(0));
3772}
3773EXPORT_SYMBOL(__kmalloc);
3774
3775void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3776{
3777    return __do_kmalloc(size, flags, (void *)caller);
3778}
3779EXPORT_SYMBOL(__kmalloc_track_caller);
3780
3781#else
3782void *__kmalloc(size_t size, gfp_t flags)
3783{
3784    return __do_kmalloc(size, flags, NULL);
3785}
3786EXPORT_SYMBOL(__kmalloc);
3787#endif
3788
3789/**
3790 * kmem_cache_free - Deallocate an object
3791 * @cachep: The cache the allocation was from.
3792 * @objp: The previously allocated object.
3793 *
3794 * Free an object which was previously allocated from this
3795 * cache.
3796 */
3797void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3798{
3799    unsigned long flags;
3800
3801    local_irq_save(flags);
3802    debug_check_no_locks_freed(objp, obj_size(cachep));
3803    if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3804        debug_check_no_obj_freed(objp, obj_size(cachep));
3805    __cache_free(cachep, objp, __builtin_return_address(0));
3806    local_irq_restore(flags);
3807
3808    trace_kmem_cache_free(_RET_IP_, objp);
3809}
3810EXPORT_SYMBOL(kmem_cache_free);
3811
3812/**
3813 * kfree - free previously allocated memory
3814 * @objp: pointer returned by kmalloc.
3815 *
3816 * If @objp is NULL, no operation is performed.
3817 *
3818 * Don't free memory not originally allocated by kmalloc()
3819 * or you will run into trouble.
3820 */
3821void kfree(const void *objp)
3822{
3823    struct kmem_cache *c;
3824    unsigned long flags;
3825
3826    trace_kfree(_RET_IP_, objp);
3827
3828    if (unlikely(ZERO_OR_NULL_PTR(objp)))
3829        return;
3830    local_irq_save(flags);
3831    kfree_debugcheck(objp);
3832    c = virt_to_cache(objp);
3833    debug_check_no_locks_freed(objp, obj_size(c));
3834    debug_check_no_obj_freed(objp, obj_size(c));
3835    __cache_free(c, (void *)objp, __builtin_return_address(0));
3836    local_irq_restore(flags);
3837}
3838EXPORT_SYMBOL(kfree);
3839
3840unsigned int kmem_cache_size(struct kmem_cache *cachep)
3841{
3842    return obj_size(cachep);
3843}
3844EXPORT_SYMBOL(kmem_cache_size);
3845
3846/*
3847 * This initializes kmem_list3 or resizes various caches for all nodes.
3848 */
3849static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3850{
3851    int node;
3852    struct kmem_list3 *l3;
3853    struct array_cache *new_shared;
3854    struct array_cache **new_alien = NULL;
3855
3856    for_each_online_node(node) {
3857
3858                if (use_alien_caches) {
3859                        new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3860                        if (!new_alien)
3861                                goto fail;
3862                }
3863
3864        new_shared = NULL;
3865        if (cachep->shared) {
3866            new_shared = alloc_arraycache(node,
3867                cachep->shared*cachep->batchcount,
3868                    0xbaadf00d, gfp);
3869            if (!new_shared) {
3870                free_alien_cache(new_alien);
3871                goto fail;
3872            }
3873        }
3874
3875        l3 = cachep->nodelists[node];
3876        if (l3) {
3877            struct array_cache *shared = l3->shared;
3878
3879            spin_lock_irq(&l3->list_lock);
3880
3881            if (shared)
3882                free_block(cachep, shared->entry,
3883                        shared->avail, node);
3884
3885            l3->shared = new_shared;
3886            if (!l3->alien) {
3887                l3->alien = new_alien;
3888                new_alien = NULL;
3889            }
3890            l3->free_limit = (1 + nr_cpus_node(node)) *
3891                    cachep->batchcount + cachep->num;
3892            spin_unlock_irq(&l3->list_lock);
3893            kfree(shared);
3894            free_alien_cache(new_alien);
3895            continue;
3896        }
3897        l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3898        if (!l3) {
3899            free_alien_cache(new_alien);
3900            kfree(new_shared);
3901            goto fail;
3902        }
3903
3904        kmem_list3_init(l3);
3905        l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3906                ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3907        l3->shared = new_shared;
3908        l3->alien = new_alien;
3909        l3->free_limit = (1 + nr_cpus_node(node)) *
3910                    cachep->batchcount + cachep->num;
3911        cachep->nodelists[node] = l3;
3912    }
3913    return 0;
3914
3915fail:
3916    if (!cachep->next.next) {
3917        /* Cache is not active yet. Roll back what we did */
3918        node--;
3919        while (node >= 0) {
3920            if (cachep->nodelists[node]) {
3921                l3 = cachep->nodelists[node];
3922
3923                kfree(l3->shared);
3924                free_alien_cache(l3->alien);
3925                kfree(l3);
3926                cachep->nodelists[node] = NULL;
3927            }
3928            node--;
3929        }
3930    }
3931    return -ENOMEM;
3932}
3933
3934struct ccupdate_struct {
3935    struct kmem_cache *cachep;
3936    struct array_cache *new[NR_CPUS];
3937};
3938
3939static void do_ccupdate_local(void *info)
3940{
3941    struct ccupdate_struct *new = info;
3942    struct array_cache *old;
3943
3944    check_irq_off();
3945    old = cpu_cache_get(new->cachep);
3946
3947    new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3948    new->new[smp_processor_id()] = old;
3949}
3950
3951/* Always called with the cache_chain_mutex held */
3952static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3953                int batchcount, int shared, gfp_t gfp)
3954{
3955    struct ccupdate_struct *new;
3956    int i;
3957
3958    new = kzalloc(sizeof(*new), gfp);
3959    if (!new)
3960        return -ENOMEM;
3961
3962    for_each_online_cpu(i) {
3963        new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3964                        batchcount, gfp);
3965        if (!new->new[i]) {
3966            for (i--; i >= 0; i--)
3967                kfree(new->new[i]);
3968            kfree(new);
3969            return -ENOMEM;
3970        }
3971    }
3972    new->cachep = cachep;
3973
3974    on_each_cpu(do_ccupdate_local, (void *)new, 1);
3975
3976    check_irq_on();
3977    cachep->batchcount = batchcount;
3978    cachep->limit = limit;
3979    cachep->shared = shared;
3980
3981    for_each_online_cpu(i) {
3982        struct array_cache *ccold = new->new[i];
3983        if (!ccold)
3984            continue;
3985        spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
3986        free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3987        spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
3988        kfree(ccold);
3989    }
3990    kfree(new);
3991    return alloc_kmemlist(cachep, gfp);
3992}
3993
3994/* Called with cache_chain_mutex held always */
3995static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3996{
3997    int err;
3998    int limit, shared;
3999
4000    /*
4001     * The head array serves three purposes:
4002     * - create a LIFO ordering, i.e. return objects that are cache-warm
4003     * - reduce the number of spinlock operations.
4004     * - reduce the number of linked list operations on the slab and
4005     * bufctl chains: array operations are cheaper.
4006     * The numbers are guessed, we should auto-tune as described by
4007     * Bonwick.
4008     */
4009    if (cachep->buffer_size > 131072)
4010        limit = 1;
4011    else if (cachep->buffer_size > PAGE_SIZE)
4012        limit = 8;
4013    else if (cachep->buffer_size > 1024)
4014        limit = 24;
4015    else if (cachep->buffer_size > 256)
4016        limit = 54;
4017    else
4018        limit = 120;
4019
4020    /*
4021     * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4022     * allocation behaviour: Most allocs on one cpu, most free operations
4023     * on another cpu. For these cases, an efficient object passing between
4024     * cpus is necessary. This is provided by a shared array. The array
4025     * replaces Bonwick's magazine layer.
4026     * On uniprocessor, it's functionally equivalent (but less efficient)
4027     * to a larger limit. Thus disabled by default.
4028     */
4029    shared = 0;
4030    if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4031        shared = 8;
4032
4033#if DEBUG
4034    /*
4035     * With debugging enabled, large batchcount lead to excessively long
4036     * periods with disabled local interrupts. Limit the batchcount
4037     */
4038    if (limit > 32)
4039        limit = 32;
4040#endif
4041    err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4042    if (err)
4043        printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4044               cachep->name, -err);
4045    return err;
4046}
4047
4048/*
4049 * Drain an array if it contains any elements taking the l3 lock only if
4050 * necessary. Note that the l3 listlock also protects the array_cache
4051 * if drain_array() is used on the shared array.
4052 */
4053static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4054             struct array_cache *ac, int force, int node)
4055{
4056    int tofree;
4057
4058    if (!ac || !ac->avail)
4059        return;
4060    if (ac->touched && !force) {
4061        ac->touched = 0;
4062    } else {
4063        spin_lock_irq(&l3->list_lock);
4064        if (ac->avail) {
4065            tofree = force ? ac->avail : (ac->limit + 4) / 5;
4066            if (tofree > ac->avail)
4067                tofree = (ac->avail + 1) / 2;
4068            free_block(cachep, ac->entry, tofree, node);
4069            ac->avail -= tofree;
4070            memmove(ac->entry, &(ac->entry[tofree]),
4071                sizeof(void *) * ac->avail);
4072        }
4073        spin_unlock_irq(&l3->list_lock);
4074    }
4075}
4076
4077/**
4078 * cache_reap - Reclaim memory from caches.
4079 * @w: work descriptor
4080 *
4081 * Called from workqueue/eventd every few seconds.
4082 * Purpose:
4083 * - clear the per-cpu caches for this CPU.
4084 * - return freeable pages to the main free memory pool.
4085 *
4086 * If we cannot acquire the cache chain mutex then just give up - we'll try
4087 * again on the next iteration.
4088 */
4089static void cache_reap(struct work_struct *w)
4090{
4091    struct kmem_cache *searchp;
4092    struct kmem_list3 *l3;
4093    int node = numa_mem_id();
4094    struct delayed_work *work = to_delayed_work(w);
4095
4096    if (!mutex_trylock(&cache_chain_mutex))
4097        /* Give up. Setup the next iteration. */
4098        goto out;
4099
4100    list_for_each_entry(searchp, &cache_chain, next) {
4101        check_irq_on();
4102
4103        /*
4104         * We only take the l3 lock if absolutely necessary and we
4105         * have established with reasonable certainty that
4106         * we can do some work if the lock was obtained.
4107         */
4108        l3 = searchp->nodelists[node];
4109
4110        reap_alien(searchp, l3);
4111
4112        drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4113
4114        /*
4115         * These are racy checks but it does not matter
4116         * if we skip one check or scan twice.
4117         */
4118        if (time_after(l3->next_reap, jiffies))
4119            goto next;
4120
4121        l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4122
4123        drain_array(searchp, l3, l3->shared, 0, node);
4124
4125        if (l3->free_touched)
4126            l3->free_touched = 0;
4127        else {
4128            int freed;
4129
4130            freed = drain_freelist(searchp, l3, (l3->free_limit +
4131                5 * searchp->num - 1) / (5 * searchp->num));
4132            STATS_ADD_REAPED(searchp, freed);
4133        }
4134next:
4135        cond_resched();
4136    }
4137    check_irq_on();
4138    mutex_unlock(&cache_chain_mutex);
4139    next_reap_node();
4140out:
4141    /* Set up the next iteration */
4142    schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4143}
4144
4145#ifdef CONFIG_SLABINFO
4146
4147static void print_slabinfo_header(struct seq_file *m)
4148{
4149    /*
4150     * Output format version, so at least we can change it
4151     * without _too_ many complaints.
4152     */
4153#if STATS
4154    seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4155#else
4156    seq_puts(m, "slabinfo - version: 2.1\n");
4157#endif
4158    seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4159         "<objperslab> <pagesperslab>");
4160    seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4161    seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4162#if STATS
4163    seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4164         "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4165    seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4166#endif
4167    seq_putc(m, '\n');
4168}
4169
4170static void *s_start(struct seq_file *m, loff_t *pos)
4171{
4172    loff_t n = *pos;
4173
4174    mutex_lock(&cache_chain_mutex);
4175    if (!n)
4176        print_slabinfo_header(m);
4177
4178    return seq_list_start(&cache_chain, *pos);
4179}
4180
4181static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4182{
4183    return seq_list_next(p, &cache_chain, pos);
4184}
4185
4186static void s_stop(struct seq_file *m, void *p)
4187{
4188    mutex_unlock(&cache_chain_mutex);
4189}
4190
4191static int s_show(struct seq_file *m, void *p)
4192{
4193    struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4194    struct slab *slabp;
4195    unsigned long active_objs;
4196    unsigned long num_objs;
4197    unsigned long active_slabs = 0;
4198    unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4199    const char *name;
4200    char *error = NULL;
4201    int node;
4202    struct kmem_list3 *l3;
4203
4204    active_objs = 0;
4205    num_slabs = 0;
4206    for_each_online_node(node) {
4207        l3 = cachep->nodelists[node];
4208        if (!l3)
4209            continue;
4210
4211        check_irq_on();
4212        spin_lock_irq(&l3->list_lock);
4213
4214        list_for_each_entry(slabp, &l3->slabs_full, list) {
4215            if (slabp->inuse != cachep->num && !error)
4216                error = "slabs_full accounting error";
4217            active_objs += cachep->num;
4218            active_slabs++;
4219        }
4220        list_for_each_entry(slabp, &l3->slabs_partial, list) {
4221            if (slabp->inuse == cachep->num && !error)
4222                error = "slabs_partial inuse accounting error";
4223            if (!slabp->inuse && !error)
4224                error = "slabs_partial/inuse accounting error";
4225            active_objs += slabp->inuse;
4226            active_slabs++;
4227        }
4228        list_for_each_entry(slabp, &l3->slabs_free, list) {
4229            if (slabp->inuse && !error)
4230                error = "slabs_free/inuse accounting error";
4231            num_slabs++;
4232        }
4233        free_objects += l3->free_objects;
4234        if (l3->shared)
4235            shared_avail += l3->shared->avail;
4236
4237        spin_unlock_irq(&l3->list_lock);
4238    }
4239    num_slabs += active_slabs;
4240    num_objs = num_slabs * cachep->num;
4241    if (num_objs - active_objs != free_objects && !error)
4242        error = "free_objects accounting error";
4243
4244    name = cachep->name;
4245    if (error)
4246        printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4247
4248    seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4249           name, active_objs, num_objs, cachep->buffer_size,
4250           cachep->num, (1 << cachep->gfporder));
4251    seq_printf(m, " : tunables %4u %4u %4u",
4252           cachep->limit, cachep->batchcount, cachep->shared);
4253    seq_printf(m, " : slabdata %6lu %6lu %6lu",
4254           active_slabs, num_slabs, shared_avail);
4255#if STATS
4256    { /* list3 stats */
4257        unsigned long high = cachep->high_mark;
4258        unsigned long allocs = cachep->num_allocations;
4259        unsigned long grown = cachep->grown;
4260        unsigned long reaped = cachep->reaped;
4261        unsigned long errors = cachep->errors;
4262        unsigned long max_freeable = cachep->max_freeable;
4263        unsigned long node_allocs = cachep->node_allocs;
4264        unsigned long node_frees = cachep->node_frees;
4265        unsigned long overflows = cachep->node_overflow;
4266
4267        seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4268               "%4lu %4lu %4lu %4lu %4lu",
4269               allocs, high, grown,
4270               reaped, errors, max_freeable, node_allocs,
4271               node_frees, overflows);
4272    }
4273    /* cpu stats */
4274    {
4275        unsigned long allochit = atomic_read(&cachep->allochit);
4276        unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4277        unsigned long freehit = atomic_read(&cachep->freehit);
4278        unsigned long freemiss = atomic_read(&cachep->freemiss);
4279
4280        seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4281               allochit, allocmiss, freehit, freemiss);
4282    }
4283#endif
4284    seq_putc(m, '\n');
4285    return 0;
4286}
4287
4288/*
4289 * slabinfo_op - iterator that generates /proc/slabinfo
4290 *
4291 * Output layout:
4292 * cache-name
4293 * num-active-objs
4294 * total-objs
4295 * object size
4296 * num-active-slabs
4297 * total-slabs
4298 * num-pages-per-slab
4299 * + further values on SMP and with statistics enabled
4300 */
4301
4302static const struct seq_operations slabinfo_op = {
4303    .start = s_start,
4304    .next = s_next,
4305    .stop = s_stop,
4306    .show = s_show,
4307};
4308
4309#define MAX_SLABINFO_WRITE 128
4310/**
4311 * slabinfo_write - Tuning for the slab allocator
4312 * @file: unused
4313 * @buffer: user buffer
4314 * @count: data length
4315 * @ppos: unused
4316 */
4317static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4318               size_t count, loff_t *ppos)
4319{
4320    char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4321    int limit, batchcount, shared, res;
4322    struct kmem_cache *cachep;
4323
4324    if (count > MAX_SLABINFO_WRITE)
4325        return -EINVAL;
4326    if (copy_from_user(&kbuf, buffer, count))
4327        return -EFAULT;
4328    kbuf[MAX_SLABINFO_WRITE] = '\0';
4329
4330    tmp = strchr(kbuf, ' ');
4331    if (!tmp)
4332        return -EINVAL;
4333    *tmp = '\0';
4334    tmp++;
4335    if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4336        return -EINVAL;
4337
4338    /* Find the cache in the chain of caches. */
4339    mutex_lock(&cache_chain_mutex);
4340    res = -EINVAL;
4341    list_for_each_entry(cachep, &cache_chain, next) {
4342        if (!strcmp(cachep->name, kbuf)) {
4343            if (limit < 1 || batchcount < 1 ||
4344                    batchcount > limit || shared < 0) {
4345                res = 0;
4346            } else {
4347                res = do_tune_cpucache(cachep, limit,
4348                               batchcount, shared,
4349                               GFP_KERNEL);
4350            }
4351            break;
4352        }
4353    }
4354    mutex_unlock(&cache_chain_mutex);
4355    if (res >= 0)
4356        res = count;
4357    return res;
4358}
4359
4360static int slabinfo_open(struct inode *inode, struct file *file)
4361{
4362    return seq_open(file, &slabinfo_op);
4363}
4364
4365static const struct file_operations proc_slabinfo_operations = {
4366    .open = slabinfo_open,
4367    .read = seq_read,
4368    .write = slabinfo_write,
4369    .llseek = seq_lseek,
4370    .release = seq_release,
4371};
4372
4373#ifdef CONFIG_DEBUG_SLAB_LEAK
4374
4375static void *leaks_start(struct seq_file *m, loff_t *pos)
4376{
4377    mutex_lock(&cache_chain_mutex);
4378    return seq_list_start(&cache_chain, *pos);
4379}
4380
4381static inline int add_caller(unsigned long *n, unsigned long v)
4382{
4383    unsigned long *p;
4384    int l;
4385    if (!v)
4386        return 1;
4387    l = n[1];
4388    p = n + 2;
4389    while (l) {
4390        int i = l/2;
4391        unsigned long *q = p + 2 * i;
4392        if (*q == v) {
4393            q[1]++;
4394            return 1;
4395        }
4396        if (*q > v) {
4397            l = i;
4398        } else {
4399            p = q + 2;
4400            l -= i + 1;
4401        }
4402    }
4403    if (++n[1] == n[0])
4404        return 0;
4405    memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4406    p[0] = v;
4407    p[1] = 1;
4408    return 1;
4409}
4410
4411static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4412{
4413    void *p;
4414    int i;
4415    if (n[0] == n[1])
4416        return;
4417    for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4418        if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4419            continue;
4420        if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4421            return;
4422    }
4423}
4424
4425static void show_symbol(struct seq_file *m, unsigned long address)
4426{
4427#ifdef CONFIG_KALLSYMS
4428    unsigned long offset, size;
4429    char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4430
4431    if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4432        seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4433        if (modname[0])
4434            seq_printf(m, " [%s]", modname);
4435        return;
4436    }
4437#endif
4438    seq_printf(m, "%p", (void *)address);
4439}
4440
4441static int leaks_show(struct seq_file *m, void *p)
4442{
4443    struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4444    struct slab *slabp;
4445    struct kmem_list3 *l3;
4446    const char *name;
4447    unsigned long *n = m->private;
4448    int node;
4449    int i;
4450
4451    if (!(cachep->flags & SLAB_STORE_USER))
4452        return 0;
4453    if (!(cachep->flags & SLAB_RED_ZONE))
4454        return 0;
4455
4456    /* OK, we can do it */
4457
4458    n[1] = 0;
4459
4460    for_each_online_node(node) {
4461        l3 = cachep->nodelists[node];
4462        if (!l3)
4463            continue;
4464
4465        check_irq_on();
4466        spin_lock_irq(&l3->list_lock);
4467
4468        list_for_each_entry(slabp, &l3->slabs_full, list)
4469            handle_slab(n, cachep, slabp);
4470        list_for_each_entry(slabp, &l3->slabs_partial, list)
4471            handle_slab(n, cachep, slabp);
4472        spin_unlock_irq(&l3->list_lock);
4473    }
4474    name = cachep->name;
4475    if (n[0] == n[1]) {
4476        /* Increase the buffer size */
4477        mutex_unlock(&cache_chain_mutex);
4478        m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4479        if (!m->private) {
4480            /* Too bad, we are really out */
4481            m->private = n;
4482            mutex_lock(&cache_chain_mutex);
4483            return -ENOMEM;
4484        }
4485        *(unsigned long *)m->private = n[0] * 2;
4486        kfree(n);
4487        mutex_lock(&cache_chain_mutex);
4488        /* Now make sure this entry will be retried */
4489        m->count = m->size;
4490        return 0;
4491    }
4492    for (i = 0; i < n[1]; i++) {
4493        seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4494        show_symbol(m, n[2*i+2]);
4495        seq_putc(m, '\n');
4496    }
4497
4498    return 0;
4499}
4500
4501static const struct seq_operations slabstats_op = {
4502    .start = leaks_start,
4503    .next = s_next,
4504    .stop = s_stop,
4505    .show = leaks_show,
4506};
4507
4508static int slabstats_open(struct inode *inode, struct file *file)
4509{
4510    unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4511    int ret = -ENOMEM;
4512    if (n) {
4513        ret = seq_open(file, &slabstats_op);
4514        if (!ret) {
4515            struct seq_file *m = file->private_data;
4516            *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4517            m->private = n;
4518            n = NULL;
4519        }
4520        kfree(n);
4521    }
4522    return ret;
4523}
4524
4525static const struct file_operations proc_slabstats_operations = {
4526    .open = slabstats_open,
4527    .read = seq_read,
4528    .llseek = seq_lseek,
4529    .release = seq_release_private,
4530};
4531#endif
4532
4533static int __init slab_proc_init(void)
4534{
4535    proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4536#ifdef CONFIG_DEBUG_SLAB_LEAK
4537    proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4538#endif
4539    return 0;
4540}
4541module_init(slab_proc_init);
4542#endif
4543
4544/**
4545 * ksize - get the actual amount of memory allocated for a given object
4546 * @objp: Pointer to the object
4547 *
4548 * kmalloc may internally round up allocations and return more memory
4549 * than requested. ksize() can be used to determine the actual amount of
4550 * memory allocated. The caller may use this additional memory, even though
4551 * a smaller amount of memory was initially specified with the kmalloc call.
4552 * The caller must guarantee that objp points to a valid object previously
4553 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4554 * must not be freed during the duration of the call.
4555 */
4556size_t ksize(const void *objp)
4557{
4558    BUG_ON(!objp);
4559    if (unlikely(objp == ZERO_SIZE_PTR))
4560        return 0;
4561
4562    return obj_size(virt_to_cache(objp));
4563}
4564EXPORT_SYMBOL(ksize);
4565

Archive Download this file



interactive