3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
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
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.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
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.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
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.
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.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
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.
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
68 * Further notes from the original documentation:
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()).
76 * At present, each engine can be growing a cache. This should be blocked.
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>
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.
89 #include <linux/slab.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/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/uaccess.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/fault-inject.h>
111 #include <linux/rtmutex.h>
112 #include <linux/reciprocal_div.h>
114 #include <asm/cacheflush.h>
115 #include <asm/tlbflush.h>
116 #include <asm/page.h>
119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
128 #ifdef CONFIG_DEBUG_SLAB
131 #define FORCED_DEBUG 1
135 #define FORCED_DEBUG 0
138 /* Shouldn't this be in a header file somewhere? */
139 #define BYTES_PER_WORD sizeof(void *)
141 #ifndef cache_line_size
142 #define cache_line_size() L1_CACHE_BYTES
145 #ifndef ARCH_KMALLOC_MINALIGN
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than the alignment of a 64-bit integer.
152 * ARCH_KMALLOC_MINALIGN allows that.
153 * Note that increasing this value may disable some debug features.
155 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
158 #ifndef ARCH_SLAB_MINALIGN
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
166 #define ARCH_SLAB_MINALIGN 0
169 #ifndef ARCH_KMALLOC_FLAGS
170 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173 /* Legal flag mask for kmem_cache_create(). */
175 # define CREATE_MASK (SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
182 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
191 * Bufctl's are used for linking objs within a slab
194 * This implementation relies on "struct page" for locating the cache &
195 * slab an object belongs to.
196 * This allows the bufctl structure to be small (one int), but limits
197 * the number of objects a slab (not a cache) can contain when off-slab
198 * bufctls are used. The limit is the size of the largest general cache
199 * that does not use off-slab slabs.
200 * For 32bit archs with 4 kB pages, is this 56.
201 * This is not serious, as it is only for large objects, when it is unwise
202 * to have too many per slab.
203 * Note: This limit can be raised by introducing a general cache whose size
204 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
207 typedef unsigned int kmem_bufctl_t;
208 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
209 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
210 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
211 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
226 unsigned short nodeid;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head;
247 struct kmem_cache *cachep;
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
259 * The limit is stored in the per-cpu structure to reduce the data cache
266 unsigned int batchcount;
267 unsigned int touched;
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
273 * [0] is for gcc 2.95. It should really be [].
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
281 #define BOOT_CPUCACHE_ENTRIES 1
282 struct arraycache_init {
283 struct array_cache cache;
284 void *entries[BOOT_CPUCACHE_ENTRIES];
288 * The slab lists for all objects.
291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
295 unsigned int free_limit;
296 unsigned int colour_next; /* Per-node cache coloring */
297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
300 unsigned long next_reap; /* updated without locking */
301 int free_touched; /* updated without locking */
305 * Need this for bootstrapping a per node allocator.
307 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309 #define CACHE_CACHE 0
311 #define SIZE_L3 (1 + MAX_NUMNODES)
313 static int drain_freelist(struct kmem_cache *cache,
314 struct kmem_list3 *l3, int tofree);
315 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
317 static int enable_cpucache(struct kmem_cache *cachep);
318 static void cache_reap(struct work_struct *unused);
321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
324 static __always_inline int index_of(const size_t size)
326 extern void __bad_size(void);
328 if (__builtin_constant_p(size)) {
336 #include "linux/kmalloc_sizes.h"
344 static int slab_early_init = 1;
346 #define INDEX_AC index_of(sizeof(struct arraycache_init))
347 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
349 static void kmem_list3_init(struct kmem_list3 *parent)
351 INIT_LIST_HEAD(&parent->slabs_full);
352 INIT_LIST_HEAD(&parent->slabs_partial);
353 INIT_LIST_HEAD(&parent->slabs_free);
354 parent->shared = NULL;
355 parent->alien = NULL;
356 parent->colour_next = 0;
357 spin_lock_init(&parent->list_lock);
358 parent->free_objects = 0;
359 parent->free_touched = 0;
362 #define MAKE_LIST(cachep, listp, slab, nodeid) \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
368 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
382 /* 1) per-cpu data, touched during every alloc/free */
383 struct array_cache *array[NR_CPUS];
384 /* 2) Cache tunables. Protected by cache_chain_mutex */
385 unsigned int batchcount;
389 unsigned int buffer_size;
390 u32 reciprocal_buffer_size;
391 /* 3) touched by every alloc & free from the backend */
393 unsigned int flags; /* constant flags */
394 unsigned int num; /* # of objs per slab */
396 /* 4) cache_grow/shrink */
397 /* order of pgs per slab (2^n) */
398 unsigned int gfporder;
400 /* force GFP flags, e.g. GFP_DMA */
403 size_t colour; /* cache colouring range */
404 unsigned int colour_off; /* colour offset */
405 struct kmem_cache *slabp_cache;
406 unsigned int slab_size;
407 unsigned int dflags; /* dynamic flags */
409 /* constructor func */
410 void (*ctor) (void *, struct kmem_cache *, unsigned long);
412 /* de-constructor func */
413 void (*dtor) (void *, struct kmem_cache *, unsigned long);
415 /* 5) cache creation/removal */
417 struct list_head next;
421 unsigned long num_active;
422 unsigned long num_allocations;
423 unsigned long high_mark;
425 unsigned long reaped;
426 unsigned long errors;
427 unsigned long max_freeable;
428 unsigned long node_allocs;
429 unsigned long node_frees;
430 unsigned long node_overflow;
438 * If debugging is enabled, then the allocator can add additional
439 * fields and/or padding to every object. buffer_size contains the total
440 * object size including these internal fields, the following two
441 * variables contain the offset to the user object and its size.
447 * We put nodelists[] at the end of kmem_cache, because we want to size
448 * this array to nr_node_ids slots instead of MAX_NUMNODES
449 * (see kmem_cache_init())
450 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
451 * is statically defined, so we reserve the max number of nodes.
453 struct kmem_list3 *nodelists[MAX_NUMNODES];
455 * Do not add fields after nodelists[]
459 #define CFLGS_OFF_SLAB (0x80000000UL)
460 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
462 #define BATCHREFILL_LIMIT 16
464 * Optimization question: fewer reaps means less probability for unnessary
465 * cpucache drain/refill cycles.
467 * OTOH the cpuarrays can contain lots of objects,
468 * which could lock up otherwise freeable slabs.
470 #define REAPTIMEOUT_CPUC (2*HZ)
471 #define REAPTIMEOUT_LIST3 (4*HZ)
474 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
475 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
476 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
477 #define STATS_INC_GROWN(x) ((x)->grown++)
478 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
479 #define STATS_SET_HIGH(x) \
481 if ((x)->num_active > (x)->high_mark) \
482 (x)->high_mark = (x)->num_active; \
484 #define STATS_INC_ERR(x) ((x)->errors++)
485 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
486 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
487 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
488 #define STATS_SET_FREEABLE(x, i) \
490 if ((x)->max_freeable < i) \
491 (x)->max_freeable = i; \
493 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
494 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
495 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
496 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
498 #define STATS_INC_ACTIVE(x) do { } while (0)
499 #define STATS_DEC_ACTIVE(x) do { } while (0)
500 #define STATS_INC_ALLOCED(x) do { } while (0)
501 #define STATS_INC_GROWN(x) do { } while (0)
502 #define STATS_ADD_REAPED(x,y) do { } while (0)
503 #define STATS_SET_HIGH(x) do { } while (0)
504 #define STATS_INC_ERR(x) do { } while (0)
505 #define STATS_INC_NODEALLOCS(x) do { } while (0)
506 #define STATS_INC_NODEFREES(x) do { } while (0)
507 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
508 #define STATS_SET_FREEABLE(x, i) do { } while (0)
509 #define STATS_INC_ALLOCHIT(x) do { } while (0)
510 #define STATS_INC_ALLOCMISS(x) do { } while (0)
511 #define STATS_INC_FREEHIT(x) do { } while (0)
512 #define STATS_INC_FREEMISS(x) do { } while (0)
518 * memory layout of objects:
520 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
521 * the end of an object is aligned with the end of the real
522 * allocation. Catches writes behind the end of the allocation.
523 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
525 * cachep->obj_offset: The real object.
526 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
527 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
528 * [BYTES_PER_WORD long]
530 static int obj_offset(struct kmem_cache *cachep)
532 return cachep->obj_offset;
535 static int obj_size(struct kmem_cache *cachep)
537 return cachep->obj_size;
540 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
542 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
543 return (unsigned long long*) (objp + obj_offset(cachep) -
544 sizeof(unsigned long long));
547 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
549 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
550 if (cachep->flags & SLAB_STORE_USER)
551 return (unsigned long long *)(objp + cachep->buffer_size -
552 sizeof(unsigned long long) -
554 return (unsigned long long *) (objp + cachep->buffer_size -
555 sizeof(unsigned long long));
558 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
560 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
561 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
566 #define obj_offset(x) 0
567 #define obj_size(cachep) (cachep->buffer_size)
568 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
569 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
570 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
575 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
578 #if defined(CONFIG_LARGE_ALLOCS)
579 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
580 #define MAX_GFP_ORDER 13 /* up to 32Mb */
581 #elif defined(CONFIG_MMU)
582 #define MAX_OBJ_ORDER 5 /* 32 pages */
583 #define MAX_GFP_ORDER 5 /* 32 pages */
585 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
586 #define MAX_GFP_ORDER 8 /* up to 1Mb */
590 * Do not go above this order unless 0 objects fit into the slab.
592 #define BREAK_GFP_ORDER_HI 1
593 #define BREAK_GFP_ORDER_LO 0
594 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
597 * Functions for storing/retrieving the cachep and or slab from the page
598 * allocator. These are used to find the slab an obj belongs to. With kfree(),
599 * these are used to find the cache which an obj belongs to.
601 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
603 page->lru.next = (struct list_head *)cache;
606 static inline struct kmem_cache *page_get_cache(struct page *page)
608 page = compound_head(page);
609 BUG_ON(!PageSlab(page));
610 return (struct kmem_cache *)page->lru.next;
613 static inline void page_set_slab(struct page *page, struct slab *slab)
615 page->lru.prev = (struct list_head *)slab;
618 static inline struct slab *page_get_slab(struct page *page)
620 BUG_ON(!PageSlab(page));
621 return (struct slab *)page->lru.prev;
624 static inline struct kmem_cache *virt_to_cache(const void *obj)
626 struct page *page = virt_to_head_page(obj);
627 return page_get_cache(page);
630 static inline struct slab *virt_to_slab(const void *obj)
632 struct page *page = virt_to_head_page(obj);
633 return page_get_slab(page);
636 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
639 return slab->s_mem + cache->buffer_size * idx;
643 * We want to avoid an expensive divide : (offset / cache->buffer_size)
644 * Using the fact that buffer_size is a constant for a particular cache,
645 * we can replace (offset / cache->buffer_size) by
646 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
648 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
649 const struct slab *slab, void *obj)
651 u32 offset = (obj - slab->s_mem);
652 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
656 * These are the default caches for kmalloc. Custom caches can have other sizes.
658 struct cache_sizes malloc_sizes[] = {
659 #define CACHE(x) { .cs_size = (x) },
660 #include <linux/kmalloc_sizes.h>
664 EXPORT_SYMBOL(malloc_sizes);
666 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
672 static struct cache_names __initdata cache_names[] = {
673 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
674 #include <linux/kmalloc_sizes.h>
679 static struct arraycache_init initarray_cache __initdata =
680 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
681 static struct arraycache_init initarray_generic =
682 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
684 /* internal cache of cache description objs */
685 static struct kmem_cache cache_cache = {
687 .limit = BOOT_CPUCACHE_ENTRIES,
689 .buffer_size = sizeof(struct kmem_cache),
690 .name = "kmem_cache",
693 #define BAD_ALIEN_MAGIC 0x01020304ul
695 #ifdef CONFIG_LOCKDEP
698 * Slab sometimes uses the kmalloc slabs to store the slab headers
699 * for other slabs "off slab".
700 * The locking for this is tricky in that it nests within the locks
701 * of all other slabs in a few places; to deal with this special
702 * locking we put on-slab caches into a separate lock-class.
704 * We set lock class for alien array caches which are up during init.
705 * The lock annotation will be lost if all cpus of a node goes down and
706 * then comes back up during hotplug
708 static struct lock_class_key on_slab_l3_key;
709 static struct lock_class_key on_slab_alc_key;
711 static inline void init_lock_keys(void)
715 struct cache_sizes *s = malloc_sizes;
717 while (s->cs_size != ULONG_MAX) {
719 struct array_cache **alc;
721 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
722 if (!l3 || OFF_SLAB(s->cs_cachep))
724 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
727 * FIXME: This check for BAD_ALIEN_MAGIC
728 * should go away when common slab code is taught to
729 * work even without alien caches.
730 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
731 * for alloc_alien_cache,
733 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
737 lockdep_set_class(&alc[r]->lock,
745 static inline void init_lock_keys(void)
751 * 1. Guard access to the cache-chain.
752 * 2. Protect sanity of cpu_online_map against cpu hotplug events
754 static DEFINE_MUTEX(cache_chain_mutex);
755 static struct list_head cache_chain;
758 * chicken and egg problem: delay the per-cpu array allocation
759 * until the general caches are up.
769 * used by boot code to determine if it can use slab based allocator
771 int slab_is_available(void)
773 return g_cpucache_up == FULL;
776 static DEFINE_PER_CPU(struct delayed_work, reap_work);
778 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
780 return cachep->array[smp_processor_id()];
783 static inline struct kmem_cache *__find_general_cachep(size_t size,
786 struct cache_sizes *csizep = malloc_sizes;
789 /* This happens if someone tries to call
790 * kmem_cache_create(), or __kmalloc(), before
791 * the generic caches are initialized.
793 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
795 while (size > csizep->cs_size)
799 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
800 * has cs_{dma,}cachep==NULL. Thus no special case
801 * for large kmalloc calls required.
803 #ifdef CONFIG_ZONE_DMA
804 if (unlikely(gfpflags & GFP_DMA))
805 return csizep->cs_dmacachep;
807 return csizep->cs_cachep;
810 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
812 return __find_general_cachep(size, gfpflags);
815 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
817 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
821 * Calculate the number of objects and left-over bytes for a given buffer size.
823 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
824 size_t align, int flags, size_t *left_over,
829 size_t slab_size = PAGE_SIZE << gfporder;
832 * The slab management structure can be either off the slab or
833 * on it. For the latter case, the memory allocated for a
837 * - One kmem_bufctl_t for each object
838 * - Padding to respect alignment of @align
839 * - @buffer_size bytes for each object
841 * If the slab management structure is off the slab, then the
842 * alignment will already be calculated into the size. Because
843 * the slabs are all pages aligned, the objects will be at the
844 * correct alignment when allocated.
846 if (flags & CFLGS_OFF_SLAB) {
848 nr_objs = slab_size / buffer_size;
850 if (nr_objs > SLAB_LIMIT)
851 nr_objs = SLAB_LIMIT;
854 * Ignore padding for the initial guess. The padding
855 * is at most @align-1 bytes, and @buffer_size is at
856 * least @align. In the worst case, this result will
857 * be one greater than the number of objects that fit
858 * into the memory allocation when taking the padding
861 nr_objs = (slab_size - sizeof(struct slab)) /
862 (buffer_size + sizeof(kmem_bufctl_t));
865 * This calculated number will be either the right
866 * amount, or one greater than what we want.
868 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
872 if (nr_objs > SLAB_LIMIT)
873 nr_objs = SLAB_LIMIT;
875 mgmt_size = slab_mgmt_size(nr_objs, align);
878 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
881 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
883 static void __slab_error(const char *function, struct kmem_cache *cachep,
886 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
887 function, cachep->name, msg);
892 * By default on NUMA we use alien caches to stage the freeing of
893 * objects allocated from other nodes. This causes massive memory
894 * inefficiencies when using fake NUMA setup to split memory into a
895 * large number of small nodes, so it can be disabled on the command
899 static int use_alien_caches __read_mostly = 1;
900 static int __init noaliencache_setup(char *s)
902 use_alien_caches = 0;
905 __setup("noaliencache", noaliencache_setup);
909 * Special reaping functions for NUMA systems called from cache_reap().
910 * These take care of doing round robin flushing of alien caches (containing
911 * objects freed on different nodes from which they were allocated) and the
912 * flushing of remote pcps by calling drain_node_pages.
914 static DEFINE_PER_CPU(unsigned long, reap_node);
916 static void init_reap_node(int cpu)
920 node = next_node(cpu_to_node(cpu), node_online_map);
921 if (node == MAX_NUMNODES)
922 node = first_node(node_online_map);
924 per_cpu(reap_node, cpu) = node;
927 static void next_reap_node(void)
929 int node = __get_cpu_var(reap_node);
932 * Also drain per cpu pages on remote zones
934 if (node != numa_node_id())
935 drain_node_pages(node);
937 node = next_node(node, node_online_map);
938 if (unlikely(node >= MAX_NUMNODES))
939 node = first_node(node_online_map);
940 __get_cpu_var(reap_node) = node;
944 #define init_reap_node(cpu) do { } while (0)
945 #define next_reap_node(void) do { } while (0)
949 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
950 * via the workqueue/eventd.
951 * Add the CPU number into the expiration time to minimize the possibility of
952 * the CPUs getting into lockstep and contending for the global cache chain
955 static void __devinit start_cpu_timer(int cpu)
957 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
960 * When this gets called from do_initcalls via cpucache_init(),
961 * init_workqueues() has already run, so keventd will be setup
964 if (keventd_up() && reap_work->work.func == NULL) {
966 INIT_DELAYED_WORK(reap_work, cache_reap);
967 schedule_delayed_work_on(cpu, reap_work,
968 __round_jiffies_relative(HZ, cpu));
972 static struct array_cache *alloc_arraycache(int node, int entries,
975 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
976 struct array_cache *nc = NULL;
978 nc = kmalloc_node(memsize, GFP_KERNEL, node);
982 nc->batchcount = batchcount;
984 spin_lock_init(&nc->lock);
990 * Transfer objects in one arraycache to another.
991 * Locking must be handled by the caller.
993 * Return the number of entries transferred.
995 static int transfer_objects(struct array_cache *to,
996 struct array_cache *from, unsigned int max)
998 /* Figure out how many entries to transfer */
999 int nr = min(min(from->avail, max), to->limit - to->avail);
1004 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1005 sizeof(void *) *nr);
1015 #define drain_alien_cache(cachep, alien) do { } while (0)
1016 #define reap_alien(cachep, l3) do { } while (0)
1018 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1020 return (struct array_cache **)BAD_ALIEN_MAGIC;
1023 static inline void free_alien_cache(struct array_cache **ac_ptr)
1027 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1032 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1038 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1039 gfp_t flags, int nodeid)
1044 #else /* CONFIG_NUMA */
1046 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1047 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1049 static struct array_cache **alloc_alien_cache(int node, int limit)
1051 struct array_cache **ac_ptr;
1052 int memsize = sizeof(void *) * nr_node_ids;
1057 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1060 if (i == node || !node_online(i)) {
1064 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1066 for (i--; i <= 0; i--)
1076 static void free_alien_cache(struct array_cache **ac_ptr)
1087 static void __drain_alien_cache(struct kmem_cache *cachep,
1088 struct array_cache *ac, int node)
1090 struct kmem_list3 *rl3 = cachep->nodelists[node];
1093 spin_lock(&rl3->list_lock);
1095 * Stuff objects into the remote nodes shared array first.
1096 * That way we could avoid the overhead of putting the objects
1097 * into the free lists and getting them back later.
1100 transfer_objects(rl3->shared, ac, ac->limit);
1102 free_block(cachep, ac->entry, ac->avail, node);
1104 spin_unlock(&rl3->list_lock);
1109 * Called from cache_reap() to regularly drain alien caches round robin.
1111 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1113 int node = __get_cpu_var(reap_node);
1116 struct array_cache *ac = l3->alien[node];
1118 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1119 __drain_alien_cache(cachep, ac, node);
1120 spin_unlock_irq(&ac->lock);
1125 static void drain_alien_cache(struct kmem_cache *cachep,
1126 struct array_cache **alien)
1129 struct array_cache *ac;
1130 unsigned long flags;
1132 for_each_online_node(i) {
1135 spin_lock_irqsave(&ac->lock, flags);
1136 __drain_alien_cache(cachep, ac, i);
1137 spin_unlock_irqrestore(&ac->lock, flags);
1142 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1144 struct slab *slabp = virt_to_slab(objp);
1145 int nodeid = slabp->nodeid;
1146 struct kmem_list3 *l3;
1147 struct array_cache *alien = NULL;
1150 node = numa_node_id();
1153 * Make sure we are not freeing a object from another node to the array
1154 * cache on this cpu.
1156 if (likely(slabp->nodeid == node))
1159 l3 = cachep->nodelists[node];
1160 STATS_INC_NODEFREES(cachep);
1161 if (l3->alien && l3->alien[nodeid]) {
1162 alien = l3->alien[nodeid];
1163 spin_lock(&alien->lock);
1164 if (unlikely(alien->avail == alien->limit)) {
1165 STATS_INC_ACOVERFLOW(cachep);
1166 __drain_alien_cache(cachep, alien, nodeid);
1168 alien->entry[alien->avail++] = objp;
1169 spin_unlock(&alien->lock);
1171 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1172 free_block(cachep, &objp, 1, nodeid);
1173 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1179 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1180 unsigned long action, void *hcpu)
1182 long cpu = (long)hcpu;
1183 struct kmem_cache *cachep;
1184 struct kmem_list3 *l3 = NULL;
1185 int node = cpu_to_node(cpu);
1186 int memsize = sizeof(struct kmem_list3);
1189 case CPU_LOCK_ACQUIRE:
1190 mutex_lock(&cache_chain_mutex);
1192 case CPU_UP_PREPARE:
1194 * We need to do this right in the beginning since
1195 * alloc_arraycache's are going to use this list.
1196 * kmalloc_node allows us to add the slab to the right
1197 * kmem_list3 and not this cpu's kmem_list3
1200 list_for_each_entry(cachep, &cache_chain, next) {
1202 * Set up the size64 kmemlist for cpu before we can
1203 * begin anything. Make sure some other cpu on this
1204 * node has not already allocated this
1206 if (!cachep->nodelists[node]) {
1207 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1210 kmem_list3_init(l3);
1211 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1212 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1215 * The l3s don't come and go as CPUs come and
1216 * go. cache_chain_mutex is sufficient
1219 cachep->nodelists[node] = l3;
1222 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1223 cachep->nodelists[node]->free_limit =
1224 (1 + nr_cpus_node(node)) *
1225 cachep->batchcount + cachep->num;
1226 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1230 * Now we can go ahead with allocating the shared arrays and
1233 list_for_each_entry(cachep, &cache_chain, next) {
1234 struct array_cache *nc;
1235 struct array_cache *shared = NULL;
1236 struct array_cache **alien = NULL;
1238 nc = alloc_arraycache(node, cachep->limit,
1239 cachep->batchcount);
1242 if (cachep->shared) {
1243 shared = alloc_arraycache(node,
1244 cachep->shared * cachep->batchcount,
1249 if (use_alien_caches) {
1250 alien = alloc_alien_cache(node, cachep->limit);
1254 cachep->array[cpu] = nc;
1255 l3 = cachep->nodelists[node];
1258 spin_lock_irq(&l3->list_lock);
1261 * We are serialised from CPU_DEAD or
1262 * CPU_UP_CANCELLED by the cpucontrol lock
1264 l3->shared = shared;
1273 spin_unlock_irq(&l3->list_lock);
1275 free_alien_cache(alien);
1279 start_cpu_timer(cpu);
1281 #ifdef CONFIG_HOTPLUG_CPU
1284 * Even if all the cpus of a node are down, we don't free the
1285 * kmem_list3 of any cache. This to avoid a race between
1286 * cpu_down, and a kmalloc allocation from another cpu for
1287 * memory from the node of the cpu going down. The list3
1288 * structure is usually allocated from kmem_cache_create() and
1289 * gets destroyed at kmem_cache_destroy().
1293 case CPU_UP_CANCELED:
1294 list_for_each_entry(cachep, &cache_chain, next) {
1295 struct array_cache *nc;
1296 struct array_cache *shared;
1297 struct array_cache **alien;
1300 mask = node_to_cpumask(node);
1301 /* cpu is dead; no one can alloc from it. */
1302 nc = cachep->array[cpu];
1303 cachep->array[cpu] = NULL;
1304 l3 = cachep->nodelists[node];
1307 goto free_array_cache;
1309 spin_lock_irq(&l3->list_lock);
1311 /* Free limit for this kmem_list3 */
1312 l3->free_limit -= cachep->batchcount;
1314 free_block(cachep, nc->entry, nc->avail, node);
1316 if (!cpus_empty(mask)) {
1317 spin_unlock_irq(&l3->list_lock);
1318 goto free_array_cache;
1321 shared = l3->shared;
1323 free_block(cachep, shared->entry,
1324 shared->avail, node);
1331 spin_unlock_irq(&l3->list_lock);
1335 drain_alien_cache(cachep, alien);
1336 free_alien_cache(alien);
1342 * In the previous loop, all the objects were freed to
1343 * the respective cache's slabs, now we can go ahead and
1344 * shrink each nodelist to its limit.
1346 list_for_each_entry(cachep, &cache_chain, next) {
1347 l3 = cachep->nodelists[node];
1350 drain_freelist(cachep, l3, l3->free_objects);
1353 case CPU_LOCK_RELEASE:
1354 mutex_unlock(&cache_chain_mutex);
1362 static struct notifier_block __cpuinitdata cpucache_notifier = {
1363 &cpuup_callback, NULL, 0
1367 * swap the static kmem_list3 with kmalloced memory
1369 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1372 struct kmem_list3 *ptr;
1374 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1377 local_irq_disable();
1378 memcpy(ptr, list, sizeof(struct kmem_list3));
1380 * Do not assume that spinlocks can be initialized via memcpy:
1382 spin_lock_init(&ptr->list_lock);
1384 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1385 cachep->nodelists[nodeid] = ptr;
1390 * Initialisation. Called after the page allocator have been initialised and
1391 * before smp_init().
1393 void __init kmem_cache_init(void)
1396 struct cache_sizes *sizes;
1397 struct cache_names *names;
1402 if (num_possible_nodes() == 1)
1403 use_alien_caches = 0;
1405 for (i = 0; i < NUM_INIT_LISTS; i++) {
1406 kmem_list3_init(&initkmem_list3[i]);
1407 if (i < MAX_NUMNODES)
1408 cache_cache.nodelists[i] = NULL;
1412 * Fragmentation resistance on low memory - only use bigger
1413 * page orders on machines with more than 32MB of memory.
1415 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1416 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1418 /* Bootstrap is tricky, because several objects are allocated
1419 * from caches that do not exist yet:
1420 * 1) initialize the cache_cache cache: it contains the struct
1421 * kmem_cache structures of all caches, except cache_cache itself:
1422 * cache_cache is statically allocated.
1423 * Initially an __init data area is used for the head array and the
1424 * kmem_list3 structures, it's replaced with a kmalloc allocated
1425 * array at the end of the bootstrap.
1426 * 2) Create the first kmalloc cache.
1427 * The struct kmem_cache for the new cache is allocated normally.
1428 * An __init data area is used for the head array.
1429 * 3) Create the remaining kmalloc caches, with minimally sized
1431 * 4) Replace the __init data head arrays for cache_cache and the first
1432 * kmalloc cache with kmalloc allocated arrays.
1433 * 5) Replace the __init data for kmem_list3 for cache_cache and
1434 * the other cache's with kmalloc allocated memory.
1435 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1438 node = numa_node_id();
1440 /* 1) create the cache_cache */
1441 INIT_LIST_HEAD(&cache_chain);
1442 list_add(&cache_cache.next, &cache_chain);
1443 cache_cache.colour_off = cache_line_size();
1444 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1445 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
1448 * struct kmem_cache size depends on nr_node_ids, which
1449 * can be less than MAX_NUMNODES.
1451 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1452 nr_node_ids * sizeof(struct kmem_list3 *);
1454 cache_cache.obj_size = cache_cache.buffer_size;
1456 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1458 cache_cache.reciprocal_buffer_size =
1459 reciprocal_value(cache_cache.buffer_size);
1461 for (order = 0; order < MAX_ORDER; order++) {
1462 cache_estimate(order, cache_cache.buffer_size,
1463 cache_line_size(), 0, &left_over, &cache_cache.num);
1464 if (cache_cache.num)
1467 BUG_ON(!cache_cache.num);
1468 cache_cache.gfporder = order;
1469 cache_cache.colour = left_over / cache_cache.colour_off;
1470 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1471 sizeof(struct slab), cache_line_size());
1473 /* 2+3) create the kmalloc caches */
1474 sizes = malloc_sizes;
1475 names = cache_names;
1478 * Initialize the caches that provide memory for the array cache and the
1479 * kmem_list3 structures first. Without this, further allocations will
1483 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1484 sizes[INDEX_AC].cs_size,
1485 ARCH_KMALLOC_MINALIGN,
1486 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1489 if (INDEX_AC != INDEX_L3) {
1490 sizes[INDEX_L3].cs_cachep =
1491 kmem_cache_create(names[INDEX_L3].name,
1492 sizes[INDEX_L3].cs_size,
1493 ARCH_KMALLOC_MINALIGN,
1494 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1498 slab_early_init = 0;
1500 while (sizes->cs_size != ULONG_MAX) {
1502 * For performance, all the general caches are L1 aligned.
1503 * This should be particularly beneficial on SMP boxes, as it
1504 * eliminates "false sharing".
1505 * Note for systems short on memory removing the alignment will
1506 * allow tighter packing of the smaller caches.
1508 if (!sizes->cs_cachep) {
1509 sizes->cs_cachep = kmem_cache_create(names->name,
1511 ARCH_KMALLOC_MINALIGN,
1512 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1515 #ifdef CONFIG_ZONE_DMA
1516 sizes->cs_dmacachep = kmem_cache_create(
1519 ARCH_KMALLOC_MINALIGN,
1520 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1527 /* 4) Replace the bootstrap head arrays */
1529 struct array_cache *ptr;
1531 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1533 local_irq_disable();
1534 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1535 memcpy(ptr, cpu_cache_get(&cache_cache),
1536 sizeof(struct arraycache_init));
1538 * Do not assume that spinlocks can be initialized via memcpy:
1540 spin_lock_init(&ptr->lock);
1542 cache_cache.array[smp_processor_id()] = ptr;
1545 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1547 local_irq_disable();
1548 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1549 != &initarray_generic.cache);
1550 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1551 sizeof(struct arraycache_init));
1553 * Do not assume that spinlocks can be initialized via memcpy:
1555 spin_lock_init(&ptr->lock);
1557 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1561 /* 5) Replace the bootstrap kmem_list3's */
1565 /* Replace the static kmem_list3 structures for the boot cpu */
1566 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1568 for_each_online_node(nid) {
1569 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1570 &initkmem_list3[SIZE_AC + nid], nid);
1572 if (INDEX_AC != INDEX_L3) {
1573 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1574 &initkmem_list3[SIZE_L3 + nid], nid);
1579 /* 6) resize the head arrays to their final sizes */
1581 struct kmem_cache *cachep;
1582 mutex_lock(&cache_chain_mutex);
1583 list_for_each_entry(cachep, &cache_chain, next)
1584 if (enable_cpucache(cachep))
1586 mutex_unlock(&cache_chain_mutex);
1589 /* Annotate slab for lockdep -- annotate the malloc caches */
1594 g_cpucache_up = FULL;
1597 * Register a cpu startup notifier callback that initializes
1598 * cpu_cache_get for all new cpus
1600 register_cpu_notifier(&cpucache_notifier);
1603 * The reap timers are started later, with a module init call: That part
1604 * of the kernel is not yet operational.
1608 static int __init cpucache_init(void)
1613 * Register the timers that return unneeded pages to the page allocator
1615 for_each_online_cpu(cpu)
1616 start_cpu_timer(cpu);
1619 __initcall(cpucache_init);
1622 * Interface to system's page allocator. No need to hold the cache-lock.
1624 * If we requested dmaable memory, we will get it. Even if we
1625 * did not request dmaable memory, we might get it, but that
1626 * would be relatively rare and ignorable.
1628 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1636 * Nommu uses slab's for process anonymous memory allocations, and thus
1637 * requires __GFP_COMP to properly refcount higher order allocations
1639 flags |= __GFP_COMP;
1642 flags |= cachep->gfpflags;
1644 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1648 nr_pages = (1 << cachep->gfporder);
1649 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1650 add_zone_page_state(page_zone(page),
1651 NR_SLAB_RECLAIMABLE, nr_pages);
1653 add_zone_page_state(page_zone(page),
1654 NR_SLAB_UNRECLAIMABLE, nr_pages);
1655 for (i = 0; i < nr_pages; i++)
1656 __SetPageSlab(page + i);
1657 return page_address(page);
1661 * Interface to system's page release.
1663 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1665 unsigned long i = (1 << cachep->gfporder);
1666 struct page *page = virt_to_page(addr);
1667 const unsigned long nr_freed = i;
1669 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1670 sub_zone_page_state(page_zone(page),
1671 NR_SLAB_RECLAIMABLE, nr_freed);
1673 sub_zone_page_state(page_zone(page),
1674 NR_SLAB_UNRECLAIMABLE, nr_freed);
1676 BUG_ON(!PageSlab(page));
1677 __ClearPageSlab(page);
1680 if (current->reclaim_state)
1681 current->reclaim_state->reclaimed_slab += nr_freed;
1682 free_pages((unsigned long)addr, cachep->gfporder);
1685 static void kmem_rcu_free(struct rcu_head *head)
1687 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1688 struct kmem_cache *cachep = slab_rcu->cachep;
1690 kmem_freepages(cachep, slab_rcu->addr);
1691 if (OFF_SLAB(cachep))
1692 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1697 #ifdef CONFIG_DEBUG_PAGEALLOC
1698 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1699 unsigned long caller)
1701 int size = obj_size(cachep);
1703 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1705 if (size < 5 * sizeof(unsigned long))
1708 *addr++ = 0x12345678;
1710 *addr++ = smp_processor_id();
1711 size -= 3 * sizeof(unsigned long);
1713 unsigned long *sptr = &caller;
1714 unsigned long svalue;
1716 while (!kstack_end(sptr)) {
1718 if (kernel_text_address(svalue)) {
1720 size -= sizeof(unsigned long);
1721 if (size <= sizeof(unsigned long))
1727 *addr++ = 0x87654321;
1731 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1733 int size = obj_size(cachep);
1734 addr = &((char *)addr)[obj_offset(cachep)];
1736 memset(addr, val, size);
1737 *(unsigned char *)(addr + size - 1) = POISON_END;
1740 static void dump_line(char *data, int offset, int limit)
1743 unsigned char error = 0;
1746 printk(KERN_ERR "%03x:", offset);
1747 for (i = 0; i < limit; i++) {
1748 if (data[offset + i] != POISON_FREE) {
1749 error = data[offset + i];
1752 printk(" %02x", (unsigned char)data[offset + i]);
1756 if (bad_count == 1) {
1757 error ^= POISON_FREE;
1758 if (!(error & (error - 1))) {
1759 printk(KERN_ERR "Single bit error detected. Probably "
1762 printk(KERN_ERR "Run memtest86+ or a similar memory "
1765 printk(KERN_ERR "Run a memory test tool.\n");
1774 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1779 if (cachep->flags & SLAB_RED_ZONE) {
1780 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1781 *dbg_redzone1(cachep, objp),
1782 *dbg_redzone2(cachep, objp));
1785 if (cachep->flags & SLAB_STORE_USER) {
1786 printk(KERN_ERR "Last user: [<%p>]",
1787 *dbg_userword(cachep, objp));
1788 print_symbol("(%s)",
1789 (unsigned long)*dbg_userword(cachep, objp));
1792 realobj = (char *)objp + obj_offset(cachep);
1793 size = obj_size(cachep);
1794 for (i = 0; i < size && lines; i += 16, lines--) {
1797 if (i + limit > size)
1799 dump_line(realobj, i, limit);
1803 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1809 realobj = (char *)objp + obj_offset(cachep);
1810 size = obj_size(cachep);
1812 for (i = 0; i < size; i++) {
1813 char exp = POISON_FREE;
1816 if (realobj[i] != exp) {
1822 "Slab corruption: %s start=%p, len=%d\n",
1823 cachep->name, realobj, size);
1824 print_objinfo(cachep, objp, 0);
1826 /* Hexdump the affected line */
1829 if (i + limit > size)
1831 dump_line(realobj, i, limit);
1834 /* Limit to 5 lines */
1840 /* Print some data about the neighboring objects, if they
1843 struct slab *slabp = virt_to_slab(objp);
1846 objnr = obj_to_index(cachep, slabp, objp);
1848 objp = index_to_obj(cachep, slabp, objnr - 1);
1849 realobj = (char *)objp + obj_offset(cachep);
1850 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1852 print_objinfo(cachep, objp, 2);
1854 if (objnr + 1 < cachep->num) {
1855 objp = index_to_obj(cachep, slabp, objnr + 1);
1856 realobj = (char *)objp + obj_offset(cachep);
1857 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1859 print_objinfo(cachep, objp, 2);
1867 * slab_destroy_objs - destroy a slab and its objects
1868 * @cachep: cache pointer being destroyed
1869 * @slabp: slab pointer being destroyed
1871 * Call the registered destructor for each object in a slab that is being
1874 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1877 for (i = 0; i < cachep->num; i++) {
1878 void *objp = index_to_obj(cachep, slabp, i);
1880 if (cachep->flags & SLAB_POISON) {
1881 #ifdef CONFIG_DEBUG_PAGEALLOC
1882 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1884 kernel_map_pages(virt_to_page(objp),
1885 cachep->buffer_size / PAGE_SIZE, 1);
1887 check_poison_obj(cachep, objp);
1889 check_poison_obj(cachep, objp);
1892 if (cachep->flags & SLAB_RED_ZONE) {
1893 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1894 slab_error(cachep, "start of a freed object "
1896 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1897 slab_error(cachep, "end of a freed object "
1900 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1901 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1905 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1909 for (i = 0; i < cachep->num; i++) {
1910 void *objp = index_to_obj(cachep, slabp, i);
1911 (cachep->dtor) (objp, cachep, 0);
1918 * slab_destroy - destroy and release all objects in a slab
1919 * @cachep: cache pointer being destroyed
1920 * @slabp: slab pointer being destroyed
1922 * Destroy all the objs in a slab, and release the mem back to the system.
1923 * Before calling the slab must have been unlinked from the cache. The
1924 * cache-lock is not held/needed.
1926 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1928 void *addr = slabp->s_mem - slabp->colouroff;
1930 slab_destroy_objs(cachep, slabp);
1931 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1932 struct slab_rcu *slab_rcu;
1934 slab_rcu = (struct slab_rcu *)slabp;
1935 slab_rcu->cachep = cachep;
1936 slab_rcu->addr = addr;
1937 call_rcu(&slab_rcu->head, kmem_rcu_free);
1939 kmem_freepages(cachep, addr);
1940 if (OFF_SLAB(cachep))
1941 kmem_cache_free(cachep->slabp_cache, slabp);
1946 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1947 * size of kmem_list3.
1949 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1953 for_each_online_node(node) {
1954 cachep->nodelists[node] = &initkmem_list3[index + node];
1955 cachep->nodelists[node]->next_reap = jiffies +
1957 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1961 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1964 struct kmem_list3 *l3;
1966 for_each_online_cpu(i)
1967 kfree(cachep->array[i]);
1969 /* NUMA: free the list3 structures */
1970 for_each_online_node(i) {
1971 l3 = cachep->nodelists[i];
1974 free_alien_cache(l3->alien);
1978 kmem_cache_free(&cache_cache, cachep);
1983 * calculate_slab_order - calculate size (page order) of slabs
1984 * @cachep: pointer to the cache that is being created
1985 * @size: size of objects to be created in this cache.
1986 * @align: required alignment for the objects.
1987 * @flags: slab allocation flags
1989 * Also calculates the number of objects per slab.
1991 * This could be made much more intelligent. For now, try to avoid using
1992 * high order pages for slabs. When the gfp() functions are more friendly
1993 * towards high-order requests, this should be changed.
1995 static size_t calculate_slab_order(struct kmem_cache *cachep,
1996 size_t size, size_t align, unsigned long flags)
1998 unsigned long offslab_limit;
1999 size_t left_over = 0;
2002 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
2006 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2010 if (flags & CFLGS_OFF_SLAB) {
2012 * Max number of objs-per-slab for caches which
2013 * use off-slab slabs. Needed to avoid a possible
2014 * looping condition in cache_grow().
2016 offslab_limit = size - sizeof(struct slab);
2017 offslab_limit /= sizeof(kmem_bufctl_t);
2019 if (num > offslab_limit)
2023 /* Found something acceptable - save it away */
2025 cachep->gfporder = gfporder;
2026 left_over = remainder;
2029 * A VFS-reclaimable slab tends to have most allocations
2030 * as GFP_NOFS and we really don't want to have to be allocating
2031 * higher-order pages when we are unable to shrink dcache.
2033 if (flags & SLAB_RECLAIM_ACCOUNT)
2037 * Large number of objects is good, but very large slabs are
2038 * currently bad for the gfp()s.
2040 if (gfporder >= slab_break_gfp_order)
2044 * Acceptable internal fragmentation?
2046 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2052 static int setup_cpu_cache(struct kmem_cache *cachep)
2054 if (g_cpucache_up == FULL)
2055 return enable_cpucache(cachep);
2057 if (g_cpucache_up == NONE) {
2059 * Note: the first kmem_cache_create must create the cache
2060 * that's used by kmalloc(24), otherwise the creation of
2061 * further caches will BUG().
2063 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2066 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2067 * the first cache, then we need to set up all its list3s,
2068 * otherwise the creation of further caches will BUG().
2070 set_up_list3s(cachep, SIZE_AC);
2071 if (INDEX_AC == INDEX_L3)
2072 g_cpucache_up = PARTIAL_L3;
2074 g_cpucache_up = PARTIAL_AC;
2076 cachep->array[smp_processor_id()] =
2077 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2079 if (g_cpucache_up == PARTIAL_AC) {
2080 set_up_list3s(cachep, SIZE_L3);
2081 g_cpucache_up = PARTIAL_L3;
2084 for_each_online_node(node) {
2085 cachep->nodelists[node] =
2086 kmalloc_node(sizeof(struct kmem_list3),
2088 BUG_ON(!cachep->nodelists[node]);
2089 kmem_list3_init(cachep->nodelists[node]);
2093 cachep->nodelists[numa_node_id()]->next_reap =
2094 jiffies + REAPTIMEOUT_LIST3 +
2095 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2097 cpu_cache_get(cachep)->avail = 0;
2098 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2099 cpu_cache_get(cachep)->batchcount = 1;
2100 cpu_cache_get(cachep)->touched = 0;
2101 cachep->batchcount = 1;
2102 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2107 * kmem_cache_create - Create a cache.
2108 * @name: A string which is used in /proc/slabinfo to identify this cache.
2109 * @size: The size of objects to be created in this cache.
2110 * @align: The required alignment for the objects.
2111 * @flags: SLAB flags
2112 * @ctor: A constructor for the objects.
2113 * @dtor: A destructor for the objects.
2115 * Returns a ptr to the cache on success, NULL on failure.
2116 * Cannot be called within a int, but can be interrupted.
2117 * The @ctor is run when new pages are allocated by the cache
2118 * and the @dtor is run before the pages are handed back.
2120 * @name must be valid until the cache is destroyed. This implies that
2121 * the module calling this has to destroy the cache before getting unloaded.
2125 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2126 * to catch references to uninitialised memory.
2128 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2129 * for buffer overruns.
2131 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2132 * cacheline. This can be beneficial if you're counting cycles as closely
2136 kmem_cache_create (const char *name, size_t size, size_t align,
2137 unsigned long flags,
2138 void (*ctor)(void*, struct kmem_cache *, unsigned long),
2139 void (*dtor)(void*, struct kmem_cache *, unsigned long))
2141 size_t left_over, slab_size, ralign;
2142 struct kmem_cache *cachep = NULL, *pc;
2145 * Sanity checks... these are all serious usage bugs.
2147 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2148 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
2149 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2155 * We use cache_chain_mutex to ensure a consistent view of
2156 * cpu_online_map as well. Please see cpuup_callback
2158 mutex_lock(&cache_chain_mutex);
2160 list_for_each_entry(pc, &cache_chain, next) {
2165 * This happens when the module gets unloaded and doesn't
2166 * destroy its slab cache and no-one else reuses the vmalloc
2167 * area of the module. Print a warning.
2169 res = probe_kernel_address(pc->name, tmp);
2172 "SLAB: cache with size %d has lost its name\n",
2177 if (!strcmp(pc->name, name)) {
2179 "kmem_cache_create: duplicate cache %s\n", name);
2186 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2189 * Enable redzoning and last user accounting, except for caches with
2190 * large objects, if the increased size would increase the object size
2191 * above the next power of two: caches with object sizes just above a
2192 * power of two have a significant amount of internal fragmentation.
2194 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2195 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2196 if (!(flags & SLAB_DESTROY_BY_RCU))
2197 flags |= SLAB_POISON;
2199 if (flags & SLAB_DESTROY_BY_RCU)
2200 BUG_ON(flags & SLAB_POISON);
2202 if (flags & SLAB_DESTROY_BY_RCU)
2206 * Always checks flags, a caller might be expecting debug support which
2209 BUG_ON(flags & ~CREATE_MASK);
2212 * Check that size is in terms of words. This is needed to avoid
2213 * unaligned accesses for some archs when redzoning is used, and makes
2214 * sure any on-slab bufctl's are also correctly aligned.
2216 if (size & (BYTES_PER_WORD - 1)) {
2217 size += (BYTES_PER_WORD - 1);
2218 size &= ~(BYTES_PER_WORD - 1);
2221 /* calculate the final buffer alignment: */
2223 /* 1) arch recommendation: can be overridden for debug */
2224 if (flags & SLAB_HWCACHE_ALIGN) {
2226 * Default alignment: as specified by the arch code. Except if
2227 * an object is really small, then squeeze multiple objects into
2230 ralign = cache_line_size();
2231 while (size <= ralign / 2)
2234 ralign = BYTES_PER_WORD;
2238 * Redzoning and user store require word alignment. Note this will be
2239 * overridden by architecture or caller mandated alignment if either
2240 * is greater than BYTES_PER_WORD.
2242 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2243 ralign = __alignof__(unsigned long long);
2245 /* 2) arch mandated alignment */
2246 if (ralign < ARCH_SLAB_MINALIGN) {
2247 ralign = ARCH_SLAB_MINALIGN;
2249 /* 3) caller mandated alignment */
2250 if (ralign < align) {
2253 /* disable debug if necessary */
2254 if (ralign > __alignof__(unsigned long long))
2255 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2261 /* Get cache's description obj. */
2262 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2267 cachep->obj_size = size;
2270 * Both debugging options require word-alignment which is calculated
2273 if (flags & SLAB_RED_ZONE) {
2274 /* add space for red zone words */
2275 cachep->obj_offset += sizeof(unsigned long long);
2276 size += 2 * sizeof(unsigned long long);
2278 if (flags & SLAB_STORE_USER) {
2279 /* user store requires one word storage behind the end of
2282 size += BYTES_PER_WORD;
2284 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2285 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2286 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2287 cachep->obj_offset += PAGE_SIZE - size;
2294 * Determine if the slab management is 'on' or 'off' slab.
2295 * (bootstrapping cannot cope with offslab caches so don't do
2298 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2300 * Size is large, assume best to place the slab management obj
2301 * off-slab (should allow better packing of objs).
2303 flags |= CFLGS_OFF_SLAB;
2305 size = ALIGN(size, align);
2307 left_over = calculate_slab_order(cachep, size, align, flags);
2311 "kmem_cache_create: couldn't create cache %s.\n", name);
2312 kmem_cache_free(&cache_cache, cachep);
2316 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2317 + sizeof(struct slab), align);
2320 * If the slab has been placed off-slab, and we have enough space then
2321 * move it on-slab. This is at the expense of any extra colouring.
2323 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2324 flags &= ~CFLGS_OFF_SLAB;
2325 left_over -= slab_size;
2328 if (flags & CFLGS_OFF_SLAB) {
2329 /* really off slab. No need for manual alignment */
2331 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2334 cachep->colour_off = cache_line_size();
2335 /* Offset must be a multiple of the alignment. */
2336 if (cachep->colour_off < align)
2337 cachep->colour_off = align;
2338 cachep->colour = left_over / cachep->colour_off;
2339 cachep->slab_size = slab_size;
2340 cachep->flags = flags;
2341 cachep->gfpflags = 0;
2342 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2343 cachep->gfpflags |= GFP_DMA;
2344 cachep->buffer_size = size;
2345 cachep->reciprocal_buffer_size = reciprocal_value(size);
2347 if (flags & CFLGS_OFF_SLAB) {
2348 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2350 * This is a possibility for one of the malloc_sizes caches.
2351 * But since we go off slab only for object size greater than
2352 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2353 * this should not happen at all.
2354 * But leave a BUG_ON for some lucky dude.
2356 BUG_ON(!cachep->slabp_cache);
2358 cachep->ctor = ctor;
2359 cachep->dtor = dtor;
2360 cachep->name = name;
2362 if (setup_cpu_cache(cachep)) {
2363 __kmem_cache_destroy(cachep);
2368 /* cache setup completed, link it into the list */
2369 list_add(&cachep->next, &cache_chain);
2371 if (!cachep && (flags & SLAB_PANIC))
2372 panic("kmem_cache_create(): failed to create slab `%s'\n",
2374 mutex_unlock(&cache_chain_mutex);
2377 EXPORT_SYMBOL(kmem_cache_create);
2380 static void check_irq_off(void)
2382 BUG_ON(!irqs_disabled());
2385 static void check_irq_on(void)
2387 BUG_ON(irqs_disabled());
2390 static void check_spinlock_acquired(struct kmem_cache *cachep)
2394 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2398 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2402 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2407 #define check_irq_off() do { } while(0)
2408 #define check_irq_on() do { } while(0)
2409 #define check_spinlock_acquired(x) do { } while(0)
2410 #define check_spinlock_acquired_node(x, y) do { } while(0)
2413 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2414 struct array_cache *ac,
2415 int force, int node);
2417 static void do_drain(void *arg)
2419 struct kmem_cache *cachep = arg;
2420 struct array_cache *ac;
2421 int node = numa_node_id();
2424 ac = cpu_cache_get(cachep);
2425 spin_lock(&cachep->nodelists[node]->list_lock);
2426 free_block(cachep, ac->entry, ac->avail, node);
2427 spin_unlock(&cachep->nodelists[node]->list_lock);
2431 static void drain_cpu_caches(struct kmem_cache *cachep)
2433 struct kmem_list3 *l3;
2436 on_each_cpu(do_drain, cachep, 1, 1);
2438 for_each_online_node(node) {
2439 l3 = cachep->nodelists[node];
2440 if (l3 && l3->alien)
2441 drain_alien_cache(cachep, l3->alien);
2444 for_each_online_node(node) {
2445 l3 = cachep->nodelists[node];
2447 drain_array(cachep, l3, l3->shared, 1, node);
2452 * Remove slabs from the list of free slabs.
2453 * Specify the number of slabs to drain in tofree.
2455 * Returns the actual number of slabs released.
2457 static int drain_freelist(struct kmem_cache *cache,
2458 struct kmem_list3 *l3, int tofree)
2460 struct list_head *p;
2465 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2467 spin_lock_irq(&l3->list_lock);
2468 p = l3->slabs_free.prev;
2469 if (p == &l3->slabs_free) {
2470 spin_unlock_irq(&l3->list_lock);
2474 slabp = list_entry(p, struct slab, list);
2476 BUG_ON(slabp->inuse);
2478 list_del(&slabp->list);
2480 * Safe to drop the lock. The slab is no longer linked
2483 l3->free_objects -= cache->num;
2484 spin_unlock_irq(&l3->list_lock);
2485 slab_destroy(cache, slabp);
2492 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2493 static int __cache_shrink(struct kmem_cache *cachep)
2496 struct kmem_list3 *l3;
2498 drain_cpu_caches(cachep);
2501 for_each_online_node(i) {
2502 l3 = cachep->nodelists[i];
2506 drain_freelist(cachep, l3, l3->free_objects);
2508 ret += !list_empty(&l3->slabs_full) ||
2509 !list_empty(&l3->slabs_partial);
2511 return (ret ? 1 : 0);
2515 * kmem_cache_shrink - Shrink a cache.
2516 * @cachep: The cache to shrink.
2518 * Releases as many slabs as possible for a cache.
2519 * To help debugging, a zero exit status indicates all slabs were released.
2521 int kmem_cache_shrink(struct kmem_cache *cachep)
2524 BUG_ON(!cachep || in_interrupt());
2526 mutex_lock(&cache_chain_mutex);
2527 ret = __cache_shrink(cachep);
2528 mutex_unlock(&cache_chain_mutex);
2531 EXPORT_SYMBOL(kmem_cache_shrink);
2534 * kmem_cache_destroy - delete a cache
2535 * @cachep: the cache to destroy
2537 * Remove a &struct kmem_cache object from the slab cache.
2539 * It is expected this function will be called by a module when it is
2540 * unloaded. This will remove the cache completely, and avoid a duplicate
2541 * cache being allocated each time a module is loaded and unloaded, if the
2542 * module doesn't have persistent in-kernel storage across loads and unloads.
2544 * The cache must be empty before calling this function.
2546 * The caller must guarantee that noone will allocate memory from the cache
2547 * during the kmem_cache_destroy().
2549 void kmem_cache_destroy(struct kmem_cache *cachep)
2551 BUG_ON(!cachep || in_interrupt());
2553 /* Find the cache in the chain of caches. */
2554 mutex_lock(&cache_chain_mutex);
2556 * the chain is never empty, cache_cache is never destroyed
2558 list_del(&cachep->next);
2559 if (__cache_shrink(cachep)) {
2560 slab_error(cachep, "Can't free all objects");
2561 list_add(&cachep->next, &cache_chain);
2562 mutex_unlock(&cache_chain_mutex);
2566 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2569 __kmem_cache_destroy(cachep);
2570 mutex_unlock(&cache_chain_mutex);
2572 EXPORT_SYMBOL(kmem_cache_destroy);
2575 * Get the memory for a slab management obj.
2576 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2577 * always come from malloc_sizes caches. The slab descriptor cannot
2578 * come from the same cache which is getting created because,
2579 * when we are searching for an appropriate cache for these
2580 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2581 * If we are creating a malloc_sizes cache here it would not be visible to
2582 * kmem_find_general_cachep till the initialization is complete.
2583 * Hence we cannot have slabp_cache same as the original cache.
2585 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2586 int colour_off, gfp_t local_flags,
2591 if (OFF_SLAB(cachep)) {
2592 /* Slab management obj is off-slab. */
2593 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2594 local_flags & ~GFP_THISNODE, nodeid);
2598 slabp = objp + colour_off;
2599 colour_off += cachep->slab_size;
2602 slabp->colouroff = colour_off;
2603 slabp->s_mem = objp + colour_off;
2604 slabp->nodeid = nodeid;
2608 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2610 return (kmem_bufctl_t *) (slabp + 1);
2613 static void cache_init_objs(struct kmem_cache *cachep,
2614 struct slab *slabp, unsigned long ctor_flags)
2618 for (i = 0; i < cachep->num; i++) {
2619 void *objp = index_to_obj(cachep, slabp, i);
2621 /* need to poison the objs? */
2622 if (cachep->flags & SLAB_POISON)
2623 poison_obj(cachep, objp, POISON_FREE);
2624 if (cachep->flags & SLAB_STORE_USER)
2625 *dbg_userword(cachep, objp) = NULL;
2627 if (cachep->flags & SLAB_RED_ZONE) {
2628 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2629 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2632 * Constructors are not allowed to allocate memory from the same
2633 * cache which they are a constructor for. Otherwise, deadlock.
2634 * They must also be threaded.
2636 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2637 cachep->ctor(objp + obj_offset(cachep), cachep,
2640 if (cachep->flags & SLAB_RED_ZONE) {
2641 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2642 slab_error(cachep, "constructor overwrote the"
2643 " end of an object");
2644 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2645 slab_error(cachep, "constructor overwrote the"
2646 " start of an object");
2648 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2649 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2650 kernel_map_pages(virt_to_page(objp),
2651 cachep->buffer_size / PAGE_SIZE, 0);
2654 cachep->ctor(objp, cachep, ctor_flags);
2656 slab_bufctl(slabp)[i] = i + 1;
2658 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2662 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2664 if (CONFIG_ZONE_DMA_FLAG) {
2665 if (flags & GFP_DMA)
2666 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2668 BUG_ON(cachep->gfpflags & GFP_DMA);
2672 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2675 void *objp = index_to_obj(cachep, slabp, slabp->free);
2679 next = slab_bufctl(slabp)[slabp->free];
2681 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2682 WARN_ON(slabp->nodeid != nodeid);
2689 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2690 void *objp, int nodeid)
2692 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2695 /* Verify that the slab belongs to the intended node */
2696 WARN_ON(slabp->nodeid != nodeid);
2698 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2699 printk(KERN_ERR "slab: double free detected in cache "
2700 "'%s', objp %p\n", cachep->name, objp);
2704 slab_bufctl(slabp)[objnr] = slabp->free;
2705 slabp->free = objnr;
2710 * Map pages beginning at addr to the given cache and slab. This is required
2711 * for the slab allocator to be able to lookup the cache and slab of a
2712 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2714 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2720 page = virt_to_page(addr);
2723 if (likely(!PageCompound(page)))
2724 nr_pages <<= cache->gfporder;
2727 page_set_cache(page, cache);
2728 page_set_slab(page, slab);
2730 } while (--nr_pages);
2734 * Grow (by 1) the number of slabs within a cache. This is called by
2735 * kmem_cache_alloc() when there are no active objs left in a cache.
2737 static int cache_grow(struct kmem_cache *cachep,
2738 gfp_t flags, int nodeid, void *objp)
2743 unsigned long ctor_flags;
2744 struct kmem_list3 *l3;
2747 * Be lazy and only check for valid flags here, keeping it out of the
2748 * critical path in kmem_cache_alloc().
2750 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
2752 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2753 local_flags = (flags & GFP_LEVEL_MASK);
2754 /* Take the l3 list lock to change the colour_next on this node */
2756 l3 = cachep->nodelists[nodeid];
2757 spin_lock(&l3->list_lock);
2759 /* Get colour for the slab, and cal the next value. */
2760 offset = l3->colour_next;
2762 if (l3->colour_next >= cachep->colour)
2763 l3->colour_next = 0;
2764 spin_unlock(&l3->list_lock);
2766 offset *= cachep->colour_off;
2768 if (local_flags & __GFP_WAIT)
2772 * The test for missing atomic flag is performed here, rather than
2773 * the more obvious place, simply to reduce the critical path length
2774 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2775 * will eventually be caught here (where it matters).
2777 kmem_flagcheck(cachep, flags);
2780 * Get mem for the objs. Attempt to allocate a physical page from
2784 objp = kmem_getpages(cachep, flags, nodeid);
2788 /* Get slab management. */
2789 slabp = alloc_slabmgmt(cachep, objp, offset,
2790 local_flags & ~GFP_THISNODE, nodeid);
2794 slabp->nodeid = nodeid;
2795 slab_map_pages(cachep, slabp, objp);
2797 cache_init_objs(cachep, slabp, ctor_flags);
2799 if (local_flags & __GFP_WAIT)
2800 local_irq_disable();
2802 spin_lock(&l3->list_lock);
2804 /* Make slab active. */
2805 list_add_tail(&slabp->list, &(l3->slabs_free));
2806 STATS_INC_GROWN(cachep);
2807 l3->free_objects += cachep->num;
2808 spin_unlock(&l3->list_lock);
2811 kmem_freepages(cachep, objp);
2813 if (local_flags & __GFP_WAIT)
2814 local_irq_disable();
2821 * Perform extra freeing checks:
2822 * - detect bad pointers.
2823 * - POISON/RED_ZONE checking
2824 * - destructor calls, for caches with POISON+dtor
2826 static void kfree_debugcheck(const void *objp)
2828 if (!virt_addr_valid(objp)) {
2829 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2830 (unsigned long)objp);
2835 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2837 unsigned long long redzone1, redzone2;
2839 redzone1 = *dbg_redzone1(cache, obj);
2840 redzone2 = *dbg_redzone2(cache, obj);
2845 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2848 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2849 slab_error(cache, "double free detected");
2851 slab_error(cache, "memory outside object was overwritten");
2853 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2854 obj, redzone1, redzone2);
2857 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2864 objp -= obj_offset(cachep);
2865 kfree_debugcheck(objp);
2866 page = virt_to_head_page(objp);
2868 slabp = page_get_slab(page);
2870 if (cachep->flags & SLAB_RED_ZONE) {
2871 verify_redzone_free(cachep, objp);
2872 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2873 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2875 if (cachep->flags & SLAB_STORE_USER)
2876 *dbg_userword(cachep, objp) = caller;
2878 objnr = obj_to_index(cachep, slabp, objp);
2880 BUG_ON(objnr >= cachep->num);
2881 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2883 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2884 /* we want to cache poison the object,
2885 * call the destruction callback
2887 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2889 #ifdef CONFIG_DEBUG_SLAB_LEAK
2890 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2892 if (cachep->flags & SLAB_POISON) {
2893 #ifdef CONFIG_DEBUG_PAGEALLOC
2894 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2895 store_stackinfo(cachep, objp, (unsigned long)caller);
2896 kernel_map_pages(virt_to_page(objp),
2897 cachep->buffer_size / PAGE_SIZE, 0);
2899 poison_obj(cachep, objp, POISON_FREE);
2902 poison_obj(cachep, objp, POISON_FREE);
2908 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2913 /* Check slab's freelist to see if this obj is there. */
2914 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2916 if (entries > cachep->num || i >= cachep->num)
2919 if (entries != cachep->num - slabp->inuse) {
2921 printk(KERN_ERR "slab: Internal list corruption detected in "
2922 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2923 cachep->name, cachep->num, slabp, slabp->inuse);
2925 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2928 printk("\n%03x:", i);
2929 printk(" %02x", ((unsigned char *)slabp)[i]);
2936 #define kfree_debugcheck(x) do { } while(0)
2937 #define cache_free_debugcheck(x,objp,z) (objp)
2938 #define check_slabp(x,y) do { } while(0)
2941 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2944 struct kmem_list3 *l3;
2945 struct array_cache *ac;
2948 node = numa_node_id();
2951 ac = cpu_cache_get(cachep);
2953 batchcount = ac->batchcount;
2954 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2956 * If there was little recent activity on this cache, then
2957 * perform only a partial refill. Otherwise we could generate
2960 batchcount = BATCHREFILL_LIMIT;
2962 l3 = cachep->nodelists[node];
2964 BUG_ON(ac->avail > 0 || !l3);
2965 spin_lock(&l3->list_lock);
2967 /* See if we can refill from the shared array */
2968 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2971 while (batchcount > 0) {
2972 struct list_head *entry;
2974 /* Get slab alloc is to come from. */
2975 entry = l3->slabs_partial.next;
2976 if (entry == &l3->slabs_partial) {
2977 l3->free_touched = 1;
2978 entry = l3->slabs_free.next;
2979 if (entry == &l3->slabs_free)
2983 slabp = list_entry(entry, struct slab, list);
2984 check_slabp(cachep, slabp);
2985 check_spinlock_acquired(cachep);
2988 * The slab was either on partial or free list so
2989 * there must be at least one object available for
2992 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
2994 while (slabp->inuse < cachep->num && batchcount--) {
2995 STATS_INC_ALLOCED(cachep);
2996 STATS_INC_ACTIVE(cachep);
2997 STATS_SET_HIGH(cachep);
2999 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3002 check_slabp(cachep, slabp);
3004 /* move slabp to correct slabp list: */
3005 list_del(&slabp->list);
3006 if (slabp->free == BUFCTL_END)
3007 list_add(&slabp->list, &l3->slabs_full);
3009 list_add(&slabp->list, &l3->slabs_partial);
3013 l3->free_objects -= ac->avail;
3015 spin_unlock(&l3->list_lock);
3017 if (unlikely(!ac->avail)) {
3019 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3021 /* cache_grow can reenable interrupts, then ac could change. */
3022 ac = cpu_cache_get(cachep);
3023 if (!x && ac->avail == 0) /* no objects in sight? abort */
3026 if (!ac->avail) /* objects refilled by interrupt? */
3030 return ac->entry[--ac->avail];
3033 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3036 might_sleep_if(flags & __GFP_WAIT);
3038 kmem_flagcheck(cachep, flags);
3043 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3044 gfp_t flags, void *objp, void *caller)
3048 if (cachep->flags & SLAB_POISON) {
3049 #ifdef CONFIG_DEBUG_PAGEALLOC
3050 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3051 kernel_map_pages(virt_to_page(objp),
3052 cachep->buffer_size / PAGE_SIZE, 1);
3054 check_poison_obj(cachep, objp);
3056 check_poison_obj(cachep, objp);
3058 poison_obj(cachep, objp, POISON_INUSE);
3060 if (cachep->flags & SLAB_STORE_USER)
3061 *dbg_userword(cachep, objp) = caller;
3063 if (cachep->flags & SLAB_RED_ZONE) {
3064 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3065 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3066 slab_error(cachep, "double free, or memory outside"
3067 " object was overwritten");
3069 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3070 objp, *dbg_redzone1(cachep, objp),
3071 *dbg_redzone2(cachep, objp));
3073 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3074 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3076 #ifdef CONFIG_DEBUG_SLAB_LEAK
3081 slabp = page_get_slab(virt_to_head_page(objp));
3082 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3083 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3086 objp += obj_offset(cachep);
3087 if (cachep->ctor && cachep->flags & SLAB_POISON)
3088 cachep->ctor(objp, cachep, SLAB_CTOR_CONSTRUCTOR);
3089 #if ARCH_SLAB_MINALIGN
3090 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3091 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3092 objp, ARCH_SLAB_MINALIGN);
3098 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3101 #ifdef CONFIG_FAILSLAB
3103 static struct failslab_attr {
3105 struct fault_attr attr;
3107 u32 ignore_gfp_wait;
3108 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3109 struct dentry *ignore_gfp_wait_file;
3113 .attr = FAULT_ATTR_INITIALIZER,
3114 .ignore_gfp_wait = 1,
3117 static int __init setup_failslab(char *str)
3119 return setup_fault_attr(&failslab.attr, str);
3121 __setup("failslab=", setup_failslab);
3123 static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3125 if (cachep == &cache_cache)
3127 if (flags & __GFP_NOFAIL)
3129 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3132 return should_fail(&failslab.attr, obj_size(cachep));
3135 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3137 static int __init failslab_debugfs(void)
3139 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3143 err = init_fault_attr_dentries(&failslab.attr, "failslab");
3146 dir = failslab.attr.dentries.dir;
3148 failslab.ignore_gfp_wait_file =
3149 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3150 &failslab.ignore_gfp_wait);
3152 if (!failslab.ignore_gfp_wait_file) {
3154 debugfs_remove(failslab.ignore_gfp_wait_file);
3155 cleanup_fault_attr_dentries(&failslab.attr);
3161 late_initcall(failslab_debugfs);
3163 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3165 #else /* CONFIG_FAILSLAB */
3167 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3172 #endif /* CONFIG_FAILSLAB */
3174 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3177 struct array_cache *ac;
3181 ac = cpu_cache_get(cachep);
3182 if (likely(ac->avail)) {
3183 STATS_INC_ALLOCHIT(cachep);
3185 objp = ac->entry[--ac->avail];
3187 STATS_INC_ALLOCMISS(cachep);
3188 objp = cache_alloc_refill(cachep, flags);
3195 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3197 * If we are in_interrupt, then process context, including cpusets and
3198 * mempolicy, may not apply and should not be used for allocation policy.
3200 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3202 int nid_alloc, nid_here;
3204 if (in_interrupt() || (flags & __GFP_THISNODE))
3206 nid_alloc = nid_here = numa_node_id();
3207 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3208 nid_alloc = cpuset_mem_spread_node();
3209 else if (current->mempolicy)
3210 nid_alloc = slab_node(current->mempolicy);
3211 if (nid_alloc != nid_here)
3212 return ____cache_alloc_node(cachep, flags, nid_alloc);
3217 * Fallback function if there was no memory available and no objects on a
3218 * certain node and fall back is permitted. First we scan all the
3219 * available nodelists for available objects. If that fails then we
3220 * perform an allocation without specifying a node. This allows the page
3221 * allocator to do its reclaim / fallback magic. We then insert the
3222 * slab into the proper nodelist and then allocate from it.
3224 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3226 struct zonelist *zonelist;
3232 if (flags & __GFP_THISNODE)
3235 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3236 ->node_zonelists[gfp_zone(flags)];
3237 local_flags = (flags & GFP_LEVEL_MASK);
3241 * Look through allowed nodes for objects available
3242 * from existing per node queues.
3244 for (z = zonelist->zones; *z && !obj; z++) {
3245 nid = zone_to_nid(*z);
3247 if (cpuset_zone_allowed_hardwall(*z, flags) &&
3248 cache->nodelists[nid] &&
3249 cache->nodelists[nid]->free_objects)
3250 obj = ____cache_alloc_node(cache,
3251 flags | GFP_THISNODE, nid);
3256 * This allocation will be performed within the constraints
3257 * of the current cpuset / memory policy requirements.
3258 * We may trigger various forms of reclaim on the allowed
3259 * set and go into memory reserves if necessary.
3261 if (local_flags & __GFP_WAIT)
3263 kmem_flagcheck(cache, flags);
3264 obj = kmem_getpages(cache, flags, -1);
3265 if (local_flags & __GFP_WAIT)
3266 local_irq_disable();
3269 * Insert into the appropriate per node queues
3271 nid = page_to_nid(virt_to_page(obj));
3272 if (cache_grow(cache, flags, nid, obj)) {
3273 obj = ____cache_alloc_node(cache,
3274 flags | GFP_THISNODE, nid);
3277 * Another processor may allocate the
3278 * objects in the slab since we are
3279 * not holding any locks.
3283 /* cache_grow already freed obj */
3292 * A interface to enable slab creation on nodeid
3294 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3297 struct list_head *entry;
3299 struct kmem_list3 *l3;
3303 l3 = cachep->nodelists[nodeid];
3308 spin_lock(&l3->list_lock);
3309 entry = l3->slabs_partial.next;
3310 if (entry == &l3->slabs_partial) {
3311 l3->free_touched = 1;
3312 entry = l3->slabs_free.next;
3313 if (entry == &l3->slabs_free)
3317 slabp = list_entry(entry, struct slab, list);
3318 check_spinlock_acquired_node(cachep, nodeid);
3319 check_slabp(cachep, slabp);
3321 STATS_INC_NODEALLOCS(cachep);
3322 STATS_INC_ACTIVE(cachep);
3323 STATS_SET_HIGH(cachep);
3325 BUG_ON(slabp->inuse == cachep->num);
3327 obj = slab_get_obj(cachep, slabp, nodeid);
3328 check_slabp(cachep, slabp);
3330 /* move slabp to correct slabp list: */
3331 list_del(&slabp->list);
3333 if (slabp->free == BUFCTL_END)
3334 list_add(&slabp->list, &l3->slabs_full);
3336 list_add(&slabp->list, &l3->slabs_partial);
3338 spin_unlock(&l3->list_lock);
3342 spin_unlock(&l3->list_lock);
3343 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3347 return fallback_alloc(cachep, flags);
3354 * kmem_cache_alloc_node - Allocate an object on the specified node
3355 * @cachep: The cache to allocate from.
3356 * @flags: See kmalloc().
3357 * @nodeid: node number of the target node.
3358 * @caller: return address of caller, used for debug information
3360 * Identical to kmem_cache_alloc but it will allocate memory on the given
3361 * node, which can improve the performance for cpu bound structures.
3363 * Fallback to other node is possible if __GFP_THISNODE is not set.
3365 static __always_inline void *
3366 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3369 unsigned long save_flags;
3372 if (should_failslab(cachep, flags))
3375 cache_alloc_debugcheck_before(cachep, flags);
3376 local_irq_save(save_flags);
3378 if (unlikely(nodeid == -1))
3379 nodeid = numa_node_id();
3381 if (unlikely(!cachep->nodelists[nodeid])) {
3382 /* Node not bootstrapped yet */
3383 ptr = fallback_alloc(cachep, flags);
3387 if (nodeid == numa_node_id()) {
3389 * Use the locally cached objects if possible.
3390 * However ____cache_alloc does not allow fallback
3391 * to other nodes. It may fail while we still have
3392 * objects on other nodes available.
3394 ptr = ____cache_alloc(cachep, flags);
3398 /* ___cache_alloc_node can fall back to other nodes */
3399 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3401 local_irq_restore(save_flags);
3402 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3407 static __always_inline void *
3408 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3412 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3413 objp = alternate_node_alloc(cache, flags);
3417 objp = ____cache_alloc(cache, flags);
3420 * We may just have run out of memory on the local node.
3421 * ____cache_alloc_node() knows how to locate memory on other nodes
3424 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3431 static __always_inline void *
3432 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3434 return ____cache_alloc(cachep, flags);
3437 #endif /* CONFIG_NUMA */
3439 static __always_inline void *
3440 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3442 unsigned long save_flags;
3445 if (should_failslab(cachep, flags))
3448 cache_alloc_debugcheck_before(cachep, flags);
3449 local_irq_save(save_flags);
3450 objp = __do_cache_alloc(cachep, flags);
3451 local_irq_restore(save_flags);
3452 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3459 * Caller needs to acquire correct kmem_list's list_lock
3461 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3465 struct kmem_list3 *l3;
3467 for (i = 0; i < nr_objects; i++) {
3468 void *objp = objpp[i];
3471 slabp = virt_to_slab(objp);
3472 l3 = cachep->nodelists[node];
3473 list_del(&slabp->list);
3474 check_spinlock_acquired_node(cachep, node);
3475 check_slabp(cachep, slabp);
3476 slab_put_obj(cachep, slabp, objp, node);
3477 STATS_DEC_ACTIVE(cachep);
3479 check_slabp(cachep, slabp);
3481 /* fixup slab chains */
3482 if (slabp->inuse == 0) {
3483 if (l3->free_objects > l3->free_limit) {
3484 l3->free_objects -= cachep->num;
3485 /* No need to drop any previously held
3486 * lock here, even if we have a off-slab slab
3487 * descriptor it is guaranteed to come from
3488 * a different cache, refer to comments before
3491 slab_destroy(cachep, slabp);
3493 list_add(&slabp->list, &l3->slabs_free);
3496 /* Unconditionally move a slab to the end of the
3497 * partial list on free - maximum time for the
3498 * other objects to be freed, too.
3500 list_add_tail(&slabp->list, &l3->slabs_partial);
3505 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3508 struct kmem_list3 *l3;
3509 int node = numa_node_id();
3511 batchcount = ac->batchcount;
3513 BUG_ON(!batchcount || batchcount > ac->avail);
3516 l3 = cachep->nodelists[node];
3517 spin_lock(&l3->list_lock);
3519 struct array_cache *shared_array = l3->shared;
3520 int max = shared_array->limit - shared_array->avail;
3522 if (batchcount > max)
3524 memcpy(&(shared_array->entry[shared_array->avail]),
3525 ac->entry, sizeof(void *) * batchcount);
3526 shared_array->avail += batchcount;
3531 free_block(cachep, ac->entry, batchcount, node);
3536 struct list_head *p;
3538 p = l3->slabs_free.next;
3539 while (p != &(l3->slabs_free)) {
3542 slabp = list_entry(p, struct slab, list);
3543 BUG_ON(slabp->inuse);
3548 STATS_SET_FREEABLE(cachep, i);
3551 spin_unlock(&l3->list_lock);
3552 ac->avail -= batchcount;
3553 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3557 * Release an obj back to its cache. If the obj has a constructed state, it must
3558 * be in this state _before_ it is released. Called with disabled ints.
3560 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3562 struct array_cache *ac = cpu_cache_get(cachep);
3565 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3567 if (use_alien_caches && cache_free_alien(cachep, objp))
3570 if (likely(ac->avail < ac->limit)) {
3571 STATS_INC_FREEHIT(cachep);
3572 ac->entry[ac->avail++] = objp;
3575 STATS_INC_FREEMISS(cachep);
3576 cache_flusharray(cachep, ac);
3577 ac->entry[ac->avail++] = objp;
3582 * kmem_cache_alloc - Allocate an object
3583 * @cachep: The cache to allocate from.
3584 * @flags: See kmalloc().
3586 * Allocate an object from this cache. The flags are only relevant
3587 * if the cache has no available objects.
3589 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3591 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3593 EXPORT_SYMBOL(kmem_cache_alloc);
3596 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3597 * @cache: The cache to allocate from.
3598 * @flags: See kmalloc().
3600 * Allocate an object from this cache and set the allocated memory to zero.
3601 * The flags are only relevant if the cache has no available objects.
3603 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3605 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3607 memset(ret, 0, obj_size(cache));
3610 EXPORT_SYMBOL(kmem_cache_zalloc);
3613 * kmem_ptr_validate - check if an untrusted pointer might
3615 * @cachep: the cache we're checking against
3616 * @ptr: pointer to validate
3618 * This verifies that the untrusted pointer looks sane:
3619 * it is _not_ a guarantee that the pointer is actually
3620 * part of the slab cache in question, but it at least
3621 * validates that the pointer can be dereferenced and
3622 * looks half-way sane.
3624 * Currently only used for dentry validation.
3626 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3628 unsigned long addr = (unsigned long)ptr;
3629 unsigned long min_addr = PAGE_OFFSET;
3630 unsigned long align_mask = BYTES_PER_WORD - 1;
3631 unsigned long size = cachep->buffer_size;
3634 if (unlikely(addr < min_addr))
3636 if (unlikely(addr > (unsigned long)high_memory - size))
3638 if (unlikely(addr & align_mask))
3640 if (unlikely(!kern_addr_valid(addr)))
3642 if (unlikely(!kern_addr_valid(addr + size - 1)))
3644 page = virt_to_page(ptr);
3645 if (unlikely(!PageSlab(page)))
3647 if (unlikely(page_get_cache(page) != cachep))
3655 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3657 return __cache_alloc_node(cachep, flags, nodeid,
3658 __builtin_return_address(0));
3660 EXPORT_SYMBOL(kmem_cache_alloc_node);
3662 static __always_inline void *
3663 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3665 struct kmem_cache *cachep;
3667 cachep = kmem_find_general_cachep(size, flags);
3668 if (unlikely(cachep == NULL))
3670 return kmem_cache_alloc_node(cachep, flags, node);
3673 #ifdef CONFIG_DEBUG_SLAB
3674 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3676 return __do_kmalloc_node(size, flags, node,
3677 __builtin_return_address(0));
3679 EXPORT_SYMBOL(__kmalloc_node);
3681 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3682 int node, void *caller)
3684 return __do_kmalloc_node(size, flags, node, caller);
3686 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3688 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3690 return __do_kmalloc_node(size, flags, node, NULL);
3692 EXPORT_SYMBOL(__kmalloc_node);
3693 #endif /* CONFIG_DEBUG_SLAB */
3694 #endif /* CONFIG_NUMA */
3697 * __do_kmalloc - allocate memory
3698 * @size: how many bytes of memory are required.
3699 * @flags: the type of memory to allocate (see kmalloc).
3700 * @caller: function caller for debug tracking of the caller
3702 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3705 struct kmem_cache *cachep;
3707 /* If you want to save a few bytes .text space: replace
3709 * Then kmalloc uses the uninlined functions instead of the inline
3712 cachep = __find_general_cachep(size, flags);
3713 if (unlikely(cachep == NULL))
3715 return __cache_alloc(cachep, flags, caller);
3719 #ifdef CONFIG_DEBUG_SLAB
3720 void *__kmalloc(size_t size, gfp_t flags)
3722 return __do_kmalloc(size, flags, __builtin_return_address(0));
3724 EXPORT_SYMBOL(__kmalloc);
3726 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3728 return __do_kmalloc(size, flags, caller);
3730 EXPORT_SYMBOL(__kmalloc_track_caller);
3733 void *__kmalloc(size_t size, gfp_t flags)
3735 return __do_kmalloc(size, flags, NULL);
3737 EXPORT_SYMBOL(__kmalloc);
3741 * krealloc - reallocate memory. The contents will remain unchanged.
3742 * @p: object to reallocate memory for.
3743 * @new_size: how many bytes of memory are required.
3744 * @flags: the type of memory to allocate.
3746 * The contents of the object pointed to are preserved up to the
3747 * lesser of the new and old sizes. If @p is %NULL, krealloc()
3748 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
3749 * %NULL pointer, the object pointed to is freed.
3751 void *krealloc(const void *p, size_t new_size, gfp_t flags)
3753 struct kmem_cache *cache, *new_cache;
3757 return kmalloc_track_caller(new_size, flags);
3759 if (unlikely(!new_size)) {
3764 cache = virt_to_cache(p);
3765 new_cache = __find_general_cachep(new_size, flags);
3768 * If new size fits in the current cache, bail out.
3770 if (likely(cache == new_cache))
3774 * We are on the slow-path here so do not use __cache_alloc
3775 * because it bloats kernel text.
3777 ret = kmalloc_track_caller(new_size, flags);
3779 memcpy(ret, p, min(new_size, ksize(p)));
3784 EXPORT_SYMBOL(krealloc);
3787 * kmem_cache_free - Deallocate an object
3788 * @cachep: The cache the allocation was from.
3789 * @objp: The previously allocated object.
3791 * Free an object which was previously allocated from this
3794 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3796 unsigned long flags;
3798 BUG_ON(virt_to_cache(objp) != cachep);
3800 local_irq_save(flags);
3801 debug_check_no_locks_freed(objp, obj_size(cachep));
3802 __cache_free(cachep, objp);
3803 local_irq_restore(flags);
3805 EXPORT_SYMBOL(kmem_cache_free);
3808 * kfree - free previously allocated memory
3809 * @objp: pointer returned by kmalloc.
3811 * If @objp is NULL, no operation is performed.
3813 * Don't free memory not originally allocated by kmalloc()
3814 * or you will run into trouble.
3816 void kfree(const void *objp)
3818 struct kmem_cache *c;
3819 unsigned long flags;
3821 if (unlikely(!objp))
3823 local_irq_save(flags);
3824 kfree_debugcheck(objp);
3825 c = virt_to_cache(objp);
3826 debug_check_no_locks_freed(objp, obj_size(c));
3827 __cache_free(c, (void *)objp);
3828 local_irq_restore(flags);
3830 EXPORT_SYMBOL(kfree);
3832 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3834 return obj_size(cachep);
3836 EXPORT_SYMBOL(kmem_cache_size);
3838 const char *kmem_cache_name(struct kmem_cache *cachep)
3840 return cachep->name;
3842 EXPORT_SYMBOL_GPL(kmem_cache_name);
3845 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3847 static int alloc_kmemlist(struct kmem_cache *cachep)
3850 struct kmem_list3 *l3;
3851 struct array_cache *new_shared;
3852 struct array_cache **new_alien = NULL;
3854 for_each_online_node(node) {
3856 if (use_alien_caches) {
3857 new_alien = alloc_alien_cache(node, cachep->limit);
3863 if (cachep->shared) {
3864 new_shared = alloc_arraycache(node,
3865 cachep->shared*cachep->batchcount,
3868 free_alien_cache(new_alien);
3873 l3 = cachep->nodelists[node];
3875 struct array_cache *shared = l3->shared;
3877 spin_lock_irq(&l3->list_lock);
3880 free_block(cachep, shared->entry,
3881 shared->avail, node);
3883 l3->shared = new_shared;
3885 l3->alien = new_alien;
3888 l3->free_limit = (1 + nr_cpus_node(node)) *
3889 cachep->batchcount + cachep->num;
3890 spin_unlock_irq(&l3->list_lock);
3892 free_alien_cache(new_alien);
3895 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3897 free_alien_cache(new_alien);
3902 kmem_list3_init(l3);
3903 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3904 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3905 l3->shared = new_shared;
3906 l3->alien = new_alien;
3907 l3->free_limit = (1 + nr_cpus_node(node)) *
3908 cachep->batchcount + cachep->num;
3909 cachep->nodelists[node] = l3;
3914 if (!cachep->next.next) {
3915 /* Cache is not active yet. Roll back what we did */
3918 if (cachep->nodelists[node]) {
3919 l3 = cachep->nodelists[node];
3922 free_alien_cache(l3->alien);
3924 cachep->nodelists[node] = NULL;
3932 struct ccupdate_struct {
3933 struct kmem_cache *cachep;
3934 struct array_cache *new[NR_CPUS];
3937 static void do_ccupdate_local(void *info)
3939 struct ccupdate_struct *new = info;
3940 struct array_cache *old;
3943 old = cpu_cache_get(new->cachep);
3945 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3946 new->new[smp_processor_id()] = old;
3949 /* Always called with the cache_chain_mutex held */
3950 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3951 int batchcount, int shared)
3953 struct ccupdate_struct *new;
3956 new = kzalloc(sizeof(*new), GFP_KERNEL);
3960 for_each_online_cpu(i) {
3961 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3964 for (i--; i >= 0; i--)
3970 new->cachep = cachep;
3972 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
3975 cachep->batchcount = batchcount;
3976 cachep->limit = limit;
3977 cachep->shared = shared;
3979 for_each_online_cpu(i) {
3980 struct array_cache *ccold = new->new[i];
3983 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3984 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3985 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3989 return alloc_kmemlist(cachep);
3992 /* Called with cache_chain_mutex held always */
3993 static int enable_cpucache(struct kmem_cache *cachep)
3999 * The head array serves three purposes:
4000 * - create a LIFO ordering, i.e. return objects that are cache-warm
4001 * - reduce the number of spinlock operations.
4002 * - reduce the number of linked list operations on the slab and
4003 * bufctl chains: array operations are cheaper.
4004 * The numbers are guessed, we should auto-tune as described by
4007 if (cachep->buffer_size > 131072)
4009 else if (cachep->buffer_size > PAGE_SIZE)
4011 else if (cachep->buffer_size > 1024)
4013 else if (cachep->buffer_size > 256)
4019 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4020 * allocation behaviour: Most allocs on one cpu, most free operations
4021 * on another cpu. For these cases, an efficient object passing between
4022 * cpus is necessary. This is provided by a shared array. The array
4023 * replaces Bonwick's magazine layer.
4024 * On uniprocessor, it's functionally equivalent (but less efficient)
4025 * to a larger limit. Thus disabled by default.
4028 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4033 * With debugging enabled, large batchcount lead to excessively long
4034 * periods with disabled local interrupts. Limit the batchcount
4039 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
4041 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4042 cachep->name, -err);
4047 * Drain an array if it contains any elements taking the l3 lock only if
4048 * necessary. Note that the l3 listlock also protects the array_cache
4049 * if drain_array() is used on the shared array.
4051 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4052 struct array_cache *ac, int force, int node)
4056 if (!ac || !ac->avail)
4058 if (ac->touched && !force) {
4061 spin_lock_irq(&l3->list_lock);
4063 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4064 if (tofree > ac->avail)
4065 tofree = (ac->avail + 1) / 2;
4066 free_block(cachep, ac->entry, tofree, node);
4067 ac->avail -= tofree;
4068 memmove(ac->entry, &(ac->entry[tofree]),
4069 sizeof(void *) * ac->avail);
4071 spin_unlock_irq(&l3->list_lock);
4076 * cache_reap - Reclaim memory from caches.
4077 * @w: work descriptor
4079 * Called from workqueue/eventd every few seconds.
4081 * - clear the per-cpu caches for this CPU.
4082 * - return freeable pages to the main free memory pool.
4084 * If we cannot acquire the cache chain mutex then just give up - we'll try
4085 * again on the next iteration.
4087 static void cache_reap(struct work_struct *w)
4089 struct kmem_cache *searchp;
4090 struct kmem_list3 *l3;
4091 int node = numa_node_id();
4092 struct delayed_work *work =
4093 container_of(w, struct delayed_work, work);
4095 if (!mutex_trylock(&cache_chain_mutex))
4096 /* Give up. Setup the next iteration. */
4099 list_for_each_entry(searchp, &cache_chain, next) {
4103 * We only take the l3 lock if absolutely necessary and we
4104 * have established with reasonable certainty that
4105 * we can do some work if the lock was obtained.
4107 l3 = searchp->nodelists[node];
4109 reap_alien(searchp, l3);
4111 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4114 * These are racy checks but it does not matter
4115 * if we skip one check or scan twice.
4117 if (time_after(l3->next_reap, jiffies))
4120 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4122 drain_array(searchp, l3, l3->shared, 0, node);
4124 if (l3->free_touched)
4125 l3->free_touched = 0;
4129 freed = drain_freelist(searchp, l3, (l3->free_limit +
4130 5 * searchp->num - 1) / (5 * searchp->num));
4131 STATS_ADD_REAPED(searchp, freed);
4137 mutex_unlock(&cache_chain_mutex);
4139 refresh_cpu_vm_stats(smp_processor_id());
4141 /* Set up the next iteration */
4142 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4145 #ifdef CONFIG_PROC_FS
4147 static void print_slabinfo_header(struct seq_file *m)
4150 * Output format version, so at least we can change it
4151 * without _too_ many complaints.
4154 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4156 seq_puts(m, "slabinfo - version: 2.1\n");
4158 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4159 "<objperslab> <pagesperslab>");
4160 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4161 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4163 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4164 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4165 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4170 static void *s_start(struct seq_file *m, loff_t *pos)
4173 struct list_head *p;
4175 mutex_lock(&cache_chain_mutex);
4177 print_slabinfo_header(m);
4178 p = cache_chain.next;
4181 if (p == &cache_chain)
4184 return list_entry(p, struct kmem_cache, next);
4187 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4189 struct kmem_cache *cachep = p;
4191 return cachep->next.next == &cache_chain ?
4192 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
4195 static void s_stop(struct seq_file *m, void *p)
4197 mutex_unlock(&cache_chain_mutex);
4200 static int s_show(struct seq_file *m, void *p)
4202 struct kmem_cache *cachep = p;
4204 unsigned long active_objs;
4205 unsigned long num_objs;
4206 unsigned long active_slabs = 0;
4207 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4211 struct kmem_list3 *l3;
4215 for_each_online_node(node) {
4216 l3 = cachep->nodelists[node];
4221 spin_lock_irq(&l3->list_lock);
4223 list_for_each_entry(slabp, &l3->slabs_full, list) {
4224 if (slabp->inuse != cachep->num && !error)
4225 error = "slabs_full accounting error";
4226 active_objs += cachep->num;
4229 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4230 if (slabp->inuse == cachep->num && !error)
4231 error = "slabs_partial inuse accounting error";
4232 if (!slabp->inuse && !error)
4233 error = "slabs_partial/inuse accounting error";
4234 active_objs += slabp->inuse;
4237 list_for_each_entry(slabp, &l3->slabs_free, list) {
4238 if (slabp->inuse && !error)
4239 error = "slabs_free/inuse accounting error";
4242 free_objects += l3->free_objects;
4244 shared_avail += l3->shared->avail;
4246 spin_unlock_irq(&l3->list_lock);
4248 num_slabs += active_slabs;
4249 num_objs = num_slabs * cachep->num;
4250 if (num_objs - active_objs != free_objects && !error)
4251 error = "free_objects accounting error";
4253 name = cachep->name;
4255 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4257 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4258 name, active_objs, num_objs, cachep->buffer_size,
4259 cachep->num, (1 << cachep->gfporder));
4260 seq_printf(m, " : tunables %4u %4u %4u",
4261 cachep->limit, cachep->batchcount, cachep->shared);
4262 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4263 active_slabs, num_slabs, shared_avail);
4266 unsigned long high = cachep->high_mark;
4267 unsigned long allocs = cachep->num_allocations;
4268 unsigned long grown = cachep->grown;
4269 unsigned long reaped = cachep->reaped;
4270 unsigned long errors = cachep->errors;
4271 unsigned long max_freeable = cachep->max_freeable;
4272 unsigned long node_allocs = cachep->node_allocs;
4273 unsigned long node_frees = cachep->node_frees;
4274 unsigned long overflows = cachep->node_overflow;
4276 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4277 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4278 reaped, errors, max_freeable, node_allocs,
4279 node_frees, overflows);
4283 unsigned long allochit = atomic_read(&cachep->allochit);
4284 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4285 unsigned long freehit = atomic_read(&cachep->freehit);
4286 unsigned long freemiss = atomic_read(&cachep->freemiss);
4288 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4289 allochit, allocmiss, freehit, freemiss);
4297 * slabinfo_op - iterator that generates /proc/slabinfo
4306 * num-pages-per-slab
4307 * + further values on SMP and with statistics enabled
4310 const struct seq_operations slabinfo_op = {
4317 #define MAX_SLABINFO_WRITE 128
4319 * slabinfo_write - Tuning for the slab allocator
4321 * @buffer: user buffer
4322 * @count: data length
4325 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4326 size_t count, loff_t *ppos)
4328 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4329 int limit, batchcount, shared, res;
4330 struct kmem_cache *cachep;
4332 if (count > MAX_SLABINFO_WRITE)
4334 if (copy_from_user(&kbuf, buffer, count))
4336 kbuf[MAX_SLABINFO_WRITE] = '\0';
4338 tmp = strchr(kbuf, ' ');
4343 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4346 /* Find the cache in the chain of caches. */
4347 mutex_lock(&cache_chain_mutex);
4349 list_for_each_entry(cachep, &cache_chain, next) {
4350 if (!strcmp(cachep->name, kbuf)) {
4351 if (limit < 1 || batchcount < 1 ||
4352 batchcount > limit || shared < 0) {
4355 res = do_tune_cpucache(cachep, limit,
4356 batchcount, shared);
4361 mutex_unlock(&cache_chain_mutex);
4367 #ifdef CONFIG_DEBUG_SLAB_LEAK
4369 static void *leaks_start(struct seq_file *m, loff_t *pos)
4372 struct list_head *p;
4374 mutex_lock(&cache_chain_mutex);
4375 p = cache_chain.next;
4378 if (p == &cache_chain)
4381 return list_entry(p, struct kmem_cache, next);
4384 static inline int add_caller(unsigned long *n, unsigned long v)
4394 unsigned long *q = p + 2 * i;
4408 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4414 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4420 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4421 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4423 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4428 static void show_symbol(struct seq_file *m, unsigned long address)
4430 #ifdef CONFIG_KALLSYMS
4431 unsigned long offset, size;
4432 char modname[MODULE_NAME_LEN + 1], name[KSYM_NAME_LEN + 1];
4434 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4435 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4437 seq_printf(m, " [%s]", modname);
4441 seq_printf(m, "%p", (void *)address);
4444 static int leaks_show(struct seq_file *m, void *p)
4446 struct kmem_cache *cachep = p;
4448 struct kmem_list3 *l3;
4450 unsigned long *n = m->private;
4454 if (!(cachep->flags & SLAB_STORE_USER))
4456 if (!(cachep->flags & SLAB_RED_ZONE))
4459 /* OK, we can do it */
4463 for_each_online_node(node) {
4464 l3 = cachep->nodelists[node];
4469 spin_lock_irq(&l3->list_lock);
4471 list_for_each_entry(slabp, &l3->slabs_full, list)
4472 handle_slab(n, cachep, slabp);
4473 list_for_each_entry(slabp, &l3->slabs_partial, list)
4474 handle_slab(n, cachep, slabp);
4475 spin_unlock_irq(&l3->list_lock);
4477 name = cachep->name;
4479 /* Increase the buffer size */
4480 mutex_unlock(&cache_chain_mutex);
4481 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4483 /* Too bad, we are really out */
4485 mutex_lock(&cache_chain_mutex);
4488 *(unsigned long *)m->private = n[0] * 2;
4490 mutex_lock(&cache_chain_mutex);
4491 /* Now make sure this entry will be retried */
4495 for (i = 0; i < n[1]; i++) {
4496 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4497 show_symbol(m, n[2*i+2]);
4504 const struct seq_operations slabstats_op = {
4505 .start = leaks_start,
4514 * ksize - get the actual amount of memory allocated for a given object
4515 * @objp: Pointer to the object
4517 * kmalloc may internally round up allocations and return more memory
4518 * than requested. ksize() can be used to determine the actual amount of
4519 * memory allocated. The caller may use this additional memory, even though
4520 * a smaller amount of memory was initially specified with the kmalloc call.
4521 * The caller must guarantee that objp points to a valid object previously
4522 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4523 * must not be freed during the duration of the call.
4525 size_t ksize(const void *objp)
4527 if (unlikely(objp == NULL))
4530 return obj_size(virt_to_cache(objp));