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
119#include <asm/cacheflush.h>
120#include <asm/tlbflush.h>
121#include <asm/page.h>
122
123/*
124 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
125 * 0 for faster, smaller code (especially in the critical paths).
126 *
127 * STATS - 1 to collect stats for /proc/slabinfo.
128 * 0 for faster, smaller code (especially in the critical paths).
129 *
130 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
131 */
132
133#ifdef CONFIG_DEBUG_SLAB
134#define DEBUG 1
135#define STATS 1
136#define FORCED_DEBUG 1
137#else
138#define DEBUG 0
139#define STATS 0
140#define FORCED_DEBUG 0
141#endif
142
143/* Shouldn't this be in a header file somewhere? */
144#define BYTES_PER_WORD sizeof(void *)
145#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
146
147#ifndef ARCH_KMALLOC_FLAGS
148#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
149#endif
150
151/* Legal flag mask for kmem_cache_create(). */
152#if DEBUG
153# define CREATE_MASK (SLAB_RED_ZONE | \
154             SLAB_POISON | SLAB_HWCACHE_ALIGN | \
155             SLAB_CACHE_DMA | \
156             SLAB_STORE_USER | \
157             SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
158             SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
159             SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
160#else
161# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
162             SLAB_CACHE_DMA | \
163             SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
164             SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
165             SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
166#endif
167
168/*
169 * kmem_bufctl_t:
170 *
171 * Bufctl's are used for linking objs within a slab
172 * linked offsets.
173 *
174 * This implementation relies on "struct page" for locating the cache &
175 * slab an object belongs to.
176 * This allows the bufctl structure to be small (one int), but limits
177 * the number of objects a slab (not a cache) can contain when off-slab
178 * bufctls are used. The limit is the size of the largest general cache
179 * that does not use off-slab slabs.
180 * For 32bit archs with 4 kB pages, is this 56.
181 * This is not serious, as it is only for large objects, when it is unwise
182 * to have too many per slab.
183 * Note: This limit can be raised by introducing a general cache whose size
184 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
185 */
186
187typedef unsigned int kmem_bufctl_t;
188#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
189#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
190#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
191#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
192
193/*
194 * struct slab
195 *
196 * Manages the objs in a slab. Placed either at the beginning of mem allocated
197 * for a slab, or allocated from an general cache.
198 * Slabs are chained into three list: fully used, partial, fully free slabs.
199 */
200struct slab {
201    struct list_head list;
202    unsigned long colouroff;
203    void *s_mem; /* including colour offset */
204    unsigned int inuse; /* num of objs active in slab */
205    kmem_bufctl_t free;
206    unsigned short nodeid;
207};
208
209/*
210 * struct slab_rcu
211 *
212 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
213 * arrange for kmem_freepages to be called via RCU. This is useful if
214 * we need to approach a kernel structure obliquely, from its address
215 * obtained without the usual locking. We can lock the structure to
216 * stabilize it and check it's still at the given address, only if we
217 * can be sure that the memory has not been meanwhile reused for some
218 * other kind of object (which our subsystem's lock might corrupt).
219 *
220 * rcu_read_lock before reading the address, then rcu_read_unlock after
221 * taking the spinlock within the structure expected at that address.
222 *
223 * We assume struct slab_rcu can overlay struct slab when destroying.
224 */
225struct slab_rcu {
226    struct rcu_head head;
227    struct kmem_cache *cachep;
228    void *addr;
229};
230
231/*
232 * struct array_cache
233 *
234 * Purpose:
235 * - LIFO ordering, to hand out cache-warm objects from _alloc
236 * - reduce the number of linked list operations
237 * - reduce spinlock operations
238 *
239 * The limit is stored in the per-cpu structure to reduce the data cache
240 * footprint.
241 *
242 */
243struct array_cache {
244    unsigned int avail;
245    unsigned int limit;
246    unsigned int batchcount;
247    unsigned int touched;
248    spinlock_t lock;
249    void *entry[]; /*
250             * Must have this definition in here for the proper
251             * alignment of array_cache. Also simplifies accessing
252             * the entries.
253             */
254};
255
256/*
257 * bootstrap: The caches do not work without cpuarrays anymore, but the
258 * cpuarrays are allocated from the generic caches...
259 */
260#define BOOT_CPUCACHE_ENTRIES 1
261struct arraycache_init {
262    struct array_cache cache;
263    void *entries[BOOT_CPUCACHE_ENTRIES];
264};
265
266/*
267 * The slab lists for all objects.
268 */
269struct kmem_list3 {
270    struct list_head slabs_partial; /* partial list first, better asm code */
271    struct list_head slabs_full;
272    struct list_head slabs_free;
273    unsigned long free_objects;
274    unsigned int free_limit;
275    unsigned int colour_next; /* Per-node cache coloring */
276    spinlock_t list_lock;
277    struct array_cache *shared; /* shared per node */
278    struct array_cache **alien; /* on other nodes */
279    unsigned long next_reap; /* updated without locking */
280    int free_touched; /* updated without locking */
281};
282
283/*
284 * Need this for bootstrapping a per node allocator.
285 */
286#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
287struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
288#define CACHE_CACHE 0
289#define SIZE_AC MAX_NUMNODES
290#define SIZE_L3 (2 * MAX_NUMNODES)
291
292static int drain_freelist(struct kmem_cache *cache,
293            struct kmem_list3 *l3, int tofree);
294static void free_block(struct kmem_cache *cachep, void **objpp, int len,
295            int node);
296static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
297static void cache_reap(struct work_struct *unused);
298
299/*
300 * This function must be completely optimized away if a constant is passed to
301 * it. Mostly the same as what is in linux/slab.h except it returns an index.
302 */
303static __always_inline int index_of(const size_t size)
304{
305    extern void __bad_size(void);
306
307    if (__builtin_constant_p(size)) {
308        int i = 0;
309
310#define CACHE(x) \
311    if (size <=x) \
312        return i; \
313    else \
314        i++;
315#include <linux/kmalloc_sizes.h>
316#undef CACHE
317        __bad_size();
318    } else
319        __bad_size();
320    return 0;
321}
322
323static int slab_early_init = 1;
324
325#define INDEX_AC index_of(sizeof(struct arraycache_init))
326#define INDEX_L3 index_of(sizeof(struct kmem_list3))
327
328static void kmem_list3_init(struct kmem_list3 *parent)
329{
330    INIT_LIST_HEAD(&parent->slabs_full);
331    INIT_LIST_HEAD(&parent->slabs_partial);
332    INIT_LIST_HEAD(&parent->slabs_free);
333    parent->shared = NULL;
334    parent->alien = NULL;
335    parent->colour_next = 0;
336    spin_lock_init(&parent->list_lock);
337    parent->free_objects = 0;
338    parent->free_touched = 0;
339}
340
341#define MAKE_LIST(cachep, listp, slab, nodeid) \
342    do { \
343        INIT_LIST_HEAD(listp); \
344        list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
345    } while (0)
346
347#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
348    do { \
349    MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
350    MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
351    MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
352    } while (0)
353
354#define CFLGS_OFF_SLAB (0x80000000UL)
355#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
356
357#define BATCHREFILL_LIMIT 16
358/*
359 * Optimization question: fewer reaps means less probability for unnessary
360 * cpucache drain/refill cycles.
361 *
362 * OTOH the cpuarrays can contain lots of objects,
363 * which could lock up otherwise freeable slabs.
364 */
365#define REAPTIMEOUT_CPUC (2*HZ)
366#define REAPTIMEOUT_LIST3 (4*HZ)
367
368#if STATS
369#define STATS_INC_ACTIVE(x) ((x)->num_active++)
370#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
371#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
372#define STATS_INC_GROWN(x) ((x)->grown++)
373#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
374#define STATS_SET_HIGH(x) \
375    do { \
376        if ((x)->num_active > (x)->high_mark) \
377            (x)->high_mark = (x)->num_active; \
378    } while (0)
379#define STATS_INC_ERR(x) ((x)->errors++)
380#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
381#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
382#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
383#define STATS_SET_FREEABLE(x, i) \
384    do { \
385        if ((x)->max_freeable < i) \
386            (x)->max_freeable = i; \
387    } while (0)
388#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
389#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
390#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
391#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
392#else
393#define STATS_INC_ACTIVE(x) do { } while (0)
394#define STATS_DEC_ACTIVE(x) do { } while (0)
395#define STATS_INC_ALLOCED(x) do { } while (0)
396#define STATS_INC_GROWN(x) do { } while (0)
397#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
398#define STATS_SET_HIGH(x) do { } while (0)
399#define STATS_INC_ERR(x) do { } while (0)
400#define STATS_INC_NODEALLOCS(x) do { } while (0)
401#define STATS_INC_NODEFREES(x) do { } while (0)
402#define STATS_INC_ACOVERFLOW(x) do { } while (0)
403#define STATS_SET_FREEABLE(x, i) do { } while (0)
404#define STATS_INC_ALLOCHIT(x) do { } while (0)
405#define STATS_INC_ALLOCMISS(x) do { } while (0)
406#define STATS_INC_FREEHIT(x) do { } while (0)
407#define STATS_INC_FREEMISS(x) do { } while (0)
408#endif
409
410#if DEBUG
411
412/*
413 * memory layout of objects:
414 * 0 : objp
415 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
416 * the end of an object is aligned with the end of the real
417 * allocation. Catches writes behind the end of the allocation.
418 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
419 * redzone word.
420 * cachep->obj_offset: The real object.
421 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
422 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
423 * [BYTES_PER_WORD long]
424 */
425static int obj_offset(struct kmem_cache *cachep)
426{
427    return cachep->obj_offset;
428}
429
430static int obj_size(struct kmem_cache *cachep)
431{
432    return cachep->obj_size;
433}
434
435static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
436{
437    BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
438    return (unsigned long long*) (objp + obj_offset(cachep) -
439                      sizeof(unsigned long long));
440}
441
442static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
443{
444    BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
445    if (cachep->flags & SLAB_STORE_USER)
446        return (unsigned long long *)(objp + cachep->buffer_size -
447                          sizeof(unsigned long long) -
448                          REDZONE_ALIGN);
449    return (unsigned long long *) (objp + cachep->buffer_size -
450                       sizeof(unsigned long long));
451}
452
453static void **dbg_userword(struct kmem_cache *cachep, void *objp)
454{
455    BUG_ON(!(cachep->flags & SLAB_STORE_USER));
456    return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
457}
458
459#else
460
461#define obj_offset(x) 0
462#define obj_size(cachep) (cachep->buffer_size)
463#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
464#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
465#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
466
467#endif
468
469#ifdef CONFIG_TRACING
470size_t slab_buffer_size(struct kmem_cache *cachep)
471{
472    return cachep->buffer_size;
473}
474EXPORT_SYMBOL(slab_buffer_size);
475#endif
476
477/*
478 * Do not go above this order unless 0 objects fit into the slab.
479 */
480#define BREAK_GFP_ORDER_HI 1
481#define BREAK_GFP_ORDER_LO 0
482static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
483
484/*
485 * Functions for storing/retrieving the cachep and or slab from the page
486 * allocator. These are used to find the slab an obj belongs to. With kfree(),
487 * these are used to find the cache which an obj belongs to.
488 */
489static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
490{
491    page->lru.next = (struct list_head *)cache;
492}
493
494static inline struct kmem_cache *page_get_cache(struct page *page)
495{
496    page = compound_head(page);
497    BUG_ON(!PageSlab(page));
498    return (struct kmem_cache *)page->lru.next;
499}
500
501static inline void page_set_slab(struct page *page, struct slab *slab)
502{
503    page->lru.prev = (struct list_head *)slab;
504}
505
506static inline struct slab *page_get_slab(struct page *page)
507{
508    BUG_ON(!PageSlab(page));
509    return (struct slab *)page->lru.prev;
510}
511
512static inline struct kmem_cache *virt_to_cache(const void *obj)
513{
514    struct page *page = virt_to_head_page(obj);
515    return page_get_cache(page);
516}
517
518static inline struct slab *virt_to_slab(const void *obj)
519{
520    struct page *page = virt_to_head_page(obj);
521    return page_get_slab(page);
522}
523
524static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
525                 unsigned int idx)
526{
527    return slab->s_mem + cache->buffer_size * idx;
528}
529
530/*
531 * We want to avoid an expensive divide : (offset / cache->buffer_size)
532 * Using the fact that buffer_size is a constant for a particular cache,
533 * we can replace (offset / cache->buffer_size) by
534 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
535 */
536static inline unsigned int obj_to_index(const struct kmem_cache *cache,
537                    const struct slab *slab, void *obj)
538{
539    u32 offset = (obj - slab->s_mem);
540    return reciprocal_divide(offset, cache->reciprocal_buffer_size);
541}
542
543/*
544 * These are the default caches for kmalloc. Custom caches can have other sizes.
545 */
546struct cache_sizes malloc_sizes[] = {
547#define CACHE(x) { .cs_size = (x) },
548#include <linux/kmalloc_sizes.h>
549    CACHE(ULONG_MAX)
550#undef CACHE
551};
552EXPORT_SYMBOL(malloc_sizes);
553
554/* Must match cache_sizes above. Out of line to keep cache footprint low. */
555struct cache_names {
556    char *name;
557    char *name_dma;
558};
559
560static struct cache_names __initdata cache_names[] = {
561#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
562#include <linux/kmalloc_sizes.h>
563    {NULL,}
564#undef CACHE
565};
566
567static struct arraycache_init initarray_cache __initdata =
568    { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
569static struct arraycache_init initarray_generic =
570    { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
571
572/* internal cache of cache description objs */
573static struct kmem_cache cache_cache = {
574    .batchcount = 1,
575    .limit = BOOT_CPUCACHE_ENTRIES,
576    .shared = 1,
577    .buffer_size = sizeof(struct kmem_cache),
578    .name = "kmem_cache",
579};
580
581#define BAD_ALIEN_MAGIC 0x01020304ul
582
583/*
584 * chicken and egg problem: delay the per-cpu array allocation
585 * until the general caches are up.
586 */
587static enum {
588    NONE,
589    PARTIAL_AC,
590    PARTIAL_L3,
591    EARLY,
592    FULL
593} g_cpucache_up;
594
595/*
596 * used by boot code to determine if it can use slab based allocator
597 */
598int slab_is_available(void)
599{
600    return g_cpucache_up >= EARLY;
601}
602
603#ifdef CONFIG_LOCKDEP
604
605/*
606 * Slab sometimes uses the kmalloc slabs to store the slab headers
607 * for other slabs "off slab".
608 * The locking for this is tricky in that it nests within the locks
609 * of all other slabs in a few places; to deal with this special
610 * locking we put on-slab caches into a separate lock-class.
611 *
612 * We set lock class for alien array caches which are up during init.
613 * The lock annotation will be lost if all cpus of a node goes down and
614 * then comes back up during hotplug
615 */
616static struct lock_class_key on_slab_l3_key;
617static struct lock_class_key on_slab_alc_key;
618
619static void init_node_lock_keys(int q)
620{
621    struct cache_sizes *s = malloc_sizes;
622
623    if (g_cpucache_up != FULL)
624        return;
625
626    for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
627        struct array_cache **alc;
628        struct kmem_list3 *l3;
629        int r;
630
631        l3 = s->cs_cachep->nodelists[q];
632        if (!l3 || OFF_SLAB(s->cs_cachep))
633            continue;
634        lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
635        alc = l3->alien;
636        /*
637         * FIXME: This check for BAD_ALIEN_MAGIC
638         * should go away when common slab code is taught to
639         * work even without alien caches.
640         * Currently, non NUMA code returns BAD_ALIEN_MAGIC
641         * for alloc_alien_cache,
642         */
643        if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
644            continue;
645        for_each_node(r) {
646            if (alc[r])
647                lockdep_set_class(&alc[r]->lock,
648                    &on_slab_alc_key);
649        }
650    }
651}
652
653static inline void init_lock_keys(void)
654{
655    int node;
656
657    for_each_node(node)
658        init_node_lock_keys(node);
659}
660#else
661static void init_node_lock_keys(int q)
662{
663}
664
665static inline void init_lock_keys(void)
666{
667}
668#endif
669
670/*
671 * Guard access to the cache-chain.
672 */
673static DEFINE_MUTEX(cache_chain_mutex);
674static struct list_head cache_chain;
675
676static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
677
678static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
679{
680    return cachep->array[smp_processor_id()];
681}
682
683static inline struct kmem_cache *__find_general_cachep(size_t size,
684                            gfp_t gfpflags)
685{
686    struct cache_sizes *csizep = malloc_sizes;
687
688#if DEBUG
689    /* This happens if someone tries to call
690     * kmem_cache_create(), or __kmalloc(), before
691     * the generic caches are initialized.
692     */
693    BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
694#endif
695    if (!size)
696        return ZERO_SIZE_PTR;
697
698    while (size > csizep->cs_size)
699        csizep++;
700
701    /*
702     * Really subtle: The last entry with cs->cs_size==ULONG_MAX
703     * has cs_{dma,}cachep==NULL. Thus no special case
704     * for large kmalloc calls required.
705     */
706#ifdef CONFIG_ZONE_DMA
707    if (unlikely(gfpflags & GFP_DMA))
708        return csizep->cs_dmacachep;
709#endif
710    return csizep->cs_cachep;
711}
712
713static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
714{
715    return __find_general_cachep(size, gfpflags);
716}
717
718static size_t slab_mgmt_size(size_t nr_objs, size_t align)
719{
720    return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
721}
722
723/*
724 * Calculate the number of objects and left-over bytes for a given buffer size.
725 */
726static void cache_estimate(unsigned long gfporder, size_t buffer_size,
727               size_t align, int flags, size_t *left_over,
728               unsigned int *num)
729{
730    int nr_objs;
731    size_t mgmt_size;
732    size_t slab_size = PAGE_SIZE << gfporder;
733
734    /*
735     * The slab management structure can be either off the slab or
736     * on it. For the latter case, the memory allocated for a
737     * slab is used for:
738     *
739     * - The struct slab
740     * - One kmem_bufctl_t for each object
741     * - Padding to respect alignment of @align
742     * - @buffer_size bytes for each object
743     *
744     * If the slab management structure is off the slab, then the
745     * alignment will already be calculated into the size. Because
746     * the slabs are all pages aligned, the objects will be at the
747     * correct alignment when allocated.
748     */
749    if (flags & CFLGS_OFF_SLAB) {
750        mgmt_size = 0;
751        nr_objs = slab_size / buffer_size;
752
753        if (nr_objs > SLAB_LIMIT)
754            nr_objs = SLAB_LIMIT;
755    } else {
756        /*
757         * Ignore padding for the initial guess. The padding
758         * is at most @align-1 bytes, and @buffer_size is at
759         * least @align. In the worst case, this result will
760         * be one greater than the number of objects that fit
761         * into the memory allocation when taking the padding
762         * into account.
763         */
764        nr_objs = (slab_size - sizeof(struct slab)) /
765              (buffer_size + sizeof(kmem_bufctl_t));
766
767        /*
768         * This calculated number will be either the right
769         * amount, or one greater than what we want.
770         */
771        if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
772               > slab_size)
773            nr_objs--;
774
775        if (nr_objs > SLAB_LIMIT)
776            nr_objs = SLAB_LIMIT;
777
778        mgmt_size = slab_mgmt_size(nr_objs, align);
779    }
780    *num = nr_objs;
781    *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
782}
783
784#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
785
786static void __slab_error(const char *function, struct kmem_cache *cachep,
787            char *msg)
788{
789    printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
790           function, cachep->name, msg);
791    dump_stack();
792}
793
794/*
795 * By default on NUMA we use alien caches to stage the freeing of
796 * objects allocated from other nodes. This causes massive memory
797 * inefficiencies when using fake NUMA setup to split memory into a
798 * large number of small nodes, so it can be disabled on the command
799 * line
800  */
801
802static int use_alien_caches __read_mostly = 1;
803static int __init noaliencache_setup(char *s)
804{
805    use_alien_caches = 0;
806    return 1;
807}
808__setup("noaliencache", noaliencache_setup);
809
810#ifdef CONFIG_NUMA
811/*
812 * Special reaping functions for NUMA systems called from cache_reap().
813 * These take care of doing round robin flushing of alien caches (containing
814 * objects freed on different nodes from which they were allocated) and the
815 * flushing of remote pcps by calling drain_node_pages.
816 */
817static DEFINE_PER_CPU(unsigned long, slab_reap_node);
818
819static void init_reap_node(int cpu)
820{
821    int node;
822
823    node = next_node(cpu_to_mem(cpu), node_online_map);
824    if (node == MAX_NUMNODES)
825        node = first_node(node_online_map);
826
827    per_cpu(slab_reap_node, cpu) = node;
828}
829
830static void next_reap_node(void)
831{
832    int node = __get_cpu_var(slab_reap_node);
833
834    node = next_node(node, node_online_map);
835    if (unlikely(node >= MAX_NUMNODES))
836        node = first_node(node_online_map);
837    __get_cpu_var(slab_reap_node) = node;
838}
839
840#else
841#define init_reap_node(cpu) do { } while (0)
842#define next_reap_node(void) do { } while (0)
843#endif
844
845/*
846 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
847 * via the workqueue/eventd.
848 * Add the CPU number into the expiration time to minimize the possibility of
849 * the CPUs getting into lockstep and contending for the global cache chain
850 * lock.
851 */
852static void __cpuinit start_cpu_timer(int cpu)
853{
854    struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
855
856    /*
857     * When this gets called from do_initcalls via cpucache_init(),
858     * init_workqueues() has already run, so keventd will be setup
859     * at that time.
860     */
861    if (keventd_up() && reap_work->work.func == NULL) {
862        init_reap_node(cpu);
863        INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
864        schedule_delayed_work_on(cpu, reap_work,
865                    __round_jiffies_relative(HZ, cpu));
866    }
867}
868
869static struct array_cache *alloc_arraycache(int node, int entries,
870                        int batchcount, gfp_t gfp)
871{
872    int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
873    struct array_cache *nc = NULL;
874
875    nc = kmalloc_node(memsize, gfp, node);
876    /*
877     * The array_cache structures contain pointers to free object.
878     * However, when such objects are allocated or transfered to another
879     * cache the pointers are not cleared and they could be counted as
880     * valid references during a kmemleak scan. Therefore, kmemleak must
881     * not scan such objects.
882     */
883    kmemleak_no_scan(nc);
884    if (nc) {
885        nc->avail = 0;
886        nc->limit = entries;
887        nc->batchcount = batchcount;
888        nc->touched = 0;
889        spin_lock_init(&nc->lock);
890    }
891    return nc;
892}
893
894/*
895 * Transfer objects in one arraycache to another.
896 * Locking must be handled by the caller.
897 *
898 * Return the number of entries transferred.
899 */
900static int transfer_objects(struct array_cache *to,
901        struct array_cache *from, unsigned int max)
902{
903    /* Figure out how many entries to transfer */
904    int nr = min(min(from->avail, max), to->limit - to->avail);
905
906    if (!nr)
907        return 0;
908
909    memcpy(to->entry + to->avail, from->entry + from->avail -nr,
910            sizeof(void *) *nr);
911
912    from->avail -= nr;
913    to->avail += nr;
914    return nr;
915}
916
917#ifndef CONFIG_NUMA
918
919#define drain_alien_cache(cachep, alien) do { } while (0)
920#define reap_alien(cachep, l3) do { } while (0)
921
922static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
923{
924    return (struct array_cache **)BAD_ALIEN_MAGIC;
925}
926
927static inline void free_alien_cache(struct array_cache **ac_ptr)
928{
929}
930
931static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
932{
933    return 0;
934}
935
936static inline void *alternate_node_alloc(struct kmem_cache *cachep,
937        gfp_t flags)
938{
939    return NULL;
940}
941
942static inline void *____cache_alloc_node(struct kmem_cache *cachep,
943         gfp_t flags, int nodeid)
944{
945    return NULL;
946}
947
948#else /* CONFIG_NUMA */
949
950static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
951static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
952
953static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
954{
955    struct array_cache **ac_ptr;
956    int memsize = sizeof(void *) * nr_node_ids;
957    int i;
958
959    if (limit > 1)
960        limit = 12;
961    ac_ptr = kzalloc_node(memsize, gfp, node);
962    if (ac_ptr) {
963        for_each_node(i) {
964            if (i == node || !node_online(i))
965                continue;
966            ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
967            if (!ac_ptr[i]) {
968                for (i--; i >= 0; i--)
969                    kfree(ac_ptr[i]);
970                kfree(ac_ptr);
971                return NULL;
972            }
973        }
974    }
975    return ac_ptr;
976}
977
978static void free_alien_cache(struct array_cache **ac_ptr)
979{
980    int i;
981
982    if (!ac_ptr)
983        return;
984    for_each_node(i)
985        kfree(ac_ptr[i]);
986    kfree(ac_ptr);
987}
988
989static void __drain_alien_cache(struct kmem_cache *cachep,
990                struct array_cache *ac, int node)
991{
992    struct kmem_list3 *rl3 = cachep->nodelists[node];
993
994    if (ac->avail) {
995        spin_lock(&rl3->list_lock);
996        /*
997         * Stuff objects into the remote nodes shared array first.
998         * That way we could avoid the overhead of putting the objects
999         * into the free lists and getting them back later.
1000         */
1001        if (rl3->shared)
1002            transfer_objects(rl3->shared, ac, ac->limit);
1003
1004        free_block(cachep, ac->entry, ac->avail, node);
1005        ac->avail = 0;
1006        spin_unlock(&rl3->list_lock);
1007    }
1008}
1009
1010/*
1011 * Called from cache_reap() to regularly drain alien caches round robin.
1012 */
1013static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1014{
1015    int node = __get_cpu_var(slab_reap_node);
1016
1017    if (l3->alien) {
1018        struct array_cache *ac = l3->alien[node];
1019
1020        if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1021            __drain_alien_cache(cachep, ac, node);
1022            spin_unlock_irq(&ac->lock);
1023        }
1024    }
1025}
1026
1027static void drain_alien_cache(struct kmem_cache *cachep,
1028                struct array_cache **alien)
1029{
1030    int i = 0;
1031    struct array_cache *ac;
1032    unsigned long flags;
1033
1034    for_each_online_node(i) {
1035        ac = alien[i];
1036        if (ac) {
1037            spin_lock_irqsave(&ac->lock, flags);
1038            __drain_alien_cache(cachep, ac, i);
1039            spin_unlock_irqrestore(&ac->lock, flags);
1040        }
1041    }
1042}
1043
1044static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1045{
1046    struct slab *slabp = virt_to_slab(objp);
1047    int nodeid = slabp->nodeid;
1048    struct kmem_list3 *l3;
1049    struct array_cache *alien = NULL;
1050    int node;
1051
1052    node = numa_mem_id();
1053
1054    /*
1055     * Make sure we are not freeing a object from another node to the array
1056     * cache on this cpu.
1057     */
1058    if (likely(slabp->nodeid == node))
1059        return 0;
1060
1061    l3 = cachep->nodelists[node];
1062    STATS_INC_NODEFREES(cachep);
1063    if (l3->alien && l3->alien[nodeid]) {
1064        alien = l3->alien[nodeid];
1065        spin_lock(&alien->lock);
1066        if (unlikely(alien->avail == alien->limit)) {
1067            STATS_INC_ACOVERFLOW(cachep);
1068            __drain_alien_cache(cachep, alien, nodeid);
1069        }
1070        alien->entry[alien->avail++] = objp;
1071        spin_unlock(&alien->lock);
1072    } else {
1073        spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1074        free_block(cachep, &objp, 1, nodeid);
1075        spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1076    }
1077    return 1;
1078}
1079#endif
1080
1081/*
1082 * Allocates and initializes nodelists for a node on each slab cache, used for
1083 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1084 * will be allocated off-node since memory is not yet online for the new node.
1085 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1086 * already in use.
1087 *
1088 * Must hold cache_chain_mutex.
1089 */
1090static int init_cache_nodelists_node(int node)
1091{
1092    struct kmem_cache *cachep;
1093    struct kmem_list3 *l3;
1094    const int memsize = sizeof(struct kmem_list3);
1095
1096    list_for_each_entry(cachep, &cache_chain, next) {
1097        /*
1098         * Set up the size64 kmemlist for cpu before we can
1099         * begin anything. Make sure some other cpu on this
1100         * node has not already allocated this
1101         */
1102        if (!cachep->nodelists[node]) {
1103            l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1104            if (!l3)
1105                return -ENOMEM;
1106            kmem_list3_init(l3);
1107            l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1108                ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1109
1110            /*
1111             * The l3s don't come and go as CPUs come and
1112             * go. cache_chain_mutex is sufficient
1113             * protection here.
1114             */
1115            cachep->nodelists[node] = l3;
1116        }
1117
1118        spin_lock_irq(&cachep->nodelists[node]->list_lock);
1119        cachep->nodelists[node]->free_limit =
1120            (1 + nr_cpus_node(node)) *
1121            cachep->batchcount + cachep->num;
1122        spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1123    }
1124    return 0;
1125}
1126
1127static void __cpuinit cpuup_canceled(long cpu)
1128{
1129    struct kmem_cache *cachep;
1130    struct kmem_list3 *l3 = NULL;
1131    int node = cpu_to_mem(cpu);
1132    const struct cpumask *mask = cpumask_of_node(node);
1133
1134    list_for_each_entry(cachep, &cache_chain, next) {
1135        struct array_cache *nc;
1136        struct array_cache *shared;
1137        struct array_cache **alien;
1138
1139        /* cpu is dead; no one can alloc from it. */
1140        nc = cachep->array[cpu];
1141        cachep->array[cpu] = NULL;
1142        l3 = cachep->nodelists[node];
1143
1144        if (!l3)
1145            goto free_array_cache;
1146
1147        spin_lock_irq(&l3->list_lock);
1148
1149        /* Free limit for this kmem_list3 */
1150        l3->free_limit -= cachep->batchcount;
1151        if (nc)
1152            free_block(cachep, nc->entry, nc->avail, node);
1153
1154        if (!cpumask_empty(mask)) {
1155            spin_unlock_irq(&l3->list_lock);
1156            goto free_array_cache;
1157        }
1158
1159        shared = l3->shared;
1160        if (shared) {
1161            free_block(cachep, shared->entry,
1162                   shared->avail, node);
1163            l3->shared = NULL;
1164        }
1165
1166        alien = l3->alien;
1167        l3->alien = NULL;
1168
1169        spin_unlock_irq(&l3->list_lock);
1170
1171        kfree(shared);
1172        if (alien) {
1173            drain_alien_cache(cachep, alien);
1174            free_alien_cache(alien);
1175        }
1176free_array_cache:
1177        kfree(nc);
1178    }
1179    /*
1180     * In the previous loop, all the objects were freed to
1181     * the respective cache's slabs, now we can go ahead and
1182     * shrink each nodelist to its limit.
1183     */
1184    list_for_each_entry(cachep, &cache_chain, next) {
1185        l3 = cachep->nodelists[node];
1186        if (!l3)
1187            continue;
1188        drain_freelist(cachep, l3, l3->free_objects);
1189    }
1190}
1191
1192static int __cpuinit cpuup_prepare(long cpu)
1193{
1194    struct kmem_cache *cachep;
1195    struct kmem_list3 *l3 = NULL;
1196    int node = cpu_to_mem(cpu);
1197    int err;
1198
1199    /*
1200     * We need to do this right in the beginning since
1201     * alloc_arraycache's are going to use this list.
1202     * kmalloc_node allows us to add the slab to the right
1203     * kmem_list3 and not this cpu's kmem_list3
1204     */
1205    err = init_cache_nodelists_node(node);
1206    if (err < 0)
1207        goto bad;
1208
1209    /*
1210     * Now we can go ahead with allocating the shared arrays and
1211     * array caches
1212     */
1213    list_for_each_entry(cachep, &cache_chain, next) {
1214        struct array_cache *nc;
1215        struct array_cache *shared = NULL;
1216        struct array_cache **alien = NULL;
1217
1218        nc = alloc_arraycache(node, cachep->limit,
1219                    cachep->batchcount, GFP_KERNEL);
1220        if (!nc)
1221            goto bad;
1222        if (cachep->shared) {
1223            shared = alloc_arraycache(node,
1224                cachep->shared * cachep->batchcount,
1225                0xbaadf00d, GFP_KERNEL);
1226            if (!shared) {
1227                kfree(nc);
1228                goto bad;
1229            }
1230        }
1231        if (use_alien_caches) {
1232            alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1233            if (!alien) {
1234                kfree(shared);
1235                kfree(nc);
1236                goto bad;
1237            }
1238        }
1239        cachep->array[cpu] = nc;
1240        l3 = cachep->nodelists[node];
1241        BUG_ON(!l3);
1242
1243        spin_lock_irq(&l3->list_lock);
1244        if (!l3->shared) {
1245            /*
1246             * We are serialised from CPU_DEAD or
1247             * CPU_UP_CANCELLED by the cpucontrol lock
1248             */
1249            l3->shared = shared;
1250            shared = NULL;
1251        }
1252#ifdef CONFIG_NUMA
1253        if (!l3->alien) {
1254            l3->alien = alien;
1255            alien = NULL;
1256        }
1257#endif
1258        spin_unlock_irq(&l3->list_lock);
1259        kfree(shared);
1260        free_alien_cache(alien);
1261    }
1262    init_node_lock_keys(node);
1263
1264    return 0;
1265bad:
1266    cpuup_canceled(cpu);
1267    return -ENOMEM;
1268}
1269
1270static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1271                    unsigned long action, void *hcpu)
1272{
1273    long cpu = (long)hcpu;
1274    int err = 0;
1275
1276    switch (action) {
1277    case CPU_UP_PREPARE:
1278    case CPU_UP_PREPARE_FROZEN:
1279        mutex_lock(&cache_chain_mutex);
1280        err = cpuup_prepare(cpu);
1281        mutex_unlock(&cache_chain_mutex);
1282        break;
1283    case CPU_ONLINE:
1284    case CPU_ONLINE_FROZEN:
1285        start_cpu_timer(cpu);
1286        break;
1287#ifdef CONFIG_HOTPLUG_CPU
1288      case CPU_DOWN_PREPARE:
1289      case CPU_DOWN_PREPARE_FROZEN:
1290        /*
1291         * Shutdown cache reaper. Note that the cache_chain_mutex is
1292         * held so that if cache_reap() is invoked it cannot do
1293         * anything expensive but will only modify reap_work
1294         * and reschedule the timer.
1295        */
1296        cancel_rearming_delayed_work(&per_cpu(slab_reap_work, cpu));
1297        /* Now the cache_reaper is guaranteed to be not running. */
1298        per_cpu(slab_reap_work, cpu).work.func = NULL;
1299          break;
1300      case CPU_DOWN_FAILED:
1301      case CPU_DOWN_FAILED_FROZEN:
1302        start_cpu_timer(cpu);
1303          break;
1304    case CPU_DEAD:
1305    case CPU_DEAD_FROZEN:
1306        /*
1307         * Even if all the cpus of a node are down, we don't free the
1308         * kmem_list3 of any cache. This to avoid a race between
1309         * cpu_down, and a kmalloc allocation from another cpu for
1310         * memory from the node of the cpu going down. The list3
1311         * structure is usually allocated from kmem_cache_create() and
1312         * gets destroyed at kmem_cache_destroy().
1313         */
1314        /* fall through */
1315#endif
1316    case CPU_UP_CANCELED:
1317    case CPU_UP_CANCELED_FROZEN:
1318        mutex_lock(&cache_chain_mutex);
1319        cpuup_canceled(cpu);
1320        mutex_unlock(&cache_chain_mutex);
1321        break;
1322    }
1323    return notifier_from_errno(err);
1324}
1325
1326static struct notifier_block __cpuinitdata cpucache_notifier = {
1327    &cpuup_callback, NULL, 0
1328};
1329
1330#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1331/*
1332 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1333 * Returns -EBUSY if all objects cannot be drained so that the node is not
1334 * removed.
1335 *
1336 * Must hold cache_chain_mutex.
1337 */
1338static int __meminit drain_cache_nodelists_node(int node)
1339{
1340    struct kmem_cache *cachep;
1341    int ret = 0;
1342
1343    list_for_each_entry(cachep, &cache_chain, next) {
1344        struct kmem_list3 *l3;
1345
1346        l3 = cachep->nodelists[node];
1347        if (!l3)
1348            continue;
1349
1350        drain_freelist(cachep, l3, l3->free_objects);
1351
1352        if (!list_empty(&l3->slabs_full) ||
1353            !list_empty(&l3->slabs_partial)) {
1354            ret = -EBUSY;
1355            break;
1356        }
1357    }
1358    return ret;
1359}
1360
1361static int __meminit slab_memory_callback(struct notifier_block *self,
1362                    unsigned long action, void *arg)
1363{
1364    struct memory_notify *mnb = arg;
1365    int ret = 0;
1366    int nid;
1367
1368    nid = mnb->status_change_nid;
1369    if (nid < 0)
1370        goto out;
1371
1372    switch (action) {
1373    case MEM_GOING_ONLINE:
1374        mutex_lock(&cache_chain_mutex);
1375        ret = init_cache_nodelists_node(nid);
1376        mutex_unlock(&cache_chain_mutex);
1377        break;
1378    case MEM_GOING_OFFLINE:
1379        mutex_lock(&cache_chain_mutex);
1380        ret = drain_cache_nodelists_node(nid);
1381        mutex_unlock(&cache_chain_mutex);
1382        break;
1383    case MEM_ONLINE:
1384    case MEM_OFFLINE:
1385    case MEM_CANCEL_ONLINE:
1386    case MEM_CANCEL_OFFLINE:
1387        break;
1388    }
1389out:
1390    return ret ? notifier_from_errno(ret) : NOTIFY_OK;
1391}
1392#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1393
1394/*
1395 * swap the static kmem_list3 with kmalloced memory
1396 */
1397static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1398                int nodeid)
1399{
1400    struct kmem_list3 *ptr;
1401
1402    ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1403    BUG_ON(!ptr);
1404
1405    memcpy(ptr, list, sizeof(struct kmem_list3));
1406    /*
1407     * Do not assume that spinlocks can be initialized via memcpy:
1408     */
1409    spin_lock_init(&ptr->list_lock);
1410
1411    MAKE_ALL_LISTS(cachep, ptr, nodeid);
1412    cachep->nodelists[nodeid] = ptr;
1413}
1414
1415/*
1416 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1417 * size of kmem_list3.
1418 */
1419static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1420{
1421    int node;
1422
1423    for_each_online_node(node) {
1424        cachep->nodelists[node] = &initkmem_list3[index + node];
1425        cachep->nodelists[node]->next_reap = jiffies +
1426            REAPTIMEOUT_LIST3 +
1427            ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1428    }
1429}
1430
1431/*
1432 * Initialisation. Called after the page allocator have been initialised and
1433 * before smp_init().
1434 */
1435void __init kmem_cache_init(void)
1436{
1437    size_t left_over;
1438    struct cache_sizes *sizes;
1439    struct cache_names *names;
1440    int i;
1441    int order;
1442    int node;
1443
1444    if (num_possible_nodes() == 1)
1445        use_alien_caches = 0;
1446
1447    for (i = 0; i < NUM_INIT_LISTS; i++) {
1448        kmem_list3_init(&initkmem_list3[i]);
1449        if (i < MAX_NUMNODES)
1450            cache_cache.nodelists[i] = NULL;
1451    }
1452    set_up_list3s(&cache_cache, CACHE_CACHE);
1453
1454    /*
1455     * Fragmentation resistance on low memory - only use bigger
1456     * page orders on machines with more than 32MB of memory.
1457     */
1458    if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1459        slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1460
1461    /* Bootstrap is tricky, because several objects are allocated
1462     * from caches that do not exist yet:
1463     * 1) initialize the cache_cache cache: it contains the struct
1464     * kmem_cache structures of all caches, except cache_cache itself:
1465     * cache_cache is statically allocated.
1466     * Initially an __init data area is used for the head array and the
1467     * kmem_list3 structures, it's replaced with a kmalloc allocated
1468     * array at the end of the bootstrap.
1469     * 2) Create the first kmalloc cache.
1470     * The struct kmem_cache for the new cache is allocated normally.
1471     * An __init data area is used for the head array.
1472     * 3) Create the remaining kmalloc caches, with minimally sized
1473     * head arrays.
1474     * 4) Replace the __init data head arrays for cache_cache and the first
1475     * kmalloc cache with kmalloc allocated arrays.
1476     * 5) Replace the __init data for kmem_list3 for cache_cache and
1477     * the other cache's with kmalloc allocated memory.
1478     * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1479     */
1480
1481    node = numa_mem_id();
1482
1483    /* 1) create the cache_cache */
1484    INIT_LIST_HEAD(&cache_chain);
1485    list_add(&cache_cache.next, &cache_chain);
1486    cache_cache.colour_off = cache_line_size();
1487    cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1488    cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1489
1490    /*
1491     * struct kmem_cache size depends on nr_node_ids, which
1492     * can be less than MAX_NUMNODES.
1493     */
1494    cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1495                 nr_node_ids * sizeof(struct kmem_list3 *);
1496#if DEBUG
1497    cache_cache.obj_size = cache_cache.buffer_size;
1498#endif
1499    cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1500                    cache_line_size());
1501    cache_cache.reciprocal_buffer_size =
1502        reciprocal_value(cache_cache.buffer_size);
1503
1504    for (order = 0; order < MAX_ORDER; order++) {
1505        cache_estimate(order, cache_cache.buffer_size,
1506            cache_line_size(), 0, &left_over, &cache_cache.num);
1507        if (cache_cache.num)
1508            break;
1509    }
1510    BUG_ON(!cache_cache.num);
1511    cache_cache.gfporder = order;
1512    cache_cache.colour = left_over / cache_cache.colour_off;
1513    cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1514                      sizeof(struct slab), cache_line_size());
1515
1516    /* 2+3) create the kmalloc caches */
1517    sizes = malloc_sizes;
1518    names = cache_names;
1519
1520    /*
1521     * Initialize the caches that provide memory for the array cache and the
1522     * kmem_list3 structures first. Without this, further allocations will
1523     * bug.
1524     */
1525
1526    sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1527                    sizes[INDEX_AC].cs_size,
1528                    ARCH_KMALLOC_MINALIGN,
1529                    ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1530                    NULL);
1531
1532    if (INDEX_AC != INDEX_L3) {
1533        sizes[INDEX_L3].cs_cachep =
1534            kmem_cache_create(names[INDEX_L3].name,
1535                sizes[INDEX_L3].cs_size,
1536                ARCH_KMALLOC_MINALIGN,
1537                ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1538                NULL);
1539    }
1540
1541    slab_early_init = 0;
1542
1543    while (sizes->cs_size != ULONG_MAX) {
1544        /*
1545         * For performance, all the general caches are L1 aligned.
1546         * This should be particularly beneficial on SMP boxes, as it
1547         * eliminates "false sharing".
1548         * Note for systems short on memory removing the alignment will
1549         * allow tighter packing of the smaller caches.
1550         */
1551        if (!sizes->cs_cachep) {
1552            sizes->cs_cachep = kmem_cache_create(names->name,
1553                    sizes->cs_size,
1554                    ARCH_KMALLOC_MINALIGN,
1555                    ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1556                    NULL);
1557        }
1558#ifdef CONFIG_ZONE_DMA
1559        sizes->cs_dmacachep = kmem_cache_create(
1560                    names->name_dma,
1561                    sizes->cs_size,
1562                    ARCH_KMALLOC_MINALIGN,
1563                    ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1564                        SLAB_PANIC,
1565                    NULL);
1566#endif
1567        sizes++;
1568        names++;
1569    }
1570    /* 4) Replace the bootstrap head arrays */
1571    {
1572        struct array_cache *ptr;
1573
1574        ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1575
1576        BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1577        memcpy(ptr, cpu_cache_get(&cache_cache),
1578               sizeof(struct arraycache_init));
1579        /*
1580         * Do not assume that spinlocks can be initialized via memcpy:
1581         */
1582        spin_lock_init(&ptr->lock);
1583
1584        cache_cache.array[smp_processor_id()] = ptr;
1585
1586        ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1587
1588        BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1589               != &initarray_generic.cache);
1590        memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1591               sizeof(struct arraycache_init));
1592        /*
1593         * Do not assume that spinlocks can be initialized via memcpy:
1594         */
1595        spin_lock_init(&ptr->lock);
1596
1597        malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1598            ptr;
1599    }
1600    /* 5) Replace the bootstrap kmem_list3's */
1601    {
1602        int nid;
1603
1604        for_each_online_node(nid) {
1605            init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1606
1607            init_list(malloc_sizes[INDEX_AC].cs_cachep,
1608                  &initkmem_list3[SIZE_AC + nid], nid);
1609
1610            if (INDEX_AC != INDEX_L3) {
1611                init_list(malloc_sizes[INDEX_L3].cs_cachep,
1612                      &initkmem_list3[SIZE_L3 + nid], nid);
1613            }
1614        }
1615    }
1616
1617    g_cpucache_up = EARLY;
1618}
1619
1620void __init kmem_cache_init_late(void)
1621{
1622    struct kmem_cache *cachep;
1623
1624    /* 6) resize the head arrays to their final sizes */
1625    mutex_lock(&cache_chain_mutex);
1626    list_for_each_entry(cachep, &cache_chain, next)
1627        if (enable_cpucache(cachep, GFP_NOWAIT))
1628            BUG();
1629    mutex_unlock(&cache_chain_mutex);
1630
1631    /* Done! */
1632    g_cpucache_up = FULL;
1633
1634    /* Annotate slab for lockdep -- annotate the malloc caches */
1635    init_lock_keys();
1636
1637    /*
1638     * Register a cpu startup notifier callback that initializes
1639     * cpu_cache_get for all new cpus
1640     */
1641    register_cpu_notifier(&cpucache_notifier);
1642
1643#ifdef CONFIG_NUMA
1644    /*
1645     * Register a memory hotplug callback that initializes and frees
1646     * nodelists.
1647     */
1648    hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1649#endif
1650
1651    /*
1652     * The reap timers are started later, with a module init call: That part
1653     * of the kernel is not yet operational.
1654     */
1655}
1656
1657static int __init cpucache_init(void)
1658{
1659    int cpu;
1660
1661    /*
1662     * Register the timers that return unneeded pages to the page allocator
1663     */
1664    for_each_online_cpu(cpu)
1665        start_cpu_timer(cpu);
1666    return 0;
1667}
1668__initcall(cpucache_init);
1669
1670/*
1671 * Interface to system's page allocator. No need to hold the cache-lock.
1672 *
1673 * If we requested dmaable memory, we will get it. Even if we
1674 * did not request dmaable memory, we might get it, but that
1675 * would be relatively rare and ignorable.
1676 */
1677static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1678{
1679    struct page *page;
1680    int nr_pages;
1681    int i;
1682
1683#ifndef CONFIG_MMU
1684    /*
1685     * Nommu uses slab's for process anonymous memory allocations, and thus
1686     * requires __GFP_COMP to properly refcount higher order allocations
1687     */
1688    flags |= __GFP_COMP;
1689#endif
1690
1691    flags |= cachep->gfpflags;
1692    if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1693        flags |= __GFP_RECLAIMABLE;
1694
1695    page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1696    if (!page)
1697        return NULL;
1698
1699    nr_pages = (1 << cachep->gfporder);
1700    if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1701        add_zone_page_state(page_zone(page),
1702            NR_SLAB_RECLAIMABLE, nr_pages);
1703    else
1704        add_zone_page_state(page_zone(page),
1705            NR_SLAB_UNRECLAIMABLE, nr_pages);
1706    for (i = 0; i < nr_pages; i++)
1707        __SetPageSlab(page + i);
1708
1709    if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1710        kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1711
1712        if (cachep->ctor)
1713            kmemcheck_mark_uninitialized_pages(page, nr_pages);
1714        else
1715            kmemcheck_mark_unallocated_pages(page, nr_pages);
1716    }
1717
1718    return page_address(page);
1719}
1720
1721/*
1722 * Interface to system's page release.
1723 */
1724static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1725{
1726    unsigned long i = (1 << cachep->gfporder);
1727    struct page *page = virt_to_page(addr);
1728    const unsigned long nr_freed = i;
1729
1730    kmemcheck_free_shadow(page, cachep->gfporder);
1731
1732    if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1733        sub_zone_page_state(page_zone(page),
1734                NR_SLAB_RECLAIMABLE, nr_freed);
1735    else
1736        sub_zone_page_state(page_zone(page),
1737                NR_SLAB_UNRECLAIMABLE, nr_freed);
1738    while (i--) {
1739        BUG_ON(!PageSlab(page));
1740        __ClearPageSlab(page);
1741        page++;
1742    }
1743    if (current->reclaim_state)
1744        current->reclaim_state->reclaimed_slab += nr_freed;
1745    free_pages((unsigned long)addr, cachep->gfporder);
1746}
1747
1748static void kmem_rcu_free(struct rcu_head *head)
1749{
1750    struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1751    struct kmem_cache *cachep = slab_rcu->cachep;
1752
1753    kmem_freepages(cachep, slab_rcu->addr);
1754    if (OFF_SLAB(cachep))
1755        kmem_cache_free(cachep->slabp_cache, slab_rcu);
1756}
1757
1758#if DEBUG
1759
1760#ifdef CONFIG_DEBUG_PAGEALLOC
1761static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1762                unsigned long caller)
1763{
1764    int size = obj_size(cachep);
1765
1766    addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1767
1768    if (size < 5 * sizeof(unsigned long))
1769        return;
1770
1771    *addr++ = 0x12345678;
1772    *addr++ = caller;
1773    *addr++ = smp_processor_id();
1774    size -= 3 * sizeof(unsigned long);
1775    {
1776        unsigned long *sptr = &caller;
1777        unsigned long svalue;
1778
1779        while (!kstack_end(sptr)) {
1780            svalue = *sptr++;
1781            if (kernel_text_address(svalue)) {
1782                *addr++ = svalue;
1783                size -= sizeof(unsigned long);
1784                if (size <= sizeof(unsigned long))
1785                    break;
1786            }
1787        }
1788
1789    }
1790    *addr++ = 0x87654321;
1791}
1792#endif
1793
1794static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1795{
1796    int size = obj_size(cachep);
1797    addr = &((char *)addr)[obj_offset(cachep)];
1798
1799    memset(addr, val, size);
1800    *(unsigned char *)(addr + size - 1) = POISON_END;
1801}
1802
1803static void dump_line(char *data, int offset, int limit)
1804{
1805    int i;
1806    unsigned char error = 0;
1807    int bad_count = 0;
1808
1809    printk(KERN_ERR "%03x:", offset);
1810    for (i = 0; i < limit; i++) {
1811        if (data[offset + i] != POISON_FREE) {
1812            error = data[offset + i];
1813            bad_count++;
1814        }
1815        printk(" %02x", (unsigned char)data[offset + i]);
1816    }
1817    printk("\n");
1818
1819    if (bad_count == 1) {
1820        error ^= POISON_FREE;
1821        if (!(error & (error - 1))) {
1822            printk(KERN_ERR "Single bit error detected. Probably "
1823                    "bad RAM.\n");
1824#ifdef CONFIG_X86
1825            printk(KERN_ERR "Run memtest86+ or a similar memory "
1826                    "test tool.\n");
1827#else
1828            printk(KERN_ERR "Run a memory test tool.\n");
1829#endif
1830        }
1831    }
1832}
1833#endif
1834
1835#if DEBUG
1836
1837static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1838{
1839    int i, size;
1840    char *realobj;
1841
1842    if (cachep->flags & SLAB_RED_ZONE) {
1843        printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1844            *dbg_redzone1(cachep, objp),
1845            *dbg_redzone2(cachep, objp));
1846    }
1847
1848    if (cachep->flags & SLAB_STORE_USER) {
1849        printk(KERN_ERR "Last user: [<%p>]",
1850            *dbg_userword(cachep, objp));
1851        print_symbol("(%s)",
1852                (unsigned long)*dbg_userword(cachep, objp));
1853        printk("\n");
1854    }
1855    realobj = (char *)objp + obj_offset(cachep);
1856    size = obj_size(cachep);
1857    for (i = 0; i < size && lines; i += 16, lines--) {
1858        int limit;
1859        limit = 16;
1860        if (i + limit > size)
1861            limit = size - i;
1862        dump_line(realobj, i, limit);
1863    }
1864}
1865
1866static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1867{
1868    char *realobj;
1869    int size, i;
1870    int lines = 0;
1871
1872    realobj = (char *)objp + obj_offset(cachep);
1873    size = obj_size(cachep);
1874
1875    for (i = 0; i < size; i++) {
1876        char exp = POISON_FREE;
1877        if (i == size - 1)
1878            exp = POISON_END;
1879        if (realobj[i] != exp) {
1880            int limit;
1881            /* Mismatch ! */
1882            /* Print header */
1883            if (lines == 0) {
1884                printk(KERN_ERR
1885                    "Slab corruption: %s start=%p, len=%d\n",
1886                    cachep->name, realobj, size);
1887                print_objinfo(cachep, objp, 0);
1888            }
1889            /* Hexdump the affected line */
1890            i = (i / 16) * 16;
1891            limit = 16;
1892            if (i + limit > size)
1893                limit = size - i;
1894            dump_line(realobj, i, limit);
1895            i += 16;
1896            lines++;
1897            /* Limit to 5 lines */
1898            if (lines > 5)
1899                break;
1900        }
1901    }
1902    if (lines != 0) {
1903        /* Print some data about the neighboring objects, if they
1904         * exist:
1905         */
1906        struct slab *slabp = virt_to_slab(objp);
1907        unsigned int objnr;
1908
1909        objnr = obj_to_index(cachep, slabp, objp);
1910        if (objnr) {
1911            objp = index_to_obj(cachep, slabp, objnr - 1);
1912            realobj = (char *)objp + obj_offset(cachep);
1913            printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1914                   realobj, size);
1915            print_objinfo(cachep, objp, 2);
1916        }
1917        if (objnr + 1 < cachep->num) {
1918            objp = index_to_obj(cachep, slabp, objnr + 1);
1919            realobj = (char *)objp + obj_offset(cachep);
1920            printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1921                   realobj, size);
1922            print_objinfo(cachep, objp, 2);
1923        }
1924    }
1925}
1926#endif
1927
1928#if DEBUG
1929static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1930{
1931    int i;
1932    for (i = 0; i < cachep->num; i++) {
1933        void *objp = index_to_obj(cachep, slabp, i);
1934
1935        if (cachep->flags & SLAB_POISON) {
1936#ifdef CONFIG_DEBUG_PAGEALLOC
1937            if (cachep->buffer_size % PAGE_SIZE == 0 &&
1938                    OFF_SLAB(cachep))
1939                kernel_map_pages(virt_to_page(objp),
1940                    cachep->buffer_size / PAGE_SIZE, 1);
1941            else
1942                check_poison_obj(cachep, objp);
1943#else
1944            check_poison_obj(cachep, objp);
1945#endif
1946        }
1947        if (cachep->flags & SLAB_RED_ZONE) {
1948            if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1949                slab_error(cachep, "start of a freed object "
1950                       "was overwritten");
1951            if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1952                slab_error(cachep, "end of a freed object "
1953                       "was overwritten");
1954        }
1955    }
1956}
1957#else
1958static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1959{
1960}
1961#endif
1962
1963/**
1964 * slab_destroy - destroy and release all objects in a slab
1965 * @cachep: cache pointer being destroyed
1966 * @slabp: slab pointer being destroyed
1967 *
1968 * Destroy all the objs in a slab, and release the mem back to the system.
1969 * Before calling the slab must have been unlinked from the cache. The
1970 * cache-lock is not held/needed.
1971 */
1972static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1973{
1974    void *addr = slabp->s_mem - slabp->colouroff;
1975
1976    slab_destroy_debugcheck(cachep, slabp);
1977    if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1978        struct slab_rcu *slab_rcu;
1979
1980        slab_rcu = (struct slab_rcu *)slabp;
1981        slab_rcu->cachep = cachep;
1982        slab_rcu->addr = addr;
1983        call_rcu(&slab_rcu->head, kmem_rcu_free);
1984    } else {
1985        kmem_freepages(cachep, addr);
1986        if (OFF_SLAB(cachep))
1987            kmem_cache_free(cachep->slabp_cache, slabp);
1988    }
1989}
1990
1991static void __kmem_cache_destroy(struct kmem_cache *cachep)
1992{
1993    int i;
1994    struct kmem_list3 *l3;
1995
1996    for_each_online_cpu(i)
1997        kfree(cachep->array[i]);
1998
1999    /* NUMA: free the list3 structures */
2000    for_each_online_node(i) {
2001        l3 = cachep->nodelists[i];
2002        if (l3) {
2003            kfree(l3->shared);
2004            free_alien_cache(l3->alien);
2005            kfree(l3);
2006        }
2007    }
2008    kmem_cache_free(&cache_cache, cachep);
2009}
2010
2011
2012/**
2013 * calculate_slab_order - calculate size (page order) of slabs
2014 * @cachep: pointer to the cache that is being created
2015 * @size: size of objects to be created in this cache.
2016 * @align: required alignment for the objects.
2017 * @flags: slab allocation flags
2018 *
2019 * Also calculates the number of objects per slab.
2020 *
2021 * This could be made much more intelligent. For now, try to avoid using
2022 * high order pages for slabs. When the gfp() functions are more friendly
2023 * towards high-order requests, this should be changed.
2024 */
2025static size_t calculate_slab_order(struct kmem_cache *cachep,
2026            size_t size, size_t align, unsigned long flags)
2027{
2028    unsigned long offslab_limit;
2029    size_t left_over = 0;
2030    int gfporder;
2031
2032    for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2033        unsigned int num;
2034        size_t remainder;
2035
2036        cache_estimate(gfporder, size, align, flags, &remainder, &num);
2037        if (!num)
2038            continue;
2039
2040        if (flags & CFLGS_OFF_SLAB) {
2041            /*
2042             * Max number of objs-per-slab for caches which
2043             * use off-slab slabs. Needed to avoid a possible
2044             * looping condition in cache_grow().
2045             */
2046            offslab_limit = size - sizeof(struct slab);
2047            offslab_limit /= sizeof(kmem_bufctl_t);
2048
2049             if (num > offslab_limit)
2050                break;
2051        }
2052
2053        /* Found something acceptable - save it away */
2054        cachep->num = num;
2055        cachep->gfporder = gfporder;
2056        left_over = remainder;
2057
2058        /*
2059         * A VFS-reclaimable slab tends to have most allocations
2060         * as GFP_NOFS and we really don't want to have to be allocating
2061         * higher-order pages when we are unable to shrink dcache.
2062         */
2063        if (flags & SLAB_RECLAIM_ACCOUNT)
2064            break;
2065
2066        /*
2067         * Large number of objects is good, but very large slabs are
2068         * currently bad for the gfp()s.
2069         */
2070        if (gfporder >= slab_break_gfp_order)
2071            break;
2072
2073        /*
2074         * Acceptable internal fragmentation?
2075         */
2076        if (left_over * 8 <= (PAGE_SIZE << gfporder))
2077            break;
2078    }
2079    return left_over;
2080}
2081
2082static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2083{
2084    if (g_cpucache_up == FULL)
2085        return enable_cpucache(cachep, gfp);
2086
2087    if (g_cpucache_up == NONE) {
2088        /*
2089         * Note: the first kmem_cache_create must create the cache
2090         * that's used by kmalloc(24), otherwise the creation of
2091         * further caches will BUG().
2092         */
2093        cachep->array[smp_processor_id()] = &initarray_generic.cache;
2094
2095        /*
2096         * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2097         * the first cache, then we need to set up all its list3s,
2098         * otherwise the creation of further caches will BUG().
2099         */
2100        set_up_list3s(cachep, SIZE_AC);
2101        if (INDEX_AC == INDEX_L3)
2102            g_cpucache_up = PARTIAL_L3;
2103        else
2104            g_cpucache_up = PARTIAL_AC;
2105    } else {
2106        cachep->array[smp_processor_id()] =
2107            kmalloc(sizeof(struct arraycache_init), gfp);
2108
2109        if (g_cpucache_up == PARTIAL_AC) {
2110            set_up_list3s(cachep, SIZE_L3);
2111            g_cpucache_up = PARTIAL_L3;
2112        } else {
2113            int node;
2114            for_each_online_node(node) {
2115                cachep->nodelists[node] =
2116                    kmalloc_node(sizeof(struct kmem_list3),
2117                        gfp, node);
2118                BUG_ON(!cachep->nodelists[node]);
2119                kmem_list3_init(cachep->nodelists[node]);
2120            }
2121        }
2122    }
2123    cachep->nodelists[numa_mem_id()]->next_reap =
2124            jiffies + REAPTIMEOUT_LIST3 +
2125            ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2126
2127    cpu_cache_get(cachep)->avail = 0;
2128    cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2129    cpu_cache_get(cachep)->batchcount = 1;
2130    cpu_cache_get(cachep)->touched = 0;
2131    cachep->batchcount = 1;
2132    cachep->limit = BOOT_CPUCACHE_ENTRIES;
2133    return 0;
2134}
2135
2136/**
2137 * kmem_cache_create - Create a cache.
2138 * @name: A string which is used in /proc/slabinfo to identify this cache.
2139 * @size: The size of objects to be created in this cache.
2140 * @align: The required alignment for the objects.
2141 * @flags: SLAB flags
2142 * @ctor: A constructor for the objects.
2143 *
2144 * Returns a ptr to the cache on success, NULL on failure.
2145 * Cannot be called within a int, but can be interrupted.
2146 * The @ctor is run when new pages are allocated by the cache.
2147 *
2148 * @name must be valid until the cache is destroyed. This implies that
2149 * the module calling this has to destroy the cache before getting unloaded.
2150 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2151 * therefore applications must manage it themselves.
2152 *
2153 * The flags are
2154 *
2155 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2156 * to catch references to uninitialised memory.
2157 *
2158 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2159 * for buffer overruns.
2160 *
2161 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2162 * cacheline. This can be beneficial if you're counting cycles as closely
2163 * as davem.
2164 */
2165struct kmem_cache *
2166kmem_cache_create (const char *name, size_t size, size_t align,
2167    unsigned long flags, void (*ctor)(void *))
2168{
2169    size_t left_over, slab_size, ralign;
2170    struct kmem_cache *cachep = NULL, *pc;
2171    gfp_t gfp;
2172
2173    /*
2174     * Sanity checks... these are all serious usage bugs.
2175     */
2176    if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2177        size > KMALLOC_MAX_SIZE) {
2178        printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2179                name);
2180        BUG();
2181    }
2182
2183    /*
2184     * We use cache_chain_mutex to ensure a consistent view of
2185     * cpu_online_mask as well. Please see cpuup_callback
2186     */
2187    if (slab_is_available()) {
2188        get_online_cpus();
2189        mutex_lock(&cache_chain_mutex);
2190    }
2191
2192    list_for_each_entry(pc, &cache_chain, next) {
2193        char tmp;
2194        int res;
2195
2196        /*
2197         * This happens when the module gets unloaded and doesn't
2198         * destroy its slab cache and no-one else reuses the vmalloc
2199         * area of the module. Print a warning.
2200         */
2201        res = probe_kernel_address(pc->name, tmp);
2202        if (res) {
2203            printk(KERN_ERR
2204                   "SLAB: cache with size %d has lost its name\n",
2205                   pc->buffer_size);
2206            continue;
2207        }
2208
2209        if (!strcmp(pc->name, name)) {
2210            printk(KERN_ERR
2211                   "kmem_cache_create: duplicate cache %s\n", name);
2212            dump_stack();
2213            goto oops;
2214        }
2215    }
2216
2217#if DEBUG
2218    WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2219#if FORCED_DEBUG
2220    /*
2221     * Enable redzoning and last user accounting, except for caches with
2222     * large objects, if the increased size would increase the object size
2223     * above the next power of two: caches with object sizes just above a
2224     * power of two have a significant amount of internal fragmentation.
2225     */
2226    if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2227                        2 * sizeof(unsigned long long)))
2228        flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2229    if (!(flags & SLAB_DESTROY_BY_RCU))
2230        flags |= SLAB_POISON;
2231#endif
2232    if (flags & SLAB_DESTROY_BY_RCU)
2233        BUG_ON(flags & SLAB_POISON);
2234#endif
2235    /*
2236     * Always checks flags, a caller might be expecting debug support which
2237     * isn't available.
2238     */
2239    BUG_ON(flags & ~CREATE_MASK);
2240
2241    /*
2242     * Check that size is in terms of words. This is needed to avoid
2243     * unaligned accesses for some archs when redzoning is used, and makes
2244     * sure any on-slab bufctl's are also correctly aligned.
2245     */
2246    if (size & (BYTES_PER_WORD - 1)) {
2247        size += (BYTES_PER_WORD - 1);
2248        size &= ~(BYTES_PER_WORD - 1);
2249    }
2250
2251    /* calculate the final buffer alignment: */
2252
2253    /* 1) arch recommendation: can be overridden for debug */
2254    if (flags & SLAB_HWCACHE_ALIGN) {
2255        /*
2256         * Default alignment: as specified by the arch code. Except if
2257         * an object is really small, then squeeze multiple objects into
2258         * one cacheline.
2259         */
2260        ralign = cache_line_size();
2261        while (size <= ralign / 2)
2262            ralign /= 2;
2263    } else {
2264        ralign = BYTES_PER_WORD;
2265    }
2266
2267    /*
2268     * Redzoning and user store require word alignment or possibly larger.
2269     * Note this will be overridden by architecture or caller mandated
2270     * alignment if either is greater than BYTES_PER_WORD.
2271     */
2272    if (flags & SLAB_STORE_USER)
2273        ralign = BYTES_PER_WORD;
2274
2275    if (flags & SLAB_RED_ZONE) {
2276        ralign = REDZONE_ALIGN;
2277        /* If redzoning, ensure that the second redzone is suitably
2278         * aligned, by adjusting the object size accordingly. */
2279        size += REDZONE_ALIGN - 1;
2280        size &= ~(REDZONE_ALIGN - 1);
2281    }
2282
2283    /* 2) arch mandated alignment */
2284    if (ralign < ARCH_SLAB_MINALIGN) {
2285        ralign = ARCH_SLAB_MINALIGN;
2286    }
2287    /* 3) caller mandated alignment */
2288    if (ralign < align) {
2289        ralign = align;
2290    }
2291    /* disable debug if not aligning with REDZONE_ALIGN */
2292    if (ralign & (__alignof__(unsigned long long) - 1))
2293        flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2294    /*
2295     * 4) Store it.
2296     */
2297    align = ralign;
2298
2299    if (slab_is_available())
2300        gfp = GFP_KERNEL;
2301    else
2302        gfp = GFP_NOWAIT;
2303
2304    /* Get cache's description obj. */
2305    cachep = kmem_cache_zalloc(&cache_cache, gfp);
2306    if (!cachep)
2307        goto oops;
2308
2309#if DEBUG
2310    cachep->obj_size = size;
2311
2312    /*
2313     * Both debugging options require word-alignment which is calculated
2314     * into align above.
2315     */
2316    if (flags & SLAB_RED_ZONE) {
2317        /* add space for red zone words */
2318        cachep->obj_offset += align;
2319        size += align + sizeof(unsigned long long);
2320    }
2321    if (flags & SLAB_STORE_USER) {
2322        /* user store requires one word storage behind the end of
2323         * the real object. But if the second red zone needs to be
2324         * aligned to 64 bits, we must allow that much space.
2325         */
2326        if (flags & SLAB_RED_ZONE)
2327            size += REDZONE_ALIGN;
2328        else
2329            size += BYTES_PER_WORD;
2330    }
2331#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2332    if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2333        && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2334        cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2335        size = PAGE_SIZE;
2336    }
2337#endif
2338#endif
2339
2340    /*
2341     * Determine if the slab management is 'on' or 'off' slab.
2342     * (bootstrapping cannot cope with offslab caches so don't do
2343     * it too early on. Always use on-slab management when
2344     * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2345     */
2346    if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2347        !(flags & SLAB_NOLEAKTRACE))
2348        /*
2349         * Size is large, assume best to place the slab management obj
2350         * off-slab (should allow better packing of objs).
2351         */
2352        flags |= CFLGS_OFF_SLAB;
2353
2354    size = ALIGN(size, align);
2355
2356    left_over = calculate_slab_order(cachep, size, align, flags);
2357
2358    if (!cachep->num) {
2359        printk(KERN_ERR
2360               "kmem_cache_create: couldn't create cache %s.\n", name);
2361        kmem_cache_free(&cache_cache, cachep);
2362        cachep = NULL;
2363        goto oops;
2364    }
2365    slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2366              + sizeof(struct slab), align);
2367
2368    /*
2369     * If the slab has been placed off-slab, and we have enough space then
2370     * move it on-slab. This is at the expense of any extra colouring.
2371     */
2372    if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2373        flags &= ~CFLGS_OFF_SLAB;
2374        left_over -= slab_size;
2375    }
2376
2377    if (flags & CFLGS_OFF_SLAB) {
2378        /* really off slab. No need for manual alignment */
2379        slab_size =
2380            cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2381
2382#ifdef CONFIG_PAGE_POISONING
2383        /* If we're going to use the generic kernel_map_pages()
2384         * poisoning, then it's going to smash the contents of
2385         * the redzone and userword anyhow, so switch them off.
2386         */
2387        if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2388            flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2389#endif
2390    }
2391
2392    cachep->colour_off = cache_line_size();
2393    /* Offset must be a multiple of the alignment. */
2394    if (cachep->colour_off < align)
2395        cachep->colour_off = align;
2396    cachep->colour = left_over / cachep->colour_off;
2397    cachep->slab_size = slab_size;
2398    cachep->flags = flags;
2399    cachep->gfpflags = 0;
2400    if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2401        cachep->gfpflags |= GFP_DMA;
2402    cachep->buffer_size = size;
2403    cachep->reciprocal_buffer_size = reciprocal_value(size);
2404
2405    if (flags & CFLGS_OFF_SLAB) {
2406        cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2407        /*
2408         * This is a possibility for one of the malloc_sizes caches.
2409         * But since we go off slab only for object size greater than
2410         * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2411         * this should not happen at all.
2412         * But leave a BUG_ON for some lucky dude.
2413         */
2414        BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2415    }
2416    cachep->ctor = ctor;
2417    cachep->name = name;
2418
2419    if (setup_cpu_cache(cachep, gfp)) {
2420        __kmem_cache_destroy(cachep);
2421        cachep = NULL;
2422        goto oops;
2423    }
2424
2425    /* cache setup completed, link it into the list */
2426    list_add(&cachep->next, &cache_chain);
2427oops:
2428    if (!cachep && (flags & SLAB_PANIC))
2429        panic("kmem_cache_create(): failed to create slab `%s'\n",
2430              name);
2431    if (slab_is_available()) {
2432        mutex_unlock(&cache_chain_mutex);
2433        put_online_cpus();
2434    }
2435    return cachep;
2436}
2437EXPORT_SYMBOL(kmem_cache_create);
2438
2439#if DEBUG
2440static void check_irq_off(void)
2441{
2442    BUG_ON(!irqs_disabled());
2443}
2444
2445static void check_irq_on(void)
2446{
2447    BUG_ON(irqs_disabled());
2448}
2449
2450static void check_spinlock_acquired(struct kmem_cache *cachep)
2451{
2452#ifdef CONFIG_SMP
2453    check_irq_off();
2454    assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2455#endif
2456}
2457
2458static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2459{
2460#ifdef CONFIG_SMP
2461    check_irq_off();
2462    assert_spin_locked(&cachep->nodelists[node]->list_lock);
2463#endif
2464}
2465
2466#else
2467#define check_irq_off() do { } while(0)
2468#define check_irq_on() do { } while(0)
2469#define check_spinlock_acquired(x) do { } while(0)
2470#define check_spinlock_acquired_node(x, y) do { } while(0)
2471#endif
2472
2473static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2474            struct array_cache *ac,
2475            int force, int node);
2476
2477static void do_drain(void *arg)
2478{
2479    struct kmem_cache *cachep = arg;
2480    struct array_cache *ac;
2481    int node = numa_mem_id();
2482
2483    check_irq_off();
2484    ac = cpu_cache_get(cachep);
2485    spin_lock(&cachep->nodelists[node]->list_lock);
2486    free_block(cachep, ac->entry, ac->avail, node);
2487    spin_unlock(&cachep->nodelists[node]->list_lock);
2488    ac->avail = 0;
2489}
2490
2491static void drain_cpu_caches(struct kmem_cache *cachep)
2492{
2493    struct kmem_list3 *l3;
2494    int node;
2495
2496    on_each_cpu(do_drain, cachep, 1);
2497    check_irq_on();
2498    for_each_online_node(node) {
2499        l3 = cachep->nodelists[node];
2500        if (l3 && l3->alien)
2501            drain_alien_cache(cachep, l3->alien);
2502    }
2503
2504    for_each_online_node(node) {
2505        l3 = cachep->nodelists[node];
2506        if (l3)
2507            drain_array(cachep, l3, l3->shared, 1, node);
2508    }
2509}
2510
2511/*
2512 * Remove slabs from the list of free slabs.
2513 * Specify the number of slabs to drain in tofree.
2514 *
2515 * Returns the actual number of slabs released.
2516 */
2517static int drain_freelist(struct kmem_cache *cache,
2518            struct kmem_list3 *l3, int tofree)
2519{
2520    struct list_head *p;
2521    int nr_freed;
2522    struct slab *slabp;
2523
2524    nr_freed = 0;
2525    while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2526
2527        spin_lock_irq(&l3->list_lock);
2528        p = l3->slabs_free.prev;
2529        if (p == &l3->slabs_free) {
2530            spin_unlock_irq(&l3->list_lock);
2531            goto out;
2532        }
2533
2534        slabp = list_entry(p, struct slab, list);
2535#if DEBUG
2536        BUG_ON(slabp->inuse);
2537#endif
2538        list_del(&slabp->list);
2539        /*
2540         * Safe to drop the lock. The slab is no longer linked
2541         * to the cache.
2542         */
2543        l3->free_objects -= cache->num;
2544        spin_unlock_irq(&l3->list_lock);
2545        slab_destroy(cache, slabp);
2546        nr_freed++;
2547    }
2548out:
2549    return nr_freed;
2550}
2551
2552/* Called with cache_chain_mutex held to protect against cpu hotplug */
2553static int __cache_shrink(struct kmem_cache *cachep)
2554{
2555    int ret = 0, i = 0;
2556    struct kmem_list3 *l3;
2557
2558    drain_cpu_caches(cachep);
2559
2560    check_irq_on();
2561    for_each_online_node(i) {
2562        l3 = cachep->nodelists[i];
2563        if (!l3)
2564            continue;
2565
2566        drain_freelist(cachep, l3, l3->free_objects);
2567
2568        ret += !list_empty(&l3->slabs_full) ||
2569            !list_empty(&l3->slabs_partial);
2570    }
2571    return (ret ? 1 : 0);
2572}
2573
2574/**
2575 * kmem_cache_shrink - Shrink a cache.
2576 * @cachep: The cache to shrink.
2577 *
2578 * Releases as many slabs as possible for a cache.
2579 * To help debugging, a zero exit status indicates all slabs were released.
2580 */
2581int kmem_cache_shrink(struct kmem_cache *cachep)
2582{
2583    int ret;
2584    BUG_ON(!cachep || in_interrupt());
2585
2586    get_online_cpus();
2587    mutex_lock(&cache_chain_mutex);
2588    ret = __cache_shrink(cachep);
2589    mutex_unlock(&cache_chain_mutex);
2590    put_online_cpus();
2591    return ret;
2592}
2593EXPORT_SYMBOL(kmem_cache_shrink);
2594
2595/**
2596 * kmem_cache_destroy - delete a cache
2597 * @cachep: the cache to destroy
2598 *
2599 * Remove a &struct kmem_cache object from the slab cache.
2600 *
2601 * It is expected this function will be called by a module when it is
2602 * unloaded. This will remove the cache completely, and avoid a duplicate
2603 * cache being allocated each time a module is loaded and unloaded, if the
2604 * module doesn't have persistent in-kernel storage across loads and unloads.
2605 *
2606 * The cache must be empty before calling this function.
2607 *
2608 * The caller must guarantee that noone will allocate memory from the cache
2609 * during the kmem_cache_destroy().
2610 */
2611void kmem_cache_destroy(struct kmem_cache *cachep)
2612{
2613    BUG_ON(!cachep || in_interrupt());
2614
2615    /* Find the cache in the chain of caches. */
2616    get_online_cpus();
2617    mutex_lock(&cache_chain_mutex);
2618    /*
2619     * the chain is never empty, cache_cache is never destroyed
2620     */
2621    list_del(&cachep->next);
2622    if (__cache_shrink(cachep)) {
2623        slab_error(cachep, "Can't free all objects");
2624        list_add(&cachep->next, &cache_chain);
2625        mutex_unlock(&cache_chain_mutex);
2626        put_online_cpus();
2627        return;
2628    }
2629
2630    if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2631        rcu_barrier();
2632
2633    __kmem_cache_destroy(cachep);
2634    mutex_unlock(&cache_chain_mutex);
2635    put_online_cpus();
2636}
2637EXPORT_SYMBOL(kmem_cache_destroy);
2638
2639/*
2640 * Get the memory for a slab management obj.
2641 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2642 * always come from malloc_sizes caches. The slab descriptor cannot
2643 * come from the same cache which is getting created because,
2644 * when we are searching for an appropriate cache for these
2645 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2646 * If we are creating a malloc_sizes cache here it would not be visible to
2647 * kmem_find_general_cachep till the initialization is complete.
2648 * Hence we cannot have slabp_cache same as the original cache.
2649 */
2650static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2651                   int colour_off, gfp_t local_flags,
2652                   int nodeid)
2653{
2654    struct slab *slabp;
2655
2656    if (OFF_SLAB(cachep)) {
2657        /* Slab management obj is off-slab. */
2658        slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2659                          local_flags, nodeid);
2660        /*
2661         * If the first object in the slab is leaked (it's allocated
2662         * but no one has a reference to it), we want to make sure
2663         * kmemleak does not treat the ->s_mem pointer as a reference
2664         * to the object. Otherwise we will not report the leak.
2665         */
2666        kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2667                   local_flags);
2668        if (!slabp)
2669            return NULL;
2670    } else {
2671        slabp = objp + colour_off;
2672        colour_off += cachep->slab_size;
2673    }
2674    slabp->inuse = 0;
2675    slabp->colouroff = colour_off;
2676    slabp->s_mem = objp + colour_off;
2677    slabp->nodeid = nodeid;
2678    slabp->free = 0;
2679    return slabp;
2680}
2681
2682static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2683{
2684    return (kmem_bufctl_t *) (slabp + 1);
2685}
2686
2687static void cache_init_objs(struct kmem_cache *cachep,
2688                struct slab *slabp)
2689{
2690    int i;
2691
2692    for (i = 0; i < cachep->num; i++) {
2693        void *objp = index_to_obj(cachep, slabp, i);
2694#if DEBUG
2695        /* need to poison the objs? */
2696        if (cachep->flags & SLAB_POISON)
2697            poison_obj(cachep, objp, POISON_FREE);
2698        if (cachep->flags & SLAB_STORE_USER)
2699            *dbg_userword(cachep, objp) = NULL;
2700
2701        if (cachep->flags & SLAB_RED_ZONE) {
2702            *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2703            *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2704        }
2705        /*
2706         * Constructors are not allowed to allocate memory from the same
2707         * cache which they are a constructor for. Otherwise, deadlock.
2708         * They must also be threaded.
2709         */
2710        if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2711            cachep->ctor(objp + obj_offset(cachep));
2712
2713        if (cachep->flags & SLAB_RED_ZONE) {
2714            if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2715                slab_error(cachep, "constructor overwrote the"
2716                       " end of an object");
2717            if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2718                slab_error(cachep, "constructor overwrote the"
2719                       " start of an object");
2720        }
2721        if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2722                OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2723            kernel_map_pages(virt_to_page(objp),
2724                     cachep->buffer_size / PAGE_SIZE, 0);
2725#else
2726        if (cachep->ctor)
2727            cachep->ctor(objp);
2728#endif
2729        slab_bufctl(slabp)[i] = i + 1;
2730    }
2731    slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2732}
2733
2734static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2735{
2736    if (CONFIG_ZONE_DMA_FLAG) {
2737        if (flags & GFP_DMA)
2738            BUG_ON(!(cachep->gfpflags & GFP_DMA));
2739        else
2740            BUG_ON(cachep->gfpflags & GFP_DMA);
2741    }
2742}
2743
2744static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2745                int nodeid)
2746{
2747    void *objp = index_to_obj(cachep, slabp, slabp->free);
2748    kmem_bufctl_t next;
2749
2750    slabp->inuse++;
2751    next = slab_bufctl(slabp)[slabp->free];
2752#if DEBUG
2753    slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2754    WARN_ON(slabp->nodeid != nodeid);
2755#endif
2756    slabp->free = next;
2757
2758    return objp;
2759}
2760
2761static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2762                void *objp, int nodeid)
2763{
2764    unsigned int objnr = obj_to_index(cachep, slabp, objp);
2765
2766#if DEBUG
2767    /* Verify that the slab belongs to the intended node */
2768    WARN_ON(slabp->nodeid != nodeid);
2769
2770    if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2771        printk(KERN_ERR "slab: double free detected in cache "
2772                "'%s', objp %p\n", cachep->name, objp);
2773        BUG();
2774    }
2775#endif
2776    slab_bufctl(slabp)[objnr] = slabp->free;
2777    slabp->free = objnr;
2778    slabp->inuse--;
2779}
2780
2781/*
2782 * Map pages beginning at addr to the given cache and slab. This is required
2783 * for the slab allocator to be able to lookup the cache and slab of a
2784 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2785 */
2786static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2787               void *addr)
2788{
2789    int nr_pages;
2790    struct page *page;
2791
2792    page = virt_to_page(addr);
2793
2794    nr_pages = 1;
2795    if (likely(!PageCompound(page)))
2796        nr_pages <<= cache->gfporder;
2797
2798    do {
2799        page_set_cache(page, cache);
2800        page_set_slab(page, slab);
2801        page++;
2802    } while (--nr_pages);
2803}
2804
2805/*
2806 * Grow (by 1) the number of slabs within a cache. This is called by
2807 * kmem_cache_alloc() when there are no active objs left in a cache.
2808 */
2809static int cache_grow(struct kmem_cache *cachep,
2810        gfp_t flags, int nodeid, void *objp)
2811{
2812    struct slab *slabp;
2813    size_t offset;
2814    gfp_t local_flags;
2815    struct kmem_list3 *l3;
2816
2817    /*
2818     * Be lazy and only check for valid flags here, keeping it out of the
2819     * critical path in kmem_cache_alloc().
2820     */
2821    BUG_ON(flags & GFP_SLAB_BUG_MASK);
2822    local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2823
2824    /* Take the l3 list lock to change the colour_next on this node */
2825    check_irq_off();
2826    l3 = cachep->nodelists[nodeid];
2827    spin_lock(&l3->list_lock);
2828
2829    /* Get colour for the slab, and cal the next value. */
2830    offset = l3->colour_next;
2831    l3->colour_next++;
2832    if (l3->colour_next >= cachep->colour)
2833        l3->colour_next = 0;
2834    spin_unlock(&l3->list_lock);
2835
2836    offset *= cachep->colour_off;
2837
2838    if (local_flags & __GFP_WAIT)
2839        local_irq_enable();
2840
2841    /*
2842     * The test for missing atomic flag is performed here, rather than
2843     * the more obvious place, simply to reduce the critical path length
2844     * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2845     * will eventually be caught here (where it matters).
2846     */
2847    kmem_flagcheck(cachep, flags);
2848
2849    /*
2850     * Get mem for the objs. Attempt to allocate a physical page from
2851     * 'nodeid'.
2852     */
2853    if (!objp)
2854        objp = kmem_getpages(cachep, local_flags, nodeid);
2855    if (!objp)
2856        goto failed;
2857
2858    /* Get slab management. */
2859    slabp = alloc_slabmgmt(cachep, objp, offset,
2860            local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2861    if (!slabp)
2862        goto opps1;
2863
2864    slab_map_pages(cachep, slabp, objp);
2865
2866    cache_init_objs(cachep, slabp);
2867
2868    if (local_flags & __GFP_WAIT)
2869        local_irq_disable();
2870    check_irq_off();
2871    spin_lock(&l3->list_lock);
2872
2873    /* Make slab active. */
2874    list_add_tail(&slabp->list, &(l3->slabs_free));
2875    STATS_INC_GROWN(cachep);
2876    l3->free_objects += cachep->num;
2877    spin_unlock(&l3->list_lock);
2878    return 1;
2879opps1:
2880    kmem_freepages(cachep, objp);
2881failed:
2882    if (local_flags & __GFP_WAIT)
2883        local_irq_disable();
2884    return 0;
2885}
2886
2887#if DEBUG
2888
2889/*
2890 * Perform extra freeing checks:
2891 * - detect bad pointers.
2892 * - POISON/RED_ZONE checking
2893 */
2894static void kfree_debugcheck(const void *objp)
2895{
2896    if (!virt_addr_valid(objp)) {
2897        printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2898               (unsigned long)objp);
2899        BUG();
2900    }
2901}
2902
2903static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2904{
2905    unsigned long long redzone1, redzone2;
2906
2907    redzone1 = *dbg_redzone1(cache, obj);
2908    redzone2 = *dbg_redzone2(cache, obj);
2909
2910    /*
2911     * Redzone is ok.
2912     */
2913    if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2914        return;
2915
2916    if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2917        slab_error(cache, "double free detected");
2918    else
2919        slab_error(cache, "memory outside object was overwritten");
2920
2921    printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2922            obj, redzone1, redzone2);
2923}
2924
2925static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2926                   void *caller)
2927{
2928    struct page *page;
2929    unsigned int objnr;
2930    struct slab *slabp;
2931
2932    BUG_ON(virt_to_cache(objp) != cachep);
2933
2934    objp -= obj_offset(cachep);
2935    kfree_debugcheck(objp);
2936    page = virt_to_head_page(objp);
2937
2938    slabp = page_get_slab(page);
2939
2940    if (cachep->flags & SLAB_RED_ZONE) {
2941        verify_redzone_free(cachep, objp);
2942        *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2943        *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2944    }
2945    if (cachep->flags & SLAB_STORE_USER)
2946        *dbg_userword(cachep, objp) = caller;
2947
2948    objnr = obj_to_index(cachep, slabp, objp);
2949
2950    BUG_ON(objnr >= cachep->num);
2951    BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2952
2953#ifdef CONFIG_DEBUG_SLAB_LEAK
2954    slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2955#endif
2956    if (cachep->flags & SLAB_POISON) {
2957#ifdef CONFIG_DEBUG_PAGEALLOC
2958        if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2959            store_stackinfo(cachep, objp, (unsigned long)caller);
2960            kernel_map_pages(virt_to_page(objp),
2961                     cachep->buffer_size / PAGE_SIZE, 0);
2962        } else {
2963            poison_obj(cachep, objp, POISON_FREE);
2964        }
2965#else
2966        poison_obj(cachep, objp, POISON_FREE);
2967#endif
2968    }
2969    return objp;
2970}
2971
2972static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2973{
2974    kmem_bufctl_t i;
2975    int entries = 0;
2976
2977    /* Check slab's freelist to see if this obj is there. */
2978    for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2979        entries++;
2980        if (entries > cachep->num || i >= cachep->num)
2981            goto bad;
2982    }
2983    if (entries != cachep->num - slabp->inuse) {
2984bad:
2985        printk(KERN_ERR "slab: Internal list corruption detected in "
2986                "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2987            cachep->name, cachep->num, slabp, slabp->inuse);
2988        for (i = 0;
2989             i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2990             i++) {
2991            if (i % 16 == 0)
2992                printk("\n%03x:", i);
2993            printk(" %02x", ((unsigned char *)slabp)[i]);
2994        }
2995        printk("\n");
2996        BUG();
2997    }
2998}
2999#else
3000#define kfree_debugcheck(x) do { } while(0)
3001#define cache_free_debugcheck(x,objp,z) (objp)
3002#define check_slabp(x,y) do { } while(0)
3003#endif
3004
3005static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3006{
3007    int batchcount;
3008    struct kmem_list3 *l3;
3009    struct array_cache *ac;
3010    int node;
3011
3012retry:
3013    check_irq_off();
3014    node = numa_mem_id();
3015    ac = cpu_cache_get(cachep);
3016    batchcount = ac->batchcount;
3017    if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3018        /*
3019         * If there was little recent activity on this cache, then
3020         * perform only a partial refill. Otherwise we could generate
3021         * refill bouncing.
3022         */
3023        batchcount = BATCHREFILL_LIMIT;
3024    }
3025    l3 = cachep->nodelists[node];
3026
3027    BUG_ON(ac->avail > 0 || !l3);
3028    spin_lock(&l3->list_lock);
3029
3030    /* See if we can refill from the shared array */
3031    if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3032        l3->shared->touched = 1;
3033        goto alloc_done;
3034    }
3035
3036    while (batchcount > 0) {
3037        struct list_head *entry;
3038        struct slab *slabp;
3039        /* Get slab alloc is to come from. */
3040        entry = l3->slabs_partial.next;
3041        if (entry == &l3->slabs_partial) {
3042            l3->free_touched = 1;
3043            entry = l3->slabs_free.next;
3044            if (entry == &l3->slabs_free)
3045                goto must_grow;
3046        }
3047
3048        slabp = list_entry(entry, struct slab, list);
3049        check_slabp(cachep, slabp);
3050        check_spinlock_acquired(cachep);
3051
3052        /*
3053         * The slab was either on partial or free list so
3054         * there must be at least one object available for
3055         * allocation.
3056         */
3057        BUG_ON(slabp->inuse >= cachep->num);
3058
3059        while (slabp->inuse < cachep->num && batchcount--) {
3060            STATS_INC_ALLOCED(cachep);
3061            STATS_INC_ACTIVE(cachep);
3062            STATS_SET_HIGH(cachep);
3063
3064            ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3065                                node);
3066        }
3067        check_slabp(cachep, slabp);
3068
3069        /* move slabp to correct slabp list: */
3070        list_del(&slabp->list);
3071        if (slabp->free == BUFCTL_END)
3072            list_add(&slabp->list, &l3->slabs_full);
3073        else
3074            list_add(&slabp->list, &l3->slabs_partial);
3075    }
3076
3077must_grow:
3078    l3->free_objects -= ac->avail;
3079alloc_done:
3080    spin_unlock(&l3->list_lock);
3081
3082    if (unlikely(!ac->avail)) {
3083        int x;
3084        x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3085
3086        /* cache_grow can reenable interrupts, then ac could change. */
3087        ac = cpu_cache_get(cachep);
3088        if (!x && ac->avail == 0) /* no objects in sight? abort */
3089            return NULL;
3090
3091        if (!ac->avail) /* objects refilled by interrupt? */
3092            goto retry;
3093    }
3094    ac->touched = 1;
3095    return ac->entry[--ac->avail];
3096}
3097
3098static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3099                        gfp_t flags)
3100{
3101    might_sleep_if(flags & __GFP_WAIT);
3102#if DEBUG
3103    kmem_flagcheck(cachep, flags);
3104#endif
3105}
3106
3107#if DEBUG
3108static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3109                gfp_t flags, void *objp, void *caller)
3110{
3111    if (!objp)
3112        return objp;
3113    if (cachep->flags & SLAB_POISON) {
3114#ifdef CONFIG_DEBUG_PAGEALLOC
3115        if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3116            kernel_map_pages(virt_to_page(objp),
3117                     cachep->buffer_size / PAGE_SIZE, 1);
3118        else
3119            check_poison_obj(cachep, objp);
3120#else
3121        check_poison_obj(cachep, objp);
3122#endif
3123        poison_obj(cachep, objp, POISON_INUSE);
3124    }
3125    if (cachep->flags & SLAB_STORE_USER)
3126        *dbg_userword(cachep, objp) = caller;
3127
3128    if (cachep->flags & SLAB_RED_ZONE) {
3129        if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3130                *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3131            slab_error(cachep, "double free, or memory outside"
3132                        " object was overwritten");
3133            printk(KERN_ERR
3134                "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3135                objp, *dbg_redzone1(cachep, objp),
3136                *dbg_redzone2(cachep, objp));
3137        }
3138        *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3139        *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3140    }
3141#ifdef CONFIG_DEBUG_SLAB_LEAK
3142    {
3143        struct slab *slabp;
3144        unsigned objnr;
3145
3146        slabp = page_get_slab(virt_to_head_page(objp));
3147        objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3148        slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3149    }
3150#endif
3151    objp += obj_offset(cachep);
3152    if (cachep->ctor && cachep->flags & SLAB_POISON)
3153        cachep->ctor(objp);
3154#if ARCH_SLAB_MINALIGN
3155    if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3156        printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3157               objp, ARCH_SLAB_MINALIGN);
3158    }
3159#endif
3160    return objp;
3161}
3162#else
3163#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3164#endif
3165
3166static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3167{
3168    if (cachep == &cache_cache)
3169        return false;
3170
3171    return should_failslab(obj_size(cachep), flags, cachep->flags);
3172}
3173
3174static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3175{
3176    void *objp;
3177    struct array_cache *ac;
3178
3179    check_irq_off();
3180
3181    ac = cpu_cache_get(cachep);
3182    if (likely(ac->avail)) {
3183        STATS_INC_ALLOCHIT(cachep);
3184        ac->touched = 1;
3185        objp = ac->entry[--ac->avail];
3186    } else {
3187        STATS_INC_ALLOCMISS(cachep);
3188        objp = cache_alloc_refill(cachep, flags);
3189        /*
3190         * the 'ac' may be updated by cache_alloc_refill(),
3191         * and kmemleak_erase() requires its correct value.
3192         */
3193        ac = cpu_cache_get(cachep);
3194    }
3195    /*
3196     * To avoid a false negative, if an object that is in one of the
3197     * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3198     * treat the array pointers as a reference to the object.
3199     */
3200    if (objp)
3201        kmemleak_erase(&ac->entry[ac->avail]);
3202    return objp;
3203}
3204
3205#ifdef CONFIG_NUMA
3206/*
3207 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3208 *
3209 * If we are in_interrupt, then process context, including cpusets and
3210 * mempolicy, may not apply and should not be used for allocation policy.
3211 */
3212static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3213{
3214    int nid_alloc, nid_here;
3215
3216    if (in_interrupt() || (flags & __GFP_THISNODE))
3217        return NULL;
3218    nid_alloc = nid_here = numa_mem_id();
3219    get_mems_allowed();
3220    if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3221        nid_alloc = cpuset_slab_spread_node();
3222    else if (current->mempolicy)
3223        nid_alloc = slab_node(current->mempolicy);
3224    put_mems_allowed();
3225    if (nid_alloc != nid_here)
3226        return ____cache_alloc_node(cachep, flags, nid_alloc);
3227    return NULL;
3228}
3229
3230/*
3231 * Fallback function if there was no memory available and no objects on a
3232 * certain node and fall back is permitted. First we scan all the
3233 * available nodelists for available objects. If that fails then we
3234 * perform an allocation without specifying a node. This allows the page
3235 * allocator to do its reclaim / fallback magic. We then insert the
3236 * slab into the proper nodelist and then allocate from it.
3237 */
3238static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3239{
3240    struct zonelist *zonelist;
3241    gfp_t local_flags;
3242    struct zoneref *z;
3243    struct zone *zone;
3244    enum zone_type high_zoneidx = gfp_zone(flags);
3245    void *obj = NULL;
3246    int nid;
3247
3248    if (flags & __GFP_THISNODE)
3249        return NULL;
3250
3251    get_mems_allowed();
3252    zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3253    local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3254
3255retry:
3256    /*
3257     * Look through allowed nodes for objects available
3258     * from existing per node queues.
3259     */
3260    for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3261        nid = zone_to_nid(zone);
3262
3263        if (cpuset_zone_allowed_hardwall(zone, flags) &&
3264            cache->nodelists[nid] &&
3265            cache->nodelists[nid]->free_objects) {
3266                obj = ____cache_alloc_node(cache,
3267                    flags | GFP_THISNODE, nid);
3268                if (obj)
3269                    break;
3270        }
3271    }
3272
3273    if (!obj) {
3274        /*
3275         * This allocation will be performed within the constraints
3276         * of the current cpuset / memory policy requirements.
3277         * We may trigger various forms of reclaim on the allowed
3278         * set and go into memory reserves if necessary.
3279         */
3280        if (local_flags & __GFP_WAIT)
3281            local_irq_enable();
3282        kmem_flagcheck(cache, flags);
3283        obj = kmem_getpages(cache, local_flags, numa_mem_id());
3284        if (local_flags & __GFP_WAIT)
3285            local_irq_disable();
3286        if (obj) {
3287            /*
3288             * Insert into the appropriate per node queues
3289             */
3290            nid = page_to_nid(virt_to_page(obj));
3291            if (cache_grow(cache, flags, nid, obj)) {
3292                obj = ____cache_alloc_node(cache,
3293                    flags | GFP_THISNODE, nid);
3294                if (!obj)
3295                    /*
3296                     * Another processor may allocate the
3297                     * objects in the slab since we are
3298                     * not holding any locks.
3299                     */
3300                    goto retry;
3301            } else {
3302                /* cache_grow already freed obj */
3303                obj = NULL;
3304            }
3305        }
3306    }
3307    put_mems_allowed();
3308    return obj;
3309}
3310
3311/*
3312 * A interface to enable slab creation on nodeid
3313 */
3314static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3315                int nodeid)
3316{
3317    struct list_head *entry;
3318    struct slab *slabp;
3319    struct kmem_list3 *l3;
3320    void *obj;
3321    int x;
3322
3323    l3 = cachep->nodelists[nodeid];
3324    BUG_ON(!l3);
3325
3326retry:
3327    check_irq_off();
3328    spin_lock(&l3->list_lock);
3329    entry = l3->slabs_partial.next;
3330    if (entry == &l3->slabs_partial) {
3331        l3->free_touched = 1;
3332        entry = l3->slabs_free.next;
3333        if (entry == &l3->slabs_free)
3334            goto must_grow;
3335    }
3336
3337    slabp = list_entry(entry, struct slab, list);
3338    check_spinlock_acquired_node(cachep, nodeid);
3339    check_slabp(cachep, slabp);
3340
3341    STATS_INC_NODEALLOCS(cachep);
3342    STATS_INC_ACTIVE(cachep);
3343    STATS_SET_HIGH(cachep);
3344
3345    BUG_ON(slabp->inuse == cachep->num);
3346
3347    obj = slab_get_obj(cachep, slabp, nodeid);
3348    check_slabp(cachep, slabp);
3349    l3->free_objects--;
3350    /* move slabp to correct slabp list: */
3351    list_del(&slabp->list);
3352
3353    if (slabp->free == BUFCTL_END)
3354        list_add(&slabp->list, &l3->slabs_full);
3355    else
3356        list_add(&slabp->list, &l3->slabs_partial);
3357
3358    spin_unlock(&l3->list_lock);
3359    goto done;
3360
3361must_grow:
3362    spin_unlock(&l3->list_lock);
3363    x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3364    if (x)
3365        goto retry;
3366
3367    return fallback_alloc(cachep, flags);
3368
3369done:
3370    return obj;
3371}
3372
3373/**
3374 * kmem_cache_alloc_node - Allocate an object on the specified node
3375 * @cachep: The cache to allocate from.
3376 * @flags: See kmalloc().
3377 * @nodeid: node number of the target node.
3378 * @caller: return address of caller, used for debug information
3379 *
3380 * Identical to kmem_cache_alloc but it will allocate memory on the given
3381 * node, which can improve the performance for cpu bound structures.
3382 *
3383 * Fallback to other node is possible if __GFP_THISNODE is not set.
3384 */
3385static __always_inline void *
3386__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3387           void *caller)
3388{
3389    unsigned long save_flags;
3390    void *ptr;
3391    int slab_node = numa_mem_id();
3392
3393    flags &= gfp_allowed_mask;
3394
3395    lockdep_trace_alloc(flags);
3396
3397    if (slab_should_failslab(cachep, flags))
3398        return NULL;
3399
3400    cache_alloc_debugcheck_before(cachep, flags);
3401    local_irq_save(save_flags);
3402
3403    if (nodeid == -1)
3404        nodeid = slab_node;
3405
3406    if (unlikely(!cachep->nodelists[nodeid])) {
3407        /* Node not bootstrapped yet */
3408        ptr = fallback_alloc(cachep, flags);
3409        goto out;
3410    }
3411
3412    if (nodeid == slab_node) {
3413        /*
3414         * Use the locally cached objects if possible.
3415         * However ____cache_alloc does not allow fallback
3416         * to other nodes. It may fail while we still have
3417         * objects on other nodes available.
3418         */
3419        ptr = ____cache_alloc(cachep, flags);
3420        if (ptr)
3421            goto out;
3422    }
3423    /* ___cache_alloc_node can fall back to other nodes */
3424    ptr = ____cache_alloc_node(cachep, flags, nodeid);
3425  out:
3426    local_irq_restore(save_flags);
3427    ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3428    kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3429                 flags);
3430
3431    if (likely(ptr))
3432        kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3433
3434    if (unlikely((flags & __GFP_ZERO) && ptr))
3435        memset(ptr, 0, obj_size(cachep));
3436
3437    return ptr;
3438}
3439
3440static __always_inline void *
3441__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3442{
3443    void *objp;
3444
3445    if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3446        objp = alternate_node_alloc(cache, flags);
3447        if (objp)
3448            goto out;
3449    }
3450    objp = ____cache_alloc(cache, flags);
3451
3452    /*
3453     * We may just have run out of memory on the local node.
3454     * ____cache_alloc_node() knows how to locate memory on other nodes
3455     */
3456    if (!objp)
3457        objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3458
3459  out:
3460    return objp;
3461}
3462#else
3463
3464static __always_inline void *
3465__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3466{
3467    return ____cache_alloc(cachep, flags);
3468}
3469
3470#endif /* CONFIG_NUMA */
3471
3472static __always_inline void *
3473__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3474{
3475    unsigned long save_flags;
3476    void *objp;
3477
3478    flags &= gfp_allowed_mask;
3479
3480    lockdep_trace_alloc(flags);
3481
3482    if (slab_should_failslab(cachep, flags))
3483        return NULL;
3484
3485    cache_alloc_debugcheck_before(cachep, flags);
3486    local_irq_save(save_flags);
3487    objp = __do_cache_alloc(cachep, flags);
3488    local_irq_restore(save_flags);
3489    objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3490    kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3491                 flags);
3492    prefetchw(objp);
3493
3494    if (likely(objp))
3495        kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3496
3497    if (unlikely((flags & __GFP_ZERO) && objp))
3498        memset(objp, 0, obj_size(cachep));
3499
3500    return objp;
3501}
3502
3503/*
3504 * Caller needs to acquire correct kmem_list's list_lock
3505 */
3506static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3507               int node)
3508{
3509    int i;
3510    struct kmem_list3 *l3;
3511
3512    for (i = 0; i < nr_objects; i++) {
3513        void *objp = objpp[i];
3514        struct slab *slabp;
3515
3516        slabp = virt_to_slab(objp);
3517        l3 = cachep->nodelists[node];
3518        list_del(&slabp->list);
3519        check_spinlock_acquired_node(cachep, node);
3520        check_slabp(cachep, slabp);
3521        slab_put_obj(cachep, slabp, objp, node);
3522        STATS_DEC_ACTIVE(cachep);
3523        l3->free_objects++;
3524        check_slabp(cachep, slabp);
3525
3526        /* fixup slab chains */
3527        if (slabp->inuse == 0) {
3528            if (l3->free_objects > l3->free_limit) {
3529                l3->free_objects -= cachep->num;
3530                /* No need to drop any previously held
3531                 * lock here, even if we have a off-slab slab
3532                 * descriptor it is guaranteed to come from
3533                 * a different cache, refer to comments before
3534                 * alloc_slabmgmt.
3535                 */
3536                slab_destroy(cachep, slabp);
3537            } else {
3538                list_add(&slabp->list, &l3->slabs_free);
3539            }
3540        } else {
3541            /* Unconditionally move a slab to the end of the
3542             * partial list on free - maximum time for the
3543             * other objects to be freed, too.
3544             */
3545            list_add_tail(&slabp->list, &l3->slabs_partial);
3546        }
3547    }
3548}
3549
3550static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3551{
3552    int batchcount;
3553    struct kmem_list3 *l3;
3554    int node = numa_mem_id();
3555
3556    batchcount = ac->batchcount;
3557#if DEBUG
3558    BUG_ON(!batchcount || batchcount > ac->avail);
3559#endif
3560    check_irq_off();
3561    l3 = cachep->nodelists[node];
3562    spin_lock(&l3->list_lock);
3563    if (l3->shared) {
3564        struct array_cache *shared_array = l3->shared;
3565        int max = shared_array->limit - shared_array->avail;
3566        if (max) {
3567            if (batchcount > max)
3568                batchcount = max;
3569            memcpy(&(shared_array->entry[shared_array->avail]),
3570                   ac->entry, sizeof(void *) * batchcount);
3571            shared_array->avail += batchcount;
3572            goto free_done;
3573        }
3574    }
3575
3576    free_block(cachep, ac->entry, batchcount, node);
3577free_done:
3578#if STATS
3579    {
3580        int i = 0;
3581        struct list_head *p;
3582
3583        p = l3->slabs_free.next;
3584        while (p != &(l3->slabs_free)) {
3585            struct slab *slabp;
3586
3587            slabp = list_entry(p, struct slab, list);
3588            BUG_ON(slabp->inuse);
3589
3590            i++;
3591            p = p->next;
3592        }
3593        STATS_SET_FREEABLE(cachep, i);
3594    }
3595#endif
3596    spin_unlock(&l3->list_lock);
3597    ac->avail -= batchcount;
3598    memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3599}
3600
3601/*
3602 * Release an obj back to its cache. If the obj has a constructed state, it must
3603 * be in this state _before_ it is released. Called with disabled ints.
3604 */
3605static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3606{
3607    struct array_cache *ac = cpu_cache_get(cachep);
3608
3609    check_irq_off();
3610    kmemleak_free_recursive(objp, cachep->flags);
3611    objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3612
3613    kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3614
3615    /*
3616     * Skip calling cache_free_alien() when the platform is not numa.
3617     * This will avoid cache misses that happen while accessing slabp (which
3618     * is per page memory reference) to get nodeid. Instead use a global
3619     * variable to skip the call, which is mostly likely to be present in
3620     * the cache.
3621     */
3622    if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3623        return;
3624
3625    if (likely(ac->avail < ac->limit)) {
3626        STATS_INC_FREEHIT(cachep);
3627        ac->entry[ac->avail++] = objp;
3628        return;
3629    } else {
3630        STATS_INC_FREEMISS(cachep);
3631        cache_flusharray(cachep, ac);
3632        ac->entry[ac->avail++] = objp;
3633    }
3634}
3635
3636/**
3637 * kmem_cache_alloc - Allocate an object
3638 * @cachep: The cache to allocate from.
3639 * @flags: See kmalloc().
3640 *
3641 * Allocate an object from this cache. The flags are only relevant
3642 * if the cache has no available objects.
3643 */
3644void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3645{
3646    void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3647
3648    trace_kmem_cache_alloc(_RET_IP_, ret,
3649                   obj_size(cachep), cachep->buffer_size, flags);
3650
3651    return ret;
3652}
3653EXPORT_SYMBOL(kmem_cache_alloc);
3654
3655#ifdef CONFIG_TRACING
3656void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3657{
3658    return __cache_alloc(cachep, flags, __builtin_return_address(0));
3659}
3660EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3661#endif
3662
3663/**
3664 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3665 * @cachep: the cache we're checking against
3666 * @ptr: pointer to validate
3667 *
3668 * This verifies that the untrusted pointer looks sane;
3669 * it is _not_ a guarantee that the pointer is actually
3670 * part of the slab cache in question, but it at least
3671 * validates that the pointer can be dereferenced and
3672 * looks half-way sane.
3673 *
3674 * Currently only used for dentry validation.
3675 */
3676int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3677{
3678    unsigned long size = cachep->buffer_size;
3679    struct page *page;
3680
3681    if (unlikely(!kern_ptr_validate(ptr, size)))
3682        goto out;
3683    page = virt_to_page(ptr);
3684    if (unlikely(!PageSlab(page)))
3685        goto out;
3686    if (unlikely(page_get_cache(page) != cachep))
3687        goto out;
3688    return 1;
3689out:
3690    return 0;
3691}
3692
3693#ifdef CONFIG_NUMA
3694void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3695{
3696    void *ret = __cache_alloc_node(cachep, flags, nodeid,
3697                       __builtin_return_address(0));
3698
3699    trace_kmem_cache_alloc_node(_RET_IP_, ret,
3700                    obj_size(cachep), cachep->buffer_size,
3701                    flags, nodeid);
3702
3703    return ret;
3704}
3705EXPORT_SYMBOL(kmem_cache_alloc_node);
3706
3707#ifdef CONFIG_TRACING
3708void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3709                    gfp_t flags,
3710                    int nodeid)
3711{
3712    return __cache_alloc_node(cachep, flags, nodeid,
3713                  __builtin_return_address(0));
3714}
3715EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3716#endif
3717
3718static __always_inline void *
3719__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3720{
3721    struct kmem_cache *cachep;
3722    void *ret;
3723
3724    cachep = kmem_find_general_cachep(size, flags);
3725    if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3726        return cachep;
3727    ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3728
3729    trace_kmalloc_node((unsigned long) caller, ret,
3730               size, cachep->buffer_size, flags, node);
3731
3732    return ret;
3733}
3734
3735#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3736void *__kmalloc_node(size_t size, gfp_t flags, int node)
3737{
3738    return __do_kmalloc_node(size, flags, node,
3739            __builtin_return_address(0));
3740}
3741EXPORT_SYMBOL(__kmalloc_node);
3742
3743void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3744        int node, unsigned long caller)
3745{
3746    return __do_kmalloc_node(size, flags, node, (void *)caller);
3747}
3748EXPORT_SYMBOL(__kmalloc_node_track_caller);
3749#else
3750void *__kmalloc_node(size_t size, gfp_t flags, int node)
3751{
3752    return __do_kmalloc_node(size, flags, node, NULL);
3753}
3754EXPORT_SYMBOL(__kmalloc_node);
3755#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3756#endif /* CONFIG_NUMA */
3757
3758/**
3759 * __do_kmalloc - allocate memory
3760 * @size: how many bytes of memory are required.
3761 * @flags: the type of memory to allocate (see kmalloc).
3762 * @caller: function caller for debug tracking of the caller
3763 */
3764static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3765                      void *caller)
3766{
3767    struct kmem_cache *cachep;
3768    void *ret;
3769
3770    /* If you want to save a few bytes .text space: replace
3771     * __ with kmem_.
3772     * Then kmalloc uses the uninlined functions instead of the inline
3773     * functions.
3774     */
3775    cachep = __find_general_cachep(size, flags);
3776    if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3777        return cachep;
3778    ret = __cache_alloc(cachep, flags, caller);
3779
3780    trace_kmalloc((unsigned long) caller, ret,
3781              size, cachep->buffer_size, flags);
3782
3783    return ret;
3784}
3785
3786
3787#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3788void *__kmalloc(size_t size, gfp_t flags)
3789{
3790    return __do_kmalloc(size, flags, __builtin_return_address(0));
3791}
3792EXPORT_SYMBOL(__kmalloc);
3793
3794void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3795{
3796    return __do_kmalloc(size, flags, (void *)caller);
3797}
3798EXPORT_SYMBOL(__kmalloc_track_caller);
3799
3800#else
3801void *__kmalloc(size_t size, gfp_t flags)
3802{
3803    return __do_kmalloc(size, flags, NULL);
3804}
3805EXPORT_SYMBOL(__kmalloc);
3806#endif
3807
3808/**
3809 * kmem_cache_free - Deallocate an object
3810 * @cachep: The cache the allocation was from.
3811 * @objp: The previously allocated object.
3812 *
3813 * Free an object which was previously allocated from this
3814 * cache.
3815 */
3816void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3817{
3818    unsigned long flags;
3819
3820    local_irq_save(flags);
3821    debug_check_no_locks_freed(objp, obj_size(cachep));
3822    if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3823        debug_check_no_obj_freed(objp, obj_size(cachep));
3824    __cache_free(cachep, objp);
3825    local_irq_restore(flags);
3826
3827    trace_kmem_cache_free(_RET_IP_, objp);
3828}
3829EXPORT_SYMBOL(kmem_cache_free);
3830
3831/**
3832 * kfree - free previously allocated memory
3833 * @objp: pointer returned by kmalloc.
3834 *
3835 * If @objp is NULL, no operation is performed.
3836 *
3837 * Don't free memory not originally allocated by kmalloc()
3838 * or you will run into trouble.
3839 */
3840void kfree(const void *objp)
3841{
3842    struct kmem_cache *c;
3843    unsigned long flags;
3844
3845    trace_kfree(_RET_IP_, objp);
3846
3847    if (unlikely(ZERO_OR_NULL_PTR(objp)))
3848        return;
3849    local_irq_save(flags);
3850    kfree_debugcheck(objp);
3851    c = virt_to_cache(objp);
3852    debug_check_no_locks_freed(objp, obj_size(c));
3853    debug_check_no_obj_freed(objp, obj_size(c));
3854    __cache_free(c, (void *)objp);
3855    local_irq_restore(flags);
3856}
3857EXPORT_SYMBOL(kfree);
3858
3859unsigned int kmem_cache_size(struct kmem_cache *cachep)
3860{
3861    return obj_size(cachep);
3862}
3863EXPORT_SYMBOL(kmem_cache_size);
3864
3865const char *kmem_cache_name(struct kmem_cache *cachep)
3866{
3867    return cachep->name;
3868}
3869EXPORT_SYMBOL_GPL(kmem_cache_name);
3870
3871/*
3872 * This initializes kmem_list3 or resizes various caches for all nodes.
3873 */
3874static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3875{
3876    int node;
3877    struct kmem_list3 *l3;
3878    struct array_cache *new_shared;
3879    struct array_cache **new_alien = NULL;
3880
3881    for_each_online_node(node) {
3882
3883                if (use_alien_caches) {
3884                        new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3885                        if (!new_alien)
3886                                goto fail;
3887                }
3888
3889        new_shared = NULL;
3890        if (cachep->shared) {
3891            new_shared = alloc_arraycache(node,
3892                cachep->shared*cachep->batchcount,
3893                    0xbaadf00d, gfp);
3894            if (!new_shared) {
3895                free_alien_cache(new_alien);
3896                goto fail;
3897            }
3898        }
3899
3900        l3 = cachep->nodelists[node];
3901        if (l3) {
3902            struct array_cache *shared = l3->shared;
3903
3904            spin_lock_irq(&l3->list_lock);
3905
3906            if (shared)
3907                free_block(cachep, shared->entry,
3908                        shared->avail, node);
3909
3910            l3->shared = new_shared;
3911            if (!l3->alien) {
3912                l3->alien = new_alien;
3913                new_alien = NULL;
3914            }
3915            l3->free_limit = (1 + nr_cpus_node(node)) *
3916                    cachep->batchcount + cachep->num;
3917            spin_unlock_irq(&l3->list_lock);
3918            kfree(shared);
3919            free_alien_cache(new_alien);
3920            continue;
3921        }
3922        l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3923        if (!l3) {
3924            free_alien_cache(new_alien);
3925            kfree(new_shared);
3926            goto fail;
3927        }
3928
3929        kmem_list3_init(l3);
3930        l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3931                ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3932        l3->shared = new_shared;
3933        l3->alien = new_alien;
3934        l3->free_limit = (1 + nr_cpus_node(node)) *
3935                    cachep->batchcount + cachep->num;
3936        cachep->nodelists[node] = l3;
3937    }
3938    return 0;
3939
3940fail:
3941    if (!cachep->next.next) {
3942        /* Cache is not active yet. Roll back what we did */
3943        node--;
3944        while (node >= 0) {
3945            if (cachep->nodelists[node]) {
3946                l3 = cachep->nodelists[node];
3947
3948                kfree(l3->shared);
3949                free_alien_cache(l3->alien);
3950                kfree(l3);
3951                cachep->nodelists[node] = NULL;
3952            }
3953            node--;
3954        }
3955    }
3956    return -ENOMEM;
3957}
3958
3959struct ccupdate_struct {
3960    struct kmem_cache *cachep;
3961    struct array_cache *new[NR_CPUS];
3962};
3963
3964static void do_ccupdate_local(void *info)
3965{
3966    struct ccupdate_struct *new = info;
3967    struct array_cache *old;
3968
3969    check_irq_off();
3970    old = cpu_cache_get(new->cachep);
3971
3972    new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3973    new->new[smp_processor_id()] = old;
3974}
3975
3976/* Always called with the cache_chain_mutex held */
3977static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3978                int batchcount, int shared, gfp_t gfp)
3979{
3980    struct ccupdate_struct *new;
3981    int i;
3982
3983    new = kzalloc(sizeof(*new), gfp);
3984    if (!new)
3985        return -ENOMEM;
3986
3987    for_each_online_cpu(i) {
3988        new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3989                        batchcount, gfp);
3990        if (!new->new[i]) {
3991            for (i--; i >= 0; i--)
3992                kfree(new->new[i]);
3993            kfree(new);
3994            return -ENOMEM;
3995        }
3996    }
3997    new->cachep = cachep;
3998
3999    on_each_cpu(do_ccupdate_local, (void *)new, 1);
4000
4001    check_irq_on();
4002    cachep->batchcount = batchcount;
4003    cachep->limit = limit;
4004    cachep->shared = shared;
4005
4006    for_each_online_cpu(i) {
4007        struct array_cache *ccold = new->new[i];
4008        if (!ccold)
4009            continue;
4010        spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4011        free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4012        spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4013        kfree(ccold);
4014    }
4015    kfree(new);
4016    return alloc_kmemlist(cachep, gfp);
4017}
4018
4019/* Called with cache_chain_mutex held always */
4020static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4021{
4022    int err;
4023    int limit, shared;
4024
4025    /*
4026     * The head array serves three purposes:
4027     * - create a LIFO ordering, i.e. return objects that are cache-warm
4028     * - reduce the number of spinlock operations.
4029     * - reduce the number of linked list operations on the slab and
4030     * bufctl chains: array operations are cheaper.
4031     * The numbers are guessed, we should auto-tune as described by
4032     * Bonwick.
4033     */
4034    if (cachep->buffer_size > 131072)
4035        limit = 1;
4036    else if (cachep->buffer_size > PAGE_SIZE)
4037        limit = 8;
4038    else if (cachep->buffer_size > 1024)
4039        limit = 24;
4040    else if (cachep->buffer_size > 256)
4041        limit = 54;
4042    else
4043        limit = 120;
4044
4045    /*
4046     * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4047     * allocation behaviour: Most allocs on one cpu, most free operations
4048     * on another cpu. For these cases, an efficient object passing between
4049     * cpus is necessary. This is provided by a shared array. The array
4050     * replaces Bonwick's magazine layer.
4051     * On uniprocessor, it's functionally equivalent (but less efficient)
4052     * to a larger limit. Thus disabled by default.
4053     */
4054    shared = 0;
4055    if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4056        shared = 8;
4057
4058#if DEBUG
4059    /*
4060     * With debugging enabled, large batchcount lead to excessively long
4061     * periods with disabled local interrupts. Limit the batchcount
4062     */
4063    if (limit > 32)
4064        limit = 32;
4065#endif
4066    err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4067    if (err)
4068        printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4069               cachep->name, -err);
4070    return err;
4071}
4072
4073/*
4074 * Drain an array if it contains any elements taking the l3 lock only if
4075 * necessary. Note that the l3 listlock also protects the array_cache
4076 * if drain_array() is used on the shared array.
4077 */
4078void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4079             struct array_cache *ac, int force, int node)
4080{
4081    int tofree;
4082
4083    if (!ac || !ac->avail)
4084        return;
4085    if (ac->touched && !force) {
4086        ac->touched = 0;
4087    } else {
4088        spin_lock_irq(&l3->list_lock);
4089        if (ac->avail) {
4090            tofree = force ? ac->avail : (ac->limit + 4) / 5;
4091            if (tofree > ac->avail)
4092                tofree = (ac->avail + 1) / 2;
4093            free_block(cachep, ac->entry, tofree, node);
4094            ac->avail -= tofree;
4095            memmove(ac->entry, &(ac->entry[tofree]),
4096                sizeof(void *) * ac->avail);
4097        }
4098        spin_unlock_irq(&l3->list_lock);
4099    }
4100}
4101
4102/**
4103 * cache_reap - Reclaim memory from caches.
4104 * @w: work descriptor
4105 *
4106 * Called from workqueue/eventd every few seconds.
4107 * Purpose:
4108 * - clear the per-cpu caches for this CPU.
4109 * - return freeable pages to the main free memory pool.
4110 *
4111 * If we cannot acquire the cache chain mutex then just give up - we'll try
4112 * again on the next iteration.
4113 */
4114static void cache_reap(struct work_struct *w)
4115{
4116    struct kmem_cache *searchp;
4117    struct kmem_list3 *l3;
4118    int node = numa_mem_id();
4119    struct delayed_work *work = to_delayed_work(w);
4120
4121    if (!mutex_trylock(&cache_chain_mutex))
4122        /* Give up. Setup the next iteration. */
4123        goto out;
4124
4125    list_for_each_entry(searchp, &cache_chain, next) {
4126        check_irq_on();
4127
4128        /*
4129         * We only take the l3 lock if absolutely necessary and we
4130         * have established with reasonable certainty that
4131         * we can do some work if the lock was obtained.
4132         */
4133        l3 = searchp->nodelists[node];
4134
4135        reap_alien(searchp, l3);
4136
4137        drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4138
4139        /*
4140         * These are racy checks but it does not matter
4141         * if we skip one check or scan twice.
4142         */
4143        if (time_after(l3->next_reap, jiffies))
4144            goto next;
4145
4146        l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4147
4148        drain_array(searchp, l3, l3->shared, 0, node);
4149
4150        if (l3->free_touched)
4151            l3->free_touched = 0;
4152        else {
4153            int freed;
4154
4155            freed = drain_freelist(searchp, l3, (l3->free_limit +
4156                5 * searchp->num - 1) / (5 * searchp->num));
4157            STATS_ADD_REAPED(searchp, freed);
4158        }
4159next:
4160        cond_resched();
4161    }
4162    check_irq_on();
4163    mutex_unlock(&cache_chain_mutex);
4164    next_reap_node();
4165out:
4166    /* Set up the next iteration */
4167    schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4168}
4169
4170#ifdef CONFIG_SLABINFO
4171
4172static void print_slabinfo_header(struct seq_file *m)
4173{
4174    /*
4175     * Output format version, so at least we can change it
4176     * without _too_ many complaints.
4177     */
4178#if STATS
4179    seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4180#else
4181    seq_puts(m, "slabinfo - version: 2.1\n");
4182#endif
4183    seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4184         "<objperslab> <pagesperslab>");
4185    seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4186    seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4187#if STATS
4188    seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4189         "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4190    seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4191#endif
4192    seq_putc(m, '\n');
4193}
4194
4195static void *s_start(struct seq_file *m, loff_t *pos)
4196{
4197    loff_t n = *pos;
4198
4199    mutex_lock(&cache_chain_mutex);
4200    if (!n)
4201        print_slabinfo_header(m);
4202
4203    return seq_list_start(&cache_chain, *pos);
4204}
4205
4206static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4207{
4208    return seq_list_next(p, &cache_chain, pos);
4209}
4210
4211static void s_stop(struct seq_file *m, void *p)
4212{
4213    mutex_unlock(&cache_chain_mutex);
4214}
4215
4216static int s_show(struct seq_file *m, void *p)
4217{
4218    struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4219    struct slab *slabp;
4220    unsigned long active_objs;
4221    unsigned long num_objs;
4222    unsigned long active_slabs = 0;
4223    unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4224    const char *name;
4225    char *error = NULL;
4226    int node;
4227    struct kmem_list3 *l3;
4228
4229    active_objs = 0;
4230    num_slabs = 0;
4231    for_each_online_node(node) {
4232        l3 = cachep->nodelists[node];
4233        if (!l3)
4234            continue;
4235
4236        check_irq_on();
4237        spin_lock_irq(&l3->list_lock);
4238
4239        list_for_each_entry(slabp, &l3->slabs_full, list) {
4240            if (slabp->inuse != cachep->num && !error)
4241                error = "slabs_full accounting error";
4242            active_objs += cachep->num;
4243            active_slabs++;
4244        }
4245        list_for_each_entry(slabp, &l3->slabs_partial, list) {
4246            if (slabp->inuse == cachep->num && !error)
4247                error = "slabs_partial inuse accounting error";
4248            if (!slabp->inuse && !error)
4249                error = "slabs_partial/inuse accounting error";
4250            active_objs += slabp->inuse;
4251            active_slabs++;
4252        }
4253        list_for_each_entry(slabp, &l3->slabs_free, list) {
4254            if (slabp->inuse && !error)
4255                error = "slabs_free/inuse accounting error";
4256            num_slabs++;
4257        }
4258        free_objects += l3->free_objects;
4259        if (l3->shared)
4260            shared_avail += l3->shared->avail;
4261
4262        spin_unlock_irq(&l3->list_lock);
4263    }
4264    num_slabs += active_slabs;
4265    num_objs = num_slabs * cachep->num;
4266    if (num_objs - active_objs != free_objects && !error)
4267        error = "free_objects accounting error";
4268
4269    name = cachep->name;
4270    if (error)
4271        printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4272
4273    seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4274           name, active_objs, num_objs, cachep->buffer_size,
4275           cachep->num, (1 << cachep->gfporder));
4276    seq_printf(m, " : tunables %4u %4u %4u",
4277           cachep->limit, cachep->batchcount, cachep->shared);
4278    seq_printf(m, " : slabdata %6lu %6lu %6lu",
4279           active_slabs, num_slabs, shared_avail);
4280#if STATS
4281    { /* list3 stats */
4282        unsigned long high = cachep->high_mark;
4283        unsigned long allocs = cachep->num_allocations;
4284        unsigned long grown = cachep->grown;
4285        unsigned long reaped = cachep->reaped;
4286        unsigned long errors = cachep->errors;
4287        unsigned long max_freeable = cachep->max_freeable;
4288        unsigned long node_allocs = cachep->node_allocs;
4289        unsigned long node_frees = cachep->node_frees;
4290        unsigned long overflows = cachep->node_overflow;
4291
4292        seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4293               "%4lu %4lu %4lu %4lu %4lu",
4294               allocs, high, grown,
4295               reaped, errors, max_freeable, node_allocs,
4296               node_frees, overflows);
4297    }
4298    /* cpu stats */
4299    {
4300        unsigned long allochit = atomic_read(&cachep->allochit);
4301        unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4302        unsigned long freehit = atomic_read(&cachep->freehit);
4303        unsigned long freemiss = atomic_read(&cachep->freemiss);
4304
4305        seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4306               allochit, allocmiss, freehit, freemiss);
4307    }
4308#endif
4309    seq_putc(m, '\n');
4310    return 0;
4311}
4312
4313/*
4314 * slabinfo_op - iterator that generates /proc/slabinfo
4315 *
4316 * Output layout:
4317 * cache-name
4318 * num-active-objs
4319 * total-objs
4320 * object size
4321 * num-active-slabs
4322 * total-slabs
4323 * num-pages-per-slab
4324 * + further values on SMP and with statistics enabled
4325 */
4326
4327static const struct seq_operations slabinfo_op = {
4328    .start = s_start,
4329    .next = s_next,
4330    .stop = s_stop,
4331    .show = s_show,
4332};
4333
4334#define MAX_SLABINFO_WRITE 128
4335/**
4336 * slabinfo_write - Tuning for the slab allocator
4337 * @file: unused
4338 * @buffer: user buffer
4339 * @count: data length
4340 * @ppos: unused
4341 */
4342ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4343               size_t count, loff_t *ppos)
4344{
4345    char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4346    int limit, batchcount, shared, res;
4347    struct kmem_cache *cachep;
4348
4349    if (count > MAX_SLABINFO_WRITE)
4350        return -EINVAL;
4351    if (copy_from_user(&kbuf, buffer, count))
4352        return -EFAULT;
4353    kbuf[MAX_SLABINFO_WRITE] = '\0';
4354
4355    tmp = strchr(kbuf, ' ');
4356    if (!tmp)
4357        return -EINVAL;
4358    *tmp = '\0';
4359    tmp++;
4360    if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4361        return -EINVAL;
4362
4363    /* Find the cache in the chain of caches. */
4364    mutex_lock(&cache_chain_mutex);
4365    res = -EINVAL;
4366    list_for_each_entry(cachep, &cache_chain, next) {
4367        if (!strcmp(cachep->name, kbuf)) {
4368            if (limit < 1 || batchcount < 1 ||
4369                    batchcount > limit || shared < 0) {
4370                res = 0;
4371            } else {
4372                res = do_tune_cpucache(cachep, limit,
4373                               batchcount, shared,
4374                               GFP_KERNEL);
4375            }
4376            break;
4377        }
4378    }
4379    mutex_unlock(&cache_chain_mutex);
4380    if (res >= 0)
4381        res = count;
4382    return res;
4383}
4384
4385static int slabinfo_open(struct inode *inode, struct file *file)
4386{
4387    return seq_open(file, &slabinfo_op);
4388}
4389
4390static const struct file_operations proc_slabinfo_operations = {
4391    .open = slabinfo_open,
4392    .read = seq_read,
4393    .write = slabinfo_write,
4394    .llseek = seq_lseek,
4395    .release = seq_release,
4396};
4397
4398#ifdef CONFIG_DEBUG_SLAB_LEAK
4399
4400static void *leaks_start(struct seq_file *m, loff_t *pos)
4401{
4402    mutex_lock(&cache_chain_mutex);
4403    return seq_list_start(&cache_chain, *pos);
4404}
4405
4406static inline int add_caller(unsigned long *n, unsigned long v)
4407{
4408    unsigned long *p;
4409    int l;
4410    if (!v)
4411        return 1;
4412    l = n[1];
4413    p = n + 2;
4414    while (l) {
4415        int i = l/2;
4416        unsigned long *q = p + 2 * i;
4417        if (*q == v) {
4418            q[1]++;
4419            return 1;
4420        }
4421        if (*q > v) {
4422            l = i;
4423        } else {
4424            p = q + 2;
4425            l -= i + 1;
4426        }
4427    }
4428    if (++n[1] == n[0])
4429        return 0;
4430    memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4431    p[0] = v;
4432    p[1] = 1;
4433    return 1;
4434}
4435
4436static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4437{
4438    void *p;
4439    int i;
4440    if (n[0] == n[1])
4441        return;
4442    for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4443        if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4444            continue;
4445        if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4446            return;
4447    }
4448}
4449
4450static void show_symbol(struct seq_file *m, unsigned long address)
4451{
4452#ifdef CONFIG_KALLSYMS
4453    unsigned long offset, size;
4454    char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4455
4456    if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4457        seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4458        if (modname[0])
4459            seq_printf(m, " [%s]", modname);
4460        return;
4461    }
4462#endif
4463    seq_printf(m, "%p", (void *)address);
4464}
4465
4466static int leaks_show(struct seq_file *m, void *p)
4467{
4468    struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4469    struct slab *slabp;
4470    struct kmem_list3 *l3;
4471    const char *name;
4472    unsigned long *n = m->private;
4473    int node;
4474    int i;
4475
4476    if (!(cachep->flags & SLAB_STORE_USER))
4477        return 0;
4478    if (!(cachep->flags & SLAB_RED_ZONE))
4479        return 0;
4480
4481    /* OK, we can do it */
4482
4483    n[1] = 0;
4484
4485    for_each_online_node(node) {
4486        l3 = cachep->nodelists[node];
4487        if (!l3)
4488            continue;
4489
4490        check_irq_on();
4491        spin_lock_irq(&l3->list_lock);
4492
4493        list_for_each_entry(slabp, &l3->slabs_full, list)
4494            handle_slab(n, cachep, slabp);
4495        list_for_each_entry(slabp, &l3->slabs_partial, list)
4496            handle_slab(n, cachep, slabp);
4497        spin_unlock_irq(&l3->list_lock);
4498    }
4499    name = cachep->name;
4500    if (n[0] == n[1]) {
4501        /* Increase the buffer size */
4502        mutex_unlock(&cache_chain_mutex);
4503        m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4504        if (!m->private) {
4505            /* Too bad, we are really out */
4506            m->private = n;
4507            mutex_lock(&cache_chain_mutex);
4508            return -ENOMEM;
4509        }
4510        *(unsigned long *)m->private = n[0] * 2;
4511        kfree(n);
4512        mutex_lock(&cache_chain_mutex);
4513        /* Now make sure this entry will be retried */
4514        m->count = m->size;
4515        return 0;
4516    }
4517    for (i = 0; i < n[1]; i++) {
4518        seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4519        show_symbol(m, n[2*i+2]);
4520        seq_putc(m, '\n');
4521    }
4522
4523    return 0;
4524}
4525
4526static const struct seq_operations slabstats_op = {
4527    .start = leaks_start,
4528    .next = s_next,
4529    .stop = s_stop,
4530    .show = leaks_show,
4531};
4532
4533static int slabstats_open(struct inode *inode, struct file *file)
4534{
4535    unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4536    int ret = -ENOMEM;
4537    if (n) {
4538        ret = seq_open(file, &slabstats_op);
4539        if (!ret) {
4540            struct seq_file *m = file->private_data;
4541            *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4542            m->private = n;
4543            n = NULL;
4544        }
4545        kfree(n);
4546    }
4547    return ret;
4548}
4549
4550static const struct file_operations proc_slabstats_operations = {
4551    .open = slabstats_open,
4552    .read = seq_read,
4553    .llseek = seq_lseek,
4554    .release = seq_release_private,
4555};
4556#endif
4557
4558static int __init slab_proc_init(void)
4559{
4560    proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4561#ifdef CONFIG_DEBUG_SLAB_LEAK
4562    proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4563#endif
4564    return 0;
4565}
4566module_init(slab_proc_init);
4567#endif
4568
4569/**
4570 * ksize - get the actual amount of memory allocated for a given object
4571 * @objp: Pointer to the object
4572 *
4573 * kmalloc may internally round up allocations and return more memory
4574 * than requested. ksize() can be used to determine the actual amount of
4575 * memory allocated. The caller may use this additional memory, even though
4576 * a smaller amount of memory was initially specified with the kmalloc call.
4577 * The caller must guarantee that objp points to a valid object previously
4578 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4579 * must not be freed during the duration of the call.
4580 */
4581size_t ksize(const void *objp)
4582{
4583    BUG_ON(!objp);
4584    if (unlikely(objp == ZERO_SIZE_PTR))
4585        return 0;
4586
4587    return obj_size(virt_to_cache(objp));
4588}
4589EXPORT_SYMBOL(ksize);
4590

Archive Download this file



interactive