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/config.h>
90 #include <linux/slab.h>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
111 #include <asm/uaccess.h>
112 #include <asm/cacheflush.h>
113 #include <asm/tlbflush.h>
114 #include <asm/page.h>
117 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
118 * SLAB_RED_ZONE & SLAB_POISON.
119 * 0 for faster, smaller code (especially in the critical paths).
121 * STATS - 1 to collect stats for /proc/slabinfo.
122 * 0 for faster, smaller code (especially in the critical paths).
124 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
127 #ifdef CONFIG_DEBUG_SLAB
130 #define FORCED_DEBUG 1
134 #define FORCED_DEBUG 0
137 /* Shouldn't this be in a header file somewhere? */
138 #define BYTES_PER_WORD sizeof(void *)
140 #ifndef cache_line_size
141 #define cache_line_size() L1_CACHE_BYTES
144 #ifndef ARCH_KMALLOC_MINALIGN
146 * Enforce a minimum alignment for the kmalloc caches.
147 * Usually, the kmalloc caches are cache_line_size() aligned, except when
148 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
149 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
150 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
151 * Note that this flag disables some debug features.
153 #define ARCH_KMALLOC_MINALIGN 0
156 #ifndef ARCH_SLAB_MINALIGN
158 * Enforce a minimum alignment for all caches.
159 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
160 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
161 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
162 * some debug features.
164 #define ARCH_SLAB_MINALIGN 0
167 #ifndef ARCH_KMALLOC_FLAGS
168 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
171 /* Legal flag mask for kmem_cache_create(). */
173 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
174 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
176 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
180 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
181 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
182 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
183 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
189 * Bufctl's are used for linking objs within a slab
192 * This implementation relies on "struct page" for locating the cache &
193 * slab an object belongs to.
194 * This allows the bufctl structure to be small (one int), but limits
195 * the number of objects a slab (not a cache) can contain when off-slab
196 * bufctls are used. The limit is the size of the largest general cache
197 * that does not use off-slab slabs.
198 * For 32bit archs with 4 kB pages, is this 56.
199 * This is not serious, as it is only for large objects, when it is unwise
200 * to have too many per slab.
201 * Note: This limit can be raised by introducing a general cache whose size
202 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
205 typedef unsigned int kmem_bufctl_t;
206 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
207 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
208 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
209 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
219 struct list_head list;
220 unsigned long colouroff;
221 void *s_mem; /* including colour offset */
222 unsigned int inuse; /* num of objs active in slab */
224 unsigned short nodeid;
230 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
231 * arrange for kmem_freepages to be called via RCU. This is useful if
232 * we need to approach a kernel structure obliquely, from its address
233 * obtained without the usual locking. We can lock the structure to
234 * stabilize it and check it's still at the given address, only if we
235 * can be sure that the memory has not been meanwhile reused for some
236 * other kind of object (which our subsystem's lock might corrupt).
238 * rcu_read_lock before reading the address, then rcu_read_unlock after
239 * taking the spinlock within the structure expected at that address.
241 * We assume struct slab_rcu can overlay struct slab when destroying.
244 struct rcu_head head;
245 struct kmem_cache *cachep;
253 * - LIFO ordering, to hand out cache-warm objects from _alloc
254 * - reduce the number of linked list operations
255 * - reduce spinlock operations
257 * The limit is stored in the per-cpu structure to reduce the data cache
264 unsigned int batchcount;
265 unsigned int touched;
268 * Must have this definition in here for the proper
269 * alignment of array_cache. Also simplifies accessing
271 * [0] is for gcc 2.95. It should really be [].
276 * bootstrap: The caches do not work without cpuarrays anymore, but the
277 * cpuarrays are allocated from the generic caches...
279 #define BOOT_CPUCACHE_ENTRIES 1
280 struct arraycache_init {
281 struct array_cache cache;
282 void *entries[BOOT_CPUCACHE_ENTRIES];
286 * The slab lists for all objects.
289 struct list_head slabs_partial; /* partial list first, better asm code */
290 struct list_head slabs_full;
291 struct list_head slabs_free;
292 unsigned long free_objects;
293 unsigned int free_limit;
294 unsigned int colour_next; /* Per-node cache coloring */
295 spinlock_t list_lock;
296 struct array_cache *shared; /* shared per node */
297 struct array_cache **alien; /* on other nodes */
298 unsigned long next_reap; /* updated without locking */
299 int free_touched; /* updated without locking */
303 * Need this for bootstrapping a per node allocator.
305 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
306 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
307 #define CACHE_CACHE 0
309 #define SIZE_L3 (1 + MAX_NUMNODES)
312 * This function must be completely optimized away if a constant is passed to
313 * it. Mostly the same as what is in linux/slab.h except it returns an index.
315 static __always_inline int index_of(const size_t size)
317 extern void __bad_size(void);
319 if (__builtin_constant_p(size)) {
327 #include "linux/kmalloc_sizes.h"
335 static int slab_early_init = 1;
337 #define INDEX_AC index_of(sizeof(struct arraycache_init))
338 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
340 static void kmem_list3_init(struct kmem_list3 *parent)
342 INIT_LIST_HEAD(&parent->slabs_full);
343 INIT_LIST_HEAD(&parent->slabs_partial);
344 INIT_LIST_HEAD(&parent->slabs_free);
345 parent->shared = NULL;
346 parent->alien = NULL;
347 parent->colour_next = 0;
348 spin_lock_init(&parent->list_lock);
349 parent->free_objects = 0;
350 parent->free_touched = 0;
353 #define MAKE_LIST(cachep, listp, slab, nodeid) \
355 INIT_LIST_HEAD(listp); \
356 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
359 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
361 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
373 /* 1) per-cpu data, touched during every alloc/free */
374 struct array_cache *array[NR_CPUS];
375 /* 2) Cache tunables. Protected by cache_chain_mutex */
376 unsigned int batchcount;
380 unsigned int buffer_size;
381 /* 3) touched by every alloc & free from the backend */
382 struct kmem_list3 *nodelists[MAX_NUMNODES];
384 unsigned int flags; /* constant flags */
385 unsigned int num; /* # of objs per slab */
387 /* 4) cache_grow/shrink */
388 /* order of pgs per slab (2^n) */
389 unsigned int gfporder;
391 /* force GFP flags, e.g. GFP_DMA */
394 size_t colour; /* cache colouring range */
395 unsigned int colour_off; /* colour offset */
396 struct kmem_cache *slabp_cache;
397 unsigned int slab_size;
398 unsigned int dflags; /* dynamic flags */
400 /* constructor func */
401 void (*ctor) (void *, struct kmem_cache *, unsigned long);
403 /* de-constructor func */
404 void (*dtor) (void *, struct kmem_cache *, unsigned long);
406 /* 5) cache creation/removal */
408 struct list_head next;
412 unsigned long num_active;
413 unsigned long num_allocations;
414 unsigned long high_mark;
416 unsigned long reaped;
417 unsigned long errors;
418 unsigned long max_freeable;
419 unsigned long node_allocs;
420 unsigned long node_frees;
421 unsigned long node_overflow;
429 * If debugging is enabled, then the allocator can add additional
430 * fields and/or padding to every object. buffer_size contains the total
431 * object size including these internal fields, the following two
432 * variables contain the offset to the user object and its size.
439 #define CFLGS_OFF_SLAB (0x80000000UL)
440 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
442 #define BATCHREFILL_LIMIT 16
444 * Optimization question: fewer reaps means less probability for unnessary
445 * cpucache drain/refill cycles.
447 * OTOH the cpuarrays can contain lots of objects,
448 * which could lock up otherwise freeable slabs.
450 #define REAPTIMEOUT_CPUC (2*HZ)
451 #define REAPTIMEOUT_LIST3 (4*HZ)
454 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
455 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
456 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
457 #define STATS_INC_GROWN(x) ((x)->grown++)
458 #define STATS_INC_REAPED(x) ((x)->reaped++)
459 #define STATS_SET_HIGH(x) \
461 if ((x)->num_active > (x)->high_mark) \
462 (x)->high_mark = (x)->num_active; \
464 #define STATS_INC_ERR(x) ((x)->errors++)
465 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
466 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
467 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
468 #define STATS_SET_FREEABLE(x, i) \
470 if ((x)->max_freeable < i) \
471 (x)->max_freeable = i; \
473 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
474 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
475 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
476 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
478 #define STATS_INC_ACTIVE(x) do { } while (0)
479 #define STATS_DEC_ACTIVE(x) do { } while (0)
480 #define STATS_INC_ALLOCED(x) do { } while (0)
481 #define STATS_INC_GROWN(x) do { } while (0)
482 #define STATS_INC_REAPED(x) do { } while (0)
483 #define STATS_SET_HIGH(x) do { } while (0)
484 #define STATS_INC_ERR(x) do { } while (0)
485 #define STATS_INC_NODEALLOCS(x) do { } while (0)
486 #define STATS_INC_NODEFREES(x) do { } while (0)
487 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
488 #define STATS_SET_FREEABLE(x, i) do { } while (0)
489 #define STATS_INC_ALLOCHIT(x) do { } while (0)
490 #define STATS_INC_ALLOCMISS(x) do { } while (0)
491 #define STATS_INC_FREEHIT(x) do { } while (0)
492 #define STATS_INC_FREEMISS(x) do { } while (0)
498 * memory layout of objects:
500 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
501 * the end of an object is aligned with the end of the real
502 * allocation. Catches writes behind the end of the allocation.
503 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
505 * cachep->obj_offset: The real object.
506 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
507 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
508 * [BYTES_PER_WORD long]
510 static int obj_offset(struct kmem_cache *cachep)
512 return cachep->obj_offset;
515 static int obj_size(struct kmem_cache *cachep)
517 return cachep->obj_size;
520 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
522 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
523 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
526 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
528 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
529 if (cachep->flags & SLAB_STORE_USER)
530 return (unsigned long *)(objp + cachep->buffer_size -
532 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
535 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
537 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
538 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
543 #define obj_offset(x) 0
544 #define obj_size(cachep) (cachep->buffer_size)
545 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
546 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
547 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
552 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
555 #if defined(CONFIG_LARGE_ALLOCS)
556 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
557 #define MAX_GFP_ORDER 13 /* up to 32Mb */
558 #elif defined(CONFIG_MMU)
559 #define MAX_OBJ_ORDER 5 /* 32 pages */
560 #define MAX_GFP_ORDER 5 /* 32 pages */
562 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
563 #define MAX_GFP_ORDER 8 /* up to 1Mb */
567 * Do not go above this order unless 0 objects fit into the slab.
569 #define BREAK_GFP_ORDER_HI 1
570 #define BREAK_GFP_ORDER_LO 0
571 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
574 * Functions for storing/retrieving the cachep and or slab from the page
575 * allocator. These are used to find the slab an obj belongs to. With kfree(),
576 * these are used to find the cache which an obj belongs to.
578 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
580 page->lru.next = (struct list_head *)cache;
583 static inline struct kmem_cache *page_get_cache(struct page *page)
585 if (unlikely(PageCompound(page)))
586 page = (struct page *)page_private(page);
587 BUG_ON(!PageSlab(page));
588 return (struct kmem_cache *)page->lru.next;
591 static inline void page_set_slab(struct page *page, struct slab *slab)
593 page->lru.prev = (struct list_head *)slab;
596 static inline struct slab *page_get_slab(struct page *page)
598 if (unlikely(PageCompound(page)))
599 page = (struct page *)page_private(page);
600 BUG_ON(!PageSlab(page));
601 return (struct slab *)page->lru.prev;
604 static inline struct kmem_cache *virt_to_cache(const void *obj)
606 struct page *page = virt_to_page(obj);
607 return page_get_cache(page);
610 static inline struct slab *virt_to_slab(const void *obj)
612 struct page *page = virt_to_page(obj);
613 return page_get_slab(page);
616 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
619 return slab->s_mem + cache->buffer_size * idx;
622 static inline unsigned int obj_to_index(struct kmem_cache *cache,
623 struct slab *slab, void *obj)
625 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
629 * These are the default caches for kmalloc. Custom caches can have other sizes.
631 struct cache_sizes malloc_sizes[] = {
632 #define CACHE(x) { .cs_size = (x) },
633 #include <linux/kmalloc_sizes.h>
637 EXPORT_SYMBOL(malloc_sizes);
639 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
645 static struct cache_names __initdata cache_names[] = {
646 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
647 #include <linux/kmalloc_sizes.h>
652 static struct arraycache_init initarray_cache __initdata =
653 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
654 static struct arraycache_init initarray_generic =
655 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
657 /* internal cache of cache description objs */
658 static struct kmem_cache cache_cache = {
660 .limit = BOOT_CPUCACHE_ENTRIES,
662 .buffer_size = sizeof(struct kmem_cache),
663 .name = "kmem_cache",
665 .obj_size = sizeof(struct kmem_cache),
669 /* Guard access to the cache-chain. */
670 static DEFINE_MUTEX(cache_chain_mutex);
671 static struct list_head cache_chain;
674 * vm_enough_memory() looks at this to determine how many slab-allocated pages
675 * are possibly freeable under pressure
677 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
679 atomic_t slab_reclaim_pages;
682 * chicken and egg problem: delay the per-cpu array allocation
683 * until the general caches are up.
693 * used by boot code to determine if it can use slab based allocator
695 int slab_is_available(void)
697 return g_cpucache_up == FULL;
700 static DEFINE_PER_CPU(struct work_struct, reap_work);
702 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
704 static void enable_cpucache(struct kmem_cache *cachep);
705 static void cache_reap(void *unused);
706 static int __node_shrink(struct kmem_cache *cachep, int node);
708 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
710 return cachep->array[smp_processor_id()];
713 static inline struct kmem_cache *__find_general_cachep(size_t size,
716 struct cache_sizes *csizep = malloc_sizes;
719 /* This happens if someone tries to call
720 * kmem_cache_create(), or __kmalloc(), before
721 * the generic caches are initialized.
723 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
725 while (size > csizep->cs_size)
729 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
730 * has cs_{dma,}cachep==NULL. Thus no special case
731 * for large kmalloc calls required.
733 if (unlikely(gfpflags & GFP_DMA))
734 return csizep->cs_dmacachep;
735 return csizep->cs_cachep;
738 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
740 return __find_general_cachep(size, gfpflags);
742 EXPORT_SYMBOL(kmem_find_general_cachep);
744 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
746 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
750 * Calculate the number of objects and left-over bytes for a given buffer size.
752 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
753 size_t align, int flags, size_t *left_over,
758 size_t slab_size = PAGE_SIZE << gfporder;
761 * The slab management structure can be either off the slab or
762 * on it. For the latter case, the memory allocated for a
766 * - One kmem_bufctl_t for each object
767 * - Padding to respect alignment of @align
768 * - @buffer_size bytes for each object
770 * If the slab management structure is off the slab, then the
771 * alignment will already be calculated into the size. Because
772 * the slabs are all pages aligned, the objects will be at the
773 * correct alignment when allocated.
775 if (flags & CFLGS_OFF_SLAB) {
777 nr_objs = slab_size / buffer_size;
779 if (nr_objs > SLAB_LIMIT)
780 nr_objs = SLAB_LIMIT;
783 * Ignore padding for the initial guess. The padding
784 * is at most @align-1 bytes, and @buffer_size is at
785 * least @align. In the worst case, this result will
786 * be one greater than the number of objects that fit
787 * into the memory allocation when taking the padding
790 nr_objs = (slab_size - sizeof(struct slab)) /
791 (buffer_size + sizeof(kmem_bufctl_t));
794 * This calculated number will be either the right
795 * amount, or one greater than what we want.
797 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
801 if (nr_objs > SLAB_LIMIT)
802 nr_objs = SLAB_LIMIT;
804 mgmt_size = slab_mgmt_size(nr_objs, align);
807 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
810 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
812 static void __slab_error(const char *function, struct kmem_cache *cachep,
815 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
816 function, cachep->name, msg);
822 * Special reaping functions for NUMA systems called from cache_reap().
823 * These take care of doing round robin flushing of alien caches (containing
824 * objects freed on different nodes from which they were allocated) and the
825 * flushing of remote pcps by calling drain_node_pages.
827 static DEFINE_PER_CPU(unsigned long, reap_node);
829 static void init_reap_node(int cpu)
833 node = next_node(cpu_to_node(cpu), node_online_map);
834 if (node == MAX_NUMNODES)
835 node = first_node(node_online_map);
837 __get_cpu_var(reap_node) = node;
840 static void next_reap_node(void)
842 int node = __get_cpu_var(reap_node);
845 * Also drain per cpu pages on remote zones
847 if (node != numa_node_id())
848 drain_node_pages(node);
850 node = next_node(node, node_online_map);
851 if (unlikely(node >= MAX_NUMNODES))
852 node = first_node(node_online_map);
853 __get_cpu_var(reap_node) = node;
857 #define init_reap_node(cpu) do { } while (0)
858 #define next_reap_node(void) do { } while (0)
862 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
863 * via the workqueue/eventd.
864 * Add the CPU number into the expiration time to minimize the possibility of
865 * the CPUs getting into lockstep and contending for the global cache chain
868 static void __devinit start_cpu_timer(int cpu)
870 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
873 * When this gets called from do_initcalls via cpucache_init(),
874 * init_workqueues() has already run, so keventd will be setup
877 if (keventd_up() && reap_work->func == NULL) {
879 INIT_WORK(reap_work, cache_reap, NULL);
880 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
884 static struct array_cache *alloc_arraycache(int node, int entries,
887 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
888 struct array_cache *nc = NULL;
890 nc = kmalloc_node(memsize, GFP_KERNEL, node);
894 nc->batchcount = batchcount;
896 spin_lock_init(&nc->lock);
902 * Transfer objects in one arraycache to another.
903 * Locking must be handled by the caller.
905 * Return the number of entries transferred.
907 static int transfer_objects(struct array_cache *to,
908 struct array_cache *from, unsigned int max)
910 /* Figure out how many entries to transfer */
911 int nr = min(min(from->avail, max), to->limit - to->avail);
916 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
926 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
927 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
929 static struct array_cache **alloc_alien_cache(int node, int limit)
931 struct array_cache **ac_ptr;
932 int memsize = sizeof(void *) * MAX_NUMNODES;
937 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
940 if (i == node || !node_online(i)) {
944 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
946 for (i--; i <= 0; i--)
956 static void free_alien_cache(struct array_cache **ac_ptr)
967 static void __drain_alien_cache(struct kmem_cache *cachep,
968 struct array_cache *ac, int node)
970 struct kmem_list3 *rl3 = cachep->nodelists[node];
973 spin_lock(&rl3->list_lock);
975 * Stuff objects into the remote nodes shared array first.
976 * That way we could avoid the overhead of putting the objects
977 * into the free lists and getting them back later.
980 transfer_objects(rl3->shared, ac, ac->limit);
982 free_block(cachep, ac->entry, ac->avail, node);
984 spin_unlock(&rl3->list_lock);
989 * Called from cache_reap() to regularly drain alien caches round robin.
991 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
993 int node = __get_cpu_var(reap_node);
996 struct array_cache *ac = l3->alien[node];
998 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
999 __drain_alien_cache(cachep, ac, node);
1000 spin_unlock_irq(&ac->lock);
1005 static void drain_alien_cache(struct kmem_cache *cachep,
1006 struct array_cache **alien)
1009 struct array_cache *ac;
1010 unsigned long flags;
1012 for_each_online_node(i) {
1015 spin_lock_irqsave(&ac->lock, flags);
1016 __drain_alien_cache(cachep, ac, i);
1017 spin_unlock_irqrestore(&ac->lock, flags);
1022 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1024 struct slab *slabp = virt_to_slab(objp);
1025 int nodeid = slabp->nodeid;
1026 struct kmem_list3 *l3;
1027 struct array_cache *alien = NULL;
1030 * Make sure we are not freeing a object from another node to the array
1031 * cache on this cpu.
1033 if (likely(slabp->nodeid == numa_node_id()))
1036 l3 = cachep->nodelists[numa_node_id()];
1037 STATS_INC_NODEFREES(cachep);
1038 if (l3->alien && l3->alien[nodeid]) {
1039 alien = l3->alien[nodeid];
1040 spin_lock(&alien->lock);
1041 if (unlikely(alien->avail == alien->limit)) {
1042 STATS_INC_ACOVERFLOW(cachep);
1043 __drain_alien_cache(cachep, alien, nodeid);
1045 alien->entry[alien->avail++] = objp;
1046 spin_unlock(&alien->lock);
1048 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1049 free_block(cachep, &objp, 1, nodeid);
1050 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1057 #define drain_alien_cache(cachep, alien) do { } while (0)
1058 #define reap_alien(cachep, l3) do { } while (0)
1060 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1062 return (struct array_cache **) 0x01020304ul;
1065 static inline void free_alien_cache(struct array_cache **ac_ptr)
1069 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1076 static int __devinit cpuup_callback(struct notifier_block *nfb,
1077 unsigned long action, void *hcpu)
1079 long cpu = (long)hcpu;
1080 struct kmem_cache *cachep;
1081 struct kmem_list3 *l3 = NULL;
1082 int node = cpu_to_node(cpu);
1083 int memsize = sizeof(struct kmem_list3);
1086 case CPU_UP_PREPARE:
1087 mutex_lock(&cache_chain_mutex);
1089 * We need to do this right in the beginning since
1090 * alloc_arraycache's are going to use this list.
1091 * kmalloc_node allows us to add the slab to the right
1092 * kmem_list3 and not this cpu's kmem_list3
1095 list_for_each_entry(cachep, &cache_chain, next) {
1097 * Set up the size64 kmemlist for cpu before we can
1098 * begin anything. Make sure some other cpu on this
1099 * node has not already allocated this
1101 if (!cachep->nodelists[node]) {
1102 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1105 kmem_list3_init(l3);
1106 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1107 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1110 * The l3s don't come and go as CPUs come and
1111 * go. cache_chain_mutex is sufficient
1114 cachep->nodelists[node] = l3;
1117 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1118 cachep->nodelists[node]->free_limit =
1119 (1 + nr_cpus_node(node)) *
1120 cachep->batchcount + cachep->num;
1121 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1125 * Now we can go ahead with allocating the shared arrays and
1128 list_for_each_entry(cachep, &cache_chain, next) {
1129 struct array_cache *nc;
1130 struct array_cache *shared;
1131 struct array_cache **alien;
1133 nc = alloc_arraycache(node, cachep->limit,
1134 cachep->batchcount);
1137 shared = alloc_arraycache(node,
1138 cachep->shared * cachep->batchcount,
1143 alien = alloc_alien_cache(node, cachep->limit);
1146 cachep->array[cpu] = nc;
1147 l3 = cachep->nodelists[node];
1150 spin_lock_irq(&l3->list_lock);
1153 * We are serialised from CPU_DEAD or
1154 * CPU_UP_CANCELLED by the cpucontrol lock
1156 l3->shared = shared;
1165 spin_unlock_irq(&l3->list_lock);
1167 free_alien_cache(alien);
1169 mutex_unlock(&cache_chain_mutex);
1172 start_cpu_timer(cpu);
1174 #ifdef CONFIG_HOTPLUG_CPU
1177 * Even if all the cpus of a node are down, we don't free the
1178 * kmem_list3 of any cache. This to avoid a race between
1179 * cpu_down, and a kmalloc allocation from another cpu for
1180 * memory from the node of the cpu going down. The list3
1181 * structure is usually allocated from kmem_cache_create() and
1182 * gets destroyed at kmem_cache_destroy().
1185 case CPU_UP_CANCELED:
1186 mutex_lock(&cache_chain_mutex);
1187 list_for_each_entry(cachep, &cache_chain, next) {
1188 struct array_cache *nc;
1189 struct array_cache *shared;
1190 struct array_cache **alien;
1193 mask = node_to_cpumask(node);
1194 /* cpu is dead; no one can alloc from it. */
1195 nc = cachep->array[cpu];
1196 cachep->array[cpu] = NULL;
1197 l3 = cachep->nodelists[node];
1200 goto free_array_cache;
1202 spin_lock_irq(&l3->list_lock);
1204 /* Free limit for this kmem_list3 */
1205 l3->free_limit -= cachep->batchcount;
1207 free_block(cachep, nc->entry, nc->avail, node);
1209 if (!cpus_empty(mask)) {
1210 spin_unlock_irq(&l3->list_lock);
1211 goto free_array_cache;
1214 shared = l3->shared;
1216 free_block(cachep, l3->shared->entry,
1217 l3->shared->avail, node);
1224 spin_unlock_irq(&l3->list_lock);
1228 drain_alien_cache(cachep, alien);
1229 free_alien_cache(alien);
1235 * In the previous loop, all the objects were freed to
1236 * the respective cache's slabs, now we can go ahead and
1237 * shrink each nodelist to its limit.
1239 list_for_each_entry(cachep, &cache_chain, next) {
1240 l3 = cachep->nodelists[node];
1243 spin_lock_irq(&l3->list_lock);
1244 /* free slabs belonging to this node */
1245 __node_shrink(cachep, node);
1246 spin_unlock_irq(&l3->list_lock);
1248 mutex_unlock(&cache_chain_mutex);
1254 mutex_unlock(&cache_chain_mutex);
1258 static struct notifier_block __cpuinitdata cpucache_notifier = {
1259 &cpuup_callback, NULL, 0
1263 * swap the static kmem_list3 with kmalloced memory
1265 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1268 struct kmem_list3 *ptr;
1270 BUG_ON(cachep->nodelists[nodeid] != list);
1271 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1274 local_irq_disable();
1275 memcpy(ptr, list, sizeof(struct kmem_list3));
1276 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1277 cachep->nodelists[nodeid] = ptr;
1282 * Initialisation. Called after the page allocator have been initialised and
1283 * before smp_init().
1285 void __init kmem_cache_init(void)
1288 struct cache_sizes *sizes;
1289 struct cache_names *names;
1293 for (i = 0; i < NUM_INIT_LISTS; i++) {
1294 kmem_list3_init(&initkmem_list3[i]);
1295 if (i < MAX_NUMNODES)
1296 cache_cache.nodelists[i] = NULL;
1300 * Fragmentation resistance on low memory - only use bigger
1301 * page orders on machines with more than 32MB of memory.
1303 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1304 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1306 /* Bootstrap is tricky, because several objects are allocated
1307 * from caches that do not exist yet:
1308 * 1) initialize the cache_cache cache: it contains the struct
1309 * kmem_cache structures of all caches, except cache_cache itself:
1310 * cache_cache is statically allocated.
1311 * Initially an __init data area is used for the head array and the
1312 * kmem_list3 structures, it's replaced with a kmalloc allocated
1313 * array at the end of the bootstrap.
1314 * 2) Create the first kmalloc cache.
1315 * The struct kmem_cache for the new cache is allocated normally.
1316 * An __init data area is used for the head array.
1317 * 3) Create the remaining kmalloc caches, with minimally sized
1319 * 4) Replace the __init data head arrays for cache_cache and the first
1320 * kmalloc cache with kmalloc allocated arrays.
1321 * 5) Replace the __init data for kmem_list3 for cache_cache and
1322 * the other cache's with kmalloc allocated memory.
1323 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1326 /* 1) create the cache_cache */
1327 INIT_LIST_HEAD(&cache_chain);
1328 list_add(&cache_cache.next, &cache_chain);
1329 cache_cache.colour_off = cache_line_size();
1330 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1331 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1333 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1336 for (order = 0; order < MAX_ORDER; order++) {
1337 cache_estimate(order, cache_cache.buffer_size,
1338 cache_line_size(), 0, &left_over, &cache_cache.num);
1339 if (cache_cache.num)
1342 BUG_ON(!cache_cache.num);
1343 cache_cache.gfporder = order;
1344 cache_cache.colour = left_over / cache_cache.colour_off;
1345 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1346 sizeof(struct slab), cache_line_size());
1348 /* 2+3) create the kmalloc caches */
1349 sizes = malloc_sizes;
1350 names = cache_names;
1353 * Initialize the caches that provide memory for the array cache and the
1354 * kmem_list3 structures first. Without this, further allocations will
1358 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1359 sizes[INDEX_AC].cs_size,
1360 ARCH_KMALLOC_MINALIGN,
1361 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1364 if (INDEX_AC != INDEX_L3) {
1365 sizes[INDEX_L3].cs_cachep =
1366 kmem_cache_create(names[INDEX_L3].name,
1367 sizes[INDEX_L3].cs_size,
1368 ARCH_KMALLOC_MINALIGN,
1369 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1373 slab_early_init = 0;
1375 while (sizes->cs_size != ULONG_MAX) {
1377 * For performance, all the general caches are L1 aligned.
1378 * This should be particularly beneficial on SMP boxes, as it
1379 * eliminates "false sharing".
1380 * Note for systems short on memory removing the alignment will
1381 * allow tighter packing of the smaller caches.
1383 if (!sizes->cs_cachep) {
1384 sizes->cs_cachep = kmem_cache_create(names->name,
1386 ARCH_KMALLOC_MINALIGN,
1387 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1391 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1393 ARCH_KMALLOC_MINALIGN,
1394 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1400 /* 4) Replace the bootstrap head arrays */
1404 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1406 local_irq_disable();
1407 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1408 memcpy(ptr, cpu_cache_get(&cache_cache),
1409 sizeof(struct arraycache_init));
1410 cache_cache.array[smp_processor_id()] = ptr;
1413 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1415 local_irq_disable();
1416 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1417 != &initarray_generic.cache);
1418 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1419 sizeof(struct arraycache_init));
1420 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1424 /* 5) Replace the bootstrap kmem_list3's */
1427 /* Replace the static kmem_list3 structures for the boot cpu */
1428 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1431 for_each_online_node(node) {
1432 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1433 &initkmem_list3[SIZE_AC + node], node);
1435 if (INDEX_AC != INDEX_L3) {
1436 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1437 &initkmem_list3[SIZE_L3 + node],
1443 /* 6) resize the head arrays to their final sizes */
1445 struct kmem_cache *cachep;
1446 mutex_lock(&cache_chain_mutex);
1447 list_for_each_entry(cachep, &cache_chain, next)
1448 enable_cpucache(cachep);
1449 mutex_unlock(&cache_chain_mutex);
1453 g_cpucache_up = FULL;
1456 * Register a cpu startup notifier callback that initializes
1457 * cpu_cache_get for all new cpus
1459 register_cpu_notifier(&cpucache_notifier);
1462 * The reap timers are started later, with a module init call: That part
1463 * of the kernel is not yet operational.
1467 static int __init cpucache_init(void)
1472 * Register the timers that return unneeded pages to the page allocator
1474 for_each_online_cpu(cpu)
1475 start_cpu_timer(cpu);
1478 __initcall(cpucache_init);
1481 * Interface to system's page allocator. No need to hold the cache-lock.
1483 * If we requested dmaable memory, we will get it. Even if we
1484 * did not request dmaable memory, we might get it, but that
1485 * would be relatively rare and ignorable.
1487 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1495 * Nommu uses slab's for process anonymous memory allocations, and thus
1496 * requires __GFP_COMP to properly refcount higher order allocations
1498 flags |= __GFP_COMP;
1500 flags |= cachep->gfpflags;
1502 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1506 nr_pages = (1 << cachep->gfporder);
1507 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1508 atomic_add(nr_pages, &slab_reclaim_pages);
1509 add_page_state(nr_slab, nr_pages);
1510 for (i = 0; i < nr_pages; i++)
1511 __SetPageSlab(page + i);
1512 return page_address(page);
1516 * Interface to system's page release.
1518 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1520 unsigned long i = (1 << cachep->gfporder);
1521 struct page *page = virt_to_page(addr);
1522 const unsigned long nr_freed = i;
1525 BUG_ON(!PageSlab(page));
1526 __ClearPageSlab(page);
1529 sub_page_state(nr_slab, nr_freed);
1530 if (current->reclaim_state)
1531 current->reclaim_state->reclaimed_slab += nr_freed;
1532 free_pages((unsigned long)addr, cachep->gfporder);
1533 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1534 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1537 static void kmem_rcu_free(struct rcu_head *head)
1539 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1540 struct kmem_cache *cachep = slab_rcu->cachep;
1542 kmem_freepages(cachep, slab_rcu->addr);
1543 if (OFF_SLAB(cachep))
1544 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1549 #ifdef CONFIG_DEBUG_PAGEALLOC
1550 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1551 unsigned long caller)
1553 int size = obj_size(cachep);
1555 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1557 if (size < 5 * sizeof(unsigned long))
1560 *addr++ = 0x12345678;
1562 *addr++ = smp_processor_id();
1563 size -= 3 * sizeof(unsigned long);
1565 unsigned long *sptr = &caller;
1566 unsigned long svalue;
1568 while (!kstack_end(sptr)) {
1570 if (kernel_text_address(svalue)) {
1572 size -= sizeof(unsigned long);
1573 if (size <= sizeof(unsigned long))
1579 *addr++ = 0x87654321;
1583 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1585 int size = obj_size(cachep);
1586 addr = &((char *)addr)[obj_offset(cachep)];
1588 memset(addr, val, size);
1589 *(unsigned char *)(addr + size - 1) = POISON_END;
1592 static void dump_line(char *data, int offset, int limit)
1595 printk(KERN_ERR "%03x:", offset);
1596 for (i = 0; i < limit; i++)
1597 printk(" %02x", (unsigned char)data[offset + i]);
1604 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1609 if (cachep->flags & SLAB_RED_ZONE) {
1610 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1611 *dbg_redzone1(cachep, objp),
1612 *dbg_redzone2(cachep, objp));
1615 if (cachep->flags & SLAB_STORE_USER) {
1616 printk(KERN_ERR "Last user: [<%p>]",
1617 *dbg_userword(cachep, objp));
1618 print_symbol("(%s)",
1619 (unsigned long)*dbg_userword(cachep, objp));
1622 realobj = (char *)objp + obj_offset(cachep);
1623 size = obj_size(cachep);
1624 for (i = 0; i < size && lines; i += 16, lines--) {
1627 if (i + limit > size)
1629 dump_line(realobj, i, limit);
1633 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1639 realobj = (char *)objp + obj_offset(cachep);
1640 size = obj_size(cachep);
1642 for (i = 0; i < size; i++) {
1643 char exp = POISON_FREE;
1646 if (realobj[i] != exp) {
1652 "Slab corruption: start=%p, len=%d\n",
1654 print_objinfo(cachep, objp, 0);
1656 /* Hexdump the affected line */
1659 if (i + limit > size)
1661 dump_line(realobj, i, limit);
1664 /* Limit to 5 lines */
1670 /* Print some data about the neighboring objects, if they
1673 struct slab *slabp = virt_to_slab(objp);
1676 objnr = obj_to_index(cachep, slabp, objp);
1678 objp = index_to_obj(cachep, slabp, objnr - 1);
1679 realobj = (char *)objp + obj_offset(cachep);
1680 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1682 print_objinfo(cachep, objp, 2);
1684 if (objnr + 1 < cachep->num) {
1685 objp = index_to_obj(cachep, slabp, objnr + 1);
1686 realobj = (char *)objp + obj_offset(cachep);
1687 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1689 print_objinfo(cachep, objp, 2);
1697 * slab_destroy_objs - destroy a slab and its objects
1698 * @cachep: cache pointer being destroyed
1699 * @slabp: slab pointer being destroyed
1701 * Call the registered destructor for each object in a slab that is being
1704 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1707 for (i = 0; i < cachep->num; i++) {
1708 void *objp = index_to_obj(cachep, slabp, i);
1710 if (cachep->flags & SLAB_POISON) {
1711 #ifdef CONFIG_DEBUG_PAGEALLOC
1712 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1714 kernel_map_pages(virt_to_page(objp),
1715 cachep->buffer_size / PAGE_SIZE, 1);
1717 check_poison_obj(cachep, objp);
1719 check_poison_obj(cachep, objp);
1722 if (cachep->flags & SLAB_RED_ZONE) {
1723 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1724 slab_error(cachep, "start of a freed object "
1726 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1727 slab_error(cachep, "end of a freed object "
1730 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1731 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1735 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1739 for (i = 0; i < cachep->num; i++) {
1740 void *objp = index_to_obj(cachep, slabp, i);
1741 (cachep->dtor) (objp, cachep, 0);
1748 * slab_destroy - destroy and release all objects in a slab
1749 * @cachep: cache pointer being destroyed
1750 * @slabp: slab pointer being destroyed
1752 * Destroy all the objs in a slab, and release the mem back to the system.
1753 * Before calling the slab must have been unlinked from the cache. The
1754 * cache-lock is not held/needed.
1756 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1758 void *addr = slabp->s_mem - slabp->colouroff;
1760 slab_destroy_objs(cachep, slabp);
1761 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1762 struct slab_rcu *slab_rcu;
1764 slab_rcu = (struct slab_rcu *)slabp;
1765 slab_rcu->cachep = cachep;
1766 slab_rcu->addr = addr;
1767 call_rcu(&slab_rcu->head, kmem_rcu_free);
1769 kmem_freepages(cachep, addr);
1770 if (OFF_SLAB(cachep))
1771 kmem_cache_free(cachep->slabp_cache, slabp);
1776 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1777 * size of kmem_list3.
1779 static void set_up_list3s(struct kmem_cache *cachep, int index)
1783 for_each_online_node(node) {
1784 cachep->nodelists[node] = &initkmem_list3[index + node];
1785 cachep->nodelists[node]->next_reap = jiffies +
1787 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1792 * calculate_slab_order - calculate size (page order) of slabs
1793 * @cachep: pointer to the cache that is being created
1794 * @size: size of objects to be created in this cache.
1795 * @align: required alignment for the objects.
1796 * @flags: slab allocation flags
1798 * Also calculates the number of objects per slab.
1800 * This could be made much more intelligent. For now, try to avoid using
1801 * high order pages for slabs. When the gfp() functions are more friendly
1802 * towards high-order requests, this should be changed.
1804 static size_t calculate_slab_order(struct kmem_cache *cachep,
1805 size_t size, size_t align, unsigned long flags)
1807 unsigned long offslab_limit;
1808 size_t left_over = 0;
1811 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1815 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1819 if (flags & CFLGS_OFF_SLAB) {
1821 * Max number of objs-per-slab for caches which
1822 * use off-slab slabs. Needed to avoid a possible
1823 * looping condition in cache_grow().
1825 offslab_limit = size - sizeof(struct slab);
1826 offslab_limit /= sizeof(kmem_bufctl_t);
1828 if (num > offslab_limit)
1832 /* Found something acceptable - save it away */
1834 cachep->gfporder = gfporder;
1835 left_over = remainder;
1838 * A VFS-reclaimable slab tends to have most allocations
1839 * as GFP_NOFS and we really don't want to have to be allocating
1840 * higher-order pages when we are unable to shrink dcache.
1842 if (flags & SLAB_RECLAIM_ACCOUNT)
1846 * Large number of objects is good, but very large slabs are
1847 * currently bad for the gfp()s.
1849 if (gfporder >= slab_break_gfp_order)
1853 * Acceptable internal fragmentation?
1855 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1861 static void setup_cpu_cache(struct kmem_cache *cachep)
1863 if (g_cpucache_up == FULL) {
1864 enable_cpucache(cachep);
1867 if (g_cpucache_up == NONE) {
1869 * Note: the first kmem_cache_create must create the cache
1870 * that's used by kmalloc(24), otherwise the creation of
1871 * further caches will BUG().
1873 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1876 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1877 * the first cache, then we need to set up all its list3s,
1878 * otherwise the creation of further caches will BUG().
1880 set_up_list3s(cachep, SIZE_AC);
1881 if (INDEX_AC == INDEX_L3)
1882 g_cpucache_up = PARTIAL_L3;
1884 g_cpucache_up = PARTIAL_AC;
1886 cachep->array[smp_processor_id()] =
1887 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1889 if (g_cpucache_up == PARTIAL_AC) {
1890 set_up_list3s(cachep, SIZE_L3);
1891 g_cpucache_up = PARTIAL_L3;
1894 for_each_online_node(node) {
1895 cachep->nodelists[node] =
1896 kmalloc_node(sizeof(struct kmem_list3),
1898 BUG_ON(!cachep->nodelists[node]);
1899 kmem_list3_init(cachep->nodelists[node]);
1903 cachep->nodelists[numa_node_id()]->next_reap =
1904 jiffies + REAPTIMEOUT_LIST3 +
1905 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1907 cpu_cache_get(cachep)->avail = 0;
1908 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1909 cpu_cache_get(cachep)->batchcount = 1;
1910 cpu_cache_get(cachep)->touched = 0;
1911 cachep->batchcount = 1;
1912 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1916 * kmem_cache_create - Create a cache.
1917 * @name: A string which is used in /proc/slabinfo to identify this cache.
1918 * @size: The size of objects to be created in this cache.
1919 * @align: The required alignment for the objects.
1920 * @flags: SLAB flags
1921 * @ctor: A constructor for the objects.
1922 * @dtor: A destructor for the objects.
1924 * Returns a ptr to the cache on success, NULL on failure.
1925 * Cannot be called within a int, but can be interrupted.
1926 * The @ctor is run when new pages are allocated by the cache
1927 * and the @dtor is run before the pages are handed back.
1929 * @name must be valid until the cache is destroyed. This implies that
1930 * the module calling this has to destroy the cache before getting unloaded.
1934 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1935 * to catch references to uninitialised memory.
1937 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1938 * for buffer overruns.
1940 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1941 * cacheline. This can be beneficial if you're counting cycles as closely
1945 kmem_cache_create (const char *name, size_t size, size_t align,
1946 unsigned long flags,
1947 void (*ctor)(void*, struct kmem_cache *, unsigned long),
1948 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1950 size_t left_over, slab_size, ralign;
1951 struct kmem_cache *cachep = NULL, *pc;
1954 * Sanity checks... these are all serious usage bugs.
1956 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
1957 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1958 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1964 * Prevent CPUs from coming and going.
1965 * lock_cpu_hotplug() nests outside cache_chain_mutex
1969 mutex_lock(&cache_chain_mutex);
1971 list_for_each_entry(pc, &cache_chain, next) {
1972 mm_segment_t old_fs = get_fs();
1977 * This happens when the module gets unloaded and doesn't
1978 * destroy its slab cache and no-one else reuses the vmalloc
1979 * area of the module. Print a warning.
1982 res = __get_user(tmp, pc->name);
1985 printk("SLAB: cache with size %d has lost its name\n",
1990 if (!strcmp(pc->name, name)) {
1991 printk("kmem_cache_create: duplicate cache %s\n", name);
1998 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1999 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2000 /* No constructor, but inital state check requested */
2001 printk(KERN_ERR "%s: No con, but init state check "
2002 "requested - %s\n", __FUNCTION__, name);
2003 flags &= ~SLAB_DEBUG_INITIAL;
2007 * Enable redzoning and last user accounting, except for caches with
2008 * large objects, if the increased size would increase the object size
2009 * above the next power of two: caches with object sizes just above a
2010 * power of two have a significant amount of internal fragmentation.
2012 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2013 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2014 if (!(flags & SLAB_DESTROY_BY_RCU))
2015 flags |= SLAB_POISON;
2017 if (flags & SLAB_DESTROY_BY_RCU)
2018 BUG_ON(flags & SLAB_POISON);
2020 if (flags & SLAB_DESTROY_BY_RCU)
2024 * Always checks flags, a caller might be expecting debug support which
2027 BUG_ON(flags & ~CREATE_MASK);
2030 * Check that size is in terms of words. This is needed to avoid
2031 * unaligned accesses for some archs when redzoning is used, and makes
2032 * sure any on-slab bufctl's are also correctly aligned.
2034 if (size & (BYTES_PER_WORD - 1)) {
2035 size += (BYTES_PER_WORD - 1);
2036 size &= ~(BYTES_PER_WORD - 1);
2039 /* calculate the final buffer alignment: */
2041 /* 1) arch recommendation: can be overridden for debug */
2042 if (flags & SLAB_HWCACHE_ALIGN) {
2044 * Default alignment: as specified by the arch code. Except if
2045 * an object is really small, then squeeze multiple objects into
2048 ralign = cache_line_size();
2049 while (size <= ralign / 2)
2052 ralign = BYTES_PER_WORD;
2054 /* 2) arch mandated alignment: disables debug if necessary */
2055 if (ralign < ARCH_SLAB_MINALIGN) {
2056 ralign = ARCH_SLAB_MINALIGN;
2057 if (ralign > BYTES_PER_WORD)
2058 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2060 /* 3) caller mandated alignment: disables debug if necessary */
2061 if (ralign < align) {
2063 if (ralign > BYTES_PER_WORD)
2064 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2067 * 4) Store it. Note that the debug code below can reduce
2068 * the alignment to BYTES_PER_WORD.
2072 /* Get cache's description obj. */
2073 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
2078 cachep->obj_size = size;
2080 if (flags & SLAB_RED_ZONE) {
2081 /* redzoning only works with word aligned caches */
2082 align = BYTES_PER_WORD;
2084 /* add space for red zone words */
2085 cachep->obj_offset += BYTES_PER_WORD;
2086 size += 2 * BYTES_PER_WORD;
2088 if (flags & SLAB_STORE_USER) {
2089 /* user store requires word alignment and
2090 * one word storage behind the end of the real
2093 align = BYTES_PER_WORD;
2094 size += BYTES_PER_WORD;
2096 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2097 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2098 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2099 cachep->obj_offset += PAGE_SIZE - size;
2106 * Determine if the slab management is 'on' or 'off' slab.
2107 * (bootstrapping cannot cope with offslab caches so don't do
2110 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2112 * Size is large, assume best to place the slab management obj
2113 * off-slab (should allow better packing of objs).
2115 flags |= CFLGS_OFF_SLAB;
2117 size = ALIGN(size, align);
2119 left_over = calculate_slab_order(cachep, size, align, flags);
2122 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2123 kmem_cache_free(&cache_cache, cachep);
2127 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2128 + sizeof(struct slab), align);
2131 * If the slab has been placed off-slab, and we have enough space then
2132 * move it on-slab. This is at the expense of any extra colouring.
2134 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2135 flags &= ~CFLGS_OFF_SLAB;
2136 left_over -= slab_size;
2139 if (flags & CFLGS_OFF_SLAB) {
2140 /* really off slab. No need for manual alignment */
2142 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2145 cachep->colour_off = cache_line_size();
2146 /* Offset must be a multiple of the alignment. */
2147 if (cachep->colour_off < align)
2148 cachep->colour_off = align;
2149 cachep->colour = left_over / cachep->colour_off;
2150 cachep->slab_size = slab_size;
2151 cachep->flags = flags;
2152 cachep->gfpflags = 0;
2153 if (flags & SLAB_CACHE_DMA)
2154 cachep->gfpflags |= GFP_DMA;
2155 cachep->buffer_size = size;
2157 if (flags & CFLGS_OFF_SLAB)
2158 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2159 cachep->ctor = ctor;
2160 cachep->dtor = dtor;
2161 cachep->name = name;
2164 setup_cpu_cache(cachep);
2166 /* cache setup completed, link it into the list */
2167 list_add(&cachep->next, &cache_chain);
2169 if (!cachep && (flags & SLAB_PANIC))
2170 panic("kmem_cache_create(): failed to create slab `%s'\n",
2172 mutex_unlock(&cache_chain_mutex);
2173 unlock_cpu_hotplug();
2176 EXPORT_SYMBOL(kmem_cache_create);
2179 static void check_irq_off(void)
2181 BUG_ON(!irqs_disabled());
2184 static void check_irq_on(void)
2186 BUG_ON(irqs_disabled());
2189 static void check_spinlock_acquired(struct kmem_cache *cachep)
2193 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2197 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2201 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2206 #define check_irq_off() do { } while(0)
2207 #define check_irq_on() do { } while(0)
2208 #define check_spinlock_acquired(x) do { } while(0)
2209 #define check_spinlock_acquired_node(x, y) do { } while(0)
2212 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2213 struct array_cache *ac,
2214 int force, int node);
2216 static void do_drain(void *arg)
2218 struct kmem_cache *cachep = arg;
2219 struct array_cache *ac;
2220 int node = numa_node_id();
2223 ac = cpu_cache_get(cachep);
2224 spin_lock(&cachep->nodelists[node]->list_lock);
2225 free_block(cachep, ac->entry, ac->avail, node);
2226 spin_unlock(&cachep->nodelists[node]->list_lock);
2230 static void drain_cpu_caches(struct kmem_cache *cachep)
2232 struct kmem_list3 *l3;
2235 on_each_cpu(do_drain, cachep, 1, 1);
2237 for_each_online_node(node) {
2238 l3 = cachep->nodelists[node];
2239 if (l3 && l3->alien)
2240 drain_alien_cache(cachep, l3->alien);
2243 for_each_online_node(node) {
2244 l3 = cachep->nodelists[node];
2246 drain_array(cachep, l3, l3->shared, 1, node);
2250 static int __node_shrink(struct kmem_cache *cachep, int node)
2253 struct kmem_list3 *l3 = cachep->nodelists[node];
2257 struct list_head *p;
2259 p = l3->slabs_free.prev;
2260 if (p == &l3->slabs_free)
2263 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
2265 BUG_ON(slabp->inuse);
2267 list_del(&slabp->list);
2269 l3->free_objects -= cachep->num;
2270 spin_unlock_irq(&l3->list_lock);
2271 slab_destroy(cachep, slabp);
2272 spin_lock_irq(&l3->list_lock);
2274 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
2278 static int __cache_shrink(struct kmem_cache *cachep)
2281 struct kmem_list3 *l3;
2283 drain_cpu_caches(cachep);
2286 for_each_online_node(i) {
2287 l3 = cachep->nodelists[i];
2289 spin_lock_irq(&l3->list_lock);
2290 ret += __node_shrink(cachep, i);
2291 spin_unlock_irq(&l3->list_lock);
2294 return (ret ? 1 : 0);
2298 * kmem_cache_shrink - Shrink a cache.
2299 * @cachep: The cache to shrink.
2301 * Releases as many slabs as possible for a cache.
2302 * To help debugging, a zero exit status indicates all slabs were released.
2304 int kmem_cache_shrink(struct kmem_cache *cachep)
2306 BUG_ON(!cachep || in_interrupt());
2308 return __cache_shrink(cachep);
2310 EXPORT_SYMBOL(kmem_cache_shrink);
2313 * kmem_cache_destroy - delete a cache
2314 * @cachep: the cache to destroy
2316 * Remove a struct kmem_cache object from the slab cache.
2317 * Returns 0 on success.
2319 * It is expected this function will be called by a module when it is
2320 * unloaded. This will remove the cache completely, and avoid a duplicate
2321 * cache being allocated each time a module is loaded and unloaded, if the
2322 * module doesn't have persistent in-kernel storage across loads and unloads.
2324 * The cache must be empty before calling this function.
2326 * The caller must guarantee that noone will allocate memory from the cache
2327 * during the kmem_cache_destroy().
2329 int kmem_cache_destroy(struct kmem_cache *cachep)
2332 struct kmem_list3 *l3;
2334 BUG_ON(!cachep || in_interrupt());
2336 /* Don't let CPUs to come and go */
2339 /* Find the cache in the chain of caches. */
2340 mutex_lock(&cache_chain_mutex);
2342 * the chain is never empty, cache_cache is never destroyed
2344 list_del(&cachep->next);
2345 mutex_unlock(&cache_chain_mutex);
2347 if (__cache_shrink(cachep)) {
2348 slab_error(cachep, "Can't free all objects");
2349 mutex_lock(&cache_chain_mutex);
2350 list_add(&cachep->next, &cache_chain);
2351 mutex_unlock(&cache_chain_mutex);
2352 unlock_cpu_hotplug();
2356 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2359 for_each_online_cpu(i)
2360 kfree(cachep->array[i]);
2362 /* NUMA: free the list3 structures */
2363 for_each_online_node(i) {
2364 l3 = cachep->nodelists[i];
2367 free_alien_cache(l3->alien);
2371 kmem_cache_free(&cache_cache, cachep);
2372 unlock_cpu_hotplug();
2375 EXPORT_SYMBOL(kmem_cache_destroy);
2377 /* Get the memory for a slab management obj. */
2378 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2379 int colour_off, gfp_t local_flags,
2384 if (OFF_SLAB(cachep)) {
2385 /* Slab management obj is off-slab. */
2386 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2387 local_flags, nodeid);
2391 slabp = objp + colour_off;
2392 colour_off += cachep->slab_size;
2395 slabp->colouroff = colour_off;
2396 slabp->s_mem = objp + colour_off;
2397 slabp->nodeid = nodeid;
2401 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2403 return (kmem_bufctl_t *) (slabp + 1);
2406 static void cache_init_objs(struct kmem_cache *cachep,
2407 struct slab *slabp, unsigned long ctor_flags)
2411 for (i = 0; i < cachep->num; i++) {
2412 void *objp = index_to_obj(cachep, slabp, i);
2414 /* need to poison the objs? */
2415 if (cachep->flags & SLAB_POISON)
2416 poison_obj(cachep, objp, POISON_FREE);
2417 if (cachep->flags & SLAB_STORE_USER)
2418 *dbg_userword(cachep, objp) = NULL;
2420 if (cachep->flags & SLAB_RED_ZONE) {
2421 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2422 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2425 * Constructors are not allowed to allocate memory from the same
2426 * cache which they are a constructor for. Otherwise, deadlock.
2427 * They must also be threaded.
2429 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2430 cachep->ctor(objp + obj_offset(cachep), cachep,
2433 if (cachep->flags & SLAB_RED_ZONE) {
2434 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2435 slab_error(cachep, "constructor overwrote the"
2436 " end of an object");
2437 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2438 slab_error(cachep, "constructor overwrote the"
2439 " start of an object");
2441 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2442 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2443 kernel_map_pages(virt_to_page(objp),
2444 cachep->buffer_size / PAGE_SIZE, 0);
2447 cachep->ctor(objp, cachep, ctor_flags);
2449 slab_bufctl(slabp)[i] = i + 1;
2451 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2455 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2457 if (flags & SLAB_DMA)
2458 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2460 BUG_ON(cachep->gfpflags & GFP_DMA);
2463 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2466 void *objp = index_to_obj(cachep, slabp, slabp->free);
2470 next = slab_bufctl(slabp)[slabp->free];
2472 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2473 WARN_ON(slabp->nodeid != nodeid);
2480 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2481 void *objp, int nodeid)
2483 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2486 /* Verify that the slab belongs to the intended node */
2487 WARN_ON(slabp->nodeid != nodeid);
2489 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2490 printk(KERN_ERR "slab: double free detected in cache "
2491 "'%s', objp %p\n", cachep->name, objp);
2495 slab_bufctl(slabp)[objnr] = slabp->free;
2496 slabp->free = objnr;
2501 * Map pages beginning at addr to the given cache and slab. This is required
2502 * for the slab allocator to be able to lookup the cache and slab of a
2503 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2505 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2511 page = virt_to_page(addr);
2514 if (likely(!PageCompound(page)))
2515 nr_pages <<= cache->gfporder;
2518 page_set_cache(page, cache);
2519 page_set_slab(page, slab);
2521 } while (--nr_pages);
2525 * Grow (by 1) the number of slabs within a cache. This is called by
2526 * kmem_cache_alloc() when there are no active objs left in a cache.
2528 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2534 unsigned long ctor_flags;
2535 struct kmem_list3 *l3;
2538 * Be lazy and only check for valid flags here, keeping it out of the
2539 * critical path in kmem_cache_alloc().
2541 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
2542 if (flags & SLAB_NO_GROW)
2545 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2546 local_flags = (flags & SLAB_LEVEL_MASK);
2547 if (!(local_flags & __GFP_WAIT))
2549 * Not allowed to sleep. Need to tell a constructor about
2550 * this - it might need to know...
2552 ctor_flags |= SLAB_CTOR_ATOMIC;
2554 /* Take the l3 list lock to change the colour_next on this node */
2556 l3 = cachep->nodelists[nodeid];
2557 spin_lock(&l3->list_lock);
2559 /* Get colour for the slab, and cal the next value. */
2560 offset = l3->colour_next;
2562 if (l3->colour_next >= cachep->colour)
2563 l3->colour_next = 0;
2564 spin_unlock(&l3->list_lock);
2566 offset *= cachep->colour_off;
2568 if (local_flags & __GFP_WAIT)
2572 * The test for missing atomic flag is performed here, rather than
2573 * the more obvious place, simply to reduce the critical path length
2574 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2575 * will eventually be caught here (where it matters).
2577 kmem_flagcheck(cachep, flags);
2580 * Get mem for the objs. Attempt to allocate a physical page from
2583 objp = kmem_getpages(cachep, flags, nodeid);
2587 /* Get slab management. */
2588 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
2592 slabp->nodeid = nodeid;
2593 slab_map_pages(cachep, slabp, objp);
2595 cache_init_objs(cachep, slabp, ctor_flags);
2597 if (local_flags & __GFP_WAIT)
2598 local_irq_disable();
2600 spin_lock(&l3->list_lock);
2602 /* Make slab active. */
2603 list_add_tail(&slabp->list, &(l3->slabs_free));
2604 STATS_INC_GROWN(cachep);
2605 l3->free_objects += cachep->num;
2606 spin_unlock(&l3->list_lock);
2609 kmem_freepages(cachep, objp);
2611 if (local_flags & __GFP_WAIT)
2612 local_irq_disable();
2619 * Perform extra freeing checks:
2620 * - detect bad pointers.
2621 * - POISON/RED_ZONE checking
2622 * - destructor calls, for caches with POISON+dtor
2624 static void kfree_debugcheck(const void *objp)
2628 if (!virt_addr_valid(objp)) {
2629 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2630 (unsigned long)objp);
2633 page = virt_to_page(objp);
2634 if (!PageSlab(page)) {
2635 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2636 (unsigned long)objp);
2641 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2643 unsigned long redzone1, redzone2;
2645 redzone1 = *dbg_redzone1(cache, obj);
2646 redzone2 = *dbg_redzone2(cache, obj);
2651 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2654 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2655 slab_error(cache, "double free detected");
2657 slab_error(cache, "memory outside object was overwritten");
2659 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2660 obj, redzone1, redzone2);
2663 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2670 objp -= obj_offset(cachep);
2671 kfree_debugcheck(objp);
2672 page = virt_to_page(objp);
2674 slabp = page_get_slab(page);
2676 if (cachep->flags & SLAB_RED_ZONE) {
2677 verify_redzone_free(cachep, objp);
2678 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2679 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2681 if (cachep->flags & SLAB_STORE_USER)
2682 *dbg_userword(cachep, objp) = caller;
2684 objnr = obj_to_index(cachep, slabp, objp);
2686 BUG_ON(objnr >= cachep->num);
2687 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2689 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2691 * Need to call the slab's constructor so the caller can
2692 * perform a verify of its state (debugging). Called without
2693 * the cache-lock held.
2695 cachep->ctor(objp + obj_offset(cachep),
2696 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2698 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2699 /* we want to cache poison the object,
2700 * call the destruction callback
2702 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2704 #ifdef CONFIG_DEBUG_SLAB_LEAK
2705 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2707 if (cachep->flags & SLAB_POISON) {
2708 #ifdef CONFIG_DEBUG_PAGEALLOC
2709 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2710 store_stackinfo(cachep, objp, (unsigned long)caller);
2711 kernel_map_pages(virt_to_page(objp),
2712 cachep->buffer_size / PAGE_SIZE, 0);
2714 poison_obj(cachep, objp, POISON_FREE);
2717 poison_obj(cachep, objp, POISON_FREE);
2723 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2728 /* Check slab's freelist to see if this obj is there. */
2729 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2731 if (entries > cachep->num || i >= cachep->num)
2734 if (entries != cachep->num - slabp->inuse) {
2736 printk(KERN_ERR "slab: Internal list corruption detected in "
2737 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2738 cachep->name, cachep->num, slabp, slabp->inuse);
2740 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2743 printk("\n%03x:", i);
2744 printk(" %02x", ((unsigned char *)slabp)[i]);
2751 #define kfree_debugcheck(x) do { } while(0)
2752 #define cache_free_debugcheck(x,objp,z) (objp)
2753 #define check_slabp(x,y) do { } while(0)
2756 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2759 struct kmem_list3 *l3;
2760 struct array_cache *ac;
2763 ac = cpu_cache_get(cachep);
2765 batchcount = ac->batchcount;
2766 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2768 * If there was little recent activity on this cache, then
2769 * perform only a partial refill. Otherwise we could generate
2772 batchcount = BATCHREFILL_LIMIT;
2774 l3 = cachep->nodelists[numa_node_id()];
2776 BUG_ON(ac->avail > 0 || !l3);
2777 spin_lock(&l3->list_lock);
2779 /* See if we can refill from the shared array */
2780 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2783 while (batchcount > 0) {
2784 struct list_head *entry;
2786 /* Get slab alloc is to come from. */
2787 entry = l3->slabs_partial.next;
2788 if (entry == &l3->slabs_partial) {
2789 l3->free_touched = 1;
2790 entry = l3->slabs_free.next;
2791 if (entry == &l3->slabs_free)
2795 slabp = list_entry(entry, struct slab, list);
2796 check_slabp(cachep, slabp);
2797 check_spinlock_acquired(cachep);
2798 while (slabp->inuse < cachep->num && batchcount--) {
2799 STATS_INC_ALLOCED(cachep);
2800 STATS_INC_ACTIVE(cachep);
2801 STATS_SET_HIGH(cachep);
2803 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2806 check_slabp(cachep, slabp);
2808 /* move slabp to correct slabp list: */
2809 list_del(&slabp->list);
2810 if (slabp->free == BUFCTL_END)
2811 list_add(&slabp->list, &l3->slabs_full);
2813 list_add(&slabp->list, &l3->slabs_partial);
2817 l3->free_objects -= ac->avail;
2819 spin_unlock(&l3->list_lock);
2821 if (unlikely(!ac->avail)) {
2823 x = cache_grow(cachep, flags, numa_node_id());
2825 /* cache_grow can reenable interrupts, then ac could change. */
2826 ac = cpu_cache_get(cachep);
2827 if (!x && ac->avail == 0) /* no objects in sight? abort */
2830 if (!ac->avail) /* objects refilled by interrupt? */
2834 return ac->entry[--ac->avail];
2837 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2840 might_sleep_if(flags & __GFP_WAIT);
2842 kmem_flagcheck(cachep, flags);
2847 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2848 gfp_t flags, void *objp, void *caller)
2852 if (cachep->flags & SLAB_POISON) {
2853 #ifdef CONFIG_DEBUG_PAGEALLOC
2854 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2855 kernel_map_pages(virt_to_page(objp),
2856 cachep->buffer_size / PAGE_SIZE, 1);
2858 check_poison_obj(cachep, objp);
2860 check_poison_obj(cachep, objp);
2862 poison_obj(cachep, objp, POISON_INUSE);
2864 if (cachep->flags & SLAB_STORE_USER)
2865 *dbg_userword(cachep, objp) = caller;
2867 if (cachep->flags & SLAB_RED_ZONE) {
2868 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2869 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2870 slab_error(cachep, "double free, or memory outside"
2871 " object was overwritten");
2873 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2874 objp, *dbg_redzone1(cachep, objp),
2875 *dbg_redzone2(cachep, objp));
2877 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2878 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2880 #ifdef CONFIG_DEBUG_SLAB_LEAK
2885 slabp = page_get_slab(virt_to_page(objp));
2886 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2887 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2890 objp += obj_offset(cachep);
2891 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2892 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2894 if (!(flags & __GFP_WAIT))
2895 ctor_flags |= SLAB_CTOR_ATOMIC;
2897 cachep->ctor(objp, cachep, ctor_flags);
2902 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2905 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2908 struct array_cache *ac;
2911 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
2912 objp = alternate_node_alloc(cachep, flags);
2919 ac = cpu_cache_get(cachep);
2920 if (likely(ac->avail)) {
2921 STATS_INC_ALLOCHIT(cachep);
2923 objp = ac->entry[--ac->avail];
2925 STATS_INC_ALLOCMISS(cachep);
2926 objp = cache_alloc_refill(cachep, flags);
2931 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2932 gfp_t flags, void *caller)
2934 unsigned long save_flags;
2937 cache_alloc_debugcheck_before(cachep, flags);
2939 local_irq_save(save_flags);
2940 objp = ____cache_alloc(cachep, flags);
2941 local_irq_restore(save_flags);
2942 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2950 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
2952 * If we are in_interrupt, then process context, including cpusets and
2953 * mempolicy, may not apply and should not be used for allocation policy.
2955 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2957 int nid_alloc, nid_here;
2961 nid_alloc = nid_here = numa_node_id();
2962 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2963 nid_alloc = cpuset_mem_spread_node();
2964 else if (current->mempolicy)
2965 nid_alloc = slab_node(current->mempolicy);
2966 if (nid_alloc != nid_here)
2967 return __cache_alloc_node(cachep, flags, nid_alloc);
2972 * A interface to enable slab creation on nodeid
2974 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2977 struct list_head *entry;
2979 struct kmem_list3 *l3;
2983 l3 = cachep->nodelists[nodeid];
2988 spin_lock(&l3->list_lock);
2989 entry = l3->slabs_partial.next;
2990 if (entry == &l3->slabs_partial) {
2991 l3->free_touched = 1;
2992 entry = l3->slabs_free.next;
2993 if (entry == &l3->slabs_free)
2997 slabp = list_entry(entry, struct slab, list);
2998 check_spinlock_acquired_node(cachep, nodeid);
2999 check_slabp(cachep, slabp);
3001 STATS_INC_NODEALLOCS(cachep);
3002 STATS_INC_ACTIVE(cachep);
3003 STATS_SET_HIGH(cachep);
3005 BUG_ON(slabp->inuse == cachep->num);
3007 obj = slab_get_obj(cachep, slabp, nodeid);
3008 check_slabp(cachep, slabp);
3010 /* move slabp to correct slabp list: */
3011 list_del(&slabp->list);
3013 if (slabp->free == BUFCTL_END)
3014 list_add(&slabp->list, &l3->slabs_full);
3016 list_add(&slabp->list, &l3->slabs_partial);
3018 spin_unlock(&l3->list_lock);
3022 spin_unlock(&l3->list_lock);
3023 x = cache_grow(cachep, flags, nodeid);
3035 * Caller needs to acquire correct kmem_list's list_lock
3037 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3041 struct kmem_list3 *l3;
3043 for (i = 0; i < nr_objects; i++) {
3044 void *objp = objpp[i];
3047 slabp = virt_to_slab(objp);
3048 l3 = cachep->nodelists[node];
3049 list_del(&slabp->list);
3050 check_spinlock_acquired_node(cachep, node);
3051 check_slabp(cachep, slabp);
3052 slab_put_obj(cachep, slabp, objp, node);
3053 STATS_DEC_ACTIVE(cachep);
3055 check_slabp(cachep, slabp);
3057 /* fixup slab chains */
3058 if (slabp->inuse == 0) {
3059 if (l3->free_objects > l3->free_limit) {
3060 l3->free_objects -= cachep->num;
3061 slab_destroy(cachep, slabp);
3063 list_add(&slabp->list, &l3->slabs_free);
3066 /* Unconditionally move a slab to the end of the
3067 * partial list on free - maximum time for the
3068 * other objects to be freed, too.
3070 list_add_tail(&slabp->list, &l3->slabs_partial);
3075 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3078 struct kmem_list3 *l3;
3079 int node = numa_node_id();
3081 batchcount = ac->batchcount;
3083 BUG_ON(!batchcount || batchcount > ac->avail);
3086 l3 = cachep->nodelists[node];
3087 spin_lock(&l3->list_lock);
3089 struct array_cache *shared_array = l3->shared;
3090 int max = shared_array->limit - shared_array->avail;
3092 if (batchcount > max)
3094 memcpy(&(shared_array->entry[shared_array->avail]),
3095 ac->entry, sizeof(void *) * batchcount);
3096 shared_array->avail += batchcount;
3101 free_block(cachep, ac->entry, batchcount, node);
3106 struct list_head *p;
3108 p = l3->slabs_free.next;
3109 while (p != &(l3->slabs_free)) {
3112 slabp = list_entry(p, struct slab, list);
3113 BUG_ON(slabp->inuse);
3118 STATS_SET_FREEABLE(cachep, i);
3121 spin_unlock(&l3->list_lock);
3122 ac->avail -= batchcount;
3123 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3127 * Release an obj back to its cache. If the obj has a constructed state, it must
3128 * be in this state _before_ it is released. Called with disabled ints.
3130 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3132 struct array_cache *ac = cpu_cache_get(cachep);
3135 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3137 if (cache_free_alien(cachep, objp))
3140 if (likely(ac->avail < ac->limit)) {
3141 STATS_INC_FREEHIT(cachep);
3142 ac->entry[ac->avail++] = objp;
3145 STATS_INC_FREEMISS(cachep);
3146 cache_flusharray(cachep, ac);
3147 ac->entry[ac->avail++] = objp;
3152 * kmem_cache_alloc - Allocate an object
3153 * @cachep: The cache to allocate from.
3154 * @flags: See kmalloc().
3156 * Allocate an object from this cache. The flags are only relevant
3157 * if the cache has no available objects.
3159 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3161 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3163 EXPORT_SYMBOL(kmem_cache_alloc);
3166 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3167 * @cache: The cache to allocate from.
3168 * @flags: See kmalloc().
3170 * Allocate an object from this cache and set the allocated memory to zero.
3171 * The flags are only relevant if the cache has no available objects.
3173 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3175 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3177 memset(ret, 0, obj_size(cache));
3180 EXPORT_SYMBOL(kmem_cache_zalloc);
3183 * kmem_ptr_validate - check if an untrusted pointer might
3185 * @cachep: the cache we're checking against
3186 * @ptr: pointer to validate
3188 * This verifies that the untrusted pointer looks sane:
3189 * it is _not_ a guarantee that the pointer is actually
3190 * part of the slab cache in question, but it at least
3191 * validates that the pointer can be dereferenced and
3192 * looks half-way sane.
3194 * Currently only used for dentry validation.
3196 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3198 unsigned long addr = (unsigned long)ptr;
3199 unsigned long min_addr = PAGE_OFFSET;
3200 unsigned long align_mask = BYTES_PER_WORD - 1;
3201 unsigned long size = cachep->buffer_size;
3204 if (unlikely(addr < min_addr))
3206 if (unlikely(addr > (unsigned long)high_memory - size))
3208 if (unlikely(addr & align_mask))
3210 if (unlikely(!kern_addr_valid(addr)))
3212 if (unlikely(!kern_addr_valid(addr + size - 1)))
3214 page = virt_to_page(ptr);
3215 if (unlikely(!PageSlab(page)))
3217 if (unlikely(page_get_cache(page) != cachep))
3226 * kmem_cache_alloc_node - Allocate an object on the specified node
3227 * @cachep: The cache to allocate from.
3228 * @flags: See kmalloc().
3229 * @nodeid: node number of the target node.
3231 * Identical to kmem_cache_alloc, except that this function is slow
3232 * and can sleep. And it will allocate memory on the given node, which
3233 * can improve the performance for cpu bound structures.
3234 * New and improved: it will now make sure that the object gets
3235 * put on the correct node list so that there is no false sharing.
3237 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3239 unsigned long save_flags;
3242 cache_alloc_debugcheck_before(cachep, flags);
3243 local_irq_save(save_flags);
3245 if (nodeid == -1 || nodeid == numa_node_id() ||
3246 !cachep->nodelists[nodeid])
3247 ptr = ____cache_alloc(cachep, flags);
3249 ptr = __cache_alloc_node(cachep, flags, nodeid);
3250 local_irq_restore(save_flags);
3252 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3253 __builtin_return_address(0));
3257 EXPORT_SYMBOL(kmem_cache_alloc_node);
3259 void *kmalloc_node(size_t size, gfp_t flags, int node)
3261 struct kmem_cache *cachep;
3263 cachep = kmem_find_general_cachep(size, flags);
3264 if (unlikely(cachep == NULL))
3266 return kmem_cache_alloc_node(cachep, flags, node);
3268 EXPORT_SYMBOL(kmalloc_node);
3272 * __do_kmalloc - allocate memory
3273 * @size: how many bytes of memory are required.
3274 * @flags: the type of memory to allocate (see kmalloc).
3275 * @caller: function caller for debug tracking of the caller
3277 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3280 struct kmem_cache *cachep;
3282 /* If you want to save a few bytes .text space: replace
3284 * Then kmalloc uses the uninlined functions instead of the inline
3287 cachep = __find_general_cachep(size, flags);
3288 if (unlikely(cachep == NULL))
3290 return __cache_alloc(cachep, flags, caller);
3294 void *__kmalloc(size_t size, gfp_t flags)
3296 #ifndef CONFIG_DEBUG_SLAB
3297 return __do_kmalloc(size, flags, NULL);
3299 return __do_kmalloc(size, flags, __builtin_return_address(0));
3302 EXPORT_SYMBOL(__kmalloc);
3304 #ifdef CONFIG_DEBUG_SLAB
3305 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3307 return __do_kmalloc(size, flags, caller);
3309 EXPORT_SYMBOL(__kmalloc_track_caller);
3314 * __alloc_percpu - allocate one copy of the object for every present
3315 * cpu in the system, zeroing them.
3316 * Objects should be dereferenced using the per_cpu_ptr macro only.
3318 * @size: how many bytes of memory are required.
3320 void *__alloc_percpu(size_t size)
3323 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3329 * Cannot use for_each_online_cpu since a cpu may come online
3330 * and we have no way of figuring out how to fix the array
3331 * that we have allocated then....
3333 for_each_possible_cpu(i) {
3334 int node = cpu_to_node(i);
3336 if (node_online(node))
3337 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3339 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3341 if (!pdata->ptrs[i])
3343 memset(pdata->ptrs[i], 0, size);
3346 /* Catch derefs w/o wrappers */
3347 return (void *)(~(unsigned long)pdata);
3351 if (!cpu_possible(i))
3353 kfree(pdata->ptrs[i]);
3358 EXPORT_SYMBOL(__alloc_percpu);
3362 * kmem_cache_free - Deallocate an object
3363 * @cachep: The cache the allocation was from.
3364 * @objp: The previously allocated object.
3366 * Free an object which was previously allocated from this
3369 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3371 unsigned long flags;
3373 BUG_ON(virt_to_cache(objp) != cachep);
3375 local_irq_save(flags);
3376 __cache_free(cachep, objp);
3377 local_irq_restore(flags);
3379 EXPORT_SYMBOL(kmem_cache_free);
3382 * kfree - free previously allocated memory
3383 * @objp: pointer returned by kmalloc.
3385 * If @objp is NULL, no operation is performed.
3387 * Don't free memory not originally allocated by kmalloc()
3388 * or you will run into trouble.
3390 void kfree(const void *objp)
3392 struct kmem_cache *c;
3393 unsigned long flags;
3395 if (unlikely(!objp))
3397 local_irq_save(flags);
3398 kfree_debugcheck(objp);
3399 c = virt_to_cache(objp);
3400 mutex_debug_check_no_locks_freed(objp, obj_size(c));
3401 __cache_free(c, (void *)objp);
3402 local_irq_restore(flags);
3404 EXPORT_SYMBOL(kfree);
3408 * free_percpu - free previously allocated percpu memory
3409 * @objp: pointer returned by alloc_percpu.
3411 * Don't free memory not originally allocated by alloc_percpu()
3412 * The complemented objp is to check for that.
3414 void free_percpu(const void *objp)
3417 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3420 * We allocate for all cpus so we cannot use for online cpu here.
3422 for_each_possible_cpu(i)
3426 EXPORT_SYMBOL(free_percpu);
3429 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3431 return obj_size(cachep);
3433 EXPORT_SYMBOL(kmem_cache_size);
3435 const char *kmem_cache_name(struct kmem_cache *cachep)
3437 return cachep->name;
3439 EXPORT_SYMBOL_GPL(kmem_cache_name);
3442 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3444 static int alloc_kmemlist(struct kmem_cache *cachep)
3447 struct kmem_list3 *l3;
3448 struct array_cache *new_shared;
3449 struct array_cache **new_alien;
3451 for_each_online_node(node) {
3453 new_alien = alloc_alien_cache(node, cachep->limit);
3457 new_shared = alloc_arraycache(node,
3458 cachep->shared*cachep->batchcount,
3461 free_alien_cache(new_alien);
3465 l3 = cachep->nodelists[node];
3467 struct array_cache *shared = l3->shared;
3469 spin_lock_irq(&l3->list_lock);
3472 free_block(cachep, shared->entry,
3473 shared->avail, node);
3475 l3->shared = new_shared;
3477 l3->alien = new_alien;
3480 l3->free_limit = (1 + nr_cpus_node(node)) *
3481 cachep->batchcount + cachep->num;
3482 spin_unlock_irq(&l3->list_lock);
3484 free_alien_cache(new_alien);
3487 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3489 free_alien_cache(new_alien);
3494 kmem_list3_init(l3);
3495 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3496 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3497 l3->shared = new_shared;
3498 l3->alien = new_alien;
3499 l3->free_limit = (1 + nr_cpus_node(node)) *
3500 cachep->batchcount + cachep->num;
3501 cachep->nodelists[node] = l3;
3506 if (!cachep->next.next) {
3507 /* Cache is not active yet. Roll back what we did */
3510 if (cachep->nodelists[node]) {
3511 l3 = cachep->nodelists[node];
3514 free_alien_cache(l3->alien);
3516 cachep->nodelists[node] = NULL;
3524 struct ccupdate_struct {
3525 struct kmem_cache *cachep;
3526 struct array_cache *new[NR_CPUS];
3529 static void do_ccupdate_local(void *info)
3531 struct ccupdate_struct *new = info;
3532 struct array_cache *old;
3535 old = cpu_cache_get(new->cachep);
3537 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3538 new->new[smp_processor_id()] = old;
3541 /* Always called with the cache_chain_mutex held */
3542 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3543 int batchcount, int shared)
3545 struct ccupdate_struct new;
3548 memset(&new.new, 0, sizeof(new.new));
3549 for_each_online_cpu(i) {
3550 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3553 for (i--; i >= 0; i--)
3558 new.cachep = cachep;
3560 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
3563 cachep->batchcount = batchcount;
3564 cachep->limit = limit;
3565 cachep->shared = shared;
3567 for_each_online_cpu(i) {
3568 struct array_cache *ccold = new.new[i];
3571 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3572 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3573 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3577 err = alloc_kmemlist(cachep);
3579 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3580 cachep->name, -err);
3586 /* Called with cache_chain_mutex held always */
3587 static void enable_cpucache(struct kmem_cache *cachep)
3593 * The head array serves three purposes:
3594 * - create a LIFO ordering, i.e. return objects that are cache-warm
3595 * - reduce the number of spinlock operations.
3596 * - reduce the number of linked list operations on the slab and
3597 * bufctl chains: array operations are cheaper.
3598 * The numbers are guessed, we should auto-tune as described by
3601 if (cachep->buffer_size > 131072)
3603 else if (cachep->buffer_size > PAGE_SIZE)
3605 else if (cachep->buffer_size > 1024)
3607 else if (cachep->buffer_size > 256)
3613 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3614 * allocation behaviour: Most allocs on one cpu, most free operations
3615 * on another cpu. For these cases, an efficient object passing between
3616 * cpus is necessary. This is provided by a shared array. The array
3617 * replaces Bonwick's magazine layer.
3618 * On uniprocessor, it's functionally equivalent (but less efficient)
3619 * to a larger limit. Thus disabled by default.
3623 if (cachep->buffer_size <= PAGE_SIZE)
3629 * With debugging enabled, large batchcount lead to excessively long
3630 * periods with disabled local interrupts. Limit the batchcount
3635 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3637 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3638 cachep->name, -err);
3642 * Drain an array if it contains any elements taking the l3 lock only if
3643 * necessary. Note that the l3 listlock also protects the array_cache
3644 * if drain_array() is used on the shared array.
3646 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3647 struct array_cache *ac, int force, int node)
3651 if (!ac || !ac->avail)
3653 if (ac->touched && !force) {
3656 spin_lock_irq(&l3->list_lock);
3658 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3659 if (tofree > ac->avail)
3660 tofree = (ac->avail + 1) / 2;
3661 free_block(cachep, ac->entry, tofree, node);
3662 ac->avail -= tofree;
3663 memmove(ac->entry, &(ac->entry[tofree]),
3664 sizeof(void *) * ac->avail);
3666 spin_unlock_irq(&l3->list_lock);
3671 * cache_reap - Reclaim memory from caches.
3672 * @unused: unused parameter
3674 * Called from workqueue/eventd every few seconds.
3676 * - clear the per-cpu caches for this CPU.
3677 * - return freeable pages to the main free memory pool.
3679 * If we cannot acquire the cache chain mutex then just give up - we'll try
3680 * again on the next iteration.
3682 static void cache_reap(void *unused)
3684 struct kmem_cache *searchp;
3685 struct kmem_list3 *l3;
3686 int node = numa_node_id();
3688 if (!mutex_trylock(&cache_chain_mutex)) {
3689 /* Give up. Setup the next iteration. */
3690 schedule_delayed_work(&__get_cpu_var(reap_work),
3695 list_for_each_entry(searchp, &cache_chain, next) {
3696 struct list_head *p;
3703 * We only take the l3 lock if absolutely necessary and we
3704 * have established with reasonable certainty that
3705 * we can do some work if the lock was obtained.
3707 l3 = searchp->nodelists[node];
3709 reap_alien(searchp, l3);
3711 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
3714 * These are racy checks but it does not matter
3715 * if we skip one check or scan twice.
3717 if (time_after(l3->next_reap, jiffies))
3720 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3722 drain_array(searchp, l3, l3->shared, 0, node);
3724 if (l3->free_touched) {
3725 l3->free_touched = 0;
3729 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3733 * Do not lock if there are no free blocks.
3735 if (list_empty(&l3->slabs_free))
3738 spin_lock_irq(&l3->list_lock);
3739 p = l3->slabs_free.next;
3740 if (p == &(l3->slabs_free)) {
3741 spin_unlock_irq(&l3->list_lock);
3745 slabp = list_entry(p, struct slab, list);
3746 BUG_ON(slabp->inuse);
3747 list_del(&slabp->list);
3748 STATS_INC_REAPED(searchp);
3751 * Safe to drop the lock. The slab is no longer linked
3752 * to the cache. searchp cannot disappear, we hold
3755 l3->free_objects -= searchp->num;
3756 spin_unlock_irq(&l3->list_lock);
3757 slab_destroy(searchp, slabp);
3758 } while (--tofree > 0);
3763 mutex_unlock(&cache_chain_mutex);
3765 /* Set up the next iteration */
3766 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3769 #ifdef CONFIG_PROC_FS
3771 static void print_slabinfo_header(struct seq_file *m)
3774 * Output format version, so at least we can change it
3775 * without _too_ many complaints.
3778 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3780 seq_puts(m, "slabinfo - version: 2.1\n");
3782 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3783 "<objperslab> <pagesperslab>");
3784 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3785 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3787 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3788 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3789 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3794 static void *s_start(struct seq_file *m, loff_t *pos)
3797 struct list_head *p;
3799 mutex_lock(&cache_chain_mutex);
3801 print_slabinfo_header(m);
3802 p = cache_chain.next;
3805 if (p == &cache_chain)
3808 return list_entry(p, struct kmem_cache, next);
3811 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3813 struct kmem_cache *cachep = p;
3815 return cachep->next.next == &cache_chain ?
3816 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3819 static void s_stop(struct seq_file *m, void *p)
3821 mutex_unlock(&cache_chain_mutex);
3824 static int s_show(struct seq_file *m, void *p)
3826 struct kmem_cache *cachep = p;
3828 unsigned long active_objs;
3829 unsigned long num_objs;
3830 unsigned long active_slabs = 0;
3831 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3835 struct kmem_list3 *l3;
3839 for_each_online_node(node) {
3840 l3 = cachep->nodelists[node];
3845 spin_lock_irq(&l3->list_lock);
3847 list_for_each_entry(slabp, &l3->slabs_full, list) {
3848 if (slabp->inuse != cachep->num && !error)
3849 error = "slabs_full accounting error";
3850 active_objs += cachep->num;
3853 list_for_each_entry(slabp, &l3->slabs_partial, list) {
3854 if (slabp->inuse == cachep->num && !error)
3855 error = "slabs_partial inuse accounting error";
3856 if (!slabp->inuse && !error)
3857 error = "slabs_partial/inuse accounting error";
3858 active_objs += slabp->inuse;
3861 list_for_each_entry(slabp, &l3->slabs_free, list) {
3862 if (slabp->inuse && !error)
3863 error = "slabs_free/inuse accounting error";
3866 free_objects += l3->free_objects;
3868 shared_avail += l3->shared->avail;
3870 spin_unlock_irq(&l3->list_lock);
3872 num_slabs += active_slabs;
3873 num_objs = num_slabs * cachep->num;
3874 if (num_objs - active_objs != free_objects && !error)
3875 error = "free_objects accounting error";
3877 name = cachep->name;
3879 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3881 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3882 name, active_objs, num_objs, cachep->buffer_size,
3883 cachep->num, (1 << cachep->gfporder));
3884 seq_printf(m, " : tunables %4u %4u %4u",
3885 cachep->limit, cachep->batchcount, cachep->shared);
3886 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3887 active_slabs, num_slabs, shared_avail);
3890 unsigned long high = cachep->high_mark;
3891 unsigned long allocs = cachep->num_allocations;
3892 unsigned long grown = cachep->grown;
3893 unsigned long reaped = cachep->reaped;
3894 unsigned long errors = cachep->errors;
3895 unsigned long max_freeable = cachep->max_freeable;
3896 unsigned long node_allocs = cachep->node_allocs;
3897 unsigned long node_frees = cachep->node_frees;
3898 unsigned long overflows = cachep->node_overflow;
3900 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3901 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
3902 reaped, errors, max_freeable, node_allocs,
3903 node_frees, overflows);
3907 unsigned long allochit = atomic_read(&cachep->allochit);
3908 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3909 unsigned long freehit = atomic_read(&cachep->freehit);
3910 unsigned long freemiss = atomic_read(&cachep->freemiss);
3912 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3913 allochit, allocmiss, freehit, freemiss);
3921 * slabinfo_op - iterator that generates /proc/slabinfo
3930 * num-pages-per-slab
3931 * + further values on SMP and with statistics enabled
3934 struct seq_operations slabinfo_op = {
3941 #define MAX_SLABINFO_WRITE 128
3943 * slabinfo_write - Tuning for the slab allocator
3945 * @buffer: user buffer
3946 * @count: data length
3949 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3950 size_t count, loff_t *ppos)
3952 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3953 int limit, batchcount, shared, res;
3954 struct kmem_cache *cachep;
3956 if (count > MAX_SLABINFO_WRITE)
3958 if (copy_from_user(&kbuf, buffer, count))
3960 kbuf[MAX_SLABINFO_WRITE] = '\0';
3962 tmp = strchr(kbuf, ' ');
3967 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3970 /* Find the cache in the chain of caches. */
3971 mutex_lock(&cache_chain_mutex);
3973 list_for_each_entry(cachep, &cache_chain, next) {
3974 if (!strcmp(cachep->name, kbuf)) {
3975 if (limit < 1 || batchcount < 1 ||
3976 batchcount > limit || shared < 0) {
3979 res = do_tune_cpucache(cachep, limit,
3980 batchcount, shared);
3985 mutex_unlock(&cache_chain_mutex);
3991 #ifdef CONFIG_DEBUG_SLAB_LEAK
3993 static void *leaks_start(struct seq_file *m, loff_t *pos)
3996 struct list_head *p;
3998 mutex_lock(&cache_chain_mutex);
3999 p = cache_chain.next;
4002 if (p == &cache_chain)
4005 return list_entry(p, struct kmem_cache, next);
4008 static inline int add_caller(unsigned long *n, unsigned long v)
4018 unsigned long *q = p + 2 * i;
4032 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4038 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4044 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4045 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4047 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4052 static void show_symbol(struct seq_file *m, unsigned long address)
4054 #ifdef CONFIG_KALLSYMS
4057 unsigned long offset, size;
4058 char namebuf[KSYM_NAME_LEN+1];
4060 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4063 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4065 seq_printf(m, " [%s]", modname);
4069 seq_printf(m, "%p", (void *)address);
4072 static int leaks_show(struct seq_file *m, void *p)
4074 struct kmem_cache *cachep = p;
4076 struct kmem_list3 *l3;
4078 unsigned long *n = m->private;
4082 if (!(cachep->flags & SLAB_STORE_USER))
4084 if (!(cachep->flags & SLAB_RED_ZONE))
4087 /* OK, we can do it */
4091 for_each_online_node(node) {
4092 l3 = cachep->nodelists[node];
4097 spin_lock_irq(&l3->list_lock);
4099 list_for_each_entry(slabp, &l3->slabs_full, list)
4100 handle_slab(n, cachep, slabp);
4101 list_for_each_entry(slabp, &l3->slabs_partial, list)
4102 handle_slab(n, cachep, slabp);
4103 spin_unlock_irq(&l3->list_lock);
4105 name = cachep->name;
4107 /* Increase the buffer size */
4108 mutex_unlock(&cache_chain_mutex);
4109 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4111 /* Too bad, we are really out */
4113 mutex_lock(&cache_chain_mutex);
4116 *(unsigned long *)m->private = n[0] * 2;
4118 mutex_lock(&cache_chain_mutex);
4119 /* Now make sure this entry will be retried */
4123 for (i = 0; i < n[1]; i++) {
4124 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4125 show_symbol(m, n[2*i+2]);
4131 struct seq_operations slabstats_op = {
4132 .start = leaks_start,
4141 * ksize - get the actual amount of memory allocated for a given object
4142 * @objp: Pointer to the object
4144 * kmalloc may internally round up allocations and return more memory
4145 * than requested. ksize() can be used to determine the actual amount of
4146 * memory allocated. The caller may use this additional memory, even though
4147 * a smaller amount of memory was initially specified with the kmalloc call.
4148 * The caller must guarantee that objp points to a valid object previously
4149 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4150 * must not be freed during the duration of the call.
4152 unsigned int ksize(const void *objp)
4154 if (unlikely(objp == NULL))
4157 return obj_size(virt_to_cache(objp));