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/kmemtrace.h>
106#include <linux/rcupdate.h>
107#include <linux/string.h>
108#include <linux/uaccess.h>
109#include <linux/nodemask.h>
110#include <linux/kmemleak.h>
111#include <linux/mempolicy.h>
112#include <linux/mutex.h>
113#include <linux/fault-inject.h>
114#include <linux/rtmutex.h>
115#include <linux/reciprocal_div.h>
116#include <linux/debugobjects.h>
117#include <linux/kmemcheck.h>
118#include <linux/memory.h>
119
120#include <asm/cacheflush.h>
121#include <asm/tlbflush.h>
122#include <asm/page.h>
123
124/*
125 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
126 * 0 for faster, smaller code (especially in the critical paths).
127 *
128 * STATS - 1 to collect stats for /proc/slabinfo.
129 * 0 for faster, smaller code (especially in the critical paths).
130 *
131 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
132 */
133
134#ifdef CONFIG_DEBUG_SLAB
135#define DEBUG 1
136#define STATS 1
137#define FORCED_DEBUG 1
138#else
139#define DEBUG 0
140#define STATS 0
141#define FORCED_DEBUG 0
142#endif
143
144/* Shouldn't this be in a header file somewhere? */
145#define BYTES_PER_WORD sizeof(void *)
146#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
147
148#ifndef ARCH_KMALLOC_FLAGS
149#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
150#endif
151
152/* Legal flag mask for kmem_cache_create(). */
153#if DEBUG
154# define CREATE_MASK (SLAB_RED_ZONE | \
155             SLAB_POISON | SLAB_HWCACHE_ALIGN | \
156             SLAB_CACHE_DMA | \
157             SLAB_STORE_USER | \
158             SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
159             SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
160             SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
161#else
162# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
163             SLAB_CACHE_DMA | \
164             SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
165             SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
166             SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
167#endif
168
169/*
170 * kmem_bufctl_t:
171 *
172 * Bufctl's are used for linking objs within a slab
173 * linked offsets.
174 *
175 * This implementation relies on "struct page" for locating the cache &
176 * slab an object belongs to.
177 * This allows the bufctl structure to be small (one int), but limits
178 * the number of objects a slab (not a cache) can contain when off-slab
179 * bufctls are used. The limit is the size of the largest general cache
180 * that does not use off-slab slabs.
181 * For 32bit archs with 4 kB pages, is this 56.
182 * This is not serious, as it is only for large objects, when it is unwise
183 * to have too many per slab.
184 * Note: This limit can be raised by introducing a general cache whose size
185 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
186 */
187
188typedef unsigned int kmem_bufctl_t;
189#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
190#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
191#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
192#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
193
194/*
195 * struct slab
196 *
197 * Manages the objs in a slab. Placed either at the beginning of mem allocated
198 * for a slab, or allocated from an general cache.
199 * Slabs are chained into three list: fully used, partial, fully free slabs.
200 */
201struct slab {
202    struct list_head list;
203    unsigned long colouroff;
204    void *s_mem; /* including colour offset */
205    unsigned int inuse; /* num of objs active in slab */
206    kmem_bufctl_t free;
207    unsigned short nodeid;
208};
209
210/*
211 * struct slab_rcu
212 *
213 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
214 * arrange for kmem_freepages to be called via RCU. This is useful if
215 * we need to approach a kernel structure obliquely, from its address
216 * obtained without the usual locking. We can lock the structure to
217 * stabilize it and check it's still at the given address, only if we
218 * can be sure that the memory has not been meanwhile reused for some
219 * other kind of object (which our subsystem's lock might corrupt).
220 *
221 * rcu_read_lock before reading the address, then rcu_read_unlock after
222 * taking the spinlock within the structure expected at that address.
223 *
224 * We assume struct slab_rcu can overlay struct slab when destroying.
225 */
226struct slab_rcu {
227    struct rcu_head head;
228    struct kmem_cache *cachep;
229    void *addr;
230};
231
232/*
233 * struct array_cache
234 *
235 * Purpose:
236 * - LIFO ordering, to hand out cache-warm objects from _alloc
237 * - reduce the number of linked list operations
238 * - reduce spinlock operations
239 *
240 * The limit is stored in the per-cpu structure to reduce the data cache
241 * footprint.
242 *
243 */
244struct array_cache {
245    unsigned int avail;
246    unsigned int limit;
247    unsigned int batchcount;
248    unsigned int touched;
249    spinlock_t lock;
250    void *entry[]; /*
251             * Must have this definition in here for the proper
252             * alignment of array_cache. Also simplifies accessing
253             * the entries.
254             */
255};
256
257/*
258 * bootstrap: The caches do not work without cpuarrays anymore, but the
259 * cpuarrays are allocated from the generic caches...
260 */
261#define BOOT_CPUCACHE_ENTRIES 1
262struct arraycache_init {
263    struct array_cache cache;
264    void *entries[BOOT_CPUCACHE_ENTRIES];
265};
266
267/*
268 * The slab lists for all objects.
269 */
270struct kmem_list3 {
271    struct list_head slabs_partial; /* partial list first, better asm code */
272    struct list_head slabs_full;
273    struct list_head slabs_free;
274    unsigned long free_objects;
275    unsigned int free_limit;
276    unsigned int colour_next; /* Per-node cache coloring */
277    spinlock_t list_lock;
278    struct array_cache *shared; /* shared per node */
279    struct array_cache **alien; /* on other nodes */
280    unsigned long next_reap; /* updated without locking */
281    int free_touched; /* updated without locking */
282};
283
284/*
285 * Need this for bootstrapping a per node allocator.
286 */
287#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
288struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
289#define CACHE_CACHE 0
290#define SIZE_AC MAX_NUMNODES
291#define SIZE_L3 (2 * MAX_NUMNODES)
292
293static int drain_freelist(struct kmem_cache *cache,
294            struct kmem_list3 *l3, int tofree);
295static void free_block(struct kmem_cache *cachep, void **objpp, int len,
296            int node);
297static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
298static void cache_reap(struct work_struct *unused);
299
300/*
301 * This function must be completely optimized away if a constant is passed to
302 * it. Mostly the same as what is in linux/slab.h except it returns an index.
303 */
304static __always_inline int index_of(const size_t size)
305{
306    extern void __bad_size(void);
307
308    if (__builtin_constant_p(size)) {
309        int i = 0;
310
311#define CACHE(x) \
312    if (size <=x) \
313        return i; \
314    else \
315        i++;
316#include <linux/kmalloc_sizes.h>
317#undef CACHE
318        __bad_size();
319    } else
320        __bad_size();
321    return 0;
322}
323
324static int slab_early_init = 1;
325
326#define INDEX_AC index_of(sizeof(struct arraycache_init))
327#define INDEX_L3 index_of(sizeof(struct kmem_list3))
328
329static void kmem_list3_init(struct kmem_list3 *parent)
330{
331    INIT_LIST_HEAD(&parent->slabs_full);
332    INIT_LIST_HEAD(&parent->slabs_partial);
333    INIT_LIST_HEAD(&parent->slabs_free);
334    parent->shared = NULL;
335    parent->alien = NULL;
336    parent->colour_next = 0;
337    spin_lock_init(&parent->list_lock);
338    parent->free_objects = 0;
339    parent->free_touched = 0;
340}
341
342#define MAKE_LIST(cachep, listp, slab, nodeid) \
343    do { \
344        INIT_LIST_HEAD(listp); \
345        list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
346    } while (0)
347
348#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
349    do { \
350    MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
351    MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
352    MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
353    } while (0)
354
355#define CFLGS_OFF_SLAB (0x80000000UL)
356#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
357
358#define BATCHREFILL_LIMIT 16
359/*
360 * Optimization question: fewer reaps means less probability for unnessary
361 * cpucache drain/refill cycles.
362 *
363 * OTOH the cpuarrays can contain lots of objects,
364 * which could lock up otherwise freeable slabs.
365 */
366#define REAPTIMEOUT_CPUC (2*HZ)
367#define REAPTIMEOUT_LIST3 (4*HZ)
368
369#if STATS
370#define STATS_INC_ACTIVE(x) ((x)->num_active++)
371#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
372#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
373#define STATS_INC_GROWN(x) ((x)->grown++)
374#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
375#define STATS_SET_HIGH(x) \
376    do { \
377        if ((x)->num_active > (x)->high_mark) \
378            (x)->high_mark = (x)->num_active; \
379    } while (0)
380#define STATS_INC_ERR(x) ((x)->errors++)
381#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
382#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
383#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
384#define STATS_SET_FREEABLE(x, i) \
385    do { \
386        if ((x)->max_freeable < i) \
387            (x)->max_freeable = i; \
388    } while (0)
389#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
390#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
391#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
392#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
393#else
394#define STATS_INC_ACTIVE(x) do { } while (0)
395#define STATS_DEC_ACTIVE(x) do { } while (0)
396#define STATS_INC_ALLOCED(x) do { } while (0)
397#define STATS_INC_GROWN(x) do { } while (0)
398#define STATS_ADD_REAPED(x,y) do { } while (0)
399#define STATS_SET_HIGH(x) do { } while (0)
400#define STATS_INC_ERR(x) do { } while (0)
401#define STATS_INC_NODEALLOCS(x) do { } while (0)
402#define STATS_INC_NODEFREES(x) do { } while (0)
403#define STATS_INC_ACOVERFLOW(x) do { } while (0)
404#define STATS_SET_FREEABLE(x, i) do { } while (0)
405#define STATS_INC_ALLOCHIT(x) do { } while (0)
406#define STATS_INC_ALLOCMISS(x) do { } while (0)
407#define STATS_INC_FREEHIT(x) do { } while (0)
408#define STATS_INC_FREEMISS(x) do { } while (0)
409#endif
410
411#if DEBUG
412
413/*
414 * memory layout of objects:
415 * 0 : objp
416 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
417 * the end of an object is aligned with the end of the real
418 * allocation. Catches writes behind the end of the allocation.
419 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
420 * redzone word.
421 * cachep->obj_offset: The real object.
422 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
423 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
424 * [BYTES_PER_WORD long]
425 */
426static int obj_offset(struct kmem_cache *cachep)
427{
428    return cachep->obj_offset;
429}
430
431static int obj_size(struct kmem_cache *cachep)
432{
433    return cachep->obj_size;
434}
435
436static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
437{
438    BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
439    return (unsigned long long*) (objp + obj_offset(cachep) -
440                      sizeof(unsigned long long));
441}
442
443static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
444{
445    BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
446    if (cachep->flags & SLAB_STORE_USER)
447        return (unsigned long long *)(objp + cachep->buffer_size -
448                          sizeof(unsigned long long) -
449                          REDZONE_ALIGN);
450    return (unsigned long long *) (objp + cachep->buffer_size -
451                       sizeof(unsigned long long));
452}
453
454static void **dbg_userword(struct kmem_cache *cachep, void *objp)
455{
456    BUG_ON(!(cachep->flags & SLAB_STORE_USER));
457    return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
458}
459
460#else
461
462#define obj_offset(x) 0
463#define obj_size(cachep) (cachep->buffer_size)
464#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
465#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
466#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
467
468#endif
469
470#ifdef CONFIG_TRACING
471size_t slab_buffer_size(struct kmem_cache *cachep)
472{
473    return cachep->buffer_size;
474}
475EXPORT_SYMBOL(slab_buffer_size);
476#endif
477
478/*
479 * Do not go above this order unless 0 objects fit into the slab.
480 */
481#define BREAK_GFP_ORDER_HI 1
482#define BREAK_GFP_ORDER_LO 0
483static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
484
485/*
486 * Functions for storing/retrieving the cachep and or slab from the page
487 * allocator. These are used to find the slab an obj belongs to. With kfree(),
488 * these are used to find the cache which an obj belongs to.
489 */
490static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
491{
492    page->lru.next = (struct list_head *)cache;
493}
494
495static inline struct kmem_cache *page_get_cache(struct page *page)
496{
497    page = compound_head(page);
498    BUG_ON(!PageSlab(page));
499    return (struct kmem_cache *)page->lru.next;
500}
501
502static inline void page_set_slab(struct page *page, struct slab *slab)
503{
504    page->lru.prev = (struct list_head *)slab;
505}
506
507static inline struct slab *page_get_slab(struct page *page)
508{
509    BUG_ON(!PageSlab(page));
510    return (struct slab *)page->lru.prev;
511}
512
513static inline struct kmem_cache *virt_to_cache(const void *obj)
514{
515    struct page *page = virt_to_head_page(obj);
516    return page_get_cache(page);
517}
518
519static inline struct slab *virt_to_slab(const void *obj)
520{
521    struct page *page = virt_to_head_page(obj);
522    return page_get_slab(page);
523}
524
525static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
526                 unsigned int idx)
527{
528    return slab->s_mem + cache->buffer_size * idx;
529}
530
531/*
532 * We want to avoid an expensive divide : (offset / cache->buffer_size)
533 * Using the fact that buffer_size is a constant for a particular cache,
534 * we can replace (offset / cache->buffer_size) by
535 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
536 */
537static inline unsigned int obj_to_index(const struct kmem_cache *cache,
538                    const struct slab *slab, void *obj)
539{
540    u32 offset = (obj - slab->s_mem);
541    return reciprocal_divide(offset, cache->reciprocal_buffer_size);
542}
543
544/*
545 * These are the default caches for kmalloc. Custom caches can have other sizes.
546 */
547struct cache_sizes malloc_sizes[] = {
548#define CACHE(x) { .cs_size = (x) },
549#include <linux/kmalloc_sizes.h>
550    CACHE(ULONG_MAX)
551#undef CACHE
552};
553EXPORT_SYMBOL(malloc_sizes);
554
555/* Must match cache_sizes above. Out of line to keep cache footprint low. */
556struct cache_names {
557    char *name;
558    char *name_dma;
559};
560
561static struct cache_names __initdata cache_names[] = {
562#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
563#include <linux/kmalloc_sizes.h>
564    {NULL,}
565#undef CACHE
566};
567
568static struct arraycache_init initarray_cache __initdata =
569    { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
570static struct arraycache_init initarray_generic =
571    { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
572
573/* internal cache of cache description objs */
574static struct kmem_cache cache_cache = {
575    .batchcount = 1,
576    .limit = BOOT_CPUCACHE_ENTRIES,
577    .shared = 1,
578    .buffer_size = sizeof(struct kmem_cache),
579    .name = "kmem_cache",
580};
581
582#define BAD_ALIEN_MAGIC 0x01020304ul
583
584/*
585 * chicken and egg problem: delay the per-cpu array allocation
586 * until the general caches are up.
587 */
588static enum {
589    NONE,
590    PARTIAL_AC,
591    PARTIAL_L3,
592    EARLY,
593    FULL
594} g_cpucache_up;
595
596/*
597 * used by boot code to determine if it can use slab based allocator
598 */
599int slab_is_available(void)
600{
601    return g_cpucache_up >= EARLY;
602}
603
604#ifdef CONFIG_LOCKDEP
605
606/*
607 * Slab sometimes uses the kmalloc slabs to store the slab headers
608 * for other slabs "off slab".
609 * The locking for this is tricky in that it nests within the locks
610 * of all other slabs in a few places; to deal with this special
611 * locking we put on-slab caches into a separate lock-class.
612 *
613 * We set lock class for alien array caches which are up during init.
614 * The lock annotation will be lost if all cpus of a node goes down and
615 * then comes back up during hotplug
616 */
617static struct lock_class_key on_slab_l3_key;
618static struct lock_class_key on_slab_alc_key;
619
620static void init_node_lock_keys(int q)
621{
622    struct cache_sizes *s = malloc_sizes;
623
624    if (g_cpucache_up != FULL)
625        return;
626
627    for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
628        struct array_cache **alc;
629        struct kmem_list3 *l3;
630        int r;
631
632        l3 = s->cs_cachep->nodelists[q];
633        if (!l3 || OFF_SLAB(s->cs_cachep))
634            continue;
635        lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
636        alc = l3->alien;
637        /*
638         * FIXME: This check for BAD_ALIEN_MAGIC
639         * should go away when common slab code is taught to
640         * work even without alien caches.
641         * Currently, non NUMA code returns BAD_ALIEN_MAGIC
642         * for alloc_alien_cache,
643         */
644        if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
645            continue;
646        for_each_node(r) {
647            if (alc[r])
648                lockdep_set_class(&alc[r]->lock,
649                    &on_slab_alc_key);
650        }
651    }
652}
653
654static inline void init_lock_keys(void)
655{
656    int node;
657
658    for_each_node(node)
659        init_node_lock_keys(node);
660}
661#else
662static void init_node_lock_keys(int q)
663{
664}
665
666static inline void init_lock_keys(void)
667{
668}
669#endif
670
671/*
672 * Guard access to the cache-chain.
673 */
674static DEFINE_MUTEX(cache_chain_mutex);
675static struct list_head cache_chain;
676
677static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
678
679static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
680{
681    return cachep->array[smp_processor_id()];
682}
683
684static inline struct kmem_cache *__find_general_cachep(size_t size,
685                            gfp_t gfpflags)
686{
687    struct cache_sizes *csizep = malloc_sizes;
688
689#if DEBUG
690    /* This happens if someone tries to call
691     * kmem_cache_create(), or __kmalloc(), before
692     * the generic caches are initialized.
693     */
694    BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
695#endif
696    if (!size)
697        return ZERO_SIZE_PTR;
698
699    while (size > csizep->cs_size)
700        csizep++;
701
702    /*
703     * Really subtle: The last entry with cs->cs_size==ULONG_MAX
704     * has cs_{dma,}cachep==NULL. Thus no special case
705     * for large kmalloc calls required.
706     */
707#ifdef CONFIG_ZONE_DMA
708    if (unlikely(gfpflags & GFP_DMA))
709        return csizep->cs_dmacachep;
710#endif
711    return csizep->cs_cachep;
712}
713
714static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
715{
716    return __find_general_cachep(size, gfpflags);
717}
718
719static size_t slab_mgmt_size(size_t nr_objs, size_t align)
720{
721    return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
722}
723
724/*
725 * Calculate the number of objects and left-over bytes for a given buffer size.
726 */
727static void cache_estimate(unsigned long gfporder, size_t buffer_size,
728               size_t align, int flags, size_t *left_over,
729               unsigned int *num)
730{
731    int nr_objs;
732    size_t mgmt_size;
733    size_t slab_size = PAGE_SIZE << gfporder;
734
735    /*
736     * The slab management structure can be either off the slab or
737     * on it. For the latter case, the memory allocated for a
738     * slab is used for:
739     *
740     * - The struct slab
741     * - One kmem_bufctl_t for each object
742     * - Padding to respect alignment of @align
743     * - @buffer_size bytes for each object
744     *
745     * If the slab management structure is off the slab, then the
746     * alignment will already be calculated into the size. Because
747     * the slabs are all pages aligned, the objects will be at the
748     * correct alignment when allocated.
749     */
750    if (flags & CFLGS_OFF_SLAB) {
751        mgmt_size = 0;
752        nr_objs = slab_size / buffer_size;
753
754        if (nr_objs > SLAB_LIMIT)
755            nr_objs = SLAB_LIMIT;
756    } else {
757        /*
758         * Ignore padding for the initial guess. The padding
759         * is at most @align-1 bytes, and @buffer_size is at
760         * least @align. In the worst case, this result will
761         * be one greater than the number of objects that fit
762         * into the memory allocation when taking the padding
763         * into account.
764         */
765        nr_objs = (slab_size - sizeof(struct slab)) /
766              (buffer_size + sizeof(kmem_bufctl_t));
767
768        /*
769         * This calculated number will be either the right
770         * amount, or one greater than what we want.
771         */
772        if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
773               > slab_size)
774            nr_objs--;
775
776        if (nr_objs > SLAB_LIMIT)
777            nr_objs = SLAB_LIMIT;
778
779        mgmt_size = slab_mgmt_size(nr_objs, align);
780    }
781    *num = nr_objs;
782    *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
783}
784
785#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
786
787static void __slab_error(const char *function, struct kmem_cache *cachep,
788            char *msg)
789{
790    printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
791           function, cachep->name, msg);
792    dump_stack();
793}
794
795/*
796 * By default on NUMA we use alien caches to stage the freeing of
797 * objects allocated from other nodes. This causes massive memory
798 * inefficiencies when using fake NUMA setup to split memory into a
799 * large number of small nodes, so it can be disabled on the command
800 * line
801  */
802
803static int use_alien_caches __read_mostly = 1;
804static int __init noaliencache_setup(char *s)
805{
806    use_alien_caches = 0;
807    return 1;
808}
809__setup("noaliencache", noaliencache_setup);
810
811#ifdef CONFIG_NUMA
812/*
813 * Special reaping functions for NUMA systems called from cache_reap().
814 * These take care of doing round robin flushing of alien caches (containing
815 * objects freed on different nodes from which they were allocated) and the
816 * flushing of remote pcps by calling drain_node_pages.
817 */
818static DEFINE_PER_CPU(unsigned long, slab_reap_node);
819
820static void init_reap_node(int cpu)
821{
822    int node;
823
824    node = next_node(cpu_to_mem(cpu), node_online_map);
825    if (node == MAX_NUMNODES)
826        node = first_node(node_online_map);
827
828    per_cpu(slab_reap_node, cpu) = node;
829}
830
831static void next_reap_node(void)
832{
833    int node = __get_cpu_var(slab_reap_node);
834
835    node = next_node(node, node_online_map);
836    if (unlikely(node >= MAX_NUMNODES))
837        node = first_node(node_online_map);
838    __get_cpu_var(slab_reap_node) = node;
839}
840
841#else
842#define init_reap_node(cpu) do { } while (0)
843#define next_reap_node(void) do { } while (0)
844#endif
845
846/*
847 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
848 * via the workqueue/eventd.
849 * Add the CPU number into the expiration time to minimize the possibility of
850 * the CPUs getting into lockstep and contending for the global cache chain
851 * lock.
852 */
853static void __cpuinit start_cpu_timer(int cpu)
854{
855    struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
856
857    /*
858     * When this gets called from do_initcalls via cpucache_init(),
859     * init_workqueues() has already run, so keventd will be setup
860     * at that time.
861     */
862    if (keventd_up() && reap_work->work.func == NULL) {
863        init_reap_node(cpu);
864        INIT_DELAYED_WORK(reap_work, cache_reap);
865        schedule_delayed_work_on(cpu, reap_work,
866                    __round_jiffies_relative(HZ, cpu));
867    }
868}
869
870static struct array_cache *alloc_arraycache(int node, int entries,
871                        int batchcount, gfp_t gfp)
872{
873    int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
874    struct array_cache *nc = NULL;
875
876    nc = kmalloc_node(memsize, gfp, node);
877    /*
878     * The array_cache structures contain pointers to free object.
879     * However, when such objects are allocated or transfered to another
880     * cache the pointers are not cleared and they could be counted as
881     * valid references during a kmemleak scan. Therefore, kmemleak must
882     * not scan such objects.
883     */
884    kmemleak_no_scan(nc);
885    if (nc) {
886        nc->avail = 0;
887        nc->limit = entries;
888        nc->batchcount = batchcount;
889        nc->touched = 0;
890        spin_lock_init(&nc->lock);
891    }
892    return nc;
893}
894
895/*
896 * Transfer objects in one arraycache to another.
897 * Locking must be handled by the caller.
898 *
899 * Return the number of entries transferred.
900 */
901static int transfer_objects(struct array_cache *to,
902        struct array_cache *from, unsigned int max)
903{
904    /* Figure out how many entries to transfer */
905    int nr = min(min(from->avail, max), to->limit - to->avail);
906
907    if (!nr)
908        return 0;
909
910    memcpy(to->entry + to->avail, from->entry + from->avail -nr,
911            sizeof(void *) *nr);
912
913    from->avail -= nr;
914    to->avail += nr;
915    return nr;
916}
917
918#ifndef CONFIG_NUMA
919
920#define drain_alien_cache(cachep, alien) do { } while (0)
921#define reap_alien(cachep, l3) do { } while (0)
922
923static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
924{
925    return (struct array_cache **)BAD_ALIEN_MAGIC;
926}
927
928static inline void free_alien_cache(struct array_cache **ac_ptr)
929{
930}
931
932static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
933{
934    return 0;
935}
936
937static inline void *alternate_node_alloc(struct kmem_cache *cachep,
938        gfp_t flags)
939{
940    return NULL;
941}
942
943static inline void *____cache_alloc_node(struct kmem_cache *cachep,
944         gfp_t flags, int nodeid)
945{
946    return NULL;
947}
948
949#else /* CONFIG_NUMA */
950
951static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
952static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
953
954static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
955{
956    struct array_cache **ac_ptr;
957    int memsize = sizeof(void *) * nr_node_ids;
958    int i;
959
960    if (limit > 1)
961        limit = 12;
962    ac_ptr = kzalloc_node(memsize, gfp, node);
963    if (ac_ptr) {
964        for_each_node(i) {
965            if (i == node || !node_online(i))
966                continue;
967            ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
968            if (!ac_ptr[i]) {
969                for (i--; i >= 0; i--)
970                    kfree(ac_ptr[i]);
971                kfree(ac_ptr);
972                return NULL;
973            }
974        }
975    }
976    return ac_ptr;
977}
978
979static void free_alien_cache(struct array_cache **ac_ptr)
980{
981    int i;
982
983    if (!ac_ptr)
984        return;
985    for_each_node(i)
986        kfree(ac_ptr[i]);
987    kfree(ac_ptr);
988}
989
990static void __drain_alien_cache(struct kmem_cache *cachep,
991                struct array_cache *ac, int node)
992{
993    struct kmem_list3 *rl3 = cachep->nodelists[node];
994
995    if (ac->avail) {
996        spin_lock(&rl3->list_lock);
997        /*
998         * Stuff objects into the remote nodes shared array first.
999         * That way we could avoid the overhead of putting the objects
1000         * into the free lists and getting them back later.
1001         */
1002        if (rl3->shared)
1003            transfer_objects(rl3->shared, ac, ac->limit);
1004
1005        free_block(cachep, ac->entry, ac->avail, node);
1006        ac->avail = 0;
1007        spin_unlock(&rl3->list_lock);
1008    }
1009}
1010
1011/*
1012 * Called from cache_reap() to regularly drain alien caches round robin.
1013 */
1014static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1015{
1016    int node = __get_cpu_var(slab_reap_node);
1017
1018    if (l3->alien) {
1019        struct array_cache *ac = l3->alien[node];
1020
1021        if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1022            __drain_alien_cache(cachep, ac, node);
1023            spin_unlock_irq(&ac->lock);
1024        }
1025    }
1026}
1027
1028static void drain_alien_cache(struct kmem_cache *cachep,
1029                struct array_cache **alien)
1030{
1031    int i = 0;
1032    struct array_cache *ac;
1033    unsigned long flags;
1034
1035    for_each_online_node(i) {
1036        ac = alien[i];
1037        if (ac) {
1038            spin_lock_irqsave(&ac->lock, flags);
1039            __drain_alien_cache(cachep, ac, i);
1040            spin_unlock_irqrestore(&ac->lock, flags);
1041        }
1042    }
1043}
1044
1045static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1046{
1047    struct slab *slabp = virt_to_slab(objp);
1048    int nodeid = slabp->nodeid;
1049    struct kmem_list3 *l3;
1050    struct array_cache *alien = NULL;
1051    int node;
1052
1053    node = numa_mem_id();
1054
1055    /*
1056     * Make sure we are not freeing a object from another node to the array
1057     * cache on this cpu.
1058     */
1059    if (likely(slabp->nodeid == node))
1060        return 0;
1061
1062    l3 = cachep->nodelists[node];
1063    STATS_INC_NODEFREES(cachep);
1064    if (l3->alien && l3->alien[nodeid]) {
1065        alien = l3->alien[nodeid];
1066        spin_lock(&alien->lock);
1067        if (unlikely(alien->avail == alien->limit)) {
1068            STATS_INC_ACOVERFLOW(cachep);
1069            __drain_alien_cache(cachep, alien, nodeid);
1070        }
1071        alien->entry[alien->avail++] = objp;
1072        spin_unlock(&alien->lock);
1073    } else {
1074        spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1075        free_block(cachep, &objp, 1, nodeid);
1076        spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1077    }
1078    return 1;
1079}
1080#endif
1081
1082/*
1083 * Allocates and initializes nodelists for a node on each slab cache, used for
1084 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1085 * will be allocated off-node since memory is not yet online for the new node.
1086 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1087 * already in use.
1088 *
1089 * Must hold cache_chain_mutex.
1090 */
1091static int init_cache_nodelists_node(int node)
1092{
1093    struct kmem_cache *cachep;
1094    struct kmem_list3 *l3;
1095    const int memsize = sizeof(struct kmem_list3);
1096
1097    list_for_each_entry(cachep, &cache_chain, next) {
1098        /*
1099         * Set up the size64 kmemlist for cpu before we can
1100         * begin anything. Make sure some other cpu on this
1101         * node has not already allocated this
1102         */
1103        if (!cachep->nodelists[node]) {
1104            l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1105            if (!l3)
1106                return -ENOMEM;
1107            kmem_list3_init(l3);
1108            l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1109                ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1110
1111            /*
1112             * The l3s don't come and go as CPUs come and
1113             * go. cache_chain_mutex is sufficient
1114             * protection here.
1115             */
1116            cachep->nodelists[node] = l3;
1117        }
1118
1119        spin_lock_irq(&cachep->nodelists[node]->list_lock);
1120        cachep->nodelists[node]->free_limit =
1121            (1 + nr_cpus_node(node)) *
1122            cachep->batchcount + cachep->num;
1123        spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1124    }
1125    return 0;
1126}
1127
1128static void __cpuinit cpuup_canceled(long cpu)
1129{
1130    struct kmem_cache *cachep;
1131    struct kmem_list3 *l3 = NULL;
1132    int node = cpu_to_mem(cpu);
1133    const struct cpumask *mask = cpumask_of_node(node);
1134
1135    list_for_each_entry(cachep, &cache_chain, next) {
1136        struct array_cache *nc;
1137        struct array_cache *shared;
1138        struct array_cache **alien;
1139
1140        /* cpu is dead; no one can alloc from it. */
1141        nc = cachep->array[cpu];
1142        cachep->array[cpu] = NULL;
1143        l3 = cachep->nodelists[node];
1144
1145        if (!l3)
1146            goto free_array_cache;
1147
1148        spin_lock_irq(&l3->list_lock);
1149
1150        /* Free limit for this kmem_list3 */
1151        l3->free_limit -= cachep->batchcount;
1152        if (nc)
1153            free_block(cachep, nc->entry, nc->avail, node);
1154
1155        if (!cpumask_empty(mask)) {
1156            spin_unlock_irq(&l3->list_lock);
1157            goto free_array_cache;
1158        }
1159
1160        shared = l3->shared;
1161        if (shared) {
1162            free_block(cachep, shared->entry,
1163                   shared->avail, node);
1164            l3->shared = NULL;
1165        }
1166
1167        alien = l3->alien;
1168        l3->alien = NULL;
1169
1170        spin_unlock_irq(&l3->list_lock);
1171
1172        kfree(shared);
1173        if (alien) {
1174            drain_alien_cache(cachep, alien);
1175            free_alien_cache(alien);
1176        }
1177free_array_cache:
1178        kfree(nc);
1179    }
1180    /*
1181     * In the previous loop, all the objects were freed to
1182     * the respective cache's slabs, now we can go ahead and
1183     * shrink each nodelist to its limit.
1184     */
1185    list_for_each_entry(cachep, &cache_chain, next) {
1186        l3 = cachep->nodelists[node];
1187        if (!l3)
1188            continue;
1189        drain_freelist(cachep, l3, l3->free_objects);
1190    }
1191}
1192
1193static int __cpuinit cpuup_prepare(long cpu)
1194{
1195    struct kmem_cache *cachep;
1196    struct kmem_list3 *l3 = NULL;
1197    int node = cpu_to_mem(cpu);
1198    int err;
1199
1200    /*
1201     * We need to do this right in the beginning since
1202     * alloc_arraycache's are going to use this list.
1203     * kmalloc_node allows us to add the slab to the right
1204     * kmem_list3 and not this cpu's kmem_list3
1205     */
1206    err = init_cache_nodelists_node(node);
1207    if (err < 0)
1208        goto bad;
1209
1210    /*
1211     * Now we can go ahead with allocating the shared arrays and
1212     * array caches
1213     */
1214    list_for_each_entry(cachep, &cache_chain, next) {
1215        struct array_cache *nc;
1216        struct array_cache *shared = NULL;
1217        struct array_cache **alien = NULL;
1218
1219        nc = alloc_arraycache(node, cachep->limit,
1220                    cachep->batchcount, GFP_KERNEL);
1221        if (!nc)
1222            goto bad;
1223        if (cachep->shared) {
1224            shared = alloc_arraycache(node,
1225                cachep->shared * cachep->batchcount,
1226                0xbaadf00d, GFP_KERNEL);
1227            if (!shared) {
1228                kfree(nc);
1229                goto bad;
1230            }
1231        }
1232        if (use_alien_caches) {
1233            alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1234            if (!alien) {
1235                kfree(shared);
1236                kfree(nc);
1237                goto bad;
1238            }
1239        }
1240        cachep->array[cpu] = nc;
1241        l3 = cachep->nodelists[node];
1242        BUG_ON(!l3);
1243
1244        spin_lock_irq(&l3->list_lock);
1245        if (!l3->shared) {
1246            /*
1247             * We are serialised from CPU_DEAD or
1248             * CPU_UP_CANCELLED by the cpucontrol lock
1249             */
1250            l3->shared = shared;
1251            shared = NULL;
1252        }
1253#ifdef CONFIG_NUMA
1254        if (!l3->alien) {
1255            l3->alien = alien;
1256            alien = NULL;
1257        }
1258#endif
1259        spin_unlock_irq(&l3->list_lock);
1260        kfree(shared);
1261        free_alien_cache(alien);
1262    }
1263    init_node_lock_keys(node);
1264
1265    return 0;
1266bad:
1267    cpuup_canceled(cpu);
1268    return -ENOMEM;
1269}
1270
1271static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1272                    unsigned long action, void *hcpu)
1273{
1274    long cpu = (long)hcpu;
1275    int err = 0;
1276
1277    switch (action) {
1278    case CPU_UP_PREPARE:
1279    case CPU_UP_PREPARE_FROZEN:
1280        mutex_lock(&cache_chain_mutex);
1281        err = cpuup_prepare(cpu);
1282        mutex_unlock(&cache_chain_mutex);
1283        break;
1284    case CPU_ONLINE:
1285    case CPU_ONLINE_FROZEN:
1286        start_cpu_timer(cpu);
1287        break;
1288#ifdef CONFIG_HOTPLUG_CPU
1289      case CPU_DOWN_PREPARE:
1290      case CPU_DOWN_PREPARE_FROZEN:
1291        /*
1292         * Shutdown cache reaper. Note that the cache_chain_mutex is
1293         * held so that if cache_reap() is invoked it cannot do
1294         * anything expensive but will only modify reap_work
1295         * and reschedule the timer.
1296        */
1297        cancel_rearming_delayed_work(&per_cpu(slab_reap_work, cpu));
1298        /* Now the cache_reaper is guaranteed to be not running. */
1299        per_cpu(slab_reap_work, cpu).work.func = NULL;
1300          break;
1301      case CPU_DOWN_FAILED:
1302      case CPU_DOWN_FAILED_FROZEN:
1303        start_cpu_timer(cpu);
1304          break;
1305    case CPU_DEAD:
1306    case CPU_DEAD_FROZEN:
1307        /*
1308         * Even if all the cpus of a node are down, we don't free the
1309         * kmem_list3 of any cache. This to avoid a race between
1310         * cpu_down, and a kmalloc allocation from another cpu for
1311         * memory from the node of the cpu going down. The list3
1312         * structure is usually allocated from kmem_cache_create() and
1313         * gets destroyed at kmem_cache_destroy().
1314         */
1315        /* fall through */
1316#endif
1317    case CPU_UP_CANCELED:
1318    case CPU_UP_CANCELED_FROZEN:
1319        mutex_lock(&cache_chain_mutex);
1320        cpuup_canceled(cpu);
1321        mutex_unlock(&cache_chain_mutex);
1322        break;
1323    }
1324    return notifier_from_errno(err);
1325}
1326
1327static struct notifier_block __cpuinitdata cpucache_notifier = {
1328    &cpuup_callback, NULL, 0
1329};
1330
1331#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1332/*
1333 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1334 * Returns -EBUSY if all objects cannot be drained so that the node is not
1335 * removed.
1336 *
1337 * Must hold cache_chain_mutex.
1338 */
1339static int __meminit drain_cache_nodelists_node(int node)
1340{
1341    struct kmem_cache *cachep;
1342    int ret = 0;
1343
1344    list_for_each_entry(cachep, &cache_chain, next) {
1345        struct kmem_list3 *l3;
1346
1347        l3 = cachep->nodelists[node];
1348        if (!l3)
1349            continue;
1350
1351        drain_freelist(cachep, l3, l3->free_objects);
1352
1353        if (!list_empty(&l3->slabs_full) ||
1354            !list_empty(&l3->slabs_partial)) {
1355            ret = -EBUSY;
1356            break;
1357        }
1358    }
1359    return ret;
1360}
1361
1362static int __meminit slab_memory_callback(struct notifier_block *self,
1363                    unsigned long action, void *arg)
1364{
1365    struct memory_notify *mnb = arg;
1366    int ret = 0;
1367    int nid;
1368
1369    nid = mnb->status_change_nid;
1370    if (nid < 0)
1371        goto out;
1372
1373    switch (action) {
1374    case MEM_GOING_ONLINE:
1375        mutex_lock(&cache_chain_mutex);
1376        ret = init_cache_nodelists_node(nid);
1377        mutex_unlock(&cache_chain_mutex);
1378        break;
1379    case MEM_GOING_OFFLINE:
1380        mutex_lock(&cache_chain_mutex);
1381        ret = drain_cache_nodelists_node(nid);
1382        mutex_unlock(&cache_chain_mutex);
1383        break;
1384    case MEM_ONLINE:
1385    case MEM_OFFLINE:
1386    case MEM_CANCEL_ONLINE:
1387    case MEM_CANCEL_OFFLINE:
1388        break;
1389    }
1390out:
1391    return ret ? notifier_from_errno(ret) : NOTIFY_OK;
1392}
1393#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1394
1395/*
1396 * swap the static kmem_list3 with kmalloced memory
1397 */
1398static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1399                int nodeid)
1400{
1401    struct kmem_list3 *ptr;
1402
1403    ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1404    BUG_ON(!ptr);
1405
1406    memcpy(ptr, list, sizeof(struct kmem_list3));
1407    /*
1408     * Do not assume that spinlocks can be initialized via memcpy:
1409     */
1410    spin_lock_init(&ptr->list_lock);
1411
1412    MAKE_ALL_LISTS(cachep, ptr, nodeid);
1413    cachep->nodelists[nodeid] = ptr;
1414}
1415
1416/*
1417 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1418 * size of kmem_list3.
1419 */
1420static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1421{
1422    int node;
1423
1424    for_each_online_node(node) {
1425        cachep->nodelists[node] = &initkmem_list3[index + node];
1426        cachep->nodelists[node]->next_reap = jiffies +
1427            REAPTIMEOUT_LIST3 +
1428            ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1429    }
1430}
1431
1432/*
1433 * Initialisation. Called after the page allocator have been initialised and
1434 * before smp_init().
1435 */
1436void __init kmem_cache_init(void)
1437{
1438    size_t left_over;
1439    struct cache_sizes *sizes;
1440    struct cache_names *names;
1441    int i;
1442    int order;
1443    int node;
1444
1445    if (num_possible_nodes() == 1)
1446        use_alien_caches = 0;
1447
1448    for (i = 0; i < NUM_INIT_LISTS; i++) {
1449        kmem_list3_init(&initkmem_list3[i]);
1450        if (i < MAX_NUMNODES)
1451            cache_cache.nodelists[i] = NULL;
1452    }
1453    set_up_list3s(&cache_cache, CACHE_CACHE);
1454
1455    /*
1456     * Fragmentation resistance on low memory - only use bigger
1457     * page orders on machines with more than 32MB of memory.
1458     */
1459    if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1460        slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1461
1462    /* Bootstrap is tricky, because several objects are allocated
1463     * from caches that do not exist yet:
1464     * 1) initialize the cache_cache cache: it contains the struct
1465     * kmem_cache structures of all caches, except cache_cache itself:
1466     * cache_cache is statically allocated.
1467     * Initially an __init data area is used for the head array and the
1468     * kmem_list3 structures, it's replaced with a kmalloc allocated
1469     * array at the end of the bootstrap.
1470     * 2) Create the first kmalloc cache.
1471     * The struct kmem_cache for the new cache is allocated normally.
1472     * An __init data area is used for the head array.
1473     * 3) Create the remaining kmalloc caches, with minimally sized
1474     * head arrays.
1475     * 4) Replace the __init data head arrays for cache_cache and the first
1476     * kmalloc cache with kmalloc allocated arrays.
1477     * 5) Replace the __init data for kmem_list3 for cache_cache and
1478     * the other cache's with kmalloc allocated memory.
1479     * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1480     */
1481
1482    node = numa_mem_id();
1483
1484    /* 1) create the cache_cache */
1485    INIT_LIST_HEAD(&cache_chain);
1486    list_add(&cache_cache.next, &cache_chain);
1487    cache_cache.colour_off = cache_line_size();
1488    cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1489    cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1490
1491    /*
1492     * struct kmem_cache size depends on nr_node_ids, which
1493     * can be less than MAX_NUMNODES.
1494     */
1495    cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1496                 nr_node_ids * sizeof(struct kmem_list3 *);
1497#if DEBUG
1498    cache_cache.obj_size = cache_cache.buffer_size;
1499#endif
1500    cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1501                    cache_line_size());
1502    cache_cache.reciprocal_buffer_size =
1503        reciprocal_value(cache_cache.buffer_size);
1504
1505    for (order = 0; order < MAX_ORDER; order++) {
1506        cache_estimate(order, cache_cache.buffer_size,
1507            cache_line_size(), 0, &left_over, &cache_cache.num);
1508        if (cache_cache.num)
1509            break;
1510    }
1511    BUG_ON(!cache_cache.num);
1512    cache_cache.gfporder = order;
1513    cache_cache.colour = left_over / cache_cache.colour_off;
1514    cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1515                      sizeof(struct slab), cache_line_size());
1516
1517    /* 2+3) create the kmalloc caches */
1518    sizes = malloc_sizes;
1519    names = cache_names;
1520
1521    /*
1522     * Initialize the caches that provide memory for the array cache and the
1523     * kmem_list3 structures first. Without this, further allocations will
1524     * bug.
1525     */
1526
1527    sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1528                    sizes[INDEX_AC].cs_size,
1529                    ARCH_KMALLOC_MINALIGN,
1530                    ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1531                    NULL);
1532
1533    if (INDEX_AC != INDEX_L3) {
1534        sizes[INDEX_L3].cs_cachep =
1535            kmem_cache_create(names[INDEX_L3].name,
1536                sizes[INDEX_L3].cs_size,
1537                ARCH_KMALLOC_MINALIGN,
1538                ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1539                NULL);
1540    }
1541
1542    slab_early_init = 0;
1543
1544    while (sizes->cs_size != ULONG_MAX) {
1545        /*
1546         * For performance, all the general caches are L1 aligned.
1547         * This should be particularly beneficial on SMP boxes, as it
1548         * eliminates "false sharing".
1549         * Note for systems short on memory removing the alignment will
1550         * allow tighter packing of the smaller caches.
1551         */
1552        if (!sizes->cs_cachep) {
1553            sizes->cs_cachep = kmem_cache_create(names->name,
1554                    sizes->cs_size,
1555                    ARCH_KMALLOC_MINALIGN,
1556                    ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1557                    NULL);
1558        }
1559#ifdef CONFIG_ZONE_DMA
1560        sizes->cs_dmacachep = kmem_cache_create(
1561                    names->name_dma,
1562                    sizes->cs_size,
1563                    ARCH_KMALLOC_MINALIGN,
1564                    ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1565                        SLAB_PANIC,
1566                    NULL);
1567#endif
1568        sizes++;
1569        names++;
1570    }
1571    /* 4) Replace the bootstrap head arrays */
1572    {
1573        struct array_cache *ptr;
1574
1575        ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1576
1577        BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1578        memcpy(ptr, cpu_cache_get(&cache_cache),
1579               sizeof(struct arraycache_init));
1580        /*
1581         * Do not assume that spinlocks can be initialized via memcpy:
1582         */
1583        spin_lock_init(&ptr->lock);
1584
1585        cache_cache.array[smp_processor_id()] = ptr;
1586
1587        ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1588
1589        BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1590               != &initarray_generic.cache);
1591        memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1592               sizeof(struct arraycache_init));
1593        /*
1594         * Do not assume that spinlocks can be initialized via memcpy:
1595         */
1596        spin_lock_init(&ptr->lock);
1597
1598        malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1599            ptr;
1600    }
1601    /* 5) Replace the bootstrap kmem_list3's */
1602    {
1603        int nid;
1604
1605        for_each_online_node(nid) {
1606            init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1607
1608            init_list(malloc_sizes[INDEX_AC].cs_cachep,
1609                  &initkmem_list3[SIZE_AC + nid], nid);
1610
1611            if (INDEX_AC != INDEX_L3) {
1612                init_list(malloc_sizes[INDEX_L3].cs_cachep,
1613                      &initkmem_list3[SIZE_L3 + nid], nid);
1614            }
1615        }
1616    }
1617
1618    g_cpucache_up = EARLY;
1619}
1620
1621void __init kmem_cache_init_late(void)
1622{
1623    struct kmem_cache *cachep;
1624
1625    /* 6) resize the head arrays to their final sizes */
1626    mutex_lock(&cache_chain_mutex);
1627    list_for_each_entry(cachep, &cache_chain, next)
1628        if (enable_cpucache(cachep, GFP_NOWAIT))
1629            BUG();
1630    mutex_unlock(&cache_chain_mutex);
1631
1632    /* Done! */
1633    g_cpucache_up = FULL;
1634
1635    /* Annotate slab for lockdep -- annotate the malloc caches */
1636    init_lock_keys();
1637
1638    /*
1639     * Register a cpu startup notifier callback that initializes
1640     * cpu_cache_get for all new cpus
1641     */
1642    register_cpu_notifier(&cpucache_notifier);
1643
1644#ifdef CONFIG_NUMA
1645    /*
1646     * Register a memory hotplug callback that initializes and frees
1647     * nodelists.
1648     */
1649    hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1650#endif
1651
1652    /*
1653     * The reap timers are started later, with a module init call: That part
1654     * of the kernel is not yet operational.
1655     */
1656}
1657
1658static int __init cpucache_init(void)
1659{
1660    int cpu;
1661
1662    /*
1663     * Register the timers that return unneeded pages to the page allocator
1664     */
1665    for_each_online_cpu(cpu)
1666        start_cpu_timer(cpu);
1667    return 0;
1668}
1669__initcall(cpucache_init);
1670
1671/*
1672 * Interface to system's page allocator. No need to hold the cache-lock.
1673 *
1674 * If we requested dmaable memory, we will get it. Even if we
1675 * did not request dmaable memory, we might get it, but that
1676 * would be relatively rare and ignorable.
1677 */
1678static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1679{
1680    struct page *page;
1681    int nr_pages;
1682    int i;
1683
1684#ifndef CONFIG_MMU
1685    /*
1686     * Nommu uses slab's for process anonymous memory allocations, and thus
1687     * requires __GFP_COMP to properly refcount higher order allocations
1688     */
1689    flags |= __GFP_COMP;
1690#endif
1691
1692    flags |= cachep->gfpflags;
1693    if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1694        flags |= __GFP_RECLAIMABLE;
1695
1696    page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1697    if (!page)
1698        return NULL;
1699
1700    nr_pages = (1 << cachep->gfporder);
1701    if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1702        add_zone_page_state(page_zone(page),
1703            NR_SLAB_RECLAIMABLE, nr_pages);
1704    else
1705        add_zone_page_state(page_zone(page),
1706            NR_SLAB_UNRECLAIMABLE, nr_pages);
1707    for (i = 0; i < nr_pages; i++)
1708        __SetPageSlab(page + i);
1709
1710    if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1711        kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1712
1713        if (cachep->ctor)
1714            kmemcheck_mark_uninitialized_pages(page, nr_pages);
1715        else
1716            kmemcheck_mark_unallocated_pages(page, nr_pages);
1717    }
1718
1719    return page_address(page);
1720}
1721
1722/*
1723 * Interface to system's page release.
1724 */
1725static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1726{
1727    unsigned long i = (1 << cachep->gfporder);
1728    struct page *page = virt_to_page(addr);
1729    const unsigned long nr_freed = i;
1730
1731    kmemcheck_free_shadow(page, cachep->gfporder);
1732
1733    if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1734        sub_zone_page_state(page_zone(page),
1735                NR_SLAB_RECLAIMABLE, nr_freed);
1736    else
1737        sub_zone_page_state(page_zone(page),
1738                NR_SLAB_UNRECLAIMABLE, nr_freed);
1739    while (i--) {
1740        BUG_ON(!PageSlab(page));
1741        __ClearPageSlab(page);
1742        page++;
1743    }
1744    if (current->reclaim_state)
1745        current->reclaim_state->reclaimed_slab += nr_freed;
1746    free_pages((unsigned long)addr, cachep->gfporder);
1747}
1748
1749static void kmem_rcu_free(struct rcu_head *head)
1750{
1751    struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1752    struct kmem_cache *cachep = slab_rcu->cachep;
1753
1754    kmem_freepages(cachep, slab_rcu->addr);
1755    if (OFF_SLAB(cachep))
1756        kmem_cache_free(cachep->slabp_cache, slab_rcu);
1757}
1758
1759#if DEBUG
1760
1761#ifdef CONFIG_DEBUG_PAGEALLOC
1762static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1763                unsigned long caller)
1764{
1765    int size = obj_size(cachep);
1766
1767    addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1768
1769    if (size < 5 * sizeof(unsigned long))
1770        return;
1771
1772    *addr++ = 0x12345678;
1773    *addr++ = caller;
1774    *addr++ = smp_processor_id();
1775    size -= 3 * sizeof(unsigned long);
1776    {
1777        unsigned long *sptr = &caller;
1778        unsigned long svalue;
1779
1780        while (!kstack_end(sptr)) {
1781            svalue = *sptr++;
1782            if (kernel_text_address(svalue)) {
1783                *addr++ = svalue;
1784                size -= sizeof(unsigned long);
1785                if (size <= sizeof(unsigned long))
1786                    break;
1787            }
1788        }
1789
1790    }
1791    *addr++ = 0x87654321;
1792}
1793#endif
1794
1795static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1796{
1797    int size = obj_size(cachep);
1798    addr = &((char *)addr)[obj_offset(cachep)];
1799
1800    memset(addr, val, size);
1801    *(unsigned char *)(addr + size - 1) = POISON_END;
1802}
1803
1804static void dump_line(char *data, int offset, int limit)
1805{
1806    int i;
1807    unsigned char error = 0;
1808    int bad_count = 0;
1809
1810    printk(KERN_ERR "%03x:", offset);
1811    for (i = 0; i < limit; i++) {
1812        if (data[offset + i] != POISON_FREE) {
1813            error = data[offset + i];
1814            bad_count++;
1815        }
1816        printk(" %02x", (unsigned char)data[offset + i]);
1817    }
1818    printk("\n");
1819
1820    if (bad_count == 1) {
1821        error ^= POISON_FREE;
1822        if (!(error & (error - 1))) {
1823            printk(KERN_ERR "Single bit error detected. Probably "
1824                    "bad RAM.\n");
1825#ifdef CONFIG_X86
1826            printk(KERN_ERR "Run memtest86+ or a similar memory "
1827                    "test tool.\n");
1828#else
1829            printk(KERN_ERR "Run a memory test tool.\n");
1830#endif
1831        }
1832    }
1833}
1834#endif
1835
1836#if DEBUG
1837
1838static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1839{
1840    int i, size;
1841    char *realobj;
1842
1843    if (cachep->flags & SLAB_RED_ZONE) {
1844        printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1845            *dbg_redzone1(cachep, objp),
1846            *dbg_redzone2(cachep, objp));
1847    }
1848
1849    if (cachep->flags & SLAB_STORE_USER) {
1850        printk(KERN_ERR "Last user: [<%p>]",
1851            *dbg_userword(cachep, objp));
1852        print_symbol("(%s)",
1853                (unsigned long)*dbg_userword(cachep, objp));
1854        printk("\n");
1855    }
1856    realobj = (char *)objp + obj_offset(cachep);
1857    size = obj_size(cachep);
1858    for (i = 0; i < size && lines; i += 16, lines--) {
1859        int limit;
1860        limit = 16;
1861        if (i + limit > size)
1862            limit = size - i;
1863        dump_line(realobj, i, limit);
1864    }
1865}
1866
1867static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1868{
1869    char *realobj;
1870    int size, i;
1871    int lines = 0;
1872
1873    realobj = (char *)objp + obj_offset(cachep);
1874    size = obj_size(cachep);
1875
1876    for (i = 0; i < size; i++) {
1877        char exp = POISON_FREE;
1878        if (i == size - 1)
1879            exp = POISON_END;
1880        if (realobj[i] != exp) {
1881            int limit;
1882            /* Mismatch ! */
1883            /* Print header */
1884            if (lines == 0) {
1885                printk(KERN_ERR
1886                    "Slab corruption: %s start=%p, len=%d\n",
1887                    cachep->name, realobj, size);
1888                print_objinfo(cachep, objp, 0);
1889            }
1890            /* Hexdump the affected line */
1891            i = (i / 16) * 16;
1892            limit = 16;
1893            if (i + limit > size)
1894                limit = size - i;
1895            dump_line(realobj, i, limit);
1896            i += 16;
1897            lines++;
1898            /* Limit to 5 lines */
1899            if (lines > 5)
1900                break;
1901        }
1902    }
1903    if (lines != 0) {
1904        /* Print some data about the neighboring objects, if they
1905         * exist:
1906         */
1907        struct slab *slabp = virt_to_slab(objp);
1908        unsigned int objnr;
1909
1910        objnr = obj_to_index(cachep, slabp, objp);
1911        if (objnr) {
1912            objp = index_to_obj(cachep, slabp, objnr - 1);
1913            realobj = (char *)objp + obj_offset(cachep);
1914            printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1915                   realobj, size);
1916            print_objinfo(cachep, objp, 2);
1917        }
1918        if (objnr + 1 < cachep->num) {
1919            objp = index_to_obj(cachep, slabp, objnr + 1);
1920            realobj = (char *)objp + obj_offset(cachep);
1921            printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1922                   realobj, size);
1923            print_objinfo(cachep, objp, 2);
1924        }
1925    }
1926}
1927#endif
1928
1929#if DEBUG
1930static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1931{
1932    int i;
1933    for (i = 0; i < cachep->num; i++) {
1934        void *objp = index_to_obj(cachep, slabp, i);
1935
1936        if (cachep->flags & SLAB_POISON) {
1937#ifdef CONFIG_DEBUG_PAGEALLOC
1938            if (cachep->buffer_size % PAGE_SIZE == 0 &&
1939                    OFF_SLAB(cachep))
1940                kernel_map_pages(virt_to_page(objp),
1941                    cachep->buffer_size / PAGE_SIZE, 1);
1942            else
1943                check_poison_obj(cachep, objp);
1944#else
1945            check_poison_obj(cachep, objp);
1946#endif
1947        }
1948        if (cachep->flags & SLAB_RED_ZONE) {
1949            if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1950                slab_error(cachep, "start of a freed object "
1951                       "was overwritten");
1952            if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1953                slab_error(cachep, "end of a freed object "
1954                       "was overwritten");
1955        }
1956    }
1957}
1958#else
1959static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1960{
1961}
1962#endif
1963
1964/**
1965 * slab_destroy - destroy and release all objects in a slab
1966 * @cachep: cache pointer being destroyed
1967 * @slabp: slab pointer being destroyed
1968 *
1969 * Destroy all the objs in a slab, and release the mem back to the system.
1970 * Before calling the slab must have been unlinked from the cache. The
1971 * cache-lock is not held/needed.
1972 */
1973static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1974{
1975    void *addr = slabp->s_mem - slabp->colouroff;
1976
1977    slab_destroy_debugcheck(cachep, slabp);
1978    if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1979        struct slab_rcu *slab_rcu;
1980
1981        slab_rcu = (struct slab_rcu *)slabp;
1982        slab_rcu->cachep = cachep;
1983        slab_rcu->addr = addr;
1984        call_rcu(&slab_rcu->head, kmem_rcu_free);
1985    } else {
1986        kmem_freepages(cachep, addr);
1987        if (OFF_SLAB(cachep))
1988            kmem_cache_free(cachep->slabp_cache, slabp);
1989    }
1990}
1991
1992static void __kmem_cache_destroy(struct kmem_cache *cachep)
1993{
1994    int i;
1995    struct kmem_list3 *l3;
1996
1997    for_each_online_cpu(i)
1998        kfree(cachep->array[i]);
1999
2000    /* NUMA: free the list3 structures */
2001    for_each_online_node(i) {
2002        l3 = cachep->nodelists[i];
2003        if (l3) {
2004            kfree(l3->shared);
2005            free_alien_cache(l3->alien);
2006            kfree(l3);
2007        }
2008    }
2009    kmem_cache_free(&cache_cache, cachep);
2010}
2011
2012
2013/**
2014 * calculate_slab_order - calculate size (page order) of slabs
2015 * @cachep: pointer to the cache that is being created
2016 * @size: size of objects to be created in this cache.
2017 * @align: required alignment for the objects.
2018 * @flags: slab allocation flags
2019 *
2020 * Also calculates the number of objects per slab.
2021 *
2022 * This could be made much more intelligent. For now, try to avoid using
2023 * high order pages for slabs. When the gfp() functions are more friendly
2024 * towards high-order requests, this should be changed.
2025 */
2026static size_t calculate_slab_order(struct kmem_cache *cachep,
2027            size_t size, size_t align, unsigned long flags)
2028{
2029    unsigned long offslab_limit;
2030    size_t left_over = 0;
2031    int gfporder;
2032
2033    for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2034        unsigned int num;
2035        size_t remainder;
2036
2037        cache_estimate(gfporder, size, align, flags, &remainder, &num);
2038        if (!num)
2039            continue;
2040
2041        if (flags & CFLGS_OFF_SLAB) {
2042            /*
2043             * Max number of objs-per-slab for caches which
2044             * use off-slab slabs. Needed to avoid a possible
2045             * looping condition in cache_grow().
2046             */
2047            offslab_limit = size - sizeof(struct slab);
2048            offslab_limit /= sizeof(kmem_bufctl_t);
2049
2050             if (num > offslab_limit)
2051                break;
2052        }
2053
2054        /* Found something acceptable - save it away */
2055        cachep->num = num;
2056        cachep->gfporder = gfporder;
2057        left_over = remainder;
2058
2059        /*
2060         * A VFS-reclaimable slab tends to have most allocations
2061         * as GFP_NOFS and we really don't want to have to be allocating
2062         * higher-order pages when we are unable to shrink dcache.
2063         */
2064        if (flags & SLAB_RECLAIM_ACCOUNT)
2065            break;
2066
2067        /*
2068         * Large number of objects is good, but very large slabs are
2069         * currently bad for the gfp()s.
2070         */
2071        if (gfporder >= slab_break_gfp_order)
2072            break;
2073
2074        /*
2075         * Acceptable internal fragmentation?
2076         */
2077        if (left_over * 8 <= (PAGE_SIZE << gfporder))
2078            break;
2079    }
2080    return left_over;
2081}
2082
2083static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2084{
2085    if (g_cpucache_up == FULL)
2086        return enable_cpucache(cachep, gfp);
2087
2088    if (g_cpucache_up == NONE) {
2089        /*
2090         * Note: the first kmem_cache_create must create the cache
2091         * that's used by kmalloc(24), otherwise the creation of
2092         * further caches will BUG().
2093         */
2094        cachep->array[smp_processor_id()] = &initarray_generic.cache;
2095
2096        /*
2097         * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2098         * the first cache, then we need to set up all its list3s,
2099         * otherwise the creation of further caches will BUG().
2100         */
2101        set_up_list3s(cachep, SIZE_AC);
2102        if (INDEX_AC == INDEX_L3)
2103            g_cpucache_up = PARTIAL_L3;
2104        else
2105            g_cpucache_up = PARTIAL_AC;
2106    } else {
2107        cachep->array[smp_processor_id()] =
2108            kmalloc(sizeof(struct arraycache_init), gfp);
2109
2110        if (g_cpucache_up == PARTIAL_AC) {
2111            set_up_list3s(cachep, SIZE_L3);
2112            g_cpucache_up = PARTIAL_L3;
2113        } else {
2114            int node;
2115            for_each_online_node(node) {
2116                cachep->nodelists[node] =
2117                    kmalloc_node(sizeof(struct kmem_list3),
2118                        gfp, node);
2119                BUG_ON(!cachep->nodelists[node]);
2120                kmem_list3_init(cachep->nodelists[node]);
2121            }
2122        }
2123    }
2124    cachep->nodelists[numa_mem_id()]->next_reap =
2125            jiffies + REAPTIMEOUT_LIST3 +
2126            ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2127
2128    cpu_cache_get(cachep)->avail = 0;
2129    cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2130    cpu_cache_get(cachep)->batchcount = 1;
2131    cpu_cache_get(cachep)->touched = 0;
2132    cachep->batchcount = 1;
2133    cachep->limit = BOOT_CPUCACHE_ENTRIES;
2134    return 0;
2135}
2136
2137/**
2138 * kmem_cache_create - Create a cache.
2139 * @name: A string which is used in /proc/slabinfo to identify this cache.
2140 * @size: The size of objects to be created in this cache.
2141 * @align: The required alignment for the objects.
2142 * @flags: SLAB flags
2143 * @ctor: A constructor for the objects.
2144 *
2145 * Returns a ptr to the cache on success, NULL on failure.
2146 * Cannot be called within a int, but can be interrupted.
2147 * The @ctor is run when new pages are allocated by the cache.
2148 *
2149 * @name must be valid until the cache is destroyed. This implies that
2150 * the module calling this has to destroy the cache before getting unloaded.
2151 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2152 * therefore applications must manage it themselves.
2153 *
2154 * The flags are
2155 *
2156 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2157 * to catch references to uninitialised memory.
2158 *
2159 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2160 * for buffer overruns.
2161 *
2162 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2163 * cacheline. This can be beneficial if you're counting cycles as closely
2164 * as davem.
2165 */
2166struct kmem_cache *
2167kmem_cache_create (const char *name, size_t size, size_t align,
2168    unsigned long flags, void (*ctor)(void *))
2169{
2170    size_t left_over, slab_size, ralign;
2171    struct kmem_cache *cachep = NULL, *pc;
2172    gfp_t gfp;
2173
2174    /*
2175     * Sanity checks... these are all serious usage bugs.
2176     */
2177    if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2178        size > KMALLOC_MAX_SIZE) {
2179        printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2180                name);
2181        BUG();
2182    }
2183
2184    /*
2185     * We use cache_chain_mutex to ensure a consistent view of
2186     * cpu_online_mask as well. Please see cpuup_callback
2187     */
2188    if (slab_is_available()) {
2189        get_online_cpus();
2190        mutex_lock(&cache_chain_mutex);
2191    }
2192
2193    list_for_each_entry(pc, &cache_chain, next) {
2194        char tmp;
2195        int res;
2196
2197        /*
2198         * This happens when the module gets unloaded and doesn't
2199         * destroy its slab cache and no-one else reuses the vmalloc
2200         * area of the module. Print a warning.
2201         */
2202        res = probe_kernel_address(pc->name, tmp);
2203        if (res) {
2204            printk(KERN_ERR
2205                   "SLAB: cache with size %d has lost its name\n",
2206                   pc->buffer_size);
2207            continue;
2208        }
2209
2210        if (!strcmp(pc->name, name)) {
2211            printk(KERN_ERR
2212                   "kmem_cache_create: duplicate cache %s\n", name);
2213            dump_stack();
2214            goto oops;
2215        }
2216    }
2217
2218#if DEBUG
2219    WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2220#if FORCED_DEBUG
2221    /*
2222     * Enable redzoning and last user accounting, except for caches with
2223     * large objects, if the increased size would increase the object size
2224     * above the next power of two: caches with object sizes just above a
2225     * power of two have a significant amount of internal fragmentation.
2226     */
2227    if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2228                        2 * sizeof(unsigned long long)))
2229        flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2230    if (!(flags & SLAB_DESTROY_BY_RCU))
2231        flags |= SLAB_POISON;
2232#endif
2233    if (flags & SLAB_DESTROY_BY_RCU)
2234        BUG_ON(flags & SLAB_POISON);
2235#endif
2236    /*
2237     * Always checks flags, a caller might be expecting debug support which
2238     * isn't available.
2239     */
2240    BUG_ON(flags & ~CREATE_MASK);
2241
2242    /*
2243     * Check that size is in terms of words. This is needed to avoid
2244     * unaligned accesses for some archs when redzoning is used, and makes
2245     * sure any on-slab bufctl's are also correctly aligned.
2246     */
2247    if (size & (BYTES_PER_WORD - 1)) {
2248        size += (BYTES_PER_WORD - 1);
2249        size &= ~(BYTES_PER_WORD - 1);
2250    }
2251
2252    /* calculate the final buffer alignment: */
2253
2254    /* 1) arch recommendation: can be overridden for debug */
2255    if (flags & SLAB_HWCACHE_ALIGN) {
2256        /*
2257         * Default alignment: as specified by the arch code. Except if
2258         * an object is really small, then squeeze multiple objects into
2259         * one cacheline.
2260         */
2261        ralign = cache_line_size();
2262        while (size <= ralign / 2)
2263            ralign /= 2;
2264    } else {
2265        ralign = BYTES_PER_WORD;
2266    }
2267
2268    /*
2269     * Redzoning and user store require word alignment or possibly larger.
2270     * Note this will be overridden by architecture or caller mandated
2271     * alignment if either is greater than BYTES_PER_WORD.
2272     */
2273    if (flags & SLAB_STORE_USER)
2274        ralign = BYTES_PER_WORD;
2275
2276    if (flags & SLAB_RED_ZONE) {
2277        ralign = REDZONE_ALIGN;
2278        /* If redzoning, ensure that the second redzone is suitably
2279         * aligned, by adjusting the object size accordingly. */
2280        size += REDZONE_ALIGN - 1;
2281        size &= ~(REDZONE_ALIGN - 1);
2282    }
2283
2284    /* 2) arch mandated alignment */
2285    if (ralign < ARCH_SLAB_MINALIGN) {
2286        ralign = ARCH_SLAB_MINALIGN;
2287    }
2288    /* 3) caller mandated alignment */
2289    if (ralign < align) {
2290        ralign = align;
2291    }
2292    /* disable debug if not aligning with REDZONE_ALIGN */
2293    if (ralign & (__alignof__(unsigned long long) - 1))
2294        flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2295    /*
2296     * 4) Store it.
2297     */
2298    align = ralign;
2299
2300    if (slab_is_available())
2301        gfp = GFP_KERNEL;
2302    else
2303        gfp = GFP_NOWAIT;
2304
2305    /* Get cache's description obj. */
2306    cachep = kmem_cache_zalloc(&cache_cache, gfp);
2307    if (!cachep)
2308        goto oops;
2309
2310#if DEBUG
2311    cachep->obj_size = size;
2312
2313    /*
2314     * Both debugging options require word-alignment which is calculated
2315     * into align above.
2316     */
2317    if (flags & SLAB_RED_ZONE) {
2318        /* add space for red zone words */
2319        cachep->obj_offset += align;
2320        size += align + sizeof(unsigned long long);
2321    }
2322    if (flags & SLAB_STORE_USER) {
2323        /* user store requires one word storage behind the end of
2324         * the real object. But if the second red zone needs to be
2325         * aligned to 64 bits, we must allow that much space.
2326         */
2327        if (flags & SLAB_RED_ZONE)
2328            size += REDZONE_ALIGN;
2329        else
2330            size += BYTES_PER_WORD;
2331    }
2332#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2333    if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2334        && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2335        cachep->obj_offset += PAGE_SIZE - size;
2336        size = PAGE_SIZE;
2337    }
2338#endif
2339#endif
2340
2341    /*
2342     * Determine if the slab management is 'on' or 'off' slab.
2343     * (bootstrapping cannot cope with offslab caches so don't do
2344     * it too early on. Always use on-slab management when
2345     * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2346     */
2347    if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2348        !(flags & SLAB_NOLEAKTRACE))
2349        /*
2350         * Size is large, assume best to place the slab management obj
2351         * off-slab (should allow better packing of objs).
2352         */
2353        flags |= CFLGS_OFF_SLAB;
2354
2355    size = ALIGN(size, align);
2356
2357    left_over = calculate_slab_order(cachep, size, align, flags);
2358
2359    if (!cachep->num) {
2360        printk(KERN_ERR
2361               "kmem_cache_create: couldn't create cache %s.\n", name);
2362        kmem_cache_free(&cache_cache, cachep);
2363        cachep = NULL;
2364        goto oops;
2365    }
2366    slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2367              + sizeof(struct slab), align);
2368
2369    /*
2370     * If the slab has been placed off-slab, and we have enough space then
2371     * move it on-slab. This is at the expense of any extra colouring.
2372     */
2373    if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2374        flags &= ~CFLGS_OFF_SLAB;
2375        left_over -= slab_size;
2376    }
2377
2378    if (flags & CFLGS_OFF_SLAB) {
2379        /* really off slab. No need for manual alignment */
2380        slab_size =
2381            cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2382
2383#ifdef CONFIG_PAGE_POISONING
2384        /* If we're going to use the generic kernel_map_pages()
2385         * poisoning, then it's going to smash the contents of
2386         * the redzone and userword anyhow, so switch them off.
2387         */
2388        if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2389            flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2390#endif
2391    }
2392
2393    cachep->colour_off = cache_line_size();
2394    /* Offset must be a multiple of the alignment. */
2395    if (cachep->colour_off < align)
2396        cachep->colour_off = align;
2397    cachep->colour = left_over / cachep->colour_off;
2398    cachep->slab_size = slab_size;
2399    cachep->flags = flags;
2400    cachep->gfpflags = 0;
2401    if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2402        cachep->gfpflags |= GFP_DMA;
2403    cachep->buffer_size = size;
2404    cachep->reciprocal_buffer_size = reciprocal_value(size);
2405
2406    if (flags & CFLGS_OFF_SLAB) {
2407        cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2408        /*
2409         * This is a possibility for one of the malloc_sizes caches.
2410         * But since we go off slab only for object size greater than
2411         * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2412         * this should not happen at all.
2413         * But leave a BUG_ON for some lucky dude.
2414         */
2415        BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2416    }
2417    cachep->ctor = ctor;
2418    cachep->name = name;
2419
2420    if (setup_cpu_cache(cachep, gfp)) {
2421        __kmem_cache_destroy(cachep);
2422        cachep = NULL;
2423        goto oops;
2424    }
2425
2426    /* cache setup completed, link it into the list */
2427    list_add(&cachep->next, &cache_chain);
2428oops:
2429    if (!cachep && (flags & SLAB_PANIC))
2430        panic("kmem_cache_create(): failed to create slab `%s'\n",
2431              name);
2432    if (slab_is_available()) {
2433        mutex_unlock(&cache_chain_mutex);
2434        put_online_cpus();
2435    }
2436    return cachep;
2437}
2438EXPORT_SYMBOL(kmem_cache_create);
2439
2440#if DEBUG
2441static void check_irq_off(void)
2442{
2443    BUG_ON(!irqs_disabled());
2444}
2445
2446static void check_irq_on(void)
2447{
2448    BUG_ON(irqs_disabled());
2449}
2450
2451static void check_spinlock_acquired(struct kmem_cache *cachep)
2452{
2453#ifdef CONFIG_SMP
2454    check_irq_off();
2455    assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2456#endif
2457}
2458
2459static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2460{
2461#ifdef CONFIG_SMP
2462    check_irq_off();
2463    assert_spin_locked(&cachep->nodelists[node]->list_lock);
2464#endif
2465}
2466
2467#else
2468#define check_irq_off() do { } while(0)
2469#define check_irq_on() do { } while(0)
2470#define check_spinlock_acquired(x) do { } while(0)
2471#define check_spinlock_acquired_node(x, y) do { } while(0)
2472#endif
2473
2474static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2475            struct array_cache *ac,
2476            int force, int node);
2477
2478static void do_drain(void *arg)
2479{
2480    struct kmem_cache *cachep = arg;
2481    struct array_cache *ac;
2482    int node = numa_mem_id();
2483
2484    check_irq_off();
2485    ac = cpu_cache_get(cachep);
2486    spin_lock(&cachep->nodelists[node]->list_lock);
2487    free_block(cachep, ac->entry, ac->avail, node);
2488    spin_unlock(&cachep->nodelists[node]->list_lock);
2489    ac->avail = 0;
2490}
2491
2492static void drain_cpu_caches(struct kmem_cache *cachep)
2493{
2494    struct kmem_list3 *l3;
2495    int node;
2496
2497    on_each_cpu(do_drain, cachep, 1);
2498    check_irq_on();
2499    for_each_online_node(node) {
2500        l3 = cachep->nodelists[node];
2501        if (l3 && l3->alien)
2502            drain_alien_cache(cachep, l3->alien);
2503    }
2504
2505    for_each_online_node(node) {
2506        l3 = cachep->nodelists[node];
2507        if (l3)
2508            drain_array(cachep, l3, l3->shared, 1, node);
2509    }
2510}
2511
2512/*
2513 * Remove slabs from the list of free slabs.
2514 * Specify the number of slabs to drain in tofree.
2515 *
2516 * Returns the actual number of slabs released.
2517 */
2518static int drain_freelist(struct kmem_cache *cache,
2519            struct kmem_list3 *l3, int tofree)
2520{
2521    struct list_head *p;
2522    int nr_freed;
2523    struct slab *slabp;
2524
2525    nr_freed = 0;
2526    while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2527
2528        spin_lock_irq(&l3->list_lock);
2529        p = l3->slabs_free.prev;
2530        if (p == &l3->slabs_free) {
2531            spin_unlock_irq(&l3->list_lock);
2532            goto out;
2533        }
2534
2535        slabp = list_entry(p, struct slab, list);
2536#if DEBUG
2537        BUG_ON(slabp->inuse);
2538#endif
2539        list_del(&slabp->list);
2540        /*
2541         * Safe to drop the lock. The slab is no longer linked
2542         * to the cache.
2543         */
2544        l3->free_objects -= cache->num;
2545        spin_unlock_irq(&l3->list_lock);
2546        slab_destroy(cache, slabp);
2547        nr_freed++;
2548    }
2549out:
2550    return nr_freed;
2551}
2552
2553/* Called with cache_chain_mutex held to protect against cpu hotplug */
2554static int __cache_shrink(struct kmem_cache *cachep)
2555{
2556    int ret = 0, i = 0;
2557    struct kmem_list3 *l3;
2558
2559    drain_cpu_caches(cachep);
2560
2561    check_irq_on();
2562    for_each_online_node(i) {
2563        l3 = cachep->nodelists[i];
2564        if (!l3)
2565            continue;
2566
2567        drain_freelist(cachep, l3, l3->free_objects);
2568
2569        ret += !list_empty(&l3->slabs_full) ||
2570            !list_empty(&l3->slabs_partial);
2571    }
2572    return (ret ? 1 : 0);
2573}
2574
2575/**
2576 * kmem_cache_shrink - Shrink a cache.
2577 * @cachep: The cache to shrink.
2578 *
2579 * Releases as many slabs as possible for a cache.
2580 * To help debugging, a zero exit status indicates all slabs were released.
2581 */
2582int kmem_cache_shrink(struct kmem_cache *cachep)
2583{
2584    int ret;
2585    BUG_ON(!cachep || in_interrupt());
2586
2587    get_online_cpus();
2588    mutex_lock(&cache_chain_mutex);
2589    ret = __cache_shrink(cachep);
2590    mutex_unlock(&cache_chain_mutex);
2591    put_online_cpus();
2592    return ret;
2593}
2594EXPORT_SYMBOL(kmem_cache_shrink);
2595
2596/**
2597 * kmem_cache_destroy - delete a cache
2598 * @cachep: the cache to destroy
2599 *
2600 * Remove a &struct kmem_cache object from the slab cache.
2601 *
2602 * It is expected this function will be called by a module when it is
2603 * unloaded. This will remove the cache completely, and avoid a duplicate
2604 * cache being allocated each time a module is loaded and unloaded, if the
2605 * module doesn't have persistent in-kernel storage across loads and unloads.
2606 *
2607 * The cache must be empty before calling this function.
2608 *
2609 * The caller must guarantee that noone will allocate memory from the cache
2610 * during the kmem_cache_destroy().
2611 */
2612void kmem_cache_destroy(struct kmem_cache *cachep)
2613{
2614    BUG_ON(!cachep || in_interrupt());
2615
2616    /* Find the cache in the chain of caches. */
2617    get_online_cpus();
2618    mutex_lock(&cache_chain_mutex);
2619    /*
2620     * the chain is never empty, cache_cache is never destroyed
2621     */
2622    list_del(&cachep->next);
2623    if (__cache_shrink(cachep)) {
2624        slab_error(cachep, "Can't free all objects");
2625        list_add(&cachep->next, &cache_chain);
2626        mutex_unlock(&cache_chain_mutex);
2627        put_online_cpus();
2628        return;
2629    }
2630
2631    if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2632        rcu_barrier();
2633
2634    __kmem_cache_destroy(cachep);
2635    mutex_unlock(&cache_chain_mutex);
2636    put_online_cpus();
2637}
2638EXPORT_SYMBOL(kmem_cache_destroy);
2639
2640/*
2641 * Get the memory for a slab management obj.
2642 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2643 * always come from malloc_sizes caches. The slab descriptor cannot
2644 * come from the same cache which is getting created because,
2645 * when we are searching for an appropriate cache for these
2646 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2647 * If we are creating a malloc_sizes cache here it would not be visible to
2648 * kmem_find_general_cachep till the initialization is complete.
2649 * Hence we cannot have slabp_cache same as the original cache.
2650 */
2651static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2652                   int colour_off, gfp_t local_flags,
2653                   int nodeid)
2654{
2655    struct slab *slabp;
2656
2657    if (OFF_SLAB(cachep)) {
2658        /* Slab management obj is off-slab. */
2659        slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2660                          local_flags, nodeid);
2661        /*
2662         * If the first object in the slab is leaked (it's allocated
2663         * but no one has a reference to it), we want to make sure
2664         * kmemleak does not treat the ->s_mem pointer as a reference
2665         * to the object. Otherwise we will not report the leak.
2666         */
2667        kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2668                   local_flags);
2669        if (!slabp)
2670            return NULL;
2671    } else {
2672        slabp = objp + colour_off;
2673        colour_off += cachep->slab_size;
2674    }
2675    slabp->inuse = 0;
2676    slabp->colouroff = colour_off;
2677    slabp->s_mem = objp + colour_off;
2678    slabp->nodeid = nodeid;
2679    slabp->free = 0;
2680    return slabp;
2681}
2682
2683static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2684{
2685    return (kmem_bufctl_t *) (slabp + 1);
2686}
2687
2688static void cache_init_objs(struct kmem_cache *cachep,
2689                struct slab *slabp)
2690{
2691    int i;
2692
2693    for (i = 0; i < cachep->num; i++) {
2694        void *objp = index_to_obj(cachep, slabp, i);
2695#if DEBUG
2696        /* need to poison the objs? */
2697        if (cachep->flags & SLAB_POISON)
2698            poison_obj(cachep, objp, POISON_FREE);
2699        if (cachep->flags & SLAB_STORE_USER)
2700            *dbg_userword(cachep, objp) = NULL;
2701
2702        if (cachep->flags & SLAB_RED_ZONE) {
2703            *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2704            *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2705        }
2706        /*
2707         * Constructors are not allowed to allocate memory from the same
2708         * cache which they are a constructor for. Otherwise, deadlock.
2709         * They must also be threaded.
2710         */
2711        if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2712            cachep->ctor(objp + obj_offset(cachep));
2713
2714        if (cachep->flags & SLAB_RED_ZONE) {
2715            if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2716                slab_error(cachep, "constructor overwrote the"
2717                       " end of an object");
2718            if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2719                slab_error(cachep, "constructor overwrote the"
2720                       " start of an object");
2721        }
2722        if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2723                OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2724            kernel_map_pages(virt_to_page(objp),
2725                     cachep->buffer_size / PAGE_SIZE, 0);
2726#else
2727        if (cachep->ctor)
2728            cachep->ctor(objp);
2729#endif
2730        slab_bufctl(slabp)[i] = i + 1;
2731    }
2732    slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2733}
2734
2735static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2736{
2737    if (CONFIG_ZONE_DMA_FLAG) {
2738        if (flags & GFP_DMA)
2739            BUG_ON(!(cachep->gfpflags & GFP_DMA));
2740        else
2741            BUG_ON(cachep->gfpflags & GFP_DMA);
2742    }
2743}
2744
2745static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2746                int nodeid)
2747{
2748    void *objp = index_to_obj(cachep, slabp, slabp->free);
2749    kmem_bufctl_t next;
2750
2751    slabp->inuse++;
2752    next = slab_bufctl(slabp)[slabp->free];
2753#if DEBUG
2754    slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2755    WARN_ON(slabp->nodeid != nodeid);
2756#endif
2757    slabp->free = next;
2758
2759    return objp;
2760}
2761
2762static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2763                void *objp, int nodeid)
2764{
2765    unsigned int objnr = obj_to_index(cachep, slabp, objp);
2766
2767#if DEBUG
2768    /* Verify that the slab belongs to the intended node */
2769    WARN_ON(slabp->nodeid != nodeid);
2770
2771    if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2772        printk(KERN_ERR "slab: double free detected in cache "
2773                "'%s', objp %p\n", cachep->name, objp);
2774        BUG();
2775    }
2776#endif
2777    slab_bufctl(slabp)[objnr] = slabp->free;
2778    slabp->free = objnr;
2779    slabp->inuse--;
2780}
2781
2782/*
2783 * Map pages beginning at addr to the given cache and slab. This is required
2784 * for the slab allocator to be able to lookup the cache and slab of a
2785 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2786 */
2787static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2788               void *addr)
2789{
2790    int nr_pages;
2791    struct page *page;
2792
2793    page = virt_to_page(addr);
2794
2795    nr_pages = 1;
2796    if (likely(!PageCompound(page)))
2797        nr_pages <<= cache->gfporder;
2798
2799    do {
2800        page_set_cache(page, cache);
2801        page_set_slab(page, slab);
2802        page++;
2803    } while (--nr_pages);
2804}
2805
2806/*
2807 * Grow (by 1) the number of slabs within a cache. This is called by
2808 * kmem_cache_alloc() when there are no active objs left in a cache.
2809 */
2810static int cache_grow(struct kmem_cache *cachep,
2811        gfp_t flags, int nodeid, void *objp)
2812{
2813    struct slab *slabp;
2814    size_t offset;
2815    gfp_t local_flags;
2816    struct kmem_list3 *l3;
2817
2818    /*
2819     * Be lazy and only check for valid flags here, keeping it out of the
2820     * critical path in kmem_cache_alloc().
2821     */
2822    BUG_ON(flags & GFP_SLAB_BUG_MASK);
2823    local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2824
2825    /* Take the l3 list lock to change the colour_next on this node */
2826    check_irq_off();
2827    l3 = cachep->nodelists[nodeid];
2828    spin_lock(&l3->list_lock);
2829
2830    /* Get colour for the slab, and cal the next value. */
2831    offset = l3->colour_next;
2832    l3->colour_next++;
2833    if (l3->colour_next >= cachep->colour)
2834        l3->colour_next = 0;
2835    spin_unlock(&l3->list_lock);
2836
2837    offset *= cachep->colour_off;
2838
2839    if (local_flags & __GFP_WAIT)
2840        local_irq_enable();
2841
2842    /*
2843     * The test for missing atomic flag is performed here, rather than
2844     * the more obvious place, simply to reduce the critical path length
2845     * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2846     * will eventually be caught here (where it matters).
2847     */
2848    kmem_flagcheck(cachep, flags);
2849
2850    /*
2851     * Get mem for the objs. Attempt to allocate a physical page from
2852     * 'nodeid'.
2853     */
2854    if (!objp)
2855        objp = kmem_getpages(cachep, local_flags, nodeid);
2856    if (!objp)
2857        goto failed;
2858
2859    /* Get slab management. */
2860    slabp = alloc_slabmgmt(cachep, objp, offset,
2861            local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2862    if (!slabp)
2863        goto opps1;
2864
2865    slab_map_pages(cachep, slabp, objp);
2866
2867    cache_init_objs(cachep, slabp);
2868
2869    if (local_flags & __GFP_WAIT)
2870        local_irq_disable();
2871    check_irq_off();
2872    spin_lock(&l3->list_lock);
2873
2874    /* Make slab active. */
2875    list_add_tail(&slabp->list, &(l3->slabs_free));
2876    STATS_INC_GROWN(cachep);
2877    l3->free_objects += cachep->num;
2878    spin_unlock(&l3->list_lock);
2879    return 1;
2880opps1:
2881    kmem_freepages(cachep, objp);
2882failed:
2883    if (local_flags & __GFP_WAIT)
2884        local_irq_disable();
2885    return 0;
2886}
2887
2888#if DEBUG
2889
2890/*
2891 * Perform extra freeing checks:
2892 * - detect bad pointers.
2893 * - POISON/RED_ZONE checking
2894 */
2895static void kfree_debugcheck(const void *objp)
2896{
2897    if (!virt_addr_valid(objp)) {
2898        printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2899               (unsigned long)objp);
2900        BUG();
2901    }
2902}
2903
2904static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2905{
2906    unsigned long long redzone1, redzone2;
2907
2908    redzone1 = *dbg_redzone1(cache, obj);
2909    redzone2 = *dbg_redzone2(cache, obj);
2910
2911    /*
2912     * Redzone is ok.
2913     */
2914    if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2915        return;
2916
2917    if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2918        slab_error(cache, "double free detected");
2919    else
2920        slab_error(cache, "memory outside object was overwritten");
2921
2922    printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2923            obj, redzone1, redzone2);
2924}
2925
2926static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2927                   void *caller)
2928{
2929    struct page *page;
2930    unsigned int objnr;
2931    struct slab *slabp;
2932
2933    BUG_ON(virt_to_cache(objp) != cachep);
2934
2935    objp -= obj_offset(cachep);
2936    kfree_debugcheck(objp);
2937    page = virt_to_head_page(objp);
2938
2939    slabp = page_get_slab(page);
2940
2941    if (cachep->flags & SLAB_RED_ZONE) {
2942        verify_redzone_free(cachep, objp);
2943        *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2944        *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2945    }
2946    if (cachep->flags & SLAB_STORE_USER)
2947        *dbg_userword(cachep, objp) = caller;
2948
2949    objnr = obj_to_index(cachep, slabp, objp);
2950
2951    BUG_ON(objnr >= cachep->num);
2952    BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2953
2954#ifdef CONFIG_DEBUG_SLAB_LEAK
2955    slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2956#endif
2957    if (cachep->flags & SLAB_POISON) {
2958#ifdef CONFIG_DEBUG_PAGEALLOC
2959        if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2960            store_stackinfo(cachep, objp, (unsigned long)caller);
2961            kernel_map_pages(virt_to_page(objp),
2962                     cachep->buffer_size / PAGE_SIZE, 0);
2963        } else {
2964            poison_obj(cachep, objp, POISON_FREE);
2965        }
2966#else
2967        poison_obj(cachep, objp, POISON_FREE);
2968#endif
2969    }
2970    return objp;
2971}
2972
2973static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2974{
2975    kmem_bufctl_t i;
2976    int entries = 0;
2977
2978    /* Check slab's freelist to see if this obj is there. */
2979    for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2980        entries++;
2981        if (entries > cachep->num || i >= cachep->num)
2982            goto bad;
2983    }
2984    if (entries != cachep->num - slabp->inuse) {
2985bad:
2986        printk(KERN_ERR "slab: Internal list corruption detected in "
2987                "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2988            cachep->name, cachep->num, slabp, slabp->inuse);
2989        for (i = 0;
2990             i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2991             i++) {
2992            if (i % 16 == 0)
2993                printk("\n%03x:", i);
2994            printk(" %02x", ((unsigned char *)slabp)[i]);
2995        }
2996        printk("\n");
2997        BUG();
2998    }
2999}
3000#else
3001#define kfree_debugcheck(x) do { } while(0)
3002#define cache_free_debugcheck(x,objp,z) (objp)
3003#define check_slabp(x,y) do { } while(0)
3004#endif
3005
3006static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3007{
3008    int batchcount;
3009    struct kmem_list3 *l3;
3010    struct array_cache *ac;
3011    int node;
3012
3013retry:
3014    check_irq_off();
3015    node = numa_mem_id();
3016    ac = cpu_cache_get(cachep);
3017    batchcount = ac->batchcount;
3018    if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3019        /*
3020         * If there was little recent activity on this cache, then
3021         * perform only a partial refill. Otherwise we could generate
3022         * refill bouncing.
3023         */
3024        batchcount = BATCHREFILL_LIMIT;
3025    }
3026    l3 = cachep->nodelists[node];
3027
3028    BUG_ON(ac->avail > 0 || !l3);
3029    spin_lock(&l3->list_lock);
3030
3031    /* See if we can refill from the shared array */
3032    if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3033        l3->shared->touched = 1;
3034        goto alloc_done;
3035    }
3036
3037    while (batchcount > 0) {
3038        struct list_head *entry;
3039        struct slab *slabp;
3040        /* Get slab alloc is to come from. */
3041        entry = l3->slabs_partial.next;
3042        if (entry == &l3->slabs_partial) {
3043            l3->free_touched = 1;
3044            entry = l3->slabs_free.next;
3045            if (entry == &l3->slabs_free)
3046                goto must_grow;
3047        }
3048
3049        slabp = list_entry(entry, struct slab, list);
3050        check_slabp(cachep, slabp);
3051        check_spinlock_acquired(cachep);
3052
3053        /*
3054         * The slab was either on partial or free list so
3055         * there must be at least one object available for
3056         * allocation.
3057         */
3058        BUG_ON(slabp->inuse >= cachep->num);
3059
3060        while (slabp->inuse < cachep->num && batchcount--) {
3061            STATS_INC_ALLOCED(cachep);
3062            STATS_INC_ACTIVE(cachep);
3063            STATS_SET_HIGH(cachep);
3064
3065            ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3066                                node);
3067        }
3068        check_slabp(cachep, slabp);
3069
3070        /* move slabp to correct slabp list: */
3071        list_del(&slabp->list);
3072        if (slabp->free == BUFCTL_END)
3073            list_add(&slabp->list, &l3->slabs_full);
3074        else
3075            list_add(&slabp->list, &l3->slabs_partial);
3076    }
3077
3078must_grow:
3079    l3->free_objects -= ac->avail;
3080alloc_done:
3081    spin_unlock(&l3->list_lock);
3082
3083    if (unlikely(!ac->avail)) {
3084        int x;
3085        x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3086
3087        /* cache_grow can reenable interrupts, then ac could change. */
3088        ac = cpu_cache_get(cachep);
3089        if (!x && ac->avail == 0) /* no objects in sight? abort */
3090            return NULL;
3091
3092        if (!ac->avail) /* objects refilled by interrupt? */
3093            goto retry;
3094    }
3095    ac->touched = 1;
3096    return ac->entry[--ac->avail];
3097}
3098
3099static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3100                        gfp_t flags)
3101{
3102    might_sleep_if(flags & __GFP_WAIT);
3103#if DEBUG
3104    kmem_flagcheck(cachep, flags);
3105#endif
3106}
3107
3108#if DEBUG
3109static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3110                gfp_t flags, void *objp, void *caller)
3111{
3112    if (!objp)
3113        return objp;
3114    if (cachep->flags & SLAB_POISON) {
3115#ifdef CONFIG_DEBUG_PAGEALLOC
3116        if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3117            kernel_map_pages(virt_to_page(objp),
3118                     cachep->buffer_size / PAGE_SIZE, 1);
3119        else
3120            check_poison_obj(cachep, objp);
3121#else
3122        check_poison_obj(cachep, objp);
3123#endif
3124        poison_obj(cachep, objp, POISON_INUSE);
3125    }
3126    if (cachep->flags & SLAB_STORE_USER)
3127        *dbg_userword(cachep, objp) = caller;
3128
3129    if (cachep->flags & SLAB_RED_ZONE) {
3130        if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3131                *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3132            slab_error(cachep, "double free, or memory outside"
3133                        " object was overwritten");
3134            printk(KERN_ERR
3135                "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3136                objp, *dbg_redzone1(cachep, objp),
3137                *dbg_redzone2(cachep, objp));
3138        }
3139        *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3140        *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3141    }
3142#ifdef CONFIG_DEBUG_SLAB_LEAK
3143    {
3144        struct slab *slabp;
3145        unsigned objnr;
3146
3147        slabp = page_get_slab(virt_to_head_page(objp));
3148        objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3149        slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3150    }
3151#endif
3152    objp += obj_offset(cachep);
3153    if (cachep->ctor && cachep->flags & SLAB_POISON)
3154        cachep->ctor(objp);
3155#if ARCH_SLAB_MINALIGN
3156    if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3157        printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3158               objp, ARCH_SLAB_MINALIGN);
3159    }
3160#endif
3161    return objp;
3162}
3163#else
3164#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3165#endif
3166
3167static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3168{
3169    if (cachep == &cache_cache)
3170        return false;
3171
3172    return should_failslab(obj_size(cachep), flags, cachep->flags);
3173}
3174
3175static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3176{
3177    void *objp;
3178    struct array_cache *ac;
3179
3180    check_irq_off();
3181
3182    ac = cpu_cache_get(cachep);
3183    if (likely(ac->avail)) {
3184        STATS_INC_ALLOCHIT(cachep);
3185        ac->touched = 1;
3186        objp = ac->entry[--ac->avail];
3187    } else {
3188        STATS_INC_ALLOCMISS(cachep);
3189        objp = cache_alloc_refill(cachep, flags);
3190        /*
3191         * the 'ac' may be updated by cache_alloc_refill(),
3192         * and kmemleak_erase() requires its correct value.
3193         */
3194        ac = cpu_cache_get(cachep);
3195    }
3196    /*
3197     * To avoid a false negative, if an object that is in one of the
3198     * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3199     * treat the array pointers as a reference to the object.
3200     */
3201    if (objp)
3202        kmemleak_erase(&ac->entry[ac->avail]);
3203    return objp;
3204}
3205
3206#ifdef CONFIG_NUMA
3207/*
3208 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3209 *
3210 * If we are in_interrupt, then process context, including cpusets and
3211 * mempolicy, may not apply and should not be used for allocation policy.
3212 */
3213static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3214{
3215    int nid_alloc, nid_here;
3216
3217    if (in_interrupt() || (flags & __GFP_THISNODE))
3218        return NULL;
3219    nid_alloc = nid_here = numa_mem_id();
3220    get_mems_allowed();
3221    if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3222        nid_alloc = cpuset_slab_spread_node();
3223    else if (current->mempolicy)
3224        nid_alloc = slab_node(current->mempolicy);
3225    put_mems_allowed();
3226    if (nid_alloc != nid_here)
3227        return ____cache_alloc_node(cachep, flags, nid_alloc);
3228    return NULL;
3229}
3230
3231/*
3232 * Fallback function if there was no memory available and no objects on a
3233 * certain node and fall back is permitted. First we scan all the
3234 * available nodelists for available objects. If that fails then we
3235 * perform an allocation without specifying a node. This allows the page
3236 * allocator to do its reclaim / fallback magic. We then insert the
3237 * slab into the proper nodelist and then allocate from it.
3238 */
3239static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3240{
3241    struct zonelist *zonelist;
3242    gfp_t local_flags;
3243    struct zoneref *z;
3244    struct zone *zone;
3245    enum zone_type high_zoneidx = gfp_zone(flags);
3246    void *obj = NULL;
3247    int nid;
3248
3249    if (flags & __GFP_THISNODE)
3250        return NULL;
3251
3252    get_mems_allowed();
3253    zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3254    local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3255
3256retry:
3257    /*
3258     * Look through allowed nodes for objects available
3259     * from existing per node queues.
3260     */
3261    for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3262        nid = zone_to_nid(zone);
3263
3264        if (cpuset_zone_allowed_hardwall(zone, flags) &&
3265            cache->nodelists[nid] &&
3266            cache->nodelists[nid]->free_objects) {
3267                obj = ____cache_alloc_node(cache,
3268                    flags | GFP_THISNODE, nid);
3269                if (obj)
3270                    break;
3271        }
3272    }
3273
3274    if (!obj) {
3275        /*
3276         * This allocation will be performed within the constraints
3277         * of the current cpuset / memory policy requirements.
3278         * We may trigger various forms of reclaim on the allowed
3279         * set and go into memory reserves if necessary.
3280         */
3281        if (local_flags & __GFP_WAIT)
3282            local_irq_enable();
3283        kmem_flagcheck(cache, flags);
3284        obj = kmem_getpages(cache, local_flags, numa_mem_id());
3285        if (local_flags & __GFP_WAIT)
3286            local_irq_disable();
3287        if (obj) {
3288            /*
3289             * Insert into the appropriate per node queues
3290             */
3291            nid = page_to_nid(virt_to_page(obj));
3292            if (cache_grow(cache, flags, nid, obj)) {
3293                obj = ____cache_alloc_node(cache,
3294                    flags | GFP_THISNODE, nid);
3295                if (!obj)
3296                    /*
3297                     * Another processor may allocate the
3298                     * objects in the slab since we are
3299                     * not holding any locks.
3300                     */
3301                    goto retry;
3302            } else {
3303                /* cache_grow already freed obj */
3304                obj = NULL;
3305            }
3306        }
3307    }
3308    put_mems_allowed();
3309    return obj;
3310}
3311
3312/*
3313 * A interface to enable slab creation on nodeid
3314 */
3315static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3316                int nodeid)
3317{
3318    struct list_head *entry;
3319    struct slab *slabp;
3320    struct kmem_list3 *l3;
3321    void *obj;
3322    int x;
3323
3324    l3 = cachep->nodelists[nodeid];
3325    BUG_ON(!l3);
3326
3327retry:
3328    check_irq_off();
3329    spin_lock(&l3->list_lock);
3330    entry = l3->slabs_partial.next;
3331    if (entry == &l3->slabs_partial) {
3332        l3->free_touched = 1;
3333        entry = l3->slabs_free.next;
3334        if (entry == &l3->slabs_free)
3335            goto must_grow;
3336    }
3337
3338    slabp = list_entry(entry, struct slab, list);
3339    check_spinlock_acquired_node(cachep, nodeid);
3340    check_slabp(cachep, slabp);
3341
3342    STATS_INC_NODEALLOCS(cachep);
3343    STATS_INC_ACTIVE(cachep);
3344    STATS_SET_HIGH(cachep);
3345
3346    BUG_ON(slabp->inuse == cachep->num);
3347
3348    obj = slab_get_obj(cachep, slabp, nodeid);
3349    check_slabp(cachep, slabp);
3350    l3->free_objects--;
3351    /* move slabp to correct slabp list: */
3352    list_del(&slabp->list);
3353
3354    if (slabp->free == BUFCTL_END)
3355        list_add(&slabp->list, &l3->slabs_full);
3356    else
3357        list_add(&slabp->list, &l3->slabs_partial);
3358
3359    spin_unlock(&l3->list_lock);
3360    goto done;
3361
3362must_grow:
3363    spin_unlock(&l3->list_lock);
3364    x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3365    if (x)
3366        goto retry;
3367
3368    return fallback_alloc(cachep, flags);
3369
3370done:
3371    return obj;
3372}
3373
3374/**
3375 * kmem_cache_alloc_node - Allocate an object on the specified node
3376 * @cachep: The cache to allocate from.
3377 * @flags: See kmalloc().
3378 * @nodeid: node number of the target node.
3379 * @caller: return address of caller, used for debug information
3380 *
3381 * Identical to kmem_cache_alloc but it will allocate memory on the given
3382 * node, which can improve the performance for cpu bound structures.
3383 *
3384 * Fallback to other node is possible if __GFP_THISNODE is not set.
3385 */
3386static __always_inline void *
3387__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3388           void *caller)
3389{
3390    unsigned long save_flags;
3391    void *ptr;
3392    int slab_node = numa_mem_id();
3393
3394    flags &= gfp_allowed_mask;
3395
3396    lockdep_trace_alloc(flags);
3397
3398    if (slab_should_failslab(cachep, flags))
3399        return NULL;
3400
3401    cache_alloc_debugcheck_before(cachep, flags);
3402    local_irq_save(save_flags);
3403
3404    if (nodeid == -1)
3405        nodeid = slab_node;
3406
3407    if (unlikely(!cachep->nodelists[nodeid])) {
3408        /* Node not bootstrapped yet */
3409        ptr = fallback_alloc(cachep, flags);
3410        goto out;
3411    }
3412
3413    if (nodeid == slab_node) {
3414        /*
3415         * Use the locally cached objects if possible.
3416         * However ____cache_alloc does not allow fallback
3417         * to other nodes. It may fail while we still have
3418         * objects on other nodes available.
3419         */
3420        ptr = ____cache_alloc(cachep, flags);
3421        if (ptr)
3422            goto out;
3423    }
3424    /* ___cache_alloc_node can fall back to other nodes */
3425    ptr = ____cache_alloc_node(cachep, flags, nodeid);
3426  out:
3427    local_irq_restore(save_flags);
3428    ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3429    kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3430                 flags);
3431
3432    if (likely(ptr))
3433        kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3434
3435    if (unlikely((flags & __GFP_ZERO) && ptr))
3436        memset(ptr, 0, obj_size(cachep));
3437
3438    return ptr;
3439}
3440
3441static __always_inline void *
3442__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3443{
3444    void *objp;
3445
3446    if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3447        objp = alternate_node_alloc(cache, flags);
3448        if (objp)
3449            goto out;
3450    }
3451    objp = ____cache_alloc(cache, flags);
3452
3453    /*
3454     * We may just have run out of memory on the local node.
3455     * ____cache_alloc_node() knows how to locate memory on other nodes
3456     */
3457    if (!objp)
3458        objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3459
3460  out:
3461    return objp;
3462}
3463#else
3464
3465static __always_inline void *
3466__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3467{
3468    return ____cache_alloc(cachep, flags);
3469}
3470
3471#endif /* CONFIG_NUMA */
3472
3473static __always_inline void *
3474__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3475{
3476    unsigned long save_flags;
3477    void *objp;
3478
3479    flags &= gfp_allowed_mask;
3480
3481    lockdep_trace_alloc(flags);
3482
3483    if (slab_should_failslab(cachep, flags))
3484        return NULL;
3485
3486    cache_alloc_debugcheck_before(cachep, flags);
3487    local_irq_save(save_flags);
3488    objp = __do_cache_alloc(cachep, flags);
3489    local_irq_restore(save_flags);
3490    objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3491    kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3492                 flags);
3493    prefetchw(objp);
3494
3495    if (likely(objp))
3496        kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3497
3498    if (unlikely((flags & __GFP_ZERO) && objp))
3499        memset(objp, 0, obj_size(cachep));
3500
3501    return objp;
3502}
3503
3504/*
3505 * Caller needs to acquire correct kmem_list's list_lock
3506 */
3507static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3508               int node)
3509{
3510    int i;
3511    struct kmem_list3 *l3;
3512
3513    for (i = 0; i < nr_objects; i++) {
3514        void *objp = objpp[i];
3515        struct slab *slabp;
3516
3517        slabp = virt_to_slab(objp);
3518        l3 = cachep->nodelists[node];
3519        list_del(&slabp->list);
3520        check_spinlock_acquired_node(cachep, node);
3521        check_slabp(cachep, slabp);
3522        slab_put_obj(cachep, slabp, objp, node);
3523        STATS_DEC_ACTIVE(cachep);
3524        l3->free_objects++;
3525        check_slabp(cachep, slabp);
3526
3527        /* fixup slab chains */
3528        if (slabp->inuse == 0) {
3529            if (l3->free_objects > l3->free_limit) {
3530                l3->free_objects -= cachep->num;
3531                /* No need to drop any previously held
3532                 * lock here, even if we have a off-slab slab
3533                 * descriptor it is guaranteed to come from
3534                 * a different cache, refer to comments before
3535                 * alloc_slabmgmt.
3536                 */
3537                slab_destroy(cachep, slabp);
3538            } else {
3539                list_add(&slabp->list, &l3->slabs_free);
3540            }
3541        } else {
3542            /* Unconditionally move a slab to the end of the
3543             * partial list on free - maximum time for the
3544             * other objects to be freed, too.
3545             */
3546            list_add_tail(&slabp->list, &l3->slabs_partial);
3547        }
3548    }
3549}
3550
3551static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3552{
3553    int batchcount;
3554    struct kmem_list3 *l3;
3555    int node = numa_mem_id();
3556
3557    batchcount = ac->batchcount;
3558#if DEBUG
3559    BUG_ON(!batchcount || batchcount > ac->avail);
3560#endif
3561    check_irq_off();
3562    l3 = cachep->nodelists[node];
3563    spin_lock(&l3->list_lock);
3564    if (l3->shared) {
3565        struct array_cache *shared_array = l3->shared;
3566        int max = shared_array->limit - shared_array->avail;
3567        if (max) {
3568            if (batchcount > max)
3569                batchcount = max;
3570            memcpy(&(shared_array->entry[shared_array->avail]),
3571                   ac->entry, sizeof(void *) * batchcount);
3572            shared_array->avail += batchcount;
3573            goto free_done;
3574        }
3575    }
3576
3577    free_block(cachep, ac->entry, batchcount, node);
3578free_done:
3579#if STATS
3580    {
3581        int i = 0;
3582        struct list_head *p;
3583
3584        p = l3->slabs_free.next;
3585        while (p != &(l3->slabs_free)) {
3586            struct slab *slabp;
3587
3588            slabp = list_entry(p, struct slab, list);
3589            BUG_ON(slabp->inuse);
3590
3591            i++;
3592            p = p->next;
3593        }
3594        STATS_SET_FREEABLE(cachep, i);
3595    }
3596#endif
3597    spin_unlock(&l3->list_lock);
3598    ac->avail -= batchcount;
3599    memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3600}
3601
3602/*
3603 * Release an obj back to its cache. If the obj has a constructed state, it must
3604 * be in this state _before_ it is released. Called with disabled ints.
3605 */
3606static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3607{
3608    struct array_cache *ac = cpu_cache_get(cachep);
3609
3610    check_irq_off();
3611    kmemleak_free_recursive(objp, cachep->flags);
3612    objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3613
3614    kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3615
3616    /*
3617     * Skip calling cache_free_alien() when the platform is not numa.
3618     * This will avoid cache misses that happen while accessing slabp (which
3619     * is per page memory reference) to get nodeid. Instead use a global
3620     * variable to skip the call, which is mostly likely to be present in
3621     * the cache.
3622     */
3623    if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3624        return;
3625
3626    if (likely(ac->avail < ac->limit)) {
3627        STATS_INC_FREEHIT(cachep);
3628        ac->entry[ac->avail++] = objp;
3629        return;
3630    } else {
3631        STATS_INC_FREEMISS(cachep);
3632        cache_flusharray(cachep, ac);
3633        ac->entry[ac->avail++] = objp;
3634    }
3635}
3636
3637/**
3638 * kmem_cache_alloc - Allocate an object
3639 * @cachep: The cache to allocate from.
3640 * @flags: See kmalloc().
3641 *
3642 * Allocate an object from this cache. The flags are only relevant
3643 * if the cache has no available objects.
3644 */
3645void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3646{
3647    void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3648
3649    trace_kmem_cache_alloc(_RET_IP_, ret,
3650                   obj_size(cachep), cachep->buffer_size, flags);
3651
3652    return ret;
3653}
3654EXPORT_SYMBOL(kmem_cache_alloc);
3655
3656#ifdef CONFIG_TRACING
3657void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3658{
3659    return __cache_alloc(cachep, flags, __builtin_return_address(0));
3660}
3661EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3662#endif
3663
3664/**
3665 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3666 * @cachep: the cache we're checking against
3667 * @ptr: pointer to validate
3668 *
3669 * This verifies that the untrusted pointer looks sane;
3670 * it is _not_ a guarantee that the pointer is actually
3671 * part of the slab cache in question, but it at least
3672 * validates that the pointer can be dereferenced and
3673 * looks half-way sane.
3674 *
3675 * Currently only used for dentry validation.
3676 */
3677int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3678{
3679    unsigned long size = cachep->buffer_size;
3680    struct page *page;
3681
3682    if (unlikely(!kern_ptr_validate(ptr, size)))
3683        goto out;
3684    page = virt_to_page(ptr);
3685    if (unlikely(!PageSlab(page)))
3686        goto out;
3687    if (unlikely(page_get_cache(page) != cachep))
3688        goto out;
3689    return 1;
3690out:
3691    return 0;
3692}
3693
3694#ifdef CONFIG_NUMA
3695void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3696{
3697    void *ret = __cache_alloc_node(cachep, flags, nodeid,
3698                       __builtin_return_address(0));
3699
3700    trace_kmem_cache_alloc_node(_RET_IP_, ret,
3701                    obj_size(cachep), cachep->buffer_size,
3702                    flags, nodeid);
3703
3704    return ret;
3705}
3706EXPORT_SYMBOL(kmem_cache_alloc_node);
3707
3708#ifdef CONFIG_TRACING
3709void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3710                    gfp_t flags,
3711                    int nodeid)
3712{
3713    return __cache_alloc_node(cachep, flags, nodeid,
3714                  __builtin_return_address(0));
3715}
3716EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3717#endif
3718
3719static __always_inline void *
3720__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3721{
3722    struct kmem_cache *cachep;
3723    void *ret;
3724
3725    cachep = kmem_find_general_cachep(size, flags);
3726    if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3727        return cachep;
3728    ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3729
3730    trace_kmalloc_node((unsigned long) caller, ret,
3731               size, cachep->buffer_size, flags, node);
3732
3733    return ret;
3734}
3735
3736#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3737void *__kmalloc_node(size_t size, gfp_t flags, int node)
3738{
3739    return __do_kmalloc_node(size, flags, node,
3740            __builtin_return_address(0));
3741}
3742EXPORT_SYMBOL(__kmalloc_node);
3743
3744void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3745        int node, unsigned long caller)
3746{
3747    return __do_kmalloc_node(size, flags, node, (void *)caller);
3748}
3749EXPORT_SYMBOL(__kmalloc_node_track_caller);
3750#else
3751void *__kmalloc_node(size_t size, gfp_t flags, int node)
3752{
3753    return __do_kmalloc_node(size, flags, node, NULL);
3754}
3755EXPORT_SYMBOL(__kmalloc_node);
3756#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3757#endif /* CONFIG_NUMA */
3758
3759/**
3760 * __do_kmalloc - allocate memory
3761 * @size: how many bytes of memory are required.
3762 * @flags: the type of memory to allocate (see kmalloc).
3763 * @caller: function caller for debug tracking of the caller
3764 */
3765static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3766                      void *caller)
3767{
3768    struct kmem_cache *cachep;
3769    void *ret;
3770
3771    /* If you want to save a few bytes .text space: replace
3772     * __ with kmem_.
3773     * Then kmalloc uses the uninlined functions instead of the inline
3774     * functions.
3775     */
3776    cachep = __find_general_cachep(size, flags);
3777    if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3778        return cachep;
3779    ret = __cache_alloc(cachep, flags, caller);
3780
3781    trace_kmalloc((unsigned long) caller, ret,
3782              size, cachep->buffer_size, flags);
3783
3784    return ret;
3785}
3786
3787
3788#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3789void *__kmalloc(size_t size, gfp_t flags)
3790{
3791    return __do_kmalloc(size, flags, __builtin_return_address(0));
3792}
3793EXPORT_SYMBOL(__kmalloc);
3794
3795void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3796{
3797    return __do_kmalloc(size, flags, (void *)caller);
3798}
3799EXPORT_SYMBOL(__kmalloc_track_caller);
3800
3801#else
3802void *__kmalloc(size_t size, gfp_t flags)
3803{
3804    return __do_kmalloc(size, flags, NULL);
3805}
3806EXPORT_SYMBOL(__kmalloc);
3807#endif
3808
3809/**
3810 * kmem_cache_free - Deallocate an object
3811 * @cachep: The cache the allocation was from.
3812 * @objp: The previously allocated object.
3813 *
3814 * Free an object which was previously allocated from this
3815 * cache.
3816 */
3817void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3818{
3819    unsigned long flags;
3820
3821    local_irq_save(flags);
3822    debug_check_no_locks_freed(objp, obj_size(cachep));
3823    if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3824        debug_check_no_obj_freed(objp, obj_size(cachep));
3825    __cache_free(cachep, objp);
3826    local_irq_restore(flags);
3827
3828    trace_kmem_cache_free(_RET_IP_, objp);
3829}
3830EXPORT_SYMBOL(kmem_cache_free);
3831
3832/**
3833 * kfree - free previously allocated memory
3834 * @objp: pointer returned by kmalloc.
3835 *
3836 * If @objp is NULL, no operation is performed.
3837 *
3838 * Don't free memory not originally allocated by kmalloc()
3839 * or you will run into trouble.
3840 */
3841void kfree(const void *objp)
3842{
3843    struct kmem_cache *c;
3844    unsigned long flags;
3845
3846    trace_kfree(_RET_IP_, objp);
3847
3848    if (unlikely(ZERO_OR_NULL_PTR(objp)))
3849        return;
3850    local_irq_save(flags);
3851    kfree_debugcheck(objp);
3852    c = virt_to_cache(objp);
3853    debug_check_no_locks_freed(objp, obj_size(c));
3854    debug_check_no_obj_freed(objp, obj_size(c));
3855    __cache_free(c, (void *)objp);
3856    local_irq_restore(flags);
3857}
3858EXPORT_SYMBOL(kfree);
3859
3860unsigned int kmem_cache_size(struct kmem_cache *cachep)
3861{
3862    return obj_size(cachep);
3863}
3864EXPORT_SYMBOL(kmem_cache_size);
3865
3866const char *kmem_cache_name(struct kmem_cache *cachep)
3867{
3868    return cachep->name;
3869}
3870EXPORT_SYMBOL_GPL(kmem_cache_name);
3871
3872/*
3873 * This initializes kmem_list3 or resizes various caches for all nodes.
3874 */
3875static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3876{
3877    int node;
3878    struct kmem_list3 *l3;
3879    struct array_cache *new_shared;
3880    struct array_cache **new_alien = NULL;
3881
3882    for_each_online_node(node) {
3883
3884                if (use_alien_caches) {
3885                        new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3886                        if (!new_alien)
3887                                goto fail;
3888                }
3889
3890        new_shared = NULL;
3891        if (cachep->shared) {
3892            new_shared = alloc_arraycache(node,
3893                cachep->shared*cachep->batchcount,
3894                    0xbaadf00d, gfp);
3895            if (!new_shared) {
3896                free_alien_cache(new_alien);
3897                goto fail;
3898            }
3899        }
3900
3901        l3 = cachep->nodelists[node];
3902        if (l3) {
3903            struct array_cache *shared = l3->shared;
3904
3905            spin_lock_irq(&l3->list_lock);
3906
3907            if (shared)
3908                free_block(cachep, shared->entry,
3909                        shared->avail, node);
3910
3911            l3->shared = new_shared;
3912            if (!l3->alien) {
3913                l3->alien = new_alien;
3914                new_alien = NULL;
3915            }
3916            l3->free_limit = (1 + nr_cpus_node(node)) *
3917                    cachep->batchcount + cachep->num;
3918            spin_unlock_irq(&l3->list_lock);
3919            kfree(shared);
3920            free_alien_cache(new_alien);
3921            continue;
3922        }
3923        l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3924        if (!l3) {
3925            free_alien_cache(new_alien);
3926            kfree(new_shared);
3927            goto fail;
3928        }
3929
3930        kmem_list3_init(l3);
3931        l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3932                ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3933        l3->shared = new_shared;
3934        l3->alien = new_alien;
3935        l3->free_limit = (1 + nr_cpus_node(node)) *
3936                    cachep->batchcount + cachep->num;
3937        cachep->nodelists[node] = l3;
3938    }
3939    return 0;
3940
3941fail:
3942    if (!cachep->next.next) {
3943        /* Cache is not active yet. Roll back what we did */
3944        node--;
3945        while (node >= 0) {
3946            if (cachep->nodelists[node]) {
3947                l3 = cachep->nodelists[node];
3948
3949                kfree(l3->shared);
3950                free_alien_cache(l3->alien);
3951                kfree(l3);
3952                cachep->nodelists[node] = NULL;
3953            }
3954            node--;
3955        }
3956    }
3957    return -ENOMEM;
3958}
3959
3960struct ccupdate_struct {
3961    struct kmem_cache *cachep;
3962    struct array_cache *new[NR_CPUS];
3963};
3964
3965static void do_ccupdate_local(void *info)
3966{
3967    struct ccupdate_struct *new = info;
3968    struct array_cache *old;
3969
3970    check_irq_off();
3971    old = cpu_cache_get(new->cachep);
3972
3973    new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3974    new->new[smp_processor_id()] = old;
3975}
3976
3977/* Always called with the cache_chain_mutex held */
3978static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3979                int batchcount, int shared, gfp_t gfp)
3980{
3981    struct ccupdate_struct *new;
3982    int i;
3983
3984    new = kzalloc(sizeof(*new), gfp);
3985    if (!new)
3986        return -ENOMEM;
3987
3988    for_each_online_cpu(i) {
3989        new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3990                        batchcount, gfp);
3991        if (!new->new[i]) {
3992            for (i--; i >= 0; i--)
3993                kfree(new->new[i]);
3994            kfree(new);
3995            return -ENOMEM;
3996        }
3997    }
3998    new->cachep = cachep;
3999
4000    on_each_cpu(do_ccupdate_local, (void *)new, 1);
4001
4002    check_irq_on();
4003    cachep->batchcount = batchcount;
4004    cachep->limit = limit;
4005    cachep->shared = shared;
4006
4007    for_each_online_cpu(i) {
4008        struct array_cache *ccold = new->new[i];
4009        if (!ccold)
4010            continue;
4011        spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4012        free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4013        spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4014        kfree(ccold);
4015    }
4016    kfree(new);
4017    return alloc_kmemlist(cachep, gfp);
4018}
4019
4020/* Called with cache_chain_mutex held always */
4021static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4022{
4023    int err;
4024    int limit, shared;
4025
4026    /*
4027     * The head array serves three purposes:
4028     * - create a LIFO ordering, i.e. return objects that are cache-warm
4029     * - reduce the number of spinlock operations.
4030     * - reduce the number of linked list operations on the slab and
4031     * bufctl chains: array operations are cheaper.
4032     * The numbers are guessed, we should auto-tune as described by
4033     * Bonwick.
4034     */
4035    if (cachep->buffer_size > 131072)
4036        limit = 1;
4037    else if (cachep->buffer_size > PAGE_SIZE)
4038        limit = 8;
4039    else if (cachep->buffer_size > 1024)
4040        limit = 24;
4041    else if (cachep->buffer_size > 256)
4042        limit = 54;
4043    else
4044        limit = 120;
4045
4046    /*
4047     * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4048     * allocation behaviour: Most allocs on one cpu, most free operations
4049     * on another cpu. For these cases, an efficient object passing between
4050     * cpus is necessary. This is provided by a shared array. The array
4051     * replaces Bonwick's magazine layer.
4052     * On uniprocessor, it's functionally equivalent (but less efficient)
4053     * to a larger limit. Thus disabled by default.
4054     */
4055    shared = 0;
4056    if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4057        shared = 8;
4058
4059#if DEBUG
4060    /*
4061     * With debugging enabled, large batchcount lead to excessively long
4062     * periods with disabled local interrupts. Limit the batchcount
4063     */
4064    if (limit > 32)
4065        limit = 32;
4066#endif
4067    err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4068    if (err)
4069        printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4070               cachep->name, -err);
4071    return err;
4072}
4073
4074/*
4075 * Drain an array if it contains any elements taking the l3 lock only if
4076 * necessary. Note that the l3 listlock also protects the array_cache
4077 * if drain_array() is used on the shared array.
4078 */
4079void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4080             struct array_cache *ac, int force, int node)
4081{
4082    int tofree;
4083
4084    if (!ac || !ac->avail)
4085        return;
4086    if (ac->touched && !force) {
4087        ac->touched = 0;
4088    } else {
4089        spin_lock_irq(&l3->list_lock);
4090        if (ac->avail) {
4091            tofree = force ? ac->avail : (ac->limit + 4) / 5;
4092            if (tofree > ac->avail)
4093                tofree = (ac->avail + 1) / 2;
4094            free_block(cachep, ac->entry, tofree, node);
4095            ac->avail -= tofree;
4096            memmove(ac->entry, &(ac->entry[tofree]),
4097                sizeof(void *) * ac->avail);
4098        }
4099        spin_unlock_irq(&l3->list_lock);
4100    }
4101}
4102
4103/**
4104 * cache_reap - Reclaim memory from caches.
4105 * @w: work descriptor
4106 *
4107 * Called from workqueue/eventd every few seconds.
4108 * Purpose:
4109 * - clear the per-cpu caches for this CPU.
4110 * - return freeable pages to the main free memory pool.
4111 *
4112 * If we cannot acquire the cache chain mutex then just give up - we'll try
4113 * again on the next iteration.
4114 */
4115static void cache_reap(struct work_struct *w)
4116{
4117    struct kmem_cache *searchp;
4118    struct kmem_list3 *l3;
4119    int node = numa_mem_id();
4120    struct delayed_work *work = to_delayed_work(w);
4121
4122    if (!mutex_trylock(&cache_chain_mutex))
4123        /* Give up. Setup the next iteration. */
4124        goto out;
4125
4126    list_for_each_entry(searchp, &cache_chain, next) {
4127        check_irq_on();
4128
4129        /*
4130         * We only take the l3 lock if absolutely necessary and we
4131         * have established with reasonable certainty that
4132         * we can do some work if the lock was obtained.
4133         */
4134        l3 = searchp->nodelists[node];
4135
4136        reap_alien(searchp, l3);
4137
4138        drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4139
4140        /*
4141         * These are racy checks but it does not matter
4142         * if we skip one check or scan twice.
4143         */
4144        if (time_after(l3->next_reap, jiffies))
4145            goto next;
4146
4147        l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4148
4149        drain_array(searchp, l3, l3->shared, 0, node);
4150
4151        if (l3->free_touched)
4152            l3->free_touched = 0;
4153        else {
4154            int freed;
4155
4156            freed = drain_freelist(searchp, l3, (l3->free_limit +
4157                5 * searchp->num - 1) / (5 * searchp->num));
4158            STATS_ADD_REAPED(searchp, freed);
4159        }
4160next:
4161        cond_resched();
4162    }
4163    check_irq_on();
4164    mutex_unlock(&cache_chain_mutex);
4165    next_reap_node();
4166out:
4167    /* Set up the next iteration */
4168    schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4169}
4170
4171#ifdef CONFIG_SLABINFO
4172
4173static void print_slabinfo_header(struct seq_file *m)
4174{
4175    /*
4176     * Output format version, so at least we can change it
4177     * without _too_ many complaints.
4178     */
4179#if STATS
4180    seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4181#else
4182    seq_puts(m, "slabinfo - version: 2.1\n");
4183#endif
4184    seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4185         "<objperslab> <pagesperslab>");
4186    seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4187    seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4188#if STATS
4189    seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4190         "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4191    seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4192#endif
4193    seq_putc(m, '\n');
4194}
4195
4196static void *s_start(struct seq_file *m, loff_t *pos)
4197{
4198    loff_t n = *pos;
4199
4200    mutex_lock(&cache_chain_mutex);
4201    if (!n)
4202        print_slabinfo_header(m);
4203
4204    return seq_list_start(&cache_chain, *pos);
4205}
4206
4207static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4208{
4209    return seq_list_next(p, &cache_chain, pos);
4210}
4211
4212static void s_stop(struct seq_file *m, void *p)
4213{
4214    mutex_unlock(&cache_chain_mutex);
4215}
4216
4217static int s_show(struct seq_file *m, void *p)
4218{
4219    struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4220    struct slab *slabp;
4221    unsigned long active_objs;
4222    unsigned long num_objs;
4223    unsigned long active_slabs = 0;
4224    unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4225    const char *name;
4226    char *error = NULL;
4227    int node;
4228    struct kmem_list3 *l3;
4229
4230    active_objs = 0;
4231    num_slabs = 0;
4232    for_each_online_node(node) {
4233        l3 = cachep->nodelists[node];
4234        if (!l3)
4235            continue;
4236
4237        check_irq_on();
4238        spin_lock_irq(&l3->list_lock);
4239
4240        list_for_each_entry(slabp, &l3->slabs_full, list) {
4241            if (slabp->inuse != cachep->num && !error)
4242                error = "slabs_full accounting error";
4243            active_objs += cachep->num;
4244            active_slabs++;
4245        }
4246        list_for_each_entry(slabp, &l3->slabs_partial, list) {
4247            if (slabp->inuse == cachep->num && !error)
4248                error = "slabs_partial inuse accounting error";
4249            if (!slabp->inuse && !error)
4250                error = "slabs_partial/inuse accounting error";
4251            active_objs += slabp->inuse;
4252            active_slabs++;
4253        }
4254        list_for_each_entry(slabp, &l3->slabs_free, list) {
4255            if (slabp->inuse && !error)
4256                error = "slabs_free/inuse accounting error";
4257            num_slabs++;
4258        }
4259        free_objects += l3->free_objects;
4260        if (l3->shared)
4261            shared_avail += l3->shared->avail;
4262
4263        spin_unlock_irq(&l3->list_lock);
4264    }
4265    num_slabs += active_slabs;
4266    num_objs = num_slabs * cachep->num;
4267    if (num_objs - active_objs != free_objects && !error)
4268        error = "free_objects accounting error";
4269
4270    name = cachep->name;
4271    if (error)
4272        printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4273
4274    seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4275           name, active_objs, num_objs, cachep->buffer_size,
4276           cachep->num, (1 << cachep->gfporder));
4277    seq_printf(m, " : tunables %4u %4u %4u",
4278           cachep->limit, cachep->batchcount, cachep->shared);
4279    seq_printf(m, " : slabdata %6lu %6lu %6lu",
4280           active_slabs, num_slabs, shared_avail);
4281#if STATS
4282    { /* list3 stats */
4283        unsigned long high = cachep->high_mark;
4284        unsigned long allocs = cachep->num_allocations;
4285        unsigned long grown = cachep->grown;
4286        unsigned long reaped = cachep->reaped;
4287        unsigned long errors = cachep->errors;
4288        unsigned long max_freeable = cachep->max_freeable;
4289        unsigned long node_allocs = cachep->node_allocs;
4290        unsigned long node_frees = cachep->node_frees;
4291        unsigned long overflows = cachep->node_overflow;
4292
4293        seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4294               "%4lu %4lu %4lu %4lu %4lu",
4295               allocs, high, grown,
4296               reaped, errors, max_freeable, node_allocs,
4297               node_frees, overflows);
4298    }
4299    /* cpu stats */
4300    {
4301        unsigned long allochit = atomic_read(&cachep->allochit);
4302        unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4303        unsigned long freehit = atomic_read(&cachep->freehit);
4304        unsigned long freemiss = atomic_read(&cachep->freemiss);
4305
4306        seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4307               allochit, allocmiss, freehit, freemiss);
4308    }
4309#endif
4310    seq_putc(m, '\n');
4311    return 0;
4312}
4313
4314/*
4315 * slabinfo_op - iterator that generates /proc/slabinfo
4316 *
4317 * Output layout:
4318 * cache-name
4319 * num-active-objs
4320 * total-objs
4321 * object size
4322 * num-active-slabs
4323 * total-slabs
4324 * num-pages-per-slab
4325 * + further values on SMP and with statistics enabled
4326 */
4327
4328static const struct seq_operations slabinfo_op = {
4329    .start = s_start,
4330    .next = s_next,
4331    .stop = s_stop,
4332    .show = s_show,
4333};
4334
4335#define MAX_SLABINFO_WRITE 128
4336/**
4337 * slabinfo_write - Tuning for the slab allocator
4338 * @file: unused
4339 * @buffer: user buffer
4340 * @count: data length
4341 * @ppos: unused
4342 */
4343ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4344               size_t count, loff_t *ppos)
4345{
4346    char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4347    int limit, batchcount, shared, res;
4348    struct kmem_cache *cachep;
4349
4350    if (count > MAX_SLABINFO_WRITE)
4351        return -EINVAL;
4352    if (copy_from_user(&kbuf, buffer, count))
4353        return -EFAULT;
4354    kbuf[MAX_SLABINFO_WRITE] = '\0';
4355
4356    tmp = strchr(kbuf, ' ');
4357    if (!tmp)
4358        return -EINVAL;
4359    *tmp = '\0';
4360    tmp++;
4361    if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4362        return -EINVAL;
4363
4364    /* Find the cache in the chain of caches. */
4365    mutex_lock(&cache_chain_mutex);
4366    res = -EINVAL;
4367    list_for_each_entry(cachep, &cache_chain, next) {
4368        if (!strcmp(cachep->name, kbuf)) {
4369            if (limit < 1 || batchcount < 1 ||
4370                    batchcount > limit || shared < 0) {
4371                res = 0;
4372            } else {
4373                res = do_tune_cpucache(cachep, limit,
4374                               batchcount, shared,
4375                               GFP_KERNEL);
4376            }
4377            break;
4378        }
4379    }
4380    mutex_unlock(&cache_chain_mutex);
4381    if (res >= 0)
4382        res = count;
4383    return res;
4384}
4385
4386static int slabinfo_open(struct inode *inode, struct file *file)
4387{
4388    return seq_open(file, &slabinfo_op);
4389}
4390
4391static const struct file_operations proc_slabinfo_operations = {
4392    .open = slabinfo_open,
4393    .read = seq_read,
4394    .write = slabinfo_write,
4395    .llseek = seq_lseek,
4396    .release = seq_release,
4397};
4398
4399#ifdef CONFIG_DEBUG_SLAB_LEAK
4400
4401static void *leaks_start(struct seq_file *m, loff_t *pos)
4402{
4403    mutex_lock(&cache_chain_mutex);
4404    return seq_list_start(&cache_chain, *pos);
4405}
4406
4407static inline int add_caller(unsigned long *n, unsigned long v)
4408{
4409    unsigned long *p;
4410    int l;
4411    if (!v)
4412        return 1;
4413    l = n[1];
4414    p = n + 2;
4415    while (l) {
4416        int i = l/2;
4417        unsigned long *q = p + 2 * i;
4418        if (*q == v) {
4419            q[1]++;
4420            return 1;
4421        }
4422        if (*q > v) {
4423            l = i;
4424        } else {
4425            p = q + 2;
4426            l -= i + 1;
4427        }
4428    }
4429    if (++n[1] == n[0])
4430        return 0;
4431    memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4432    p[0] = v;
4433    p[1] = 1;
4434    return 1;
4435}
4436
4437static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4438{
4439    void *p;
4440    int i;
4441    if (n[0] == n[1])
4442        return;
4443    for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4444        if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4445            continue;
4446        if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4447            return;
4448    }
4449}
4450
4451static void show_symbol(struct seq_file *m, unsigned long address)
4452{
4453#ifdef CONFIG_KALLSYMS
4454    unsigned long offset, size;
4455    char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4456
4457    if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4458        seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4459        if (modname[0])
4460            seq_printf(m, " [%s]", modname);
4461        return;
4462    }
4463#endif
4464    seq_printf(m, "%p", (void *)address);
4465}
4466
4467static int leaks_show(struct seq_file *m, void *p)
4468{
4469    struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4470    struct slab *slabp;
4471    struct kmem_list3 *l3;
4472    const char *name;
4473    unsigned long *n = m->private;
4474    int node;
4475    int i;
4476
4477    if (!(cachep->flags & SLAB_STORE_USER))
4478        return 0;
4479    if (!(cachep->flags & SLAB_RED_ZONE))
4480        return 0;
4481
4482    /* OK, we can do it */
4483
4484    n[1] = 0;
4485
4486    for_each_online_node(node) {
4487        l3 = cachep->nodelists[node];
4488        if (!l3)
4489            continue;
4490
4491        check_irq_on();
4492        spin_lock_irq(&l3->list_lock);
4493
4494        list_for_each_entry(slabp, &l3->slabs_full, list)
4495            handle_slab(n, cachep, slabp);
4496        list_for_each_entry(slabp, &l3->slabs_partial, list)
4497            handle_slab(n, cachep, slabp);
4498        spin_unlock_irq(&l3->list_lock);
4499    }
4500    name = cachep->name;
4501    if (n[0] == n[1]) {
4502        /* Increase the buffer size */
4503        mutex_unlock(&cache_chain_mutex);
4504        m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4505        if (!m->private) {
4506            /* Too bad, we are really out */
4507            m->private = n;
4508            mutex_lock(&cache_chain_mutex);
4509            return -ENOMEM;
4510        }
4511        *(unsigned long *)m->private = n[0] * 2;
4512        kfree(n);
4513        mutex_lock(&cache_chain_mutex);
4514        /* Now make sure this entry will be retried */
4515        m->count = m->size;
4516        return 0;
4517    }
4518    for (i = 0; i < n[1]; i++) {
4519        seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4520        show_symbol(m, n[2*i+2]);
4521        seq_putc(m, '\n');
4522    }
4523
4524    return 0;
4525}
4526
4527static const struct seq_operations slabstats_op = {
4528    .start = leaks_start,
4529    .next = s_next,
4530    .stop = s_stop,
4531    .show = leaks_show,
4532};
4533
4534static int slabstats_open(struct inode *inode, struct file *file)
4535{
4536    unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4537    int ret = -ENOMEM;
4538    if (n) {
4539        ret = seq_open(file, &slabstats_op);
4540        if (!ret) {
4541            struct seq_file *m = file->private_data;
4542            *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4543            m->private = n;
4544            n = NULL;
4545        }
4546        kfree(n);
4547    }
4548    return ret;
4549}
4550
4551static const struct file_operations proc_slabstats_operations = {
4552    .open = slabstats_open,
4553    .read = seq_read,
4554    .llseek = seq_lseek,
4555    .release = seq_release_private,
4556};
4557#endif
4558
4559static int __init slab_proc_init(void)
4560{
4561    proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4562#ifdef CONFIG_DEBUG_SLAB_LEAK
4563    proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4564#endif
4565    return 0;
4566}
4567module_init(slab_proc_init);
4568#endif
4569
4570/**
4571 * ksize - get the actual amount of memory allocated for a given object
4572 * @objp: Pointer to the object
4573 *
4574 * kmalloc may internally round up allocations and return more memory
4575 * than requested. ksize() can be used to determine the actual amount of
4576 * memory allocated. The caller may use this additional memory, even though
4577 * a smaller amount of memory was initially specified with the kmalloc call.
4578 * The caller must guarantee that objp points to a valid object previously
4579 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4580 * must not be freed during the duration of the call.
4581 */
4582size_t ksize(const void *objp)
4583{
4584    BUG_ON(!objp);
4585    if (unlikely(objp == ZERO_SIZE_PTR))
4586        return 0;
4587
4588    return obj_size(virt_to_cache(objp));
4589}
4590EXPORT_SYMBOL(ksize);
4591

Archive Download this file



interactive